text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1703.02873v1 [cs.SE] 7 Mar 2017
Redundancy Suppression In Time-Aware Dynamic Binary
Instrumentation
Pansy Arafa
Hany Kashif
University of Waterloo
University of Waterloo
University of Waterloo
[email protected]
[email protected]
[email protected]
ABSTRACT
Software tracing techniques are well-established and used
by instrumentation tools to extract run-time information
for program analysis and debugging. Dynamic binary instrumentation as one tool instruments program binaries to
extract information. Unfortunately, instrumentation causes
perturbation that is unacceptable for time-sensitive applications. Consequently we developed DIME*, a tool for dynamic binary instrumentation that considers timing constraints. DIME* uses Pin and a rate-based server approach
to extract information only as long as user-specified constraints are maintained. Due to the large amount of redundancies in program traces, DIME* reduces the instrumentation overhead by one to three orders of magnitude compared
to native Pin while extracting up to 99% of the information.
We instrument VLC and PostgreSQL to demonstrate the
usability of DIME*.
1.
INTRODUCTION
Program profiling is essential for analyzing program performance and understanding run-time behavior [36, 6, 30].
A program profiler extracts program information during execution; such as memory access patterns, code coverage, register usage, and dynamic call context trees. Other examples
include collection of run-time statistics on instruction usage, or cache hits and misses. In order to extract this runtime information, the profiler instruments the program. Instrumentation naturally causes perturbation to the program
under analysis. Such perturbation can result in erroneous
performance-analysis data or inaccurate program profile.
Instrumentation of performance-sensitive applications such
as video players must not only preserve functional correctness, but also performance constraints. Thus, time-sensitive
applications require specialized program-profiling tools in
order to honor their timing constraints. Generally, there
exist two instrumentation approaches; hardware based and
software based. Hardware-based tracing methods [23, 29]
cause significant perturbation to the traced program [26].
Also, these methods collect low-level data and, hence, require higher-level support to provide traces at a higherlevel of abstraction [24, 25]. Software-based instrumentation
methods [22, 17] insert code in the original program, either
statically or dynamically, to enable tracing, which results
in modifying the program’s timing behavior. Consequently,
recent line of research focuses on time-aware instrumentation techniques that respect the timing constraints of the
program [12, 16, 15, 4].
Time-aware instrumentation extracts program informa-
Sebastian Fischmeister
tion while preserving both functional and timing properties. Fischmeister and Lam [12] propose a time-aware static
instrumentation technique that preserves the worst-case execution time (WCET) of the instrumented program. Kashif
et al. [16] apply program transformation techniques to increase the effectiveness of time-aware instrumentation. In [15],
Kashif et al. introduce INSTEP; a static instrumentation
framework for preserving extra-functional properties. All
these approaches are impractical for large code bases with library dependencies as they require static analysis and worstcase execution time analysis. They also do not support
multi-threaded applications. In [4], Arafa et al. propose
DIME; a time-aware dynamic binary instrumentation framework using rate-based resource allocation. DIME limits the
instrumentation time to a predefined budget per each time
period. If the instrumentation consumes the budget before
the time period ends, DIME will disable the instrumentation
till the beginning of the next time period. While DIME reduces overhead, it only extracts partial (incomplete) traces
due to disabling instrumentation on budget consumption.
In this paper, we introduce DIME*, which offers the same
advantages as the original version DIME, but provides higher
instrumentation coverage. The idea is to attempt to trace
only untraced information in successive runs of the program
under analysis. Even with multiple runs, running a program
multiple times on top of DIME* is less time-consuming than
native Pin. We investigate different implementations and
compare them qualitatively. We quantitatively evaluate the
performance of DIME* compared to native Pin using the
SPEC benchmarks [14]. Our results show that DIME* extracts up to 99% of the tracing information while reducing
the overhead of native Pin by up to three orders of magnitude. For instance, dealII, a SPEC benchmark [14], originally runs for four minutes without any instrumentation.
Extracting the branch profile of dealII using native Pin takes
nine days of CPU time, while it only takes a maximum of
30 minutes using DIME*. We also apply our approach to
VLC [3] and PostgreSQL [2] as case studies to demonstrate
the scalability and applicability of DIME*.
2.
BACKGROUND
This section provides a basic overview of time-aware instrumentation, dynamic binary instrumentation, and the basics of DIME.
2.1
Time-aware Instrumentation
Instrumentation is the process of inserting extra code inside a program to extract information during execution. In-
strumentation can be static or dynamic. Static instrumentation means inserting the instrumentation code before running the program, whereas dynamic techniques inserts the
instrumentation code during the program execution. Timeaware instrumentation [12, 16, 15, 4] is a mechanism that
preserves functional correctness and timing properties when
instrumenting programs. The static technique in [12] instruments a program at code locations that do not modify
the program’s WCET and at the same time preserves the
program’s original behavior. Static time-aware instrumentation shifts the program’s execution time profile towards
its deadline. The authors in [16] apply code transformation techniques to the program under analysis to increase
instrumentation coverage. They duplicate or create basic
blocks in the program to increase the locations at which instrumentation code can be inserted while preserving timing
constraints. The authors in [15] introduce INSTEP; an instrumentation framework for preserving multiple competing
extra-functional properties. INSTEP uses cost models and
constraints of the extra-functional properties in addition to
the user’s instrumentation intent to transform the input program into an instrumented program that respects the specified constraints. All the mentioned work on time-aware
instrumentation is based on static source-code instrumentation techniques. Static source-code instrumentation requires
performing WCET analysis of the input program to guide
the placement of instrumentation code inside the input program. It also needs WCET analysis after instrumenting the
program to validate that timing constraints are met. Static
instrumentation techniques are sound and effective, but the
need for running WCET analysis before and after instrumentation reduces the applicability to only hard real-time
applications where WCET analysis is common. Additionally, static instrumentation requires the availability of the
source code including all library dependencies. For example,
the VLC media player [3] has approximately 600 000 lines of
code and uses libraries with more than three million lines of
code. Thus, it is impractical to statically analyze the source
code of a multi-threaded application like VLC along with its
library dependencies. In [4], Arafa et al. propose DIME; a
time-aware dynamic binary instrumentation framework.
2.2
Dynamic Binary Instrumentation
Dynamic binary instrumentation (DBI) instruments the
program’s binary during execution. Unlike static instrumentation, DBI does not require preprocessing of the program
under analysis. On the other hand, DBI incurs higher runtime overhead compared to static instrumentation since DBI
decides the placement of instrumentation code at runtime.
Pin [21] is a DBI instrumentation framework that provides
a cross-platform API for building program-profiling tools.
Pin has lower run-time overhead than that of the other DBI
frameworks like DynamoRio [8], and Valgrind [28]. Pin is
easily extensible and transparent i.e., it maintains the same
instruction and data addresses, and the same register and
memory values compared to the original program. Pin can
trace statically unknown indirect-jump targets, dynamically
generated code, and dynamically loaded libraries. It can
also handle mixed code and data, and variable-length instructions. To build an analysis tool using Pin (pintool),
two types of routines should be implemented. The analysis
routine contains the code to be inserted in the program during execution, whereas the instrumentation routine decides
on the locations of inserting the analysis-routine calls. The
analysis routine is the main source of Pin’s overhead which
varies according to the invocation frequency of the analysis routines and their complexity. On the other hand, the
dynamic compilation and the execution of the instrumentation routine represent a minor source of run-time overhead. Pin is known for its efficiency [21, 4]; it uses a just-intime (JIT) compiler to insert and optimize instrumentation
code. The unit of compilation is the trace; a straight-line
sequence of instructions that have a single entry point and
may have multiple exits. When the program starts execution, Pin compiles the first trace and generates a modified
one which is almost identical to the original. The modified trace enables Pin to regain control when needed. Pin
transfers control to the generated trace, then regains control
when a branch exits the trace. Afterwards, Pin compiles the
new trace and continues execution. Whenever the JIT compiler fetches some code to compile, the pintool is allowed to
instrument the code before compilation. Pin saves the compiled code and its instrumentation in a cache in case it gets
re-executed [35, 31, 21]. Pin supports different granularities for both the instrumentation routine and the analysis
routine, e.g., trace, routine, and instruction granularities.
The instrumentation-routine granularity defines when Pin
should execute the instrumentation routine. Similarly, the
granularity of the analysis routine tells Pin where to insert
the analysis-routine calls. DIME [4] is implemented as an
extension to Pin [21].
2.3
DIME
DIME [4] is a dynamic binary time-aware instrumentation tool that respects the timing properties of the program.
DIME, as a tool for instrumenting soft real-time applications, is practical, scalable, and supports multi-threaded applications. It guarantees less run-time overhead compared
to Pin especially for the profiling tools with heavy-weight
analysis routines. DIME uses rate-based resource allocation
to limit the instrumentation time to a pre-specified budget
B per time period T . Instrumentation is enabled (i.e. allowed to execute) for a total of tins time units in every time
period T . The total instrumentation time tins per period
T should not exceed the instrumentation budget B. If the
budget is fully consumed before the end of the time period
T , DIME will disable instrumentation. At the beginning of
the next period T , the budget resets to B time units and
the instrumentation is re-enabled. This process repeats until the program terminates. Specifically, DIME limits the
execution time of the analysis routine to the budget B [4].
The reason is that the analysis routine is the main source of
run-time overhead, and the overhead of the instrumentation
routine is negligible [21].
Figure 1 further illustrates the rate-based DBI approach.
The X-axis represents the program’s execution time tprog
and the Y-axis shows the remaining instrumentation budget
(B −tins ). The program starts execution in the DBI-enabled
state i.e., full instrumentation budget is available. In the
first time period [0, T ) of the program’s execution, instrumentation code executes and reduces the available budget.
Once the instrumentation has fully consumed the budget,
the framework will switch to the DBI-disabled state and will
prevent further instrumentation. At time T , the budget is
reset, and the framework returns back to the DBI-enabled
state [4].
Remaining Budget
(B - tins)
DBI
Enabled
Budget Reset,
DBI Re-enabled
For each inst rum enta tion point {
InsertCall ( budget_check ) ;
i f ( version == V_BASE ) {
// c h e c k s w i t c h i n g t o V INSTRUMENT
I n s e r t V e r s i o n C a s e ( 1 , V_INSTRUMENT ) ; }
e l s e i f ( version == V_INSTRUMENT ) {
// c h e c k s w i t c h i n g t o V BASE
I n s e r t V e r s i o n C a s e ( 0 , V_BASE ) ; }
switch ( version ) {
case V_BASE :
break ; //Do Nothing
case V_INSTRUMENT :
...
InsertCall ( a na l ys i s_ r ou t in e ) ;
break ; }
2
B −
4
6
0 −
8
T
2T
3T
Program Execution Time (tprog)
DBI
Disabled
Figure 1: Rate-based DBI [4].
10
12
14
16
}
Listing 1: Instrumentation rtn of Trace Version [4].
In [4], the authors studied three implementations of DIME;
Trace Version, Trace Version Conditional , and Trace Conditional . Trace Version checks for budget availability at
each instrumentation point (through an extra analysis routine) and makes use of Pin’s trace versioning APIs to enable
and disable instrumentation. The implementation of Trace
Version will be discussed later in this section. Trace Version
Conditional has a similar implementation to Trace Version
but with a reduced frequency of budget checking in the DBIdisabled state. In Trace Conditional , the instrumentation
routine checks the available budget using a simple if statement in the instrumentation routine. Although Trace Version has the highest budget-checking overhead (compared to
the other implementations), it strictly respects the budget
and has full budget utilization. Trace Version Conditional ,
compared to Trace Version, has lower budget-checking overhead but lower budget utilization. Trace Conditional has the
lowest budget-checking overhead which leads to loose budget respect. Also, Trace Conditional has another source of
high run-time overhead which is the overshoots. An overshoot occurs when the instrumentation time exceeds the instrumentation budget. DIME shows an average reduction
in overhead by 12, 7, and 3 folds, for the three implementations consecutively, compared to native Pin. Moreover,
Trace Version provides good instrumentation coverage when
the instrumentation-budget value is reasonable (10% for example). Instrumentation coverage is the ratio of the output
of DIME to that of native Pin. For the reasons mentioned,
we focus on Trace Version in this paper although the suggested approaches work for the other two implementations.
Trace Version uses Pin’s trace versioning APIs to check
for budget at each instrumentation point, and enable/disable instrumentation accordingly. These APIs allow dynamic switching between multiple types (versions) of instrumentation at runtime. There is two instrumentation versions
in DIME*; V INSTRUMENT refers to enabled instrumentation, and V BASE when instrumentation is disabled. When
Pin switches versions, it creates a new trace starting from
the current instruction. Listing 1 shows a pseudocode outline of Trace Version (favoring readability over optimality).
To clarify, let Trace 1 be the sequence of instructions in
Listing 2. Assume that Trace 1 has version = V INSTRUMENT,
and “For each instrumentation point” means “For each instruction”
in this example. Pin calls the instrumentation routine at
every trace. The instrumentation routine inserts an inlined
call to budget check() before every instruction in the trace.
According to the switch case in Listing 1, the instrumentation routine inserts a call to the analysis routine before every
instruction in the trace. The API InsertVersionCase() guarantees that the execution of the inserted analysis routine will
occur, only if the output of budget check() matches the ID of
the current version (which is 1 in this case). Directly after
the execution of the instrumentation routine, budget check(),
that is inserted before the first instruction, is executed. Assume that its output is 1 which means that the budget is
currently larger than zero. In this case, Pin will execute the
analysis routine that is inserted before the first instruction.
Afterwards, Pin executes budget check() that is inserted before the second instruction. Assume that the budget is now
fully consumed, so budget check() returns 0. Since the output
mismatches the ID of the current trace, Pin will switch version to V BASE (i.e. disable instrumentation). Accordingly,
Pin will create a new trace Trace 2, with version = V BASE,
starting from the second instruction. Pin will then execute
the instrumentation routine to instrument Trace 2 according to its version. Also, the analysis routine inserted before
the second and the third instruction will be ignored i.e., not
executed.
mov eax , dword ptr [ rsp+0x30 ]
and eax , 0 x10000000
mov dword ptr [ rsp+0x3c ] , eax
Listing 2: Example (1) of a trace.
3.
REDUNDANCY SUPPRESSION IN DIME*
Both native Pin and DIME generate tracing information
that might contain redundancies. Naik et al. in [27] pointed
out that, in many applications, output traces contain many
redundancies. For example, in many analysis tools, instrumenting each instruction once is sufficient since the information extracted is the same regardless of the number of
times the instruction is executed or instrumented. In other
words, an instruction can be executed and, hence, instrumented several times. Each time, the same information is
extracted causing redundancies. Examples of these tools
include branch profiling tools used for extracting code coverage and memory profiling tools used for building memory
access patterns.
Since DIME extracts partial tracing information, it can
obtain higher instrumentation coverage through the avoidance of tracing redundant information. To respect the timing properties of a program, DIME disables the instrumentation when the instrumentation budget is consumed. Accordingly, DIME generates partial tracing information compared
to native Pin. In other words, there exists a trade-off between the instrumentation budget and the instrumentation
coverage. Hence, multiple runs of DIME are required to increase the instrumentation coverage and optimally achieve
full coverage. From a performance point of view, it is preferable to minimize the number of required runs. For DIME,
this implies obtaining the maximum possible coverage from
each single run without violating the timing constraints. Accordingly, DIME should avoid tracing redundant instrumentation. In this paper, we specifically focus on the type of
analysis tools that do not require tracing redundant information. DIME* should utilize the available instrumentation
budget for extracting unique (non-redundant) information.
To prohibit redundant instrumentation, DIME* should be
able to identify the instrumented code regions. The minimum piece of information needed to identify a code region
is the starting address. Thus, the basic idea is to enable
DIME* to save the starting addresses of instrumented code
regions in a log, and DIME* should then check the log before instrumenting a new code region. For the approach to
be efficient, DIME* should:
• Prevent re-instrumentation of a code region in the current run and all subsequent runs of the program under
analysis.
• Avoid increasing run-time overhead.
• Avoid creating large-sized logs which increase DIME*’s
memory consumption. Searching a large log may also
result in increased run-time overhead.
Both steps, saving to the log and searching it, take place in
the instrumentation routine, so its overhead is expected to
be negligible. We avoided adding these steps to the analysis
routine which is the main source of overhead in Pin.
In what follows, we discuss our approach for suppressing
redundancies in DIME*.
3.1
Granularity of Logged Code Regions
The first design aspect that we discuss is the granularity
of the code regions to be recorded in the log. We can log
addresses of code regions either at the instruction level or
the trace level. It is inefficient to log the address of each
instrumented instruction, since
1. This requires frequent access to the log which adds to
run-time overhead.
2. This results in a large log size which consumes memory.
3. This leads to searching a large log which can delay
program execution and add to the run-time overhead.
An alternative to logging instruction address is to log addresses at a coarser granularity, the trace level. The instrumentation routine analyzes traces to insert analysis-routine
calls. Recall that a trace is a sequence of program instructions that has a single entry point and may have multiple
exit points. If Pin detects a jump to an instruction in the
middle of a trace, Pin will create a new trace beginning at
the target instruction. So, the instructions inside a trace
are always in series i.e., uninterrupted by instructions from
another trace. Thus, DIME* will save the trace starting address in addition to the length of the instrumented portion in
the trace (hTrace Address, Trace Lengthi). Specifically, DIME*
will save the relative starting address of the trace with respect to the trace’s image. This guarantees that saved addresses are deterministic between successive runs (especially
for the traces of shared libraries). On the other side, as
mentioned earlier in Section 2.3, trace version switching can
cause Pin to create a new trace. Thus, some trace addresses
might only exist in a subset of the runs.
3.2
Efficient Log Search
The second design aspect is saving the trace addresses
and the trace lengths in a manner that allows for efficient
searching of the log. In this section, we propose three approaches for saving trace addresses and length. We compare
among them qualitatively and quantitatively deriving unexpected results. We choose one of these approaches to provide DIME* with the capability of suppressing redundant
instrumentations.
3.2.1
Hash-Table Log
The first approach uses a hash-table as the log for saving instrumented traces. In this approach, DIME* saves
the trace address to identify instrumented traces. Whenever DIME* instruments a trace, it adds the trace’s address
to the hash-table. Also, before instrumenting any trace,
DIME* searches for the trace address in the hash-table.
Let A be the current trace and B be a trace in the log
L; Then, DIME* will only instrument A, iff (address(A) 6=
address(B))∀B ∈ L. The advantages of using a hash-table
for logging traces are:
• Fast logging (average case: constant; worst case: linear
in the hash-table size).
• Fast searching (average case: constant; worst case: linear in the hash-table size).
• Low number of false negatives (false negatives will occur if DIME* prohibits instrumentation of an uninstrumented trace). For instance, let A be the current
trace, where address(A) = 100 and length(A)=80. If
the log contains trace B, where address(B) = 100 and
length(B)=20, DIME* will prohibit the instrumentation of A. Thus, the instructions in the address range
120 to 180 will not be instrumented in any run.
The disadvantage of this approach is that:
• Using a hash-table enables DIME* to only compare
trace addresses while ignoring the trace length. This
results in false positives. A false positive will occur if
DIME* allows instrumentation of a previously instrumented trace. For example, let A be the current trace,
where address(A) = 150 and length(A)=20. DIME*
may fail to find A in the log, although the log contains
trace B such that address(B) = 100 and length(B)
= 80. This means that trace A is previously instrumented as a part of trace B. Note that one trace being part of another happens due to the creation of new
traces through version switching as explained earlier.
Section 3.2.4 presents experimental results that support the
listed advantages and disadvantages for the three approaches.
3.2.2
BST Log
The second approach is using a binary search tree (BST)
to log the addresses of the instrumented traces along with
their lengths. Being sorted, the BST facilitates jumping to
a specific range of addresses. When DIME* instruments
a trace, it adds the trace to the log such that the trace
address is the key and the trace length is the value. Before instrumenting a trace, DIME* searches the BST using
the trace address. If not found, DIME* will jump to the
log-entry that has the first smaller trace address compared
to the current trace address. DIME* will then decide if
the current address lies within the trace of the discovered
log-entry. Let A be the current trace and B be a trace
in the log L. DIME* will not instrument A, if ∃B ∈ L
s.t. (address(B) ≤ address(A) < address(B) + length(B)).
The advantage of using a BST for logging traces is:
• Less false positives compared to the hash-table approach due to considering the lengths of the logged
traces.
The disadvantages of this approach, on the other hand, are:
• Slower than the hash-table approach in the average
case; the complexity of both saving and searching is
log(N ).
• Relatively high false negatives. Consider the following
example. Assume the current trace is A = h100, 200i
(i.e., address(A)=100 and length(A)=200), and the log
entry B = h50, 80i in the log L. This approach will
prevent instrumenting trace A since its starting address lies within the log entry B. This, however, will
consequently prevent DIME* from instrumenting the
uninstrumented portion of trace A i.e., from address
130 to address 300.
Additionally, after the program execution and before saving the log to a file for use in subsequent runs, DIME*
merges directly consecutive traces leading to a smaller log
size. For example, if the log contains two log entries h100, 50i
and h150, 50i, DIME* will merge them into one log entry
h100, 200i. Merging log entries decreases the log size and,
therefore, reduces the search time leading to less run-time
overhead in subsequent runs of DIME*.
3.2.3
Merger BST
The third approach utilizes a BST as well, but it addresses
the second disadvantage of the previous approach. Using
this approach, DIME* will prohibit instrumentation, only
if the whole current trace is part of a log entry. Otherwise,
DIME* allows instrumentation and merges the current trace
with the log entry if needed. Let A be the current trace and
B be a trace in the log L. DIME* will not instrument A,
if ∃B ∈ L s.t. (address(B) ≤ address(A) < address(B) +
length(B)∧address(A)+length(A) < address(B)+length(B)).
For example, let the current trace be A = h100, 200i, and the
log entry contains B = h50, 80i. In this approach, DIME*
instruments trace A. Afterwards, DIME* merges trace A
and B into one log entry h50, 250i to avoid redundancies in
the log. The advantages of this approach are:
• Less false negatives, compared to the BST approach.
The disadvantages are:
• Slower than the hash-table approach in the average
case; the complexity of both saving and searching is
log(N ).
• Higher false positives than the BST approach since it
allows re-instrumentation of some portions of a trace.
As mentioned before, Section 3.2.4 provides experi-
mentation that discusses these observations.
Note that a trace address and length will be saved in the
log, only if the trace is actually instrumented i.e., the trace’s
version is V INSTRUMENT.
3.2.4
Evaluation of the Different Approaches
In this section, we describe our experiments to evaluate
the three log-search approaches. According to the results,
we choose which approach to extend DIME with the feature
of suppressing redundant instrumentation output.
We experiment with two SPEC2006 C benchmark [14]
programs (lbm and mcf ) for three runs. We later, in Section 4, use more SPEC benchmarks to evaluate the performance of modified DIME* over up to eight runs. The
experiment uses a branch-profiling analysis tool which has
a heavy-weight analysis routine. It prints out the jump,
call, and return instructions in addition to the source address and the destination address. The tool is based on the
branch target addr pintool that is available as a part of the
Pin’s kit v2.12-56759. The pintool is modified to extract the
branch profile of the whole program instead of only a part
of it. A branch profiler is useful for investigating the code
coverage of a program. Each experiment, in this section and
Section 4, is conducted once due to the very long execution
time when instrumenting the benchmarks on top of native
Pin. For example, povray benchmark originally executes in
2.5 minutes, but on top of native Pin, it consumes four days
of CPU time. The execution time of the other benchmark
programs, on top of Native Pin and DIME*, will be discussed in Section 4. However, the runs of each DIME* experiment can be considered as repetitions since all the runs
operate identically. In the initialization stage, there exists
one minor difference between the first run and the following ones. The first run starts with an empty log while the
following runs read the log from a file before launching and
instrumenting the program.
The evaluation of the proposed approaches is based on the
following metrics:
1. Instrumentation Coverage: the ratio of the instrumentation output of DIME* to that of native Pin.
Note that we consider only unique (non-redundant)
traces. Increasing the instrumentation coverage is the
main objective of the proposed approaches.
2. False positives: a false positive will occur if DIME*
permits the instrumentation of a previously instrumented trace or trace-portion. This metric measures
the ratio of false positives to the total number of instrumented traces in the current run. The ratio of
false positives indicates the efficiency of the log searching approach in identifying previously instrumented
traces. As the ratio of false positives decreases, the
budget utilization increases and the number of required
runs to maintain high coverage decreases.
3. False negatives: a false negative will take place when
DIME* refuses to instrument a trace which was not
instrumented before. The metric measures the ratio
of false negatives to the total number of traces that
got rejected by DIME* in the current run. This value
includes the trace portions as well i.e., a part of the
trace is instrumented but the other part is not. As the
false negatives increase, the ability of the approach to
maintain high coverage decreases.
4. Slow-down factor of the instrumented program:
the ratio of the execution time of the dynamically instrumented benchmark to the execution time of the natively running benchmark. This metric examines the
ability of DIME* to reduce run-time overhead, compared to Pin, while saving to and searching the log to
suppress instrumentation redundancy. It also checks
if the three approaches introduce different run-time
overhead. Low run-time overhead is essential for the
instrumentation of time-sensitive systems as discussed
before.
5. Overshoots: an overshoot will occur when actual
instrumentation time exceeds the budget. The magnitude of the overshoots shows how strictly modified
DIME* respects the instrumentation budget.
test r12 , r12
setnz byte ptr [ rsp+0x3b ]
jnz 0 x7ffff7de26a8
mov eax , dword ptr [ rsp+0x30 ]
and eax , 0 x10000000
...
call 0 x7ffff7df2850
another log-entry. Hence, DIME* allows instrumentation
of this trace. Assume that enough instrumentation budget
is available to instrument all the instructions in the trace.
Thus, the trace address along with its length are saved in
the log as h34192, 62i. The instrumentation-routine inserts
analysis-routine calls for all the instructions in the trace.
Assume that in the first run of DIME*, the first three instructions only execute and a jump (through jnz) occurs. In
the second run, DIME* (BST and BST-Merger) prohibits
instrumentation for the trace 34206 (starting from the mov
instruction) since it lies inside the logged trace h34192, 62i.
Accordingly, no information is extracted starting from the
address 34206 since these instructions do not execute in the
first run and DIME* prevents their instrumentation in the
following runs. Although, the program runs with the same
inputs, this can occur due to non-deterministic execution of
some shared libraries such as libc and the Linux loader. For
such shared libraries, execution can slightly change according to the processor state. In such cases, the BST and the
Merger BST approaches fail to extract some information,
thus decreasing their instrumentation coverage.
Listing 3: Example (2) of a trace.
Run−1
Run−2
Run−3
Coverage Percentage (%)
100
80
False Positives (%)
40
60
40
Run−1
Run−2
Run−3
30
20
10
0
Htable
lbm
Merger
lbm
Htable
mcf
BST
mcf
Merger
mcf
Approach − Benchmark
20
0
BST
lbm
(a) False Positives
Htable
lbm
BST
lbm
Merger
lbm
Htable
mcf
BST
mcf
Merger
mcf
60
Approach − Benchmark
Figure 2 shows the instrumentation coverage of the three
approaches with lbm and mcf consecutively. The hash-table
approach guarantees the highest instrumentation coverage.
After three runs, it achieves 97% of the instrumentation coverage of native Pin for lbm benchmark, and 98% for mcf. The
coverage of BST is 83% and 80% for lbm and mcf consecutively. Finally, BST-Merger generates 90% and 88% of the
instrumentation output for lbm and mcf consecutively. The
low ratio of false negatives of the hash-table approach is one
reason for achieving the highest coverage. The hash-table
approach is a conservative one which favors re-instrumenting
some trace portions over uninstrumenting them. Also, some
scenarios lead to a decreased instrumentation coverage for
the BST and Merger BST approaches compared to the hashtable approach. As mentioned previously, a trace can have
multiple exits e.g. can include multiple jump instructions.
Listing 3 is an example of a trace with multiple exits (contains jnz and call instructions). Assume the starting address
of the trace is 34192 and the trace length is 62. Assume
DIME* encounters this trace for the first time, and cannot
find the address 34192 in the log as a key or as a part of
False Negitives (%)
Figure 2: Instrumentation Coverage
Run−1
Run−2
Run−3
50
40
30
20
10
0
Htable
lbm
BST
lbm
Merger
lbm
Htable
mcf
BST
mcf
Merger
mcf
Approach − Benchmark
(b) False Negatives
Figure 3: Ratio of false positives and ratio of false negatives
The ratio of false positives is shown in Figure 3a. The
hash-table approach has the highest ratio with both lbm and
mcf benchmarks. The BST-Merger approach has moderate
values of false positives, whereas BST has approximately
zero false positives. This means that BST accurately identifies the previously instrumented traces and efficiently utilizes the budget to instrument other traces. On the other
hand, BST has a high ratio of false negatives, as shown in
Figure 3b, which is an undesirable feature. BST-Merger
sustains approximately zero false negatives, and hash-table
has negligible values of false negatives ratio. The scenarios discussed in Section 3.2 explain the values in Figures 3a
and 3b. Note that false-negatives ratio is more critical than
false positives. Although false positives cause instrumentation redundancies, it is safer. False negatives prevent code
portions from being instrumented in any run which can dramatically decrease the instrumentation coverage. The results show that there is a trade-off between false positives
and false negatives. In such case, we prefer the approach
that maintains low ratio of false negatives even if it has high
ratio of false positives. Thus, the hash-table and the BSTMerger approaches outperform the BST one.
Slow Down Factor
68x
Run−1
Run−2
Run−3
67x
2x
the three DIME* modifications are suitable for dynamically
instrumenting time-sensitive systems. Comparing the three
approaches to each other, none of them shows a significant
overhead-decrease over the others. Consequently, run-time
overhead is not a factor that differentiates among the three
approaches.
Figure 5 shows the overshoots’ magnitude for the three
proposed approaches of DIME* over the execution time of
the mcf benchmark while instrumenting it using the DIME*
version of the branch-profiling pintool. The three approaches
respect the instrumentation budget; the values for the most
frequent overshoots lie below 4 microseconds. There is not
significant differences in the overshoots’ magnitudes among
the three approaches of DIME*. Thus, this metric is also
not a factor to favor one approach over the others.
Non-intuitively, the evaluation metrics reveal that the simplest approach, which is the hash-table one, results in the
best instrumentation coverage results. Moreover, the hashtable approach provides low values of false negatives, maintains low run-time overhead, and respects the instrumentation budget. Accordingly, we choose the hash-table approach to support instrumentation-redundancy suppression
in DIME*.
1x
4.
0
Native Pin
Htable
BST
Merger
Approach
4.1
(a) lbm benchmark
Slow Down Factor
251x
Run−1
Run−2
Run−3
250x
2x
1x
0
Native Pin
Htable
BST
Merger
Approach
(b) mcf benchmark
Figure 4: Slow-down factors.
Figure 4a presents the slow-down factors of native Pin and
the proposed approaches of DIME* with lbm benchmark. On
top of native Pin, lbm runs 68x slower than the native execution. On the other hand, the hash-table approach of Pin
achieves a slow-down of 1.4x, 1x, and 1x for three consecutive runs. The overhead of the BST approach is 1.5x, 1x,
and 1x, while that of BST-Merger is 1.5x, 1x, 1x for three
runs. In Figure 4b, native Pin slows down the execution
251x with mcf benchmark. Whereas, the slow-down factors
of the hash-table approach for the three runs are 1.7x, 1x,
and 1x. These of the BST are 1.6x, 1.7x, and 1x, and BSTMerger achieves slow-down of 1.9x, 1x, and 1x. To sum up,
DIME* reduces the run-time overhead by at least 45 folds
for lbm and 132 folds for mcf. These numbers reveal that
the three modifications of DIME* are able to dramatically
reduce the run-time overhead of native Pin. Thus, all of
PERFORMANCE EVALUATION
This section presents the experimentation of DIME* and
discusses its performance.
Experimental Setup
We experiment with the SPEC2006 benchmark suite [14]
including C and C++ integer and floating point programs.
The experiments run on top of a workstation hosting a quadcore i7 3.4 GHz Intel processors with 8 MB of cache, and
16 GB of RAM. The operating system is an Ubuntu 12.04
patched with a real-time kernel v3.2.0-23 to convert Linux
into a fully preemptible kernel. To obtain accurate results,
we inhibit task migration between cores and lock core speed
to the maximum frequency. The experimentation environment also maintains a real-time scheduling policy and priority. These modifications guarantee accurate results for
performance evaluation and are not mandatory for DIME*
correctness. The instrumentation objective is extracting the
branch-profile of the program. We use the Pin kit v2.1256759 and gcc v4.6.3. The time-period parameter is set to
one second and the instrumentation budget is set to 0.1 seconds for all the benchmark programs. Note that the experimentation included more benchmarks, however, these extra
benchmarks are not reported since the execution time on
top of native Pin exceeded twenty days.
We evaluate the performance of DIME* using the following metrics:
1. Instrumentation Coverage: the ratio of the amount
of unique extracted traces by DIME* to the amount
of those extracted by native Pin. This metric demonstrates the ability of DIME* to extract the maximum
possible amount of information while respecting the
instrumentation budget.
2. Slow-down factor of the instrumented program:
the ratio of the execution time of the dynamically instrumented benchmark to its original execution time
(without instrumentation). This metric reflects the reduction of run-time overhead of DIME* compared to
native Pin.
12
●
8
●
●
●
4
2
0
● ●
●
●
●●
●●● ●●●
●
● ●
●
●
●● ●
●
●●
●●
●
●
●●
● ●● ●
●
●●
●
● ●●●
● ●●●
●●
●
●●
●● ●
●
●●●● ●
●
●
●
●● ●
●●
●●
●●
●
●
●●●●●●
●
●●
●
●
● ●
●●
●●
●●
●
● ●
●
● ●● ●●
●●
●●
●●●●
●●●
●●
●
●
●
●●●
●●
●
●●
●
●
●●
●●
●● ●●●●
●
●●●●
●
●
●
●
●
● ●
● ●●●●●
●
●
●
●
●
●
●●
●●●●
●
●
●●●●
●●●
●
●
●
●● ●
●
● ●
●
●●
●● ●●
●
●●●●●
●●
● ●
●
● ●●●●
●
●●●
●
●
●
●
●● ●●
●
●
●
●●●
●
●
●
●
●
●
●
●●● ● ●
●●
●●●
●
●●●●
●
●
●●
●●
● ●●●
● ●●
●●
●● ●●
●
●●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●●●●● ●
●●
●
● ● ● ●● ●●
●
●●● ●
●●
●
●
●●
●
●● ●●
●●
●
● ●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●●
● ●●
●
●
●●● ●
●
● ● ●●
●●●●
●● ●
●●●
●
●●●
●●●●
●●
●
●
●● ●
●
●● ●●
●●
●
●
●●●
●
●● ●
●
●●●
●●
●●●
●●
●●
●●
●
●●● ●
●
●
●
●
●
●●
●●●
●●
● ●●●● ●
●●
●●
●●
●
●●●●
●
●
●● ●
●
●
●●
● ●●●●●
●
●
●
●
●
●
●
●● ●
●
●●
●●
●●
● ●●
●
●
●
●●
●
●●
●●
●●●
●
●
●
●●
●
●●
●
●
●●●
●●
●●
●●●
●
●
●●
●●●
●● ●●
● ●●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ● ●
●
●
●● ● ●
●
●●
●
● ● ● ● ●●
● ●●
●
0
100
200
300
400
500
600
8
Overshoot (usec)
●
6
Overshoot (usec)
Overshoot (usec)
10
8
6
4
2
0
700
●
●
●
●●
● ●● ●● ● ●
●●●●
● ●
●●●
●
● ● ●●
●● ●
●
●●
●●●
●● ●●●●●
● ●●
●●●
●●●●●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●●
●
●●● ●
●
●●
●
●●
●
●●
●
●● ●●●●
●
●
●
●
●●
●●
●
●●●
●
●●●
●
●●
●●
●
●●
●●●
●●
●●
●
●
●
●●●
●
●●
●
●
●
●
●● ●
●
●
●
●● ●
●
●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●●●● ●
●
●●●
●●
●
●● ●● ●
●●
●
●
●
●●
●●
●
● ●●
●
●
●●
●●
●
●
●●
●
●
●●
●
●●
●● ●
●
●
●
●●
●
●●
●
●●●
●
●
●●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
● ●
●
●
●
●
●
●
●
●
●●●
●
●●●
●●
●
●●
●●
●
●
●●
●
●
●●●
●
●●●●
●
●
●
●●
●
● ●●
●●
●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●● ● ●
●●
●●
●
●
●●
●
●●
●●
●● ●●
●●
●●●
●
●●●
●
●●●
●
●●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●●●●●●
●
●
●
●● ●●
●●● ●
● ●●
●
● ●●
● ●
●
●●●●●●
●●●●●●
●
●●
●
●●●●
●
●
●
●●●
●● ●
●●●●
●●●
● ●
●●●
●●●●●
●●●
●
●●
●●●●
●●
●●
●
●●●
●
●
●
●
●
●●
● ●●●●●●
●●●●●
●● ●●
●
●
0
100
Execution Time
200
300
400
500
600
●
6
●
●
●
●●
●
●
●●●
●
●
●
●
●
● ●
●
●●● ●●●●
●
●
●●
●●
●
●●
●
●●●
● ●●
●●
●●●●●
●●●
●●
● ● ●●●●●
●●
● ●●●●● ●
●● ●●
●
●
● ●●
●
●
●
●●
●
●
●
●
●●
●●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
● ●●
●●
●
● ● ●●
●●●
●● ●
●
●●
●
●●
●
●
●
●●● ●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●●● ●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●●●●
●● ● ●
●●
● ●●●
●
●● ●●● ●●
●
●
●
●
●
●● ● ●●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●● ● ● ●● ●●
● ● ●●
●
●●
●
●●
● ● ● ●●
●
●
● ●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●● ●
●●
●● ●● ●
●●
●
●●
●●
●●
● ●
●
●●●●●●
●
●
●
●
●
●
●●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●● ●
●● ●●● ● ●●
●●●
● ● ●● ●●
●
●●
●
●
●
●
● ●●
●
●● ●
●
●●●
●
●
●●
●
●●
●●
●
●
●● ●
●●●●
●●●
●●
● ●
●
●●●●●
● ●●
●●
●●●●
●●
●
●
●
●
●
●
●●●
●●●
●●
●
●●
● ● ●●●
● ●
●●●●
●●
● ●●
●
●●
●●●●
●●
● ●●
●●
●
●
●
●●●
●
●●
●●
●
●●
●●
●●
●
●
●
●
●●
●
●
●●●
●
● ●●
●
●●
●●● ●
●
●
●●●●
●
●●● ●
● ●●
● ●
●
●●●
●
●
●●●●
●
●
●
● ●●●●●●●
●
●●
●
4
2
0
700
0
100
200
Execution Time
300
400
500
600
700
Execution Time
(a) Hash-table
(b) BST
(c) Merger BST
Figure 5: Overshoots of the three approaches in the first run with the mcf benchmark.
4.2
Experimental Results
Native Pin
100
Coverage Percentage (%)
90
80
70
60
795x
1733x
2971x
2438x
1738x
67x
251x
208x
Slow−down Factor
Run−1
Run−2
Run−3
Run−4
Run−5
Run−6
Run−7
Run−8
DIME*
8
6
50
4
40
2
0
30
namd
xalan
deal
povray
sphinx3
lbm
mcf
milc
Benchmark
20
Figure 7: Slow-down factors of native Pin and DIME*.
10
milc
mcf
lbm
sphinx3
povray
dealII
xalancbmk
namd
0
Benchmark
Figure 6: Instrumentation Coverage
DIME* is capable of maintaining high instrumentation
coverage. Figure 6 shows the ratio of the amount of the
extracted instrumentation output through multiple runs of
DIME* with respect to that extracted by native Pin. DIME*
is capable of extracting 97% and 99% of the instrumentation output in four runs for dealII and mcf benchmarks consequently. In five runs, DIME* generates a coverage of 99%
for both namd and lbm. It extracts 92% and 97% in the sixth
run for milc and povray consequently. DIME* also extracts
98% when instrumenting xalancbmk for seven runs, and 91%
of the instrumentation output of sphinx3 after eight runs.
DIME* outperforms native Pin in terms of run-time overhead. It reduces the run-time overhead of dynamic instrumentation by one to three orders of magnitude compared to
native Pin. Figure 7 shows the slow-down factor of native
Pin, and the average slow-down factor of the eight runs of
DIME* for each benchmark program. Note that the average
slow-down factor is the geometric mean. Native Pin dramatically slows down the program execution. The average slow
down of native Pin is 706x the original benchmark execution, with maximum value of 2971x and minimum value of
67x. The benchmarks’ continuous execution time on top of
native Pin ranges from four hours to nine days with average
of four days. On the other hand, the benchmarks on top of
DIME* takes from three to 42 minutes with average of nine
minutes. The average slow-down of DIME* is 1.5x, with
maximum value of 8x and minimum value of 1x.
Running a program multiple times on top of DIME* is
less time-consuming than native Pin. More importantly,
DIME* allows the program to maintain its timing properties
(or extra-functional properties in general). Respecting such
properties is essential for time-sensitive systems. Additionally, multiple runs of DIME* provide very high instrumentation coverage compared to native Pin.
5.
CASE STUDIES
This section demonstrates the scalability and the practicality of DIME* through two case studies.
5.1
VLC Case Study
This case study reveals the usability of DIME* for instrumenting multi-threaded time-sensitive software. VLC [3],
developed by the VideoLan organization, is a multi-platform
media player. It consists of approximately 600 000 lines of
code and uses dependencies of approximately three million
lines of code. In this case study, we aim to extract the call
context tree of VLC v2.0.8 while playing a high definition,
29.97 fps, 720x480, 1 Mbps bitrate video. The call context
tree is useful for performance analysis and applying program
optimizations [32]. The tool, used in this case study, is based
on the DebugTrace pintool which is available in Pin’s v2.12
kit. We use the tool to extract the call traces, then we use
the call traces to build a call context tree. The platform is an
Ubuntu 12.04 machine hosting a quad-core i7 2.6 GHz Intel
processors with 8 GB of RAM. For DIME*, we set the time
period to one second and the budget to 10% which enables
VLC to run the video smoothly.
Table 1 shows the number of video blocks that VLC decodes for viewing frames, and the extracted percentage of
the nodes and the edges of the call-context tree. The table
lists the number of decoded blocks without instrumentation,
with native Pin, and with DIME*. It also shows the coverage of DIME* compared to Pin.
Our conclusion is that DIME* dramatically decreases the
run-time overhead of dynamic instrumentation and maintains almost the same instrumentation coverage. On top of
native Pin, VLC fails to maintain a continuous video playback. The video is unwatchable; only 37 video blocks (out
of 594) are decoded. VLC generates multiple errors of dropping video frames due to very slow processing. On the other
hand, DIME* enables VLC to play the video continuously.
VLC decodes 574, 594, and 594 video blocks (out of 594) for
the three runs consecutively. DIME* extracts 98% of the
call context tree nodes and 96% of the edges.
the quality of service (QoS) increases. The processing time
reported excludes connection-establishing time. The table
also shows the coverage of DIME* with respect to that of
Pin.
In this case study, DIME* extracts 97% of the analysis data while reducing the run-time overhead by 68 folds
compared to native Pin. Originally, PostgreSQL processes
all the transactions in 46 seconds. Native Pin causes a
slowdown of 96x; PostgreSQL executes the same number
of transactions in 1.2 hours. On the other hand, DIME*
maintains slowdown of only 1.4x, 1.4x, and 1.3x in three
runs consecutively.
Decoded Blocks Nodes(%) Edges(%)
Original
594
N/A
N/A
Native Pin
37
100%
100%
DIME* (run 1)
574
95.7%
93.6%
(run 2)
594
98%
96.2%
(run 3)
594
98.2%
96.5%
Table 1: Results for the VLC case study
A program can be instrumented at the source code level
either automatically or manually. In automatic instrumentation, a tool parses the program, may generate a CFG, and
eventually insert instrumentation points. Multiple works [12,
16, 15] investigate static source-code time-aware instrumentation tools. On the other hand, manual instrumentation
requires that the developer specifies the instrumentation locations [33]. Manual instrumentation is highly flexible, but
the induced effect of instrumentation on the timing behavior
is hard to estimate by the developer.
Some instrumentation tools are also capable of inserting
instrumentation points to binary executables, either statically or dynamically. QPT [19], EEL [20], and ATOM [34]
are examples of static binary instrumentation tools. Static
instrumentation is based on static analysis and, hence, cannot react to application changes at run time. Dynamic binary instrumentation, on the other hand, does not require
any pre-processing of the program under analysis.
Example of dynamic binary instrumentation tools that
use code transformation during program execution include
Dyninst [10] and Vulcan [11]. Most of these instrumentation tools, however, modify the native behavior of the program under analysis [9]. Other tools have software code
caches and are able to dynamically compile binaries such as
Pin [21], DynamoRIO [8], and Valgrind [28].
Superpin [38] is a parallelized version of Pin that reduces
the overhead of dynamic binary instrumentation by utilizing
multiple cores. The performance gain is limited by the number of cores and shared memory structures. Yu et al. [39]
introduce MT-profiler, a multi-threaded profiling framework
built on top of DynamoRIO to instrument parallel programs.
Zhao et al. [40] introduce PiPA, a pipelined profiling and
analysis tool for parallelizing dynamic instrumentation on
multi-core platforms. Moseley et al [25] present a probebased application monitor which is injected to the original
application under analysis. The approach we present in this
work can make use of such parallelization techniques to reduce overhead as well.
Upton et al. [37] introduce a buffering system for Pin to reduce overhead. Kumar et al. [18] present an instrumentation
5.2
PostgreSQL Case Study
DIME* is useful for extracting sufficient analysis information while maintaining high quality of service (QoS). PostgreSQL [2] is a powerful object-relational database management system that supports SQL standards. PostgreSQL 9.1
consists of approximately one million lines of code. When an
application sends a database request to PostgreSQL, a connection will be established, and a parser will check the query
syntax and create a query tree. Then, an optimizer generates a query plan from the tree and sends the plan to the
executor. Finally, the executor steps through the commands
in the query plan to retrieve the required information or to
make the required updates to the database. Pgbench [1] is
a benchmarking tool for testing the performance of PostgreSQL. It runs a sequence of transactions multiple times in
multiple database sessions. The objective of this case study
is the extraction of the branch-profile of PostgreSQL while
processing a total of 800 000 transactions. PostgreSQL runs
16 database sessions where each session consists of 50 000
transactions. The case study runs on an Ubuntu 12.04 platform hosting a quad-core i7 2.6 GHz Intel processors with
8 GB of RAM. We set the budget of DIME* to 7% and
the time period to one second. These values enable DIME*
to extract sufficient information while keeping high performance of PostgreSQL.
Table 2 shows the performance of PostgreSQL while (1)
running natively, (2) running on top of native Pin, and (3)
running on top of DIME*. The total time consumed to process the 800 000 transactions is an indication of the performance. As the processing time increases, the degradation in
Total Time (sec) Coverage(%)
Original
46
N/A
Native Pin
4419
100%
DIME* (run 1)
65
42%
(run 2)
65
75%
(run 3)
61
97%
Table 2: Results for the PostgreSQL case study
6.
RELATED WORK
optimizer for program monitoring and profiling. While these
instrumentation approaches focus on reducing the overhead
of instrumentation, they are unaware of the program’s timing requirements. They, however, are also orthogonal to our
proposed approach.
The work by Arnold and Ryder [5] is the most relevant to
reducing the overhead of dynamic instrumentation following
DIME. The authors’ approach involves duplicating code regions and using counter-based sampling to switch between
the instrumented and non-instrumented versions of the code.
Code duplication results in a large increase in code space.
Since event-based sampling only samples events according
to their frequency of occurrence, this results in a reduced
instrumentation overhead. However, as explained in [7],
event-based sampling can result in sampling bursts which
can cause high degradation in performance Other samplingbased approaches are also used for performance optimizations [13]. These approaches either apply optimizations specific to the instrumentation objective or use compiler-specific
information to perform optimizations.
7.
CONCLUSION
Dynamic binary instrumentation tools are used to analyze program behavior and extract debugging information
at runtime. Most applications, however, cannot leverage
existing tools for analysis or debugging purposes due to the
performance degradation resulting from the instrumentation
process. Time-sensitive applications, for instance, require a
bound on the execution time overhead.
In this paper, we propose DIME* that performs instrumentation only within a user-specified instrumentation budget. DIME* suppresses logging of redundant information to
reduce the instrumentation overhead. The results show a
reduction in the overhead of the instrumentation process by
one to three orders of magnitude compared to native Pin
while achieving up to 99% of the instrumentation coverage.
DIME* was able to extract 97% of the call context tree of the
VLC video player while playing a high definition video. VLC
fails to provide a watchable video while being instrumented
using native Pin. DIME* was also used for the branch profiling of the PostgreSQL database management system and
was able to extract 97% of the instrumentation information
in three runs. DIME* extracts this information in less than
two minutes per run while native Pin takes 1.2 hours to
extract the information. The presented case studies show
the scalability of DIME* and its ability to limit the instrumentation overhead while achieving a high instrumentation
coverage.
8.
REFERENCES
[1] Pgbench: Benchmarking Tool for PostgreSQL.
http://wiki.postgresql.org/wiki/Pgbench.
[2] PostgreSQL Global Development Group.
http://www.postgresql.org/.
[3] VLC Media Player.
http://www.videolan.org/vlc/index.html.
[4] P. Arafa, H. Kashif, and S. Fischmeister. Dime:
Time-aware dynamic binary instrumentation using
rate-based resource allocation. In Proc. of the 13th
International Conference on Embedded Software
(EMSOFT), Montreal, Canada, Sept 2013.
[5] M. Arnold and B. G. Ryder. A Framework for
Reducing the Cost of Instrumented Code. In Proc. of
the ACM SIGPLAN Conf. on Programming language
design and implementation (PLDI), 2001.
[6] S. Bock, B. Childers, R. Melhem, D. Mosse, and
Y. Zhang. Analyzing the Impact of Useless
Write-backs on the Endurance and Energy
Consumption of PCM Main Memory. In Performance
Analysis of Systems and Software (ISPASS), 2011
IEEE International Symposium on, pages 56–65, 2011.
[7] B. Bonakdarpour, S. Navabpour, and S. Fischmeister.
Sampling-based Runtime Verification. In Proc. of the
17th Intl. Conf. on Formal Methods (FM), Jun. 2011.
[8] D. Bruening, T. Garnett, and S. Amarasinghe. An
Infrastructure for Adaptive Dynamic Optimization. In
Proc. of the Intl. Symp. on Code Generation and
Optimization (CGO), 2003.
[9] D. Bruening, Q. Zhao, and S. Amarasinghe.
Transparent Dynamic Instrumentation. SIGPLAN
Not., 47(7), Mar. 2012.
[10] B. Buck and J. K. Hollingsworth. An API for Runtime
Code Patching. Int. J. High Perform. Comput. Appl.,
14(4), Nov. 2000.
[11] A. Edwards, H. Vo, and A. Srivastava. Vulcan: Binary
Transformation in a Distributed Environment.
Technical report, 2001.
[12] S. Fischmeister and P. Lam. Time-Aware
Instrumentation of Embedded Software. IEEE
Transactions on Industrial Informatics, 2010.
[13] N. Froyd, J. Mellor-Crummey, and R. Fowler.
Low-overhead Call Path Profiling of Unmodified,
Optimized Code. In Proc. of the 19th Annual Intl.
Conf. on Supercomputing (ICS), 2005.
[14] J. L. Henning. SPEC CPU2000: Measuring CPU
Performance in the New Millennium. Computer, 33(7),
2000.
[15] H. Kashif, P. Arafa, and S. Fischmeister. INSTEP: A
Static Instrumentation Framework for Preserving
Extra-functional Properties. In Proc. of the 19th IEEE
Intl. Conf. on Embedded and Real-Time Computing
Systems and Applications (RTCSA), Aug. 2013.
[16] H. Kashif and S. Fischmeister. Program
transformation for time-aware instrumentation. In
Proc. of the 17th IEEE Intl. Conf. on Emerging
Technologies & Factory Automation (ETFA), Sep.
2012.
[17] M. Kim, M. Viswanathan, S. Kannan, I. Lee, and
O. Sokolsky. Java-MaC: A Run-Time Assurance
Approach for Java Programs. Form. Methods Syst.
Des., 2004.
[18] N. Kumar, B. R. Childers, and M. L. Soffa. Low
Overhead Program Monitoring and Profiling. In Proc.
of the 6th ACM SIGPLAN-SIGSOFT workshop on
Program Analysis for Software Tools and Engineering
(PASTE), 2005.
[19] J. Larus. Efficient program tracing. Computer, 26(5),
1993.
[20] J. R. Larus and E. Schnarr. EEL:
Machine-Independent Executable Editing. SIGPLAN
Not., 30, 1995.
[21] C.-K. Luk, R. Cohn, R. Muth, H. Patil, A. Klauser,
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
G. Lowney, S. Wallace, V. J. Reddi, and
K. Hazelwood. Pin: Building Customized Program
Analysis Tools with Dynamic Instrumentation. In
Proc. of the ACM SIGPLAN Conf. on Programming
Language Design and Implementation (PLDI), 2005.
J. M. Mellor-Crummey and T. J. LeBlanc. A Software
Instruction Counter. In Proc. of the 3rd Intl. Conf. on
Architectural Support for Programming Languages and
Operating Systems (ASPLOS), 1989.
L. J. Moore and A. R. Moya. Non-Intrusive Debug
Technique for Embedded Programming. In Proc. of
the 14th Intl. Symp. on Software Reliability
Engineering (ISSRE), 2003.
P. Mork. Techniques for Debugging Parallel Programs.
Technical report, University of Miskolc.
T. Moseley, A. Shye, V. Reddi, D. Grunwald, and
R. Peri. Shadow Profiling: Hiding Instrumentation
Costs with Parallelism. In Intl. Symp. on Code
Generation and Optimization(CGO), Mar. 2007.
T. Mytkowicz, A. Diwan, M. Hauswirth, and
P. Sweeney. We have it Easy, but do we have it Right?
IEEE Intl. Symp. on Parallel and Distributed
Processing, 2008.
M. Naik, H. Yang, G. Castelnuovo, and M. Sagiv.
Abstractions from Tests. SIGPLAN Not., 47(1), Jan.
2012.
N. Nethercote and J. Seward. Valgrind: A Framework
for Heavyweight Dynamic Binary Instrumentation.
SIGPLAN Not., 42(6), Jun. 2007.
W. Omre. Debug and Trace for Multicore SoCs.
Technical report, ARM, 2008.
A. Rico, A. Duran, F. Cabarcas, Y. Etsion,
A. Ramirez, and M. Valero. Trace-driven Simulation
of Multithreaded Applications. In Performance
Analysis of Systems and Software (ISPASS), 2011
IEEE International Symposium on, pages 87–96, 2011.
A. Ruiz-Alvarez and K. Hazelwood. Evaluating the
Impact of Dynamic Binary Translation Systems on
Hardware Cache Performance. In IEEE Intl. Symp. on
Workload Characterization (IISWC), 2008.
M. Serrano and X. Zhuang. Building Approximate
Calling Context from Partial Call Traces. In Proc. of
the 7th Annual IEEE/ACM Intl. Symp. on Code
Generation and Optimization (CGO), 2009.
B. Simon, D. Bouvier, T.-Y. Chen, G. Lewandowski,
R. McCartney, and K. Sanders. Common Sense
Computing (Episode 4): Debugging. Computer
Science Education, 18(2), 2008.
A. Srivastava and A. Eustace. ATOM: A System for
Building Customized Program Analysis Tools.
SIGPLAN Not., 39, 1994.
G.-R. Uh, R. Cohn, B. Yadavalli, R. Peri, and
R. Ayyagari. Analyzing Dynamic Binary
Instrumentation Overhead. 2007.
D. Upton and K. Hazelwood. Finding Cool Code: An
Analysis of Source-level Causes of Temperature
Effects. In Performance Analysis of Systems and
Software (ISPASS), 2011 IEEE International
Symposium on, pages 117–118, 2011.
D. Upton, K. Hazelwood, R. Cohn, and G. Lueck.
Improving Instrumentation Speed via Buffering. In
Proc. of the Workshop on Binary Instrumentation and
Applications (WBIA), 2009.
[38] S. Wallace and K. Hazelwood. SuperPin: Parallelizing
Dynamic Instrumentation for Real-Time Performance.
In Intl. Symp. on Code Generation and Optimization
(CGO), Mar. 2007.
[39] Z. Yu, W. Zhang, and X. Tu. MT-Profiler: A Parallel
Dynamic Analysis Framework Based on Two-stage
Sampling. In Proc. of the 9th Intl. Conf. on Advanced
Parallel Processing Technologies (APPT), 2011.
[40] Q. Zhao, I. Cutcutache, and W.-F. Wong. PiPA:
Pipelined Profiling and Analysis on Multicore
Systems. ACM Trans. Archit. Code Optim., 7(3), Dec.
2010.
| 6cs.PL
|
1
On extractable shared information
Johannes Rauh∗, Pradeep Kr. Banerjee∗, Eckehard Olbrich∗ , Jürgen Jost∗, and Nils Bertschinger†
∗ Max
Planck Institute for Mathematics in the Sciences, Leipzig, Germany
{jrauh,pradeep,olbrich,jjost}@mis.mpg.de
† Frankfurt
Institute for Advanced Studies, Frankfurt, Germany
arXiv:1701.07805v3 [cs.IT] 10 Nov 2017
[email protected]
Abstract
We consider the problem of quantifying the information shared by a pair of random variables X1 , X2 about
another variable S. We propose a new measure of shared information, called extractable shared information, that is
left monotonic; that is, the information shared about S is bounded from below by the information shared about f (S)
for any function f . We show that our measure leads to a new nonnegative decomposition of the mutual information I(S; X1 X2 ) into shared, complementary and unique components. We study properties of this decomposition and
show that a left monotonic shared information is not compatible with a Blackwell interpretation of unique information.
We also discuss whether it is possible to have a decomposition in which both shared and unique information are left
monotonic.
Keywords: Information decomposition; multivariate mutual information; left monotonicity; Blackwell order
I. I NTRODUCTION
A series of recent papers have focused on the bivariate information decomposition problem [1]–[6]. Consider
three random variables S, X1 , X2 with finite alphabets S, X1 and X2 , respectively. The total information that the
pair (X1 , X2 ) convey about the target S can have aspects of shared or redundant information (conveyed by both
X1 and X2 ), of unique information (conveyed exclusively by either X1 or X2 ), and of complementary or synergistic
information (retrievable only from the the joint variable (X1 , X2 )). In general, all three kinds of information may
be present concurrently. One would like to express this by decomposing the mutual information I(S; X1 X2 ) into
a sum of nonnegative components with a well-defined operational interpretation. One possible application area is
in the neurosciences. In [7], it is argued that such a decomposition can provide a framework to analyze neural
information processing using information theory that can integrate and go beyond previous attempts.
For the general case of k finite source variables (X1 , . . . , Xk ), Williams and Beer [3] proposed the partial
information lattice framework that specifies how the total information about the target S is shared across the singleton
sources and their disjoint or overlapping coalitions. The lattice is a consequence of certain natural properties of
2
shared information (sometimes called the Williams–Beer axioms). In the bivariate case (k = 2), the decomposition
has the form
I(S; X1 X2 ) = SI(S; X1 , X2 ) + CI(S; X1 , X2 ) + U I(S; X1 \X2 ) + U I(S; X2 \X1 ),
{z
} |
{z
} |
{z
} |
{z
}
|
shared
unique (X1 wrt X2 )
complementary
(1)
unique (X2 wrt X1 )
I(S; X1 ) = SI(S; X1 , X2 ) + U I(S; X1 \X2 ),
(2)
I(S; X2 ) = SI(S; X1 , X2 ) + U I(S; X2 \X1 ),
(3)
where SI(S; X1 , X2 ), U I(S; X1 \X2 ), U I(S; X2 \X1 ), and CI(S; X1 , X2 ) are nonnegative functions that depend
continuously on the joint distribution of (S, X1 , X2 ). The difference between shared and complementary information is the familiar co-information [8] (or interaction information [9]), a symmetric generalization of the mutual
information for three variables,
CoI(S; X1 , X2 ) = I(S; X1 ) − I(S; X1 |X2 ) = SI(S; X1 , X2 ) − CI(S; X1 , X2 ).
Equations (1) to (3) leave only a single degree of freedom, i.e., it suffices to specify either a measure for SI, for
CI or for U I.
Williams and Beer not only introduced the general partial information framework, but also proposed a measure of
SI to fill this framework. While their measure has subsequently been criticized for “not measuring the right thing”
[4]–[6], there has been no successful attempt to find better measures, except for the bivariate case (k = 2) [1],
[4]. One problem seems to be the lack of a clear consensus on what an ideal measure of shared (or unique or
complementary) information should look like and what properties it should satisfy. In particular, the Williams–Beer
axioms only put crude bounds on the values of the functions SI, U I and CI. Therefore, additional axioms have
been proposed by various authors [4]–[6]. Unfortunately, some of these properties contradict each other [5], and
the question for the right axiomatic characterization is still open.
The Williams–Beer axioms do not say anything about what should happen when the target variable S undergoes
a local transformation. In this context, the following left monotonicity property was proposed in [5]:
(LM) SI(S; X1 , X2 ) ≥ SI(f (S); X1 , X2 ) for any function f .
(left monotonicity)
Left monotonicity for unique or complementary information can be defined similarly. The property captures the
intuition that shared information should only decrease if the target performs some local operation (e.g., coarse
graining) on her variable S. As argued in [2], left monotonicity of shared and unique information are indeed desirable
properties. Unfortunately, none of the measures of shared information proposed so far satisfy left monotonicity.
In this contribution, we study a construction that enforces left monotonicity. Namely, given a measure of shared
information SI, define
SI(S; X1 , X2 ) := sup
f :S→S ′
SI(f (S); X1 , X2 ),
(4)
where the supremum runs over all functions f : S → S ′ from the domain of S to an arbitrary finite set S ′ . By
construction, SI satisfies left monotonicity, and SI is the smallest function bounded from below by SI that satisfies
left monotonicity.
3
Changing the definition of shared information in the information decomposition framework Equations (1)–(3)
leads to new definitions of unique and complementary information:
∗
U I (S; X1 \X2 ) := I(S; X1 ) − SI(S; X1 , X2 ),
∗
U I (S; X2 \X1 ) := I(S; X2 ) − SI(S; X1 , X2 ),
∗
∗
∗
CI (S; X1 , X2 ) := I(S; X1 X2 ) − SI(S; X1 , X2 ) − U I (S; X1 \X2 ) − U I (S; X2 \X1 ).
∗
In general, U I (S; X1 \ X2 ) 6= U I(S; X1 \ X2 ) := supf :S→S ′ U I(f (S); X1 \ X2 ). Thus, our construction cannot
enforce left monotonicity for both U I and SI in parallel.
∗
∗
Lemma 2 shows that SI, U I and CI are nonnegative and thus define a nonnegative bivariate decomposition.
We study this decomposition in Section IV. In Theorem 1, we show that our construction is not compatible with a
decision-theoretic interpretation of unique information proposed in [1]. In Section V, we ask whether it is possible
to find an information decomposition in which both shared and unique information measures are left monotonic. Our
construction cannot directly be generalized to ensure left monotonicity of two functions simultaneously. Nevertheless,
it is possible that such a decomposition exists, and in Proposition 5, we prove bounds on the corresponding shared
information measure.
Our original motivation for the definition of SI was to find a bivariate decomposition in which the shared
information satisfies left monotonicity. However, one could also ask whether left monotonicity is a required property
of shared information, as put forward in [2]. In contrast, [4] argue that redundancy can also arise by means of a
mechanism. Applying a function to S corresponds to such a mechanism that singles out a certain aspect from S.
Even if all the Xi share nothing about the whole S, they might still share information about this aspect of S,
which means that the shared information will increase. With this intuition, we can interpret SI not as an improved
measure of shared information, but as a measure of extractable shared information, because it asks for the maximal
amount of shared information that can be extracted from S by further processing S by a local mechanism. More
generally, one can apply a similar construction to arbitrary information measures. We explore this idea in Section III
and discuss probabilistic generalizations and relations to other information measures. In Section VI, we apply our
construction to existing measures of shared information.
II. P ROPERTIES
OF I NFORMATION
D ECOMPOSITIONS
A. The Williams–Beer Axioms
Although we are mostly concerned with the case k = 2, let us first recall the three axioms that Williams and
Beer [3] proposed for a measure of shared information for arbitrarily many arguments:
(S) SI(S; X1 , . . . , Xk ) is symmetric under permutations of X1 , . . . , Xk ,
(SR) SI(S; X1 ) = I(S; X1 ),
(Symmetry)
(Self-redundancy)
(M) SI(S; X1 , . . . , Xk−1 , Xk ) ≤ SI(S; X1 , . . . , Xk−1 ),
with equality if Xi = f (Xk ) for some i < k and some function f .
(Monotonicity)
Any measure of SI satisfying these axioms is nonnegative. Moreover, the axioms imply the following:
4
(RM) SI(S; X1 , . . . , Xk ) ≥ SI(S; f1 (X1 ), . . . , fk (Xk )) for all functions f1 , . . . , fk .
(right monotonicity)
Williams and Beer also defined a function
Imin (S; X1 , . . . , Xk ) =
X
PS (s) min
s
i
nX
PXi |S (xi |s) log
xi
PS|Xi (s|xi ) o
PS (s)
(5)
and showed that Imin satisfies their axioms.
B. The C OPY example and the Identity Axiom
Let X1 , X2 be independent uniformly distributed binary random variables, and consider the copy function
C OPY (X1 , X2 ) := (X1 , X2 ). One point of criticism of Imin is the fact that X1 and X2 share Imin (C OPY (X1 , X2 ); X1 , X2 ) =
1 bit about C OPY (X1 , X2 ) according to Imin , even though they are independent. [4] argue that the shared information
about the copied pair should equal the mutual information:
(Id) SI(C OPY (X1 , X2 ); X1 , X2 ) = I(X1 ; X2 ).
(Identity)
Ref. [4] also proposed a bivariate measure of shared information that satisfies (Id). Similarly, the measures of
bivariate shared information proposed in [1] satisfies (Id). However, (Id) is incompatible with a nonnegative
information decomposition according to the Williams–Beer axioms for k ≥ 3 [2].
On the other hand, Ref. [5] uses an example from game theory to give an intuitive explanation how even
independent variables X1 and X2 can have nontrivial shared information. However, in any case the value of 1 bit
assigned by Imin is deemed to be too large.
C. The Blackwell property and property (∗)
One of the reasons that it is so difficult to find good definitions of shared, unique or synergistic information
is that a clear operational idea behind these notions is missing. Starting from an operational idea about decision
problems, Ref. [1] proposed the following property for the unique information, which we now propose to call
Blackwell property:
(BP) For a given joint distribution PSX1 X2 , U I(S; X1 \X2 ) vanishes if and only if there exists a random variable
X1′ such that S − X2 − X1′ is a Markov chain and PSX1′ = PSX1 .
(Blackwell property)
In other words, the channel S → X1 is a garbling or degradation of the channel S → X2 . Blackwell’s theorem [10]
implies that this garbling property is equivalent to the fact that any decision problem in which the task is to predict S
can be solved just as well with the knowledge of X2 as with the knowledge of X1 . We refer to Section 2 in [1]
for the details.
Ref. [1] also proposed the following property:
(∗) SI and U I depend only on the marginal distributions PSX1 and PSX2 of the pairs (S, X1 ) and (S, X2 ).
This property was in part motivated by (BP), which also depends only on the channels S → X1 and S → X2 and
thus on PSX1 and PSX2 . Most information decompositions proposed so far satisfy property (∗).
5
III. E XTRACTABLE I NFORMATION M EASURES
One can interpret SI as a measure of extractable shared information. We explain this idea in a more general
setting.
For fixed k, let IM (S; X1 , . . . , Xk ) be an arbitrary information measure that measures one aspect of the
information that X1 , . . . , Xk contain about S. At this point, we do not specify what precisely an information
measure is, except that it is a function that assigns a real number to any joint distributions of S, X1 , . . . , Xk . The
notation is, of course, suggestive of the fact that we mostly think about one of the measures SI, U I or CI, in
which the first argument plays a special role. However, IM could also be the mutual information I(S; X1 ), the
entropy H(S), or the coinformation CoI(S; X1 , X2 ). We define the corresponding extractable information measure
as
IM (S; X1 , . . . , Xk ) := sup IM (f (S); X1 , . . . , Xk ),
(6)
f
where the supremum runs over all functions f : S 7→ S ′ from the domain of S to an arbitrary finite set S ′ . The
intuition is that IM is the maximal possible amount of IM one can “extract” from (X1 , . . . , Xk ) by transforming S.
Clearly, the precise interpretation depends on the interpretation of IM .
This construction has the following general properties:
1) Most information measures satisfy IM (O; X1 , . . . , Xk ) = 0 when O is a constant random variable. Thus, in
this case, IM (S; X1 , . . . , Xk ) ≥ 0. Thus, for example, even though the coinformation can be negative, the
extractable coinformation is never negative.
2) Suppose that IM satisfies left monotonicity. Then, IM = IM . For example, entropy H and mutual information
I satisfy left monotonicity, and so H = H and I = I. Similarly, as shown in [2], the measure of unique
fI = U
fI.
fI defined in [1] satisfies left monotonicity, and so U
information U
3) In fact, IM is the smallest left monotonic information measure that is at least as large as IM .
The next result shows that our construction preserves monotonicity properties of the other arguments of IM .
It follows that, by iterating this construction, one can construct an information measure that is monotonic in all
arguments.
Lemma 1. Let f1 , . . . , fk be fixed functions. If IM satisfies IM (S; f1 (X1 ), . . . , fk (Xk )) ≤ IM (S; X1 , . . . , Xk )
for all S, then IM (S; f1 (X1 ), . . . , fk (Xk )) ≤ IM (S; X1 , . . . , Xk ) for all S.
Proof. Let f ∗ = arg maxf IM (f (S); f1 (X1 ), . . . , fk (Xk )) . Then,
IM (S; f1 (X1 ), . . . , fk (Xk )) = IM (f ∗ (S); f1 (X1 ), . . . , fk (Xk ))
(a)
≤ IM (f ∗ (S); X1 , . . . , Xk ) ≤ sup IM (f (S); X1 , . . . , Xk ) = IM (S; X1 , . . . , Xk ),
f
where (a) follows from the assumptions.
6
As a generalization to the construction, instead of looking at “deterministic extractability,” one can also look at
“probabilistic extractability” and replace f by a stochastic matrix. This leads to the definition
IM (S; X1 , . . . , Xk ) := sup IM (S ′ ; X1 , . . . , Xk ),
(7)
PS ′ |S
where the supremum now runs over all random variables S ′ that are independent of X1 , . . . , Xk given S. The
function IM is the smallest function bounded from below by IM that satisfies
(PLM) IM (S; X1 , X2 ) ≥ IM (S ′ ; X1 , X2 ) whenever S ′ is independent of X1 , X2 given S.
(probabilistic left monotonicity)
An example of this construction is the intrinsic conditional information I(X; Y ↓ Z) := minPZ ′ |Z I(X; Y |Z ′ ),
which was defined in [11] to study the secret-key rate, which is the maximal rate at which a secret can be generated
by two agents knowing X or Y , respectively, such that a third agent who knows Z has arbitrarily small information
about this key. The min instead of the max in the definition implies that I(X; Y ↓ Z) is “anti-monotone” in Z.
In this paper, we restrict ourselves to the deterministic notions, since many of the properties we want to discuss
can already be explained using deterministic extractability. Moreover, the optimization problem (6) is a finite
optimization problem and thus much easier to solve than Equation (7).
IV. E XTRACTABLE S HARED I NFORMATION
We now specialize to the case of shared information. The first result is that when we apply our construction to
a measure of shared information that belongs to a bivariate information decomposition, we again obtain a bivariate
information decomposition.
Lemma 2. Suppose that SI is a measure of shared information, coming from a nonnegative bivariate information
decomposition (satisfying Equations (1) to (3)). Then, SI defines a nonnegative information decomposition; that is,
the derived functions
∗
U I (S; X1 \X2 ) := I(S; X1 ) − SI(S; X1 , X2 ),
∗
U I (S; X2 \X1 ) := I(S; X2 ) − SI(S; X1 , X2 ),
and
∗
∗
∗
CI (S; X1 , X2 ) := I(S; X1 X2 ) − SI(S; X1 , X2 ) − U I (S; X1 \X2 ) − U I (S; X2 \X1 )
are nonnegative. These quantities relate to the original decomposition by
a) SI(S; X1 , X2 ) ≥ SI(S; X1 , X2 ),
∗
b) CI (S; X1 , X2 ) ≥ CI(S; X1 , X2 ),
∗
c) U I(f ∗ (S); X1 \X2 ) ≤ U I (S; X1 \X2 ) ≤ U I(S; X1 \X2 ),
where f ∗ is a function that achieves the supremum in Equation (4).
7
Proof.
a) SI(S; X1 , X2 ) ≥ SI(S; X1 , X2 ) ≥ 0,
∗
b) CI (S; X1 , X2 ) = SI(S; X1 , X2 ) − CoI(S; X1 , X2 ) ≥ SI(S; X1 , X2 ) − CoI(S; X1 , X2 )
≥ CI(S; X1 , X2 ) ≥ 0,
∗
c) U I (S; X1 \X2 ) = I(S; X1 ) − SI(S; X1 , X2 ) ≤ I(S; X1 ) − SI(S; X1 , X2 ) = U I(S; X1 \X2 ),
∗
U I (S; X1 \X2 ) = I(S; X1 ) − SI(S; X1 , X2 ) ≥ I(f ∗ (S); X1 ) − SI(f ∗ (S); X1 , X2 )
= U I(f ∗ (S); X1 \X2 ) ≥ 0,
where we have used the data processing inequality.
Lemma 3.
1) If SI satisfies (∗), then SI also satisfies (∗).
2) If SI is right monotonic, then SI is also right monotonic.
Proof. (1) is direct, and (2) follows from Lemma 1.
Without further assumptions on SI, we cannot say much about when SI vanishes. However, the condition that
∗
U I vanishes has strong consequences.
∗
Lemma 4. Suppose that U I (S; X1 \X2 ) vanishes, and let f ∗ be a function that achieves the supremum in
Equation (4). Then, there is a Markov chain X1 — f ∗ (S) — S. Moreover, U I(f ∗ (S); X1 \X2 ) = 0.
∗
Proof. Suppose that U I (S; X1 \X2 ) = 0. Then, I(S; X1 ) = SI(S; X1 , X2 ) = SI(f ∗ (S); X1 , X2 ) ≤ I(f ∗ (S); X1 ) ≤
I(S; X1 ). Thus, the data processing inequality holds with equality. This implies that X1 − f ∗ (S) − S is a Markov
chain. The identity U I(f ∗ (S); X1 \X2 ) = 0 follows from the same chain of inequalities.
∗
Theorem 1. If U I has the Blackwell property, then U I does not have the Blackwell property.
Proof. As shown in the example in the appendix, there exist random variables S, X1 , X2 and a function f that
satisfy
1) S and X1 are independent given f (S).
2) The channel f (S) → X1 is a garbling of the channel f (S) → X2 .
3) The channel S → X1 is not a garbling of the channel S → X2 .
We claim that f solves the optimization problem (4). Indeed, for an arbitrary function f ′ ,
SI(f ′ (S); X1 , X2 ) ≤ I(f ′ (S); X1 ) ≤ I(S; X1 ) = I(f (S); X1 ) = SI(f (S); X1 , X2 ).
Thus, f solves the maximization problem (4).
8
If U I satisfies the Blackwell property, then (2) and (3) imply U I(f (S); X1 \X2 ) = 0 and U I(S; X1 \X2 ) > 0.
On the other hand,
∗
U I (S; X1 \ X2 ) = I(S; X1 ) − SI(S; X1 , X2 ) = I(S; X1 ) − SI(f (S); X1 , X2 )
= I(S; X1 ) − I(f (S); X1 ) + U I(f (S); X1 \X2 ) = 0.
∗
Thus, U I does not satisfy the Blackwell property.
Corollary 2. There is no bivariate information decomposition in which U I satisfies the Blackwell property and SI
satisfies left monotonicity.
∗
Proof. If SI satisfies left monotonicity, then SI = SI. Thus, U I = U I cannot satisfy the Blackwell property by
Theorem 1.
V. L EFT M ONOTONIC I NFORMATION D ECOMPOSITIONS
Is it possible to have an extractable information decomposition? More precisely, is it possible to have an
information decomposition in which all information measures are left monotonic? The obvious strategy of starting
with an arbitrary information decomposition and replacing each partial information measure by its extractable
analogue does not work, since this would mean increasing all partial information measures (unless they are
extractable already), but then their sum would also increase. For example, in the bivariate case, when SI is replaced
by a larger function SI, then U I needs to be replaced by a smaller function, due to the constraints (2) and (3).
As argued in [2], it is intuitive that U I be left monotonic. As argued above (and in [5]), it is also desirable that
SI be left monotonic. The intuition for synergy is much less clear. In the following, we restrict our focus to the
bivariate case and study the implications of requiring both SI and U I to be left monotonic. Proposition 5 gives
bounds on the corresponding SI measure.
Proposition 5. Suppose that SI, U I and CI define a bivariate information decomposition, and suppose that SI
and U I are left monotonic. Then,
SI(f (X1 , X2 ); X1 , X2 ) ≤ I(X1 ; X2 )
(8)
for any function f .
Before proving the proposition, let us make some remarks. Inequality (8) is related to the identity axiom. Indeed, it
is easy to derive Inequality (8) from the identity axiom and from the assumption that SI is left monotonic. Although
Inequality (8) may not seem counterintuitive at first sight, none of the information decompositions proposed so far
satisfy this property (the function If from [12] satisfies left monotonicity and has been proposed as a measure of
shared information, but it does not lead to a nonnegative information decomposition).
9
Proof. If SI is left monotonic, then
SI(f (X1 , X2 ); X1 , X2 ) ≤ SI(C OPY (X1 , X2 ); X1 , X2 )
= I(C OPY (X1 , X2 ); X1 ) − U I(C OPY (X1 , X2 ); X1 \X2 ).
If U I is left monotonic, then
U I(C OPY (X1 , X2 ); X1 \X2 ) ≥ U I(X1 ; X1 \X2 ) = I(X1 ; X1 ) − SI(X1 ; X1 , X2 ).
Note that I(X1 ; X1 ) = H(X1 ) = I(C OPY (X1 , X2 ); X1 ) and
SI(X1 ; X1 , X2 ) = I(X1 ; X2 ) − U I(X1 ; X2 \X1 ) = I(X1 ; X2 ).
Putting these inequalities together, we obtain SI(f (X1 , X2 ); X1 , X2 ) ≤ I(X1 ; X2 ).
VI. E XAMPLES
In this section, we apply our construction to Williams and Beer’s measure, Imin [3], and to the bivariate measure
f proposed in [1].
of shared information, SI,
First, we make some remarks on how to compute the extractable information measure (under the assumption that
one knows how to compute the underlying information measure itself). The optimization problem (4) is a discrete
optimization problem. The search space is the set of functions from the support S of S to some finite set S ′ . For
the information measures that we have in mind, we may restrict to surjective functions f , since the information
measures only depend on events with positive probabilities. Thus, we may restrict to sets S ′ with |S ′ | ≤ |S|.
Moreover, the information measures are invariant under permutations of the alphabet S. Therefore, the only thing
that matters about f is which elements from S are mapped to the same element in S ′ . Thus, any function f : S → S ′
corresponds to a partition of S, where s, s′ ∈ S belong to the same block if and only if f (s) = f (s′ ), and it suffices
to look at all such partitions. The number of partitions of a finite set S is the Bell number B|S| .
The Bell numbers increase super-exponentially, and for larger sets S, the search space of the optimization
problem (4) becomes quite large. For smaller problems, enumerating all partitions in order to find the maximum
is still feasible. For larger problems, one would need a better understanding about the optimization problem. For
reference, some Bell numbers include:
n
3
4
6
10
Bn
5
15
203
115975
.
As always, symmetries may help, and so in the C OPY example discussed below, where |S| = 4, it suffices to study
six functions instead of B4 = 15.
We now compare the measure I min , an extractable version of Williams and Beer’s measure Imin (see Equation (5)
f an extractable version of the measure SI
f proposed in [1]. For the latter, we briefly
above), to the measure SI,
recall the definitions. Let ∆ be the set of all joint distributions of random variables (S, X1 , X2 ) with given state
10
TABLE I
Shared information about f (X1 , X2 ) for various functions f (in bits).
f
C OPY
A ND/O R
Imin
I min
f
SI
f
SI
1
1
0
1/2
3/4 log 4/3
3/4 log 4/3
3/4 log 4/3
3/4 log 4/3
X OR
0
0
0
0
S UM
1/2
1/2
1/2
1/2
X1
0
0
0
0
f1
1/2
1/2
0
0
spaces S, X1 , X2 . Fix P = PSX1 X2 ∈ ∆. Define ∆P as the set of all distributions QSX1 X2 that preserves the
marginals of the pairs (S, X1 ) and (S, X2 ), that is,
∆P ..= QSX1 X2 ∈ ∆ : QSX1 = PSX1 , QSX2 = PSX2 , ∀ (S, X1 , X2 ) ∈ ∆ .
Then, define the functions
fI(S; X1 \X2 ) ..= min IQ (S; X1 |X2 ),
U
Q∈∆P
fI(S; X2 \X1 ) ..= min IQ (S; X2 |X1 ),
U
Q∈∆P
f
SI(S;
X1 , X2 ) ..= max CoIQ (S; X1 , X2 ),
Q∈∆P
f
CI(S;
X1 , X2 ) ..= I(S; X1 X2 ) − min IQ (S; X1 X2 ),
Q∈∆P
where the index Q in IQ or CoIQ indicates that the corresponding quantity is computed with respect to the joint
f satisfies the Blackwell property and the identity axiom [1].
distribution Q. The decomposition corresponding to SI
f=
f SI
f can be characterized as the smallest measure
fI is left monotonic, but SI
f is not [2]. In particular, SI
6 SI.
U
f is the smallest left monotonic measure of shared
of shared information that satisfies property (∗). Therefore, SI
information that satisfies property (∗).
Let X1 = X2 = {0, 1} and let X1 , X2 be independent uniformly distributed random variables. Table I collects
values of shared information about f (X1 , X2 ) for various functions f (in bits). The function f1 : {00, 01, 10, 11} →
{0, 1, 2} is defined as
f1 (X1 , X2 ) :=
X1 ,
2,
if X2 = 1,
if X2 = 0.
The S UM function is defined as f (X1 , X2 ) := X1 + X2 . Table I contains (up to symmetry) all possible non-trivial
functions f . The values for the extractable measures are derived from the values of the corresponding non-extractable
measures. Note that the values for the extractable versions differ only for C OPY from the original ones. In these
examples, I min = Imin , but as shown in [5], Imin is not left monotonic in general.
11
VII. C ONCLUSIONS
We introduced a new measure of shared information that satisfies the left monotonicity property with respect to
local operations on the target variable. Left monotonicity corresponds to the idea that local processing will remove
information in the target variable and thus should lead to lower values of measures which quantify information about
the target variable. Our measure fits the bivariate information decomposition framework; that is, we also obtain
corresponding measures of unique and synergistic information. However, we also have shown that left monotonicity
for the shared information contradicts the Blackwell property of the unique information, which limits the value of
a left monotonic measure of shared information for information decomposition.
We also presented an alternative interpretation of the construction used in this paper. Starting from an arbitrary
measure of shared information SI (which need not be left monotonic), we interpret the left monotonic measure SI
as the amount of shared information that can be extracted from S by local processing.
Our initial motivation for the construction of SI was the question to which extent shared information originates
from the redundancy between the predictors X1 and X2 or is created by the mechanism that generated S. These
two different flavors of redundancy were called source redundancy and mechanistic redundancy, respectively, in [4].
While SI cannot be used to completely disentangle source and mechanistic redundancy, it can be seen as a measure
of the maximum amount of redundancy that can be created from S using a (deterministic) mechanism. In this sense,
we believe that it is an important step forward towards a better understanding of this problem and related questions.
A PPENDIX : C OUNTEREXAMPLE IN T HEOREM 1
Consider the joint distribution
f (s)
s
x1
x2
Pf (S)SX1 X2
0
0
0
0
1/4
0
1
0
1
1/4
0
0
1
0
1/8
0
1
1
0
1/8
1
2
1
1
1/4
and the function f : {0, 1, 2} → {0, 1} with f (0) = f (1) = 0 and f (2) = 1. Then, X1 and X2 are independent
uniform binary random variables, and f (S) = A ND (X1 , X2 ). In addition, S − f (S) − X1 is a Markov chain. By
symmetry, the joint distributions of the pairs (f (S), X1 ) and (f (S), X2 ) are identical, and so the two channels
f (S) → X1 and f (S) → X2 are identical, and, hence, trivially, one is a garbling of the other. However, one can
check that the channel S → X1 is not a garbling of the channel S → X2 .
This example is discussed in more detail in [13].
R EFERENCES
[1] N. Bertschinger, J. Rauh, E. Olbrich, J. Jost, and N. Ay, “Quantifying unique information,” Entropy, vol. 16, no. 4, pp. 2161–2183, 2014.
[2] J. Rauh, N. Bertschinger, E. Olbrich, and J. Jost, “Reconsidering unique information: Towards a multivariate information decomposition,”
in Proc. IEEE ISIT, 2014, pp. 2232–2236.
12
[3] P. Williams and R. Beer, “Nonnegative decomposition of multivariate information,” arXiv:1004.2515v1, 2010.
[4] M. Harder, C. Salge, and D. Polani, “A bivariate measure of redundant information,” Phys. Rev. E, vol. 87, p. 012130, Jan 2013.
[5] N. Bertschinger, J. Rauh, E. Olbrich, and J. Jost, “Shared information — new insights and problems in decomposing information in
complex systems,” in Proc. ECCS 2012.
Springer, 2013, pp. 251–269.
[6] V. Griffith and C. Koch, “Quantifying synergistic mutual information,” in Guided Self-Organization: Inception, M. Prokopenko, Ed.
Springer Berlin Heidelberg, 2014, vol. 9, pp. 159–190.
[7] M. Wibral, V. Priesemann, J. W. Kay, J. T. Lizier, and W. A. Phillips, “Partial information decomposition as a unified approach to the
specification of neural goal functions,” Brain and Cognition, vol. 112, pp. 25 – 38, 2017, perspectives on Human Probabilistic Inferences
and the ’Bayesian Brain’. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S027826261530021X
[8] A. J. Bell, “The co-information lattice,” in Proc. Fourth Int. Symp. Independent Component Analysis and Blind Signal Separation (ICA
03), 2003.
[9] W. McGill, “Multivariate information transmission,” IRE Trans. Inf. Theory, vol. 4, no. 4, pp. 93–111, 1954.
[10] D. Blackwell, “Equivalent comparisons of experiments,” The Annals of Mathematical Statistics, vol. 24, no. 2, pp. 265–272, 1953.
[11] U. Maurer and S. Wolf, “The intrinsic conditional mutual information and perfect secrecy,” in Proc. IEEE ISIT, 1997.
[12] V. Griffith, E. K. P. Chong, R. G. James, C. J. Ellison, and J. P. Crutchfield, “Intersection information based on common randomness,”
Entropy, vol. 16, no. 4, pp. 1985–2000, 2014.
[13] J. Rauh, P. K. Banerjee, E. Olbrich, J. Jost, N. Bertschinger, and D. Wolpert, “Coarse-graining and the blackwell order,” arXiv:1701.07805,
2017.
| 7cs.IT
|
Department of Computer Science and Software Engineering
Comparative Studies of 10 Programming Languages
within 10 Diverse Criteria
Jiang Li
Concordia University
Montreal, Quebec, Concordia
[email protected]
Mingzhi Liu
Concordia University
Montreal, Quebec, Concordia
[email protected]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Sleiman Rabah
Concordia University
Montreal, Quebec, Concordia
[email protected]
Yuanwei Lai
Concordia University
Montreal, Quebec, Concordia
[email protected]
1/139
This page was intentionally left blank
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
2/139
Abstract
There are many programming languages in the world today.Each language has their
advantage and disavantage. In this paper, we will discuss ten programming languages:
C++, C#, Java, Groovy, JavaScript, PHP, Schalar, Scheme, Haskell and AspectJ. We
summarize and compare these ten languages on ten different criterion. For example,
Default more secure programming practices, Web applications development, OO-based
abstraction and etc. At the end, we will give our conclusion that which languages are
suitable and which are not for using in some cases. We will also provide evidence and
our analysis on why some language are better than other or have advantages over the
other on some criterion.
1 Introduction
Since there are hundreds of programming languages existing nowadays, it is impossible and inefficient
to put effort on analyzing each languages. But we can classify the some representative categories of
languages and make deep research on them according to some certain criteria. Thus our research
problem is aiming to compare and contrast 10 languages according to 10 specified criteria with the
purpose of determining the suitability and applicability of the languages for each criterion, distinguish
them their pros and cons, evaluate and explore the related features on those languages, illustrate the
points either with code examples or related work. In our project, we will evaluate our languages based
on following criteria:
1. Default more secure programming practices
2. Web applications development
3. Web services design and composition
4. OO-based abstraction
5. Reflection
6. Aspect-orientation
7. Functional programming
8. Declarative programming
9. Batch scripting
10. UI prototype design
Depends on your choice of languages, some of them may have something in common on certain
aspect while some part may totally different.
1.1
Related work
In order to complete our comparison work, we do some relevant search among conference papers, text
books, Wikipedia, official websites and discuss with classmate and teacher in courses.
1.2
Overview
The rest of this paper is organized as follows. First we introduce the formatting basics in SecCOMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
3/139
tion 1.3 and Section 1.4. We then briey introduce the languages being compared in Section 1.4, 1.5,
1.6, 1.7, 1.8, 1.9, 1.10, 1.11, 1.12, and Section 1.13. Next, we present our analysis of the criteria in
Section 2 for pair-wise comparison of the assigned languages. We then move on to macro analysis and
synthesis of our results in a consolidated form in Section 3. We conclude and outline our future work
plans in Section 4 and Section 4.1 respectively.
1.3
Programming Language
1.4
C++
C++ is a statically typed, free-form, multi-paradigm, compiled, general-purpose programming
language. Some people say that C++ is a middle language because it has the features of high-level and
low-level language. As one of the most popular programming languages in the world, C++ is widely
used in the software industry.[1] C++ is also used for hardware designto analyze structure. Some of its
application domains include systems software, application software, device drivers, embedded
software, high-performance server and client applications, and entertainment software such as video
games.
1.5
AspectJ
AspectJ is a general-purpose Aspect-Oriented extension to java programming language [2]. It was
created at Palo Alto Research Center Incorporated (PARC), now it is an open source project and part
of the Eclipse Foundation. AspectJ has everything that Java has and more which means every valid
Java program is also a valid AspectJ program [2]. The main goal of AspectJ development is
modularizing crosscutting concerns such as logging, error checking and handling, synchronization,
context-sensitive behavior, performance optimizations, monitoring and logging, debugging support,
and multi-object protocols [4].
Aspect-oriented programming (AOP) is a programming paradigm built on top of the objectoriented paradigm and aims to modularize crosscutting concerns [2] by isolating secondary functions
from the program’s business logic [3]. AOP enhances code readability and reuse.
AspectJ compiler produces java bytecode, an AspectJ program can run on any Java compatible
virtual machine. The runtime library “aspectjrt.jar” is required to run any AspectJ program. AspectJ
development tool (AJDT) is a plug-in for the Eclipse IDE which can be used to compile and run
AspectJ programs.
1.6
Haskell
Haskell is an advanced, standardized, general-purpose purely functional programming language
incorporating many recent innovations in programming language design. Haskell provides higherorder functions, non-strict semantics, static polymorphic typing, user-defined algebraic datatypes,
pattern-matching, list comprehensions, a module system, a monadic I/O system, and a rich set of
primitive datatypes, including lists, arrays, arbitrary and fixed precision integers, and floating-point
numbers. In Haskell, a function is a primary control construct of the programming language. It allows
rapid development of robust, concise, correct software. Haskell is easier to produce flexible,
maintainable high-quality software due to its strong support for integration with other languages, builtin concurrency and parallelism, debuggers, profilers, rich libraries and an active community. [5]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
4/139
1.7
PHP
PHP is a wide use general purpose scripting language which used to make dynamic interact web
pages. It can embedded with HTML source document used in server side. Influenced by C, Perl, Java,
C++, thus it support multiple paradigm in programming, such as object-oriented (OO) and imperative.
In the meantime, its type system is loosing typing and dynamic type checking. As the updates of
version, it supports more new features to make the PHP more functional and diversify. Now, the latest
version is PHP 5.33, and will be used in the following.
1.8
Scheme
Scheme is a general-purpose, functional and multi-paradigm programming language. Scheme
derives some of its dialects and features from LSIP. Scheme is primarily intended to be a functional
programming language; it supports lambda calculus, lexical scope and recursion. Today, Scheme is
almost every where: it is used in many software development projects such as text editors, compilers
optimization, expert systems, etc [9].
There are many implementations of Scheme providing different features [6] based on IEEE Scheme
standards. The Gambit project provides the Gambit-C compiler as part of Gambit programming
system. Gambit-C generates portable C code and executable [7]. PLT Scheme (known as Racket
project) is one of Scheme implementations which provides a massive set of libraries for supporting
many features such as GUI, macros, classes and objects and many more [8].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
5/139
1.9
Groovy
Groovy is an agile and dynamic language for the Java Virtual Machine. It has some features
similar to those of Python, Ruby, Perl, and Smalltalk. It can be used as a scripting language for the
Java Platform.[10] Groovy uses a Java-like bracket syntax. Groovy compiles into Java bytecode and
extends the Java API and libraries. It runs on Java 1.4 or newer. Most Java code is also syntactically
valid Groovy.Groovy support XML and HTML.
1.10
Java
Java is a powerful, platform independent, object-oriented, strongly-type, interpreted and compiled,
general-purpose programming language with build-in automatic memory management. It is a
programming language originally developed by James Gosling at Sun Microsystems (which is now a
subsidiary of Oracle Corporation) and released in 1995 as a core component of Sun Microsystems'
Java platform. The language derives much of its syntax from C and C++ but has a simpler object
model and fewer low-level facilities. It is intended to let application developers "write once, run
anywhere". Java is currently one of the most popular programming languages in use, and is widely
used from application software to web applications. [11]
1.11
JavaScript
JavaScript is a small, lightweight, prototype-based object-oriented, interpreted, cross-platform
scripting
language.
Today,
it
is
the
most
popular
http://en.wikipedia.org/wiki/Scripting_languagescripting language on the internet, and works in all
major browsers, such as IE, Firefox, Chrome, Opera, and Safari. JavaScript was designed to add
interactivity to HTML pages and usually is embedded directly into HTML pages. But it may also be
used at outside webpage, such as server side.
1.12
Scala
Scala is a general purpose programming languages which support multiple paradigms. It extends
the object-oriented characteristic with functional extension. It integrated many other languages
features to itself. It is designed to express common programming patterns in a concise, elegant, and
type-safe way [12]. The Scala run its code on the JVM, which is byte code compatible with Java. That
is to say you can utilize all the libraries or existing resource in Java. This not only can benefit the java
programmer to make productive and efficient product using Scala, it gives a higher start point for
Scala completive with other programming languages. Although it seems Scala integrate with Java
seamlessly, Scala is not a subset of java, it has much more features rather than Java. Now the latest
version of Scala is 2.80. And will be used in the following.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
6/139
1.13
C#
C# is modern, general-purpose and multi-paradigm programming language enclosing objectoriented, imperative, functional, generic, event-driven and component-oriented programming styles
(DLLs and Assemblies). It was designed by Andres Hejlsberg (the creator of Turbo Pascal), developed
by Microsoft and first released in July 2000. C# was developed specially for the .NET platform with a
main goal to provide a simple, powerful and a strongly-typed programming language allowing
programmers to quickly build a wide range of applications for the .NET platform [13].
The .NET platform is composed of a Common Language Runtime (CLR) and a large rich class
library known as the .NET Framework (DLLs files). The .NET Framework provides a wide powerful
range of features, among them: multi-threading, user interface prototyping, database connectivity, web
application and service-oriented application development.
The CLR is the core component of the .NET platform and the execution environment in which all
the managed code runs. The CLR is the Microsoft’s JVM equivalent that supports several
programming languages and performs services such as memory management, exception handling,
garbage collection, security and interaction with the operating-system services. The CLR provides a
common development environment enabling developers to build applications using different
languages such as C#, VB, C++, F# and Python, etc.
C# code source is compiled to an intermediate language presentation called Microsoft Intermediate
Language (MSIL) and it is the Java byte-code’s equivalent. MSIL is translated into machine code by
the CLR at run-time. C# has an advantage over Java which is high interoperability with other
languages such as C/C++, Python, VB.NET, etc [14].
C# can be used to develop the following types of applications [14]:
• Command line application, aka console applications with a text-based interface (text user
interface), that can be run using a Command Line Interface e.g MS DOS.
• Windows applications for developing GUI using Windows Forms or the new Windows
Presentation Foundation (WPF) which is first released with the .NET Framework 3.0 that
enables rendering user interfaces.
• Web applications such as Websites using ASP.NET technology, Web Services and serviceoriented application using Windows Communication Foundation (WCF) framework.
The .NET Framework needs to be installed in order to run application written in C#.
Microsoft’s Visual C# Integrated Development Environment (IDE) is used to create such applications.
It provides a set of built in tools such as a C# compiler, a user interface designer for web and GUI
development and also a powerful debugger. C# has been approved as a standard by the European
Computer Manufacturers Association (ECMA) (ECMA-334) [15] [13].
Mono is an open source project which aims to create a cross-platform implementation of the .NET
platform based on the EMCA standard.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
7/139
2 Analysis
2.1
AspectJ vs C#
2.1.1 Source code size
The following examples consist of the Hello World program; they also show AsepctJ and C# syntax.
We conclude that both languages have nearly the same syntax. Thus, the source code size is little
bigger in AspectJ in this examples but also the have nearly the same compilation time.
/**
* Example illustrating a simple aspect
*
* @author Sleiman Rabah
*/
public aspect HelloFromAspectJ {
// Intercepts the main method execution
pointcut mainMethod() : execution(public static void main(String[]));
// Will be executed at the end of the main method
after() returning : mainMethod() {
System.out.println("Hello from AspectJ");
}
}
/**
* A Java class.
*
* @author Sleiman Rabah
*/
public class HelloWorld {
public static void main(String[] args) {
System.out.println("Hello world!");
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
8/139
Hello word program in C# showing its syntax. Hello word program in C# showing its syntax.
using System;
// A "Hello World" program in C#
namespace HelloWorld
{
/// <summary>
/// <autor>Sleiman Rabah</autor>
/// </summary>
class Hello
{
static void Main()
{
System.Console.WriteLine("Hello World!");
}
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
9/139
2.1.2 Default more secure programming practices
“Our civilization runs on software”
__ Bjarne Stroustrup
Our reliance on software to automate things has resulted in including software in every industry from
healthcare, education, aviation to defense. But the question remains how often programs don’t correctly work
due to software bugs? And, can we develop better more robust software? Most of software bugs come from data
mishandling (data conversion), unnecessary code and unsecure input and output handling.
AspectJ and C# are both strongly-typed languages and were designed to be secure; hence their type systems
play a very important role in developing secure programs by ensuring the type-safety. AspectJ is statically typed
language, but it does support some kind of dynamic typing such as down casting. Since .NET 4.0, C# supports
dynamic-type checking after the introduction of the dynamic keyword. Dynamic type-checking can be disabled
in C# by using an unsafe code block marked as unsafe (using the keyword unsafe)[16].
AspectJ and C# also derive a set of features from C and C++ except pointers which has eliminated a major
problem: manual memory management.
The JVM and the CLR have similar run-time services. As part of their specification, they manage code
execution, automatic memory management (Garbage Collection, memory allocation and de-allocation) and
exception handling. Garbage collection is an automatic memory management mechanism; it eliminates some
bugs related to manual memory management such as dangling pointers and double free bugs (freeing a pointer
twice) [17].
Both AspectJ and C# provide an exception handling mechanism allowing preventing application from
crashing at run-time. Contrarily to AspectJ/Java primitive data type, C# has value types (used to store values)
which are objects found in the System namespace [18].
Run-time bounds checking are provided by both Aspect and C# [19]. Bounds checking are used to control data
structure manipulation such as arrays.
To ensure total safety in AspectJ applications, programmers have to be well experimented when
expressing crosscutting. A minimal logical mistake can result in total system failure [20].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
10/139
Default more secure programming features, AspectJ vs C#
Feature
Memory management
Bounds checking
Static Type checking
Dynamic Type checking
Type safety
Exception handling
Compiled/Interpreted
Conditional compilation
Assertions
AspectJ
Yes, provided by the JVM
C#
Yes, provided by the CLR
Yes, also for array bounds
Yes, for arrays and data
(raise IndexOutOfRange
structure.
exception), buffer overflow.
It can be disabled in C# [25]
Yes, at compile time.
Yes at compile time.
Partial dynamic checking,
Yes, since C# 4.0
i.e down casting. Also,
(.NET System.Dynamic
when using reflection.
namespace)
Yes, C# is strongly typed.
Yes, Based on Java’s access The unsafe keyword can be
rules.
used for allowing pointers
use.
Yes, exceptions can be
thrown in pointcuts.
Yes, Within tryHandlers are used to catch
catch/finally block.
exceptions in AspectJ.
Multiple catch blocks is
Within try-catch/finally
supported in C#.
block in Java code.
Compiled, and the JavaCompiled, Interpreted by
byte code interpreted by a
the CLR
JVM compliant.
Yes (using SCoPE
Yes (using preprocessor
compiler, conditional
directives)[21]
pointcut evaluator) [23]
Yes, temporal assertions at
Yes, Managed code
run-time[24]
assertions [22]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
11/139
2.1.3 Web applications development
ASP.NET/C#: As part of the .NET Framework, ASP.NET is a new web development model and
the successor of Aspect Server Pages (ASP) technology. ASP.NET enables developers to build a wide
variety of secure server-side and browser-based applications such as e-commerce, dynamic websites,
and e-learning applications. ASP.NET is compiled and run on the .NET platform where it takes
advantage of the CLR secure and powerful environment and its multi-programming languages
support. ASP.NET is more and more widely used in software industry thanks to the Visual Studio
.NET IDE which is excellent for rapidly building variety of applications in a Rapid Application
Development (RAD) environment.
C# language is mainly and heavily used to develop such applications due to its power, features and
simplicity. There are plenty of open source web applications project written in C# and ASP.NET that
can be found in the Codeplex [26] repositories such as Customer Relationship Management (CRM)
“e.g ASP.NET CRM” [27], Content Management System (CMS) and e-commerce such as
nopCommerce which is the best available open source .NET based solution, also written in C# [28].
ASP.NET has received a lot of criticism because it was first released with a
Technology called Web Forms technology which aims to separate HTML code and stuff from
application logic, and reduce code by using the data binding capabilities of the server-side .NET
control. Web Forms applications are not flexible, but difficult to test and not extendable because they
use what is called server-based forms. ASP.NET MVC was released first in 2007 as an alternative to
the Web Forms that enables creation MVC-based web applications.
AspectJ/J2EE: Java 2 Platform Enterprise Edition (J2EE) technology is not only suited to build
dynamic websites but also enterprise applications and server applications. Its simplicity, complete
portability on any operating-system and design made it widely popular. J2EE includes/uses APIs such
as Java Database Connectivity (JDBC) Database Connectivity which allows applications to interact
with databases, Remote Method Invocation (RMI) for building distributed applications and many
more. Java security and portability makes J2EE to be the first choice for developing online banking
systems and transactional/e-commerce web applications.
AspectJ can be applied/coupled to J2EE in the development of web applications where aspects are
used to isolate secondary concerns (such as error-handling and logging), improve the flexibility of
your code and reduce the code source size [29]. AspectJ/J2EE applications are clean, easy to maintain
and to debug. SpringSource dm Server is open source solution which is used to develop enterprise
applications that use AspectJ in order to either simplify enterprise application or implement different
crosscutting functionalities. AspectJ is heavily used with Spring Framework [30]. AspectJ can be used
in Java Servlet [31]
AspectJ/J2EE vs C#/ASP.NET:
J2EE is more mature than ASP.NET [32], thus, it supports AOP using different tools/language
such as AspectJ and JBoss AOP framework and it has been adopted by various firms and foundation
such as IBM, SAP, Oracle and Apache Software Foundation for developing a wide range of server
applications such as WebSphere.
Because of Microsoft’s products bad reputation in term of security, Java applications became more
popular to adopt in developing critical web applications, but it seems that the new .NET infrastructure
has been evolved where new security features have been introduced to the .NET framework and CLR
environment such as role-based security and authentication.
ASP.NET applications are deployed on Microsoft’s Internet Information Services (IIS) and requires
the .NET platform to be installed.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
12/139
Web applications: AspectJ vs C#
Feature
Dynamic Web Pages
Web Server
Web Framework/Libraries
Session management
Security
Model-View-Controller
Database interaction
Secure Sockets Layer
support
AspectJ
Yes, injection Aspectj code
in JSP pages
Tomcat, JBoss, GlassFish
Using J2EE, Spring or
Struts framework.
Yes, i.e using Spring which
provides
SessionManagementFilter
class and concurrency
control [33].
Yes, Java web applications
run in what is called
container, Each application
has its container. Thus, the
JVM add
By adding AspectJ’s
aspects to
J2EE/Spring/Struts
applications.
C#
Yes, as code behind
ASP.NET web pages.
IIS on Windows OS
Mono (Apache, Nginx)
Using .NET Framework on
Windows and Mono
ASP.NET Linux and Unixbased OSs.
Yes, provided by ASP.NET
State management.
Yes (ASP.NET Security
Architecture [sr-4])
ASP.NET MVC 1.0/2.0
Yes, using ADO.NET
library
Yes, SSL certificate can be
Yes using OpenSSL library. generated and used to on
IIS server.
Yes, using the JDBC library
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
13/139
2.1.4 Web services design and composition
AspectJ doesn’t provide an API for building web services, but it can be very useful in combination
with other framework like spring [30] for building web applications. Web Services, can be seen also as
distributed distant programs that communicate together via HTTP.
A distributed system is a collection of independent computers that appears to its users as a single
coherent system [35]. One program can be piped to another and so on resulting in kind of a complex
program’s network. As results, many issues will obviously appear and force programmers to take in
consideration. Some of these issues are: communication, fault tolerance, synchronization and security
[36].
AOP is the solution to reduce distributed systems’ complexity and improve their efficiency by
modularizing crosscutting concerns mentioned before. A group of aspects can be added to an existing
system with a main task to observe the system’s pipeline. For example, “when part of the system fails,
a new behaviour emerges which uses certain aspects along with the rest of the system” [36].
.NET Framework provides the System.Web.Services namespace which enables programmers to write
XML Web Services (WS) that are available over the Web. Windows Communication Foundation (WCF) is
an API for building service-oriented applications (SOA) in a unified programming model; it was
released with .NET 3.0. WCF is based on the SOA principles which enable distributed applications to
communicate together via SOAP messages. Visual Studio IDE is a great tool to create Web Services:
the WSDL file is generated and maintained automatically as soon as the code changes, also it provides
tools for testing and debugging a WS at run-time [37]. Microsoft BizTalk Server provides services to
services that enable different distributed applications/web services to communicate together. BizTalk
is based on adapters that support different communication protocol such as HTTP, FTP and SOAP
[37].
Unfortunately, the .NET solution is not flexible enough and it is far from benefiting from the AOP
approach to modularize crosscutting concerns. There are no mainstream AOP solutions that are
suitable for the .NET framework. AspectJ is a cross-platform language that can run on any platform
and can be coupled with Java’s technology for building either web applications or distributed ones.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
14/139
Web Services: AspectJ vs C#
Feature
Web Services
including(SOAP, WSDL,
UDDI)
AspectJ
Yes. i.e using Apache
XML-RPC for Java library
[38] or Spring Web
Services, JAX-WS 2.1)
Web services security
Yes. HTTPS, above
libraries support “XMLbased standards” such as
WS-Policy, WS-Security
and WS-Transfer standards)
Web Services pipeline
watching.
Using AspectJ and J2EE
framework, two or more
web services can
communicate together by
linking them as a network.
Web Services composition
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
C#
Yes.
using ASPT.NET web
services and .NET’s
System.Web.Services
namespaces for Windowsbased WS [42] and Mono
project’s Web Services on
other OSs (Linux/Unix,
Mac)
Yes, HTTPS Encryption
and signing with SSL[39],
WS-Policy, WS-Security
standards support)
Using ASP.NET web
services. Also a network of
web services is possible by
making two or more WS
talking together.
15/139
2.1.5 OO-based abstraction
AspectJ is an aspect-oriented programming language; it is an extension to OO.
“Aspects are the central unit in AspectJ” [2]. Aspects and classes have many similarities: they share
some common elements such as methods and fields, but also aspects introduce the one that is related
to AOP: pointcuts and advices. Like classes, an aspect can be extended, but only abstract aspects can
be extended. Unlike classes, the main goal of using aspects is to modularize crosscutting concerns.
C# is a pure object-oriented programming language; in C# everything is an object. Every declared
type (object) implicitly or explicitly inherits from the System.Object class. It supports all object
orientation concepts: abstraction (classes), encapsulation, polymorphism and inheritance. Also, C#
supports interfaces, abstracts classes and has introduced partial classes (the class’s source code can be
split over two or more files) [41].
Multiple inheritance is not supported in C#. C# provides properties, also called accessors (Java’s
getter and setter equivalent) which are class members that provide access either to read or write private
fields’ values.
C# is suitable to build object-oriented-based applications. C# is designed to be an object-oriented
while AspectJ is designed as an aspect-oriented.
OO-based abstraction, Aspectj vs C#
Feature
Central unit
AspectJ
Structural elements
Pointcuts, Advices, InterType declarations for
declaring aspect’s members
(fields, methods, and
constructors)
Aspects, abstract Aspects
Multiple inheritance
No (Aspect Inheritance)
Polymorphism
Partial Classes
Structs support
Instantiation
Aspectual polymorphism
[44]. Method overloading,
method overriding, virtual
method. By default AspectJ
methods are virtual (nonstatic methods). Using
Inter-type declaration.
No, but it provides Intertype declarations (aka open
classes) as alternative.
No, doesn’t support Structs
declaration.
Aspect instantiation: not
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
C#
Classes, interfaces, abstract
classes
Constructors, Destructors,
Methods, Fields (attributes),
Delegates, Properties,
Indexes, Events, Finalizes,
Operators, Nested Classes.
No, can be done by using
Interface implementation as
work around.
Yes, through inheritance,
based on Base Type and Sub
Type relation. Method
overloading, method
overriding. C# provides the
virtual
Keyword which is derived
from C++.
Yes, separation of code
(class definition split into
two or more source files.
Yes, derived from C++ with
enhanced features.
Yes, Class instantiation
16/139
Extension/Inheritance
Accessibility control
directly instantiated. You
cannot use the new keyword
to instantiate an aspect.
Yes, “Aspect extension”
Aspects can extend classes
and implement interfaces,
but Aspects can extend only
abstract Aspects.
Yes, privileged for Aspects,
private, public and
protected for inter-type
members.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
using the new keyword.
Yes, “class extension”. A
class may extend another
class and may implement
one or more interface.
Extension methods feature
is supported since C# 3.0
but it requires a static class
and static method to do so
[43].
Yes, C# provides private,
public, protected, internal,
sealed and protected
internal keywords.
17/139
2.1.6 Reflection
In AspectJ, join point information is accessed via reflection. There are two types [2]:
• dynamic information (dynamic crosscutting): this type of information changes after each call
of the same join points. Dynamic information result in the program’s dynamic flow.
• static information (static crosscutting): is said fixed and doesn’t change between multiple
execution.
AspectJ provides three objects to be used in advice body:
thisJointPoint, thisjoinPointStaticPart, thisEnclosingJoinPointStaticPart. These objects are similar to
this keyword in C# or Java which provides access to the current object being executed [2] [45].
Since AspectJ 5, a new reflection API is provided similar to java.lang.reflect package. To use this
API, code should be compiled by the AspectJ 5 compiler and runs under Java 5 environment [46].
.NET components are pre-compiled modules with a .DLL extension (Dynamic Link Library), they
are also known as assemblies which are the smallest unit of execution in the CLR environment.
Thanks to the .NET CLR dynamic services, especially the loader manager, these components can be
loaded at run-time into the memory and used by different applications directly or through reflection
[47].
C# allows reflective programming in different manners. Reflection in C# can be used for loading
assemblies (DLLs) in order to instantiate some of their types. Custom Attributes, which are a special
form of customizable code annotation and an extension to the reflection model “similar to Java 1.5
annotation”, contain metadata information that can be attached to classes or methods and they are
accessed at run-time through reflection.
The .NET framework provides the System.Reflection.Emit namespace for code generation at run-time
“emitting dynamic methods and assemblies”. This namespace provides developers many capabilities
such as define assemblies at run-time and saving them to disk, define new types and create instances
of these types also at run-time and so on [47].
There are some Java’s equivalent libraries available that can be used in AspectJ applications such
as ASM http://asm.ow2.org and BCEL http://jakarta.apache.org/bcel. These libraries provide similar
services to the .NET’s Reflection.Emit namespace for generating byte-code at run-time.
Accessing information dynamically is not suggested in certain applications; reflection has a poor
performance and it is unsafe without static type checking [2].
C# provides more features to do reflective programming than AspectJ. AspectJ is not flexible
enough to create or modify aspects on the fly (at run-time). For code security reason, AspectJ’s joint
cuts have to be well tested since they are plugged in a way to the program flow which is difficult to do
using reflection.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
18/139
Reflection, AspectJ vs C#
Feature
Access to program’s
metadata
AspectJ
Yes, since AspectJ 5.0.
Annotation-based
development style.
C#
Yes, accessing Attributes
through reflection using
System. Reflection
namespace.
Yes, using .NET’s
System.Reflection.Emit
namespace.
AspectJ doesn’t provide such
feature. Can be done using the
following librairies.
ASM http://asm.ow2.org and
BCEL
http://jakarta.apache.org/bcel.
Dynamic method invocation Using thisJoinPoint to access
Yes, method invocation at
the current join point for the
run-time through reflection
advice through reflection.
[48].
Generation code at run-time
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
19/139
2.1.7 Aspect-orientation
AspectJ, natively, is an aspect-oriented programming language. AspectJ implements the joint-point
model and inter-type declaration implementation. AspectJ has been adopted as a mainstream
programming language and there are numerous articles that have been written about it [49]. It is an
active project and it is present in many important development projects and domains such as banking
systems and web transactional applications where security and logging are implemented in a modular
way using aspects [4].
C# doesn’t officially provide an implementation of AOP. There are some open source projects [51]
but it seems there are no longer active. As it mentioned on the AspectDNG project’s website: “the
project is not supported any more due to .NET rapid changes” [50]. Due to the .NET platform closed
source and its proprietary license.
Aspect-orientation, AspectJ vs C#
Feature
Aspect-oriented
programming
Modularity
AspectJ
Natively, it is an aspectoriented programming
language
Isolate crosscutting
concerns in a modular way.
C#
Partially, there are some nomainstream projects
implementing AOP for C#.
It is Difficult some times
impossible to isolate
crosscutting concerns from
the program’s business
logic.
Can produce pure
uncoupled code
Can’t isolate code that
crosscut over modules.
Partially, can be done using
DLLs in order to reduce
code coupling degree.
Not applicable since there is
no mainstream AO solution
for C#.
Code reusability
Security
Securing UI by detecting
Single-Thread UI’s rule and
web services pipelines
failure.
Errors handling/Logging
Cleaner way to handle and
log exceptions
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Exceptions are handled in
try-catch blocks and logged
within this scope.
20/139
2.1.8 Functional programming
C# and AspectJ are both imperative programming language which means a program written in
one of them consists of sequence of steps/commands that should be executed one by one (step by
step), so programs can be seen as a sequence of operations. C# supports generic classes, generic
delegates which are object-oriented and type-safe functions pointers (similar to function pointers
in C++) and anonymous methods [52] which is an advantage over AspectJ (see Example #1 below).
AspectJ doesn’t support delegates but does support generic classes. Also, AspectJ doesn’t
allow passing a function as parameter to another function. Some functional programming style in
AspectJ can be done either by using libraries or by using higher-order functions and lazy
functional programming [53]. Interfaces and inner classes can be used as a work around to pass
functions as parameters [54]. This practice is not encouraged since this will result in a code much
bigger and complex which is not easy to understand by someone who is not familiar with
functional programming.
C# has more features for supporting functional programming than AspectJ. Language
integrated query (LINQ) has been released with the .NET Framework 3.5 which provides a
powerful way to query data in a functional programming style [55].
Recursion is another concept of functional programming that requires functions, also methods,
to invoke themselves. C# does support recursion without issues [56] while using it in AspectJ has
resulted in major problem: infinite recursion. “if no special precautions are taken, aspects which
advise other aspects can easily and unintentionally advise themselves.” [57]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
21/139
Functional programming features: AspectJ vs C#
Feature
Type inference
Lambda expression
Anonymous methods
Higher-order functions
Closure
First-class
functions(delegates)
Recursion
AspectJ
C#
No, planned for Java 7
Yes, functional
programming’s features
have been introduced in C#
3.0
Yes, since C# 3.0 using
LINQ library and
System.Linq.Expression
namespace
Yes, since C# 2.0 where
functional programming
features has been
introduced.
Yes using C# delegates.
Using interfaces as a work
around.
No, use inner classes as
work around
Using interfaces as work
around
No, not popular enough an
open source project for
OpenJDK [58])
No, AspectJ doesn’t support
delegates, lazy functional
programming can be used
as work around [53].
Not secure as in C# [57]
due to infinite methods
loop.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Yes, since C# 3.0
Yes, delegates are C++’s
function pointers
equivalent.
Yes it is fully supported.
Can be done using C#’s
methods.
22/139
Example #1: An example showing how to use anonymous functions in C#
using System;
namespace Program
{
/// <summary>
/// <autor>Sleiman Rabah</autor>
/// </summary>
class Program
{
static void Main()
{
// Anonymous function to calculate the square of
// two integer.
//the first int is the x' type
//the second int is the return type
Func<int, int> square = x => x * x;
Console.WriteLine(square(7)); // Will output 49
}
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
23/139
2.1.9 Declarative programming
AspectJ and C# enable programmers to specify metadata that can be associated to some elements
in the source code and can be retrieved/consulted at run-time through reflection. They both provide a
mechanism to define declarative tags that can be added to classes, methods or variables. This
mechanism enables declarative programming style to be used in both languages. It consists in
declaring data and rules rather than coding them explicitly.
AspectJ provides what is called Annotations [59] while C# provides Attributes
[60]. Programmers can use pre-defined Annotations or Attributes and also they can create their own
customized ones.
C# does better support declarative programming than AspectJ since it provides ways to mix
imperative and declarative code using LINQ. LINQ can be used to query data from databases (LINQ
to SQL), XML (LINQ to XML) and many more [55] (see Example #2 below).
Also, C# 4.0 has introduced dynamic programming with the dynamic keyword which can be used
to work on data type.
Some new technologies have introduced new ways to build UIs based on declarative programming
principles such as XAML which was invented by Microsoft for the .NET Framework [61]. Also there
are plenty of open source project which are equivalent to XAML and provide the same features as
XAML does. Among them we find YAML [62] which is widely used to build AspectJ Swing GUIs.
Using these languages, developers can define UI’s elements and their properties outside the source
code.
Declarative programming, AspectJ vs C#
Feature
Tags on
methods/fields/classes
LINQ equivalent?
Dynamic programming
(Dynamic type checking)
Declarative programming
based on XML [41]
AspectJ
C#
AspectJ/Java Annotations
C# Attributes
Yes, system.data.linq
library should be used and
referenced.
Yes – since C# 3.0 (var
No, AspectJ doesn’t provide
keyword) and C# 4.0
such features.
(dynamic keyword)
YAML, SwiXML
Using XAML technology
which is available since
.NET 3.0 and WPF
Additional libraries, lamdaJ
[63]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
24/139
Example #2: An example showing declarative programming using C# and LINQ
using System;
using System.Collections.Generic;
// Importing the LINQ DLL
using System.Linq;
using System.Text;
namespace LinqSample
{
/// <summary>
/// <autor>Sleiman Rabah</autor>
/// </summary>
class Program
{
static void Main(string[] args)
{
// Check if a number in a list of integer is odd using LINQ
// if yes it will result in a list of odd numbers
List<int> collection = new List<int> { 1, 2, 3, 4, 5, 6, 7 };
var results = collection.Where(num => num % 2 != 0);
foreach (var num in results)
{
Console.WriteLine("Results:" + num); // Will output 1,3, 5, 7
}
// Using "SQL" style to print element from an array
// which are <= to 7 using LINQ
int[] numbers = { 3, 2, 13, 15, 5, 7};
var lowNums =
from n in numbers
where n <= 7
select n;
Console.WriteLine("Numbers <= 7:");
foreach (var num in lowNums)
{
Console.WriteLine(num);
}
}
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
25/139
2.1.10 Batch scripting
Console applications, or command line applications (CLA), are programs that take commands and
arguments as input, process them and output result/display messages at the command line. CLA often
are executed from the command line and are said text-based applications which use a TUI (stands for
Text User Interface) to interact with users/other programs using system calls in order to read their
commands from the command line and achieve certain tasks like processing files, querying databases,
etc. Before commands processing, commands and arguments are checked and validated to be sure that
they meet certain constraints. Example of validation can be arguments’ data type, commands names
whether if they are syntactically correct, arguments count for a given command and so on.
CLA can be done using both AspectJ/Java and C# [64] they both provide this possibility:
commands and arguments are received as parameters at the program’s entry point which is the main
method, usually they are received as an array of strings.
The only difference is that the main method name has to be capitalized due to the C#/.NET naming
convention (Pascal case) [65].
The method main signature is as follow:
AspectJ: public static void main(string[] args)
C#: public static void Main(string[] args)
Building CLA using both languages is a secure solution since they both are type-safe and they
provide an exception handling mechanism which can be used when processing commands. While
processing a file for example, an exception can be thrown and the program will stop executing.
These type applications have some limitations [66]: they are not flexible enough because they are
compiled and usually deployed as executables with extension .exe in case of C# and a .jar in case of
Java. If any changes in the code have been done in an existing application/program, it has to be
recompiled and re-deployed for the changes to take place which is often a disadvantage since all users
have to be notified about changes and required to update their installed copies. There are also other
limitations: they are not user-friendly and they usually target a set of trained/experimented users. On
the other hands, because they are efficient and they run faster, console applications are suitable for
applications that need bigger processing resources especially because there are no graphics
rendering/drawing and so on [66].
Also, both languages AspectJ/Java and C# enable to developers to write programs that can
invoke/execute external commands and programs. One program can execute another program while it
runs, to do so, a program creates a process to run the desired program (as illustrated in example #3
below ).
Pipeline can be used to redirect CLA output to another command or CLA. For example, under
Linux operating system, we can run a CLA and use the grep command. e.g
[prompt]$ java myProgram | grep [search pattern]
AspectJ can be coupled with Java to build a wide variety of CLA or added to an existing
application in order to advantage of AOP in such application. Exception can be logged in files using
AspectJ to be later analyzed and debugged.
CLA written in AspectJ or C# can be automated and scheduled to be executed at a given point in
time. Script automation is a useful and common way to automate things such as daily reports
generation, job alerts and so on. This can be done using Scheduled Tasks tool in Windows or using
crontab on Linux/Unix system.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
26/139
AspectJ is more suitable/better to build and deploy portable (cross-platform) CLA since such applications
can profit from the AOP’s features and AspectJ’s portability.
Batch scripting, AspectJ vs C#
Feature
Run external
programs/external
commands
Tasks automation
Accept command line
arguments
Need to be recompiled
after changing the code
source
AspectJ
Yes, using Java’s Process
and Runtime classes.
Yes, can be executed as
stand alone and executed on
a scheduled basis.
Yes as parameters for the
main method.
C#
Yes, using Process class in
System.Diagnostics
namespace.
Yes, can be executed as
stand alone and executed on
a scheduled basis.
Yes as parameters for the
Main method. i.e
i.e public static void
main(string[] args)
public static void
Main(string[] args)
Yes, code should be
recompiled in order to
source code changes take
effect.
Yes, code should be
recompiled in order to
source code changes take
effect.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
27/139
Example #3 A simple example showing how to execute a program from another program in C#
using System;
using System.Diagnostics;
using System.ComponentModel;
namespace LaunchProcessSample
{
/// <summary>
/// <autor>Sleiman Rabah</autor>
/// </summary>
class ProcessLauncher
{
public static void Main(string[] args)
{
// instantiate the System.Diagnostics.Process class
Process myProcess = new Process();
try
{
// Tell whether the new process will be executed
// using Shell or not
myProcess.StartInfo.UseShellExecute = false;
// Fill out the process name to be started:
// often it is a program e.g: notepad
myProcess.StartInfo.FileName = "notepad.exe";
myProcess.StartInfo.CreateNoWindow = true;
// Run/launch the process
myProcess.Start();
}
catch (Exception e)
{
Console.WriteLine("A problem has occured while
starting the process " + e.Message);
}
}
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
28/139
2.1.11 UI prototype design
Abstract Window Toolkit (AWT) was the first Java’s technology/component set for building
graphical user interfaces (GUI). AWT UIs were very slow to run with limited functionalities due to the
small number of provided components. Just-in-time compilers (JIT) were introduced in order to
improve applications’ execution and performance and Java Foundation Classes (JFC) technology was
released also to improve productivity and to provide a new rich set of components. JFC provides a
group of features and graphic functionalities that can be mixed together for building powerful and
extendable desktop applications. Among these features are internationalization which enables to build
applications that support different languages and cultural conventions, pluggable look-and-feel which
allows to choose a look among wide choices of look and feel and themes on different
platform/operating systems (GTK+, Motif, Windows, Macintosh) and many more. As part of JFC,
Java Swing is an API which provides a set of GUI components such as buttons, tables, text box and
features for printing, drag and drop, customizing components layout and many more [67].
AspectJ-based Swing enables developers to build either document interface (MDI) using the
JdesktopPane container and JinternalFrames or single document interface (SDI) GUIs.
The NetBeans IDE provides a user interface designer (UID) allowing developers to
design/prototype GUIs. Also, there are some plug-in for the Eclipse IDE which can be used to design
interfaces with Swing.
On the other hand, the .NET Framework was first released with the Windows Forms (WinForms)
API as a part of it. WinForms is not flexible enough and is said platform-dependent since it is built
upon the Windows API (the famous WinAPI or Win32 API) which resulted in a bad design and it
doesn’t support the model-view-controller architecture. To resolve this problem and push further the
UI development, Microsoft has introduced the Windows Presentation Foundation (WPF) a new
revolutionary technology for building platform-independent GUIs, [68]. WPF provides much better
functionalities than Swing including animations, data templates, effects, templates and many more.
Also, WPF is based on the XAML (Extensible Application Markup Language) which is a new
declarative language where developers can either manually or using the UI designer describe/declare
the properties of the UI’s controls and components without the use of the traditional imperative
programming [68]. The main goal of XAML is to isolate the graphical content from the code which
resulted in a cleaner and better understandable and maintainable code.
WinForms and WPF are not part of .NET’s Base Class Library (BCL) which means they are not
standardized as of the ECMA/ISO standards. To build applications using the WinForms the
System.Windows.Forms.dll library should be used and referenced, same for WPF where the
System.Windows.dll is required. AspectJ-based Swing is built-in Java and no libraries/references are
needed to do so (it can be directly imported using the import keyword i.e import javax.swing.*)[68].
Visual Studio C# comes with a user interface designer (UID) and Toolbox which enable users to
add control/components such as buttons, text box, menus, toolbars and so on to the design interface
and set their properties and their events handlers.
Swing, WinForms and WPF are popular and used in different development projects, but nowadays,
Microsoft’s technologies are more popular and suitable to build desktop applications. WinForms or
WPF have a high interoperability with the MS SQL Server database system. C# has a lot of
advantages over AspectJ-based Swing for prototyping UIs, a WPF code can be simply embedded in a
web ASP.NET application without any configuration since WPF can be executed in any browser [69].
There are some discussions and rumors over the
web that AspectJ/Java Swing is dead and it is no
longer used in development projects: before Java 1.5, Java Swing left a bad perception behind.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
29/139
Java Swing library has a single thread-safety rule: it uses a single thread to access all UI’s
components through the event-dispatching thread [2, chap.11]. This restriction is a problem in
complex applications, for example, if a thread is performing database queries and you want to update
the UI components. AspectJ resolves this problem: you can write an aspect to detect any violations to
this rule [2, chap.11].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
30/139
Comparing Java Swing and Windows Forms/WPF APIs
Feature
Graphical user interface
Built-in?
Look and feel
2D and 3D support
Drag and Drop support
Components’ Layout
Management
Multiple document
interface (MDI)
Model-View-Controller
Declarative GUI
development [41]
Performance
IDE/UI designer
Rich Web UI Development
Deployment
Single thread of execution
Event handling mechanism
AspectJ
Yes, using Java’s built-in
Swing and AWT libraries.
C#
Yes, using .NET’s Windows
Forms and WPF libraries.
Built on top of the Base Class
Library
You need to import/reference
Yes, part of Java
System.Windows.Forms.dll for
Foundation Classes
WinForms and
System.Windows.dll WPF. WPF
only on Windows platform [70]
Yes such as GTK+, Motif, Not flexible enough in WinForms.
Windows, Macintosh, etc
Full customizable UI support in
themes.
WPF
Yes – Java 2D API
Yes, using WPF framework
Yes, using Java.awt.dnd
Partially, only using WPF
package.
framework [71]
Yes, this can be done
Yes, i.e Border/Grid layout,
automatically using Visual
etc.
Studio’s UI designer.
Yes, using,
Yes, provided by AWT
System.Windows.Forms.MdiLayout
class.
Yes in components such as WinForms no, WPF Model-ViewJTable and JList, etc.
ViewModel [72]
Glade XML, SwiXML: not
popular yet as much as
Using XAML technology
XAML
Improved execution on windows
Relatively Slow
Fast
Visual Studio C# and
NetBeans, Plugins for
SharpDevelop for Windows,
Eclipse IDE
MonoDevelop .
Silverlight framework on windows
Yes, using JavaFX
and Moonlight implementation for
framework.
Linux and other Unix-based OSs.
.NET Assemblies: (executables
Jar archive (.jar files)
and libraries, .exe and DLLs)
Windows Forms (Not by default,
but by applying the STAThread
Yes, Swing library’s singleattribute to the Main method).
thread rule, thread-safe[73]
WPF use by default a single thread
of execution.
Only Event Listeners,
Yes, Event Handlers and Delegates
provided by java.awt.event
support
package
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
31/139
2.2
C++ vs Groovy
2.2.1 Source code size
Hello world in C++
#include <iostream>
using namespace std;
int main(int argc, char *argv[])
{
cout << "Hello, World!\n";
}
Hello world in Groovy
println "Hello, world!"
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
32/139
2.2.2 Default more secure programming practices
First of all, C++’s type strength is strong. C++ language places severe restrictions on the intermixing
that is permitted to occur, preventing the compiling or running of source code which uses data in what
is considered to be an invalid way. Secondly, C++’s type safety is unsafe. C++ can discourage or
prevent some type errors. For example, C++ can prevent result from attempting to perform operations
on values that are not of the appropriate data type [74]. Thirdly, C++’s type checking is static. C++
allows many type errors to be caught early in the development cycle. It eliminates the need to repeat
type checks every time the program is executed. Program execution may also be made more efficient
by omitting runtime type checks and enabling other optimizations. But it has some disadvantage, E.g,
C++ will reject some programs that may be well-behaved at run-time, but that cannot be statically
determined to be well-typed [74]. Finally, C++ doesn’t have garbage collector, in some sense, it will
cause dangling pointer bugs, double free bugs and certain kinds of memory leaks.
First of all, Groovy’s type strength is also strong which means that Groovy specify one or more
restrictions on how operations involving values having been executed. Secondly, Groovy’s type safety
is also safe. It is just like C++ [74]. Finally, Groovy’s type checking is dynamic. Compare to C++,
Groovy is more flexible but fewer guarantees, because Groovy accepts and attempts to execute some
programs which may be thought as invalid by C++. Groovy may result in runtime type errors.
Compare to C++, Groovy make fewer "compile-time" checks on the source code. On the other hand,
Groovy’ type checking is more sophisticated [74].
2.2.3 Web applications development
C++ uses Wt and CppCMS as an open source web application framework to support the development
of web applications. It supports Ajax ,template framework ,caching frameworks and security
framework, but doesn’t support DB migration framework [76]. It also support MVC framework [75]. It
provides SQL library and Wt::Dbo for ORM and uses boost. test in testing framework [76] [77].
This is a example of web application in C++:
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
33/139
// Displays the current date and time in a Web browser.
#include <iostream>
using std::cout;
#include <ctime> // definitions of time_t, time, localtime and asctime
using std::time_t;
using std::time;
using std::localtime;
using std::asctime;
int main()
{
time_t currentTime; // variable for storing time
cout << "Content-Type: text/html\n\n"; // output HTTP header
// output XML declaration and DOCTYPE
cout << "<?xml version = \"1.0\"?>"
<< "<!DOCTYPE html PUBLIC \"-//W3C//DTD XHTML 1.1//EN\" "
<< "\"http://www.w3.org/TR/xhtml11/DTD/xhtml11.dtd\">";
time( ¤tTime ); // store time in currentTime
// output html element and some of its contents
cout << "<html xmlns = \"http://www.w3.org/1999/xhtml\">"
<< "<head><title>Current date and time</title></head>"
<< "<body><p>" << asctime( localtime( ¤tTime ) )
<< "</p></body></html>";
return 0;
} // end main
Groovy uses Grails as an open source web application framework to support the development of web
applications.Grails supports Ajax ,template framework and caching frameworks, but dosenot support
form validation framework.It uses active record patten technology for MVC framework.It provides
GORM, Hibernate for ORM and uses multiple plugins in DB migration framework [78].
Compare to C++,Grails will take more time to write a program and is more complexity.In some
condition,the speed of excution is slow. It is neither high performance nor easy to program.But Grails
is more secure and have more feature [79].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
34/139
2.2.4 Web services design and composition
Groovy, as befits a sprightly young scripting language for the Java Virtual Machine, can do XML in a
way that is relatively free of bloat and allows the developer to focus on the real problem at hand [80].
Groovy also support WS,SOAP and WSDL [81] [82] [83]. GroovyWS comes with two sets of APIS
that are briefly described below using a simple example:Publishing a web-service and Consuming a
web service [84].
Compare to C++ in XML, the best part about the Groovy code is not that it is shorter than the
equivalent C++ code — only five line. The Groovy code is that it is far more expressive. You can
write directly with the XML.The following code show parsing an XML in Groovy,with shorter code:
def langs = new XmlParser().parse("languages.xml")
println "type = ${langs.attribute("type")}"
langs.language.each{
println it.text()
}
//output:
type = current
Java
Groovy
JavaScript
C++ is implemented with gSOAP toolkit for SOAP/XML Web services and generic (non-SOAP) C++
XML data bindings. The toolkit analyzes WSDLs and XML schemas (separately or as a combined set)
and maps the XML schema types and the SOAP messaging protocols to easy-to-use and efficient C
and C++ code 85. It also supports exposing (legacy) C++ applications as SOAP/XML Web services by
auto-generating XML serialization code and WSDL specifications. Or you can simply use it to
automatically convert XML to/from C++ data. The toolkit supports options to generate pure ANSI C
or C++ with or without STL [86][87].
The gSOAP toolkit is speed, reliability and flexibility, coupled with a proven track record and used by
some of the largest technology vendors makes it an ideal platform to develop applications using Web
services and XML processing.The gSOAP toolkit offers a comprehensive and transparent C++ XML
data binding solution through autocoding techniques [86]. This saves developers substantial time to
implement SOAP/XML Web services in C/C++. In addition, the use of XML data bindings
significantly simplifies the use of XML in applications by automatically mapping XML to C/C++ data
types.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
35/139
2.2.5 OO-based abstraction
Groovy is a pure OO language in which everything is an object. Groovy enables one class to extend
another, just as interfaces can. Groovy also support abastract class and interface. In Groovy a class can
derive from only one class, but a class can implement multiple interfaces ,in other words, it supports
multiple inheritance of types, but only single inheritance of implementation [87] [88].
C++ support OO programming paradigm, it supports inheritance, abstract class and interface. It also
supports encapsulation.Full multiple inheritance, including virtual inheritance.Support abstract
class,interface and encapsulation [91]. C++ methods can be declared as virtual functions, which means
the method to be called is determined by the run-time type of the object [89][90].
2.2.6 Reflection
Many programming languages provide built-in reflection mechanism. For example, in Java there is
special package java.lang.reflect and in Groovy, org.codehaus.groovy.reflection. But unfortunately
C++ doesn't support reflection [92].
Reflection requires some metadata about types to be stored somewhere that can be queried. Since C++
compiles to native machine code and undergoes heavy changes due to optimization, high level view of
the application is pretty much lost in the process of compilation, consequently, it won't be possible to
query them at run time.
Groovy has the capability of reflection, which enables you to gather information about your script at
runtime, including peer into any class or object and obtain detailed information on its properties,
methods, interfaces, and constants [84] For example,Groovy can examine the interfaces, public fields
and their types, check if a class is a class or an interface. When a class is unknown at compile
time,Groovy can use reflection to create objects. Groovy can also use reflection on the reflection
classes themselves [93].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
36/139
2.2.7 Aspect-orientation
Bascally, C++does not support aspect-orientation directly.But AspectC++, an aspect-oriented
extension of C++ languages,support aspect-orientation. It is based on source-to-source translation,
translating AspectC++ source code to C++.It allows modularizing cross-cutting concerns in a single
module, an aspect. AspectC++ provides a join point API to provide and access to information about
the join point [94][95].
Compare to Groovy, A major different is the modified join point model, which allows to have class,
object and control flow join points. The result is a more coherent language design. While it was
necessary to make some visible changes to the syntax and grammar of Groovy, we have preserved
most language concepts [96]. This should enable users experienced with Groovy and C++ to get
familiar with AspectC++ without much effort.However,there are some disavantage on AspectC++.The
prototype implementation of an AspectC++ compiler is still in an early stage. The compiler cannot yet
deal correctly with all AspectC++ language constructs in all possible contexts [97].
Groovy supports AOP. Adding AOP-like features into Groovy code is as simple as implementing the
GroovyInterceptable interface.Any Groovy object that implements this interface will automatically
have all of its method calls routed through its invokeMethod()Groovy AOP can provide a shortcut for
getter and setter [98].
Tt is easy to implement the powerful features of AOP in Groovy. It's not too difficult to envision how
such a capability could be built into an application to handle a lot of the boiler plate code that tends to
crowd non AOP applications.
2.2.8 Functionnal programming
Groovy is not a functional-only programming language,but it support functional programming.
Groovy's functions can be used to define functions which contain no imperative steps [99]. Groovy
use lazy evaluation and they typically hide away the hard bits so you don't need to know what magic is
happening on your behalf.
FC++ is a library for doing functional programming in C++. The library provides a general framework
to support various functional programming aspects such as higher-order polymorphic functions,
currying, lazy evaluation, [100] and lambda. In addition to the framework, FC++ also provides a large
library of useful functions and data types [101].
2.2.9 Declarative programming
Declarative programming is often defined as any style of programming that is not imperative. A
number of other common definitions exist that attempt to give the term a definition other than simply
contrasting it with imperative programming.Both C++ and Groovy do not support the declarative
programming paradigm [102], since they are both support imperative paradigm,which requires an
explicitly provided algorithm.In the declarative programming,the program is structured as a collection
of properties to find in the expected result, not as a procedure to follow.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
37/139
2.2.10 Batch scripting
Groovy is a scripting language and can be used for command line shell scripting. The simplest way to
invoke an external process in Groovy is to use the execute() command on a string [103]. For example,
to execute maven from a groovy script run this: "cmd /c mvn".execute().The 'cmd /c' at the start
invokes the Windows command shell [105]. The command line is short and clear.Since mvn.bat is a
batch script you need this. For Unix you can invoke the system shell [104][84].
C++ is a scripting language and can be used for command line shell scripting.The code use to write
command in C++ is more complex and sophosticated and Groovy.Here is the example:
string getStdoutFromCommand(string cmd)
{
// setup
string data;
FILE *stream;
char buffer[MAX_BUFFER];
// do it
stream = popen(cmd.c_str(), "r");
while ( fgets(buffer, MAX_BUFFER, stream) != NULL )
data.append(buffer);
pclose(stream);
// exit
return trim(data);
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
38/139
2.2.11 UI prototype design
One of the application development for C++ doing UI is Qt.The Qt toolkit is a C++ class library and a
set of tools for building multiplatform GUI programs using a “write once, compile anywhere”
approach. Qt lets programmers use a single source tree for applications that will run on Windows 95 to
XP.
Compare to Swing/AWT,the program code used in Qt is considerably more intuitive. This is because
Swing enforces the use of a Model-View-Controller architecture (MVC) while Qt supports, but does
not enforce, such an approach. Qt supports the development of sophisticated user interfaces.Qt has
better tool than Swing/AWT. Qt does not enforce particular programming paradigms as Swing/AWT
does [106].
Groovy uses Swing/AWT to provide a graphical user interface (GUI) for Groovy programs. The
Abstract Window Toolkit (AWT) is Java's original platform-independent windowing, graphics, and
user-interface widget toolkit. Swing was developed to provide a more sophisticated set of GUI
component than the earlier Abstract Window Toolkit [107].
Compare to Qt,the program code used in Swing/AWT is less intuitive and flexible. Swing/AWT also
supports the development of sophisticated user interfaces. Swing/AWT has some problems with
runtime- and memory-efficiency which are also problem of Groovy [109]. Swing/AWT may be
appropriate for certain projects, especially those without GUIs or with limited GUI functionality. Qt is
an overall superior solution, particularly for GUI applications [108].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
39/139
2.3
Haskell vs Java
2.3.1 Source code size
Not that the “source code size” criteria is present as one of the official criteria. However, code size
may effect the compile performance. While a language are interpreted, the larger code may take more
compile time. Moreover, due to the store space and memory size limitation, large source code may
also overall running performance. The following code examples show the code size of Haskell and
Java.
module Main where
main :: IO ()
main = putStrLn "Hello, World!"
Listing 1: Example Haskell program
public class HelloWorld {
public static void main(String args[]) {
System.out.println("Hello world!");
}
}
Listing 2: Example Java Program
From these two sample codes, It is clear that source code size of Java (see Listing 2) is longer than that
of Haskell (see Listing 1)
2.3.2 Default more secure programming practices
I am talking secure programming practices by type system, memory management, and exception
handling.
2.1.2.1 Type system
Haskell has a strong, static, and automatically inferred type system based on Hindly-milner type
inference.
Firstly, all types in Haskell are strong. Type system of Haskell prevent program from certain kinds of
type errors coming from trying to write expressions that don't make sense. Another aspect of Haskell's
view of strong typing is that it will not automatically coerce values from one type to another. For
example, the following codes are valid in Java:
int a = 10;
double b = (double) a;
Listing 3: Java explicit casts
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
40/139
But, this is invalid in Haskell, and Haskell compiler will raise a compilation error.
Secondly, in Haskell, type checking is performed during compile-time. The compiler knows the type
of every value and expression at compile time, before any code is executed. A Haskell compiler or
interpreter will detect type error, and reject our code with an error message running. Haskell's
combination of strong and static typing makes it impossible for type errors to occur at runtime.
Finally, a Haskell compiler can automatically deduce the types of almost all expressions in a program.
Haskell allows programmer to explicitly declare the type of any value, but the presence of type
inference means that this is almost always optional, not something we are required to do.[110]
The Java programming language is also a strongly and statically typed language. Java require all
variables and expressions to have a defined type that is known at compile time. However, other than
Haskell, Java supports the use of explicit casts of arithmetic values to other arithmetic types like
shown in Codes3. Moreover, Java does not support automatically inferred type system. That means
programmers has to declare the types they intend a method or function to use.
2.1.2.2 Memory management
Haskell's computation model is very different from that of conventional languages. Data are
immutable that means the only way to store every next operation's result is to create new value.[111]
Especially, every iteration of a recursive computation creates new values. Therefore, Haskell
computations produce a lot of memory garbage - much more than conventional imperative languages.
However, GHC is able to efficiently manage garbage collection. The trick is that immutable data never
points to younger values. [110] Due to this key property, Haskell uses a simplified garbage collection.
At anytime GHC can scan the last values created and free those that are not pointed to from the same
set.
Java 2 uses a great automatic memory management in the object lifecycle, thereby keeping the
developer from the complexity of explicit memory management. In Java, memory is only allocated to
objects, and there is not explicit allocation of memory. At runtime, Java employs a garbage collector
that reclaims the resources used by an object once it determines that object is not used in the future.
This automatic process makes it safe to throw away unneeded object references because the garbage
collector does not collect the object if it is still needed elsewhere. Therefore, in Java the act of letting
go of unneeded references never runs the risk of deallocating memory prematurely. [112]
2.1.2.3 Exception Handling
Haskell can use monad Maybe and Either type to achieve exception handling. The Maybe type
encapsulates an optional value. A value of type Maybe a either contains a value of type a (represented
as Just a), or it is empty (represented as Nothing). Maybe easily handle errors or exceptional cases
without resorting to drastic measures such as error [113]. The Either type is similar to the Maybe type,
with one key difference: it can carry attached data both for an error and a success ("the Right
answer").[114] Appendix 1 show how Maybe and Either handle Haskell exception.
In Java, all exceptions are objects. That is when an exception is thrown, an object is thrown. However,
not any object can be thrown -- only those objects whose classes descend from Throwable. Throwable
serves as the base class for an entire family of classes, declared in java.lang, and all exception is
subclass of Throwable. Throwable has two direct subclasses, Exception and Error. Exceptions are
thrown to signal abnormal conditions that can often be handled by some catcher, though it's possible
they may not be caught and therefore could result in a dead thread. Errors usually introduce system
internal situation, and are thrown for more serious problems, such as OutOfMemoryError, that may
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
41/139
not be so easy to handle. In general, Java code only throw exceptions, not errors. Figure 1 show A
partial view of the Throwable family.
Figure 1: A partial view of the Throwable family.
In addition to throwing objects whose classes are declared in java.lang, Java programmer can also
throw objects of own design by declaring new class as a subclass of some member of the Throwable
family. To catch an exception, Java uses a try block with one or more catch clauses. Each catch clause
specifies one exception type that it is prepared to handle. The try block places a fence around some
code that may throw exceptions. If the code delimited by the try block throws an exception, the
associated catch clauses will be examined by the Java virtual machine. If the virtual machine finds a
catch clause that is prepared to handle the thrown exception, the program continues execution starting
with the first statement of that catch clause. Appendix 2 is a sample code to show how Java handle a
exception.
2.3.3 Web applications development
Web applications are nature distributed applications running on more than one computer and can be
access through a network or server. Specifically, web applications are accessed with a web browser
and are popular because of the ease of using the browser as a user client. [115]
There are many approaches to Haskell web programming, such as Web Authoring System Haskell
(WASH), Haskell Application Server (HAppS), Happstack, and Haskell Server Pages (HSP). WASH
is a domain-specific embedded language with type-safe forms handling and threads continuation
through client. This gives good back-button and session splitting properties. HApps is a complete
system including web server in one program with using of XSLT for output. Happstack is the
successor to HAppS. It is a complete system including a web server and database system in one
program. It has many template options including HSP, HStringTemplate, Hamlet, XSLT, and more!
Haskell Server Pages (HSP) uses preprocessor with dynamic compilation to make XML tags into
Haskell expressions. Moreover, another simple, portable, and relatively light-weight approach to
Haskell web programming is a CGI library and an XHTML combinator library. Here is a very simple
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
42/139
example which just outputs some static HTML using CGI library and XHTML. Appendix 3 shows
two simple examples. The first one just outputs some static HTML, and the second one show how to
get user input. [116]
Java uses J2EE platform to create web applications. There are too many Java technologies can be used
to create web applications, such as Java servlet API, JavaServer Pages, JavaServer Pages Standard Tag
Library, JavaServer Faces Technology, Java Message Service API, and so on. In a Java web
application, components are either Java servlets, JSP pages, or Web service endpoints. The interaction
between a Web client and a Web application is illustrated in Figure 2.
Figure 2, Java Web Application Request Handling [10]
The client sends an HTTP request to the Web server. A Web server that implements Java Servlet and
JavaServer Pages technology converts the request into an HTTPServletRequest object and then
delivered to a Web component, which can interact with JavaBeans components or a database to get
dynamic content. The Web component can then generate an HTTPServletResponse or it can pass the
request to another Web component. Finally a Web component generates a HTTPServletResponse
object. The Web server converts this object to an HTTP response and returns it to the client. Appendix
4 show JavaServlet handle HttpServlet Request and send HttpServlet Response.
2.3.4 Web services design and composition
Since a large Haskell function is composed of a number of smaller functions, it is ideal for composing
services. In order to compose service operations, it is necessary to enable the viewing of services as
side-effecting functions in Haskell. Many features in Haskell also provide an ideal platform on which
various data-processing applications can be elegantly constructed, and this motivates the ability to
create atomic services in Haskell. The Haskell Application Interoperation Framework/Architecture
(HAIFA) is a library providing the basic components for web-based interoperability, in the form of a
service modeling framework. Firstly, the Simple Object Access Protocol (SOAP) can help to set webservice architecture. Secondly, HaXML, the most well know XML library for Haskell, provides
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
43/139
Haskell2Xml type-class, which performs serialization and deserialization of Haskell types to XML.
After created a suitable XML serializer, user can build a client side service interoperation, and then
publish the Haskell function as service[117].
Java has been a powerful development platform for Service-Oriented Architecture (SOA) since 2006.
Java EE 5, released in May 2006, significantly enhanced the power and usability of the Web Services
capabilities on the application server. Java API for XML-based RPC (JAX-RPC) is a technology for
building Web services and clients that use remote procedure calls (RPC) and XML. Often used in a
distributed client-server model, an RPC mechanism enables clients to execute procedures on other
systems. JAX-RPC provides an easy to develop programming model for development of SOAP based
Web services [118] In JAX-RPC, a remote procedure call is represented by an XML-based protocol
such as SOAP. The SOAP specification defines the envelope structure, encoding rules, and
conventions for representing remote procedure calls and responses. These calls and responses are
transmitted as SOAP messages (XML files) over HTTP. JAX-RPC can provide a big advantage for
both client side and service side -- the platform independence of the Java programming language.
Moreover, JAX-RPC is not restrictive. That is a JAX-RPC client can access a Web service not running
on the Java platform, and vice versa. [119]
2.3.5 OO-based abstraction
Contrast subtype polymorphism of object-oriented languages, Haskell provides type-class-bounded
and parametric polymorphism. Ad-hoc polymorphism (user-defined overloading) can be handled by a
powerful abstraction mechanism provided by type classes. The basic idea behind type classes is that
class declarations allow one to group together related methods (overloaded functions), and instance
declarations prove that a type is in the class, by providing appropriate definitions for the methods.
Here
are
some
standard
Haskell
declarations.[120]
class Eq a where (==)::a->a->Bool
instance Eq Int where (==) = primIntEq
instance Eq a => Eq [a] where
(==) [] [] = True
(==) (x:xs) (y:ys) = (x==y) && (xs==ys)
(==) _ _ = False
Listing 4: standard Haskell declarations[120]
And then, we can extend the type class hierarchy by introducing new subclasses.
class Eq a => Ord a where (<)::a->a->Bool)
instance Ord Int where ...
instance Ord a => Ord [a] where ...
Listing 5: introducing new subclasses[120]
The above class declaration introduces a new subclass Ord which inherits all methods of its superclass
Eq. For brevity, we ignore the straightforward instance bodies.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
44/139
Moreover, the OO features are introduced in Haskell as the O'Haskell library, based on the HList
library of extensible polymorphic records with first-class labels and subtyping. Not only O'Haskell
provides the conventional OO features, it has also language-engineered several features that are either
bleeding-edge or unattainable in mainstream OO languages [121]: for example, first-class classes and
class closures; statically type-checked collection classes with bounded polymorphism of implicit
collection arguments; multiple inheritance with user-controlled sharing; safe co-variant argument
subtyping.[122]The Appendix 5 show an O'Haskell sample code.
Java languages designed mainly for Object-Oriented programming. It support all features of OO
programming technique, such as data abstraction, encapsulation, modularity, polymorphism, and
inheritance. The Appendix 6 show an Java sample code
2.3.6 Reflection
Haskell supports some kinds of reflection such as monads. A monadic-style functional program allows
the imperative, behavioral view of a computational effect to be identified with a declarative, databased
view in a uniform way. Monadic reflection looks like a formal bridge between the two views. It
refines the identification into an observational isomorphism, with explicit reification and reflection
functions mediating between views. Such a separation allows the programmer to reason robustly about
monadic effects according to the declarative view, while implementing the imperative view much
more efficiently in terms of lower-level control and state manipulations. [123] Monadic reflection is
essentially a grammar for describing layered monads or monad layering. In Haskell describing also
means constructing monads. This is a higher level system so the code looks like functional but the
result is monad composition - meaning that without actual monads (which are non-functional) there's
nothing real / runnable at the end of the day.
Reflection is a feature in the Java programming language. It allows an executing Java program to
observe, examine, and modify internal properties of the program. For example, it's possible for a Java
class to obtain the names of all its members and display them. One tangible use of reflection is in
JavaBeans, where software components can be manipulated visually via a builder tool. The tool uses
reflection to obtain the properties of Java components (classes) as they are dynamically loaded. The
following code show how Java reflection working
// Without reflection
String stringObj = "This is a string";
system.out.println("Length of string is: "+stringObj.length());
// With reflection
Class cls = Class.forName("java.lang.String");
Method method = cls.getDeclaredMethod("length", new Class[]{});
int length = (Integer)method.invoke(stringObj, new Object[]{});
system.out.println("Length of string is: "+length);
Listing 6: Java Reflection
2.3.7 Aspect-orientation
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
45/139
Type classes can help Haskell to achieve Aspect-Oriented Programming (AOP). Type classes provide
some form of type-safe cast to encode AOP in the setting of a strongly typed language by exploiting
GHC’s overlapping instances. AOP Haskell extends the Haskell syntax by supporting top-level aspect
definitions of the form:
N@advice #f1, …, fn# :: ( C => t ) = e
In the definitions, N is the name of the aspect, and each fi is function symbols as joinpoint. C => t is
type annotation following the Haskell syntax for types. e is the advice body following Haskell syntax
for expressions. The advice will be applied if the type of joinpoint fi is an instance of t such that
constrains C are satisfied. And then, by turning advice into type class instance and instrument
joinpoints with calls to a “weaving” function – a process that intercept calls to joinpoints and re-direct
the control flow to the advice bodies - AOP idioms can be translated to type classes as supported by
the Glasgow Haskell Compiler (GHC). Appendix 7 show a AOP Haskell example [120]
Java has a simple and practical extension, AspectJ, which provide aspect-oriented programming
(AOP) capabilities for Java. This allows developers to reap the benefits of modularity for concerns that
cut across the natural units of modularity. Like any other aspect-oriented compiler, the AspectJ
compiler includes a weaving phase that unambiguously resolves the cross-cutting between classes and
aspects. AspectJ implements this additional phase by first weaving aspects with regular code and then
calling the standard Java compiler.[124] Appendix 8 show a AspectJ example and its running output
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
46/139
2.3.8 Functional programming
Haskell is an advanced purely functional programming language. The underlying model of
computation is mathematical concept of a function, and programs are executed by evaluating
expressions. In Haskell, functions are first-class, which means that they are treated like any other
values and can be passed as arguments to other functions or be returned as a result of a function. Being
first-class also means that it is possible to define and manipulate functions nested in code blocks.
Special attention needs to be given to nested functions, called closures that reference local variables
from their scope. If such a function escapes their block after being returned from it, the local variables
must be retained in memory, as they might be needed later when the function is called. Usually it is
not possible to determine statically the moment of release, which means that an automatic memory
management technique has to be used. The following code is a simple Haskell function define.
hyp :: Float → Float → Float
hyp x y = sqrt (x*x + y*y)
Listing 7: function define in Haskell
Basically, Java does not allow composition of functions. In functional programming languages such as
Haskell, functions are first-class, while in Java, class is first-class. However, a extension library,
Functional Java provide optional functional programming in Java. The library implements several
advanced programming concepts that assist in achieving composition-oriented development.[125] The
following code is a example of Functional Java
final Array<Integer> a = array(3, 2, 1);
final Array<Integer> b = a.map({int i => i + 2});
arrayShow(intShow).println(b); //{5,4,3}
Listing 8: Functional Java
2.3.9 Declarative programming
In computer science, declarative programming is a programming paradigm that expresses the logic of
a computation without describing its control flow. [126]That means the language focus on what need
to do and not how to do. As a pure functional language, Haskell support declarative programming. The
following code shows how to write code in declaration style.
filter :: (a → Bool ) → [a] → [a]
filter p [ ] = [ ]
filter p (x:xs ) | p x = x : rest
| otherwise = rest
where
rest = filter p xs
Listing 8: filter function with Declaration style
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
47/139
From this example, we can see that declaration style define a function by many equations, and each
equation uses pattern matching and/or guards to identify the cases it covers.
Java is an imperative programming language which gives a list of instructions to execute in a
particular order. However, JDK 1.5 release a powerful language construct, annotation. Annotation is
a generic mechanism for adding tags with data (metadata) with program elements such as classes,
methods, fields, parameters, local variables, and packages.[127] By annotation, declarative
programming style can be introduced into Java language. The following codes show how to declaring
an annotation and how to use it.
// Declaring an Annotation
Package njunit.annotation;
import java.lang.annotation.*;
@Retention(RetentionPolicy.RUNTIME)
@Target({ElementType.METHOD})
public @interface UnitTest {
String value();
}
//Using an @UnitTest annotation
import njunit.annotation.*;
public class Example {
@UnitTest(value="Test 1. This test will pass.")
public void pass() {
assert 10 > 5;
}
@UnitTest("Test 2. This test will fail.")
public void fail() {
assert 10 < 5;
}
}
Listing 9: Example of Java Annotation[130]
2.3.10 Batch scripting
Both Haskell and Java can create batch scripting because they can process external commands in
automated mater. For example, Haskell can use function system and rawSystem to executing an
external command. The basic synopsis of system and rawSystem are:
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
48/139
system :: String
IO ExitCode
rawSystem :: String
[String] ->IO ExitCode
Listing 10: synopisis of system and rawSystem[128]
In Linux system we can play them like:
system “ls -l”
rawSystem “ls” [“-l”]
Listing 11: Example of system and rawSystem
In Java, Runtime class can be used to executing an external command. The following is an example.
import java.io.*;
public class someClass {
public static void main(String[] args)
{
try
{
Process proc = Runtime.getRuntime().exec("ls -l");
InputStream p = proc.getInputStream();
BufferedReader reader = new BufferedReader(new
InputStreamReader(p));
String line;
while((line=reader.readLine()) != null){
System.out.println(line);
}
}
catch (Exception e)
{
System.out.println(e.getMessage());
e.printStackTrace();
}
}
}
Listing 12: Example of Java Runtime
2.3.11 UI prototype design
There are several GUI toolkits available for Haskell. However, there is no standards one and all are
more or less incomplete. Generally, there are three level Haskell GUI library, low-level, mediumlevel, and high-level. Low-level,such as GLFW, GLUT, TclHaskell, Win32, and X11, are going well,
but it is doing everything in the IO monad. High-level abstractions, such as FG, FranTk, Fruit,
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
49/139
Fudgets, Grapefruit, GuiTV, Phooey, and wxFruit, are pretty experimental[116]. The medium-level,
such as wxHaskell, Gtk2Hs, hoc, and qtHaskell, can support Haskell GUI.[128]
wxHaskell is a portable and native GUI library for Haskell, which provides a interface to wxWidgets
toolkit[129].
Gtk2Hs is a GUI library for Haskell based on Gtk+ which is an extensive and mature multi-platform
toolkit for creating graphical user interfaces. [131]
HOC (documentation at sourceforge) - provides a Haskell to Objective-C binding which allows users
to access to the Cocoa library on MacOS X [132]
qtHaskell is is a set of Haskell bindings for the Qt Widget Library from Nokia[133]
Appendix 11 shows how to build GUI using Gtk2Hs.[134]
In Java, there are two components to support GUI design, AWT and Swing. AWT, Abstract Window
Toolkit, is Java's original platform-independent windowing, graphics, and user-interface widget
toolkit. The AWT is now part of the Java Foundation Classes (JFC) — the standard API for providing
a graphical user interface (GUI) for a Java program.
Swing is a built-in GUI component technology of the Java platform. It is successor to AWT. Swing
replaces some function of AWT such as using javax.swing.JTextField instead of java.awt.TextField.
On the other hand, Swing builds on AWT. For example JtextField is descendant of java.awt.Container.
Appendix 10 give a sample code of Java Swing
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
50/139
2.4
PHP vs Scala
2.4.1 Source code size
Hello world in PHP
<?php
echo "Hello, world!";
?>
Hello world in Scala
object HelloWorld extends Application {
println("Hello, world!")
}
2.4.2 Default more secure programming practices
PHP is a loosing typing and dynamic checking language and the type safety is not that safe. It enforces
run-time checking which can provide flexibility to try evaluating compile time error type.
PHP does not require (or support) explicit type definition in variable declaration; a variable’s type is
determined by the context in which the variable is used [135].
PHP 5.3 introduces garbage collection (GC) to handle the memory allocation and de-allocation
problems. In PHP, when there is no variable points to the object, those kinds of objects would be
considered as garbage, and it will automated freed in the memory, which avoid the memory leak.
When one PHP session is ended, the memory space occupied by those sessions will be destroyed. All
the objects in the current applications would also be destroyed in the meantime. Generally, the process
of GC in PHP is followed by the start of one session. When session is end, it automatically destroys
those files [136].
PHP is popular with its strong extension API after PHP 3.0, which support PHP wrapping with some
other existing library functions for certain purpose. This rich API will help PHP integrate with other
more completely. As PHP is usually used as a web server module, you should be aware that PHP
might be used in threaded an environment which means that global variables might lead to rare
conditions [137]. When small PHP program’s leak would not be a big problem to OS. If PHP associate
with server service(i.e Apache server), it would be a long term running period, then the memory leak
would be huge problem to the OS. Even if your program is well-written format to avoid leak issue, it
is necessary to have some memory management to avoid them. With the help of extension API, PHP
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
51/139
can avoid that leak problem. In extension API (the Zend engine), it provides a set of wrapper function
(Like emalloc (), efree () similar to the method malloc (), free () in C language) to handle the memory
problems [138].
PHP do support exception handling like java using “try-catch-throw” style to deal with various of
types of errors [139].
Static typing, or compile-time type checking, helps you define and verify interface contracts at
compile time. Although Scala is static typing languages, unlike some of the other statically typed
languages, does not expect you to provide redundant type information. You don’t have to specify a
type in most cases, and you certainly don’t have to repeat it. At the same time, Scala will infer the type
(Type inference) and verify proper usage of references at compile time. Scala is considered as a type
safety languages. You will gain two benefits from Scala type system [140]:
1. The compile time checking will give you confidence that the complied code meets certain
defense expectation.
2. Help you express the expectation on your API in a complier-verifiable format
Scala runs on top of JVM which means that your code is compiled into bytecode. That in turn means
that JVM uses garbage collection as your Scala program is running - garbage collector (GC) cleans up
the memory when objects you allocated in your code become obsolete [141]. Scala also supports a
java-like style exception handling, which using “try {…} catch {…}” block to handle error
exceptions. But it would not so effectively to handle the run-time exception [142]
In sum up on this criterion, PHP is dynamic type with weak type strength, while Scala is static type
with strong type strength and support type inference. Although dynamic type provides flexibility to
PHP type system, it also brings unsafe factors. In the memory management part, Scala is only support
the garbage collection type, while PHP can gain support from the extension API(zend engine) to have
flexibility deal with specified problems(multiple-threading problems). Both languages provide similar
syntax exception handling.
2.4.3
Web applications development
PHP is associated with HTML to serve as dynamic website development tools.PHP runs on the
Apache server to collect user data via HTML form and process them on the server side. After
interpreted by web server, it generate website. The PHP web application will collect the user
information and store them in the session, thus it is necessary to verify the validity of user information,
especially in case of tamping by hacker. Thus, it requires some security issues at the server end. For
instance, safety handling the user input in registration validates the user URL etc. This technique
would make scripts safer. There are two common forms to for PHP to organize website safety: “one
script for serves all” and “one script serve one function”. Whatever forms you take, validate every
element from outside the server end.
It comes with the following possible benefits for developing web application using PHP [143]:
• It is especially design for the web, generally for the server end.
• It is a robust and proven platform.
• Since this is an open-source, it is constantly upgraded through community development.
• It is also highly customizable, and it is easily adaptable to suit for all uses.
• The web application in PHP reduces the web application development costs further.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
52/139
Lift is a free leading web application framework that aims to deliver benefits similar to Ruby on Rails.
But it is written in Scala, not Ruby. The use of Scala means that any existing Java library and Web
container can be used in running Lift applications [144]. Lift applications benefit from native support
for advanced web development techniques such as Comet and Ajax. And lift is often used for
commercial websites. Now the latest version of Lift 2.0, which provide more industrial optimization
and support.
It comes with following possible benefits that Lift-based application [144]:
• Lift is like a breath of fresh air: concise, elegant and robust
• Lift's high performance and scalability
• Lift's built-in support for REST and other web services
• Lift's use of Scala's type-safety so your tests can focus on business logic.
• Built-in security means more time focusing on your application and less time being
defensive about parameter tampering, SQL injection, Cross Site Scripting and other
nasty attacks.
2.4.4 Web services design and composition
Web services are typically application programming interfaces (API) or Web APIs that are accessed
via Hypertext Transfer Protocol (HTTP) and executed on a remote system hosting the requested
services. Web services tend to fall into one of two camps: big Web services and RESTful Web
services.” Big Web services" use Extensible Markup Language (XML) messages that follow the
SOAP standard and have been popular with traditional enterprise [145].
In PHP 5, all XML extensions have been rewritten to use the superb libxml2 XML toolkit. PHP 5
comes with a SOAP extension ext/soap, which has even more features than PEAR::SOAP, and is
written in C instead of PEAR::SOAP, which is written in PHP. [14] PHP 5 do supports web service
design and composition, and support different protocols such as OAuth, SCA, SOAP and XML-RPC
[146]. PHPXMLRPC is PHP library for implanting the function of XML-RPC.It designed for ease of
use, flexicibility, and completeness. But this library does not consider the speed and memory
management issues. With the support for XML-RPC, it enable distributed systems and system
interoperability, it allows applications call the methods in other place no matter what machine is, what
language written with. Such as C++ function can call the PHP methods. This achieves the purpose of
web service. New XML support makes PHP the best language available for processing XML and,
coupled with new SOAP support, an ideal platform for creating and using Web Services [147].
Scala do support the Web service building.
Scala's notion of pattern matching naturally extends to the processing of XML data with the help of
right-ignoring sequence patterns. In addition, Scala’s XM; support is partly from library, with some
built-in syntax support. It allows using inline-XML that makes XML in Scala in a convenient form.
Since Scala can seamlessly integrate with Java’s library, thus Scala can using the XML-RPC in Java,
which is a java implementation of XML-RPC protocol for web service purpose. In this context,
sequence comprehensions are useful for formulating queries. These features make Scala especially
ideal for web services [148].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
53/139
[18] Listing 1 How XML-RPC work
2.4.5 OO-based abstraction
Both PHP and Scala do support OO-based abstraction.
PHP 5 support Object Oriented programming (OOP) paradigm, it supports single inheritance, interface
(abstract method and class). It also supports encapsulation.
The purpose of inheritance in PHP is to inherit the characteristic of the base class that does make code
reusable and structural. Using keyword “new” to instantiate a new object from non-abstract class and
keyword “extends” to declare which class inherits from which class. That’s somewhat similar way to
Java. PHP also support abstract class, which any class extends an abstract class must create an
interface (Interface in PHP will define your own common structures and can not be instantiated itself)
of the parent abstract methods. Additionally, in order to achieve some level of encapsulation purpose,
PHP use some keywords (public, private, protected) to manipulate the visibility of class members
(properties or methods). In the source code example A1 [149], it simplify illustrate an abstract class
example.
Scala is a pure OO language, everything in Scala is an object. It supports single inheritance and traits,
which is like optional implemented interfaces. It is quit similar to the interfaces in Java, traits are used
to define objects by specifying the signature of the supported methods. It is allowed for traits to be
partially implemented, while there is not allowed to implement the abstract class in PHP. This implies
that you can define default implementation for some methods. [151] Compared with the classes, trait
may not require constructor parameter. It also provides similar way to encapsulate data members as
PHP did. In the source code example A2 [152], it give a trait version that implement abstract class.
To sum up, two languages have similar way in OO-based abstraction. Compared to PHP, Scala have
more advanced OOP characteristics (such as Case Classes, Traits, and Companion Objects) which
make the data member more flexible in the OO framework.
2.4.6 Reflection
Reflection is the process by which a computer program can observe and modify its own structure and
behavior. A reflection-oriented program component can monitor the execution of an enclosure of code
and can modify itself according to a desired goal related to that enclosure. This is typically
accomplished by dynamically assigning program code at runtime [153].
PHP 5 comes with a complete and rich reflection API that adds the ability to reverse-engineer classes
[154], interfaces, functions, methods and extensions. Additionally, the reflection API offers ways to
retrieve doc comments for functions, classes and methods. Using reflection is actually let
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
54/139
you know how things actually worked. Reflection itself is handled through various classes, the root of
which is the ReflectionClass. There's other class like ReflectionExtension, ReflectionFunction,
ReflectionMethod, ReflectionObject, Reflection Parameter [155].
Scala supports the same reflection capabilities that Java and .NET support. [156] The syntax in Scala
would be somewhat different depending on the cases. The following code [157] segment shows that
some reflection methods in JVM, through java.lang.Object and java.lang.Class:
Listing2 Reflections on
Scala using Traits
trait T[A] {
val vT: A
def mT = vT
}
class C extends T[String] {
val vT = "T"
val vC = "C"
def mC = vC
class C2
trait T2
}
val c = new C
val clazz = c.getClass
// method from java.lang.Object
val clazz2 = classOf[C]
// Scala method: classOf[C] ~ C.class
val methods = clazz.getMethods
// method from java.lang.Class<T>
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
55/139
2.4.7 Aspect-orientation
Aspect Oriented Programming is new programming paradigm which enables you to effectively
implement crosscutting concerns in new and also existing programming without changing the original
code. Unfortunately, it is possible in complied programming languages, but not for scripting language.
Since the PHP is a scripting language, the written code is interpreted, not complied in the server end.
In PHP5, it brings to new approach to solve the weaving problems.
To solve this problem of weaving before executing the program, some implementations, of the AspectOriented Programming paradigm in PHP, use another script to read the file that is interpreted firstly,
and then all Advices are added on the designated locations in the code. After this, the newly created
file is interpreted. This requires another script, time to read the original file, a lot of find and replace
operations and finally interpreting a non-optimized new file. This can really slow down the execution
time of the PHP code, especially with large programs with a lot of Aspects. Another common used
solution is to add some libraries to the PHP interpreter. These libraries then take care of the weaving
before executing the file. This solution doesn’t remarkably slow down the interpretation and execution
process, but needs some extra skills to create the library. Furthermore the PHP installation on the
server is modified to enable these libraries. That can result in less stable services. In the source code
example A3 [158], it is the second way which provide external library to help achieve the AOP
purpose into PHP.
Scala have the ability to interoperate with AspectJ, an extension of Java, which supports Aspectoriented programming [159]. When using AspectJ in Scala, we have mentioned some detail issues:
how to reference Scala execution points and how to invoke Scala code as advice Scala traits is a better
choice rather than using AspectJ for the purpose of AOP [159]. However, Scala doesn’t have a point
cut language, like AspectJ. Using aspect would be more suitable for more “pervasive modification”
(e.g. Tracing, police enforcement, security) while using trait is no need to worry about other language
(traits is not from the Java or .NET) and fits your entire requirement. [160] In the source code example
A4[161], it is an aspect that logs methods calls to the Complex class.
In sum up, Scala provide two ways to achieve AOP purpose: Built in approach (using traits) or Aspect
approach (using Aspect), while PHP also provide two ways to handle AOP, but seems rather
complicated than Scala did.
2.4.8 Functional programming
Since PHP 5.3 support Closure, then PHP can support functional programming with more
convenience. Before the Closure come, we need to implement curry function (curry is not built-in
function in PHP, need to be defined by developer) to achieve the same goal as Closure, which is a
technique to transform functions from multiple arguments to one argument. Assume the following
code segment [162]:
$add2 = curry ($add, 2); // returns a new function reference, to add (), bound with 2 as the first
//argument
$add2 (8);
// return 10
Listing 3 using curry method for functional programming purpose
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
56/139
Scala support functional programming paradigm.Scala is also a functional language in the sense that
every function is treating as value. Scala provides a lightweight syntax for defining anonymous
functions, it supports higher-order functions, it allows functions to be nested, and supports currying.
Scala's case classes and its built-in support for pattern matching model algebraic types used in many
functional programming languages. [163] In the following code segment, it is the code example of
high-order function of Scala [165]:
def increment(x:Int) = x + 1
def handler( f:Int => Int) = f(2)
println( handler( increment ))
//pass the function as value
Listing 4 High order function
Scala uses the keyword lazy defers the initialization of a variable until this variable is used. Evaluation
of delimited continuations is supported in version 2.8. Tail call optimization is not supported
completely, because the JVM lacks tail call support. In simple cases, the Scala compiler can optimize
tail calls into loops [166]. In the following code segment, it is code example of lazy evaluation in Scala
[167] (Since Scala is not default lazy evaluation, it requires keyword “lazy” to evaluate when in
needed):
Listing 5 Lazy
def lazyMultiply(x: => y:) = { lazy val y = x * x }
evaluation
2.4.9 Declarative programming
Compared to imperative programming, the main point of declarative programming to focus on what to
do, and do not concern on how the steps should be.
PHP is a scripting language which can support declarative programming paradigm under the help of
annotation. Annotation can be used by tools or libraries which do not affect the semantic structure of
program; it adds metadata to the properties (classes, methods) [169]. In addition, PHP can embelled
with HTML, which is a declarative user markup language just to tell what the server end should do to
support the interface [170],
Since Java can support declarative programming through annotation and reflection [171], and Scala do
support both techniques, thus, we consider Scala do also support declarative programming, may have
somewhat different in code syntax.
2.4.10
Batch Scripting
PHP is a scripting language which more than just web application and can be used for command line
shell scripting. PHP supports a CLI SAPI (a layer that PHP interact with other web server (i.e. Apache
and M$ IIS)) through as of PHP 4.3.0. The main focus of this SAPI is for developing shell
applications with PHP. That is to say we can use CLI as command-line tools same as the standalone
server application. This make PHP look more like traditional scripting languages. It is possible to
write shell scripts using CGI version of PHP, although CGI have some significant different (In CLI, no
header written to the output) with the CLI in web server integration. For the reason of accessibility
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
57/139
(default installed) and consistent, developer prefer using CLI to write PHP shell scripts. [40] In the
following, it tells PHP to execute specified files:
Listing 6
$ php -f comp6411script.php
execute the
.php file in
command line
Scala is similar to PHP, which its program can also run as a shell script as a batch command. For
instance, The “scalac” command compiles one or multiple Scala source files and generates Java
bytecode on the base of JVM, which the Scala complier work similar to Javc(java complier) [172]
> scalac HelloWorld.scala
Listing 7 execute the .scala file in command line
2.4.11
UI prototype design
PHP itself does not support UI prototype design generally, PHP was embelled into HTML file to
invoke HTML’s features to obtain the graphical interface purpose. Strictly speaking, we should say
PHP does not support UI prototype design directly. But after the coming of the PHP-GTK, it reduces
the dependence on the UI purpose with HTML. That is to say, we can use PHP create stand-alone GUI
application. PHP-GTK [173] is an extension for the PHP programming language that implements
language bindings for GTK+. It provides an object-oriented interface to GTK+ classes and functions
and greatly simplifies writing client-side cross-platform GUI applications.PHP. We can use this tools
create windows, buttons, and text without help by another program (i.e. browser, and text editor).
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
58/139
// show an window using gtk library
<?php
$window= new GtkWindow();
$window->connect_object(‘destroy’,array(‘gtk’, ‘main_quit’));
$windown->show();
gtk::main();
?>
Listing 8: A
Simple
Application
in PHP-GTK
[174]
Scala’s
graphical user
interface
(GUI) development is heavily depending on the Java GUI library. It works well with Java AWT,
swing. But there are some swing Scala library to choose like: ScalaGUI [174], scala-swing (Scala
API).
package example
import scala.gui._
// import from Scala library
object application extends scala.gui.Application {
val mainWindow = new container.Window {
val press = new widget.Button {
text = "Press me, please"
subscribe(this)
toplevel eventloop {
case this.Click() =>
field.text = "Wow! Someone pressed me."
}
}
Listing 9: A
simple
segment for
create a
window
and button
using
ScalaGUI
But
in
general case,
invoking
Java GUI is
more
convince and with rich library choice. Because there are not much Scala built-in GUI library for use in
the API.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
59/139
2.5
Scheme vs JavaScript
2.5.1
Source code size
Hello world in Scheme
(display "Hello World!")
Scheme Code sample
Hello world in JavaScript
document.write('Hello World!');
Javascript Code sample
2.5.2 Default more secure programming practices
Type system:
Scheme uses strongly but dynamically typed variables. In Scheme, a variable can refer to a
value of any type, and the majority of its type checking is performed at run-time. However, at runtime, Scheme rejects any operations which attempt to disregard data types. However, Javascript uses
weakly and dynamically typed variables. That is, In Javascript, any kind of variables can be stored into
a variable. Programmers just need to simply declare a variable without assigning it a type
Memory Management:
Both Scheme and Javascript uses automatic garbage collection for heap memory management.
That is, Scheme memory management system can scan and reclaim all dead objects, i.e. objects that
will not be used in future, automatically. This frees the programmer from having to explicitly
deallocate memory themselves.
Exception Handling:
A continuation reifies an instance of a computational process at a given point in the process's
execution. It contains information such as the process's current stack (including all data whose lifetime
is within the process e.g. "local variables"), as well the process' point in the computation [176]. In
Scheme, Continuations can form the basis for implementing a simple throw/try/catch-style exception
handling mechanism in just a few lines. However, things get rather more complicated if the
mechanism is to work in programs that themselves use continuations. Therefore, several Schemes
come with built-in exception handling capabilities [177]
Javascript provide an exception handling mechanism allowing preventing application from
crashing at run-time. Programmers can use try-catch to handle exception.
2.5.3
Web applications development
Many implementation of Scheme can be used to develop a web application, and Schemes with an
FFI (foreign function interface) can also support web programming. For example, openScheme and
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
60/139
RScheme have build-in CGI, incoming and outgoing HTTP, and HTML generation [178]. PLT
Scheme has several web programming packages available, such as complete web server, CGI, MIME,
and cookie handling. Gauche also has cgi and html generation bits in its standard library. JScheme can
work in a similar way to Java Server Pages. Mod_lisp is an Apache module to easily write web
applications in Lisp/Scheme[179]. LAML is a Scheme-based set of libraries for server side web
programming. [180]
Originally, Javascript for a web application are created to run on client-side, refer to Client-side
Javascript (CSJA). Now, server-side Javascript (SSJA) is available. The first implementation of SSJS
was Netscape's LiveWire, which was included in their Enterprise Server 2.0 back in 1996. Since then,
a number of other companies have followed suit in offering an alternative to the usual server-side
technologies, such as ASP of Microsoft and Jaxer of Aptana. These system support Javascript for
accessing database, file system, send email and so on.
2.5.4
Web services design and composition
Racket (or PLT Scheme) programming language has a massive set of libraries and tools which are
suitable for specific domains such as web applications development, querying databases (MySQL.plt)
and batch scripting and user interface prototyping [181].
Scheme does support HTTP transactions also it does support web development aside CGI
programming. Scheme support also XML documents parsing and writing which is an advantage for
supporting web services transactions [182]. Racket provides the xml library for parsing and generating
XML documents which are represented in a structure type called X-pression [183].
Scheme does support web services [184]. Using Scheme, you can call web services methods
through CGI programming facilities. Using Scheme, Amazon Web services (AWS) can be called in
order to retrieve data from Amazon’s database [185].
There is an implementation of Scheme called Gambit which enables to write a stand alone executable
web service. This implementation requires the Gambit-C which is a compiler that generates portables
C code [186].
JavaScript is a scripting language which runs in web browser as part of them. In web applications,
mostly, it is used on the client-side to ensure data validation and interaction with Document Object
Model by embedding it within HTML tags.
There are some implementations for JavaScript that run on the server-side. Thus, using Ajax API,
a web application can communicate asynchronously with a web server in an interactive mode. This
communication enables you to retrieve data form the web server using the XMLHttpRequest object
[187].
JavaScript enables to call web services using Ajax (the XMLHttpRequest object) in order to
retrieve customized data. Yahoo! and Amazon web services are a popular example for building
XMLHttpRequest-based web services client in JavaScript [188].
2.5.5
OO-based abstraction
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
61/139
Scheme support Object-Oriented programming, because it always pass message within the body,
just like objects passing. In Scheme, it provides a procedural way to express its OOP features.
Instantiate an object instance is represented as a procedure which project operation to methods while a
method is somewhat like put the parameter into operation then perform operation on instance [189].
And it also support inheritance and polymorphism (by virtual method).In the listing 1 [190], it shows a
class y inherit class x.Some Schemes implementation have their own built-in OO system (i.e.
MacScheme, Feel, Oaklisp, XScheme, and PC-Scheme [191]). Most of these are similar to the
Common Lisp Object System (CLOS), while others are conceptually closer to C++ and Java. Kawa is
fully integrated with the Java object system, i.e. it allows Java classes to be defined and extended in
Scheme [192]. The OOP syntax in different implementation of Scheme may vary depends on how it
interoperates with the OO concept. In addition, the Scheme provides local encapsulation by closures.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
62/139
(define (y)
(let ((super (new-part x))
(self 'nil))
(let ((y-state 2)
)
(define (get-state) y-state)
(define (set-self! object-part)
(set! self object-part)
(send 'set-self! super object-part))
(define (self message)
(cond ((eqv? message 'get-state) get-state)
((eqv? message 'set-self!) set-self!)
(else (method-lookup super message))))
self))) ; end y
Listing 1 Class inheritance: Class y inherit from x
JavaScript support OO-based abstraction (inheritance, polymorphism, encapsulation), although it
is not a pure OO languages. Unlike Scheme, JavaScript have similar way in syntax with traditional OO
language (i.e. java) which would be friendly adapted by java developer. In JavaScript, the
straightforward way for constructing OO is the build-in object data type. In JavaScript, objects are
implemented as a collection of named properties. Being an interpreted language, JavaScript allows for
the creation of any number of properties in an object at any time. Everything is object in JavaScript,
except for the primate data types and it classifies 3 types object (native objects, user-defined objects,
and host objects (for browser purpose)). In JavaScript, each Object can inherit properties from another
object, just like you extend the class for reuse purpose, called its prototype. When evaluating an
expression to retrieve a property, JavaScript will check the properties which defined in the object
firstly. If it is not found, it then looks at the object's prototype to see if the property is defined there.
This continues up the prototype chain until reaching the root prototype [193] . Each object is
associated with a prototype which comes from the constructor function from which it is created. The
polymorphism achieved by calling correct name properties to involved the appropriate function for
corresponding prototype. In the listing 2 [194], it shows a segment of polymorphism include using
prototype and methods calls:
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
63/139
function A()
{
this.x = 1;
}
A.prototype.DoIt = function() // Define Method
{
this.x += 1;
}
function B()
{
this.x = 1;
}
B.prototype.DoIt = function() // Define Method
{
this.x += 2;
}
a = new A;
b = new B;
a.DoIt();
b.DoIt();
document.write(a.x + ', ' + b.x);
Listing 2 Polymorphism in JavaScript
2.5.6
Reflection
Scheme dose not support the reflection directly, But since scheme is dynamic type, it provide
scheme predicates to determine the object type or the equivalence of two objects (using eq?, eqv?, and
equal? as predicates) However, it is somewhat different than we expected, which rule is as follow:
Two objects of different types (booleans, the empty list, pairs, numbers, characters, strings, vectors,
symbols, and procedures) are distinct. Two objects of the same type with different contents or values
are distinct. Thus we need to find a way to fix it. Scheme enables us using marco to reach the function
of reflections. Some research paper [195] implements a marco-implemented extension PLT-Scheme for
object-oriented programming which can interoperate with Java. That’s means we can invoked Java’s
library within the PLT-scheme environment, so that invoked the reflection API is possible.
Unlike the complicated implementation of Scheme in reflection, JavaScript provide a simple way
to do that: do-in statement. Compared to other Scheme, it does not require extra libaries, no
namespace, no particular implemented classes, or any other help from other languages. The syntax is
just stratiforward:
For (var member in obj){ alert(member); };
Listing 3 for-in statement
The Listing 4 [196] shows a example for reflection which return the object’s properties and sort them.
function getMembers(obj) {
var members = new Array();
var i = 0;
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
64/139
for (var member in obj) {
members[i] = member;
i++;
}
return members.sort();
}
Listing 4 reflection in JavaScript
2.5.7
Aspect-orientation
PLT Scheme supports AOP and its macro system provides especially powerful support for
linguistic extensions. Macro system provides a lightweight implementation of aspects.In Scheme, we
can easily define language extensions using its macro system [200]. Scheme macros are effectively
functions that rewrite syntax trees; they are more powerful than lexical macros. Scheme macros can
define and export the new syntactic forms “around” and “fluid-around” which is frequently used in
AOP [199].
The figures below shows an implementation of aspect-scheme for statically -scoped aspects, with base
language PLT Scheme.
(module aspect-scheme mzscheme
;; previous dynamic aspects part elided
;; statically-scoped aspects
(define-syntax (around stx)
(syntax-case stx ()
[( pc adv body)
(syntax
(w-c-m ’static-aspects
(cons (make-aspect pc adv) (current-static-aspects))
body))]))
(define-syntax (lambda/static stx)
(syntax-case stx ()
[( params body . . . )
(syntax
(let ([aspects (current-static-aspects)])
(_ params
(w-c-m ’static-aspects aspects
(begin body . . . )))))]))
JavaScript is not really a place where AOP is used heavily or needed, since the use of callbacks
already provides some level of separation of concerns. AOP may have limited utility in JavaScript.But
with the JavaScript libraries such as jQuery, Javasrcipt can achieve decent separation of concerns
[198]. jQuery also supports the use of custom events, which may be used in a similar manner to some
AOP advice types. Indeed, the Ajax Events available for hooking into using jQuery already provide
some sort of AOP-like functionality [197].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
65/139
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
66/139
2.5.8
Functional programming
Scheme is a functional programming and support closures with environment.. But scheme is not a
pure functional programming which means that there are no variables, no assignment or side effects.
Scheme use static scoping; nonlocal references in a function are resolved at the point of function
definition. Static scoping is implemented by associating a closure [201].
Scheme characteristics:
· Supports functional programming - but not on an exclusive basis
· Functions are first class data objects
· Uses static binding of free names in procedures and functions
· Types are checked and handled at run time - no static type checking
· Parameters are evaluated before being passed - no lazyness
JavaScript uses functional as a library for functional programming. It defines the standard higherorder functions such as map, reduce, and select [202]. It also defines functions such as curry, rcurry,
and partial for partial function application; and compose, guard, and until for function-level
programming. And all these functions accept strings, such as 'x -> x+1', 'x+1', or '+1' as synonyms for
the more verbose function(x) {return x+1}. All functions in Javascript are objects [203].
Javascript has the ability to write anonymous functions, or functions without a name. It can also pass
functions as argument to other functions.Here is an example.
var passFunAndApply = function (fn,x,y,z) { return fn(x,y,z); };
var sum = function(x,y,z) {
return x+y+z;
};
alert( passFunAndApply(sum,3,4,5) ); // 12
JavaScript is not one of the languages that use a variety of techniques to optimize function
calls.Because of that, invoking a function in javascript is slow. Most current JavaScript
implementations are slow with recursion and closures…two cornerstones of functional programming.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
67/139
2.5.9
Declarative programming
Although scheme support funtional programming and the funtional is a subset of declarative language,
the scheme language doesn’t support declarative programming. Declarative language suits when there
is an implicit "behind the scenes" procedure or process that's going to do something uniform with the
assertions or statements presented [204].
Scheme focus on what information is desired and what transformations are required. A functional
language can help organize the expression of computation [205].
JavaScript is an implementation of the ECMAScript language standard and is typically used to enable
programmatic access to computational objects within a host environment. JavaScript supports all the
structured programming syntax in C (e.g., if statements, while loops, switch statements, etc.) [205].
Meanwhile, Structured programming can be seen as a subset or subdiscipline of imperative
programming, one of the major programming paradigms.
However, imperative programming which requires an explicitly provided algorithm is used in
opposition to declarative programming, which expresses what needs to be done, without prescribing
how to do it in terms of sequences of actions to be taken. Therefore, We could show Javascript doesn’t
support declarative programming [205].
2.5.10
Batch Scripting
The Racket project provides useful libraries that enable developers to write scripts and automate tasks
such as file processing (text or XML files) zip files creation using Racket’s gzip library and so on.
Thus, Racket has what is called port which is an Input and Output stream for receiving or redirecting
data to a file or a terminal [206]. This is a helpful technique that allows scripts to receive arguments as
parameters.
Rackets files could have on of these extensions (rkt, .rktl, .rktd, .scrbl, .plt, .ss, .scm) and can run on
Windows or on a Unix-based system as executables. This feature enables Racket to be used for batch
scripting.
Natively, JavaScript it was designed to run inside a web browser. For security reason, JavaScript
doesn’t provide support for Batch scripting. Internet Explorer (IE) browser provides ActiveXObject
which enables to execute an external program from a web browser as shown in the code snippet
below. To do so, the security settings should be changed in order to enable ActiveX support. This
approach is not supported by other web browsers. In addition, JavaScript enables you to automate
some of web browser’s functionalities. For example opening pop-up windows or bookmarking a web
page [205].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
68/139
Running Notepad using JavaScript from Internet Explorer.
<script language="javascript" type="text/javascript">
function runNotepad()
{
var shell = new ActiveXObject("WScript.shell");
shell.run("notepad.exe", 1, true);
}
</script>
2.5.11 UI prototype design
There are many GUI and graphics libraries to support scheme for UI prototype design such as Racket
Graphics Toolkit.
Javascript is designed to add interactivity to HTML pages. So, it can support DHTML to create user
interface and handle user action.
3
Consolidated Analysis and Synthesis of the Results
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
69/139
Default more secure programming practices
Criteria 1
Features/PL
Memory
management
C++
Yes(Manual memory
management)
Bounds checking
Static Type
checking
Dynamic Type
checking
Type safety:
Strong/weak
Yes [232]
No[233]
N/A – not considered
Exception handling
N/A - not considered
Compiled/
Interpreted
N/A - not considered
Conditional
compilation
N/A - not considered
Assertions
N/A - not considered
AspectJ
Yes, provided by the
JVM
Garbage collection,
provided by the GHC
Yes, for arrays and data
structure.
Yes, run time bounds
checking
Yes check on compile
time
PHP
Default using GC(by
PHP 5.3) or Manual(by
extension API)
Yes, run time bounds
checking
No, PHP is dynamic
type checking [213]
No, Haskell does not
support casting
Yes [214] (no type
cascading)
Partial dynamic
checking, i.e down
casting.
Yes, , strongly type
Not that safe(running
time type checking may
not efficient to avoid
type error)
Yes, , strongly type
Yes, but more complex
than Java
Yes, using try-catch and
can throw any kind of
exception
Yes, using try-catch
with build-in Exception
class
Compiled or Interpreted
by the GHC
Interpreted by web
server(i.e. Apache)
Compiled to byteCode,
and than interpreted by
JVM
Yes, but need to set in
GHC options [207]
Yes(but rarely used)
Yes, but need some
tricks[1]
User self-defined
Support(by PHP 4)
Yes, Java support assert
Yes, at compile time.
Partial dynamic
checking, i.e down
casting. Also, when
using reflection.
Yes, Based on Java’s
access rules.
Yes, exceptions can be
thrown in pointcuts.
Handlers are used to
catch exceptions in
AspectJ. Within trycatch/finally block in
Java code.
Compiled, and the Javabyte code interpreted by
a JVM compliant.
Yes (using SCoPE
compiler, conditional
pointcut evaluator)
Yes, temporal assertions
at run-time[24]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Haskell
70/139
Scheme
Garbage collection
Yes, run time bounds
checking
Yes, check on compile
time
Default more secure programming practices
Criteria 1
Features/PL
Memory
management
Bounds checking
Static Type
checking
Dynamic Type
checking
Type safety:
Strong/weak
Exception handling
C#
Yes, provided by the
CLR
Yes, also for array
bounds (raise
IndexOutOfRange
exception), buffer
overflow. It can be
disabled in C# [25]
Yes at compile time.
Yes, since C# 4.0
(.NET System.Dynamic
namespace) and
dynamic keyword
Yes, C# is strongly
typed. The unsafe
keyword can be used for
allowing pointers use.
Yes, Within trycatch/finally block.
Multiple catch blocks is
supported in C#.
Compiled/
Interpreted
Compiled, Interpreted
by the CLR
Conditional
compilation
Yes (using preprocessor
directives)[21]
Yes, Managed code
assertions [22]
Assertions
Groovy
JavaScript
Yes (Garbage collection
Garbage collection
Java
Garbage collection,
provided by the JVM
Scala
Garbage collection by JVM
Yes, run time bounds
checking
Yes, run time bounds
checking
Yes, run time bounds
checking
Yes [233]
No
Yes, check on compile
time
Yes, Scala support static
type checking[215]
Yes [233]
a variable can hold an
object of any type and
cannot be restricted
Partial dynamic
checking, i.e down
casting.
Strong[215] (with type
cascading )
No, weakly type
Yes, , strongly type
Safe, it is strongly type
system
Yes, using try-catch and
can throw any kind of
exception
Yes, using try-catch
with build-in Exception
class
Yes, using try-catch, but
exceptions are not checked
so effectively
N/A - not considered
N/A - not considered
Try-catch block
N/A - not considered
Interpreted
N/A - not considered
Not support
N/A - not considered
User self-defined
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Compiled to byteCode,
and than interpreted by
JVM
Yes, but need some
tricks [208]
Yes, Java support assert
71/139
Complied by JVM
Yes(with the help of
compiler flag) [217]
Support, using
assert, require and assume
Web applications development
Criteria 2
Features/PL
C++
Dynamic Web
Pages
CSP
Web Server
Web
Framework/Librari
es
Session
management
Security
Model-ViewController
Database
interaction
Secure Sockets
Layer support
Using Platinum,
Reason, Evocosm, ACF.
N/A - not considered
N/A - not considered
Wt/CppCMS/ffead
N/A - not considered
N/A - not considered
N/A - not considered
AspectJ
Haskell
Yes, injection Aspectj
code in JSP pages
Haskell Server Pages
(HSP)
Tomcat, JBoss,
GlassFish
Haskell Application
Server (HAppS)
Using J2EE, Spring or
Struts framework.
Yes, i.e using Spring
which provides
SessionManagementFilt
er class and concurrency
control [33].
Yes, Java web
applications run in what
is called container, Each
application has its
container. Thus, the
JVM add
By adding AspectJ’s
aspects to
J2EE/Spring/Struts
applications.
Yes, using the JDBC
library
Yes using OpenSSL
library.
N/A - not considered
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Snap, Turbinado web
framework
Yes, Snap have build-in
session management.
Yes
Turbinado MVC
framework
PHP
Support, the code
behind php web page
Apache, Microsoft IIS
Many (i.e. Zend,
Symfony, etc…)
Yes, using $_SESSION
variable
Not that safe(require
extra session
management to avoid
hacking and hijacking)
Symfony/Mojavi/CakeP
HP[218]
Mysql
Yes (HaskellDB)
Yes, (hOpenSSL)
Yes(use open SSL
function) [219]
Yes(store in user
session)
72/139
Scheme
PLT Scheme can create
web page directly
PLT Scheme web server
SHP framework
PLT Scheme libsm6
package
Yes, PLT provides a
Scheme interface to
some of the OpenSSL
functionality through its
openssl collection
SHP framework support
MVC
Yes (some package in
PLaneT)
Yes, openSSl
Web applications development
Criteria 2
Features/PL
Dynamic Web
Pages
Web Server
Web
Framework/Librari
es
Session
management
Security
Model-ViewController
Database
interaction
Secure Sockets
Layer support
C#
Yes, as code behind in
ASP.NET web pages.
IIS on Windows OS
Mono (Apache, Nginx)
Using .NET Framework
on Windows and Mono
ASP.NET Linux and
Unix-based OSs.
Yes, provided by
ASP.NET State
management.
Yes (ASP.NET Security
Architecture [34])
ASP.NET MVC 1.0/2.0
Yes, using ADO.NET
library
Yes, SSL certificate can
be generated and used to
on IIS server.
Groovy
JavaScript
Java
JSP
Using Server-Side
JavaScript (SSJS) [jh-4]
JSP
IIS, Apache
Tomcat, JBoss,
GlassFish
N/A - not considered
Mainly using Grails
Framework.
JAXER
J2EE framework
N/A - not considered
Yes, using Jaxer.session
Yes, are represented by
an HttpSession object.
N/A - not considered
Yes
Yes
Yes, can create MVC
structure using Jaxer
J2EE/Spring/Struts
N/A - not considered
Yes, Jaxer.DB
Yes (JDBC)
N/A - not considered
Yes
Yes, JavaSSL [209]
Grails
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
73/139
Scala
Support
Tomcat, Glassfish
Not much(mainly use
Lift)
Yes, Session Actors
Safe
Spring/Pinky
JDBC
Yes, can using java ssl
support
Web services design and composition
Criteria 3
Features/PL
Web Services
including(SOAP,
WSDL, UDDI)
Web services
security
C++
Yes. Supported by
gSOAP toolkit.
Yes.Support HTTPS
and WS-Security:
authentication,tokens,
digital signatures
Web Services
composition
N/A - not considered
AspectJ
Yes. i.e using Apache
XML-RPC for Java
library [38] or Spring
Web Services, JAX-WS
2.1)
Yes. HTTPS, above
libraries support “XMLbased standards” such as
WS-Policy, WSSecurity and WSTransfer standards)
Web Services pipeline
watching.
Using AspectJ and J2EE
framework, two or more
web services can
communicate together
by linking them as a
network. An application
written in AspectJ can
ensure the WS pipeline
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Haskell
Yes. i.e using HAIFA
PHP
Yes(XML-RPC for PHP
[220] )
Scheme
Yes, Scheme supports
web services but it
requires CGI processing
utilities [244].
Yes, hopenSSL help
Haskell for HTTPS
encrypt.
Partially support(limited
by soap interaction with
WSDL) [221]
Yes, PLT Scheme
provides an openssl
collection as part of it
[255].
N/A - not considered
Yes(declared by
PHParray or policy
files or inline with
WDSL [222])
74/139
Yes, Scheme web
services can be built as a
web services network
where one or more web
services can
communicate together.
Web services design and composition
Criteria 3
Features/PL
Web Services
including(SOAP,
WSDL, UDDI)
Web services
security
Web Services
composition
C#
Yes.
using ASPT.NET web
services and .NET’s
System.Web.Services
namespaces for
Windows-based WS and
Mono project’s Web
Services on other OSs
(Linux/Unix, Mac)
Yes, HTTPS Encryption
and signing with
SSL[39], WS-Policy,
WS-Security standards
support)
Using ASP.NET web
services. Also a network
of web services is
possible by making two
or more WS talking
together.
Groovy
Yes. (Supported by
Grails)
Yes, rely on simple
keystores or Spring
Security for
authorization and
authentication.(since
version 0.5)
N/A - not considered
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
JavaScript
Yes, it is possible using
an
AJAX/XMLHttpReques
t-based web service
client [sr-13]. Gmail and
Google services are
good examples.
Java
Yes. i.e using Apache
XML-RPC for Java
library [JL-6] or Spring
Web Services, JAX-WS
2.1)
Scala
Yes(Apache XMLRPC[223])
Yes, Ajax can be used
with any dynamic web
programming language
that support HTTPS
protocol (SSL).
Yes. HTTPS, above
libraries support “XMLbased standards” such as
WS-Policy, WSSecurity and WSTransfer standards)
Web Services pipeline
watching.
support
No, JavaScript doesn’t
support web services
composition. We can
only consume web
services using
JavaScript.
Yes(by Apache Axis2)
[224]
J2EE WS composition.
75/139
OO-based abstraction
Criteria 4
Features/PL
Central unit
C++
Classes
Structural elements
N/A - not considered
AspectJ
Aspects, abstract
Aspects
Pointcuts, Advices,
Inter-Type declarations
for declaring aspect’s
members (fields,
methods, and
constructors)
Multiple
inheritance
Yes [235]
No (Aspect Inheritance)
Aspectual
polymorphism [44].
Method overloading,
method overriding,
virtual method. By
N/A - not considered
default AspectJ methods
are virtual (non-static
methods). Using Intertype declaration.
Partial Classes
No, but it provides
Inter-type declarations
N/A - not considered
(aka open classes) as
alternative.
Structs support
No, doesn’t support
N/A - not considered
Structs declaration.
Instantiation
Aspect instantiation: not
directly instantiated.
N/A - not considered
You cannot use the new
keyword to instantiate
an aspect.
Extension/
Yes, “Aspect extension”
Inheritance
N/A - not considered
Aspects can extend
COMP 6411 - A Comparative studies of programming languages
classes and implement
Haskell
Type class
PHP
Classes(by PHP 5)
Scheme
Meta Classes
Methods, fields
Type, Methods,
Member
Meta Object Protocol,
simple-object.
Yes, We have multiple
inheritance if the
instance consults
multiple components
(not including itself) for
behavior.
Yes, Haskell support
multiple type class
inheritance
No(implemented by
interface)
Yes. The class methods
defined by a Haskell
class correspond to
virtual functions. type
class based overloading.
Yes or No, because PHP
is weakly type, and does
not care about variable
type. PHP does not
support same function
name with different
parameter
N/A - not considered
Public/private/protected
N/A - not considered
No, all member function
has to write together
N/A - not considered
Support(by Symphony
struts)
N/A - not considered
Instantiate class object
using keyword “new”
N/A - not considered
Polymorphism
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
No. Haskell does not
support it
No, doesn’t support
Structs declaration.
Yes, use instance
keyword
Yes, Haskell supports a
notion of class extension
76/139
Accessibility
control
Public/private keyword
interfaces, but Aspects
can extend only abstract
Aspects.
Yes, privileged for
Aspects, private, public
and protected for intertype members.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
NO, module system
must be used to hide or
reveal components of a
class.
Yes(wrapping with C++
classes) [225]
77/139
N/A - not considered
OO-based abstraction
Criteria 4
Features/PL
Central unit
Structural elements
Multiple
inheritance
Polymorphism
Partial Classes
Structs support
Instantiation
C#
Classes, interfaces,
abstract classes
Constructors,
Destructors, Methods,
Fields (attributes),
Delegates, Properties,
Indexes, Events,
Finalizes, Operators,
Nested Classes.
No, can be done by
using Interface
implementation as work
around.
Yes, through
inheritance, based on
Base Type and Sub Type
relation. Method
overloading, method
overriding. C# provides
the virtual
Keyword which is
derived from C++.
Yes, separation of code
(class definition split
into two or more source
files.
Yes, derived from C++
with enhanced features.
Yes, Class instantiation
using the new keyword.
Groovy
JavaScript
Java
Classes
Function
Class
Constructors, Methods,
fields, traits, case class
N/A - not considered
Variable, function
Constructors, Methods,
Members.
No, can be done by
using Interface
implementation.
Yes, it support multiple
inheritance
No, but can implement
multiple interface
No(can use traits to
implement)
N/A - not considered
Yes, basic feature for
OOP. Method
overloading. By default
methods are virtual.
Yes, Scala support type
polymorphism. Yes,
Scala support function
overloading
N/A - not considered
N/A - not considered
N/A - not considered
N/A - not considered
N/A - not considered
Yes, can declare
interface in a file and
implement it in another
file.
No, doesn’t support
Structs declaration.
N/A - not considered
N/A - not considered
Yes, Class instantiation
using the new keyword.
Extension/
Inheritance
Yes, “class extension”.
A class may extend
another class and may
N/A - not considered
implement one or more
COMP 6411 - A Comparative studies of programming languages
interface. Extension
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Scala
Classes
Yes, a class can extend
a abstract class.
N/A - not considered
78/139
Public/package/private/
protected
Yes(by Scala 2.73
provide partial
functions)
Support(require external
build tool for Scala)
[226]
Instantiate object using
keyword “new”
Accessibility
control
methods feature is
supported since C# 3.0
but it requires a static
class and static method
to so [43].
Yes, C# provides
private, public,
protected, internal,
sealed and protected
internal keywords.
Public/private keyword
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
N/A - not considered
Yes, private, public and
protected for inter-type
members.
79/139
Can extend abstract
class
Reflection
Criteria 5
Features/PL
Access to
program’s
metadata
C++
Yes. [234]
Generation code
at run-time
Yes. [236]
Dynamic
invocation
N/A - not considered
AspectJ
Yes, since AspectJ
5.0. Annotation-based
development style.
Haskell
Yes, Monadic
reflection
AspectJ doesn’t
provide such feature.
Can be done using the
following librairies.
ASM
http://asm.ow2.org
and BCEL
http://jakarta.apache.
org/bcel
Using thisJoinPoint to
access the current join
point for the advice
through reflection.
No, compile-time
reflection
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
N/A - not considered
PHP
Yes(by using
annotations and
reflection API
supported by PHP
5.3)
Yes(after PHP 5.3, it
provide runtime
configuration )
N/A - not considered
80/139
Scheme
Yes, access through
library.
Yes[
N/A - not considered
Reflection
Criteria 5
Features/PL
Access to
program’s
metadata
Generation code
at run-time
Dynamic
invocation
C#
Yes, accessing
Attributes through
reflection using
System. Reflection
namespace.
Yes, using .NET’s
System.Reflection.Emi
t namespace.
Yes, method
invocation at run-time
through reflection
[48].
Groovy
JavaScript
Yes, can access
directly
Java
Yes, since Java 5.0.
Annotation-based
development style.
Yes, compile-time
reflection
Yes, such as JUnit.
Yes [237]
N/A - not considered
N/A - not considered
N/A - not considered
Yes [236].
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Scala
Yes(by reflection
API)
No, compile
time
81/139
N/A - not considered
Aspect orientation
Criteria 6
Features/PL
Aspect-oriented
programming
C++
AspectJ
Natively, it is an
aspect-oriented
programming
language
Require extra
compiler supported
Modularity
isolate crosscutting
but need extra
compiler support.
Isolate crosscutting
concerns in a modular
way.
Haskell
AOP Haskell is an
AOP extension for
Haskell
Yes, same as AspectJ.
Code reusability
N/A - not considered
Security
Secure(base on java
platform)
Errors
handling/Logging
N/A - not considered
Can produce pure
uncoupled code
Securing UI by
detecting SingleThread UI’s rule and
web services pipelines
failure.
N/A - not considered
Cleaner way to handle
and log exceptions
N/A - not considered
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
PHP
Require extra library
supported
isolate crosscutting
but need extra aop
library support.
invoked external
library to achieve
Scheme
Yes, as the family of
languages
with
AOP
features includes not only
academic languages such as
Scheme and ML but also
industrially
popular
languages such as Python
and Perl, defining aspects in
this context takes on
immediacy and importance
Yes, object based message
passing, Many Schemes
provide
their
own
modularization facilities.
N/A - not considered
Require extra library
support
N/A - not considered
N/A - not considered
Provide exception
API by PHP 5.0 or
wrapper functions
82/139
N/A - not considered
Aspect orientation
Criteria 6
Features/PL
Aspect-oriented
programming
Modularity
C#
Partially, there are
some no-mainstream
projects implementing
AOP for C#.
Difficult some times
impossible to isolate
crosscutting concerns
from the program’s
business logic.
Groovy
Natively, it is an
aspect-oriented
programming
language
Same as AspectJ,
provide crosscutting
for handling on aspect
JavaScript
Java
Yes, AspectJS is an AspectJ is extension
open source and free of Java for AOP
framework for AOP in
Javascript
Yes, we use the main
module as the central
point through which
Same as AspectJ
data is delivered to
and
from
other
modules
Scala
Natively
AspectJ)
Support(by
Same as AspectJ,
provide crosscutting
for handling on aspect
Code reusability
Security
Errors
handling/Logging
Can’t isolate code that
crosscut over
modules. Partially,
N/A - not considered
can be done using
DLLs in order to
reduce code coupling
degree.
Not applicable since
Require extra
there is no mainstream
complier support
AO solution for C#.
Exceptions are
handled in try-catch
N/A - not considered
blocks and logged
within this scope.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
N/A - not considered
Same as AspectJ
Less coupled code
N/A - not considered
Same as AspectJ
Secured by its access
control
N/A - not considered
Same as AspectJ
N/A - not considered
83/139
Functional programming
Criteria 7
Features/PL
Type inference
C++
AspectJ
No, planned for Java 7
N/A - not considered
Lambda
expression
Anonymous
methods
Higher-order
functions
Yes [238] [239]
Yes [239]
Closure
N/A - not considered
First-class
functions(delegat
es)
Recursion
N/A - not considered
Yes
Using interfaces as a
work around.
No, use inner classes
as work around
Using interfaces as
work around
No, not popular
enough an open
source project for
OpenJDK [58])
No, AspectJ doesn’t
support delegates,
lazy functional
programming can be
used as work around
[53].
Not secure as in C#
[57] due to infinite
methods loop.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Haskell
Yes, Haskell is not
necessary to write for
each variable
Yes
Yes, using a concise
syntax
Yes, Haskell can pass
function as parameter
and return function
Yes,
PHP
Scheme
Not support
Yes(since PHP5.3)
Yes(since PHP5.3)
[227]
Yes, PHP can assign
function to variable
which can be pass as
parameters
Yes(since PHP5.3)
Yes [282]
Yes [281] [283]
N/A - not considered
Yes[JL - 3]
Yes, function is first
class object
N/A - not considered
Yes, Haskell using
recursion to support
loop
Yes
84/139
Yes [284]
Functional programming
Criteria 7
Features/PL
Type inference
Lambda
expression
Anonymous
methods
Higher-order
functions
Closure
First-class
functions(delegat
es)
Recursion
C#
Groovy
Yes, functional
programming’s
N/A - not considered
features have been
introduced in C# 3.0
Yes, since C# 3.0
using LINQ library,
No [239] [240]
System.Linq.Expressio
n namespace
Yes, since C# 2.0
where this feature has
N/A - not considered
been introduced.
Yes using C#
delegates.
JavaScript
N/A - not considered
Yes [281]
N/A - not considered
Yes [240]
Yes [285]
N/A - not considered
N/A - not considered
N/A - not considered
N/A - not considered
Yes
Yes [286]
Yes, since C# 3.0
Yes, delegates are
C++’s function
pointers equivxalent.
Yes it is fully
supported. Can be
done using C#’s
methods.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Java
Functional Java – a
extension of Java, but
each variable has to
assign a type
No, Java does not
support it
No, Java has
anonymous classes,
but do not support
anonymous functions
Java cannot pass
function and return
function
No, but Java simulate
some features of
closures [211]
No, function is not
first class object for
Java
Scala
Yes
Yes, defined in a very
succinct fashion [227]
Yes, Scala provides a
relatively lightweight
syntax for defining
anonymous functions
[277]
Yes, can pass function
and return function
[227]
Yes
Yes [278]
Yes, Java function can
N/A - not considered
call itself
85/139
Declarative programming
Criteria 8
Features/PL
Tags on
methods/fields/
classes
LINQ equivalent?
Dynamic
programming
(Dynamic type
checking)
Declarative
programming
based on XML
[41]
C++
AspectJ
Haskell
PHP
N/A
AspectJ/Java
Annotations
Annotations
annotations
N/A
Additional libraries,
lamdaJ [63]
Yes, basic feature
Yes (require extra
implementation
support).
N/A
No, AspectJ doesn’t
provide such features.
No, Haskell is static
type checking
Yes
Yes, Scheme is
dynamic typing
N/A
YAML, SwiXML
Yes, HAXML
Symphony YAML
XML Schema
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
86/139
Scheme
LINQ for R6RS
Scheme[287]
Declarative programming
Criteria 8
Features/PL
Tags on
methods/fields/
classes
C#
Using C# Attributes
Yes, system.data.linq
LINQ equivalent? library should be used
and referenced.
Dynamic
Yes – since C# 3.0
programming
(var keyword) and C#
(Dynamic type
4.0 (dynamic
checking)
keyword)
Declarative
Using XAML
programming
technology which is
based on XML
available since .NET
[41]
3.0 and WPF
Groovy
N/A - not considered
N/A - not considered
N/A - not considered
JavaScript
Java
Scala
N/A - not considered
Annotations
annotations
N/A - not considered
Additional libraries,
lamdaJ
No
N/A - not considered
YAML, SwiXML
N/A - not considered
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
N/A - not considered
87/139
Yes
No(static type
checking)
Bad support on XML
[228]
Batch scripting
Criteria 9
Features/PL
Run external
programs/external
commands
C++
Yes [241]
Tasks automation
Yes(by COM objects)
Accept command
line arguments
Yes [242]
Need to be
recompiled after
changing the code
source
Not compared
AspectJ
Yes, using Java’s
Process and Runtime
classes.
Haskell
Yes, using system or
rawsystem function
Yes, can be executed
as stand alone and
executed on a
scheduled basis.
Yes
Yes as parameters for
the main method.
i.e public static void
main(string[] args)
Yes
Yes, code should be
recompiled in order to
source code changes
take effect.
Yes
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
PHP
Yes, such as exec,
system
Yes(use cur
command)
Yes(by PHP CLI
script)
Yes [230]
88/139
Scheme
Yes, using
scheme/system [279]
library which allows
running external
commands and
programs.
Yes, Racket’s files
can be executed as
standalone programs
on Windows or on
any Unix-based OSs.
Yes, on Unix-based
terminal or MS DOS
using scheme/cmdline
from Racket libraries
[279].
Yes, need to
recompile Racket
scripts.
Batch scripting
Criteria 9
Features/PL
C#
Run external
Yes, using Process
programs/external class in
commands
System.Diagnostics
namespace.
Tasks automation Yes, can be executed
as stand alone and
executed on a
scheduled basis.
Accept command Yes as parameters for
line arguments
the Main method. i.e
public static void
Main(string[] args)
Has to recompile Yes, code should be
after changing the recompiled in order to
code source
source code changes
take effect.
Groovy
Yes [243]
Yes(by Groovy
Monkey)
Yes(using groovy
script and listen
mode)
N/A - not considered
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
JavaScript
Only with Internet
Explorer using
ActiveXObject.
Java
Yes, using Runtime
class
Yes,
Yes, nearly in every
web browsers with
support for JavaScript.
No, JavaScript run
with a browser and
not from the
command line.
No, no need to
recompile JavaScript
after changing the
source code
Scala
Yes
Yes
Yes
Yes
Yes
89/139
Yes
UI prototype design
Criteria 10
Features/PL
Graphical user
interface
C++
Yes, using Qt, a C++ class
library.
Built-in?
N/A - not considered
Look and feel
N/A - not considered
2D and 3D support
Yes using OpenGL
AspectJ
Haskell
Yes, using Java’s built-in
Swing and AWT libraries.
WxHaskell, Gtk2Hs,
HOC, qtHaskell and so on
Yes, part of Java
Foundation Classes
Yes such as GTK+, Motif,
Windows, Macintosh, etc
themes.
No, all are extension of
Haskell
N/A - not considered
Yes, using Java.awt.dnd
package.
Components’ Layout
Management
N/A - not considered
Yes, i.e Border/Grid
layout, etc.
Multiple document
interface (MDI)
Model-ViewController
Yes using QT
Support but not enforce.
Declarative GUI
development [41]
N/A - not considered
Performance
N/A - not considered
IDE/UI designer
N/A - not considered
Rich Web UI
Development
Deployment
N/A - not considered
N/A - not considered
Single thread of
execution
N/A - not considered
Event handling
mechanism
N/A - not considered
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
Yes, provided by AWT
NetBeans, Plugins for
Eclipse IDE
Yes, using JavaFX
framework.
Jar archive (.jar files)
Yes, Swing library’s
single-thread rule, threadsafe[73]
Only Event Listeners,
provided by java.awt.event
package
or
Yes, Racket Graphics
Toolkit, PLT MrEd
Graphical Toolbox
Yes(using PHP-GTK+ by
PHP 4.0)
No,
Yes
Yes, FRAN [212]
Yes, Gtk2Hs support D&D
Yes, using layout
container
Yes, wxHaskell has basic
MDI support
Yes in components such as
JTable and JList, etc.
Glade XML, SwiXML:
not popular yet as much as
XAML
Relatively Slow
Yes(PHP-GTK
embelled with HTML)
Scheme
More complex than Java
Yes – Java 2D API
Drag and Drop
support
PHP
Yes, PHP-GTK 2
N/A - not considered
Yes
Yes, using frame%
Yes, Racket GUI
Application Framework
Yes
Yes, FranTk
Yes – sgl library to support
3D
Yes, pasteboard in MrEd
[1]
Yes, specifies the layout of
a window by assigning
each
GUI element to a parent
containerrs [2]
GTK+
Working well on both
Linux and windows
Relatively Slow
Glade
Zend, Dreamware
Using XAML
Provide good memory
performance
WinScheme Editor
HsWTK
Yes
.hs file
Yes, Event Listeners
Yes.
Yes, PHP does not support
multiple thread
N/A - not considered
Yes, Event Handlers
N/A
N/A
90/139
UI prototype design
Criteria 10
Features/PL
Graphical user
interface
Built-in?
Look and feel
2D and 3D support
Drag and Drop
support
Components’ Layout
Management
Multiple document
interface (MDI)
Model-ViewController
Declarative GUI
development [41]
Performance
IDE/UI designer
Rich Web UI
Development
Deployment
Single thread of
execution
Event handling
mechanism
C#
Groovy
JavaScript
Java
Yes, using .NET’s Windows Forms
and WPF libraries.
Built on top of the Base Class Library
You need to import/reference
System.Windows.Forms.dll for
WinForms and System.Windows.dll
WPF. WPF only on Windows
platform [70]
Yes, using Java’s built-in
Swing and AWT libraries.
Support DHTML to
implement GUI
Yes, using Java’s built-in Swing
and AWT libraries.
Yes(using
ScalaGUI)
Javascript has not built in GUI
library.
Yes, part of Java Foundation
Classes
Yes(using ScalaGUI, not much
library as Java GUI)
Very flexible, and easy to
implement powerful UI on
webpage,
Javascript can support Html to
show 2D graphic
Yes such as GTK+, Motif,
Windows, Macintosh, etc
themes.
Yes – Java 2D API and Java 3D
API
N/A - not considered
No, does not support
Yes, java.awt.dnd
N/A - not considered
N/A - not considered
Yes, base on DHTML and Css
Yes, Several AWT and Swing
classes provide layout managers
N/A - not considered
Yes, using Java Swing library.
No
Yes, using JDesktopPane
N/A - not considered
Yes,enforce by Swing
No
Yes in components such as
JTable and JList, etc.
N/A - not considered
N/A - not considered
No,
using XAML
Javascript is a slow language
Relatively Slow
NetBeans, Plugins for Eclipse
IDE
N/A - not considered
Not flexible enough in WinForms.
Full customizable UI support in WPF
N/A - not considered
Yes, using WPF framework
N/A - not considered
Partially, only using WPF framework
[71]
Yes, this can be done automatically
using Visual Studio’s UI designer.
Yes, using,
System.Windows.Forms.MdiLayout
class.
WinForms no, WPF Model-ViewViewModel [72]
Using XAML technology
Improved execution on windows Fast
Visual Studio C# and SharpDevelop
for Windows, MonoDevelop .
Silverlight framework on windows
and Moonlight implementation for
Linux and other Unix-based OSs.
.NET Assemblies: (executables and
libraries, .exe and DLLs)
Windows Forms (Not by default, but
by applying the STAThread
attribute to the Main method).
WPF use by default a single thread of
execution.
Yes, Event Handlers and Delegates
support
N/A - not considered
No
N/A - not considered
Yes, Javascript can only run in
single thread
N/A - not considered
Yes, using JavaFX framework.
Scala
java
GUI
N/A - not considered
N/A - not considered
XAML
Slow as java graphic
Scala for eclipse, Net beans
Silverlight
framework
on
windows
and
Moonligh
implementation for Linux and
other Unix-based OSs. [
Jar archive (.jar files)
N/A - not considered
N/A - not considered
Yes, Javascript can only run in
single thread
Yes, Swing library’s singlethread rule, thread-safe[46]
N/A - not considered
N/A - not considered
Yes, can handle event coming
from webpage
Only Event Listeners
N/A - not considered
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
91/139
o
3.1
Criteria 1: Default more secure programming practices
The comparison tables show that the compared languages have different type system. Some of
them are strongly typed and others are weakly typed. PHP, Groovy and JavaScript are dynamic
typed which is not secure enough because type checking is performed at run-time. In C# 4.0, it
is possible to use dynamic typing using the dynamic keyword. C#, Java, Haskell and Scalar
are said strongly typed.
Among these languages, only C++ doesn’t provide automatic memory management
mechanism such as garbage collection and memory layout, etc which is not safe. In C++
dangling pointers, memory leak and double pointers free are well known issues where
programmers are called to pay more attention to prevent them.
All languages provide an exception handling mechanism for detecting errors and prevent
applications from crashing. C# and Java have the best exception handling implementation by
providing a set of customized exception classes.
Haskell support exception handling, but it is a complex task to achieve.
3.2
Criteria 2: Web applications development
We have found that all 10 languages support web applications development. However, we
have found that ASP.NET/C#, PHP, Java/J2EE are particularly suited for easy web design. It is
difficult to say which one is better than others. These three technologies provide powerful
feature and have been adopted in large web applications development project. Thus, during the
last decade they also gained popularity.
PHP is an easiest language to learn, but it seems that it is not safe enough and doesn’t provide a
strong type safety since it is dynamically typed.
We have found out also that the remaining languages can also be used in web applications as
CGI scripts, but are quite inconvenient and error-prone to program in them, maintain, and
deploy. Hence, they require external libraries which are hard to configure, maintain and
deploy.
3.3
Criteria 3: Web services design and composition
PHP, Java, C#, Groovy and Scala enable developers to write web services. As shown in the
comparison tables, we have found that ASP.NET/C#, PHP, Java/J2EE are particularly suited
for developing WS. It is seems that Java/J2EE is better than others languages. Java is crossplatform and can be coupled with AspectJ to implement a secure, stable composed web
services.
ASP.NET/C# is an easiest language to learn. Visual Studio is a powerful IDE which provides
helpful features for building and debugging web services applications.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
92/139
Also, we have found that is possible to build such applications using the remaining languages.
But it is painful since they are natively not designed/suited for kind of applications and some
external libraries are required to do so.
3.4
Criteria 4: OO-based abstraction
According to our comparison table, Java, C#, Groovy, Scala are pure object-oriented
programming languages. C++, PHP 5.0 also support OO development but they are not pure
OOP and don’t provide all OO features found in Java and C#.
C# and Java are particularly more suited for OO development. It is difficult to say which one
is better than others. But C# has introduced more OO features such as Properties, Indexers,
delegates, etc.
Java and C# have a similar syntax and they are easy to learn, and both are strongly typed.
Other partially support OOP, but in some of them OOP has been introduced lately or as
extension libraries. However, they are not suited to be adopted in a large OO development
project.
3.5
Criteria 5: Reflection
Our comparison tables let us conclude that all compared languages support reflection and
provide different mechanism to implement it. Some of them don’t require external libraries
especially C# which provides the System.Reflection.Emit namespace that allow to generate
code at run-time. AspectJ and Java indirectly provide this feature using external libraries such
as ASM and BCEL.
For others languages, reflection is supported but not mainly used in development.
3.6
Criteria 6: Aspect-orientation
Based on our research, AspectJ was the first aspect-oriented language. Since it extends Java, it gained a
wide popularity and has become mainstream language for AOP. Other languages may support AO if
there are extensions available. For example, Groovy, Haskell and Scala have already mainstream
extension for AOP. Unfortunately, there are no mainstream AOP extension for C#.
To learn AspectJ, programmers have to be familiar with Java language since it inherits syntax and
features from Java.
C++ is not suited to AOP. Dealing with pointers in AO it is not safe enough and can result in memory
issues.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
93/139
3.7
Criteria 7: Functional programming
Obviously, Haskell and Scheme are natively functional programming languages. They provide all
functional programming features, especially Haskell which is pure functional language. Basically Java,
Groovy, AspectJ and Scala don’t support functional programming. Type inference is planned to be
released with Java 7.
C# 3.0 has been introduced with the release of LINQ which provides functional programming
capabilities such as lambda expression, anonymous methods, and type inference and high-order
functions.
3.8
Criteria 8: Declarative programming
Our tables prove that Haskell is the best language to support declarative programming. Haskell
is pure functional programming language with minimized side effect.
Other languages provide what is called annotation mechanism which is considered as
declarative approach since programmers provide information used by the program at run-time.
A new declarative methodology has been introduced as extension to OOP based on XML
technology. XAML was the first one introduced by Microsoft which enables developers to
describe structured values and objects before coding.
C# is the best language which is suited for mixing OOP and declarative style using LINQ
services especially for querying database and XML files.
3.9
Criteria 9: Batch scripting
According to our comparison studies, batch scripting is possible in all languages except
JavaScript which is possible using ActiveXObject and Internet Explorer browser.
PHP, C#, Scala and Java are the best. A program written in one of them can be automated
(scheduled using Scheduled tasks in Windows and Crontab in a Unix-based system). PHP and
C# are used to automate report generation such as sales report, job alerts, etc.
PHP is more suited one since is not compiled, so if some changes has occurred on the source
code, there is no need to recompile an already deployed script.
3.10
Criteria 10: UI prototype design
After comparing more than ten features, we have found that is possible to create UI in any one
of these 10 languages. C#, Java, Scala, Groovy and AspcetJ are particularly suited for GUI
design.
Java, Scala, Groovy and AspectJ GUIs are based on Java AWT/Swing libraries.
There are a lot of extension libraries to support GUI design in Haskell but these libraries are
not standardized and it is hard to make them working together.
We conclude that PHP and JavaScript are more suitable for web pages UI design since they
can easily mixed/embedded in HTML code.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
94/139
C# is the most suitable for designing and developing GUI. Visual Studio IDE provides a UI
designer with a big set of pre-built component. Thus, C# can use WPF technology to create
powerful highly customizable UIs.
4 Conclusion
We have performed a thorough and detailed comparison of our selected programming
languages to study within the specified criteria. We have found that, by their native design,
each language is suited for a specific field. Some of them can be fully applied to the compared
criteria without extensions.
C# and Java have demonstrated that they are the more suitable for both web and desktop
applications development. They are strongly typed which enable to write more secure
programs. On the other hands, AspectJ as an extension of Java language is the best mainstream
aspect-oriented language. PHP and JavaScript are particularly suitable for web development.
Haskell and Scheme are better for functional programming. C++ is more suitable for system
development and desktop applications. Scala and Groovy have extended Java.
4.1
Future work
We'd like to refine our analysis of our languages within the stated criteria further as we become
more familiar with them over time. We also plan on expanding our analysis onto other criteria
and languages and provide more programming snippets as proof-of-concept illustrations.
4.2
Acknowledgments
We would like to acknowledge the following people and entities who made this work possible:
- Faculty of Engineering and Computer Science, Concordia University, Montreal, Canada.
- Concordia University Libraries for access to the invaluable digital libraries of ACM, IEEE,
Springer and others to do our research.
- Wikipedia contributors with the wealth of information.
- Our poor families, wives, husbands, children, parents, and pets to help us to get through
the suering and sleepless nights and oer all their help and understanding while we were
away from them while doing this project.
Our POD, Yi Ji, for the introductions into AspectJ and Java reflection.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
95/139
Acronym and abbreviation
IDE: Integrated Development Environment
CLR: Common Language Runtime
JVM: Java Virtual Machine
DOS: Disk Operating System
GUI: Graphical User Interface
WPF: Windows Presentation Foundation
API: Application Programming Interface
WCF: Windows Communication Foundation
RAD: Rapid Application Development
API: Application Programming Interface
JDBC: Java Database Connectivity
SOA: Service Oriented Architecture
J2EE: Java 2 Enterprise Edition
TUI: Text User Interface
CLA: Command Line Application
MDI:
JFC:
LINQ:
UI:
BCL:
Multiple Document Interface
Java Foundation Classes
Language Integrated Query
User Interface
Basic Class Library
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
96/139
References
[1] The C++ Resources Network, http://www.cplusplus.com/
[2] AspectJ in Action, Practical Aspect-Oriented Programming, Ramnivas Laddad.
ISBN: 1930110936
[3] Programming paradigm, Wikipedia, http://en.wikipedia.org/wiki/Programming_paradigm
[4] AspectJ projecrt, Eclipse foundation, http://www.eclipse.org/aspectj/
[5] Haskell project, http://www.haskell.org/
[6] Scheme implementation choices, http://web.mit.edu/~axch/www/scheme/choices.html
[7] Gambit project, http://dynamo.iro.umontreal.ca/~gambit/wiki/index.php/Main_Page
[8] PLT Scheme project, http://www.plt-scheme.org/
[9] The Scheme programming language, http://www.scheme.com/tspl3/
[10] Groovy (programming language), Wikipedia,
http://en.wikipedia.org/wiki/Groovy_(programming_language)
[11] Java programming language,
http://en.wikipedia.org/wiki/Java_%28programming_language%29#Practices
[12] [17] Introducing the Scala, “The Scala Programming languages”.
Retrieve from: http://www.scala-lang.org/node/25
[13] Visual C# Developer Center, http://msdn.microsoft.com/en-ca/vcsharp/default.aspx
[14] C# programming language , wikiepdia,
http://en.wikipedia.org/wiki/C_Sharp_%28programming_language%29
[15] ECMA standard, http://www.ecma-international.org/default.htm
[16] Unsafe Code Tutorial,
http://msdn.microsoft.com/en-us/library/aa288474%28VS.71%29.aspx
[17] Garbage collection (computer science), Wikipedia,
http://en.wikipedia.org/wiki/Garbage_collection_%28computer_science%29
[18] Data Types (C# Programming Guide),
http://msdn.microsoft.com/en-us/library/ms173104%28VS.80%29.aspx
[19] Bounds checking, Wikipedia, http://en.wikipedia.org/wiki/Bounds_checking
[20] COMP 6411 lecture notes, Joey Paquet and Serguei A. Mokhov.
[21]Comaparison of Java and C#,
http://en.wikipedia.org/wiki/Comparison_of_Java_and_C_Sharp#Conditional_compilation
[22]Assertions in Managed Code,
http://msdn.microsoft.com/en-us/library/ttcc4x86%28v=VS.71%29.aspx
[23]Static Conditional Pointcut Evaluator for AspectJ,
http://www.graco.c.u-tokyo.ac.jp/ppp/index.php?Projects%2Fscope
[24] J-LO, the Java Logical Observer, A tool for runtime-checking temporal assertions,
http://www.sable.mcgill.ca/~ebodde/rv//JLO/
[25] MSDN’s blog, Array Bounds Check Elimination in the CLR,
http://blogs.msdn.com/b/clrcodegeneration/archive/2009/08/13/array-bounds-check-eliminationin-the-clr.aspx
[26] Project Hosting for Open Source Software, http://www.codeplex.com/
[27] Asp.net CRM, http://crm.codeplex.com/
[28] nopCommerce. Open Source online shop e-commerce solution,
http://nopcommerce.codeplex.com/
[29] Lesiecki, N. "Applyinq AspectJ to J2EE application development," Software, IEEE , vol.23, no.1,
pp.24-32, Jan.-Feb. 2006 doi: 10.1109/MS.2006.1
[30] SpringSourse Tool Suite, http://www.springsource.org/
[31] Frequently Asked Questions, Eclipse network resource
http://www.eclipse.org/aspectj/doc/released/faq.php#q:aspectjandj2ee
[53-] [32] J2EE vs. Microsoft.NET,
http://media.techtarget.com/tss/static/articles/pdf/J2EE-vs-DotNET.pdf
[33] SpringSource, Spring project,
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
97/139
http://static.springsource.org/spring-security/site/docs/3.0.x/reference/session-mgmt.html
[34]ASP.NET Security Architecture, http://msdn.microsoft.com/en-us/library/yedba920.aspx
[35] Tanenbaum, A. S. and van Steen, M. “Distributed Systems: Principles and Paradigms”
[36] Sasa Subotic and Judith Bishop :"Emergent behaviour of aspects in high performance and
distributed computing" ,Year of Publication: 2005,ISBN:1-59593-258-5
[37] Conceptual Overview, http://msdn.microsoft.com/en-us/library/ms731190.aspx
[38] Apache XML-RPC project, http://ws.apache.org/xmlrpc/
[39] HTTP Security and ASP.NET Web Services,
http://msdn.microsoft.com/en-us/library/ms996415.aspx
[41] Object-Oriented Programming (C# and Visual Basic), http://msdn.microsoft.com/enus/library/dd460654.aspx
[42] System.Web.Services Namespace, http://msdn.microsoft.com/en-us/library/9xe4bs0s.aspx
[43] Extension Methods (C# Programming Guide),
http://msdn.microsoft.com/en-us/library/bb383977.aspx
[44] Erik Ernst and David H. Lorenz: "Aspects and polymorphism in AspectJ", Year of Publication:
2003, ISBN:1-58113-660-9
[45] New Reflection Interfaces,
http://www.eclipse.org/aspectj/doc/next/adk15notebook/reflection.html
[46] AspectJ, Eclipse network resource, http://www.eclipse.org/aspectj/doc/released/faq.php
[47] Reflection Overview,
http://msdn.microsoft.com/en-us/library/f7ykdhsy%28v=VS.80%29.aspx
[48] MSDN, Reflection the C# programming guide,
http://msdn.microsoft.com/en-us/library/ms173183%28VS.80%29.aspx
[49] ASpectJ publications,
http://dev.eclipse.org/viewcvs/indextech.cgi/aspectj-home/publications.html
[50] Tigris.org Open Source Software Engineering Tools, http://aspectdng.tigris.org/
[51] Operating system + middleware, http://www.dcl.hpi.uni-potsdam.de/research/loom/
[52] "Delegates and functional programming in C#", David R. Naugler Southeast Missouri State
University, Cape Girardeau, MO, 2004
[53] "Lazy functional programming in Java ", Anthony H. Dekker Defence Science and Technology
Organisation, Canberra ACT SSN:0362-1340
[54] "Functional programming in Java", David R. Naugler, Southeast Missouri State University Cape
Girardeau MO, ISSN:1937-4771
[55] Functional Programming for Everyday .NET Development,
http://msdn.microsoft.com/en-us/magazine/ee309512.aspx
[56] C# Recursion, http://www.meshplex.org/wiki/C_Sharp/Recursion
[57] Avoiding Infinite Recursion with Stratified Aspects,
http://www.sable.mcgill.ca/~ebodde/meta/
[58] Project Lambda, http://openjdk.java.net/projects/lambda/
[59] An Annotation Based Development Style,
http://www.eclipse.org/aspectj/doc/next/adk15notebook/ataspectj.html
[60] Introduction to Attributes, http://msdn.microsoft.com/en-us/library/Aa288059
[61] XAML Overview (WPF), http://msdn.microsoft.com/en-us/library/ms752059.aspx
[62] YAML, Wikipedia, http://en.wikipedia.org/wiki/YAML
[63] Lambdaj, http://code.google.com/p/lambdaj/
[41] Louridas, P.; , "Declarative GUI Programming in Microsoft Windows," Software, IEEE , vol.24,
no.4, pp.16-19, July-Aug. 2007 doi: 10.1109/MS.2007.105
[38] [64] Basic Console Application (C#), http://msdn.microsoft.com/en-us/library/bb251798.aspx
[65] Capitalization Styles, http://msdn.microsoft.com/en-us/library/x2dbyw72%28v=VS.71%29.aspx
[66] Text-based (computing), Wikipedia, http://en.wikipedia.org/wiki/Text-based_%28computing%29
[67] Java SE Desktop Overview, http://java.sun.com/javase/technologies/desktop/
[68] Windows Forms, Wikipedia, http://en.wikipedia.org/wiki/Windows_Forms
[69] Windows Presentation Foundation on the Web: Web Browser Applications,
http://msdn.microsoft.com/en-us/library/aa480223.aspx#wpfandwbas_topic1
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
98/139
[70] Mono project, WPF, http://www.mono-project.com/WPF
[71] MSDN, .Net framework 4, Drag and Drop Overview,
http://msdn.microsoft.com/en-us/library/ms742859.aspx
[72] WPF Apps With The Model-View-ViewModel Design Pattern, http://msdn.microsoft.com/enus/magazine/dd419663.aspx
[73] Threads and Swing, http://java.sun.com/products/jfc/tsc/articles/threads/threads1.html
[74] Comparison of programming languages, Wikipedia,
http://en.wikipedia.org/wiki/Comparison_of_programming_languages
[75] Web Applications, http://msdn.microsoft.com/en-us/library/ms235434(VS.80).aspx
[76] WT, http://www.webtoolkit.eu/wt
[77] C++ Applications, http://www2.research.att.com/~bs/applications.html
[78] Web Development in Groovy using Groovlets, http://www.javabeat.net/articles/58-webdevelopment-in-groovy-using-groovlets-1.html
[79] Groovy and Grails – A Getting Started Guide, http://www.indicthreads.com/1481/groovy-andgrails-a-getting-started-guide/
[80] Groovying XML, http://www.techbookreport.com/tutorials/groovy_xml_01.html
[81] GroovyWS,The Groovy Resources Network, http://groovy.codehaus.org/GroovyWS
[82] Groovy SOAP, The Groovy Resources Network, http://groovy.codehaus.org/Groovy+SOAP
[83] Practically Groovy: Building, parsing, and slurping XML,
http://www.ibm.com/developerworks/java/library/j-pg05199/index.html
[84] Savage, W. J. (W. John). Groovy programming : an introduction for Java developers. Amsterdam
; Boston : Morgan Kaufmann Publishers.
[85] WS-Eventing for WCF , http://www.codeproject.com/KB/WCF/WSEventing.aspx
[86] The gSOAP Toolkit for SOAP Web Services and XML-Based Applications ,
http://gsoap2.sourceforge.net/
[87] Walkthrough: Creating an XML Web Service Using C++ and the CLR,
http://msdn.microsoft.com/en-us/library/a86z84tw(VS.80).aspx
[88] JN3025-Inheritance, The Groovy Resources Network, http://groovy.codehaus.org/JN3025Inheritance
[89] Abstract Classes (C++), http://msdn.microsoft.com/en-us/library/c8whxhf1.aspx
[90] Abstract classes(IBM) ,
http://publib.boulder.ibm.com/infocenter/comphelp/v8v101/index.jsp?topic=/com.ibm.xlcpp8a.doc/lang
uage/ref/cplr142.htm
[91] Liang, Y. Daniel. Introduction to programming with C++. Upper Saddle River, NJ ; Montreal
: Prentice Hall.
[92] Q&A for professional and enthusiast programmers ,
http://stackoverflow.com/questions/359237/why-does-c-not-have-reflection
[93] JN3535-Reflection, The Groovy Resources Network, http://groovy.codehaus.org/JN3535Reflection
[94] AspectC++, Wikipedia, http://en.wikipedia.org/wiki/AspectC%2B%2B
[95] The Home of AspectC++, http://www.aspectc.org/Home.1.0.html
[96] Groovy AOP, http://chanwit.blogspot.com/2007/12/groovy-aop-part-4-getter-and-setter.html
[97] Easy AOP with GroovyInterceptable, http://www.justinspradlin.com/programming/easy-aop-withgroovyinterceptable/
[98] Suman Roychoudhury, Jeff Gray, Jing Zhang, Purushotham Bangalore, and Anthony
Skjellum. A Program Transformation Technique to Support AOP within C++ Templates. Dept. of
Computer and Information Sciences, University of Alabama at Birmingham, Alabama, USA
[99] Functional Programming with Groovy, The Groovy Resources Network,
http://groovy.codehaus.org/Functional+Programming+with+Groovy
[100] Functional Programming in C++, http://www.cc.gatech.edu/~yannis/fc++/
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
99/139
[101] Boost.FC++, http://www.cc.gatech.edu/~yannis/fc++/boostpaper/fcpp.html
[102] Imperative programming, Wikipedia, http://en.wikipedia.org/wiki/Imperative_programming
[103] How to make System command calls in Java/Groovy,
http://stackoverflow.com/questions/2701547/how-to-make-system-command-calls-in-java-groovy
[104] Windows Services, The Groovy Resources Network,
http://groovy.codehaus.org/Windows+Services
[105] robertbody C++ & *.BAT, http://www.robertbody.com/prog/cpp-bat.html
[106] GUI Programming with Groovy, The Groovy Resources Network,
http://groovy.codehaus.org/GUI+Programming+with+Groovy
[107] The Groovy Resources Network, http://groovy.codehaus.org/api/index.html
[108] Abdul-Jawad, Bashar. Groovy and Grails recipes. Berkeley, Calif. : Apress ; New York, N.Y. :
Distributed to the book trade worldwide by Springer-Verlag New York.
[109] König, Dierk. Groovy in action. Greenwich, [Conn.] : Manning.
[110] Bryan O'Sullivan, Don Stewart, and John Goerzen(2008). Real World Haskell, (Chapter
17) O'Reilly Media
[111] Haskell project, memory management,
http://www.haskell.org/haskellwiki/GHC/Memory_Management
[112] Sun Microsystems (April 2006), Memory Management in the Java HotSopt Virtual
Machine
[113] Haskell project, documentation,
http://www.haskell.org/ghc/docs/6.12.2/html/libraries/base-4.2.0.1/Data-Maybe.html
[114] Bryan O'Sullivan, Don Stewart, and John Goerzen(2008). Real World Haskell, (Chapter
19) O'Reilly Media
[115] Wikipedia, Web application, http://en.wikipedia.org/wiki/Web_application
[116] Practical web programming in Haskell, Haskell wiki
http://www.haskell.org/haskellwiki/Practical_web_programming_in_Haskell
[117] Eric Armstrong , Jennifer Ball, Stephanie Bodoff, Debbie Bode, Carson, Ian Evans, Dale Green,
Kim Haase, Eric Jendrock , J2EE 1.4 Tutalial, 2004
[118] Simon Foster, University of Sheffield, HAIFA : An XML Based Interoperability
Solution for Haskell
[119] Java API for XML-Based RPC (JAX-RPC),
http://java.sun.com/webservices/jaxrpc/overview.html
[120] Martin Sulzmann, Meng Wang, Aspect-Oriented Programming with Type Classes
[121] Haskell's overlooked object system, http://homepages.cwi.nl/~ralf/OOHaskell/
[122] Oleg Kiselyov, Ralf lammel, Haskell's overlooked object system, 10 Sep, 2005[123] Andrzej
Filinski, Monadic Reflection in Haskell, Datalogisk institut,
Københavns Universitet, 2006
[124] Aspect-Oriented Programming in Java,
http://www.voelter.de/data/articles/aop/aop.html
[125] Functional Java project, http://functionaljava.org/
[126] Declarative programming, Wikipedia, http://en.wikipedia.org/wiki/Declarative_programming
[127]
Declarative
Programming
in
Java
using
Annotations
and
Reflection,
http://www.riedquat.de/articles/javaDecl
[128] System.Cmd, http://www.haskell.org/ghc/docs/6.12.1/html/libraries/process-1.0.1.2/SystemCmd.html
[129]
Applications
and
libraries/GUI
libraries,
http://www.haskell.org/haskellwiki/Applications_and_libraries/GUI_libraries
[130] Declarative Programming in Java, http://onjava.com/pub/a/onjava/2004/04/21/declarative.html
[131] Haskell, http://www.haskell.org/haskellwiki/WxHaskell
[132] Gtk2Hs, http://haskell.org/gtk2hs/
[133] Haskell/GUI, Wikepedia, http://en.wikibooks.org/wiki/Haskell/GUI
[134] QtHaskell, http://qthaskell.berlios.de/
[135] Garbage collection, Retrieve from:
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
100/139
http://phparch.cn/index.php/php/62-articles-and-reviews/299-PHP%E4%B9%8B%E5%9E%83%E5%9C%BE%E5%9B%9E%E6%94%B6
[136] Basic memory management, PHP Manual,
Retrieve from: http://www.php.net/manual/en/internals2.memory.management.php
[137] [13] Web service, Wikipedia, Retrieve: http://en.wikipedia.org/wiki/Web_service#cite_note-0
[138] A.Gutsmans, S.S Bakken and D.Rethans,Chpater 15.2 Memory Management, PHP 5 Power
programming, Indianapolis: Prentice Hall
[139] PHP exception handling,
Retrieve from: http://www.microshell.com/programming/php/php-exception-handling/
[140] D.Wampler and A. Payne (2009), Chapter 5 sensible typing ,Programming in Scala,O’REILLY
[141] M.Huniewicz, Scala-performance-garbage collection analysis,
Retrieve from:http://blog.m1key.me/2010/04/scala-performance-garbage-collection.html
[142] J.EICHAR, Exception handling,
Retrieve from:http://daily-scala.blogspot.com/2009/09/exception-handling.html
[143] PHP Web application, Retrieve from: http://en.wikipedia.org/wiki/PHP
[144] D.Polak, Lift 1.0 released, Scala Blog
Retrieve from: http://www.scala-blogs.org/2009/02/lift-10-released.html
[145] A.Gutsmans, S.S Bakken and D.Rethans, chapter 1.3.1 XML and Web Services, chapter 8.7
PHP’s SOAP Extension, PHP 5 Power programming, Indianapolis: Prentice Hall
[146] Web service, PHP Manual, Retrieve from: http://www.php.net/manual/en/refs.webservice.php
[147] A.Gutsmans, S.S Bakken and D.Rethans, Forword, PHP 5 Power programming(pp.xii),
Indianapolis: Prentice Hall
[148] http://www.scala-lang.org/node/25
[149] XML-RPC working graph, Retrieve from: http://www.xmlrpc.com/
[150] K.Waterson, Abstract Classes, Retrieve from:
http://phpro.org/tutorials/Object-Oriented-Programming-with-PHP.html
[151] Tours of Traits, Retrieve from: http://www.scala-lang.org/node/126
[152] Tours of Traits, Retrieve from: http://www.scala-lang.org/node/126
[153] Reflection (Computer Science), Wikipedia,
Retrieve from: http://en.wikipedia.org/wiki/Reflection_%28computer_science%29#PHP
[154] Reflection, Retrieve from: http://www.tuxradar.com/practicalphp/16/4/0
[155] The Reflection class, PHP Manual.
Retrieve from:http://www.php.net/manual/en/class.reflection.php
[156] Chapter 12 Scala type system, programming Scala
[157] Sample code example for Scala reflection,
Retrieve from: http://gpiancastelli.altervista.org/scala-it/esempi/cap-12/jvm-script.scala
[158] D.Sheiko, Aspect-Oritented Programming in PHP,
Retrieved from: http://www.weberdev.com/ViewArticle/Aspect-Oriented-Programming-and-PHP
[159] D.Wampler and A. Payne (2009), chapter 14: Scala tools, libraries, and IDE support: Java library
interpretability, Programming in Scala (pp.369-381), O’REILLY
[160] D. Wamper, Trait vs. Aspects in Scala, 2008,
Retrieve from: http://blog.objectmentor.com/articles/2008/09/27/traits-vs-aspects-in-scala
[161] D.Wampler and A. Payne (2009), chapter 14: Scala tools, libraries, and IDE support: Java library
interpretability, Programming in Scala (pp.378-379), O’REILLY
Available on: http://gpiancastelli.altervista.org/scala-it/esempi/cap-14/aspectj/complex.scala
http://gpiancastelli.altervista.org/scala-it/esempi/cap-14/aspectj/complex-main.scala
http://gpiancastelli.altervista.org/scala-it/esempi/cap-14/aspectj/LogComplex.aj
[162] T.K Nieleson, The state of functional programming in PHP,
Retrieve:http://www.sitepoint.com/blogs/2007/12/15/the-state-of-functional-programming-in-php/
[163][164][165]E.Begoli, Scala vs. F#: Comparing with functional programming features, 2010,
Retrieve:http://www.developer.com/article.php/3883051/Scala-vs-F-Comparing-FunctionalProgramming-Features.htm
[166] Scala (Programming language), Wikipedia
Retrieve: http://en.wikipedia.org/wiki/Scala_(programming_language)#Functional_programming
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
101/139
[168] F.Kleine and S.Schmidt, Declarative Development using Annotations in PHP [pdf document],
international PHP 2007 conference-spring edition
[169] Declarative Programming, Wikipedia
Retrieve from: http://en.wikipedia.org/wiki/Declarative_programming
[170] Declarative Programming in Java using Annotation and reflection,
Retrieve from:http://www.riedquat.de/articles/javaDecl
[171] A.Gutsmans, S.S Bakken and D.Rethans, Chapter 16.2 PHP shell scripts,PHP 5 Power
programming, Indianapolis: Prentice Hall
[172] A.Gutsmans, S.S Bakken and D.Rethans, Chapter 14 Scala Tools, Libraries and IDE Support:
CommandLineTool, PHP 5 Power programming, Indianapolis: Prentice Hall
[173] S.Mattocks (2010), introduction, Pro PHP-GTK, Apress
[174] S.Mattocks (2010), Chapter1 introduction php-gtk, Pro PHP-GTK (p11), Apress
[175] ScalaGUI, ScalaWiki, Retrieve from: http://scala.sygneca.com/code/scalagui
[176] Continuation, Wikipedia, http://en.wikipedia.org/wiki/Continuation
[177] Scheme-faq-programming, http://community.schemewiki.org/?scheme-faq-programming
[178] Rscheme, Wikipedia, http://en.wikipedia.org/wiki/RScheme
[179] Mod_lisp,
http://www.fractalconcept.com/asp/npl5/sdataQG97Qx90SRn$DM==/sdataQuvY9x3g$ecX
[180] The LAML, http://www.cs.aau.dk/~normark/laml/
[181] The Racket project, http://racket-lang.org/
[182] http://classes.eclab.byu.edu/330/wiki/index.cgi?XMLandScheme
[183] http://docs.racket-lang.org/xml/index.html
[184] http://okmij.org/ftp/Scheme/search-mslib.scm
[185] http://classes.eclab.byu.edu/330/wiki/index.cgi?SchemeAndAmazon
[186] http://www.devx.com/opensource/Article/42778/1763/page/5
[187] Aptana project, Wikipedia, http://en.wikipedia.org/wiki/Aptana
[188] Make Yahoo! Web Service REST Calls with JavaScript and XMLHttpRequest,
http://developer.yahoo.com/javascript/howto-ajax.html
[189] N.Adams and J.Rees, Object-oriented programming in scheme
[190] A class y that inherits from x, http://www.cs.aau.dk/~normark/prog3-03/html/notes/oop-schemeself-demo-note-program-2.html
[191] How do I do object-oriented programming in Scheme, http://www.faqs.org/faqs/schemefaq/part1/section-6.html
[192] Object-Oriented programming systems in Scheme, http://www.cs.indiana.edu/schemerepository/code.oop.html
[193] JavaScript Object-Oriented Programming, http://articles.sitepoint.com/article/orientedprogramming-2/4
[194] Object Oriented Programming in JavaScript, http://mckoss.com/jscript/object.htm
[195] K.E Gary and Mathew Flatt, Compiling Java to PLT Scheme
[196] Reflection in Javascript, http://lpetr.org/blog/archives/reflection-in-javascript
[197] Scheme (programming language), wikipedia,
http://en.wikipedia.org/wiki/Scheme_(programming_language)
[198] Lambda calculus, Wikipedia, http://en.wikipedia.org/wiki/Lambda_calculus
[199] Functional programming, Wikipedia, http://en.wikipedia.org/wiki/Functional_programming
[200] Frequently Asked Questions for comp.lang.functional,
http://www.cs.nott.ac.uk/~gmh/faq.html
[201] Functional Javascript , http://www.hunlock.com/blogs/Functional_Javascript
[202] Category:Functional languages, http://en.wikipedia.org/wiki/Category:Functional_languages
[203] The Relation Reflection Scheme, http://onlinelibrary.wiley.com/doi/10.1002/malq.200710035/pdf
[204] JavaScript, Wikipedia, http://en.wikipedia.org/wiki/JavaScript
[205] Structured programming, Wikipedia, http://en.wikipedia.org/wiki/Structured_programming
[205] http://www.tildemark.com/programming/javascript/adding-bookmark-this-to-your-websiteusing-javascr.html
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
102/139
[206] The Racket project, http://racket-lang.org/
[207] Adding Bookmark This link to your website using javascript,
http://www.haskell.org/ghc/docs/6.10.2/html/users_guide/options-phases.html
[208] Comparison of Java and C Sharp, Wikipedia,
http://www.javapractices.com/topic/TopicAction.do?Id=64
[209] Bpsinfo.com, http://www.bpsinfo.com/javassl/
[210] Haskell/Polymorphism, Wikibooks, http://en.wikibooks.org/wiki/Haskell/Polymorphism
[211] Closure (computer science), Wikipedia,
http://en.wikipedia.org/wiki/Closure_%28computer_science%29#Java
[212] Fran version 1.16, http://conal.net/Fran/
[213][214][215][216] Comparison of programming languages, Wikipedia,
http://en.wikipedia.org/wiki/Comparison_of_programming_languages#Type_systems
[217] Meta-Programming with Scala: Conditional Compilation and Loop Unrolling,
http://michid.wordpress.com/2008/10/29/meta-programming-with-scala-conditionalcompilation-and-loop-unrolling/
[218] Top 10 PHP MVC frameworks, http://www.mustap.com/phpzone_post_73_top-10-phpmvc-frameworks
[219] PHP introduciton, http://www.php.net/manual/en/intro.openssl.php
[220] http://phpxmlrpc.sourceforge.net/
[221] S.Abeysinghe, PHP Web Services: Getting Started, retrieve: http://wso2.org/library/3032
[222] The apache software foundation, http://ws.apache.org/axis2/
[223] Using Scala to update LiveJournal tags, http://rafaelnaufal.com/blog/2009/05/23/usingscala-to-update-livejournal-tags-part-i/
[224] WS-SecurityPolicy With PHP, http://www.dimuthu.org/blog/2008/11/19/wssecuritypolicy-with-php/
[225] http://devzone.zend.com/article/4486
[226] scala-on-struts, http://github.com/leonm/scala-on-struts
[227] Scala vs. F#: Comparing Functional Programming Features,
http://www.developer.com/article.php/3883051/Scala-vs-F-Comparing-FunctionalProgramming-Features.htm
[228] Working with Scala’s XML Support, http://www.codecommit.com/blog/scala/workingwith-scalas-xml-support
[230] http://www.web-tech-india.com/articles/php/compiling_php_apache/#why
[231] http://days2010.scala-lang.org/node/92
[232] Comparison of programming languages, Wikipedia,
http://en.wikipedia.org/wiki/Comparison_of_programming_languages
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
103/139
[233] Type system, Wikipedia, http://en.wikipedia.org/wiki/Type_system
[234] Reflection (computer science), Wikipedia,
http://en.wikipedia.org/wiki/Reflection_(computer_science)
[235] Multiple Inheritance and virtual Base Classes ,
http://www.deitel.com/articles/cplusplus_tutorials/20060225/virtualBaseClass/
[236] Reflection in C++, http://msdn.microsoft.com/en-us/library/y0114hz2(VS.80).aspx
[237] Introduction, http://www.garret.ru/cppreflection/docs/reflect.html
[238] Functional Programming in C++ , http://www.cc.gatech.edu/~yannis/fc++/
[239] Functional programming, Wikipedia,
http://en.wikipedia.org/wiki/Functional_programming
[240] Functional Programming with Groovy ,
http://groovy.codehaus.org/Functional+Programming+with+Groovy
[241] Calling external program in C++ ,
http://www.velocityreviews.com/forums/t287554-calling-external-program-in-c.html
[242] shell commands and c++ , http://forums.macrumors.com/archive/index.php/t-92081.html
[243] Executing external processes in Groovy ,
http://startbigthinksmall.wordpress.com/2010/04/23/antexec-executing-external-processes-ingroovy/
[244] http://classes.eclab.byu.edu/330/wiki/index.cgi?SchemeAndAmazon
[255] http://www.plt-scheme.org/software/openssl/
[266] http://javascriptsoapclient.codeplex.com/
[277] http://www.scala-lang.org/node/133
[278] http://en.wikipedia.org/wiki/First-class_function
[279] http://docs.racket-lang.org/scheme/index.html?q=arguments
[280] The Relation Reflection Scheme,
http://onlinelibrary.wiley.com/doi/10.1002/malq.200710035/pdf
[281] Scheme (programming language), wikipedia,
http://en.wikipedia.org/wiki/Scheme_(programming_language)
[282] Lambda calculus, http://en.wikipedia.org/wiki/Lambda_calculus
[283] Functional programming, http://en.wikipedia.org/wiki/Functional_programming
[284] Frequently Asked Questions for comp.lang.functional,
http://www.cs.nott.ac.uk/~gmh/faq.html
[285] Functional Javascript , http://www.hunlock.com/blogs/Functional_Javascript
[286] Category:Functional languages,
http://en.wikipedia.org/wiki/Category:Functional_languages
[287] Wisdom and Wonder,
http://www.wisdomandwonder.com/link/1027/linq-for-r6rs-scheme
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
104/139
Appendix
A Source code Examples
C# Console application:
Simple calculator written in C# using Windows Forms:
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
105/139
namespace COMP6411.GuiPrototype
{
partial class Form1
{
/// <summary>
/// Required designer variable.
///
/// <autor>Sleiman Rabah</autor>
/// </summary>
private System.ComponentModel.IContainer components =
null;
/// <summary>
/// Clean up any resources being used.
/// </summary>
/// <param name="disposing">true if managed resources
should be disposed; otherwise, false.</param>
protected override void Dispose(bool disposing)
{
if (disposing && (components != null))
{
components.Dispose();
}
base.Dispose(disposing);
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
106/139
#region Windows Form Designer generated code
/// <summary>
/// Required method for Designer support - do not modify
/// the contents of this method with the code editor.
/// </summary>
private void InitializeComponent()
{
this.btnCalculate = new System.Windows.Forms.Button();
this.label1 = new System.Windows.Forms.Label();
this.txtNumber1 = new System.Windows.Forms.TextBox();
this.label2 = new System.Windows.Forms.Label();
this.txtNumber2 = new System.Windows.Forms.TextBox();
this.txtResult = new System.Windows.Forms.TextBox();
this.label3 = new System.Windows.Forms.Label();
this.groupBox1 = new System.Windows.Forms.GroupBox();
this.lstErrors = new System.Windows.Forms.ListBox();
this.btnClear = new System.Windows.Forms.Button();
this.groupBox1.SuspendLayout();
this.SuspendLayout();
this.btnCalculate.Location =
new System.Drawing.Point(211, 174);
this.btnCalculate.Name = "btnCalculate";
this.btnCalculate.Size = new System.Drawing.Size(75, 35);
this.btnCalculate.TabIndex = 0;
this.btnCalculate.Text = "Calculate";
this.btnCalculate.UseVisualStyleBackColor = true;
this.btnCalculate.Click += new
System.EventHandler(this.btnCalculate_Click);
//
// label1
//
this.label1.AutoSize = true;
this.label1.Location = new System.Drawing.Point(6, 83);
this.label1.Name = "label1";
this.label1.Size = new System.Drawing.Size(74, 17);
this.label1.TabIndex = 1;
this.label1.Text = "Number 1:";
//
// txtNumber1
//
this.txtNumber1.Location = new System.Drawing.Point(86,
78);
this.txtNumber1.Name = "txtNumber1";
this.txtNumber1.Size = new System.Drawing.Size(100, 22);
this.txtNumber1.TabIndex = 2;
//
// label2
//
this.label2.AutoSize = true;
this.label2.Location = new System.Drawing.Point(6, 119);
this.label2.Name = "label2";
this.label2.Size = new System.Drawing.Size(74, 17);
this.label2.TabIndex = 3;
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
107/139
this.label2.Text = "Number 2:";
//
// txtNumber2
//
this.txtNumber2.Location =
new System.Drawing.Point(86, 116);
this.txtNumber2.Name = "txtNumber2";
this.txtNumber2.Size = new System.Drawing.Size(100, 22);
this.txtNumber2.TabIndex = 4;
//
// txtResult
//
this.txtResult.Location =
new System.Drawing.Point(86, 175);
this.txtResult.Name = "txtResult";
this.txtResult.Size = new System.Drawing.Size(100, 22);
this.txtResult.TabIndex = 5;
//
// label3
//
this.label3.AutoSize = true;
this.label3.Location = new System.Drawing.Point(25, 175);
this.label3.Name = "label3";
this.label3.Size = new System.Drawing.Size(55, 17);
this.label3.TabIndex = 6;
this.label3.Text = "Results";
//
// groupBox1
//
this.groupBox1.Controls.Add(this.btnClear);
this.groupBox1.Controls.Add(this.txtResult);
this.groupBox1.Controls.Add(this.label3);
this.groupBox1.Controls.Add(this.btnCalculate);
this.groupBox1.Controls.Add(this.txtNumber1);
this.groupBox1.Controls.Add(this.label1);
this.groupBox1.Controls.Add(this.label2);
this.groupBox1.Controls.Add(this.txtNumber2);
this.groupBox1.Location = new System.Drawing.Point(28, 24);
this.groupBox1.Name = "groupBox1";
this.groupBox1.Size = new System.Drawing.Size(310, 215);
this.groupBox1.TabIndex = 7;
this.groupBox1.TabStop = false;
this.groupBox1.Text = "Enter two numbers:";
//
// lstErrors
//
this.lstErrors.FormattingEnabled = true;
this.lstErrors.ItemHeight = 16;
this.lstErrors.Location =
new System.Drawing.Point(12, 267);
this.lstErrors.Name = "lstErrors";
this.lstErrors.Size = new System.Drawing.Size(362, 116);
this.lstErrors.TabIndex = 8;
//
// btnClear
//
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
108/139
this.btnClear.Location =
new System.Drawing.Point(211, 129);
this.btnClear.Name = "btnClear";
this.btnClear.Size = new System.Drawing.Size(75, 29);
this.btnClear.TabIndex = 7;
this.btnClear.Text = "Clear";
this.btnClear.UseVisualStyleBackColor = true;
this.btnClear.Click += new
System.EventHandler(this.btnClear_Click);
//
// Form1
//
this.AutoScaleDimensions =
new System.Drawing.SizeF(8F, 16F);
this.AutoScaleMode =
System.Windows.Forms.AutoScaleMode.Font;
this.ClientSize = new System.Drawing.Size(392, 424);
this.Controls.Add(this.lstErrors);
this.Controls.Add(this.groupBox1);
this.Name = "Form1";
this.Text = "COMP6411 - GUI Prototype";
this.groupBox1.ResumeLayout(false);
this.groupBox1.PerformLayout();
this.ResumeLayout(false);
}
#endregion
private
private
private
private
private
private
private
private
private
private
System.Windows.Forms.Button btnCalculate;
System.Windows.Forms.Label label1;
System.Windows.Forms.TextBox txtNumber1;
System.Windows.Forms.Label label2;
System.Windows.Forms.TextBox txtNumber2;
System.Windows.Forms.TextBox txtResult;
System.Windows.Forms.Label label3;
System.Windows.Forms.GroupBox groupBox1;
System.Windows.Forms.ListBox lstErrors;
System.Windows.Forms.Button btnClear;
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
109/139
using
using
using
using
System;
System.Collections.Generic;
System.Linq;
System.Windows.Forms;
namespace COMP6411.GuiPrototype
{
/// <summary>
/// <autor>Sleiman Rabah</autor>
/// </summary>
static class Program
{
/// <summary>
/// The main entry point for the application.
/// </summary>
[STAThread]
static void Main()
{
Application.EnableVisualStyles();
Application.SetCompatibleTextRenderingDefault(false);
Application.Run(new Form1());
}
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
110/139
using System.ComponentModel;
using System.Data;
UI Application:
usingC#
System.Drawing;
using System.Linq;
using System.Text;
using System.Windows.Forms;
namespace COMP6411.GuiPrototype
{
/// <summary>
///
/// A UI from which calculates two integer.
///
/// <author> Sleiman Rabah</author>
/// </summary>
public partial class Form1 : Form
{
public Form1()
{
InitializeComponent();
}
/// <summary>
/// Calculate the results based on the user entred values.
/// </summary>
/// <param name="sender"></param>
/// <param name="e"></param>
private void btnCalculate_Click(object sender, EventArgs e)
{
int result = 0;
try
{
if (this.txtNumber1.Text != ""
|| this.txtNumber2.Text != "")
{
result = Int32.Parse(txtNumber1.Text) +
Int32.Parse(this.txtNumber2.Text);
this.txtResult.Text = result.ToString();
}
}
catch (Exception ex)
{
this.lstErrors.Items.Add(ex.Message);
}
}
/// <summary>
/// Clear the UI fields.
/// </summary>
/// <param name="sender"></param>
/// <param name="e"></param>
private void btnClear_Click(object sender, EventArgs e)
{
this.txtNumber2.Text = String.Empty;
this.txtNumber1.Text = String.Empty;
this.txtResult.Text = String.Empty;
this.lstErrors.Items.Clear();
}
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
111/139
C# Console application:
Launching processes/executing a program from a C# code.
using System;
using System.Diagnostics;
using System.ComponentModel;
namespace COMP6411.ProcessLauncher
{
/// <summary>
/// <autor>Sleman Rabah</autor>
/// </summary>
class ProcessLauncher
{
// Notepad command.
public static string szNotepade = "-note";
// Calculator command.
public static string szCalculator = "-calc";
public static string szNone = "-err";
public static void Main(string[] args)
{
try
{
if (IsValid(args))
{
// instantiate the System.Diagnostics.Process class
Process myProcess = new Process();
// Tell whether the new process will be executed
// using Shell or not
myProcess.StartInfo.UseShellExecute = false;
// Fill out the process name to be started:
// often it is a program e.g: notepad
if (args[0].Equals(ProcessLauncher.szCalculator))
{
myProcess.StartInfo.FileName = "calc.exe";
}
else if
(args[0].Equals(ProcessLauncher.szNotepade))
{
myProcess.StartInfo.FileName = "notepad.exe";
}
else
{
myProcess.StartInfo.FileName = "no_one.exe";
}
myProcess.StartInfo.CreateNoWindow = true;
// Run/launch the processs
myProcess.Start();
}
else
{
// If arguments are invalid, display usage.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
112/139
Console.WriteLine("Usage: ProcessLauncher.exe
[arg1]");
Console.WriteLine("\t -calc to run
Calculator");
Console.WriteLine("\t -note to run Notepad");
Console.WriteLine("\t -err to raise an
error");
}
}
catch (Exception e)
{
Console.WriteLine("A problem has occured while
starting the process " + e.Message);
}
}
/// <summary>
/// Validates user commands.
/// </summary>
/// <param name="args">an array containing user's
commands</param>
/// <returns>boolean indicating whether commands are valid
or not</returns>
public static bool IsValid(string[] args)
{
if (args.Length == 1 )
{
if ((args[0].Equals(ProcessLauncher.szCalculator)
||args[0].Equals(ProcessLauncher.szNotepade)) ||
args[0].Equals(ProcessLauncher.szNone))
{
return true;
}
}
return false;
}
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
113/139
UI application in AspectJ and Java Swing:
Same application as shown above (using C# and Windows Forms)
package com.comp6411.aspect.gui;
import java.awt.BorderLayout;
import java.awt.FlowLayout;
import java.awt.GridLayout;
import java.awt.event.ActionEvent;
import java.awt.event.ActionListener;
import javax.swing.JButton;
import javax.swing.JLabel;
import javax.swing.JPanel;
import javax.swing.JTextArea;
import javax.swing.JTextField;
import javax.swing.WindowConstants;
/**
*
* A class for building a Swing-based GUI.
* It calculates two integers.
*
* @author Sleiman Rabah
*/
public class CalculatorSample extends javax.swing.JFrame {
private JPanel jPanel1;
private JButton btnCalculate;
private JTextArea txtErrors;
private JTextField txtNumber2;
private JTextField txtNumber1;
private JTextField txtResult;
private JLabel lblNumber2;
private JLabel lblNumber1;
private JLabel lblResult;
private JPanel jPanel2;
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
114/139
private JButton btnClear;
/**
* Auto-generated main method to display this JFrame
*/
public static void main(String[] args) {
CalculatorSample inst = new CalculatorSample();
inst.setLocationRelativeTo(null);
inst.setVisible(true);
}
public CalculatorSample() {
super();
initGUI();
}
private void initGUI() {
try {
setDefaultCloseOperation(WindowConstants.DISPOSE_ON_CLOSE);
{
jPanel1 = new JPanel();
jPanel1.setLayout(new GridLayout(0, 2));
getContentPane().add(jPanel1, BorderLayout.NORTH);
// -lblResult = new JLabel();
lblResult.setText("Result: ");
// -lblNumber1 = new JLabel();
lblNumber1.setText("Enter Number 1:");
// -txtNumber2 = new JTextField();
// -lblNumber2 = new JLabel();
lblNumber2.setText("Enter Number 2:");
// -txtNumber1 = new JTextField();
//-txtResult= new JTextField();
// -btnCalculate = new JButton();
btnCalculate.setText("Calculate");
btnCalculate.addActionListener(new ActionListener() {
public void actionPerformed(ActionEvent e) {
calculateResult();
}
});
// -btnClear = new JButton();
btnClear.setText("Clear");
btnClear.addActionListener(new ActionListener() {
public void actionPerformed(ActionEvent e) {
clearFields();
}
});
// -jPanel1.add(lblNumber1);
jPanel1.add(txtNumber1);
jPanel1.add(lblNumber2);
jPanel1.add(txtNumber2);
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
115/139
jPanel1.add(lblResult);
jPanel1.add(txtResult);
jPanel1.add(btnCalculate);
jPanel1.add(btnClear);
jPanel2 = new JPanel();
FlowLayout jPanel2Layout = new FlowLayout();
jPanel2Layout.setAlignOnBaseline(true);
getContentPane().add(jPanel2, BorderLayout.SOUTH);
jPanel2.setLayout(jPanel2Layout);
jPanel2.setPreferredSize(new java.awt.Dimension(492, 136));
txtErrors = new JTextArea();
jPanel2.add(txtErrors);
txtErrors.setPreferredSize(new java.awt.Dimension(450, 102));
}
pack();
this.setSize(500, 417);
this.setResizable(false);
} catch (Exception e) {
// add your error handling code here
e.printStackTrace();
}
}
/**
* Clear text fields.
*/
public void clearFields() {
this.txtNumber1.setText("");
this.txtNumber2.setText("");
this.txtErrors.setText("");
this.txtResult.setText("");
}
/**
* Calculate the result and update the UI fields.
*/
public void calculateResult() {
try {
int result = 0;
result = Integer.parseInt(this.txtNumber1.getText()) +
Integer.parseInt(this.txtNumber2.getText());
this.txtResult.setText(""+ result);
} catch (Exception e) {
this.txtErrors.setText("An error has occured, only number are allowed: "+
e.getMessage());
}
}
}
package com.comp6411.aspect.gui;
import java.util.logging.Level;
import java.util.logging.Logger;
import org.aspectj.lang.Signature;
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
116/139
/**
*
* An aspect that intercepts the GUI methods and handles its exceptions.
*
* @author Sleiman Rabah
*
*/
public aspect AspectGuiSample {
Logger logger = Logger.getLogger("Log MethodEntries");
/**
* Pointcut on public methods.
*/
pointcut all_publics():
call(public * CalculatorSample.*(..));
/**
* Pointcut to handle exception.
*/
pointcut exceptionHandler() :
call(* *.*(..)) && !within(AspectGuiSample);
/**
* Catch and log any exception.
*
* @param ex the exception thrown in ProcessLauncher.java
*/
after() throwing(Throwable ex) : exceptionHandler(){
logger.setLevel(Level.WARNING);
Signature methodSignature = thisJoinPoint.getSignature();
System.err.println("AspectJ has caught an exception in method: "+
methodSignature.getDeclaringTypeName()
+ "." + methodSignature.getName());
System.err.println("AspectJ-Exception Trace:" + ex.getMessage());
//-logger.info(methodSignature.getDeclaringType().getName());
}
/**
* Advice to log methods entering: All called Swing methods will be displayed.
* NOTICE: extra stuff can be done if desired before entring a method.
*/
before(): all_publics() {
Signature methodSignature = thisJoinPoint.getSignature();
System.out.println("Entring method:" + methodSignature.getName()+ "()");
}
/**
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
117/139
* Advice to log methods' execution ending: also, All called Swing methods will be displayed.
*
* NOTICE: extra stuff can be done if desired after returning from a method.
*/
after() returning() : all_publics() {
Signature methodSignature = thisJoinPoint.getSignature();
System.out.println("Leaving method: "+ methodSignature.getName()+ "()");
}
}
Console application in AspectJ:
package com.comp6411.aspect.batchscripting;
import java.io.IOException;
import java.io.InputStream;
import java.io.OutputStream;
/**
* A class which creates a process and runs a program in it.
*
* @author Sleiman Rabah
*/
public class ProcessLauncher {
/**
* Notepad command.
*/
public static String szNotepade = "-note";
/**
* Calculator command.
*/
public static String szCalculator = "-calc";
public static String szNone = "-err";
public static void main(String[] args) throws IOException {
if (ProcessLauncher.isValid(args)) {
if (args[0].equals(ProcessLauncher.szCalculator)) {
new ProcessLauncher().launchProcess("calc.exe");
}
else if (args[0].equals(ProcessLauncher.szNotepade)) {
new ProcessLauncher().launchProcess("notepad.exe");
} else {
new ProcessLauncher().launchProcess("no_one.exe");
}
} else {
// If arguments are invalid, display usage.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
118/139
System.out.println("Usage: ProcessLauncher [arg1]");
System.out.println("\t -calc to run Calculator");
System.out.println("\t -note to run Notepad");
System.out.println("\t -err to raise an error");
}
}
/**
*
* @param iPorgramToRun The name of the program to be executed.
* @throws IOException Throws an exception if failed to execute the process.
*/
public void launchProcess(String iPorgramToRun) throws IOException {
Runtime myRunTime = Runtime.getRuntime();
Process myProcess = myRunTime.exec(iPorgramToRun);
}
/**
* A method which validates entered command (user's commands)
*
* @param args An array containing user's commands
* @return Boolean indicating whether commands are valid or not
*/
public static boolean isValid(String[] args) {
if (args.length == 1) {
if ( args[0].equals(ProcessLauncher.szCalculator) ||
args[0].equals(ProcessLauncher.szNotepade) ||
args[0].equals(ProcessLauncher.szNone)
){
return true;
}
}
return false;
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
119/139
package com.comp6411.aspect.batchscripting;
import java.util.logging.Level;
import java.util.logging.Logger;
import org.aspectj.lang.Signature;
/**
* An aspect which handles ProcessLaunchers's exceptions.
*
* @author Sleiman Rabah
*/
public aspect AspectProcessLauncher {
Logger logger = Logger.getLogger("Log MethodEntries");
/**
* Pointcut on public methods.
*/
pointcut all_publics():
call(public * ProcessLauncher.*(..));
/**
* Pointcut to handle exception.
*/
pointcut exceptionHandler() :
call(* *.*(..)) && !within(AspectProcessLauncher);
/**
* Catch and log any exception.
*
* @param ex the exception thrown in ProcessLauncher.java
*/
after() throwing(Throwable ex) : exceptionHandler(){
logger.setLevel(Level.WARNING);
Signature methodSignature = thisJoinPoint.getSignature();
System.err.println("AspectJ has caught an exception: "+ methodSignature.getName());
System.err.println("AspectJ-Exception Trace:" + ex.getMessage());
//-logger.info(methodSignature.getDeclaringType().getName());
}
/**
* Advice to log methods entering.
*/
before(): all_publics() {
Signature methodSignature = thisJoinPoint.getSignature();
System.out.println("Entring method:" + methodSignature.getName()+ "()");
}
/**
* Advice to log methods' execution ending.
*/
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
120/139
after() returning() : all_publics() {
Signature methodSignature = thisJoinPoint.getSignature();
System.out.println("Leaving method: "+ methodSignature.getName()+ "()");
}
}
A1.PHP Criteria 4
The following program was compiled and run using the Zend Development Environment
available at http://www.zend.com/en/downloads/ with the following options. To execute it run
the following: File-NEWFILE, then write code in Editor, Menu-Debug-Go.
//Inheritance, Abstract Class, interface, encapsulation sample in PHP
<?php
abstract class mathematics{
/*** child class must define these methods ***/
abstract protected function getMessage();
abstract protected function addTwo($num1);
/**
* method common to both classes
**/
public function showMessage() {
echo $this->getMessage();
}
} /*** end of class ***/
class myMath extends mathematics{
/**
* Prefix to the answer
* @return string
**/
protected function getMessage(){
return "The anwser is: ";
}
/**
* add two to a number
* @access public
* @param $num1 A number to be added to
* @return int
**/
public function addTwo($num1) {
return $num1+2;
}
} /*** end of class ***/
/*** a new instance of myMath ***/
$myMath = new myMath;
/*** show the message ***/
$myMath->showMessage();
/*** do the math ***/
echo $myMath->addTwo(4);
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
121/139
?>
A2. Scala Criteria4
The following program was compiled and run using the Scala-ide for eclipse (Eclipse 3.51
with JVM) which is available at http://www.scala-ide.org/ and
http://www.eclipse.org/downloads/packages/release/galileo/sr2 with the following options.
After installed the Scala PDT, build a scala project, and create file in the default package, put
code in the scala file, then run them as Scala application.
//Inheritance, interface, encapsulation in Scala
package com.example
trait Similarity {
def isSimilar(x: Any): Boolean
def isNotSimilar(x: Any): Boolean = !isSimilar(x)
}
class Point(xc: Int, yc: Int) extends Similarity {
var x: Int = xc
var y: Int = yc
def isSimilar(obj: Any) =
obj.isInstanceOf[Point] &&
obj.asInstanceOf[Point].x == x
}
object TraitsTest extends Application {
val p1 = new Point(2, 3)
val p2 = new Point(2, 4)
val p3 = new Point(3, 3)
println(p1.isNotSimilar(p2))
println(p1.isNotSimilar(p3))
println(p1.isNotSimilar(2))
}
A3. PHP Criteria 6
The following program was compiled and run using the Zend Development Environment
available at http://www.zend.com/en/downloads/ with the following options. To execute it run
the following: Menu->File->NEWFILE, then write code in Editor, Menu->Debug->Go.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
122/139
<?PHP
include("aop.lib.php"); // aop.lib.php is an external supported AOP library
// , provide in the .zipfile with the submitted report
$aspect1 = new Aspect();
$pc1 = $aspect1->pointcut("call Sample::Sample or call Sample::Sample2");
$pc1->_before("print 'PreProcess<br />';");
$pc1->_after("print 'PostProcess<br />';");
$pc1->destroy();
Class Sample {
var $aspect;
function Sample($aspect1) {
$this->aspect = &$aspect1;
Advice::_before($this->aspect);
print 'Some business logic of Sample<br />';
Advice::_after($this->aspect);
}
function Sample2() {
Advice::_before($this->aspect);
print 'Some business logic of Sample2<br />';
Advice::_after($this->aspect);
}
}
$Sample = new Sample(&$aspect1);
$Sample->Sample2();
?>
A4.
Scal
a
Crite
ria 6
The
followi
ng
progra
m was
compil
ed and
run
using
the
Scalaide for
eclipse
(Eclips
e 3.51
with
JVM
and
AJDT
plugin)
which
is available at
http://www.scala-ide.org/ , http://www.eclipse.org/downloads/packages/release/galileo/sr2 ,
and http://www.eclipse.org/ajdt/downloads with the following options. After installed the Scala
pdt and AJDT plug-in, create the Aspect project, then create a package , then put two .scala file
into package. Then create an aspect in the package named LogComplex.aj. Finally, then run as
Aspect program.
// code-examples/ToolsLibs/aspectj/complex.scala
package example.aspectj
case class Complex(real: Double, imaginary: Double) {
def +(that: Complex) =
new Complex(real + that.real, imaginary + that.imaginary)
def -(that: Complex) =
new Complex(real - that.real, imaginary - that.imaginary)
}
// code-examples/ToolsLibs/aspectj/complex-main.scala
package example.aspectj
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
123/139
object ComplexMain {
def main(args: Array[String]) {
val c1 = Complex(1.0, 2.0)
val c2 = Complex(3.0, 4.0)
val c12 = c1 + c2
println(c12)
}
}
// code-examples/ToolsLibs/aspectj/LogComplex.aj
// define point cut, advice for scala file.
package example.aspectj;
public aspect LogComplex {
public pointcut newInstances(double real, double imag):
execution(Complex.new(..)) && args(real, imag);
public pointcut plusInvocations(Complex self, Complex other):
execution(Complex Complex.$plus(Complex)) && this(self) && args(other);
before(double real, double imag): newInstances(real, imag) {
System.out.println("new Complex(" + real + "," + imag + ") called.");
}
before(Complex self, Complex other): plusInvocations(self, other) {
System.out.println("Calling " + self + ".+(" + other + ")");
}
after(Complex self, Complex other) returning(Complex c):
plusInvocations(self, other) {
System.out.println("Complex.+ returned " + c);
}
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
124/139
C++ vs Grouvy:
A.1 C++ Criteria 4
The following program was compiled and run using the GCC compiler in VC2008 which is
available at http://www.microsoft.com/downloads/details.aspx?FamilyID=9b2da534-3e034391-8a4d-074b9f2bc1bf&displaylang=en. To execute it run the following code in the cpp
file..
#ifndef FRUIT_H
#define FRUIT_H
#include<string>
Using namespace std;
Class Fruit
{
Public:
Virtual void identify(){cout<<”Fruit”<<endl;}
}
Class Apple:public Fruit
{
Public:
Void identify(){cout<<”Apple”<<endl;}
}
Class Orange:public Fruit
{
Public:
Void identify(){cout<<”Orange”<<endl;}
}
#include<iostream>
Using namespace std;
#include<iomanip>
#include “Fruit.h”
Int main()
{
Fruit f;
Apple a;
Orange o;
f.identify();
a.identify();
o.identify();
return 0;
}
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
125/139
A.2 Groovy Criteria 4
The following program was compiled and run using the Groovy for eclipse (Eclipse 3.51 with
JVM) which is available at http://groovy.codehaus.org/ and
http://www.eclipse.org/downloads/packages/release/galileo/sr2 with the following options.
After
installed the Groovy, build a Groovy project, and create a Groovy class in the default package,
put code in the class, then run them as Groovy application.
abstract class A{
public int prev //field
int signature //property
abstract String sayFly(int k)//abastract method
}
class B extends A{
String sayBirds(int n){ "There are $n birds!" }
String sayFly(int k){"There are $k flys!"}
}
def b= new B()
//def a= new A()
assert b.sayBirds(17) == 'There are 17 birds!'
assert b.sayFly(10) == 'There are 10 flys!'
b.signature= 19
assert b.signature == 19 //property 'signature' from A acts as
part of B
assert b.getSignature() == 19
A.3 C++ Criteria 6
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
126/139
The following program was compiled and run using the GPL under VC2008 which is available
at http://groovy.codehaus.org/ and http://www.aspectc.org/Download.2.0.html with the
following options. After installed the AspectC++, build a project, and create a class in the
default package, put code in the class, then run them as AspectC++ application.
aspect Logging
{
ostream * _out ; // ordinary attributes
public:
void bind_stream (ostream *o) { _out = o; } // member
function
pointcut virtual logged_classes () = 0; // pure virtual
pointcut
// some advice
advice execution(" % ...::%(...) ") && within(
logged_classes ()) :
before () {
*_out << "executing " << JoinPoint :: signature () << endl;
}
};
A.4 Groovy Criteria 6
The following program was compiled and run using the Groovy for eclipse (Eclipse 3.51 with
JVM) which is available at http://groovy.codehaus.org/ and
http://www.eclipse.org/downloads/packages/release/galileo/sr2 with the following options.
After
installed the Groovy, build a Groovy project, and create a Groovy class in the default package,
put code in the class, then run them as Groovy application.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
127/139
class SimplePOGO implements GroovyInterceptable {
void simpleMethod1(){
System.out.println("simpleMethod1() called")
}
void simpleMethod2(String param1, Integer param2){
System.out.println("simpleMethod2(${param1},${param2})
called")
System.out.println("sleeping...")
Timer.sleep(2000)
}
def invokeMethod(String name, args){
System.out.println("time before ${name} called: ${new Date()}")
//Get the method that was originally called.
def calledMethod = SimplePOGO.metaClass.getMetaMethod(name,
args)
//The "?" operator first checks to see that the "calledMethod" is not
//null (i.e. it exists).
calledMethod?.invoke(this, args)
System.out.println("time after ${name} called: ${new Date()}\n")
}
}
simplePogo = new SimplePOGO()
simplePogo.simpleMethod1()
simplePogo.simpleMethod2("stringParam", 24)
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
128/139
A.1 Haskell, Criteria: Default secure programming practices,
Exception handling part
The following program show how Maybe handle exception
divBy :: Integral a => a -> [a] -> Maybe [a]
divBy _ [] = Just []
divBy _ (0:_) = Nothing
divBy numerator (denom:xs) =
case divBy numerator xs of
Nothing -> Nothing
Just results -> Just ((numerator `div` denom) : results)
Use of Maybe [6]
The following program show how Either handle exception
divBy :: Integral a => a -> [a] -> Either String [a]
divBy _ [] = Right []
divBy _ (0:_) = Left "divBy: division by 0"
divBy numerator (denom:xs) =
case divBy numerator xs of
Left x -> Left x
Right results -> Right ((numerator `div` denom) : results)
Use of Either [6]
A.2 Java, Criteria: Default secure programming practices, Exception
Handling part
The following program show Java exception handling.
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
129/139
public class sample{
public int divide( int a, int b){
if ( b == 0) throw new DivisionByZeroException( );
return a/b;
}
public static void main (String arg[]){
try{
int result = divide( 5, 0);
}
catch(DivisionByZeroException e){
system.out.println(e.getMessage());
}
}
}
class DivisionByZeroException extends Exception{
DivisionByZeroException( ){
super(“Division by 0”);
}
DivisionByZeroException(String msg){
Super( msg );
}
}
Java exception handling
A.3 Haskell, Criteria: Web application
import Network.CGI
import Text.XHtml
page :: Html
page = body << h1 << "Hello World!"
cgiMain :: CGI CGIResult
cgiMain = output $ renderHtml page
main :: IO ()
main = runCGI $ handleErrors cgiMain
Haskell web application 1 – out put text [7]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
130/139
import Network.CGI
import Text.XHtml
inputForm = form << [paragraph << ("My name is " +++ textfield "name"),
submit "" "Submit"]
greet n = paragraph << ("Hello " ++ n ++ "!")
page t b = header << thetitle << t +++ body << b
cgiMain = do mn <- getInput "name"
let x = maybe inputForm greet mn
output $ renderHtml $ page "Input example" x
main = runCGI $ handleErrors cgiMain
-- Get the value of an input variable, for example from a form.
-- If the variable has multiple values, the first one is returned.
getInput :: String -> CGI (Maybe String)
Haskell web application 2 — Get user input [7]
A.4 Java, Criteria: Web application
import java.io.IOException;
import java.io.PrintWriter;
import javax.servlet.ServletException;
import javax.servlet.http.HttpServlet;
import javax.servlet.http.HttpServletRequest;
import javax.servlet.http.HttpServletResponse;
public final class Hello extends HttpServlet {
public void doGet(
HttpServletRequest request, HttpServletResponse
response)throws IOException, ServletException {
response.setContentType("text/html");
PrintWriter writer = response.getWriter();
writer.println("<html>");
writer.println("<head>");
writer.println("<title>A Sample Application</title>");
writer.println("</head>");
writer.println("<body>");
writer.println("Hello world");
writer.println("</body>");
writer.println("</html>");
}
}
Java web application — Output a text [7]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
131/139
A.5 Haskell, Criteria: OO-based abstraction
shape newx newy self
= do
-- Create references for private state
x <- newIORef newx
y <- newIORef newy
-- Return object as record of methods
returnIO $ getX
.=. readIORef x
.*. getY
.=. readIORef y
.*. setX
.=. writeIORef x
.*. setY
.=. writeIORef y
.*. moveTo
.=. (\newx newy -> do
(self # setX) newx
(self # setY) newy )
.*. rMoveTo .=. (\deltax deltay ->
do
x <- self # getX
y <- self # getY
(self # moveTo) (x + deltax) (y + deltay) )
.*. emptyRecord
Object generator for shapes [13]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
132/139
rectangle newx newy width height self
= do
-- Invoke object generator of superclass
super <- shape newx newy self
-- Create references for extended state
w <- newIORef width
h <- newIORef height
-- Return object as record of methods
returnIO $
getWidth .=.
readIORef w
.*.
getHeight .=.
readIORef h
.*.
setWidth .=.
(\neww -> writeIORef w neww)
.*.
setHeight .=.
(\newh -> writeIORef h newh)
.*.
draw .=.
do -- Implementation of the abstract draw method
putStr "Drawing a Rectangle at:(" <<
self # getX << ls "," << self # getY <<
ls "), width " << self # getWidth <<
ls ", height " << self # getHeight << ls "\n"
-- Rectangle records start from shape records
.*. super
Object generator for rectangles [13]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
133/139
circle newx newy newradius self
= do
super <- shape newx newy self
...
returnIO ... .*. super
Object generator for circles [13]
myOOP = do
-- Construct objects
s1 <- mfix (rectangle (10::Int) (20::Int) 5 6)
s2 <- mfix (circle (15::Int) 25 8)
-- Create a homogeneous list of different shapes
let scribble = consLub s1 (consLub s2 nilLub)
-- Loop over list with normal monadic map
mapM_ (\shape -> do
shape # draw
(shape # rMoveTo) 100 100
shape # draw)
scribble
Object construction and invocation as a monadic sequence [13]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
134/139
A.6 Java, Criteria: OO-based abstraction
public class Shape{
protected int x;
protected int y;
public Shape(int x, int y){
this.x = x;
this.y = y;
}
public int getX(){ return x; }
public int getY(){ return y; }
public void setX(int x){ this.x = x; }
public void setY(int y){ this.y = y; }
public void moveTo(int x, int y){
this.x = x;
this.y = y;
}
public void rMoveTo(int deltaX, int deltaY){
this.x += deltaX;
this.y += deltaY;
}
}
Object generator for shapes
public class Rectangle extends Shape{
private int w;
private int h;
public Rectangle (int x, int y, int w, int h){
super(x, y);
this.w = w;
this.h = h;
}
public int getW(){ return w; }
public int getH(){ return h; }
public void setW(int w){ this.w = w; }
public void setH(int h){ this.h = h; }
public void draw(){
system.out.println("Drawing a Rectangle at: " + x +", " + y + "Width: " +
w + "Height: " + h);
}
}
Object generator for rectangle
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
135/139
public class Circle extends Shape{
// similar to class Rectangle
}
Object generator for circle
public class Main{
public static void main(String arg[]){
Shape s1 = new Rectangle(10, 20, 5, 6);
Shape s2 = new Cirle(15, 25, 8);
s1.draw();
s2.draw();
s1.moveTo(20, 30);
s1.rMoveTo(10, 20);
}
}
Main class
A.7 Haskell, Criteria: Aspect-oriented
import List(sort)
insert x [] = [x]
insert x (y:ys)
| x <= y = x:y:ys
| otherwise = y : insert x ys
insertionSort [] = []
insertionSort xs =
insert (head xs) (insertionSort (tail xs))
-- sortedness aspect
N1@advice #insert# :: Ord a => a -> [a] -> [a] =
\x -> \ys ->
let zs = proceed x ys
in if (isSorted ys) && (isSorted zs)
then zs else error "Bug"
where
isSorted xs = (sort xs) == xs
-- efficiency aspect
N2@advice #insert# :: Int -> [Int] -> [Int] =
\x -> \ys ->
if x == 0 then x:ys
else proceed x ys
AOP Haskell Example[12]
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
136/139
A.8 Java, Criteria: Aspect-oriented
package com.reg.dev.aspects;
public class HelloWorld {
public void print() {
System.out.println("Hello World");
}
public static void main(String [] args) {
HelloWorld hw = new HelloWorld();
hw.print();
}
}
//the aspect pointcuts are dynamically weaved to create the following output:
Entering print
Hello World
Exiting print
AspectJ Example and output
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
137/139
A.9 Haskell, Criteria: UI Prototype design
import Graphics.UI.Gtk
main :: IO ()
main = do
initGUI
window <- windowNew
hbox <- hBoxNew True 10
button1 <- buttonNewWithLabel "Button 1"
button2 <- buttonNewWithLabel "Button 2"
set window [windowDefaultWidth := 200, windowDefaultHeight := 200,
containerBorderWidth := 10, containerChild := hbox]
boxPackStart hbox button1 PackGrow 0
boxPackStart hbox button2 PackGrow 0
onDestroy window mainQuit
widgetShowAll window
mainGUI
Gtk2Hs sample and output
A.10 Java, Criteria: UI Prototype design
Sample code output
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
138/139
public class Example extends javax.swing.JFrame {
public Example(){ initComponents();}
private void initComponents() {
mainFrame = new javax.swing.JFrame();
button1 = new javax.swing.JButton();
button2 = new javax.swing.JButton();
javax.swing.GroupLayout mainFrameLayout = new
javax.swing.GroupLayout(mainFrame.getContentPane());
mainFrame.getContentPane().setLayout(mainFrameLayout);
mainFrameLayout.setHorizontalGroup(
mainFrameLayout.createParallelGroup(javax.swing.
GroupLayout.Alignment.LEADING)
.addGap(0, 400, Short.MAX_VALUE));
mainFrameLayout.setVerticalGroup(
mainFrameLayout.createParallelGroup(javax.swing.
GroupLayout.Alignment.LEADING)
.addGap(0, 300, Short.MAX_VALUE));
setDefaultCloseOperation(javax.swing.WindowConstants.EXIT_ON_CLOSE);
setTitle("Packing");
button1.setText("Button 1");
button2.setText("Button2");
javax.swing.GroupLayout layout = new
javax.swing.GroupLayout(getContentPane());
getContentPane().setLayout(layout);
layout.setHorizontalGroup(
layout.createParallelGroup(javax.swing.GroupLayo
ut.Alignment.LEADING)
.addGroup(layout.createSequentialGroup()
.addContainerGap()
.addComponent(button1,javax.swing.GroupLayout.PREFERRED_SIZE, 152,
javax.swing.GroupLayout.PREFERRED_SIZE)
.addGap(18, 18, 18)
.addComponent(button2, javax.swing.GroupLayout.PREFERRED_SIZE, 152,
javax.swing.GroupLayout.PREFERRED_SIZE)
.addContainerGap(javax.swing.GroupLayout.DEFAULT_SIZE, Short.MAX_VALUE)));
layout.setVerticalGroup(
layout.createParallelGroup(javax.swing.GroupLayout.Alignment.LEADING)
.addGroup(layout.createSequentialGroup()
.addContainerGap()
.addGroup(layout.createParallelGroup(javax.swing.GroupLayout.Alignment.BASELINE)
.addComponent(button1, javax.swing.GroupLayout.PREFERRED_SIZE, 223,
javax.swing.GroupLayout.PREFERRED_SIZE)
.addComponent(button2, javax.swing.GroupLayout.PREFERRED_SIZE, 223,
javax.swing.GroupLayout.PREFERRED_SIZE))
.addContainerGap(javax.swing.GroupLayout.DEFAULT_SIZE, Short.MAX_VALUE)));
pack();
}
public static void main(String args[]) {
java.awt.EventQueue.invokeLater(new Runnable() {
public void run() { new Example().setVisible(true); }});
}
private javax.swing.JButton button1, button2;
private javax.swing.JFrame mainFrame;
}
Java Swing sample
Java Swing Sample codes
COMP 6411 - A Comparative studies of programming languages
Sleiman Rabah, Jiang Li, Mingzhi Liu, Yuanwei Lai
139/139
| 6cs.PL
|
1
Coordinate Dual Averaging for Decentralized Online
Optimization with Nonseparable Global Objectives*
arXiv:1508.07933v2 [math.OC] 20 May 2016
Soomin Lee, Angelia Nedić, and Maxim Raginsky
Abstract—We consider a decentralized online convex optimization problem in a network of agents, where each agent controls
only a coordinate (or a part) of the global decision vector. For
such a problem, we propose two decentralized variants (ODAC and ODA-PS) of Nesterov’s primal-dual algorithm with dual
averaging. In ODA-C, to mitigate the disagreements on the
primal-vector updates, the agents implement a generalization of
the local information-exchange dynamics recently proposed by Li
and Marden [1] over a static undirected graph. In ODA-PS, the
agents implement the broadcast-based push-sum dynamics [2]
over a time-varying sequence of uniformly connected digraphs.
We show that√the regret bounds in both cases have sublinear
growth of O( T ),√ with the time horizon T , when the stepsize
is of the form 1/ t and the objective functions are Lipschitzcontinuous convex functions with Lipschitz gradients. We also
implement the proposed algorithms on a sensor network to
complement our theoretical analysis.
I. I NTRODUCTION
Decentralized optimization has recently been receiving significant attention due to the emergence of large-scale distributed algorithms in machine learning, signal processing,
and control applications for wireless communication networks,
power networks, and sensor networks; see, for example, [3]–
[8]. A central generic problem in such applications is decentralized resource allocation for a multiagent system, where
the agents collectively solve an optimization problem in the
absence of full knowledge about the overall problem structure.
In such settings, the agents are allowed to communicate
to each other some relevant estimates so as to learn the
information needed for an efficient global resource allocation.
The decentralized structure of the problem is reflected in the
agents’ local view of the underlying communication network,
where each agent exchanges messages only with its neighbors.
In recent literature on control and optimization, an extensively studied decentralized resource allocation problem is one
where the system objective function f (x) P
is given as a sum
n
of local objective functions, i.e., f (x) = i=1 fi (x) where
fi is known only to agent i; see, for example [9]–[25]. In
this case, the objective function is separable across the agents,
but the agents are coupled through the resource allocation
*The work has been partially supported by the National Science Foundation
under grant no. CCF 11-11342 and by the Office of Naval Research under
grant no. N00014-12-1-0998.
S. Lee and M. Raginsky are with the Department of Electrical and Computer
Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, 61801
USA. A. Nedić is with the Department of Industrial and Enterprise Systems
Engineering, University of Illinois at Urbana-Champaign, Urbana, IL, 61801
USA. Their emails are {lee203,angelia,maxim}@illinois.edu.
Portions of this work were presented at the 2015 American Control Conference and at the 2015 International Symposium on Mathematical Programming.
vector x. Each agent maintains and updates its own copy
of the allocation/decision vector x, while trying to estimate
an optimal decision for the system problem. The vector x is
assumed to lie in (a subset of) Rd , where d may or may not
coincide with the number of agents n.
Another decentralized resource allocation problem is the
one where the system objective function f (x)
Pn may not admit a natural decomposition of the form
i=1 fi (x), and
the resource allocation vector x = (x1 , . . . , xn ) ∈ Rn is
distributed among the agents, where each agent i is responsible
for maintaining and updating only a coordinate (or a part)
xi of the whole vector x. Such decentralized problems have
been considered in [26]–[30] (see also the textbook [31]). In
the preceding work, decentralized approaches converge when
the agents are using weighted averaging, or when certain
contraction conditions are satisfied. Recently, Li and Marden
[1] have proposed a different algorithm with local updates,
where each agent i keeps estimates for the variables xj , j 6= i,
that are controlled by all the other agents in the network.
The convergence of this algorithm relies on some contraction
properties of the iterates. Note that all the aforementioned
algorithms were developed for offline optimization problems.
Our work in this paper is motivated by the ideas of Li and
Marden [1] and also by the broadcast-based subgradient push,
which was originally developed by Kempe et al. [2] and later
extended in [32] and in [15], [16] to distributed optimization.
Specifically, we use the local information exchange model
of [1] and [2], [15], [16], [32], but employ a different online
decentralized algorithm motivated by the work of Nesterov
[33]. We call these algorithms ODA-C (Online Dual Averaging with Circulation-based communication) and ODA-PS
(Online Dual Averaging with Push-Sum based communication), respectively.
In contrast with existing methods, our algorithms have
the following distinctive features: (1) We consider an online
convex optimization problem with nondecomposable system
objectives, which are functions of a distributed resource allocation vector. (2) In our algorithms, each agent maintains
and updates its private estimate of the best global allocation
vector at each time, but contributes only one coordinate to the
network-wide decision vector. (3) We provide regret bounds
in terms of the true global resource allocation vector x (rather
than some estimate on x by a single agent). For both ODA-C
and ODA-PS,
√ we show that the regret has sublinear growth
of √order O( T ) in time T with the stepsize of the form
1/ t + 1.
Our proposed algorithm ODA-PS is closest to recent papers
[34], [35]. The papers proposed a decentralized algorithm for
2
online convex optimization which is very similar to ODAPS in a sense that they also introduce online subgradient
estimations in primal [34] or dual [35] space into information
aggregation using push-sum. In these papers, the agents share
a common decision set in Rd , the objective functions
Pnare separable across the agents at each time (i.e., ft (x) = i=1 fti (x)
for all t), and the regret is analyzed in terms of each agent’s
own copy of the whole decision vector x ∈ Rd . Moreover,
an additional assumption is made in [34] that the objective
functions are strongly convex.
The paper is organized as follows. In Section II, we formalize the problem and describe how the agents interact. In
Section III, we provide an online decentralized dual-averaging
algorithm in a generic form and establish a basic regret bound
which can be used later for particular instantiations, namely,
for the two algorithms ODA-C and ODA-PS. These algo√
rithms are analyzed in Sections IV, where we establish O( T )
regret bounds under mild assumptions. In Section VI, we
demonstrate our analysis by simulations on a sensor network.
We conclude the paper with some comments in Section VII.
Notation: All vectors are column vectors. For vectors associated with agent i, we use a subscript i such as, for example,
xi , zi , etc. We will write xki to denote the kth coordinate value
of a vector xi . We will work with the Euclidean norm, denoted
by k·k. We will use e1 , . . . , en to denote the unit vectors in the
standard Euclidean basis of Rn . We use 1 to denote a vector
with all entries equal to 1, while I is reserved for an identity
matrix of a proper size. For any n ≥ 1, the set of integers
{1, . . . , n} is denoted by [n]. We use σ2 (A) to denote the
second largest singular value of a matrix A.
II. P ROBLEM FORMULATION
Consider a multiagent system (network) consisting of n
agents, indexed by elements of the set V = [n]. Each agent
i ∈ V takes actions in an action space X, which is a closed and
bounded interval of the real line.1 At each time, the multiagent
system incurs a time-varying cost ft , which comes from a fixed
class F of convex functions f : Xn → R.
The communication among agents in the network is governed by either one of the two following models:
(G1) An undirected connected graph G = (V, E): If agents
i and j are connected by an edge (which we denote
by i ↔ j), then they may exchange information with
one another. Thus, each agent i ∈ V may directly
communicate only with the agents in its neighborhood
Ni , {j ∈ V : i ↔ j} ∪ {i}. Note that agent i is always
contained in its own neighborhood.
(G2) Time-varying digraphs G(t) = (V, E(t)), for t ≥ 1: If
there exists a directed link from agent j to i at time
t (which we denote by (j, i)), agent j may send its
information to agent i. We use the notation Niin (t) and
Niout (t) to denote the in and out neighbors of agent i at
time t, respectively. That is,
Niin (t) , {j | (j, i) ∈ E(t)} ∪ {i},
1 Everything easily generalizes to X being a compact convex subset of a
multidimensional space Rd ; we mainly stick to the scalar case for simplicity.
Niout (t) , {j | (i, j) ∈ E(t)} ∪ {i}.
In this case, we assume that there always exists a selfloop (i, i) for all agent i ∈ V. Therefore, agent i is
always contained in its own neighborhood. Also, we use
di (t) to denote the out degree of node i at time t. i.e.,
di (t) , |Niout (t)|.
We assume B-strong connectivity of the graphs G(t)
with some scalar B > 0, i.e., a graph with the following
edge set
tB
[
EB (t) =
E(i)
i=(t−1)B+1
is strongly connected for every t ≥ 1. In other words,
the union of the edges appearing for B consecutive time
instances periodically constructs a strongly connected
graph. This assumption is required to ensure that there
exists a path from one node to every other node infinitely
often even if the underlying network topology is timevarying.
The network interacts with an environment according to the
protocol shown in Figure 1. We leave the details of the signal
generation process vague for the moment, except to note that
the signals received by all agents at time t may depend on all
the information available up to time t (including f1 , . . . , ft , as
well as all of the local information exchanged in the network).
Moreover, the environment may be adaptive, i.e., the choice
of the function ft may depend on all of the data generated by
the network up to time t.
Parameters: base action space X; network graph G = (V, E);
function class F
For each round t = 1, 2, . . .:
(1) Each agent i ∈ V selects an action xi (t) ∈ X
(2) Each agent i ∈ V exchanges local information with its
neighbors Ni
(3) The environment selects the current objective ft ∈ F , and
each agent receives a signal about ft
Fig. 1.
Online optimization with global objectives and local information.
Let us denote the network action at time t by
x(t) = (x1 (t), . . . , xn (t)) ∈ Xn .
(1)
We consider the network regret R(T ) at an arbitrary time
horizon T ≥ 1:
R(T ) ,
T
X
t=1
ft (x(t)) − infn
y∈X
T
X
ft (y).
(2)
t=1
Thus, R(T ) is the difference between the total cost incurred
by the network at time T and the smallest total cost that could
have been achieved with a single action in Xn in hindsight (i.e.,
with perfect advance knowledge of the sequence f1 , . . . , fT )
and without any restriction on the communication between the
agents. The problem is to design the rule (or policy) each agent
i ∈ V should use to determine its action xi (t) based on the
3
local information available to it at time t, such that the regret
in (2) is (a) sublinear as a function of the time horizon T and
(b) exhibits “reasonable” dependence on the number of agents
n and on the topology of the communication graphs.
The regret in (2) is defined over the true network actions
of individual agents, i.e., xi (t)’s, rather than in terms of
some estimates of x(t) by individual agents. This notion
of regret, which, to the best of our knowledge has been
first introduced in [30], is inspired by the literature on team
decision theory and decentralized control problems: The online
optimization is performed by a team of cooperating agents
facing a time-varying sequence of global objective functions
ft , which are
(in contrast to decomposable
P nondecomposable
i
i
objectives
i ft (x), where ft is only revealed to agent i).
Communication among agents is local, as dictated by the
network topology, so no agent has all the information in order
to compute a good global decision vector x(t). By comparing
the cumulative performance of the decentralized system to the
best centralized decision achievable in hindsight, the regret
in (2) captures the effect of decentralization. It also calls for
analysis techniques that are different from existing methods in
the literature.
III. T HE BASIC ALGORITHM AND REGRET BOUND
We now introduce a generic algorithm for solving the
decentralized online optimization problem defined in Section
II. The algorithm uses the dual-averaging subgradient method
of Nesterov [33] as an optimization subroutine.
Each agent i ∈ V generates a sequence {xi (t), zi (t)}∞
t=1 in
n
X × Rn , where the primal iterates
xi (t) = (x1i (t), . . . , xni (t)) ∈ Xn
and the dual iterates
zi (t) = (zi1 (t), . . . , zin (t)) ∈ Rn
are updated recursively as follows:
1 k
k
δ ui (t) + Fi,t
(mi (t)) , k ∈ [n]
ri i
xi (t + 1) = Πψ
Xn (Gi,t (zi (t + 1)), α(t))
zik (t + 1) =
Euclidean norm k · k, i.e., for any x, y ∈ Xn we have
1
¯
(5)
ψ(y) ≥ ψ(x) + h∇ψ(x),
y − xi + kx − yk2 ,
2
¯ denotes an arbitrary subgradient of ψ.
where ∇ψ
The dual iterate zi (t) computed by agent i at time t will
be an estimate of the “running average of the subgradients”
as seen by agent i, and will constitute an approximation
of the true centralized dual-averaging subgradient update of
Nesterov’s algorithm. The messages from Ni entering into the
dual-space dynamics are crucial for mitigating any disagreement between the agents’ local estimates of what the network
action should be. The primal iterate xi (t) of agent i at time
t is an approximation of the true centralized primal point for
the subgradient evaluation.
Note that in (3a) the local update ui (t) based on the signal
about ft affects affects only the ith coordinate of the dual
iterate zi (t + 1), while all other coordinates with k 6= i remain
untouched except for the averaging. The action of agent i at
time t is then given by
xi (t) = xii (t),
i.e., by the ith component of the vector xi (t).
A concrete realization of the algorithm (3a)-(3b) requires
specification of the rules for computing the local update ui (t),
k
the messages exchanged by the agents, and the mappings Fi,t
and Gi,t . In this paper, we present two different instantiations
of this algorithm, namely, the circulation-based method inspired by [1] and the push-sum based method inspired by [2],
[15], [16], [32]. We call these algorithms ODA-C (Online Dual
Averaing with Circulation-based communication) and ODAPS (Online Dual Averaing with Push-Sum based communication) and detail them in Section IV and V, respectively.
We now present a basic regret bound that can be used
for any generic algorithm of the form (3a)-(3b) under the
following assumption:
Assumption 1: All functions f ∈ F are Lipschitz continuous with a constant L:
|f (x) − f (y)| ≤ Lkx − yk
(3a)
(3b)
with the initial condition zi (0) = 0 for all i ∈ V. In the
dual update (3a), δik is the Kronecker delta symbol, ri > 0
is a positive weight parameter, ui (t) ∈ R is a local update
computed by agent i at time t based on the received signal
about ft , mi (t) are the messages received by agent i at time t
[from agents in Ni under the model (G1) or from Niin (t) under
k
the model (G2)], and Fi,t
, k ∈ [n], are real-valued mappings
that perform local averaging of mi (t). In the primal update
(3b), Gi,t : Rn → Rn is a mapping on dual iterates, {α(t)}∞
t=0
is a nonincreasing sequence of positive step sizes, and the
n
n
mapping Πψ
Xn : R × (0, ∞) → X is defined by
1
ψ
(4)
ΠXn (z, α) , arg min hz, xi + ψ(x) ,
α
x∈Xn
where ψ : Xn → R+ is a nonnegative proximal function.
We assume that ψ is 1-strongly convex with respect to the
for all x, y ∈ Xn .
n
Theorem 1: Let {xi (t)}∞
t=1 ⊂ X , i ∈ V, be the sequences of the agents’ primal iterates, let {u(t)}∞
t=1 with
u(t) = (u1 (t), . . . , un (t)) be the sequence of the agents’ local
n
updates, and let {x̄(t)}∞
t=1 ⊂ X be generated as
!
t
X
ψ
x̄(t + 1) = ΠXn
u(s), α(t) .
(6)
s=0
Then, under Assumption 1, the network regret R(T ) in (2)
can be upper-bounded in terms of u(t) and x̄(t) as follows:
for each T ≥ 1,
T
C
1X
α(t − 1)ku(t)k2 +
R(T ) ≤
2 t=1
α(T )
|
{z
}
(E1)
+L
T X
n
X
T
X
√
kxi (t) − x̄(t)k + nDX
k∇ft (x̄(t)) − u(t)k,
t=1 i=1
|
t=1
{z
(E2)
}
|
{z
(E3)
}
4
IV. ODA-C AND ITS REGRET BOUND
where DX , supx,y∈X |x − y| is the diameter of the set X, and
C , supx∈Xn |ψ(x)|.
Remark Since ψ is a continuous function on the compact set
Xn , C < ∞ by the Weierstrass theorem.
We now introduce a decentralized online optimization algorithm which uses a circulation-based framework for its dual
update rule (3a). We refer to this algorithm as ODA-C (Online
Dual Averaing with Circulation-based communication). ODAC uses the network model (G1) for its communication.
Proof: For any t and any y ∈ Xn we can write
A. ODA-C
ft (x(t)) − ft (y)
Fix a vector r = (r1 , . . . , rn ) of positive weights and a
nonnegative n×n matrix M , such that Mij 6= 0 only if j ∈ Ni ,
satisfying the following symmetry condition:
= ft (x(t)) − ft (x̄(t)) + ft (x̄(t)) − ft (y)
≤ h∇ft (x(t)), x(t) − x̄(t)i + h∇ft (x̄(t)), x̄(t) − yi
≤ Lkx(t) − x̄(t)k + h∇ft (x̄(t)), x̄(t) − yi,
(7)
where the second step follows from convexity of ft , while the
last step uses the fact that all f ∈ F are L-Lipschitz. Recalling
that x(t) is the network action vector (see (1)), we have the
following for the first term in (7):
kx(t) − x̄(t)k =
n
X
ri Mij = rj Mji ,
≤
zik (t + 1) =
xi (t) − x̄i (t) ei
kxi (t) − x̄(t)k,
(8)
i=1
where the equality follows from the definition of x(t) in (1)
and x̄(t) = (x̄1 (t), . . . , x̄n (t)).
The second term in (7) can be further expanded as
h∇ft (x̄(t)), x̄(t) − yi
= hu(t), x̄(t) − yi + h∇ft (x̄(t)) − u(t), x̄(t) − yi.
(9)
Now, from relation (6) we obtain
)
( t
X
1
ψ(x) .
hu(s), xi +
x̄(t + 1) = arg min
α(t)
x∈Xn
s=0
Therefore, by [36, Lemma 3], we can write
(10)
For the second term on the right-hand side of (9), we have
h∇ft (x̄(t)) − u(t), x̄(t) − yi
≤ kx̄(t) − ykk∇ft (x̄(t)) − u(t)k
√
≤ nDX k∇ft (x̄(t)) − u(t)k.
xi (t + 1) =
1 k
δ ui (t) + zik (t)
ri i
n
X
k
k
+
Mij vj→i
(t) − vi→j
(t) , k ∈ [n] (13a)
j=1
ψ
ΠXn (zi (t
+ 1), α(t)) ,
Combining the estimates in Eqs. (7)-(11) and taking the
supremum over all y ∈ Xn , we get the desired result.
Theorem 1 indicates that the regret will be small provided that
(E1) The squared norms ku(t)k2 remain bounded.
(E2) The agents’ primal variables xi (t) do not drift too much
from the centralized vector x̄(t).
(E3) The vectors u(t) stay close to the gradients ∇ft (x̄(t)).
This theorem plays an important role in the sequel, since it
provides guidelines for designing the update rule ui (t) and the
k
mappings Fi,t
(·) and Gi,t (·). We will also see later that the
centralized vector x̄(t) represents a “mean field” of the primal
iterates xi (t) for i ∈ V at time t.
(14)
and feeds this signal back into the dynamics (13a). Note,
however, that the execution of the algorithm will not change if
the agents never directly learn the full function ft , nor even the
full gradient ∇ft (x(t)), but instead receive the local gradient
signal ∇ft (xi (t)). The messages vi→j (t) take the form
k
vi→j
(t) = zik (t)
(11)
(13b)
1
n
where (vj→i
(t), . . . , vj→i
(t)) ∈ Rn represents a vector of
messages transmitted by agent j to agent i, provided that
j ∈ Ni . Since i ∈ Ni , we may include the previous dual iterate
k
zi (t) and the outgoing messages vi→j
(t) in mi (t). The dual
update rule (13a) is inspired by the state dynamics proposed
by Li and Marden [1], whereas the primal update rule (13b)
is exactly what one has in Nesterov’s scheme [33].
To complete the description of the algorithm, we must speck
ify the update policies {ui (t)} and the messages {vi→j
(t)}.
We assume that all agents receive a complete description of
ft . Agent i then computes
ui (t) = h∇ft (xi (t)), ei i, i ∈ [n], t ≥ 0.
T
T
X
1X
ψ(y)
hu(t), x̄(t) − yi ≤
α(t − 1)ku(t)k2 +
.
2 t=1
α(T )
t=1
(12)
Then, ODA-C uses the following instantiation of the update
rules in (3a)-(3b):
i=1
n
X
i, j ∈ V.
(15)
for all t and all agents i, j ∈ V with j ∈ Ni .
B. Regret of ODA-C with local gradient signals
Let z̄(t) = (z̄ 1 (t), . . . , z̄ n (t)). Our regret analysis rests on
the following simple but important fact:
Lemma 1: The weighted sum
z̄(t) ,
n
X
ri zi (t)
i=1
evolves according to the linear dynamics
z̄(t + 1) = z̄(t) + u(t),
where u(t) = u1 (t), . . . , un (t) .
(16)
5
Remark We observe that the relation in (16) holds regardless
k
k
of the choices of decisions vj→i
(t) and vi→j
(t). Moreover,
we point out that if u(t) = ∇ft (x(t)), then the combination
of (16) and (13b) will reduce to a centralized online variant
of Nesterov’s scheme [37].
Proof: Let V k (t) denote the n × n matrix with entries
k
k
k
[V (t)]ij = vj→i
(t) − vi→j
(t). Then
z̄ k (t + 1) =
n
X
≤ nL2 .
It remains to estimate term (E3) in Theorem 1. To that end,
we write
k∇ft (x̄(t)) − u(t)k
n
X
=
h∇ft (x̄(t)) − ∇ft (xi (t)), ei iei
i=1
ri zik (t + 1)
≤
i=1
=
n
X
ri
i=1
zik (t)
n
X
1
Mij [V k (t)]ij
+ δik ui (t) +
ri
j=1
= z̄ k (t) + uk (t) + tr[M̃ V k (t)],
where M̃ is an n × n matrix with entries M̃ij = ri Mij .
Since M̃ is a symmetric matrix, by (12), and V k (t) is skewsymmetric, tr[M̃ V k (t)] = 0, so we obtain (16).
Lemma 1 indicates that the vector z̄(t) can be seen as a “mean
field” of the local dual iterates zi (t) for i ∈ V at time t. Also,
if we define
x̄(t + 1) , Πψ
Xn (z̄(t + 1), α(t)),
then from relation (16) we have
x̄(t + 1) =
Πψ
Xn
t
X
!
u(s), α(t) ,
s=0
which coincides with relation (6) in Theorem 1. This allows
us to make use of Theorem 1 in analyzing the regret of this
algorithm. Furthermore, the definition of x̄(t) and relation (14)
indicate that u(t) will stay close to the centralized gradient
∇ft (x̄(t)), and as a consequence, the errors (E1) and (E3) in
Theorem 1 will remain small.
We now particularize the bound in Theorem 1 to this
scenario under the following additional assumption:
Assumption 2: All functions f ∈ F are differentiable and
have Lipschitz continuous gradients with constant G:
k∇f (x) − ∇f (y)k ≤ Gkx − yk,
∀f ∈ F; x, y ∈ Xn .
Theorem 2: Under Assumptions 1–2, the regret of any
algorithm of the form (13a)-(13b), and with u(t) computed
according to (14), can be upper-bounded as follows:
√
nGDX
T
X
t=1
α(t − 1)
n
X
n
X
2
|h∇ft (xi (t)), ei i|
i=1
≤
n
X
i=1
k∇ft (xi (t))k2
≤G
n
X
kx̄(t) − xi (t)k,
i=1
where we have exploited the fact that the gradients of all f ∈
F are G-Lipschitz.
Now, by construction,
kx̄(t) − xi (t)k
ψ
= Πψ
Xn (z̄(t), α(t − 1)) − ΠXn (zi (t), α(t − 1))
≤ α(t − 1)kz̄(t) − zi (t)k,
where the last step follows from the fact that the map
z 7→ Πψ
Xn (z, α) is α-Lipschitz (see, e.g., [33, Lemma 1]).
Substituting these estimates into the bound in Theorem 1, we
get the result.
This bound indicates that, if the network-wide disagreement
term behaves nicely, the regret R(T ) will be sublinear in T
with a proper choice of the step size α(t). We illustrate this
more specifically in the following corollary.
Corollary 1: Suppose that the policies for computing
k
{ui (t)} and {vi→j
(t)} are such that, for all t and for any
sequence f1 , . . . , fT ∈ F,
n
X
kzi (t) − z̄(t)k ≤ K
i=1
for some finite constant K > 0 (which may depend on n
and on other problem parameters). Then, the regret of the
algorithm (13a)-(13b) is bounded by
2
T
X
√
nL
C
R(T ) ≤
+ K L + nGDX
α(t − 1) +
.
2
α(T
)
t=1
kzi (t) − z̄(t)k.
i=1
Proof: The terms on the right-hand side of the bound in
Theorem 1 can be further estimated as follows. Since each
ft ∈ F is L-Lipschitz,
ku(t)k2 =
k∇ft (x̄(t)) − ∇ft (xi (t))k
i=1
1
In particular, if we choose α(t) = √t+1
for t ≥ 0, then the
√
regret is of the order O( T ):
√
√
√
R(T ) ≤ nL2 + 2K L + nGDX
T + C T + 1.
T
nL2 X
C
R(T ) ≤
α(t − 1) +
2 t=1
α(T )
+ L+
n
X
C. Full regret analysis
We now show that the network-wide disagreement term is
indeed upper-bounded by some constant. We recall that Mij 6=
0 only if j ∈ Ni . In addition to this, we posit the following
assumptions on the pair (r, M ).
Assumption 3: The positive weights r1 , . . . , rn sum to one:
n
X
i=1
ri = 1 and ri > 0 for each i ∈ [n].
6
The matrix M is row-stochastic, i.e.,
n
X
≤
Therefore,
Mij = 1 for each i ∈ [n].
2
j=1
The conditions we have imposed on the pair (r, M ) are
equivalent to saying that M is the transition probability matrix
of a reversible random walk on G with invariant distribution
r = (r1 , . . . , rn ) [38]. Let
zk (t) = (z1k (t), . . . , znk (t)), k ∈ [n], t ≥ 0,
(17)
and
ri . We state the following bound for
Pn r∗ , min1≤i≤n
2
kz
(t)
−
z̄(t)k
:
i
i=1
Lemma 2: Under Assumptions 1 and 3, for the policy in
(14)-(15) we have
n
X
nL2
√
kzi (t) − z̄(t)k2 ≤ 3
r∗ (1 − 1 − λ)2
i=1
i=1
is the r-weighted `2 -norm of the vector f ∈ Rn , and where λ
denotes the spectral gap of M [38], i.e.,
kf k2r − kM f k2r
.
kf k2r
, hr,f i=0
inf
n
Proof: From the definitions of zi (t), z̄(t), and zk (t), we
have
n
X
kzi (t) − z̄(t)k2 =
i=1
n
X
kzk (t) − z̄ k (t)1k2 .
(18)
k=1
Thus, we upper-bound the quantity on the right-hand side.
From (15), we can rewrite the dynamics (13a) as follows:
zk (t + 1) = M zk (t) +
1
uk (t)ek ,
rk
(19)
where zk (t) is defined in (17). By unrolling the dynamics (19)
and (16) from time 0 to t and recalling that zi (0) = 0 for all
i, we obtain:
t−1
1 X t−s−1
z (t) =
M
uk (s)ek .
rk s=0
k
(20)
Moreover, by the definition of z̄(t) in Eq. (16), we have
z̄ k (t) =
t−1
1 X
rk uk (s).
rk s=0
(21)
t−1
1 X
M t−s−1 ek − hr, ek i1 |uk (s)|.
rk s=0
(22)
By the properties of Markov matrices [38], for any f ∈ Rn ,
M t f − hr, f i1
2
≤
1
M t f − hr, f i1
r∗
2
r
k
2
kz (t) − z̄ (t)1k ≤
!2
t−1
t−s−1
1 X
(1 − λ) 2 |uk (s)|
3/2
r∗
s=0
2
which proves the stated result.
Lemma 2 captures the effect of the underlying network topology via the spectral gap λ (also known as the Fiedler value),
which captures the algebraic connectivity of the network.
Since G is assumed to be connected, λ > 0.
By combining Theorem 2 and Lemma 2, we can now
provide a regret bound for ODA-C:
Theorem 3: Let Assumptions 1–3 hold. With the choice
1
α(t) = √t+1
for all t ≥ 0, and under the policy (14)-(15), the
distributed algorithm ODA-C achieves the following regret:
!
√
√
2
nGDX
2
R(T ) ≤ nL 1 + 3/2
1+
T
√
L
r∗ (1 − 1 − λ)
√
+ C T + 1,
Proof: By Lemma 2, the averaging policy (15) satisfies
n
X
nL2
√
.
kzi (t) − z̄(t)k2 ≤ 3
r∗ (1 − 1 − λ)2
i=1
Hence, by Jensen’s inequality,
v
u n
n
X
u X
kzi (t) − z̄(t)k ≤ tn
kzi (t) − z̄(t)k2
i=1
i=1
Note that rk = hr, ek i. From (20) and (21), we have
kzk (t) − z̄ k (t)1k ≤
k
(23)
L
√
,
r∗3 (1 − 1 − λ)2
where Assumption 1 is used in the last inequality. From this
and relation (18), we obtain
n
X
nL2
√
kzi (t) − z̄(t)k2 ≤ 3
,
r∗ (1 − 1 − λ)2
i=1
v
u n
uX
kf kr , t
ri fi2
f ∈R
M t−s−1 ek − hr, ek i1
(1 − λ)t−s−1
2
≤
kek − hr, ek i1kr
r∗
(1 − λ)t−s−1
=
rk (1 − rk )
r∗
(1 − λ)t−s−1
.
≤
r∗
From relations (22) and (23), we obtain
≤
for every t ≥ 1, where
λ=
(1 − λ)t
kf − hr, f i1k2r .
r∗
nL
.
√
− 1 − λ)
Therefore, the conditions of Corollary 1 hold with
nL
K = 3/2
,
√
r∗ (1 − 1 − λ)
and the stated result follows.
This shows that, for any fixed communication network G
satisfying
Assumption 3, the worst-case regret is bounded by
√
O( T ). The constants also capture the dependence on the
algebraic connectivity of the network via the spectral gap λ,
as well as on the network size n.
≤
3/2
r∗ (1
7
V. ODA-PS AND ITS R EGRET B OUND
Also, we denote
We now introduce another decentralized online optimization
algorithm which uses the push-sum communication protocol
for its dual update rule (3a). We refer to this algorithm
as ODA-PS (Online Dual Averaing with Push-Sum based
communication). ODA-PS uses the network model (G2) for
its communication.
A(t − 1 : t) , I,
for all t ≥ 1.
B. Regret of ODA-PS with local gradient signals
For the regret analysis, we first study the dynamics of the
dual iterates zi (t) and its “mean field” z̄(t) in the following
lemma. We remind that z̄(t) = (z̄ 1 (t), . . . , z̄ n (t)) and
zk (t) = (z1k (t), . . . , znk (t)), k ∈ [n].
A. ODA-PS
For ODA-PS, each agent i maintains an additional scalar
sequence {wi (t)}∞
t=1 ⊂ R. Then, this algorithm particularizes
the update rule in (3a)-(3b) as
Lemma 3: Let zi (0) = 0 for all i ∈ V.
(a) The weighted sum
n
n
X
wi (t + 1) =
[A(t)]ij wj (t)
z̄(t) =
(24a)
j=1
zik (t + 1) = nδik ui (t) +
n
X
[A(t)]ij zjk (t), k ∈ [n]
evolves according to the linear dynamics
(24b)
z̄(t + 1) = z̄(t) + u(t),
j=1
xi (t + 1) = Πψ
Xn
zi (t + 1)
, α(t)
wi (t + 1)
(24c)
where the weight matrix A(t) is defined by the out-degrees of
the in-neighbors, i.e.,
1/dj (t) whenever j ∈ Niin (t)
(25)
[A(t)]ij =
0
otherwise.
The matrix A(t) is column stochastic by construction.
Note that the above update rules are based on a simple
broadcast communication. Each agent i broadcasts (or pushes)
the quantities wi (t)/di (t) and zi (t)/di (t) to all of the nodes in
its out-neighborhood Niout (t). Then, in (24a)-(24b) each agent
simply sums all the received messages to obtain wi (t + 1)
and zi (t + 1). The update rule (24c) can be executed locally.
Unlike ODA-C, the averaging matrix A(t) in ODA-PS does
not require symmetry due to this broadcast-based nature of the
push-sum protocol. However, the asymmetry requires uniformity of the positive weights ri across all agents (cf. Eq. (3a)).
Here we simply use ri = 1/n.
To complete the description of the algorithm, we must
specify the update policies {ui (t)}. As in ODA-C, we assume
that the signal agent i gets from the environment at time t is
simply the i-th coordinate of the gradient of ft at the agents
primal variable xi (t). Thus, we define:
ui (t) = h∇ft (xi (t)), ei i, i ∈ [n], t ≥ 0,
(26)
i.e., the update performed by agent i at time t is the simply
the i-th coordinate of the gradient of ft at the agent’s primal
variable xi (t).
We assume that each agent i initializes its updates with
wi (0) = 1 and zi (0) = 0, while ui (0) can be any arbitrary
value in X. We also recall that the local action of agent i at
time t is given by the ith coordinate of xi (t), i.e.,
i
x (t) =
xii (t).
For notational convenience, let us denote the products of the
weight matrices A(t), . . . , A(s) by A(t : s), i.e.,
A(t : s) , A(t) · · · A(s) for all t ≥ s ≥ 0.
1X
zi (t)
n i=1
where u(t) = (u1 (t), . . . , un (t)).
(b) For any i, k ∈ [n], the iterates in (24b) evolve according
to the following dynamics
zik (t) = n
t−1
X
[A(t − 1 : s + 1)]ik uk (s).
s=0
Proof:
(a) From relation (24b), we have for all k ∈ [n]
n
1X k
z (t + 1)
n i=1 i
n
n
X
1 X k
=
nδi ui (t) +
[A(t)]ij zjk (t)
n i=1
j=1
z̄ k (t + 1) =
= uk (t) +
n
n
1X k X
[A(t)]ij
zj (t)
n j=1
i=1
= uk (t) + z̄ k (t),
where the last equality follows from the columnstochasticity of the matrix A(t). The desired result follows by stacking up the scalar relation above over k.
(b) By stacking up the equation (24b) over i, we have for all
t ≥ 1 and k ∈ [n]
zk (t + 1) = A(t)zk (t) + nuk (t)ek .
By unrolling this equation from time 0 to t, we obtain
zk (t) = A(t − 1 : 0)zk (0)
+n
=n
t−1
X
uk (s)A(t − 1 : s + 1)ek
s=0
t−1
X
uk (s)A(t − 1 : s + 1)ek ,
s=0
where the equalities follows from A(t − 1 : t) = I and
the initial condition zi (0) = 0 for all i ∈ V. We get the
desired result by taking the i-th component of this vector.
8
where we can always choose
Lemma 3 tells us that the vector z̄(t) acts as a “mean field”
of the dual iterates zi (t). Also, if we define
x̄(t + 1) , Πψ
Xn (z̄(t + 1), α(t)) ,
then from Lemma 3(a) we can see that
x̄(t + 1) ,
Πψ
Xn
t
X
!
If in addition each G(t) is regular, we may choose
√
β = 2 2, θ = (1 − 1/4n3 )1/B ,
or
β=
u(s), α(t) ,
s=1
which coincides with relation (6) in Theorem 1.
We now particularize the bound in Theorem 1 in this
scenario under the additional assumption on the Lipschitz
continuous gradients (Assumption 2 in Section IV).
Theorem 4: Under Assumptions 1-2, the regret of the algorithm (24a)-(24c) with the local update ui (t) of agent i
computed according to (26) can be upper-bounded as follows:
for all T ≥ 1,
θ = (1 − 1/nnB )1/B .
β = 4,
√
2,
θ = max σ2 (A(t)),
t≥0
whenever supt≥0 σ2 (A(t)) < 1.
(b) The quantity
γ = inf min [A(t : 0)1]i
t≥0
1≤i≤n
satisfies
1
.
nnB
Moreover, if the graphs G(t) are regular, we have γ = 1.
T
The
next lemma provides an upper-bound for
2 X
C
nL
2
Pn
zi (t)
α(t − 1) +
R(T ) ≤
2 t=1
α(T )
i=1 wi (t) − z̄(t) .
Lemma 5: Let the sequences {zi (t)} and {wi (t)} be genT
n
X
X
√
zi (t)
to the algorithm (24a)-(24b). Recall that
+ L + nGDX
− z̄(t) . erated according
α(t − 1)
Pn
wi (t)
z̄(t) = n1 i=1 zi (t). Then, we have for all t ≥ 1,
t=1
i=1
2
n
2
X
Proof: Since the definition of u(t) in ODA-PS (cf. Eq.
2βL
zi (t)
2
≤
n
−
z̄(t)
,
(26)) coincides with that in ODA-C (cf. Eq. (14)), we can
wi (t)
γθ(θ − 1)
i=1
reuse all the derivations in the proof of Theorem 2 except for
where the constants β, γ and θ are as defined in Lemma 4.
the network-wide disagreement term:
Proof: From the definitions of zi (t), z̄(t) and zk (t), we
kx̄(t) − xi (t)k
have
2
zi (t)
n
n X
n k
2
ψ
ψ
X
X
zi (t)
zi (t)
= ΠXn (z̄(t), α(t − 1)) − ΠXn
, α(t − 1)
k
− z̄(t) =
− z̄ (t) . (28)
wi (t)
wi (t)
wi (t)
i=1
i=1
k=1
zi (t)
≤ α(t − 1) z̄(t) −
,
(27)
Thus, we can upper-bound the quantity on the right-hand side.
wi (t)
By inspecting equation (24a), it is easy to see that for any
where the last inequality follows from the α-Lipschitzian
i ∈ V and t ≥ 1, we have
ψ
property of the map z 7→ ΠXn (z, α) [33, Lemma 1].
n
n
X
X
This bound tells us that the regret R(T ) will be sublinear in T
wi (t) =
[A(t − 1 : 0)]i` wi (0) =
[A(t − 1 : 0)]i` .
with proper choice of the step size α(t) if the network-wide
`=1
`=1
disagreement term behaves nicely. Note that we can also make
From this and Lemma 3, we have the following chain of
use of Corollary 1 here if we can show
relations:
n
X
zi (t)
zik (t)
≤ K,
z̄(t) −
− z̄ k (t)
w
(t)
i
i=1
wi (t)
Pt−1
t−1
for some constant K > 0.
n s=0 [A(t − 1 : s + 1)]ik uk (s) X
Pn
=
−
uk (s)
`=1 [A(t − 1 : 0)]i`
s=0
Pn
Pn
C. Full regret analysis
t−1
X
: s + 1)]ik − `=1 [A(t − 1 : 0)]i`
`=1 [A(t − 1P
=
uk (s)
n
We now show that the network-wide disagreement term in
`=1 [A(t − 1 : 0)]i`
s=0
Theorem 4 is indeed upper-bounded by some constant. For
P
t−1
n
X
− 1 : s + 1)]ik − φi (t − 1))
doing this, we first restate a lemma from [16].
`=1 ([A(t
Pn
≤
uk (s)
Lemma 4: Let the graph sequence {G(t)} be B-strongly
`=1 [A(t − 1 : 0)]i`
s=0
!
connected. Then the following statements are valid.
Pn
(φ
(t
−
1)
−
[A(t
−
1
:
0)]
)
i
i`
n
(a) There is a sequence {φ(t)} ⊆ R of stochastic vectors
+ `=1 Pn
`=1 [A(t − 1 : 0)]i`
such that the matrix difference A(t : s)−φ(t)10 for t ≥ s
t−1
decays geometrically, i.e., for all i, j ∈ [n].
X
βθt−s−2 + βθt−1
≤
u
(s)
,
(29)
k
γ
|[A(t : s)]ij − φi (t)| ≤ βθt−s for all t ≥ s ≥ 0,
s=0
γ≥
9
where the inequalities follow from adding and subtracting
φi (t − 1) and from Lemma 4. From relation (26), we have
|uk (s)|2 = |h∇fs (xk (s)), ek i|2 ≤ k∇fs (xk (s))k2 ≤ L2 .
Combining this and the fact that βθt−s−2 ≥ βθt−1 for all
s = 0, . . . , t − 1, we further have
1
2
4
5
3
1
2
4
5
3
1
2
4
5
3
Fig. 2. Time-varying communication topology changing in cycle of three
used for ODA-PS
t−1
X
2βθt−s−2
2βL
zik (t)
− z̄ k (t) ≤
|uk (s)|
≤
.
wi (t)
γ
γθ(θ
− 1)
s=0
Substituting this estimate in relation (28), we get the desired
result.
By combining Theorem 4 and Lemma 5, we can now provide
the regret bound of ODA-PS:
Theorem 5: Let Assumptions 1–2 hold. With the choice
1
α(t) = √t+1
for all t ≥ 0, and under the policy (26), the
distributed algorithm ODA-PS achieves the following regret:
√
√
√
nGDX
4β n
2
R(T ) ≤ nL 1 + 1 +
T
L
γθ(1 − θ)
√
+ C T + 1,
where the constants β, γ and θ are as defined in Lemma 4.
Proof: By Jensen’s inequality, we have
v
u n
n
2
X
u X zi (t)
zi (t)
− z̄(t) ≤ tn
− z̄(t) .
wi (t)
wi (t)
i=1
i=1
Hence, using Lemma 5, we can estimate the network-wide
disagreement term as follows:
s
2
n
X
zi (t)
2βL
− z̄(t) ≤ n3
wi (t)
γθ(θ − 1)
i=1
=
2βn3/2 L
.
γθ(θ − 1)
Thus, the conditions of Corollary 1 with this modified
network-wide agreement hold with
√
K=n n
2βL
.
γθ(θ − 1)
and the stated result follows.
The bound shows that, for any time-varying sequence of Bstrongly connected
√ digraphs, the worst-case regret of ODA-PS
is of order O( T ). The constants also capture the dependence
on the properties of the underlying network, i.e., the number
of nodes n and as well as the connectivity period B.
The sensors are assumed to have a linear model of r(x) =
Ax, where A ∈ Rm×p and m < p.2 At each time t, each
sensor i ∈ V estimates its portion xi (t) ∈ Rpi of the target
vector x ∈ Rp , and then takes a measurement qti ∈ Rmi , which
is corrupted by observation error and possibly by modeling
error. We assume all sources of errors can be represented as
an additive noise, i.e.,
qt = Ax(t) + ζt ,
Pn
where qt ∈ Rm with m = i=1 mi is a stacked vector of
all qti ’s and ζt ∼ N (0, P ), where P is the noise covariance
matrix.
The regret is computed with respect to the least-squares
estimate of the target locations at time T , i.e.,
x̂ = arg min
x∈Xp
T
X
ft (x),
t=1
where ft (x) = 12 kAx − qt k2 . and we set X ∈ [−20, 20].
For ODA-C, we experiment with a n = 5 node cycle graph
whose communication topology is given as:
1↔2↔3↔4↔5↔1
We set ri = 1/5, Mii = 1/2 for all i, and Mij = 1/4 if i ↔ j.
For ODA-PS, we experiment with a time-varying sequence of
digraphs with n = 5 nodes whose communication topology
is changing periodically with period 3. The graph sequence
is, therefore, 3-strongly connected. In Figure 2, we depict
the repetition of the 3 corresponding graphs. The averaging
matrices A(t) (cf. Eq. (25)) can be determined accordingly.
We ran our algorithms once for each T ∈ [1000]. That is,
for a given T , the iterates in the algorithms are updated from
1
t = 1 to t = T . We used step size α(t) = √t+1
for both
algorithms.
In Figure 3, we depict the average regret R(T )/T over time
T of the distributed sensing problem when ODA-C and ODAPS are used, respectively. It shows that the regret is sublinear
for both algorithms and the average R(T )/T goes to zero as
the time increases.
VI. S IMULATION R ESULTS
Consider the problem of estimating some target vector
x ∈ Rp using measurements from a network of n sensors.
Each sensor i is in charge of estimating
a subvector xi ∈ Rpi
Pn
of x, where pi p and p =
i=1 pi is some very large
number. An example includes the localization of multiple
targets, where in this case x ∈ Rp becomes a stacked vector
of all target locations. When there are a number of spatially
dispersed targets, we can certainly benefit from distributed
sensing.
VII. C ONCLUSION
We have studied an online optimization problem in a
multiagent network. We proposed two decentralized variants
of Nesterov’s primal-dual algorithm, namely, ODA-C using
circulation-based dynamics for time-invariant networks and
2 Although target localization is usually formulated as a nonlinear estimation problem [39], for considerations of simplicity one often employs a linearized model using a first-order Taylor expansion around the measurements;
see, e.g., [40], [41].
10
3000
2500
2500
2000
R(T )/T
R(T )/T
2000
1500
1500
1000
1000
500
500
0
0
250
500
750
Iteration T
1000
0
0
250
500
750
Iteration T
1000
Fig. 3. The Average Regret R(T )/T vs. Iterations for Online Distributed
Active Sensing using ODA-C (left) and ODA-PS (right)
ODA-PS using broadcast-based push-sum protocol for timevarying networks. We have established a generic regret bound
and provided its refinements for certain information
exchange
√
T
)
when
the
policies. The regret is shown
to
grow
as
O(
√
step size is α(t) = 1/ t + 1. For ODA-C, the bound is
valid for a static connectivity graph and a row-stochastic
matrix of weights M = [Mij ] which is reversible with respect
to a strictly positive probability vector r. For ODA-PS, the
bound is valid for a uniformly strongly connected sequence
of digraphs and column-stochastic matrices of weights A(t)
whose components are based on the out-degrees of neighbors.
Simulation results on a sensor network exhibit the desired
theoretical properties of the two algorithms.
R EFERENCES
[1] N. Li and J. R. Marden, “Designing games for distributed optimization,”
IEEE J. Sel. Topics Signal Proc., vol. 7, pp. 230–242, 2013.
[2] D. Kempe, A. Dobra, and J. Gehrke, “Gossip-based computation of
aggregate information,” in 44th Annual IEEE Symposium on Foundations
of Computer Science, vol. 44, 2003, pp. 482–491.
[3] F. Bullo, J. Cortés, and S. Martı́nez, Distributed Control of Robotic
Networks. Applied Mathematics Series. Princeton University Press,
2009.
[4] M. Mesbahi and M. Egerstedt, Graph Theoretic Methods for Multiagent
Networks. Princeton, NJ, USA: Princeton University Press, 2010.
[5] S. Kar and J. Moura, “Distributed consensus algorithms in sensor
networks: Quantized data and random link failures,” IEEE Trans. Signal
Process., vol. 58, pp. 1383 –1400, 2010.
[6] A. Martinoli, F. Mondada, G. Mermoud, N. Correll, M. Egerstedt,
A. Hsieh, L. Parker, and K. Stoy, Distributed Autonomous Robotic
Systems. Springer Tracts in Advanced Robotics, Springer-Verlag, 2013.
[7] B. Zhang, A. Lam, A. Dominguez-Garcia, and D. Tse, “Optimal
distributed voltage regulation in power distribution networks,” 2015,
http://arxiv.org/abs/1204.5226.
[8] T.-H. Chang, A. Nedić, and A. Scaglione, “Distributed constrained optimization by consensus-based primal-dual perturbation method,” IEEE
Transactions on Automatic Control, vol. 59, pp. 1524–1538, 2014.
[9] A. Nedić and A. Ozdaglar, “On the rate of convergence of distributed
subgradient methods for multi-agent optimization,” in Proceedings of
IEEE CDC, 2007, pp. 4711–4716.
[10] B. Johansson, M. Rabi, and M. Johansson, “A simple peer-to-peer
algorithm for distributed optimization in sensor networks,” in 46th IEEE
Conference on Decision and Control, 2007, pp. 4705 –4710.
[11] A. Nedić and A. Ozdaglar, “Distributed subgradient methods for multiagent optimization,” IEEE Transactions on Automatic Control, vol. 54,
pp. 48–61, 2009.
[12] S. S. Ram, A. Nedić, and V. V. Veeravalli, “Distributed stochastic
subgradient projection algorithms for convex optimization,” Journal of
Optimization Theory and Applications, vol. 147, pp. 516–545, 2010.
[13] K. Srivastava and A. Nedić, “Distributed asynchronous constrained
stochastic optimization,” IEEE Journal of Selected Topics in Signal
Processing, vol. 5, pp. 772–790, 2011.
[14] K. Tsianos, S. Lawlor, and M. Rabbat, “Consensus-based distributed
optimization: Practical issues and applications in large-scale machine
learning,” in 50th Allerton Conference on Communication, Control, and
Computing, 2012, pp. 1543–1550.
[15] ——, “Push-sum distributed dual averaging for convex optimization,”
in 51st Annual Conference on Decision and Control, 2012, pp. 5453 –
5458.
[16] A. Nedić and A. Olshevsky, “Distributed optimization over time-varying
directed graphs,” IEEE Transactions on Automatic Control, vol. 60, pp.
601–615, 2015.
[17] ——, “Stochastic gradient-push for strongly convex functions on timevarying directed graphs,” 2014, http://arxiv.org/abs/1406.2075.
[18] E. Wei and A. Ozdaglar, “Distributed alternating direction method of
multipliers,” in 51st IEEE Conference on Decision and Control and
European Control Conference, 2012, pp. 5445–5450.
[19] ——, “On the O(1/k) convergence of asynchronous distributed alternating direction method of multipliers,” in IEEE Global Conference on
Signal and Information Processing, 2013, pp. 551–554.
[20] D. Jakovetic, J. Xavier, and J. Moura, “Cooperative convex optimization in networked systems: Augmented lagrangian algorithms with directed gossip communication,” IEEE Transactions on Signal Processing,
vol. 59, pp. 3889–3902, 2011.
[21] D. Jakovetic, J. Xavier, and J. M. F. Moura, “Fast distributed gradient
methods,” IEEE Transactions on Automatic Control, vol. 59, no. 5, pp.
1131–1146, May 2014.
[22] Q. Ling and A. Ribeiro, “Decentralized dynamic optimization through
the alternating direction method of multiplier,” IEEE Transactions on
Signal Processing, vol. 62, pp. 1185–1197, 2014.
[23] B. Gharesifard and J. Cortés, “Distributed continuous-time convex optimization on weight-balanced digraphs,” IEEE Transactions on Automatic
Control, vol. 59, pp. 781–786, 2014.
[24] D. Mateos-Nuñez and J. Cortés, “Distributed online convex optimization
over jointly connected digraphs,” IEEE Transactions on Network Science
and Engineering, vol. 1, pp. 23–37, 2014.
[25] P. D. Lorenzo and G. Scutari, “Next: In-network nonconvex optimization,” 2016, http://arxiv.org/abs/1602.00591.
[26] J. Tsitsiklis, “Problems in decentralized decision making and computation,” Ph.D. dissertation, Dept. of Electrical Engineering and Computer
Science, Massachusetts Institute of Technology, 1984.
[27] J. Tsitsiklis, D. Bertsekas, and M. Athans, “Distributed asynchronous
deterministic and stochastic gradient optimization algorithms,” IEEE
Transactions on Automatic Control, vol. 31, pp. 803–812, 1986.
[28] J. N. Tsitsiklis and M. Athans, “Convergence and asymptotic agreement
in distributed decision problems,” IEEE Transactions on Automatic
Control, vol. 29, pp. 42–50, 1984.
[29] S. Li and T. Basar, “Distributed learning algorithms for the computation
of noncooperative equilibria,” Automatica, vol. 23, pp. 523–533, 1987.
[30] M. Raginsky, N. Kiarashi, and R. Willett, “Decentralized online convex
programming with local information,” in Proceedings of the American
Control Conference, 2011, pp. 5363–5369.
[31] D. Bertsekas and J. Tsitsiklis, Parallel and Distributed Computation:
Numerical Methods. Belmont, MA: Athena Scientific, 1997.
[32] F. Benezit, V. Blondel, P. Thiran, J. Tsitsiklis, and M. Vetterli, “Weighted
gossip: Distributed averaging using non-doubly stochastic matrices,”
in IEEE International Symposium on Information Theory Proceedings
(ISIT), 2010, pp. 1753 – 1757.
[33] Y. Nesterov, “Primal-dual subgradient methods for convex problems,”
Math. Program., Ser. B, vol. 120, pp. 221–259, 2009.
[34] M. Akbari, B. Gharesifard, and T. Linder, “Distributed subgradient-push
online convex optimization on time-varying directed graphs,” in 52nd
Allerton Conference on Communication, Control, and Computing, 2014,
pp. 264–269.
[35] S. Hosseini, A. Chapman, and M. Mesbahi, “Online distributed optimization on dynamic networks,” 2014, http://arxiv.org/abs/1412.7215.
[36] J. C. Duchi, A. Agarwal, and M. J. Wainwright, “Dual averaging for
distributed optimization: convergence analysis and network scaling,”
IEEE Transactions on Automatic Control, vol. 57, pp. 592–606, 2012.
[37] L. Xiao, “Dual averaging methods for regularized stochastic learning and
online optimization,” J. Machine Learning Res., vol. 11, pp. 2543–2596,
2010.
[38] D. A. Levin, Y. Peres, and E. L. Wilmer, Markov Chains and Mixing
Times. Amer. Math. Soc., 2008.
[39] P. Stoica and J. Li, “Source localization from range-difference measurements,” IEEE Signal Processing Mag., pp. 63–66, November 2006.
[40] L. Kleeman and R. Kuc, “Mobile robot sonar for target localization
and classification,” International Journal of Robotics Research, vol. 14,
no. 4, pp. 295–318, Aug 1995.
[41] S. S. Ponda, “Trajectory Optimization for Target Localization Using
Small Unmanned Aerial Vehicles,” Master’s thesis, Massachusetts Institute of Technology, 2008.
| 3cs.SY
|
arXiv:1707.02151v2 [math.GR] 15 Nov 2017
POINCARÉ PROFILES OF GROUPS AND SPACES
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
Abstract. We introduce a spectrum of monotone coarse invariants for metric measure spaces called Poincaré profiles. The two
extremes of this spectrum determine the growth of the space, and
the separation profile as defined by Benjamini–Schramm–Timár.
In this paper we focus on properties of the Poincaré profiles of
groups with polynomial growth, and of hyperbolic spaces, where
we deduce a striking connection between these profiles and conformal dimension. One application of our results is that there is
a collection of hyperbolic Coxeter groups, indexed by a countable
dense subset of (1, ∞), such that Gs does not coarsely embed into
Gt whenever s < t.
Contents
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
Introduction
Notation and framework
Poincaré constants
Poincaré profiles for metric measure spaces
Regular maps and large scale equivalence
Extremal profiles: growth and separation
Dependency on p
Poincaré profiles of groups with polynomial growth
Upper bounds and large-scale dimension
Trees
Lower bounds for hyperbolic spaces with boundary
Poincaré inequalities
12. Upper bounds for hyperbolic spaces with hyperplanes
13. Applications to buildings and symmetric spaces
References
2
10
12
16
18
23
27
28
34
36
38
45
52
53
Date: November 16, 2017.
The first and third authors were supported by the grant ANR-14-CE25-0004
“GAMME”. The first and second authors were supported in part by the NSF grant
DMS-1440140 while in residence at the Mathematical Sciences Research Institute
in Berkeley, California, during the Fall 2016 semester. The second author was also
supported in part by EPSRC grant EP/P010245/1.
1
2
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
1. Introduction
A monotone coarse invariant of a collection of metric spaces X is
a function Λ from X to a partially ordered set (P, ≤) with the property that Λ(X) ≤ Λ(Y ) whenever there is a coarse embedding of X
into Y . The asymptotic dimension and the growth function are natural and well-studied examples of such invariants, and a more recent
example is the separation profile of bounded degree graphs introduced
by Benjamini–Schramm–Timár [BST12]. In this paper, we introduce a
new family of monotone coarse invariants called the Lp -Poincaré profiles.
We will only define the Poincaré profiles of graphs in the introduction; however, our results naturally extend to compactly generated
locally compact groups and Riemannian manifolds with bounded geometry. The majority of the paper is presented in a more general
context which includes all of these spaces.
Inspired by work of the first author [Hum17], which gives an equivalent definition of the separation profile in terms of the Cheeger constant,
for each p ∈ [1, ∞] we define the p-Poincaré constant of a finite graph
Γ with vertex set V Γ and edge set EΓ to be
)
(
||∇f
||
p
: f ∈ Map(V Γ → R), f 6≡ fΓ
hp (Γ) = inf
||f − fΓ ||p
where ∇f (x) = max {|f (x) − f (y)| : xy ∈ EΓ}, ||·||p is the usual pP
norm in R|V Γ| and fΓ is the average |V Γ|−1 x∈V Γ f (x). It is worth
noting that for bounded degree graphs hp (Γ) is bi-Lipschitz equivalent
1
to λ1,p (Γ) p , where λ1,p (Γ) denotes the smallest non-zero eigenvalue of
the p-Laplacian on Γ (see Remark 3.8 below).
Now we define the Lp -Poincaré profile of an infinite graph X to
be
ΛpX (r) = sup {|Γ| hp (Γ) : Γ ≤ X, |V Γ| ≤ r} .
We consider Poincaré profiles up to the natural order . where f . g
if there exists a constant C such that f (r) ≤ Cg(Cr + C) + C for all
r, and f ≃ g if f . g and g . f . Often, the constant C will depend
on p; to emphasise this we will use the notations .p and ≃p .
A lower bound on the Lp -Poincaré profile corresponds to a “p-Poincaré inequality” for functions on a finite subgraph of the corresponding size.1 Poincaré inequalities have been intensively studied, particularly in the case of balls in doubling metric spaces, see [SC02, HK00]
1Technically
these Poincaré inequalities are Neumann-type, rather than
Dirichlet-type Poincaré inequalities which consider only functions which are 0 on
POINCARÉ PROFILES OF GROUPS AND SPACES
3
and references therein. For finite graphs, there is a vast literature
linking Cheeger constants and spectral gaps to such inequalities when
p = 1, 2, see [Chu97, SC97]. Discrete Poincaré inequalities on balls in
metric spaces have been studied before by, for example, Holopainen–
Soardi [HS97] and Gill–Lopez [GL15]. Our approach differs in that we
are working in a situation where global Poincaré inequalities do not
necessarily hold, where measures need not be doubling, and where we
have to consider inequalities on all subsets, not just balls.
Our first important result is that these Poincaré profiles are monotone coarse invariants.
Theorem 1. Let X, Y be graphs with bounded degree. If there is a
regular map r : V X → V Y , then for all p ∈ [1, ∞], ΛpX .p ΛpY .
A map r : V X → V Y is said to be regular if it is Lipschitz and
supy∈V Y |r −1 (y)| < ∞. In particular every quasi-isometric or coarse
embedding is regular. Thus for each p the Lp -Poincaré profile is a
well-defined coarse invariant of a finitely generated group G.
1.1. Extremal cases. In the cases p = ∞ and p = 1 the Poincaré
profile is easily understood in terms of the growth and separation profile
respectively.
Recall the growth function of a graph X: γX (k) is the maximum
number of vertices contained in a closed ball B(x, k) of radius k centred
at some vertex x ∈ V X. We define the inverse growth function:
κX (r) is the smallest positive k such that γX (k) > r.
At one extreme, p = ∞, the Poincaré profile detects inverse growth.
s
Proposition 2. For any bounded degree graph X, Λ∞
X (r) ≃ sup κ (s) .
X
1≤s≤r
From this, we may easily deduce Theorem 1 in the case p = ∞. At
the other extreme we show that the L1 -Poincaré profile is equivalent
to the separation profile, as introduced by Benjamini–Schramm–Timár
[BST12]. The perspective we adopt of studying Poincaré profiles up to
regular maps is inspired by their observation that separation is monotone under regular maps.
We recall that the separation profile of an infinite graph X may be
defined by sepX (r) = max {|Γ| h(Γ)} where the maximum is taken over
all subgraphs Γ of X with at most r vertices, and h(Γ) is the Cheeger
constant [Hum17].
Proposition 3. For any bounded degree graph X, Λ1X (r) ≃ sepX (r).
the boundary of the subgraph in the ambient space. See Remark 4.2 for more
details.
4
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
Remark 4. The case of p = 2 is also natural, being the largest spectral
gap among subgraphs of a given size. The spectral gap can be used
to bound mixing times of random walks on the subgraph. A related
spectral profile was considered by Goel–Montenegro–Tetali [GMT06].
1.2. Relating profiles. The following results are classical, and are
likely to be easy exercises for experts; for completeness we present full
proofs.
Proposition 5. Let 1 ≤ p ≤ q < ∞. There exists a constant C =
C(p, q) such that for every bounded degree graph X and every r we have
ΛpX (r) ≤ CΛqX (r).
In the opposite direction we have the following.
Proposition 6. If Γ is a finite graph and p ∈ [1, ∞), then hp (Γ)p ≤
2p h1 (Γ).
Asymptotically this is sharp for balls in the 3-regular tree, as we
will see in section 10. Proposition 5 cannot be extended to the case
q = ∞ since there are bounded degree graphs containing expanders:
combining the above propositions with results in [Hum17] we see that
for every p ∈ [1, ∞), ΛpX (r)/r 6→ 0 as r → ∞ if and only if X contains
an expander, while a bounded degree graph Y has at most exponential
growth, so always satisfies Λ∞
Y (r) . r/ log(r).
1.3. Polynomial growth. Gromov’s celebrated polynomial growth
theorem asserts that every finitely generated group with polynomial
growth is virtually nilpotent. Results of Bass–Guivarc’h then show
that for every group G of polynomial growth there is an integer d such
that γG (r) ≃ r d [Gro81, Bas72, Gui73].
Theorem 7. Let G be a finitely generated group such that γG (r) ≃ r d .
d−1
Then for all p ∈ [1, ∞], ΛpG (r) ≃p r d .
When G is virtually abelian and p = 1 this follows from [BST12]; in
all other cases it is new. To prove the lower bound on ΛpG (r) we calculate a lower bound on the separation profile using a Poincaré inequality
and apply Propositions 3 and 5. For the upper bound we use a general
result, Proposition 9.5, which holds for any bounded degree graph with
finite Assouad–Nagata dimension (cf. [Hum17, Theorem 1.5]).
Recall that by a classical result of Heintze [Hei74], every simply
connected negatively curved homogeneous Riemannian manifold M is
isometric to a connected Lie group of the form N ⋊ R equipped with
a left-invariant Riemannian metric, where N is a simply connected
nilpotent Lie group and the action of R on N is contracting. We
POINCARÉ PROFILES OF GROUPS AND SPACES
5
immediately deduce from Theorem 7 that for every p ∈ [1, ∞], the
d−1
Lp -Poincaré profile of such a manifold is bounded from below by r d ,
where d is the homogeneous dimension of N. As a special case of this
we deduce the following lower bounds for rank one symmetric spaces.
Corollary 8. Let K ∈ {R, C, H, O} be a real division algebra, and let
X = Hm
K be a rank-one symmetric space for m ≥ 2 (and m = 2 when
K = O). Then, for all 1 ≤ p < ∞, we have ΛpX (r) &p r (Q−1)/Q where
Q = (m + 1) dimR K − 2.
It is worth noting at this point that Q is the conformal dimension of
the boundary of X. For large p this bound is far from optimal as we
will see in the next section.
1.4. Hyperbolic spaces. We begin by considering the case of an infinite 3-regular tree.
Theorem 9. Let T be the infinite 3-regular tree. Then ΛpT (r) ≃p r
for all p ∈ [1, ∞).
p−1
p
,
Note that when p = ∞, ΛpT (r) ≃ r/ log(r) by Proposition 6.1. Using
Theorem 9, together with results of Chou and Benjamini–Schramm on
embeddings of trees into elementary amenable groups with exponential
growth and non-amenable groups respectively [Cho80, BS97] we obtain
the following corollary.
Corollary 10. Let G be a finitely generated elementary amenable group
with exponential growth or a finitely generated infinite non-amenable
p−1
group. Then for all p ∈ [1, ∞), ΛpG (r) &p r p .
We continue with a striking result which suggests a connection between conformal dimension of boundaries and a phase transition in the
Poincaré profiles of hyperbolic groups.
Theorem 11. Let G be a finitely generated hyperbolic group with equivariant conformal dimension Q (see Definition 12.4). For every ε > 0
( Q−1
r Q +ǫ if p ≤ Q
p
ΛG (r) .
p−1
if p > Q.
r p
If the equivariant conformal dimension
Q−1
r Q
Q−1
1
p
ΛG (r) .
r Q log Q (r)
p−1
r p
is attained, we have:
if 1 ≤ p < Q
if p = Q
if p > Q.
6
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
In certain cases we are able to prove that these upper bounds are
sharp. Our key examples are rank-one symmetric spaces, and a collection of groups Gm,n = hs1 , . . . , sm | sni , [s1 , s2 ], . . . , [sm−1 , sm ], [sm , s1 ]i,
m ≥ 5, n ≥ 3 which occur as uniform lattices in the isometry groups
of associated Fuchsian buildings ∆m,n , as studied by Bourdon and
Bourdon–Pajot amongst others [Bou97, BP99]. Following the terminology of [Cap14] we call these Bourdon buildings. These groups are
virtually torsion free [HW99], and commensurable to hyperbolic Coxeter groups when n is even [Hag06].
Theorem 12. Let X = Hm
K be a rank-one symmetric space for K ∈
{R, C, H, O} and m ≥ 2 (with m = 2 when K = O), or let X be one of
the groups Gm,n ; in either case, let Q be the Ahlfors regular conformal
dimension of the boundary of X. Then
Q−1
if p < Q
r Q
Q−1
1
p
ΛX (r) ≃
r Q log(r) Q if p = Q
p−1
if p > Q
r p
, these sharp bounds for the separation profile
In the case of Λ1Hm
R
appear in [BST12]. It is interesting to note that uniform lattices G in
p−1
P SL(2, R) satisfy Λ1G (r) ≃ log(r) and ΛpG (r) ≃ r p for all p > 1, while
p−1
non-uniform lattices H satisfy Λ1H (r) ≃ r p for all p ≥ 1. We have no
other examples of this distinction between uniform and non-uniform
lattices for any other p or for any groups of higher rank.
We do not define the conformal dimension here, but comment that
for a rank-one symmetric space Hm
K we have Q = (m+1) dimR K−2 and
for the groups Gm,n we have Q = 1 + log(n − 1)/arccosh((m − 2)/m),
which takes a dense set of values in (1, ∞) as m, n vary.
The upper bound in Theorem 12 is obtained by constructing specific
functions on the boundary using an embedding of the space into a
real hyperbolic space. The lower bound in Theorem 12 for rank-one
symmetric spaces with p < Q follows from Corollary 8. For the groups
Gm,n , and for the sharp case p = Q, we require the following more
general result.
Theorem 13. Suppose that X is a visual Gromov hyperbolic graph
with a visual metric ρ on ∂∞ X that is Ahlfors Q-regular and admits a
p-Poincaré inequality. Then for all q ≥ p, ΛqX (r) & r (Q−1)/Q .
Moreover, if (∂∞ X, ρ) admits a Q-Poincaré inequality, then ΛQ
X (r) &
1−1/Q
1/Q
r
log(r) .
POINCARÉ PROFILES OF GROUPS AND SPACES
7
Here a “p-Poincaré inequality” is in the sense of Heinonen–Koskela
[HK98], namely an analytic property of the compact metric space ∂∞ X.
Such inequalities hold on boundaries of rank-one symmetric spaces, see
e.g. [Jer86, HK98, MT10], so we can apply this lower bound to obtain
an alternative proof of Corollary 8. For the groups Gm,n , the Poincaré
inequalities are constructed in [BP99].
The sharp lower bounds on ΛQ
X come from showing a suitable Poincaré
inequality on annuli B(o, 2r) \ B(o, r) in X. It is interesting to observe
that for p < Q, p = Q, and p > Q, the sharp lower bounds on Λp X are
realised by embedded spheres, annuli and trees respectively.
Finally, Theorems 9 and 12, together with the embedding theorem
of Bonk–Schramm [BS00], imply that for every hyperbolic group G
p−1
there is some p0 such that for all p > p0 , we have ΛpG (r) ≃ r p . The
relationship between the infimal such p0 and the conformal dimension
of the boundary of G is one of the most intriguing aspects of these
profiles.
1.5. Consequences. By applying Theorem 12 to the groups Gm,n , we
find a new collection of functions which can be obtained as separation
profiles of finitely generated groups:
Corollary 14. There exists a dense subset A of (0, 1) such that for all
α ∈ A there is a hyperbolic group Gα with sepGα (r) ≃ r α .
The key purpose of a monotone coarse invariant is to be able to distinguish situations in which one space cannot be coarsely embedded
into another. There are few general tools to do this; asymptotic dimension is one and growth (or equivalently, the L∞ -Poincaré profile) is
another. Here we present and discuss some results of this form which
cannot be obtained by studying growth and/or asymptotic dimension.
Corollary 15. If there is a coarse embedding of HkC into HlR , then
l > 2k. Likewise, if there is a coarse embedding of HkH into HlR , then
l > 4k + 2.
To prove the analogous result for quasi-isometric embeddings, one
can use the conformal dimension of the boundary, however, a coarse
embedding does not necessarily induce a well-defined map between
boundaries [BR13] so this approach cannot be expected to work. Using asymptotic dimension as an invariant one could only deduce that
l ≥ 2k in the first case and l ≥ 4k in the second.
By [BS00], every hyperbolic group quasi-isometrically embeds into
some HkR . A natural obstruction to a coarse embedding Gk → HkR
8
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
is that the asymptotic dimension of Gk is greater than k. Poincaré
profiles provide a different obstruction.
Corollary 16. For every k there is a hyperbolic group Gk of asymptotic
dimension 2 which does not coarsely embed into HkR .
We can take Gk to be Gm(k),5 for some appropriately chosen m(k)
and apply Theorem 12.
It is in general very difficult to prove a statement of the form “a
hyperbolic group H is not isomorphic to a subgroup of a hyperbolic
group G”. One may use torsion or asymptotic dimension in certain
cases, here we show that the Poincaré profiles can exclude subgroups
when the two methods listed above fail.
Corollary 17. There exists a collection of (torsion-free) hyperbolic
groups (Gq )q∈Q with asymptotic dimension 2 such that whenever i < j
there is no coarse embedding from Gi to Gj . In particular, Gi is not
virtually a subgroup of Gj .
Indeed, the groups Gm,n are virtually torsion-free so we may choose
the Gq in Corollary 17 to be torsion-free. By results of Gersten, finitely
presented subgroups of hyperbolic groups with cohomological dimension 2 (which equals the asymptotic dimension for torsion-free hyperbolic groups [BM91, BL07]) are hyperbolic but not necessarily quasiconvex. The fact that Gj is not a quasi-convex subgroup of Gi is
immediate by considering the conformal dimension.
Remark 18. By a recent result of Pansu [Pan16], if a hyperbolic
group H coarsely embeds into a hyperbolic group G, then the “Lp cohomological dimension” of H is less than or equal to the conformal
dimension of the boundary of G. In the cases of the Bourdon buildings and rank-one symmetric spaces, these two numbers turn out to
coincide. This provides an alternative proof of Corollaries 15, 16 and
17.
1.6. About the proofs. The proof of the theorems described in the
previous section employ a variety of techniques. In particular, the
arguments needed for getting upper bounds are completely different
than those for obtaining lower bounds.
1.6.1. Upper bounds. For groups with polynomial growth, our sharp
upper bounds are obtained via an argument adapted from [Hum17]
based on the fact that these groups have finite Assouad–Nagata dimension ([Hum17, Theorem 1.5] deals with the separation profile, corresponding to p = 1). Finite Assouad–Nagata dimension is a quantitative strengthening of finite asymptotic dimension. Contrary to the
POINCARÉ PROFILES OF GROUPS AND SPACES
9
latter, finite Assouad–Nagata dimension is not monotone under coarse
embedding (only quasi-isometric embedding) as it is sensitive to distortion of the metric. We come up with a new notion called finite measured
dimension (see Definition 9.1), weaker than finite asymptotic dimension, which should be of independent interest. Our motivation here
is that it is well adapted for providing upper bound on the Poincaré
profiles.
Obtaining upper bounds for hyperbolic groups is trickier, and occupies all of §12. We need three different arguments, depending whether
p lies below, above, or equals the conformal dimension. We show that
hyperbolic groups admit in some sense “many hyperplanes”, by using
a theorem of Bonk–Schramm [BS00] to embed the group into a real hyperbolic space, which has an abundance of hyperplanes. We crucially
use a version of Helly’s theorem for CAT(0) spaces. Our argument for
small p is largely inspired from [BST12] where the separation profile of
the real hyperbolic plane is computed. It consists in a symmetrization
argument. For large p, we construct for every finite set A, a p-Dirichlet
function whose restriction to A provides a good test function.
1.6.2. Lower bounds. Obtaining lower bounds for groups with polynomial growth follows from well-known Poincaré inequalities in balls. It is
interesting that the functional analytic interpretation of the separation
profile gives us new estimates for nilpotent groups without effort.
The lower bounds for hyperbolic Lie groups and small p are obtained by considering parabolic closed nilpotent subgroups. The cases
of Bourdon–Pajot buildings, and of the cases when p = Q, are more
interesting and more subtle. For p < Q, we exploit the fact that their
visual boundary satisfies (infinitesimal) Poincaré inequalities. We “pull
down” these inequalities on a sphere of large radius in the space using
a discretization argument developed in the first part of the paper. To
get the sharp lower bound in the p = Q case, we use the Poincaré
inequalities on spheres and a curve counting argument to find a new
Poincaré inequalities on annuli. A similar but simpler curve counting
argument gives the lower bound for the 3-regular tree, and hence all
spaces in which it embeds.
1.7. Structure of the paper. The paper splits roughly into three
parts. The first part introduces Poincaré profiles as monotone regular (in particular coarse) invariants. After introducing our notations
and fixing the class of metric measure spaces under consideration, we
present the more general definition of Poincaré constants in Section 3
and explain some basic properties. We then introduce Poincaré profiles
and prove Theorem 1 in Sections 4 and 5 respectively.
10
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
The second part deals with relationships between Poincaré profiles.
The descriptions of extremal profiles (Propositions 2 and 3) and the
connection with separation (Proposition 6) are proved in Section 6,
and the dependence on p (Proposition 5) is discussed in Section 7.
The final part is dedicated to calculating profiles using the technology
developed in the rest of the paper. Groups with polynomial growth
(Theorem 7) are considered in Sections 8 and 9. For hyperbolic spaces,
trees (Theorem 9), lower bounds (Theorem 13) and upper bounds are
in Sections 10, 11 and 12 respectively, with applications (in particular,
Theorem 12) discussed in Section 13.
1.8. Acknowledgements. We are grateful to Laurent Saloff-Coste
and Anne Thomas for comments on earlier versions of this paper, and
for pointing out a number of related references.
2. Notation and framework
We first introduce notation to be used throughout the paper.
Suppose f, g : S → [0, ∞) where S = N or S = [0, ∞). We write
f u,v,... g if there exists a constant C > 0 depending only on u, v, . . .
such that f (x) ≤ Cg(x) for all x ∈ S. If f u,v,... g and g u,v,... f
then we write f ≍u,v,... g. We drop the subscripts if the constants are
understood.
We write f .u,v,... g if there exists a constant C > 0 depending only
on u, v, . . . such that f (x) ≤ Cg(Cx + C) + C for all x ∈ S; similarly,
we write f ≃u,v,... g if f .u,v,... g and g .u,v,... f .
Given a subset A of a metric space (X, d) and some M ≥ 0 we define
the closed M-neighbourhood of A to be
[A]M = {x ∈ X : d(x, A) ≤ M} .
Given a point x ∈ X and r ≥ 0 we denote by B(x, r) the closed metric
ball of radius r centred at x.
Let (Z, ν) be a measure space with positive finite measure. We denote the averaged integral by
Z
Z
1
− f dν =
f dν.
ν(Z) Z
Z
Given a function f ∈ Lp (X, µ), another measure µ′ such that f ∈
Lp (X, µ′ ) and a measurable subset Z ⊆ X we write
Z
p1
p
′
||f ||p,µ′ =
|f (z)| dµ (z)
and
X
POINCARÉ PROFILES OF GROUPS AND SPACES
||f ||Z,p =
Z
p
|f (z)| dµ(z)
Z
p1
11
.
The L∞ norms ||·||∞,µ′ and ||·||Z,∞ are defined analogously.
Given a graph Γ = (V Γ, EΓ) and a subset A ⊂ V Γ, the full (or
induced) subgraph of Γ with vertex set A is the graph with vertex set
A and edge set {xy ∈ E : x, y ∈ A}.
The purpose of the remainder of this section is to introduce the class
of spaces we will consider in this paper.
Definition 2.1. A metric measure space is a triple (X, d, µ) where
µ is a non-trivial, locally finite, Borel measure on a complete, separable
metric space (X, d).
The key examples are: graphs of bounded degree, Riemannian manifolds with bounded geometry and compactly generated locally compact
groups, so we will make the following standing assumptions.
We will assume throughout the paper that any metric measure space
(X, d, µ) satisfies the following properties:
• X has bounded packing on large scales2: if there exists
r0 ≥ 0 such that for all r ≥ r0 , there exists Kr > 0 such that
∀x ∈ X, µ(B(x, 2r)) ≤ Kr µ(B(x, r)).
We then say that X has bounded packing on scales ≥ r0 .
• X is k-geodesic for some k > 0: for every pair of points x, y ∈
X there is a sequence x =Px0 , . . . , xn = y such that d(xi−1 , xi ) ≤
k for all i and d(x, y) = ni=1 d(xi−1 , xi ).
Up to rescaling the metric and/or the measure we will assume that
X is 1-geodesic and has bounded packing on scales ≥ r0 = 1.
A subspace Z ⊂ X is always assumed to be 1-thick (a union of
closed balls of radius 1), so in particular it has positive measure. We
equip Z with the subspace measure and the induced 1-distance
( n
)
X
d(z, z ′ ) = inf
d(zi−1 , zi )
i=1
where the infimum is taken over all sequences z = z0 , . . . , zn = z ′ , such
that each zi ∈ Z and d(zi−1 , zi ) ≤ 1.
Note that (as in the case of a disconnected subgraph) the induced
1-distance will take values in [0, ∞].
Remark 2.2. In the case of (the vertex set of) a bounded degree
graph X, d is the shortest path metric and µ is the (vertex) counting
2If
r0 = 0, then we simply say that X has bounded packing.
12
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
measure. Subspaces Z are (vertex sets of) 1-thick subgraphs equipped
with the vertex counting measure and their own shortest path metric
(the induced 1-distance).
In a locally compact group G with compact generating set K, we
equip G with a Haar measure (which is unique up to scaling) and the
word metric d = dK .
The reason for working with thick sets is justified by the following
easy lemma (see [Tes08, Lemma 8.4]).
Lemma 2.3. Assume X has bounded packing on scales ≥ r0 , and let
A ⊂ X be r-thick for some r ≥ r0 . Then for all u > 0,
µ([A]u ) u µ(A).
3. Poincaré constants
Let (X, d) be a metric space and let a > 0. Given a measurable
function f : X → R, we define its upper gradient at scale a to be
|∇a f |(x) =
sup
|f (y) − f (y ′ )|.
y,y ′ ∈B(x,a)
Remark 3.1. We have slightly modified the notation from [Tes08],
where the upper gradient was referred to as the “local norm of the
gradient” and was denoted by |∇f |a . The changes in this paper are for
brevity; in what follows k|∇a f |kp will simply be denoted by k∇a f kp .
Definition 3.2. Let (Z, d, ν) be a metric measure space with finite
measure and fix a scale a > 0. We define the Lp -Poincaré constant
at scale a of Z to be
k∇a f kp
hpa (Z) = inf
,
f
kf kp
p
fZ :=
Rwhere the infimum is taken over all f ∈ L (Z, ν) such that
− f dν = 0 and f 6≡ 0. We adopt the convention that hp
a (Z) = 0
Z
whenever ν(Z) = 0.
Before continuing we list some basic properties of the Poincaré constant.
Lemma 3.3. Let (Z, d, ν) be a metric measure space with finite measure.
(i) Let θ : (0, ∞) → (0, ∞) be a non-decreasing function, and let
(Z ′ , d′ , ν ′ ) be a metric measure space such that (Z, ν) = (Z ′ , ν ′ ),
and d′ (z1 , z2 ) ≤ θ(d(z1 , z2 )) for all z1 , z2 . Then for all a > 0,
hpa (Z) ≤ hpθ(a) (Z ′ ).
POINCARÉ PROFILES OF GROUPS AND SPACES
13
(ii) Let (Z ′ , d′ , ν ′ ) be a metric measure space where (Z, d) = (Z ′ , d′)
and there exists some M ≥ 1 such that M −1 ν(A) ≤ ν ′ (A) ≤
Mν(A) for every measurable A ⊆ Z. Then for all a > 0,
hpa (Z ′ ) ≤ 2M 2/p hpa (Z).
Proof. Part (i) is immediate.R For part (ii), let f : X R→ R be a measurable function such that − f dµ = 0 and let m = − f dν ′ . We see
that
kf kp,ν ≤ 2kf − mkp,ν ≤ 2M 1/p kf − mkp,ν ′
The first inequality above is the C = −m case of inequality (3.5) proved
in Lemma 3.4. On the other hand
k∇f kp,ν ≤ M 1/p k∇f kp,ν ′ ,
so we are done.
To obtain a sensible definition of the Lp -Poincaré constant it is necessary to only consider functions whose average is zero and to choose
a notion of gradient. In both cases there are multiple ways to do this.
3.1. Choice of average. Given a measure space (Z, ν) with finite
positive measure, there are multiple ways to define the “average” of a
measurable function f : (Z, ν) → R:
R
(1) the average fZ = −Z f dν,
(2) a median mf : any value such that ν({f < mf }) ≤ ν(Z)/2 and
ν({f > mf }) ≤ ν(Z)/2,
(3) a p-energy minimizer: any value cp such that inf c ||f − c||p
is attained for c = cp .
There is a simple comparison between the average and any energy
minimizer, so choosing (1) or (3) gives comparable Poincaré constants.
Lemma 3.4. Let (Z, ν) be a measure space with finite positive measure,
and let f : (Z, ν) → R be a measurable function. For every p ∈ [1, ∞)
we have ||f − cp ||p ≤ ||f − fZ ||p ≤ 2 ||f − cp ||p .
Proof. For any C ∈ R we have
kf − fZ kp ≤ kf + Ckp + kC + fZ kp
Z
1
C+
= kf + Ckp + ν(Z)
f (z)dν(z)
ν(Z) Z
Z
−1+1/p
|C + f (z)|dν(z)
≤ kf + Ckp + ν(Z)
1/p
(3.5)
Z
≤ kf + Ckp + ν(Z)
= 2kf + Ckp .
−1+1/p
kC + f kp k1kp/(p−1)
14
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
In addition, if C = cp , then kf − cp kp ≤ kf − fZ kp by definition.
In the case of p = 1, this lemma combines with the following to show
that taking either averages or medians will yield comparable Poincaré
constants.
Lemma 3.6. Let (Z, ν) be a measure space with finite positive measure
ν and let f : Z → R be a measurable function. Then a value c is a
1-energy minimizer c1 of f if and only if it is a median mf .
Proof. For c′ > c, a calculation gives:
(3.7) kf − c′ k1 − kf − ck1 = (c′ − c) ν({f ≤ c}) − ν({f ≥ c′ })
Z
+
(c + c′ − 2f )dν.
{c<f <c′ }
If c minimizes kf − ck1 , (3.7) gives
0 ≤ (c′ − c) ν({f ≤ c}) − ν({f ≥ c′ }) + (c′ − c)ν({c < f < c′ }).
Letting c′ → c, we get ν({f > c}) ≤ ν({f ≤ c}). The same argument
applied to −f gives ν({f < c}) ≤ ν({f ≥ c}), so c is a median of f .
Conversely, if c is a median for f , (3.7) gives
kf − c′ k1 − kf − ck1
= (c′ − c) (ν({f ≤ c}) − ν({f > c}))
Z
′
′
+ (c − c)ν({c < f < c }) +
(c + c′ − 2f )dν.
{c<f <c′ }
Z
≥ (c′ − c)( 21 ν(Z) − 12 ν(Z)) +
(2c′ − 2f ) ≥ 0,
{c<f <c′ }
so increasing c cannot lower kf − ck1 . The same argument applied to
−f gives that the median c is also a minimizer for kf − ck1 .
Remark 3.8. For Γ a finite graph of constant degree d, λ1,p (Γ), the
first eigenvalue
may be calculated to be the inP of the p-Laplacianon Γ,
P
p
p
fimum of
/
xy∈EΓ |f (x) − f (y)|
x∈V Γ |f (x) − cp (f )| d over all
non-constant f with cp (f ) the energy minimizer of f (see [Bou12]).
Thus by Lemma 3.4 we have λ1,p (Γ) ≍ hp (Γ)1/p .
3.2. Comparison with Lipschitz gradient. Classical Poincaré inequalities on balls in Rn involve the Lp -norms of the usual gradient
vector ∇f . For general metric spaces this makes no sense, but it is
possible to define an analogue of the point-wise norm |∇f |. Given
this, one can define what it means for a metric measure space to satisfy a Poincaré inequality in this infinitesimal sense (see Section 11).
POINCARÉ PROFILES OF GROUPS AND SPACES
15
Let (Z, d, ν) be a metric measure space with finite (positive) measure.
We define the Lipschitz gradient to be
|f (x) − f (y)|
Lipx (f ) = lim sup sup
.
h
h→0 y∈B(x,h)
Given a metric space (Z, d) we can define
hpLip (Z) = inf
||Lipx (f )||p
||f ||p
where the infimum is taken over all non-constant Lipschitz functions
f : Z → R with average 0.
Following §10.2 and §10.3 from [Tes08], one can show that—under
suitable assumptions on a metric measure space—the Poincaré constant
relative to the Lipschitz norm (for Lipschitz functions) is equivalent to
the Poincaré constant with respect to the gradient at some fixed scale
α > 0.
Here, we will focus on one direction (the only one required in the
paper, namely in the proof of Theorem 11.1) which relies solely on a
bounded packing assumption:
Proposition 3.9. Let (Z, d, ν) be a metric measure space with finite
measure ≥ 1, let a > 0 and let C ≥ 1. Assume that for all x ∈ Z,
ν(B(x, 2a)) ≤ Cν(B(x, a/2)). Then,
hpLip (Z) C,a,p hpa (Z).
Proof. We first need the following lemma:
Lemma 3.10. Assume hpa (Z) ≤ 1/8. Let (Px )x be a family of probability measures on Z, such that Px is supported in B(x, a) for every
x ∈ Z. Then there exists f ∈ L∞ such that
k∇a f kp
≤ 4hpa (Z),
kP f − (P f )Z kp
R
where P f (x) := f dPx .
Proof. We start with f with average 0, fZ = 0, such that
1
k∇a f kp
≤ 2hpa (Z) ≤ .
kf kp
4
Observe that
kf − P f kp ≤ k∇a f kp ≤ 41 kf kp ,
from which we deduce that
Z
Z
kf kp
kf − P f kp
≤
.
|(P f )Z | = − P f = − P f − f ≤
1/p
ν(Z)
4ν(Z)1/p
Z
Z
16
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
So k(P f )Z kp ≤ 41 kf kp and then we deduce by the triangle inequality
that
kf kp
kP f − (P f )Z kp ≥
.
2
The rest of the proof of the proposition is similar to that of [Tes08,
Theorem 10.9]. For the convenience of the reader we sketch it. Define
a 1-Lipschitz map θ : Z × Z → R+ by θ(x, y) = d(y, B(x, a)c). For
U ⊂ Z write
Z
θ(x, y)
Px (U) =
dν(y),
U K(x)
R
where K(x) = B(x,a) θ(x, z)dν(z). Note that K(x) ≍C ν(B(x, a)), and
that by assumption, ν(B(x, a)) ≍C ν(B(y, a)) as soon as d(x, y) < a.
Since θ is 1-Lipschitz with respect to x, we see that for all f ∈ L∞ (Z),
Lipx (P f ) C |∇a f |.
Note that if hpa (Z) > 81 , then the statement of the proposition follows
trivially. Hence we can assume that hpa (X) ≤ 81 . By Lemma 3.10 we
deduce that there exists some function f such that
kLipx (P f )kp
C hpa (Z).
kP f − (P f )Z kp
Hence the proposition follows.
4. Poincaré profiles for metric measure spaces
Our goal in this section is to generalise the Poincaré profile to the
class of metric measure spaces defined in Section 2.
Definition 4.1. Let (X, d, µ) be a metric measure space satisfying our
standing assumptions, and fix some number a ≥ 2. We define the
Lp -Poincaré profile ΛpX,a (r) of X at scale a to be the supremum of
µ(A)hpa (A) over all subspaces A ⊂ X satisfying µ(A) ≤ r. If no such
subspace exists, define ΛpX,a (r) = 0.
Recall that by assumption, we only consider 1-thick subsets of X to
be subspaces.
Remark 4.2. As mentioned in the introduction, strictly speaking, we
have defined the Lp -Neumann-Poincaré profile. We could alternatively
define Lp -Dirichlet-Poincaré profiles using Dirichlet-Poincaré inequalities (considering the infimum over all functions which are 0 on the
boundary of Γ in X, rather than those which have average 0). As defined above, the monotone coarse invariant we obtain detects only if
the space has infinite diameter. A small modification to the definition
POINCARÉ PROFILES OF GROUPS AND SPACES
17
(taking the infimum of µ(A)hpa (A) over all subspaces A ⊂ X satisfying
µ(A) ≥ r) yields a coarse invariant (it is not even monotone under
quasi–isometric embeddings) which detects isoperimetry (and in particular, Følner amenability) in the case p = 1. Dirichlet-type Poincaré
inequalities were introduced in [Cou00, Section 7.2] where they are
called Sobolev inequalities (see also [Tes08] for a related notion of Lp isoperimetric profile). They have been extensive studied in the cases
p = 1, where they are equivalent to isoperimetric inequalities, and
p = 2, where they govern the asymptotic behaviour of the probability
of return of the simple symmetric random walk.
We first prove that the Poincaré profile does not actually depend on
the choice of a.
Proposition 4.3. Assume that (Z, d, ν) is a finite metric measure
space. Then for all a ≥ 2 and all p ∈ [1, ∞) we have
hpa (Z) ≍a hp2 (Z).
Proof. We claim that for any t ≥ 0,
(4.4)
ν ({|∇a f | ≥ t}) a ν
t
|∇2 f | ≥
5a
,
and ν({|∇2 f | ≥ t}) ≤ ν({|∇a f | ≥ t}). Together these inequalities immediately imply the proposition. The second inequality is obvious. Let
z ∈ Z, and let x, y ∈ B(z, a). Then one can easily check that our standing assumption implies that there exists a 1-path x = x0 , . . . , xn = y
within B(z, a) such that n ≤ 5a. By the triangle inequality, this means
1
that for at least one 1 ≤ i ≤ n, |f (xi )−f (xi−1 )| ≥ 5a
|f (x)−f (y)|. Now
1
′
′
for all z ∈ B(xi , 1) this implies that |∇2 f |(z ) ≥ 5a |f (x)−f (y)|. Hence
there is a 1-thick subset which is 2a-dense in the set {|∇a f | ≥ t)} on
which |∇2 f |(z ′ ) ≥ 5at . Thus, the left-hand inequality in (4.4) follows
from Lemma 2.3.
Corollary 4.5. Assume that (X, d, µ) satisfies our standing assumptions. Then for all a, a′ ≥ 2 and all p ∈ [1, ∞) we have
ΛpX,a ≍a,a′ ΛpX,a′ .
Moreover, by Lemma 3.3, choosing a bi-Lipschitz equivalent metric and/or measure does not affect the Lp -Poincaré profile ΛpX,a for
sufficiently large a (up to ≃). In particular this means that for a compactly generated locally compact group, the Lp -Poincaré profile does
not depend on the choice of Haar measure or on the choice of compact
generating set.
18
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
In light of Corollary 4.5, we now refer to ΛpX as the Lp -Poincaré
profile of X, without the need to specify a scale.
5. Regular maps and large scale equivalence
The goal of this section is to prove Theorem 1. Firstly, we formally
introduce the notion of a coarse regular map and prove that the definition coincides with regular maps for bounded degree graphs.
5.1. Regular maps. We introduce a natural generalisation of a regular map between graphs suited to the context of metric measure
spaces. In this section we show that Poincaré profiles are monotone
non-decreasing under coarse regular maps.
Definition 5.1. A map F : (X, d, µ) → (X ′ , d′ , µ′ ) is said to be coarse
regular if it satisfies the following properties:
(i) F is coarse Lipschitz: there exists an increasing function ρ+ :
[0, ∞) → [0, ∞) such that for all x, y ∈ X,
d(F (x), F (y)) ≤ ρ+ (d(x, y));
(ii) F is coarsely measure preserving: there exists δ0 such that for
all δ ≥ δ0 and for all (1-thick) subspaces A ⊂ X,
µ([A]δ ) ≍δ µ′ ([F (A)]δ ) ≍δ µ([F −1 (F (A))]δ ).
The parameters of F are the constant δ0 as well as the function ρ+ .
Remark 5.2. Coarse regular maps between spaces with bounded packing on large scales are stable under composition.
In applications, coarse regular maps often are embeddings of the
following kind.
Definition 5.3. A coarse regular map F : (X, d, µ) → (Y, d, ν) is called
a large-scale embedding if it is also a coarse embedding; there exists
a function ρ− such that limt→∞ ρ− (t) = ∞ and for all x, y ∈ X,
ρ− (d(x, y)) ≤ d(F (x), F (y)).
If, in addition, [F (X)]C = Y for some C ≥ 0 (in other words, if F is a
coarse equivalence), then F is called a large-scale equivalence.
It is easy to see that the relation “there exists a large scale equivalence from X to Y ” is an equivalence relation among metric measure
spaces.
POINCARÉ PROFILES OF GROUPS AND SPACES
19
Lemma 5.4. Let X, X ′ be simplicial graphs of bounded degree equipped
with the shortest path metrics and vertex counting measures. A map
F : V X → V X ′ is regular in the sense of [BST12] if and only if it is
coarsely regular as a map between metric measure spaces.
Proof. If F is regular, then by definition there exists a constant K such
that F is K-Lipschitz, the image of every set of measure m has measure
at most m and the pre-image of every set of measure m has measure at
most Km. Since X and X ′ have bounded degree, F is coarsely regular.
Suppose F is coarsely regular, then it is ρ+ (1)-Lipschitz. Fix some
suitable δ0 , let x′ = F (x) ∈ F (V X) and notice that the (1-thick)
subspace A = [x]1 satisfies
F −1 (x′ ) ≤ [F −1 (F (A))]δ δ |[A]δ | ≤ |[x]δ+1 | δ 1.
Thus, F is regular.
The following proposition is the main goal of this section, and will
be proved in §5.2.
Proposition 5.5. Let F : X → X ′ be a coarsely regular map between
metric measure spaces which satisfy our standing assumptions. Then
for all p ∈ [1, ∞),
ΛpX .p ΛpX ′ .
Theorem 1 follows immediately from Lemma 5.4 and this proposition. Note that by Proposition 4.3 it suffices to prove ΛpX,a .p ΛpX ′ ,a′ .
for some a, a′ ≥ 2.
An important consequence of Proposition 5.5 is the following.
Proposition 5.6. Let G and H be compactly generated locally compact
groups, and let φ : H → G be a proper continuous morphism (i.e. ker φ
is compact and φ(H) is a closed subgroup). We assume that both G
and H are equipped with left-invariant Haar measures and word metrics
with respect to some compact symmetric generating sets. Then, for all
p ∈ [1, ∞), ΛpH .p ΛpG . If φ(H) is co-compact then ΛpH ≃p ΛpG .
Proof. The morphism φ is a large-scale embedding hence it is coarsely
regular. If φ(H) is co-compact then φ is a large-scale equivalence. The
result then follows from Proposition 5.5.
5.2. Proof of Theorem 1. The argument behind the proof is as follows: given a coarse regular map F : X → X ′ which is ρ+ -coarse
Lipschitz and coarsely measure preserving for all δ ≥ δ0 , and a subspace Z ⊆ X, we define M = max {ρ+ (1), δ0 } and build metric measure
20
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
space discretizations Y of Z and Y ′ of the 1-thick subspace [[F (Z)]M ]1 .
By the definition of a coarse regular map and Lemma 2.3,
µX (Z) ≍M µX ([Z]M ) ≍M µX ′ ([[F (Z)]M ]1 ).
We then show that the process of taking a discretization yields spaces
with equal measure and comparable Poincaré constants, and finally
prove that Y and Y ′ have comparable Poincaré constants.
The first step of the proof consists in constructing discretizations of
our spaces. We fix some b ≥ M (which we refer to as the discretization
parameter ). We let Y ⊂ Z be a maximal 3b-separated
subset of Z. By
S
maximality Z is covered by the union of balls y∈Y B(y, 9b). We pick
(measurably) a set Ay for each y ∈ Y with the following properties:
• B(y, b) ⊂ Ay ⊂ B(y, 9b);
• (Ay )y∈Y forms a measurable partition of Z.
We equip Y with the subspace distance and the measure νY (y) = ν(Ay ).
Let π : Z → Y be defined by “π(z) is the only y ∈ Y such that z ∈ Ay ”.
Note that π is surjective, and a right-inverse of the inclusion j : Y → Z.
Moreover, π −1 (y) = Ay for every y ∈ Y .
Remark 5.7. Observe that the choice of b ensures that Y has bounded
packing at all scales ≥ 0, and that both π and j are large-scale equivalences. In particular, if Y ′ is a similar discretization of [[F (Z)]M ]1 ,
then Ψ = π ′ ◦ F ◦ j is a coarse regular map. Moreover, if one chooses
the discretization parameter b′ large enough, then Ψ is surjective.
Our next goal is to compare the Poincaré constant of a subspace
with that of its discretization.
Lemma 5.8. Let (Z, d, ν) be a metric measure space with finite measure. Suppose (Y, d, νY ) is a discretization (with parameter b ≥ 2) of
Z as above. Then for all a ≥ b,
hpa (Y ) .a hp20a (Z),
and hpa (Z) ≤ hp20a (Y ).
R
− f dν = 0. We define φ ∈ ℓ∞ (Y )
Proof. LetRf ∈ L∞ (Z) be such
that
Z
R
by φ(y) = −Ay f dν. Clearly −Y φdνY = 0 and kφ ◦ πkZ,p = kφkY,p. Write
POINCARÉ PROFILES OF GROUPS AND SPACES
R
f (z) = φ(π(z)) + −Aπ(z) (f (z) − f (w))dν(w). Then
Z Z
−
kf kZ,p ≤ kφ ◦ πkZ,p +
≤ kφkY,p +
Z
≤ kφkY,p +
Z
p
(f (z) − f (w)) dν(w) dν(z)
Aπ(z)
Z
Z Z
−
21
|f (z) − f (w)|p dν(w)dν(z)
Aπ(z)
|∇10a f |(z)
p
Z
= kφkY,p + k∇10a f kp .
!1/p
!1/p
1/p
On the other hand, it is immediate from the definitions that |∇a φ|(y) ≤
|∇20a f |(z) for all z ∈ Ay .
We now prove the first inequality. If hp20a (Z) ≤ 21 , then for any
ǫ ∈ (0, 1/6) we can find f as above so that
k∇20a f kp
1
k∇20a f kp
2
≥
.
≥ + ǫ ≥ hp20a (Z) + ǫ ≥
3
2
kf kp
kφkp + k∇20a f kp
Thus k∇20a f kp ≤ 2kφkp and
hp20a (Z) + ǫ ≥
1
k∇a φkp
≥ hpa (Y ).
3kφkp
3
Since ǫ was arbitrary, hpa (Y ) ≤ 3hp20a (Z). Moreover, it is easy to see that
hpa (Y ) .a 1 (a much more general statement is proved in Proposition
7.1), so if hp20a (Z) ≥ 12 , then hpa (Y ) .a hp20a (Z).
R
∞
− ψdνY =
The other direction
is
easier:
given
ψ
∈
ℓ
(Y
),
such
that
Y
P
0 we define g = y∈Y ψ(y)1Ay , where 1Ay denotes the characteristic
R
function of Ay . We clearly have − gdν = 0 and kgkp = kψkp . Hence we
are left with comparing the gradients.
Z
X
p
sup |g(z ′ ) − g(z ′′ )|p dν(z)
k∇r gkp =
ν(Ay )−
Ay z ′ ,z ′′ ∈B(z,a)
Y
≤
X
ν(Ay )
Y
≤
X
Y
sup
|g(z ′) − g(z ′′ )|p
z ′ ,z ′′ ∈B(y,10a)
νY (y)
sup
|ψ(y ′ ) − ψ(y ′′ )|p
y ′ ,y ′′ ∈B(y,20a)∩Y
= k∇20a ψkpp .
Now we compare the Poincaré constants of discrete spaces related
by a sufficiently nice surjective coarse regular map.
22
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
Lemma 5.9. Let π : (Y, d, ν) → (Y ′ , d, ν ′ ) be a map between two discrete metric measure spaces with finite (non-degenerate) measures, and
assume that:
• π is surjective;
• ν(π −1 (y ′)) = ν ′ (y ′).
Then for all a ≥ 0 and C such that d(y, z) ≤ a implies d(π(y), π(z)) ≤
Ca, we have
hpa (Y ) ≤ hpCa (Y ′ ).
R
Proof. Choose f ′ R∈ ℓ∞ (Y ′ ) such that f ′ dµ′ = 0 and let f = f ′ ◦ π.
Clearly, we have f dµ = 0, and
kf kp = kf ′ kp .
Moreover, for every y ∈ Y , if y1 , y2 ∈ B(y, a) then π(y1 ), π(y2 ) ∈
B(π(y), Ca). So a straightforward computation shows that
k∇a f kp ≤ k∇Ca f ′ kp .
As a result we obtain a version of Proposition 5.5 in the uniformly
discrete case.
Corollary 5.10. Let Ψ : (Y, d, ν) → (Y ′ , d, ν ′) be a surjective coarse
regular map between uniformly discrete spaces, which have bounded
packing at any scale. Then, for all a > 0, there exists C > 0 such
that
hpa (Y ) a hpCa (Y ′ ).
Proof. The assumptions imply that Ψ is coarse Lipschitz, surjective,
and such that ν(Ψ−1 (y ′)) ≍a ν ′ (y ′ ). Hence the corollary follows from
Lemmas 3.3(ii) and 5.9: if we push the measure ν forward with Ψ to
obtain a measure Ψ∗ ν on Y ′ we have
hpa (Y, ν) hpCa (Y ′ , Ψ∗ ν) hpCa (Y ′ , ν ′ ).
Combining these results we are in a position to prove Proposition
5.5.
Proof of Proposition 5.5. Let Z be a 1-thick subspace of X and define
Z ′ = [[F (Z)]M ]1 where M = max {δ0 , ρ+ (1)}. Then Z ′ is a 1-thick
subspace of X ′ and µ(Z) ≍M µ′ (Z ′ ). Let b, b′ be sufficiently large
that the discretizations Y of Z and Y ′ of Z ′ satisfy the hypotheses of
Lemma 5.8 for some suitable a = a(b, b′ ) ≥ 2 and so that Ψ = π ′ ◦ F ◦ j
is surjective. Note that b and b′ may be chosen independently of the
choice of subspace Z of X, hence a does not depend on Z.
POINCARÉ PROFILES OF GROUPS AND SPACES
23
Applying Corollary 5.10 we see that there exists a constant C depending only on a such that hpa (Y ) a hpCa (Y ′ ). Now, by Lemma 5.8
hpa (Z) a,M hpC ′ a (Z ′ ) where a, M, C ′ do not depend on Z.
Thus there is some M ′ depending only on M and Y ′ such that
ΛpX,a (r)
=
sup {µ(Z)hpa (Z) : µ(Z) ≤ r}
.a,M sup {µ′ (Z ′ )hpC ′ a (Z ′ ) : Z ′ = [[F (Z)]M ]1 , µ(Z) ≤ r}
≤
ΛpX ′ ,C ′ a (M ′ r).
We conclude using Corollary 4.5.
6. Extremal profiles: growth and separation
6.1. Growth and the L∞ -Poincaré profile. In this section we give
the proof of Proposition 2. Recall our standing assumptions: a metric measure space (X, d, µ) is 1-geodesic and has bounded packing at
scales ≥ r0 = 1. Recall also that the growth function γX (k) is the
supremum of the measures of balls of radius k in X, and the inverse
growth function κX (n) is the infimal s such that there exists a ball
B ⊂ X of radius s with measure > n. By assumption subspaces are
1-thick and equipped with a 1-geodesic metric.
Proposition 6.1. Let (X, d, µ) be a metric measure space with unbounded growth function γX : [1, ∞) → (0, ∞), and let a ≥ 2. Then
s
∞
ΛX,a (r) ≃a sup
: γX (1) ≤ s ≤ r ,
κX (s)
where we interpret sup ∅ to be 0.
o
n
In all our applications, the function sup κXs(s) : γX (1) ≤ s ≤ r will
be equivalent to κXr(r) but in general this may not be the case. The
proof requires a lemma.
Lemma 6.2. Let Z be a subspace of X with diameter m and let a ≥ 2.
4a
Then h∞
a (Z) ≤ m , and if every y, z ∈ Z can be joined by a 1-path of
1
length ≤ 2m then h∞
a (Z) ≥ 2m .
Proof. Choose x, y ∈ Z such that d(x, y) ≥ m − δ, and define f (z) =
d(x, z). It is clear that f (x) = 0 and f (y) ≥ m − δ, so ||f − fZ ||∞ ≥
m−δ
, while ||∇a f ||∞ ≤ 2a by the triangle inequality. Thus h∞
a (Z) ≤
2
4a
for
all
δ
>
0.
m−δ
For the second inequality, fix δ > 0 and let f ∈ L∞ (Z) satisfy
inf z∈Z f (z) = 0. Choose y, z so that (f (z)−f (y))+δ ≥ supz∈Z |f (z)| =
||f ||∞ .
24
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
By our hypothesis there exists a sequence of points y = z0 , . . . , zk = z
such that k ≤ 2m and d(zi , zi+1 ) ≤ 1 for all i. Therefore, ∇a f (zi ) ≥
1
1
(||f ||∞ − δ) for some i, so ||∇a f ||∞ ≥ 2m
(||f ||∞ − δ). Since we have
2m
1
||f − fZ ||∞ ≤ ||f ||∞ , letting δ → 0, we see that h∞
a (Z) ≥ 2m .
Proof of Proposition 6.1. The upper bound on Λ∞
X,a (r) follows immediately from Lemma 6.2. Indeed, if µ(Z) ≤ r then
µ(Z)h∞
a (Z) .a
µ(Z)
µ(Z)
≤
,
diam(Z)
κ(µ(Z))
so if µ(Z) ≥ γX (1) we are done. We can ignore Z with µ(Z) bounded by
any fixed constant like γX (1) since any f ∈ L∞ (Z) has a representative
with k∇a f k∞ ≤ 2kf k∞ , and so µ(Z)h∞
a (Z) ≤ 2µ(Z) is bounded too.
We now prove the lower bound. Let t ≥ 2 and choose xt ∈ X
such that µ(B(xt , t)) ≥ 21 γX (t). Define Zt to be the 1-thick subspace
[B(xt , t − 1)]1 . By Lemma 2.3 there is a constant C (which does not
depend on t) such that µ(Zt ) ≤ µ(B(xt , t)) ≤ γX (t) ≤ Cµ(Zt ).
γX (t)
2a
1
∞
.
By Lemma 6.2 h∞
a (Zt ) ∈ [ 4t , (t−1) ], so µ(Zt )ha (Zt ) ≍a
t
There exists s0 ≥ γX (1) so that for all s ≥ s0 , κX (s) ≥ 3. On any
bounded interval in [γX (1), ∞), κX is ≥ 1 and so s/κX (s) is bounded,
thus we may assume that s and r satisfy s0 ≤ s ≤ r. Repeating
the above argument, we see that γX (t)/γX (t − 1) has a uniform upper
bound which is independent of t. Let t = κX (s) − 1 ≥ 2, and so
γX (t) ≤ s ≤ γX (t + 2) γX (t). Thus for r ≥ s0 ,
Λ∞
X,a (r) a
s
γX (t)
.
t
κX (s)
6.2. Separation profiles of metric measure spaces. We wish to
extend the Cheeger constant definition of separation [Hum17] to the
setting of metric measure spaces (X, d, µ) which are 1-geodesic and has
bounded packing at scales ≥ 1.
Given a subspace A ⊂ X (which as usual we assume is 1-thick and
equipped with the induced measure and induced 1-geodesic metric) we
define the boundary at scale a ≥ 2 of A to be
∂a A = [A]a ∩ [Ac ]a
with the usual notation Ac = X \ A. For clarity, given a subspace Z of
X and A ⊂ Z, we also define the boundary at scale a of A in Z to be
∂aZ A = Z ∩ ∂a (A).
Definition 6.3. Let (Z, d, ν) be a metric measure space, where ν(Z)
is finite and let a ≥ 2. We define the Cheeger constant at scale a
POINCARÉ PROFILES OF GROUPS AND SPACES
25
of Z to be
ν(Z)
ν(∂a Ω)
.
: ν(Ω) ≤
ha (Z) = inf
ν(Ω)
2
Let (X, d, µ) be a metric measure space. We define the function
sepX,a (r) = sup {µ(Z)ha (Z)}, where the supremum is taken over all
(1-thick) subspaces Z ⊆ X with µ(Z) ≤ r, and is 0 if no such subspaces
exist.
Remark 6.4. If Γ is a finite graph of bounded degree D then the
boundary at scale a has comparable size to the vertex boundary, so
the usual (vertex) Cheeger constant h(Γ) satisfies h(Γ) ≍a,D ha (Γ).
As a result, if X is an infinite graph of bounded degree D, then
sepX,a ≃a,D sepX , where sepX is the usual separation function for
graphs. (See [Hum17, Propositions 2.2, 2.4].)
6.3. Comparing Cheeger and L1 -Poincaré constants. Our next
goal is to prove Proposition 3. Along the way we will also prove Proposition 6.
Proposition 6.5. Let (X, d, µ) be a metric measure space and let a ≥
2. Then
1
sepX,a ≤ Λ1X,a ≤ sepX,a .
2
We prove this by comparing the Cheeger constant and the L1 -Poincaré constant.
Proposition 6.6. Let (X, d, µ) be a metric measure space. The following co-area formula holds for every non-negative measurable function
f : X → R.
Z
Z
µ (∂a {f > t}) dt
(6.7)
|∇a f |(x)dµ(x) =
X
R+
Proof. For every measurable subset A ⊂ X, we have
Z
(6.8)
µ(∂a A) =
|∇a 1A |(x)dµ(x).
X
Thus, (6.7) follows by integrating over X the following local equalities
Z
|∇a 1{f >t} |(x)dt.
(6.9)
|∇a f |(x) =
R+
It remains to show that these equalities hold for all x ∈ X.
Notice that |∇a 1{f >t} (x)| = 1 if and only if there exists y, y ′ ∈
B(x, a) with f (y) > t and f (y ′) ≤ t. In particular, |∇a 1{f >t} (x)|
26
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
equals one for t ∈ (inf B(x,a) f, supB(x,a) f ) and equals zero for t ∈
/
[inf B(x,a) f, supB(x,a) f ]. Hence,
Z
|∇a 1{f >t} |(x)dt = sup f − inf f = |∇a f |(x),
B(x,a)
R+
B(x,a)
which proves (6.9).
Using this co-area formula we can prove the required relation between
ha (Z) and h1a (Z).
Proposition 6.10. Let (Z, d, ν) be a metric measure space with finite
positive measure ν and let a ≥ 2. Then
h1a (Z) ≤ ha (Z) ≤ 2h1a (Z).
Proof. Let Ω ⊂ Z such that ν(Ω) ≤ ν(Z)/2. We deduce from (6.8)
that
k∇a f k1 = ν(∂a Ω),
where f = 1Ω . On the other hand,
ν(Ω)
ν(Ω)
+ (ν(Z) − ν(Ω))
≥ ν(Ω).
kf − fZ k1 = ν(Ω) 1 −
ν(Z)
ν(Z)
Hence h1a (Z) ≤ ha (Z).
By Lemmas 3.4 and 3.6, for each δ > 0 we may choose f ∈ L1 (Z, ν)
(with median 0) such that
k∇a f k1
≤ 2h1a (Z) + δ.
kf k1
Let f+ = max{f, 0} and f− = min{f, 0}. For any s, s′ , t, t′ > 0 if
′
s+s′
≤ C then st ≤ C or st′ ≤ C. Since kf k1 = kf− k1 + kf+ k1 and
t+t′
k∇a f k1 = k∇a f+ k1 + k∇a f− k1 , we deduce that up to replacing f by
−f , we have
k∇a f+ k1
≤ 2h1a (Z) + δ.
kf+ k1
Hence using (6.7) and the fact that
Z
ν({f > t})dt,
kf+ k1 =
R+
we conclude that there exists some t ≥ 0 such that the subset Ωt =
{f > t} satisfies
ν(∂a Ωt )
≤ 2h1a (Z) + δ.
ν(Ωt )
This proves the second inequality.
ha (Z) ≤
POINCARÉ PROFILES OF GROUPS AND SPACES
27
Proof of Proposition 6: The first half of the above proof can easily be
adapted to prove that 21−p hpa (Z)p ≤ ha (Z). Hence, hpa (Z)p ≤ 2p h1a (Z).
7. Dependency on p
In this section we prove Proposition 5. One trivial upper bound can
always be put on Poincaré constants.
Proposition 7.1. Let (Z, d, ν) be a metric measure space with ν(Z)
finite. Assume there is no z ∈ Z with ν({z}) > 23 µ(Z). For all p ∈
1
[1, ∞) and all a ≥ 2, hpa (Z) ≤ 2 · 3 p .
Proof. By our standing assumptions (Definition 2.1), ν is measure isomorphic to a real interval and an at-most-countable collection of atoms.
It is then easy to find a subset Y ⊂ Z satisfying 13 ν(Z) ≤ ν(Y ) ≤
2
ν(Z). Let f be the characteristic function of Y .
3
1
and ||∇a f ||pp ≤ ν(Z), thus hpa (Z) ≤ 2·3 p .
Then ||f − fY ||pp ≥ ν(Z)
3·2p
Equipped with this we are now able to study the relationship between
different Poincaré profiles of the same space and prove Proposition 5.
Proposition 7.2. Let (Z, d, ν) be a metric measure space with ν(Z)
finite. Assume there is no z ∈ Z with ν({z}) > 32 ν(Z). Then for all
1 ≤ p ≤ q < ∞ and all a ≥ 2,
hqa (Z) p,q hpa (Z).
For all metric measure spaces (X, d, µ) (where µ is possibly infinite),
and all 1 ≤ p ≤ q < ∞,
ΛqX p,q ΛpX .
Proof. Our goal is to prove that for any function g : Z → R, there is a
function f : Z → R such that
||∇a g||q
||∇a f ||p
p,q
≥ hpa (Z).
||g − gZ ||q
||f − fZ ||p
Taking the infimum over all g would then yield the desired result. From
this, we see that it suffices to consider all functions g which satisfy the
upper bound ||∇a g||q ≤ 6 ||g − gZ ||q given by Proposition 7.1. By (3.5)
we have that for all C ∈ R, 6 ||g − gZ ||q ≤ 12 ||g − C||q .
For a ∈ R and p ≥ 1, write {a}p = sign(a)|a|p . For each C, define
f C : Z → R by f C (z) = {g(z) + C}q/p , for some C ∈ R. Since fZC is a
continuous function of C, we fix C so that fZC = 0. Set f = f C .
For each z ∈ Z let (g + C)a (z) = sup {|g(z ′ ) + C| : d(z, z ′ ) ≤ a}.
28
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
By the mean value theorem (see e.g. Matoušek [Mat97, Lemma 4]),
for every s, t ∈ R and α ≥ 1,
|{s}α − {t}α | ≤ α(|s|α−1 + |t|α−1 )|s − t|.
For each z ∈ Z we apply this to s = g(x) + C, t = g(y) + C, α =
all pairs of points x, y ∈ B(z, a) and see that
q−p
2q
|∇a f |(z) ≤ (g + C)a (z) p |∇a g|(z).
p
q
p
for
By the definition of ∇a , (g + C)a (z) ≤ |g(z) + C| + |∇a g| (z), so taking
pth powers and integrating, we see that
||g + C||qq hpa (Z)p = ||f ||pp hpa (Z)p = ||f − fZ ||pp hpa (Z)p
Z
≤
|∇a f | (z)p dν
Z p Z
2q
(|g(z) + C| + |∇a g| (z))q−p |∇a g|(z)p dν
≤
p
Z
p
Z
(⋆)
2q
q−p
q
p
q−p
|g(z) + C| |∇a g| (z) dν + ||∇a g||q
2
≤
p
Z
(†) 2q q p
≤ p ||g + C||qq−p ||∇a g||pq + 12q−p ||g + C||qq−p ||∇a g||pq
p
p,q ||g + C||qq−p ||∇a g||pq ,
where (⋆) follows from (s + t)α ≤ 2α (sα + tα ) for any s, t, α > 0,
and (†) follows from Hölder’s inequality and ||∇a g||q ≤ 12 ||g + C||q .
Rearranging, taking pth roots, and applying (3.5) we have
||g − gZ ||q ≤ 2 ||g + C||q p,q
||∇a g||q
.
hpa (Z)
Remark 7.3. There are graphs X of bounded degree containing expanders, and by Propositions 7.2 and 3,
ΛpX (rn ) p Λ1X (rn ) ≍ sepX (rn ) rn
on some unbounded subsequence (rn ) [Hum17], but Λ∞
X (r) ≃ r/ log(r)
by Proposition 2, so one should not expect universal constants (independent of p, q) in the above proposition.
8. Poincaré profiles of groups with polynomial growth
The goal of this section is to prove the lower bound in Theorem 7.
Given a compactly generated locally compact group group G, with
compact symmetric generating set K, let d = dK be the associated
POINCARÉ PROFILES OF GROUPS AND SPACES
29
word metric and let µ be a left-invariant Haar measure. We refer to the
triple (G, d, µ) as a metric measure CGLC group. By Lemma 3.3
and Corollary 4.5, the Lp -Poincaré profile of G is well-defined (up to
≃).
Theorem 8.1. Let (G, d, µ) be a metric measure CGLC group. If there
exists some m > 0 such that γ(r) ≍ r m , then for every p ∈ [1, ∞],
m−1
ΛpG (r) &p r m .
Note that the p = ∞ case follows immediately from Proposition 6.1.
Moreover, by Proposition 7.2 ΛpG &p Λ1G for all p ∈ [1, ∞). Using
Proposition 6.5 we see that Theorem 8.1 follows from
Theorem 8.2. Let (G, d, µ) be a metric measure CGLC group. If
µ(B(1, r)) ≍ r m then for all a, r sufficiently large, there is a subset Br
of B(1, r) with measure at least 12 µ(B(1, r)) satisfying ha (Br ) a r −1 .
This theorem will be our goal for the section. The proof is in three
parts: the first part gives a general Poincaré inequality satisfied by any
compactly generated locally compact group. Secondly we refine this
inequality for groups with polynomial growth. In the third part we use
this Poincaré inequality (specifically in the L1 setting) to obtain lower
bounds on the Cheeger constant at scale a of metric balls.
8.1. A Poincaré inequality. Poincaré inequalities are well known to
hold for groups with polynomial growth, see for example [SC02]. In this
subsection we present a generalisation of [Kle10, Theorem 2.2] (attributed to Saloff-Coste and explicitly appearing in the L2 case in [DSC93])
to compactly generated locally compact groups in our framework. The
proof below is also similar in nature to [HK00, Proposition 11.17] which
is attributed to Varopoulos [Var87].
Theorem 8.3. Let (G, d, µ) be a metric measure CGLC group. Let
∆ : G → R be the modular function on G; i.e., for U ⊂ G and g ∈ G,
µ(Ug) = µ(U)∆(g). Define ∆(K) = supg∈K ∆(g).
For any p ≥ 1, a ≥ 1, for any metric ball B = B(x0 , R) of radius R
and any function f ∈ L1 (G) we have the following:
Z
Z
(2R)p µ(2B)∆(K)2R
p
|f (x) − fB | dµ(x) ≤
|∇a f | (x)p dµ(x),
µ(B)
B
3B
where for λ > 0, λB = B(x0 , λR).
Proof. We may assume x0 = e. Recall that
|∇a f | (x) = sup {|f (y) − f (z)| : y, z ∈ B(x, a)} .
30
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
If a ≤ a′ then |∇a f | (x) ≤ |∇a′ f | (x) so it suffices to prove the result
above for a = 1.
For every z ∈ 2B, we choose a geodesic γz : {0, 1, . . . , k} → G with
γz (0) = e and γz (k) = z.
For x, y ∈ B(R), let z = x−1 y, and let |γz | = k be the length of the
corresponding path. Then by the triangle and Hölder’s inequality,
p
|f (x) − f (y)| ≤
|γz |
X
p
≤ |γz |
|∇1 f |(xγz (i))
p−1
i=1
|γz |
X
|∇1 f |(xγz (i))p .
i=1
For fixed z ∈ 2B, consider the map F : (x, i) 7→ (xγz (i), i). This is
clearly injective, so (x, i) 7→ xγz (i) is at most 2R-to-1, and
Z X
|γz |
p
|∇1 f |(xγz (i)) dµ(x) =
B i=1
|γz | Z
X
i=1
=
|∇1 f |(xγz (i))p dµ(x)
B
|γz | Z
X
i=1
|∇1 f |(x)p ∆(γz (i)−1 )dµ(x)
B·γz (i)
≤ 2R sup ∆(g)
g∈2B
Z
|∇1 f |(x)p dµ(x).
3B
P
Since 2B = K 2R we have g∈2B ∆(g) = ∆(K 2R ) ≤ ∆(K)2R , so
Z
Z Z
p
dµ(x2 )
p
dµ(x1 )
|f (x1 ) − f (x2 )|
|f − fB | dµ ≤
µ(B)
B
B
B
Z
1
≤
|f (x1 ) − f (x2 )|p dµ(x1 )dµ(x2 )
µ(B) B×B
Z
Z
|γz |
X
(2R)p−1
|∇1 f |(xγz (i))p dµ(z)dµ(x)
≤
µ(B) x∈B z∈2B i=1
Z
Z
(2R)p ∆(K)2R
≤
|∇1 f |(x)p dµ(x)dµ(z)
µ(B)
z∈2B x∈3B
Z
p
2R
(2R) ∆(K) µ(2B)
≤
|∇1 f |(x)p dµ(x).
µ(B)
3B
8.2. CGLC groups with polynomial growth. We begin by refining
the above Poincaré inequality.
Lemma 8.4. If lim inf r→∞ 1r log(µ(B(1, r))) = 0, then G is unimodular.
POINCARÉ PROFILES OF GROUPS AND SPACES
31
Proof. Suppose G is not unimodular, then there exists some g ∈ G
such that ∆(g) > 1. Since ∆ is multiplicative, there is some k ∈ K
with ∆(k) > 1.
Now, for each n, Kk n ⊆ B(1, n + 1), so µ(B(1, n + 1)) > ∆(k)n µ(K),
and therefore lim inf r→∞ 1r log(µ(B(1, r))) > 0.
From this we obtain the following refinement of a special case of
Theorem 8.3.
Corollary 8.5. If G has polynomial growth then there exists a constant
C such that, for any p ≥ 1 and a ≥ 1, for any metric ball B = B(x0 , R)
of radius R and any function f ∈ Lp (G) we have the following:
Z
Z
p
p
|∇a f | (x)p dµ(x).
(8.6)
|f (x) − fB | dµ(x) ≤ CR
B
3B
Using this refined Poincaré inequality (specifically the case p = 1)
we will now present a proof of Theorem 8.2 via a series of lemmas.
The goal is to prove that any subset A of B such that both A ∩ B
and Ac ∩ B have measure proportional to B must have large boundary
inside B. It is not sufficient to apply the Poincaré inequality (8.6) to
the characteristic function of A inside B as we cannot distinguish the
contribution coming from the boundary of A in B with that coming
from the boundary of B in X. The solution is to apply the Poincaré
inequality (8.6) “deep inside” B.
From this we will show that there is a large subset of B with sufficiently large Cheeger constant. This step is modelled on ideas from
[Hum17] relating the cut size and Cheeger constant definitions of separation.
Definition 8.7. Let X be a metric space, let x ∈ X, and let r, s ∈ R+
with s > r. The (r, s)-corona around x is the set Cr,s (x) = B(x, s) \
B(x, r).
Lemma 8.8. Let (G, d, µ) be a metric measure CGLC group with
γ(r) ≍ r m . For each δ ∈ (0, 1) there exists some ǫ > 0 such that
for every x ∈ G and r sufficiently large, we have µ(Cr,(1+ǫ)r (x)) ≤
δµ(B(x, r)).
Proof. By [Tes07, Lemma 24], there exist constants α, β > 0 independent of r such that µ (Cr−s,r (x)) ≥ αµ(Cr,r+s(x)) for every x ∈ G
whenever 4β < s ≤ r.
Let ǫ′ ∈ (0, 1), let r ≥ 8β and for each 1 ≤ i ≤ k = ⌊− log2 ǫ′ ⌋, let
bi = µ(C(1−2i ǫ′ )r,r ).
By construction bi+1 ≥ (1 + α)bi for all i ≥ 1, so bk ≥ (1 + α)k−1b1 .
32
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
Fix δ ∈ (0, 1). If µ(Cr,(1+ǫ′ )r ) > δµ(B(x, r)), then µ(Cr,(1+ǫ′)r ) ≥
δbk ≥ δ(1 + α)k−1 b1 . But, by [Tes07, Lemma 24], b1 ≥ αµ(Cr,(1+ǫ′)r ),
so αδ(1 + α)k−1 ≤ 1.
1
Thus k ≤ log1+α ( αδ
) + 1, which implies that
ǫ′ ≥ ǫα,δ :=
αδ
.
4 log1+α (2)
The conclusion of the lemma holds for all ǫ < ǫα,δ .
Lemma 8.9. Let (G, d, µ) be a metric measure CGLC group with
γ(r) ≍ r m . There exist constants r0 , ǫ, k > 0 such that the following
holds for all r ≥ r0 .
For any A ⊂ B(x, r) with 41 γ(r) ≤ µ(A) ≤ 12 γ(r), there exists a point
w ∈ B(x, r) such that B(w, 3ǫr) ⊂ B(x, r), and such that µ(B(w, ǫr) ∩
A) ≥ kγ(r) and µ(B(w, ǫr) ∩ Ac ) ≥ kγ(r).
Proof. By Lemma 8.8, for all r sufficiently large, and ǫ sufficiently
1
small the corona C(1−3ǫ)r,r (x) has size < 10
γ(r) for every x ∈ X. Now
80
fix k > 0 such that γ(ǫr) ≥ 3 kγ(r) for all r ≥ r0 . Applying Lemma
8.8 with δ = k2 we deduce that
k
k
γ(ǫr) ≤ γ(r),
2
2
holds whenever ǫ is sufficiently small and r sufficiently large.
3
Since µ(A ∩ B(x, (1 − 3ǫ)r)) ≥ 20
γ(r), there exists a point y ∈
3
γ(ǫr) ≥ 2kγ(r). SimiB(x, (1 − 3ǫ)r)) such that µ(A ∩ B(y, ǫr)) ≥ 20
larly, there is some z ∈ B(x, (1 − 3ǫ)r)) such that µ(Ac ∩ B(z, ǫr)) ≥
2kγ(r). Now, by our choice of k, for every v ∈ B(x, (1 − 3ǫ)r),
(8.10)
γ(ǫr + 1) − γ(ǫr) ≤
max {µ(A ∩ B(v, ǫr)), µ(Ac ∩ B(v, ǫr))} ≥ 21 γ(ǫr) ≥ 2kγ(r).
Since y, z ∈ B(x, (1 − ǫ)r)) there is a sequence y = v0 , v1 , . . . , vl = z
such that d(vi−1 , vi ) = 1, l ≤ 2r and {vi } ⊂ B(x, (1 − ǫ)r)). By (8.10),
we see that the measure of the symmetric difference of B(vi , ǫr) and
B(vi+1 , ǫr) is at most kγ(r) for all i.
Choose i maximal such that µ(A ∩ B(vi , ǫr)) ≥ 2kγ(r). If i = l
then we choose w = vl and the proof is complete. If i < l then
µ(Ac ∩ B(vi+1 , ǫr)) ≥ 2kγ(r), but since the symmetric difference of
B(vi , ǫr) and B(vi+1 , ǫr) has measure at most kγ(r), we see that µ(Ac ∩
B(vi , ǫr)) ≥ kγ(r) and we set w = vi .
With this lemma we can show that large subsets of balls have large
boundaries inside the ball.
POINCARÉ PROFILES OF GROUPS AND SPACES
33
Proposition 8.11. Let (G, d, µ) be a metric measure CGLC group
with γ(r) ≍ r m . For every a sufficiently large, there exists a constant
k = k(a) such that for every ball B of radius r, and any subspace
A ⊂ B(x, r) with 14 γ(r) ≤ µ(A) ≤ 12 γ(r), we have µ(∂aB A) ≥ kr m−1 .
Proof. Let A ⊂ B(x, r) = B be such that 41 µ(B) ≤ µ(A) ≤ 21 µ(B). By
Lemma 8.9 there exists some w ∈ B(x, (1−3ǫ)r) such that µ(B(w, ǫr)∩
A), µ(B(w, ǫr) ∩ Ac ) ≥ kµ(A). Applying the Poincaré inequality (8.6)
with p = 1 to the characteristic function 1A on the ball B(w, ǫr) we
see that
1
kµ(A) ≤ Cǫrµ(∂aB(w,3ǫr) A).
2
Since B(w, 3ǫr) ⊆ B we deduce that there exists a constant k ′ > 0
(independent of r) such that
k′
µ(∂aB A) ≥ µ(B).
r
The last step in this argument ensures that there is a large subset of
the ball with suitable Cheeger constant at scale a. This is a generalisation of a similar result for graphs presented in [Hum17].
Proposition 8.12. Let (X, d, µ) be a metric measure space such that
inf x∈X µ(B(x, 1)) = c > 0, and let a, r ≥ 2. If there exists a constant λ = λ(a, r) ≤ 14 and a ball B = B(x, r) of radius r, such
that for any subspace A ⊂ B with 14 µ(B) ≤ µ(A) ≤ 21 µ(B), we have
µ(∂aB A) ≥ λµ(A), then there exists some 1-thick subspace B ′ of B such
that µ(B ′ ) ≥ 12 µ(B) and ha (B ′ ) ≥ λ2 .
Proof. Fix B as above. Given any subset A0 of B such that µ(A0 ) ≤
1
µ(B) and µ(∂a A0 ) < λ2 µ(B), we have µ(A0 ) < 41 µ(B) by assumption.
2
Let m be the supremum of the measures of all subsets A0 satisfying
.
the above, and let A1 be such a subset with measure at least m − 2c
λ
Define A′ = [B \ [A1 ]a ]1 ⊆ B \ A1 . Note that A′ is 1-thick and µ(A′ ) ≥
µ(B) − µ(A1 ) − µ(∂a A1 ) ≥ 85 µ(B).
We wish to show that ha (A′ ) ≥ λ/2. Suppose for a contradiction that
′
there exists some subset E ⊂ A′ with µ(E) ≤ 21 µ(A′ ) and µ(∂aA E) <
′
λ
µ(E). Since ∂aB E ⊆ ∂aA E ∪∂aB A1 , we have µ(∂aB (E ∪A1 )) < λ2 (µ(E)+
2
µ(A1 )). From this we deduce that µ(E) + µ(A1 ) > 12 µ(B), or, if this is
not the case, then µ(E) ≤ 2c
by the choice of A1 .
λ
′
In the second case we are done: ∂aA E contains a ball of radius 1, so
′
c ≤ µ(∂aA E) < λ2 µ(E) ≤ c which is a contradiction.
Otherwise µ(E ∪ A1 ) ∈ ( 21 µ(B), 43 µ(B)) so E ′ = B \ (E ∪ A1 ) satisfies
µ(E ′ ) ∈ ( 41 µ(B), 12 µ(B)) and µ(∂aB E ′ ) = µ(∂aB (E ∪ A1 )) < λ2 µ(B),
which is also a contradiction.
34
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
Proof of Theorem 8.2. This follows immediately from Propositions 8.11
and 8.12 with λ = k/r.
9. Upper bounds and large-scale dimension
The goal of this section is to obtain upper bounds on the Poincaré
profiles of a metric measure space which is finite dimensional in the
sense of the definition below. In doing so, we will prove that the lower
bound for groups of polynomial growth in section 8 is sharp to complete
the proof of Theorem 7.
Definition 9.1. Let (X, d, µ) be a metric measure space. We say X
has measurable dimension at most n (mdim(X) ≤ n) if, for
S all
r ≥ 0 we can write X = X0 ∪ · · · ∪ Xn and decompose each Xi = Xij
so that each Xij is 1-thick, sup(µ(Xij )) < ∞ and d(Xij , Xij ′ ) ≥ r
whenever j 6= j ′ .
If mdim(X) ≤ n we define the function γn (r) to be the infimal value
of sup(µ(Xij )) + 1 taken over all decompositions of X satisfying the
above hypotheses.
Notice that γn (r) is non-decreasing as a function of r.
A simple comparison can be made with asymptotic dimension when
the metric measure space has bounded geometry: for all r ≥ 0 there
exists some Cr such that µ(B(x, r)) ≤ Cr for all x ∈ X.
Lemma 9.2. Let (X, d, µ) be a metric measure space with bounded
geometry. Then the asymptotic dimension of X is at least mdim(X).
Proof. Suppose asdim(X) ≤ n. This implies that for all r ≥ 0 one
S can
decompose X = X0′ ∪ . . . ∪ Xn′ and further decompose each Xi′ = Xij′
so that sup diam(Xij′ ) = Kr < ∞ and d(Xij , Xij ′ ) ≥ r + 2 whenever
j 6= j ′ .
S
Define Xij = y∈X ′ B(y, 1). Each Xij is 1-thick, it has diameter at
ij
most L = Kr + 2 and d(Xij , Xij ′ ) ≥ r whenever j 6= j ′ . Since X has
bounded geometry, µ(Xij ) ≤ CL for all i, j.
Lemma 9.3. Let (X, d, µ) and (Y, d′ , µ′) be metric measure spaces and
suppose Y has bounded packing at scales ≥ 1. If there exists a coarsely
regular map F : X → Y , then mdim(X) ≤ mdim(Y ). Moreover, for
all suitable n we have γnX .n γnY .
Proof.SSuppose
mdim(Y ) ≤ n. Then for all r ≥ 0 one can write
n S
r
Y = i=0 j Yij where each Yijr is 1-thick, µ′ (Yijr ) ≤ C for some C
and all i, j, and d′ (Yijr , Yijr ′ ) > ρ+ (r + 2) whenever j 6= j ′ .
Let Xijr = [F −1 (Yijr )]1 . By Definition 5.1(i), d(Xijr , Xijr ′ ) > r whenever j 6= j ′ , and by (ii) µ(Xijr ) ≍ µ′ ([Yijr ]1 ) µ′ (Yijr ) by Lemma 2.3.
POINCARÉ PROFILES OF GROUPS AND SPACES
35
Remark 9.4. One can remove the assumption that Y has bounded
packing at scales ≥ 1 by removing the assumption that each Xij is
1-thick in the definition of measurable dimension.
Proposition 9.5. Let (X, d, µ) be a metric measure space with µ(X) =
∞ and measurable dimension at most n. For all δ > 0,
ΛpX (r) .n sup {γn (t + δ)/t : γn (t) ≤ r/(4n + 4)} .
Proof. If γn is bounded then µ is bounded, which is a contradiction.
Choose s > 4(n + 1)γn (0) and assume µ(A) = s ≤ r. Fix δ > 0
and find t so that 4(n + 1)γn (t) ≤ µ(A) ≤ 4(n + 1)γn (t + δ). Select a
decomposition of X into sets Xijt as above where µ(Xijt ) ≤ γn (t) for all
i, j.
1
Then there exists some i such that µ(A ∩ Xi ) ≥ n+1
µ(A) ≥ 4γn (t).
Without loss of generality, assume i = 0. Choose J so that
[
µ(A)
µ(A)
t
X0′ :=
X0j
satisfies
≤ µ(A ∩ X0′ ) ≤
.
4(n
+
1)
2(n
+
1)
j∈J
Set X0′′ = X0t \ X0′ and let ft : A → R be the function f (x) =
1
min {t, dX (x, X0′ )}.
t
R
p
Now ft is 1t -Lipschitz, so A |∇2 f |p ≤ 2tp µ(A). Since f takes values
in [0, 1] and has value 0 on XR0′ and value 1 on X0′′ each of measure
1
µ(A).
≥ µ(A)/4(n + 1), we see that A |f − fR |p dµ(x) ≥ ( 21 )p 4(n+1)
4
1
p
2
Thus, ha (A) ≤ t (n + 1) n t . As this holds for every measurable
A ⊂ X of finite measure the result follows.
Remark 9.6. Under nice circumstances, for instance when a space
X has a cobounded isometry group, and finite asymptotic dimension
where the Kr can be bounded by an affine function of r (sometimes
called linearly controlled or asymptotic Assouad–Nagata dimension),
the function γn (sr + δ)/sr is equivalent (up to ≃) to r/κ(r) where κ is
the inverse growth function. This is easily deduced from the argument
in the proof of Proposition 6.1.
Proof of Theorem 7. Let (G, d, µ) be a CGLC metric measure group
with µ(B(1, r)) ≍ r m . Such groups have finite asymptotic Assouad–
Nagata dimension ([Bre14, Theorem 1.2] and [HP13, Theorem 5.5]), so
m−1
by Proposition 9.5, ΛpG (r) . r m for all p ≥ 1. The lower bound is
proved in Theorem 8.1.
Example 9.7. As another example, for X equal to the product of two
3-regular trees we have ΛpX (r) ≃ r/ log(r) for all p ∈ [1, ∞]: The case
p = ∞ follows immediately from Proposition 2. By [BST12, Theorem
36
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
3.1] and Proposition 6.5, the lower bound holds when p = 1, so the
lower bound for general p follows from Proposition 7.2. For the upper bound, X has exponential growth, a cobounded isometry group,
and asymptotic Assouad–Nagata dimension 2, so by Proposition 9.5,
ΛpX (r) . r/ log(r) for all p ≥ 1.
10. Trees
In this section, we calculate the Poincaré profile for regular trees.
Theorem 10.1 (Theorem 9). Let T be the infinite 3-regular tree. Then
for every p ∈ [1, ∞), ΛpT (r) ≍p r (p−1)/p .
For p = 1 this is immediate from [BST12]. This theorem immediately
implies the following corollary for groups admitting quasi-isometric embeddings of such trees.
Corollary 10.2. If (G, d, µ) is a CGLC measure group which is nonamenable, non-unimodular, or is compact-by-elementary amenable and
has exponential growth, then for any p ≥ 1, ΛpG (r) &G,p r (p−1)/p .
Proof. In the first two cases this follows from [BS97], and in the third
from [Cho80].
In this section, for a graph X, and a function f : V X → R, we
define |∇f | : EX → R as |∇f |(e) = |f (x) − f (y)| where e ∈ EX has
endpoints x, y ∈ V X. If X has maximum vertex degree d then for each
p ≥ 1,
!1/p
X
k∇2 f kp ≍d k∇f kp =
.
|∇f |(e)p
e∈EX
A key step in proving Theorem 10.1 is to reduce to an estimate on
complete graphs in the spirit of, for example, Spielman [Spi15, Section
4.7].
Proposition 10.3. For any r ∈ N, r ≥ 2 and p ∈ [1, ∞), letting Kr
denote the complete graph on r vertices, we have
k∇f kp
1/p
: f : V Kr → R, f 6≡ fKr p r 1/p .
r ≤ inf
kf − fKr kp
POINCARÉ PROFILES OF GROUPS AND SPACES
37
Proof. Let f : V Kr → R be any non-constant function on Kr . Then
kf −
fKr kpp
=
X
x
1X
f (y)
f (x) −
r y
p
1 X X
≤ p
|f (x) − f (y)|
r x
y
!p
!
1 X X
|f (x) − f (y)|p r p−1
≤ p
r x
y
= r −1 k∇f kpp .
This proves the first inequality; the second can be seen by considering
a function which is 1 and −1 on one vertex each, and zero everywhere
else.
Proof of Theorem 10.1. First we show the upper bound, which is relatively simple.
Suppose A ⊂ T is a graph of size |A| = r; we can find a vertex x so
that on deleting this vertex, all remaining connected components have
size ≤ r/2. Group these components into sets U, V of size ∈ [r/4, 3r/4].
Let f : A → [−1, 1] be identically −1 on U, 1 on V and 0 on x.
Clearly kf − fA kpp ≥ 41 r, and since ∇f is only non-zero on edges
adjacent to x, k∇f kpp ≤ 3. Thus hp (A) ≤ (12/r)1/p and
Λp (r) = sup |A|hp (A) ≤ 12r (p−1)/p .
|A|≤r
Second, we show the lower bound.
For any r > 0 there exists a ball B = B(x0 , t) ⊂ T of size ≍ 2t ≍ r,
so we can assume r = |B| and it then suffices to show that hp (B)
|B|−1/p , with constant independent of B.
Let Kr be the complete graph on r vertices. Suppose that a nonconstant function f : B → R is given. Consider f as a function on the
complete graph Kr = K|B| . In light of Proposition 10.3, to show that
hp (B) |B|−1/p , it suffices to show that
X
1 X
|∇f (e)|p ≥
|f (x) − f (y)|p,
2
2|B|
e∈EB
x,y∈B
for then hp (B) |B|−2/p r 1/p |B|−1/p .
Now for each x, y ∈ B, let γxy P
be the simple path in T joining x to
y. Observe that |f (x) − f (y)| ≤ e∈γxy |∇f (e)|.
38
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
For each e ∈ EB, let Ne be the number of such simple paths that
pass through e. Observe that Ne ≍ 2t · 2t−d(x0 ,e) , where d(x0 , e) is the
distance from the centre of the ball to the edge e.
Using Hölder’s inequality, we have
p
X
X
X
|f (x) − f (y)|p ≤
|∇f (e)|
x,y∈B
e⊂γxy
x,y∈B
=
X
x,y∈B
≤
X
x,y∈B
X
e⊂γxy
X
e⊂γxy
1/(p−1)
p
|∇f (e)|Ne−1/p Ne1/p
|∇f (e)|p Ne−1
X
e⊂γxy
p−1
Ne1/(p−1)
For each simple path, Ne
takes values in (two) geometric series,
with ratio depending only on p and maximum value (22t )1/(p−1) ≍
|B|2/(p−1) , and so the sum inside the second parentheses above is also
|B|2/(p−1) . Thus,
X
X X
|f (x) − f (y)|p
|∇f (e)|pNe−1 |B|2
x,y∈B
x,y∈B e∈γxy
≤ 2|B|2
X
|∇f (e)|p ,
e∈B
and so we are done.
11. Lower bounds for hyperbolic spaces with boundary
Poincaré inequalities
In this section we find lower bounds on Poincaré profiles for hyperbolic groups whose boundaries admit Poincaré inequalities in the sense
of Heinonen and Koskela (Theorem 13). In section 13 we will apply
these results to rank 1 symmetric spaces, and a family of hyperbolic
buildings studied by Bourdon and Pajot.
Suppose a metric space (Z, ρ) is Ahlfors Q-regular, i.e. there is a
measure µ on Z so that for every ball B(z, r) in Z with r ≤ diam(Z), we
have µ(B(z, r)) ≍ r Q . (We may take µ to be the Hausdorff Q-measure
on Z.) For p, q ≥ 1, we say (Z, ρ) admits a (q, p)-Poincaré inequality
(with constant L ≥ 1) if for every Lipschitz function f : Z → R and
POINCARÉ PROFILES OF GROUPS AND SPACES
every ball B(z, r) ⊂ Z,
Z
1/q
Z
q
−
|f − fB(z,r) | dµ
≤ Lr −
B(z,r)
p
(Lipx f ) dµ(x)
B(z,Lr)
R
where for U ⊂ Z, fU = −U f dµ =
1
µ(U )
U
1/p
,
f dµ, and
|f (y) − f (x)|
.
r
y∈B(x,r)
Lipx f = lim sup sup
r→0
R
39
If q = 1, we say Z admits a p-Poincaré inequality. By Hölder’s
inequality, if Z admits a p-Poincaré inequality, it admits a q-Poincaré
inequality for all q ≥ p. Moreover, since Z is doubling, it will admit
(q, q)-Poincaré inequalities for all q ≥ p by [HK00, Theorem 5.1].
A geodesic metric measure space (X, d, µ) is Gromov hyperbolic
if it is δ-hyperbolic for some δ ≥ 0: for every geodesic triangle T =
(γ1 , γ2, γ3 ), we have γ1 ⊆ [γ2 ∪ γ3 ]δ . It is visual if there exists x0 ∈ X
and C ≥ 0 so that every x ∈ X belongs to a C-quasi-geodesic ray
γ : [0, ∞) → X with γ(0) = x0 . Gromov hyperbolic metric spaces have
a boundary at infinity ∂∞ X which comes with a family of metrics: if
X is visual with respect to x0 , a visual metric ρ on ∂∞ X based at
x0 ∈ X with visibility parameter ǫ > 0 is a metric satisfying ρ(·, ·) ≍
exp(−ǫ(·|·)x0 ), where (·|·)x0 denotes the Gromov product with respect
to x0 . For more background and discussion, see [BS00, BP03].
We can now state the first main result of this section (cf. Theorem 13).
Theorem 11.1. Suppose that X is a visual Gromov hyperbolic graph
with a visual metric ρ on ∂∞ X that is Ahlfors Q-regular and admits a
p-Poincaré inequality. Then for all q ≥ p, ΛqX (r) & r 1−1/Q .
By taking discretizations, one can apply this result to rank-1 symmetric spaces, amongst other examples.
Proof. Consider ∂∞ X with the metric ρ, which admits a p-Poincaré
inequality with some constant L ≥ 1. As a consequence, (∂∞ X, ρ)
is quasi-convex, so ρ is bi-Lipschitz equivalent to a geodesic metric.
Therefore we may assume that ρ is geodesic, and so our standing assumptions hold.
Following Bourdon–Pajot [BP03, Section 2.1], we ensure that Z =
(∂∞ X, ρ) has diameter 1/2 by rescaling, and define a graph Γ which
approximates Z: Γ has vertex set {zti : t ∈ N, 1 ≤ i ≤ k(t)} where for
k(t)
each t ∈ N, Γt = {zt1 , . . . , zt } is a maximal e−t -separated net in Z.
To each zti we associate a ball B(zti , e−t ) ⊂ Z, and we join zti and zuj by
an edge if and only if |t − u| ≤ 1 and B(zti , e−t ) ∩ B(zuj , e−u ) 6= ∅.
40
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
By Bourdon–Pajot [BP03, Proposition 2.1, Corollary 2.4], Γ, with
the path metric d, is a bounded degree hyperbolic graph which is quasiisometric to X, and so it suffices to show the Poincaré profile bound
for Γ.
We now consider the sequence Zt = (Z, ρt , µt ) of metric measure
spaces, where ρt = 6et ρ, and µt = eQt µ. Note that µt (Zt ) ≍ eQt . We
deduce from the Poincaré inequality satisfied by Z that Zt satisfies for
any Lipschitz function f on Zt , for all q ≥ p
1/q
Z
Z
1/q
q
q
t
− |f − fZt | dµt
e − (Lipx f ) dµt (x)
,
and therefore that
hqLip (Zt ) e−t
with constant independent of t. By Proposition 3.9, this implies that
hq2 (Zt ) e−t .
Now equip Γt with the counting measure and the distance induced from
its inclusion in Zt . Since Γt is a maximal 6-separated subset of Zt , we
can find a partition
G
Aγ ,
Zt =
γ∈Γt
where
Bρt (γ, 2) ⊂ Aγ ⊂ Bρt (γ, 18).
By the Ahlfors regularity of Zt , µ(Aγ ) ≍ 1. Hence by Lemmas 5.8 and
3.3(ii), we deduce that
hq40 (Γt , ρt ) e−t .
In order to conclude, we need to show that there exists a constant
C such that two vertices x, y ∈ Γt such that ρt (x, y) ≤ 40 satisfy
d(x, y) ≤ C (where d(x, y) is their distance in Γ). Indeed, that will
show that
hqC (Γt , d) e−t ,
(11.2)
and since |Γt | ≍ eQt ,
ΛqΓ (r) r 1−1/Q .
By [BP03, Lemma 2.2], for x, y ∈ Γ corresponding to balls Bx , By ⊂
Z, e−(x|y) ≍ diam(Bx ∪ By ), where (x|y) denotes the Gromov product
with respect to the base point z11 . For x, y ∈ Γt , we have (x|y) equal
to t − 21 d(x, y) up to a uniform additive error, and diam(Bx ∪ By ) ≍
e−t + ρ(x, y), so
1
e−t e 2 d(x,y) ≍ diam(Bx ∪ By ) ≍ e−t + ρ(x, y).
POINCARÉ PROFILES OF GROUPS AND SPACES
41
e−t implies that d(x, y) 1, which completes the
Thus, ρ(x, y) ≤ 40
6
proof of Theorem 11.1.
We will see in section 13 that for many spaces, Theorem 11.1 gives
sharp lower bounds for ΛqX when q ∈ [1, Q). For q = Q, however, one
can do better.
Theorem 11.3. Suppose that X is a visual Gromov hyperbolic graph
with a visual metric ρ on ∂∞ X that is Ahlfors Q-regular and admits a
1−1/Q
Q-Poincaré inequality. Then ΛQ
log(r)1/Q .
X (r) & r
Proof. We continue with the notation of the proof of Theorem 11.1.
Given s < t ∈ N, let Bs,t be the full subgraph of Γ containing the
layers Γs+1 , Γs+2 , . . . , Γt . (Later we will take s = ⌊t/2⌋.) The strategy
of the proof is to use the Poincaré inequality in each layer to get a
stronger constant for all of Bs,t .
Let us be given a function f : Bs,t → R, i.e. a function on V Bs,t .
For x ∈ Γ, define ix ∈ N to satisfy x ∈ Γix . Given x ∈ Γ and i ≤ ix ,
let πi (x) ∈ Γi be (one of) the points in Γi so that the point in Z
corresponding to x lies in the ball of radius e−i corresponding to πi (x);
the allowed choices of πi (x) are all at distance 1 from each other.
For i = s + 1, . . . , t, there are i − s layers in Bs,t with labels ≤ i.
Lemma 11.4. There is an assignment Bs,t → N that maps each x ∈
Bs,t to a layer cx ∈ {s + 1, . . . , ix }, so that for any z ∈ Bs,t and any c, i
with c ≤ iz ≤ i ≤ t we have
eQ(i−iz )
eQ(t−s)
(11.5)
|{x ∈ Γi : πiz (x) = z and cx = c}|
≤
,
i−s
t−s
where the constant of ‘ ’ is independent of s, t, z, c and i.
This follows from a colouring argument that we defer until later.
Similarly to the proofs in Section 10, we bound
p
kf −
(11.6)
fBs,t kpp
1 X
f (x) −
f (y)
=
|Bs,t| y
x
1 XX
≤
|f (x) − f (y)|p.
|Bs,t| x y
X
(We refrain from setting p = Q at present to clarify the role this plays
in the proof.)
At a cost of multiplying by 2, we can restrict to sum only over x, y
where ix ≤ iy . In particular, cx ≤ iy . Given such x, y, we consider the
path αx that follows x, πix −1 (x), . . . , πcx (x), and also the path βx,y that
follows along πcx (y), πcx +1 (y), . . . , πiy −1 (y), y.
42
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
Continuing from (11.6), since
X
|f (x) − f (y)| ≤
|∇1 f |(z)
z∈αx
!
+ |f (πcx (x)) − f (πcx (y))|
X
+
!
|∇1 f |(z) ,
z∈βx,y
we use the inequality (a+ b+ c)p ≤ 3p (ap + bp + cp ) to find the following:
1 X
(11.7) kf − fBs,t kpp
|Bs,t | x,y
ix ≤iy
X
|∇1 f |(z)
z∈αx
!p
+
p
1 XX
1 X
|∇1 f |(z) .
|f (πcx (x)) − f (πcx (y))|p +
|Bs,t | x,y
|Bs,t| x,y z∈β
ix ≤iy
x,y
ix ≤iy
We denote the resulting three terms of the sum by S1 , S2 , and S3 . For
each z ∈ Bs,t , let Mz be the number of pairs (x, y) so that αx passes
through z, and likewise Nz for βx,y . Let us bound the first term of
(11.7), S1 .
!p
1 X X
S1 =
|∇1 f |(z)Mz−1/p Mz1/p
|Bs,t| x,y z∈α
x
ix ≤iy
1 X
≤
|Bs,t | x,y
ix ≤iy
X
z∈αx
|∇1 f |(z)p Mz−1
!
X
Mz1/(p−1)
z∈αx
!p−1
when p > 1. If z ∈ Γs+j for some j ∈ {1, . . . , t − s}, then by (11.5) the
number of possible choices of x is
t
X
i=s+j
j
eQ(i−s−j)
j Q(t−s−j)
e
i−s
t−s
and there are ≤ |Bs,t | possible choices of y so that z ∈ αx . Thus
Mz
j eQ(t−s) |Bs,t |
j Q(t−s−j)
e
|Bs,t| = Qj ·
.
t−s
e
t−s
POINCARÉ PROFILES OF GROUPS AND SPACES
43
P
For any p > 1, j≥1 (je−Qj )1/(p−1) is bounded by some constant depending only on Q and p. Whether p > 1 or p = 1, we get that
!
eQ(t−s) X X
S1
|∇1 f |(z)p Mz−1
t − s x,y z∈α
x
ix ≤iy
Q(t−s)
=
e
t−s
X
z
|∇1 f |(z)p
X
x,y:ix ≤iy ,z∈αx
Mz−1 =
eQ(t−s)
k∇1 f kpp .
t−s
A very similar calculation lets us bound S3 : if z ∈ Γs+j for some
j
|Bs,t| possible choices
j ∈ {1, . . . , t − s}, then by (11.5) there are t−s
Q(t−s−j)
of x and e
possible choices of y so that z ∈ βx,y . Thus
j eQ(t−s) |Bs,t |
j
Q(t−s−j)
|B
|
·
e
=
·
,
Nz
s,t
t−s
eQj
t−s
and the rest of the calculation goes through as before to give S3
1 Q(t−s)
e
k∇1 f kpp .
t−s
It remains to bound S2 . Suppose we have x′ , y ′ ∈ Γs+j for some
j ∈ {1, . . . , t − s}. Let Px′ ,y′ be the number of pairs x, y ∈ Bs,t so that
ix ≤ iy and πcx (x) = x′ and πcx (y) = y ′ . Using again (11.5), we can
bound Px′ ,y′ by the product of the number of choices of x, which is
1 Q(t−s−j)
t−s
e
, and the number of choices of y, which is eQ(t−s−j) .
Thus
1 X
|f (πcx (x)) − f (πcx (y))|p
S2 =
|Bs,t | x,y
ix ≤iy
t−s
1 X
=
|Bs,t | j=1
(11.8)
X
Px′ ,y′ |f (x′ ) − f (y ′)|p
x′ ,y ′ ∈Γs+j
t−s
e2Q(t−s) X −2Qj X
e
(t − s)|Bs,t| j=1
x′ ,y ′ ∈Γ
|f (x′ ) − f (y ′)|p .
s+j
Fixing for a moment our choice of j, let fj be the average value of
f restricted to Γs+j . Assuming Z = (∂∞ X, ρ) satisfies a p-Poincaré
inequality, we apply (11.2) to Γs+j to obtain:
X
X
p
|f (x′ ) − fj | ep(s+j)
|∇C f |(x′ )p .
x′ ∈Γs+j
x′ ∈Γs+j
44
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
Applying this twice, we have that
X
X
|f (x′ ) − f (y ′)|p ≤ 2p
(|f (x′ ) − fj |p + |f (y ′) − fj |p )
x′ ,y ′ ∈Γs+j
x′ ,y ′ ∈Γs+j
ep(s+j) |Γs+j |
X
|∇C f |(x′ )p .
x′ ∈Γs+j
Since |Γs+j | ≍ eQ(s+j), and |Bs,t| ≍ eQt , on substituting this back in to
(11.8), we get
t−s
e2Q(t−s) X −2Qj p(s+j) Q(s+j) X
S2
e
e
e
|∇C f |(x′ )p
(t − s)|Bs,t| j=1
′
x ∈Γs+j
≍
t−s
eQt+(p−Q)s X (p−Q)j X
e
|∇C f |(x′ )p .
t − s j=1
′
x ∈Γs+j
Provided p = Q, this simplifies to
eQt
k∇C f kQ
Q.
t−s
Our bounds for S1 and S3 are dominated by our bounds for S2 , so
we set s = ⌊t/2⌋ and conclude by (11.7) that
S2
kf − fBs,t kQ
Q
eQt
|Bs,t |
k∇C f kQ
k∇C f kQ
Q ≍
Q.
t
log |Bs,t |
It remains to show the colouring argument giving (11.5).
Proof of Lemma 11.4. Recall that we are defining a colouring map Bs,t →
{s + 1, . . . , t}, x 7→ cx .
For each i ∈ {s+1, . . . , t}, the vertices of Γi correspond to a maximal
e−i -separated net in Z. By Ahlfors Q-regularity, there exists C so that
the number of e−i separated points in any r-ball in Z is ≤ C(r/e−i )Q =
Cr Q eiQ . So if we let ri = 12 (i − s)1/Q C −1/Q e−i , we guarantee that any
ri -ball in Z meets at most (i − s) points corresponding to vertices of
Γi .
Define Γi → {s + 1, . . . , i}, x 7→ cx to be any mapping so that no
two points at distance ≤ ri in Z are mapped to the same value. The
existence of such a mapping follows from Zorn’s lemma applied to the
collection of all such partially defined functions.
Doing this for each i, we obtain our mapping Bs,t → {s+1, . . . , t}. To
verify that (11.5) holds, observe that for any z ∈ Bs,t and c, i satisfying
c ≤ iz ≤ i ≤ t the set {x ∈ Γi : πix (x) = z and cx = iz } is an ri separated set in B(z, e−iz ) ⊂ Z, therefore by Ahlfors regularity it has
POINCARÉ PROFILES OF GROUPS AND SPACES
cardinality
−iz Q
Q
e−iz
e
eQ(i−iz )
eQ(t−s)
=
≤
.
ri
(i − s)1/Q e−i
(i − s)
t−s
45
12. Upper bounds for hyperbolic spaces with hyperplanes
In this section we present an approach to finding upper bounds on
the Lp -Poincaré profiles of hyperbolic spaces. Our hypotheses are as
follows:
(1) (X, d, µ) is a δ-hyperbolic geodesic metric measure space, and
it is visual with respect to a given point x0 ∈ X: there exists
C ≥ 0 so that every x ∈ X belongs to a C-quasi-geodesic ray
γ : [0, ∞) → X with γ(0) = x0 .
(2) There exists a constant h(X) > 0 (called the volume entropy)
and a constant C ≥ 0 such that for every R > 0, h(X)R − C ≤
loge (µ(BR (x0 ))) ≤ h(X)R + C.
(3) There is a visual metric ρ on ∂∞ X based at x0 ∈ X with visibility parameter ǫ > 0; i.e., ρ(·, ·) ≍ exp(−ǫ(·|·)x0 ), where (·|·)x0
denotes the Gromov product with respect to x0 .
For our last hypothesis, we require the following notion.
Definition 12.1. Let (X, d) be a metric space and x0 ∈ X. For C ≥ 1,
a subset A ⊆ X is said to be a C-asymptotic shadow of x0 if, for
every x ∈ A there is a C-quasi-geodesic ray γx : [0, ∞) → X with
γx (0) = x0 and d(γx (rx )) = x for some rx , and γx [rx , ∞) ⊆ A. (Recall
that a C-quasi-geodesic ray is a (C, C)-quasi-isometric embedding of
[0, ∞).)
The final hypothesis only needs to hold for large a, where a ≥ 2 is
the constant of thickness in Definition 4.1. Let Isomµ (X) be the group
of µ-preserving isometries of X.
(4) There exist constants κ, N, C > 0 such that for any a-thick
subspace Z of X with measure at least N, there is some ψ ∈
Isomµ (X), and there exist two measurable subsets H ± of X
which are C-asymptotic shadows of x0 , and satisfy the inequalities ρ(∂∞ H + , ∂∞ H − ) ≥ κ, µ(ψ(Z)∩H + ) ≥ κµ(Z) and µ(ψ(Z)∩
H − ) ≥ κµ(Z).
These properties are satisfied for suitable geometric actions of a hyperbolic group, as we will see in subsection 12.2.
Proposition 12.2. If G is a non-elementary hyperbolic group which
acts geometrically and on a space (X, d, µ) and preserving µ, then for
46
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
any x0 ∈ X and visual metric ρ on ∂∞ X based at x0 with visual parameter ǫ, (X, d, µ) satisfies properties (1)–(4) for suitable δ, C and h(X).
Moreover, (∂∞ X, ρ) is Ahlfors h(X)/ǫ-regular.
Properties (1)–(3) are already known to hold in this generality, so
our efforts will be focused on property (4). Given these properties, we
find the following bounds on the Poincaré profile of X. Note that
Theorem 12.3. Suppose X satisfies conditions (1)–(4) above for some
fixed δ, C, ǫ, κ, N and set Q = h(X)/ǫ. Then we have the following
bounds on ΛpX :
Q−1
if p < Q,
r Q
p−1
1
p
ΛX,a (r) .δ,C,κ,N
r p log(r) p if p = Q,
p−1
if p > Q.
r p
To find the best bound possible for the Poincaré profiles ΛpG of a
hyperbolic group G, it is natural to consider the following concept.
Definition 12.4. The equivariant conformal dimension of a hyperbolic group G is defined to be the infimum of the Hausdorff dimension of (∂∞ X, ρ) where ∂∞ X is the boundary of a space X on which
G acts geometrically and ρ is a visual metric on ∂∞ X. We say the
equivariant conformal dimension is attained if the infimum is realised.
Equivalently, we minimise h(X)/ǫ over all such actions, metrics and
permissible visibility parameters.
Using Proposition 12.2 and Theorem 12.3 we are now ready to prove
Theorem 11.
Corollary 12.5. Let G be a hyperbolic group and let Q be its equivariant conformal dimension. Then, for any ǫ > 0,
( Q−1
r Q +ǫ if p ≤ Q
p
ΛG (r) .
p−1
if p > Q.
r p
If the equivariant conformal dimension
Q−1
r Q
Q−1
1
p
ΛG (r) .
r Q log Q (r)
p−1
r p
is attained, we have:
if 1 ≤ p < Q
if p = Q
if p > Q.
12.1. Helly’s theorem and centrepoints. Inspired by the arguments presented in [BST12, Section 4], we show that finite measure
thick subsets of real hyperbolic spaces have “medians”. To find a suitable centrepoint of a subset, we use Helly’s theorem (cf. [MTTV97]).
POINCARÉ PROFILES OF GROUPS AND SPACES
47
The version suitable for our needs is the following variation on a result
of Ivanov [Iva14].
Theorem 12.6 (Ivanov). Let X be a uniquely geodesic space of compact topological dimension k < ∞ (for example, a CAT(0) space of
geometric dimension k). Let H be a (possibly infinite) collection of
closed convex subsets of X, with the property that there exists a compact convex set Y ⊂ X so thatTfor any HT
1 , . . . , Hk+1 ∈ H we have
Y ∩ H1 ∩ · · · ∩ Hk+1 6= ∅. Then H∈H H ⊃ H∈H H ∩ Y 6= ∅.
Proof. If not, then for any y ∈ Y there exists Hy ∈ H with y ∈
/ Hy .
Since Y is compact, for some y1 , . . . , ym we have that {X \ Hyi }i=1,...,m
is a finite subcover of the open cover {X \Hy }y∈Y of Y . By assumption,
any k + 1 of the finite collection of convex sets {Y, Hy1 , . . . , Hym } have
non-empty intersection (in Y ), and so Helly’s Theorem [Iva14, Theorem 1.1] implies that there exists y ∈ Y ∩ Hy1 ∩ · · · ∩ Hym 6= ∅. This is
a contradiction, since y is not covered by {X \ Hyi }i=1,...,m .
Lemma 12.7. (Centrepoint theorem) Let a > 0. There exists a constant c = c(k, a) > 0 such that for any a-thick subset Z of HkR with
finite measure, there is a point x ∈ HkR such that for any half-space H
of HkR containing x, we have µ(H ∩ Z) ≥ cµ(Z).
S
Proof. By assumption Z = i∈I B(zi , a) for some {zi }i∈I ⊂ Z. Let Z ′
be an 2a-separated 4a-net in {zi : i ∈ I}. It follows that |Z ′ | ≍a µ(Z)
since |Z ′ | µ(B(zi , a)) ≤ µ(Z) ≤ |Z ′ | µ(B(zi , 5a)) for some (any) zi .
Let Y be a large closed (convex) ball containing Z ′ . Let Z be the
k
set of all closed half-spaces of HkR containing more than k+1
|Z ′ | of the
points in Z ′ . Thus the intersection of any k + 1 of the sets in Z has
non-empty intersection with Y .
k
Applying Theorem 12.6, and
T the fact that HR has geometric dimension k, there exists some x ∈ H∈Z H. Thus for any half-space H ⊂ HkR
k
|Z ′ | we have x ∈ H. It is a short exercise to see
with |H ∩ Z ′ | > k+1
k
that x is contained in every half-space H such that |Z ′ ∩ H| > k+1
|Z ′ | if
1
and only if every half-space H containing x satisfies |Z ′ ∩ H| > k+1
|Z ′ |.
′
Let H be a half-space containing x and let ZH
= Z ′ ∩ H. It is clear
1
′
that µ(B(z, r) ∩ H) ≥ 2 µ(B(z, r)) for any z ∈ ZH and any r ≥ 0, so
µ(Z ∩ H) ≥
µ(B(z, a)) ′
|Z | ≍k,a µ(Z).
2(k + 1)
We can use a measure-preserving isometry to move such a centrepoint
x to the origin o ∈ HkR in the Poincaré ball model, and now show that
hypothesis (4) of Theorem 12.3 is satisfied for HkR .
48
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
Lemma 12.8. There exist constants κ, C > 0 so that for any athick subset Z ⊂ HkR , and o ∈ HkR a centrepoint of Z, there exist
C-asymptotic shadows of o denoted by H − , H + ⊂ HkR so that we have
ρ(∂∞ H − , ∂∞ H + ) ≥ κ and that µ(Z ∩ H − ), µ(Z ∩ H + ) ≥ κµ(Z).
Proof. Fix a > 0 and c = c(k, a) > 0 the constants from Lemma 12.7.
Let H ⊂ HkR be a hyperplane containing o, and let α > 0. We denote
by H α the union of all two-sided geodesics passing through o and with
end points in the α-neighbourhood of the boundary ∂∞ H ⊂ ∂∞ HkR =
Sk−1 .
We start with an argument inspired by the proof of [BST12, Proposition 4.1]. Consider for every r > 0 the sphere Sr = {x ∈ HkR , d(x, o) =
r} equipped with its Riemannian measure νr . Note that
νr (Sr ∩ H α ) = η(α)νr (Sr )
for some increasing function η satisfying limα→0 η(α) = 0. We now fix
α > 0 so that η(α) ≤ 2c .
Recall that hyperplanes passing through o are characterized by their
normal vector at o, and therefore are parametrized by the projective
space P k−1 . We consider the Lebesgue probability measure ν on P k−1.
Given θ ∈ P k−1 we define Hθ to be the hyperplane through o with
normal vector θ. Recall that Z ⊂ HkR is a measurable subset of finite
measure, so for each r
Z
νr (Sr ∩ H α )
νr (Z ∩ Hθα ∩ Sr )dν(θ) = νr (Z ∩ Sr )
= νr (Z ∩ Sr )η(α).
νr (Sr )
P k−1
Integrating over r, we deduce that
Z
µ(Z ∩ Hθα )dν(θ) = µ(Z)η(α),
P k−1
and so for some hyperplane HZ we have µ(Z ∩ HZα ) ≤ µ(Z)η(α) ≤
c
µ(Z).
2
Let H − , H + be the two connected components of the complement of
α
HZ ; these are convex and asymptotic shadows of o, and satisfy µ(H − ∩
Z), µ(H + ∩ Z) ≥ 2c µ(Z). Moreover, ρ(∂∞ H − , ∂∞ H + ) ≥ 2α.
12.2. Hyperbolic groups and centrepoints. In this subsection, we
prove Proposition 12.2.
Proof of Proposition 12.2. Property (1) follows from a standard argument with the Arzela–Ascoli theorem, see e.g. [BS00]. Property (2)
follows from [Coo93, Theorem 7.2], and (∂∞ X, ρ) is Ahlfors Q-regular
with Q = 1ǫ h(X). Property (3) is the definition of a visual metric, so
it remains only to show that property (4) is satisfied.
POINCARÉ PROFILES OF GROUPS AND SPACES
49
We require a probably well-known basic fact about convex hulls of
quasi-convex subsets of real hyperbolic spaces. Recall that a subset
Y of a geodesic metric space is K-quasi-convex if every geodesic that
connect a pair of points of Y lies within the K-neighbourhood of Y .
It turns out that in real hyperbolic spaces, quasi-convex subsets are
“nearly” convex in a stronger sense:
Lemma 12.9. Given K ≥ 0, there exists N = N(K, k) such that for
every K-quasi-convex subset Z ⊂ HkR , the convex hull of Z is contained
in the N-neighbourhood of Z.
Proof. Note that in Klein model of HkR , the hyperbolic convex hull
coincides with the Euclidean one. By Carathéodory’s theorem, we
deduce that any point of the convex hull of Z is a convex combination
of some points z1 , . . . , zm ∈ Z, with m ≤ k + 1. Using the quasiconvexity of Z, the lemma follows by induction on m.
We now show that (4) holds for X. Let X be a δX -hyperbolic Cayley
graph of the hyperbolic group G. By a result of Bonk–Schramm [BS00],
there exists constants k ∈ N, λψ ≥ 1, Cψ ≥ 0 and a (λψ , Cψ )-quasiisometric embedding ψ : X → HkR . By post-composing ψ with an
appropriate element of Isomµ (HkR ) if necessary, we may assume ψ(1) =
o, the origin in the Poincaré ball model of HkR .
Given a finite subset Y of V X, define Y ′ ⊂ HkR to be the closed
2-neighbourhood of ψ(Y ). By Lemma 12.7, there is a constant c =
c(k) > 0 and a point x′ ∈ HkR such that for any half-space H of HkR
containing X we have µ(H ∩ Y ′ ) ≥ cµ(Y ′ ). Such x′ is contained in the
convex hull of ψ(Y ), so by Lemma 12.9, dHkR (x′ , ψ(x)) ≤ N(k) for some
x ∈ X. By applying a left-translation in G (by an element g) we may
assume x = 1, while by applying an isometry φ ∈ Isom(HkR ), we may
assume x′ = o. Define f = φ ◦ ψ ◦ g −1 : X → HkR and let ∂∞ f be the
induced map ∂∞ f : ∂∞ X → Sk−1 , where ∂∞ X is endowed with a visual
metric ρ based at 1 and Sk−1 = ∂∞ HkR is endowed with the Euclidean
(visual) metric ρEuc .
By Lemma 12.8, there exist constants κ, C and C-asymptotic shadows of o denoted H ± so that ρEuc (∂∞ H − , ∂∞ H + ) ≥ 4κ and that
µ(Y ′ ∩ H ± ) ≥ 4κµ(Y ′ ). Since f (1) = o, it follows that
ρ(∂∞ f −1 [∂∞ H − ]κ , ∂∞ f −1 [∂∞ H + ]κ ) ≥ κ′
for some κ′ > 0 which does not depend on the choices of φ and g used
to construct f . (It is not a priori obvious that either of ∂∞ f −1 [∂∞ H ± ]κ
is non-empty.)
50
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
±
Define HX
to be the set of all points y ∈ X \ B(1, R) contained in
the A-neighbourhood of the set of all geodesic rays in X from 1 to a
point in ∂∞ f −1 ([∂∞ H ± ]κ ), where A and R are determined below.
We claim that there exist A, R so that if y ∈ Y satisfies dX (1, y) ≥
±
R and B(f (y), 2) ∩ H ± 6= ∅, then y ∈ HX
. Let z ∈ H ± satisfy
dHkR (z, f (y)) ≤ 2, and let γ be the unique geodesic ray in HkR starting at
o and containing z (we assume y is sufficiently far from 1 that z 6= o);
denote the boundary point of γ by ζ. Since X is CX -visual for some
CX , there exists a CX -quasi-geodesic ray β in X from 1 that contains
y; denote by η the boundary point of β in ∂∞ X. The Gromov product
of ζ and ∂∞ f (η) (relative to o) is bounded from below by dHkR (o, f (y))
up to a uniform additive error, so by insisting that dX (1, y) ≥ R is
sufficiently large, we may assume that ρEuc (ζ, ∂∞ f (η)) ≤ κ, hence
∂∞ f (η) ∈ [∂∞ H ± ]κ . By the Morse Lemma, β is contained in a uniform neighbourhood of a geodesic ray from 1 to η, and hence for a
±
suitable choice of A will be contained in HX
outside B(1, R). For these
±
choices of R, A we have that y ∈ HX as desired.
From this, and the fact that f is a quasi-isometry with fixed con±
stants, it follows that there exist η, η ′ > 0 so that |Y ∩ HX
| ≥ ηµ(Y ′ ∩
±
′
′
H ) ≥ ηκµ(Y ) ≥ ηκη |Y |.
The proof of Proposition 12.2 is complete.
12.3. Upper bounds for the Poincaré profile.
Proof of Theorem 12.3. Let x0 ∈ X and a ≥ 2 be fixed so that (4)
holds. Let Z be an a-thick subspace of X of sufficiently large finite
measure (to be determined later). Apply (4) to move Z; without loss
of generality we may assume that ψ = id. Let H ± be the corresponding
C-asymptotic shadows of x0 .
Define ∂∞ φ : (∂∞ X, ρ) → [0, 1] by
∂∞ φ(z) = min{1, max{0, κ3 ρ(z, ∂∞ H − ) − 1}};
this is a κ3 -Lipschitz function so that ∂∞ φ is zero on [∂∞ H − ]κ/3 and one
on [∂∞ H + ]κ/3 .
We choose a function φ : X → [0, 1] by setting φ(x) = ∂∞ φ(η) where
η ∈ ∂∞ X is the endpoint of some C-quasi-geodesic γx : [0, ∞) → X
with γx (0) = x0 and γx (t) = x for some t. Regardless of the choices
made in defining this function we have the following control: for any
x, y ∈ X with d(x, y) ≤ C ′ there exists K = K(δ, C, C ′, ρ, κ) so that
(12.10)
|φ(x) − φ(y)| ≤ K exp(−ǫd(x, x0 )).
POINCARÉ PROFILES OF GROUPS AND SPACES
51
By a similar argument, there exists L > 0 so that if d(x, x0 ) ≥ L
and x ∈ H − then the endpoint η of γx used to define φ(x) satisfies ρ(η, ∂∞ H − ) ≤ κ/3, and so φ(x) = 0. Likewise, if x ∈ H + and
d(x, x0 ) ≥ L then φ(x) = 1.
By assuming that µ(Z) is greater than κ2 µ(B(x0 , L)), we know—by
assumption (4)—that µ(Z ∩ H − \ B(x0 , L)) and µ(Z ∩ H + \ B(x0 , L))
are both ≥ κ2 µ(Z). Switching the roles of H ± if necessary, we assume
φZ ≥ 1/2 and so
(12.11)
||φ − φZ ||pZ,p ≥ |φZ |p µ(Z ∩ H − \ B(x0 , L)) ≥ 2−p−1κµ(Z).
We now bound ||∇a φ||B,p on the ball B = B(x0 , r). Since we have
µ(B(x0 , R)) ≍ exp(h(X)R), (12.10) gives
Z r
p
(12.12)
||∇a φ||B,p K,κ,p
exp(h(X)t) exp(−pǫt)dt.
t=0
We now consider the three cases for p separately. (Recall that
h(X) = ǫQ.)
Case 1, p > Q: Equation (12.12) gives that ||∇a φ||pX,p is bounded
by some constant D only depending on K, κ and p, so (12.11) gives
hpa (Z) K,κ,p µ(Z)−1/p for any subspace Z and the case p > Q follows.
Case 2, p < Q: The function φ is no longer a p-Dirichlet, but its gradient is well-behaved. Indeed, (12.10) gives |∇a φ|(x) exp(−ǫd(x, x0 )),
so
(12.13)
k∇a φkpB,p k exp(−ǫd(·, x0 ))kpZ,p.
Now we wish to put an upper bound on ||∇a φ||pZ,p / ||φ − φZ ||pZ,p , which
by (12.11) is bounded by ||∇a φ||pZ,p /µ(Z) up to a uniform multiplicative
error. Thus by (12.13) it suffices to maximize ||exp(−ǫd(·, x0 ))||pZ,p
among all sets Z with the same measure.
But, up to a uniform multiplicative error, ||exp(−ǫd(·, x0 ))||pZ,p is
maximised when d(·, x0 ) is minimised as a function from Z to R.
Clearly this occurs when Z is a metric ball centred at x0 .
By (12.12), for Z = B(x0 , r),
(12.14)
||∇a φ||pZ,p exp(h(X)r) · exp(−pǫr) ≍ µ(Z) · µ(Z)−p/Q ,
thus hpa (Z) µ(Z)−1/Q and the bound on ΛpX,a (µ(Z)) follows.
Case 3, p = Q: If p = Q then the same argument as in Case 2
shows that inequality (12.14) is maximised for a metric ball, so
Z r
p
||∇a φ||Z,p
exp(h(X)t) exp(−pǫt)dt = r ≍ log(µ(Z)),
t=0
52
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
so hpa (Z) log(µ(Z))1/p ·µ(Z)−1/p and thus the bound on ΛpX for p = Q
follows.
13. Applications to buildings and symmetric spaces
We use results from Sections 11 and 12 to calculate Poincaré profiles
of buildings and rank-one symmetric spaces (Theorem 12).
Bourdon and Pajot [BP99] showed that a family of Fuchsian buildings earlier studied by Bourdon [Bou97] have boundaries that admit
1-Poincaré inequalities.
Definition 13.1. Let m ≥ 5, n ≥ 3 be given. Let R be the regular,
right-angled hyperbolic polygon with m sides. Let I = Im,n be the
Fuchsian building where the chambers are isometric to R, each edge
is adjacent to n copies of R, and the vertex links are copies of the
complete bipartite graph with n, n vertices.
The group
Gm,n = hs1 , . . . , sm | sni , [si , si+1 ] ∀ii,
where indices are modulo m, acts cellularly and geometrically on Im,n .
By [BP99, Theorem 1.1], ∂∞ Gm,n = ∂∞ Im,n carries an Ahlfors Qm,n regular metric, where Qm,n = 1 + log(n − 1)/arccosh((m − 2)/m) ∈
(1, ∞), and which admits a 1-Poincaré inequality in the sense of Heinonen–Koskela (Section 11).
The apartments in Im,n are each copies of the hyperbolic plane tiled
by right-angled regular m-gons. As such, they have separation at least
log(r); the boundary geometry lets us find much larger lower bounds.
Theorem 13.2. Given m ≥ 5, n ≥ 3, and p ∈ [1, ∞),
1−1/Qm,n
if p < Qm,n
r
p
1−1/Q
1/Q
m,n
m,n
ΛIm,n (r) ≃p r
log(r)
if p = Qm,n
r 1−1/p
if p > Qm,n .
Proof. The lower bounds follow from Theorem 11.1 for p < Qm,n , Theorem 11.3 for p = Qm,n and Corollary 10.2 for p ≥ Qm,n . The upper
bounds follow from Corollary 12.5.
Finally, we calculate the Poincaré profiles of rank-one symmetric
spaces. The case of p = 1 for HkR is dealt with by [BST12, Proposition
4.1] and Proposition 6.5, but all other cases are new.
Theorem 13.3. Let K ∈ {R, C, H, O} be a real division algebra, and
let X = Hm
K be a rank-one symmetric space for m ≥ 2 (and m = 2
POINCARÉ PROFILES OF GROUPS AND SPACES
53
when K = O). Let Q = (m + 1) dimR K − 2, then
Q−1
if p < Q
r Q
Q−1
1
p
ΛHm (r) ≃
r Q log(r) Q if p = Q
K
p−1
r p
if p > Q
Proof. The boundary of a rank-one symmetric space carries a visual
metric that is Ahlfors Q-regular for the given exponent, and satisfies
a 1-Poincaré inequality. The result then follows from Theorem 11.1,
Theorem 11.3, Corollary 10.2, and Theorem 12.3.
References
[Bas72]
[BL07]
[BM91]
[Bou97]
[Bou12]
[BP99]
[BP03]
[BR13]
[Bre14]
[BS97]
[BS00]
[BST12]
[Cap14]
[Cho80]
[Chu97]
H. Bass. The degree of polynomial growth of finitely generated nilpotent
groups. Proc. London Math. Soc. (3), 25:603–614, 1972.
S. V. Buyalo and N. D. Lebedeva. Dimensions of locally and asymptotically self-similar spaces. Algebra i Analiz, 19(1):60–92, 2007.
M. Bestvina and G. Mess. The boundary of negatively curved groups.
J. Amer. Math. Soc., 4(3):469–481, 1991.
M. Bourdon. Immeubles hyperboliques, dimension conforme et rigidité
de Mostow. Geom. Funct. Anal., 7(2):245–268, 1997.
M. Bourdon. Un théorème de point fixe sur les espaces Lp . Publ. Mat.,
56(2):375–392, 2012.
M. Bourdon and H. Pajot. Poincaré inequalities and quasiconformal
structure on the boundary of some hyperbolic buildings. Proc. Amer.
Math. Soc., 127(8):2315–2324, 1999.
M. Bourdon and H. Pajot. Cohomologie lp et espaces de Besov. J. Reine
Angew. Math., 558:85–108, 2003.
O. Baker and T. R. Riley. Cannon-Thurston maps do not always exist.
Forum Math. Sigma, 1:e3, 11, 2013.
E. Breuillard. Geometry of locally compact groups of polynomial growth
and shape of large balls. Groups Geom. Dyn., 8(3):669–732, 2014.
I. Benjamini and O. Schramm. Every graph with a positive Cheeger
constant contains a tree with a positive Cheeger constant. Geom. Funct.
Anal., 7(3):403–419, 1997.
M. Bonk and O. Schramm. Embeddings of Gromov hyperbolic spaces.
Geom. Funct. Anal., 10(2):266–306, 2000.
I. Benjamini, O. Schramm, and Á. Timár. On the separation profile of
infinite graphs. Groups Geom. Dyn., 6(4):639–658, 2012.
P.-E. Caprace. Automorphism groups of right-angled buildings: simplicity and local splittings. Fund. Math., 224(1):17–51, 2014.
C. Chou. Elementary amenable groups. Illinois J. Math., 24(3):396–
407, 1980.
F. R. K. Chung. Spectral graph theory, volume 92 of CBMS Regional
Conference Series in Mathematics. Published for the Conference Board
of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 1997.
54
[Coo93]
DAVID HUME, JOHN M. MACKAY, AND ROMAIN TESSERA
M. Coornaert. Mesures de Patterson-Sullivan sur le bord d’un espace
hyperbolique au sens de Gromov. Pacific J. Math., 159(2):241–270,
1993.
[Cou00]
T. Coulhon. Random walks and geometry on infinite graphs. In Lecture
notes on analysis in metric spaces (Trento, 1999), Appunti Corsi Tenuti
Docenti Sc., pages 5–36. Scuola Norm. Sup., Pisa, 2000.
[DSC93]
P. Diaconis and L. Saloff-Coste. Comparison theorems for reversible
Markov chains. Ann. Appl. Probab., 3(3):696–730, 1993.
[GL15]
J. T. Gill and M. Lopez. Discrete approximations of metric measure
spaces of controlled geometry. J. Math. Anal. Appl., 431(1):73–98, 2015.
[GMT06] S. Goel, R. Montenegro, and P. Tetali. Mixing time bounds via the
spectral profile. Electron. J. Probab., 11:no. 1, 1–26, 2006.
[Gro81]
M. Gromov. Groups of polynomial growth and expanding maps. Inst.
Hautes Études Sci. Publ. Math., (53):53–73, 1981.
[Gui73]
Y. Guivarc’h. Croissance polynomiale et périodes des fonctions harmoniques. Bull. Soc. Math. France, 101:333–379, 1973.
[Hag06]
F. Haglund. Commensurability and separability of quasiconvex subgroups. Algebr. Geom. Topol., 6:949–1024, 2006.
[Hei74]
E. Heintze. On homogeneous manifolds of negative curvature. Mathematische Annalen, 211:23–34, 1974.
[HK98]
J. Heinonen and P. Koskela. Quasiconformal maps in metric spaces with
controlled geometry. Acta Math., 181(1):1–61, 1998.
[HK00]
P. Hajlasz and P. Koskela. Sobolev met Poincaré. Mem. Amer. Math.
Soc., 145(688):x+101, 2000.
[HP13]
J. Higes and I. Peng. Assouad-Nagata dimension of connected Lie
groups. Math. Z., 273(1-2):283–302, 2013.
[HS97]
I. Holopainen and P. M. Soardi. p-harmonic functions on graphs and
manifolds. Manuscripta Math., 94(1):95–110, 1997.
[Hum17] D. Hume. A continuum of expanders. Fund. Math., 238:143–152, 2017.
[HW99]
T. Hsu and D. T. Wise. On linear and residual properties of graph
products. Michigan Math. J., 46(2):251–259, 1999.
[Iva14]
S. Ivanov. On Helly’s theorem in geodesic spaces. Electron. Res. Announc. Math. Sci., 21:109–112, 2014.
[Jer86]
D. Jerison. The Poincaré inequality for vector fields satisfying
Hörmander’s condition. Duke Math. J., 53(2):503–523, 1986.
[Kle10]
B. Kleiner. A new proof of Gromov’s theorem on groups of polynomial
growth. J. Amer. Math. Soc., 23(3):815–829, 2010.
[Mat97]
J. Matoušek. On embedding expanders into lp spaces. Israel J. Math.,
102:189–197, 1997.
[MT10]
J. M. Mackay and J. T. Tyson. Conformal dimension: theory and application, volume 54 of University Lecture Series. American Mathematical
Society, Providence, RI, 2010.
[MTTV97] G. L. Miller, S.-H. Teng, W. Thurston, and S. A. Vavasis. Separators
for sphere-packings and nearest neighbor graphs. J. ACM, 44(1):1–29,
1997.
[Pan16]
P.
Pansu.
Large
scale
conformal
maps.
Preprint,
page arXiv:1604.01195, 2016.
POINCARÉ PROFILES OF GROUPS AND SPACES
[SC97]
[SC02]
[Spi15]
[Tes07]
[Tes08]
[Var87]
55
L. Saloff-Coste. Lectures on finite Markov chains. In Lectures on probability theory and statistics (Saint-Flour, 1996), volume 1665 of Lecture
Notes in Math., pages 301–413. Springer, Berlin, 1997.
L. Saloff-Coste. Aspects of Sobolev-type inequalities, volume 289 of London Mathematical Society Lecture Note Series. Cambridge University
Press, Cambridge, 2002.
D. Spielman. Spectral Graph Theory. 2015. Available from
http://www.cs.yale.edu/homes/spielman/561.
R. Tessera. Volume of spheres in doubling metric measured spaces and
in groups of polynomial growth. Bull. Soc. Math. France, 135(1):47–64,
2007.
R. Tessera. Large scale Sobolev inequalities on metric measure spaces
and applications. Rev. Mat. Iberoam., 24(3):825–864, 2008.
N. Th. Varopoulos. Fonctions harmoniques sur les groupes de Lie. C.
R. Acad. Sci. Paris Sér. I Math., 304(17):519–521, 1987.
Mathematical Institute, University of Oxford, Oxford, OX2 6GG.
E-mail address: [email protected]
School of Mathematics, University of Bristol, Bristol, BS8 1TX.
E-mail address: [email protected]
Université Paris-Sud, Orsay, France.
E-mail address: [email protected]
| 4math.GR
|
Attention-Based Models for Speech Recognition
Dzmitry Bahdanau
Jacobs University Bremen, Germany
arXiv:1506.07503v1 [cs.CL] 24 Jun 2015
Jan Chorowski
University of Wrocław, Poland
[email protected]
Dmitriy Serdyuk
Université de Montréal
Kyunghyun Cho
Université de Montréal
Yoshua Bengio
Université de Montréal
CIFAR Senior Fellow
Abstract
Recurrent sequence generators conditioned on input data through an attention
mechanism have recently shown very good performance on a range of tasks including machine translation, handwriting synthesis [1, 2] and image caption generation [3]. We extend the attention-mechanism with features needed for speech
recognition. We show that while an adaptation of the model used for machine
translation in [2] reaches a competitive 18.7% phoneme error rate (PER) on the
TIMIT phoneme recognition task, it can only be applied to utterances which are
roughly as long as the ones it was trained on. We offer a qualitative explanation of
this failure and propose a novel and generic method of adding location-awareness
to the attention mechanism to alleviate this issue. The new method yields a model
that is robust to long inputs and achieves 18% PER in single utterances and 20%
in 10-times longer (repeated) utterances. Finally, we propose a change to the attention mechanism that prevents it from concentrating too much on single frames,
which further reduces PER to 17.6% level.
1
Introduction
Recently, attention-based recurrent networks have been successfully applied to a wide variety of
tasks, such as handwriting synthesis [1], machine translation [2], image caption generation [3] and
visual object classification [4].1 Such models iteratively process their input by selecting relevant
content at every step. This basic idea significantly extends the applicability range of end-to-end
training methods, for instance, making it possible to construct networks with external memory [6, 7].
We introduce extensions to attention-based recurrent networks that make them applicable to speech
recognition. Learning to recognize speech can be viewed as learning to generate a sequence (transcription) given another sequence (speech). From this perspective it is similar to machine translation
and handwriting synthesis tasks, for which attention-based methods have been found suitable [2, 1].
However, compared to machine translation, speech recognition principally differs by requesting
much longer input sequences (thousands of frames instead of dozens of words), which introduces a
challenge of distinguishing similar speech fragments2 in a single utterance. It is also different from
handwriting synthesis, since the input sequence is much noisier and does not have as clear structure.
For these reasons speech recognition is an interesting testbed for developing new attention-based
architectures capable of processing long and noisy inputs.
Application of attention-based models to speech recognition is also an important step toward building fully end-to-end trainable speech recognition systems, which is an active area of research. The
1
2
An early version of this work was presented at the NIPS 2014 Deep Learning Workshop [5].
Explained in more detail in Sec. 2.1.
1
dominant approach is still based on hybrid systems consisting of a deep neural acoustic model, a triphone HMM model and an n-gram language model [8, 9]. This requires dictionaries of hand-crafted
pronunciation and phoneme lexicons, and a multi-stage training procedure to make the components
work together. Excellent results by an HMM-less recognizer have recently been reported, with the
system consisting of a CTC-trained neural network and a language model [10]. Still, the language
model was added only at the last stage in that work, thus leaving open a question of how much an
acoustic model can benefit from being aware of a language model during training.
In this paper, we evaluate attention-based models on a phoneme recognition task using the widelyused TIMIT dataset. At each time step in generating an output sequence (phonemes), an attention
mechanism selects or weighs the signals produced by a trained feature extraction mechanism at potentially all of the time steps in the input sequence (speech frames). The weighted feature vector then
helps to condition the generation of the next element of the output sequence. Since the utterances
in this dataset are rather short (mostly under 5 seconds), we measure the ability of the considered
models in recognizing much longer utterances which were created by artificially concatenating the
existing utterances.
We start with a model proposed in [2] for the machine translation task as the baseline. This model
seems entirely vulnerable to the issue of similar speech fragments but despite our expectations it
was competitive on the original test set, reaching 18.7% phoneme error rate (PER). However, its
performance degraded quickly with longer, concatenated utterances. We provide evidence that this
model adapted to track the absolute location in the input sequence of the content it is recognizing, a
strategy feasible for short utterances from the original test set but inherently unscalable.
In order to circumvent this undesired behavior, in this paper, we propose to modify the attention
mechanism such that it explicitly takes into account both (a) the location of the focus from the
previous step, as in [6] and (b) the features of the input sequence, as in [2]. This is achieved by
adding as inputs to the attention mechanism auxiliary convolutional features which are extracted by
convolving the attention weights from the previous step with trainable filters. We show that a model
with such convolutional features performs significantly better on the considered task (18.0% PER).
More importantly, the model with convolutional features robustly recognized utterances many times
longer than the ones from the training set, always staying below 20% PER.
Therefore, the contribution of this work is three-fold. For one, we present a novel purely neural
speech recognition architecture based on an attention mechanism, whose performance is comparable to that of the conventional approaches on the TIMIT dataset. Moreover, we propose a generic
method of adding location awareness to the attention mechanism. Finally, we introduce a modification of the attention mechanism to avoid concentrating the attention on a single frame, and thus
avoid obtaining less “effective training examples”, bringing the PER down to 17.6%.
2
2.1
Attention-Based Model for Speech Recognition
General Framework
An attention-based recurrent sequence generator (ARSG) is a recurrent neural network that stochastically generates an output sequence (y1 , . . . , yT ) from an input x. In practice, x is often processed
by an encoder which outputs a sequential input representation h = (h1 , . . . , hL ) more suitable for
the attention mechanism to work with.
In the context of this work, the output y is a sequence of phonemes, and the input x = (x1 , . . . , xL0 )
is a sequence of feature vectors. Each feature vector is extracted from a small overlapping window
of audio frames. The encoder is implemented as a deep bidirectional recurrent network (BiRNN),
to form a sequential representation h of length L = L0 .
At the i-th step an ARSG generates an output yi by focusing on the relevant elements of h:
αi = Attend(si−1 , αi−1 , h)
gi =
L
X
(1)
αi,j hj
(2)
yi ∼ Generate(si−1 , gi ),
(3)
j=1
2
yi+1
yi
si
si-1
si+1
gi
×
αi-1,j-1
×
αi-1,j
hj-1
×
gi+1
αi
αi-1,j+1
×
αi,j-1
×
αi,j
×
αi,j+1
hj+1
hj
Figure 1:
Two steps of the proposed attention-based recurrent sequence generator (ARSG) with a hybrid attention mechanism (computing
α), based on both content (h) and location (previous α) information. The
dotted lines correspond to Eq. (1),
thick solid lines to Eq. (2) and dashed
lines to Eqs. (3)–(4).
where si−1 is the (i − 1)-th state of the recurrent neural network to which we refer as the generator, αi ∈ RL is a vector of the attention weights, also often called the alignment [2]. Using the
terminology from [4], we call gi a glimpse. The step is completed by computing a new generator
state:
si = Recurrency(si−1 , gi , yi )
(4)
Long short-term memory units (LSTM, [11]) and gated recurrent units (GRU, [12]) are typically
used as a recurrent activation, to which we refer as a recurrency. The process is graphically illustrated in Fig. 1.
Inspired by [6] we distinguish between location-based, content-based and hybrid attention mechanisms. Attend in Eq. (1) describes the most generic, hybrid attention. If the term αi−1 is dropped
from Attend arguments, i.e., αi = Attend(si−1 , h), we call it content-based (see, e.g., [2] or [3]).
In this case, Attend is often implemented by scoring each element in h separately and normalizing
the scores:
αi,j
ei,j = Score(si−1 , hj ),
, L
X
= exp(ei,j )
exp(ei,j ) .
(5)
(6)
j=1
The main limitation of such scheme is that identical or very similar elements of h are scored equally
regardless of their position in the sequence. This is the issue of “similar speech fragments” raised
above. Often this issue is partially alleviated by an encoder such as e.g. a BiRNN [2] or a deep
convolutional network [3] that encode contextual information into every element of h . However,
capacity of h elements is always limited, and thus disambiguation by context is only possible to a
limited extent.
Alternatively, a location-based attention mechanism computes the alignment from the generator
state and the previous alignment only such that αi = Attend(si−1 , αi−1 ). For instance, Graves
[1] used the location-based attention mechanism using a Gaussian mixture model in his handwriting
synthesis model. In the case of speech recognition, this type of location-based attention mechanism
would have to predict the distance between consequent phonemes using si−1 only, which we expect
to be hard due to large variance of this quantity.
For these limitations associated with both content-based and location-based mechanisms, we argue
that a hybrid attention mechanism is a natural candidate for speech recognition. Informally, we
would like an attention model that uses the previous alignment αi−1 to select a short list of elements
from h, from which the content-based attention, in Eqs. (5)–(6), will select the relevant ones without
confusion.
2.2
Proposed Model: ARSG with Convolutional Features
We start from the ARSG-based model with the content-based attention mechanism proposed in [2].
This model can be described by Eqs. (5)–(6), where
ei,j = w> tanh(W si−1 + V hj + b).
w and b are vectors, W and V are matrices.
3
(7)
We extend this content-based attention mechanism of the original model to be location-aware by
making it take into account the alignment produced at the previous step. First, we extract k vectors
fi,j ∈ Rk for every position j of the previous alignment αi−1 by convolving it with a matrix
F ∈ Rk×r :
fi = F ∗ αi−1 .
(8)
These additional vectors fi,j are then used by the scoring mechanism ei,j :
ei,j = w> tanh(W si−1 + V hj + U fi,j + b)
2.3
(9)
Score Normalization: Sharpening and Smoothing
There are three potential issues with the normalization in Eq. (6).
First, when the input sequence h is long, the glimpse gi is likely to contain noisy information from
many irrelevant feature vectors hj , as the normalized scores αi,j are all positive and sum to 1. This
makes it difficult for the proposed ARSG to focus clearly on a few relevant frames at each time
i. Second, the attention mechanism is required to consider all the L frames each time it decodes a
single output yi while decoding the output of length T , leading to a computational complexity of
O(LT ). This may easily become prohibitively expensive, when input utterances are long (and issue
that is less serious for machine translation, because in that case the input sequence is made of words,
not of 20ms acoustic frames).
The other side of the coin is that the use of softmax normalization in Eq. (6) prefers to mostly focus
on only a single feature vector hj . This prevents the model from aggregating multiple top-scored
frames to form a glimpse gi .
Sharpening There is a straightforward way to address the first issue of a noisy glimpse by “sharpening” the scores αi,j . One way to sharpen the weights is to introduce an inverse temperature β > 1
to the softmax function such that
, L
X
ai,j = exp(βei,j )
exp(βei,j ) ,
j=1
or to keep only the top-k frames according to the scores and re-normalize them. These sharpening
methods, however, still requires us to compute the score of every frame each time (O(LT )), and
they worsen the second issue, of overly narrow focus.
We also propose and investigate a windowing technique. At each time i, the attention mechanism
considers only a subsequence h̃ = (hpi −w , . . . , hpi +w−1 ) of the whole sequence h, where w L is
the predefined window width and pi is the median of the alignment αi−1 . The scores for hj ∈
/ h̃ are
not computed, resulting in a lower complexity of O(L + T ). This windowing technique is similar
to taking the top-k frames, and similarly, has the effect of sharpening.
The proposed sharpening based on windowing can be used both during training and evaluation.
Later, in the experiments, we only consider the case where it is used during evaluation.
Smoothing We observed that the proposed sharpening methods indeed helped with long utterances. However, all of them, and especially selecting the frame with the highest score, negatively
affected the model’s performance on the standard development set which mostly consists of short utterances. This observations let us hypothesize that it is helpful for the model to aggregate selections
from multiple top-scored frames. In a sense this brings more diversity, i.e., more effective training
examples, to the output part of the model, as more input locations are considered. To facilitate this
effect, we replace the unbounded exponential function of the softmax function in Eq. (6) with the
bounded logistic sigmoid σ such that
, L
X
ai,j = σ(ei,j )
σ(ei,j ) .
j=1
This has the effect of smoothing the focus found by the attention mechanism.
4
Phoneme Error Rate [%]
Dependency of error rate on beam search width.
Baseline
19
Conv Feats
Smooth Focus
Dataset
18
●
dev
17
16
●
●
●
●
●
●
●
1
2
5
10
20
50
100
●
1
●
2
test
●
●
●
●
●
●
●
●
●
●
●
●
5
10
20
50
100
1
2
5
10
20
50
100
Beam width
Figure 2: Decoding performance w.r.t. the beam size. For rigorous comparison, if decoding failed
to generate heosi, we considered it wrongly recognized without retrying with a larger beams size.
The models, especially with smooth focus, perform well even with a beam width as small as 1.
3
Related Work
Speech recognizers based on the connectionist temporal classification (CTC, [13]) and its extension,
RNN Transducer [14], are the closest to the ARSG model considered in this paper. They follow
earlier work on end-to-end trainable deep learning over sequences with gradient signals flowing
through the alignment process [15]. They have been shown to perform well on the phoneme recognition task [16]. Furthermore, the CTC was recently found to be able to directly transcribe text from
speech without any intermediate phonetic representation [17].
The considered ARSG is different from both the CTC and RNN Transducer in two ways. First,
whereas the attention mechanism deterministically aligns the input and the output sequences, the
CTC and RNN Transducer treat the alignment as a latent random variable over which MAP (maximum a posteriori) inference is performed. This deterministic nature of the ARSG’s alignment
mechanism allows beam search procedure to be simpler. Furthermore, we empirically observe that
a much smaller beam width can be used with the deterministic mechanism, which allows faster
decoding (see Sec. 4.2 and Fig. 2). Second, the alignment mechanism of both the CTC and RNN
Transducer is constrained to be “monotonic” to keep marginalization of the alignment tractable. On
the other hand, the proposed attention mechanism can result in non-monotonic alignment, which
makes it suitable for a larger variety of tasks other than speech recognition.
A hybrid attention model using a convolution operation was also proposed in [6] for neural Turing
machines (NTM). At each time step, the NTM computes content-based attention weights which are
then convolved with a predicted shifting distribution. Unlike the NTM’s approach, the hybrid mechanism proposed here lets learning figure out how the content-based and location-based addressing
be combined by a deep, parametric function (see Eq. (9).)
Sukhbaatar et al. [18] describes a similar hybrid attention mechanism, where location embeddings
are used as input to the attention model. This approach has an important disadvantage that the model
cannot work with an input sequence longer than those seen during training. Our approach, on the
other hand, works well on sequences many times longer than those seen during training (see Sec. 5.)
4
Experimental Setup
We closely followed the procedure in [16]. All experiments were performed on the TIMIT corpus
[19]. We used the train-dev-test split from the Kaldi [20] TIMIT s5 recipe. We trained on the
standard 462 speaker set with all SA utterances removed and used the 50 speaker dev set for early
stopping. We tested on the 24 speaker core test set. All networks were trained on 40 mel-scale filterbank features together with the energy in each frame, and first and second temporal differences,
yielding in total 123 features per frame. Each feature was rescaled to have zero mean and unit
variance over the training set. Networks were trained on the full 61-phone set extended with an
extra “end-of-sequence” token that was appended to each target sequence. Similarly, we appended
an all-zero frame at the end of each input sequence to indicate the end of the utterance. Decoding
was performed using the 61+1 phoneme set, while scoring was done on the 39 phoneme set.
4.1
Training Procedure
One property of ARSG models is that different subsets of parameters are reused different number of
times; L times for those of the encoder, LT for the attention weights and T times for all the other
5
FDHC0_SX209: Michael colored the bedroom wall with crayons.
h#
m
ay
kcl k
kcl k el
ah
l
er
dcl d r
bcl b eh
ix
dh
dcl
m
ux
w
ao
th kcl k
l w ix
r
ey
aa
n
s
h#
Figure 3: Alignments produced by the baseline model. The vertical bars indicate ground truth
phone location from TIMIT. Each row of the upper image indicates frames selected by the attention
mechanism to emit a phone symbol. The network has clearly learned to produce a left-to-right
alignment with a tendency to look slightly ahead, and does not confuse between the repeated “kclk” phrase. Best viewed in color.
parameters of the ARSG. This makes the scales of derivatives w.r.t. parameters vary significantly,
and we handle it by using an adaptive learning rate algorithm, AdaDelta [21] which has two hyperparameters and ρ. All the weight matrices were initialized from a normal Gaussian distribution
with its standard deviation set to 0.01. Recurrent weights were furthermore orthogonalized.
As TIMIT is a relatively small dataset, proper regularization is crucial. We used the adaptive weight
noise as a main regularizer [22]. We first trained our models with a column norm constraint [23] with
the maximum norm 1 until the lowest development negative log-likelihood is achieved.3 During this
time, and ρ are set to 10−8 and 0.95, respectively. At this point, we began using the adaptive
weight noise, and scaled down the model complexity cost LC by a factor of 10, while disabling
the column norm constraints. Once the new lowest development log-likelihood was reached, we
fine-tuned the model with a smaller = 10−10 , until we did not observe the improvement in the
development phoneme error rate (PER) for 100K weight updates. Batch size 1 was used throughout
the training.
4.2
Details of Evaluated Models
We evaluated the ARSGs with different attention mechanisms. The encoder was a 3-layer BiRNN
with 256 GRU units in each direction, and the activations of the 512 top-layer units were used as the
representation h. The generator had a single recurrent layer of 256 GRU units. Generate in Eq. (3)
had a hidden layer of 64 maxout units. The initial states of both the encoder and generator were
treated as additional parameters.
Our baseline model is the one with a purely content-based attention mechanism (See Eqs. (5)–(7).)
The scoring network in Eq. (7) had 512 hidden units. The other two models use the convolutional
features in Eq. (8) with k = 10 and r = 201. One of them uses the smoothing from Sec. 2.3.
Decoding Procedure A left-to-right beam search over phoneme sequences was used during decoding [24]. Beam search was stopped when the “end-of-sequence” token heosi was emitted. We
started with a beam width of 10, increasing it up to 40 when the network failed to produce heosi with
the narrower beam. As shown in Fig. 2, decoding with a wider beam gives little-to-none benefit.
5
Results
All the models achieved competitive PERs (see Table 1). With the convolutional features, we see
3.7% relative improvement over the baseline and further 5.9% with the smoothing.
To our surprise (see Sec. 2.1.), the baseline model learned to align properly. An alignment produced by the baseline model on a sequence with repeated phonemes (utterance FDHC0 SX209) is
presented in Fig. 3 which demonstrates that the baseline model is not confused by short-range repetitions. We can also see from the figure that it prefers to select frames that are near the beginning or
3
Applying the weight noise from the beginning of training caused severe underfitting.
6
Table 1: Phoneme error rates (PER). The bold-faced PER corresponds to the best error rate with an
attention-based recurrent sequence generator (ARSG) incorporating convolutional attention features
and a smooth focus.
Model
Dev
Test
Baseline Model
15.9% 18.7%
16.1% 18.0%
Baseline + Conv. Features
Baseline + Conv. Features + Smooth Focus
15.8% 17.6%
RNN Transducer [16]
N/A
17.7%
HMM over Time and Frequency Convolutional Net [25] 13.9% 16.7%
Figure 4: Results of force-aligning the concatenated utterances. Each dot represents a single utterance created by either concatenating multiple copies of the same utterance, or of different, randomly
chosen utterances. We clearly see that the highest robustness is achieved when the hybrid attention
mechanism is combined with the proposed sharpening technique (see the bottom-right plot.)
even slightly before the phoneme location provided as a part of the dataset. The alignments produced
by the other models were very similar visually.
5.1
Forced Alignment of Long Utterances
The good performance of the baseline model led us to the question of how it distinguishes between
repetitions of similar phoneme sequences and how reliably it decodes longer sequences with more
repetitions. We created two datasets of long utterances; one by repeating each test utterance, and the
other by concatenating randomly chosen utterances. In both cases, the waveforms were cross-faded
with a 0.05s silence inserted as the “pau” phone. We concatenated up to 15 utterances.
First, we checked the forced alignment with these longer utterances by forcing the generator to emit
the correct phonemes. Each alignment was considered correct if 90% of the alignment weight lies
inside the ground-truth phoneme window extended by 20 frames on each side. Under this definition,
all phones but the heosi shown in Fig. 3 are properly aligned.
The first column of Fig. 4 shows the number of correctly aligned frames w.r.t. the utterance length
(in frames) for some of the considered models. One can see that the baseline model was able to
decode sequences up to about 120 phones when a single utterance was repeated, and up to about 150
phones when different utterances were concatenated. Even when it failed, it correctly aligned about
50 phones. On the other hand, the model with the hybrid attention mechanism with convolutional
features was able to align sequences up to 200 phones long. However, once it began to fail, the
model was not able to align almost all phones. The model with the smoothing behaved similarly to
the one with convolutional features only.
We examined failed alignments to understand these two different modes of failure. Some of the
examples are shown in the Supplementary Materials.
We found that the baseline model properly aligns about 40 first phones, then makes a jump to the
end of the recording and cycles over the last 10 phones. This behavior suggests that it learned to
track its approximate location in the source sequence. However, the tracking capability is limited to
the lengths observed during training. Once the tracker saturates, it jumps to the end of the recording.
7
●
Phoneme error rates on long utterances
Baseline
Phoneme error rate [%]
24
Decoding algorithm
Conv Feats
Smooth Focus
Keep 20
●
●
●
●
22
●
●
●
●
●
●
●
●
Keep 50
●
Win ± 150
Win ± 75
●
●
●
20
●
●
●
18
●
●
●
●
●
●
●
●
●
●
●
●
●
●
3
6
9
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
Dataset
●
●
3
6
9
●
Mixed Utt.
3
6
9
Same Utt.
Number of repetitions
Figure 5: Phoneme error rates obtained on decoding long sequences. Each network was decoded with alignment sharpening techniques that produced proper forced alignments. The proposed
ARSG’s are clearly more robust to the length of the utterances than the baseline one is.
In contrast, when the location-aware network failed it just stopped aligning – no particular frames
were selected for each phone. We attribute this behavior to the issue of noisy glimpse discussed in
Sec. 2.3. With a long utterance there are many irrelevant frames negatively affecting the weight assigned to the correct frames. In line with this conjecture, the location-aware network works slightly
better on the repetition of the same utterance, where all frames are somehow relevant, than on the
concatenation of different utterances, where each misaligned frame is irrelevant.
To gain more insight we applied the alignment sharpening schemes described in Sec. 2.3. In the
remaining columns of Fig. 4, we see that the sharpening methods help the location-aware network
to find proper alignments, while they show little effect on the baseline network. The windowing
technique helps both the baseline and location-aware networks, with the location-aware network
properly aligning nearly all sequences.
During visual inspection, we noticed that in the middle of very long utterances the baseline model
was confused by repetitions of similar content within the window, and that such confusions did not
happen in the beginning. This supports our conjecture above.
5.2
Decoding Long Utterances
We evaluated the models on long sequences. Each model was decoded using the alignment sharpening techniques that helped to obtain proper forced alignments. The results are presented in Fig. 5.
The baseline model fails to decode long utterances, even when a narrow window is used to constrain
the alignments it produces. The two other location-aware networks are able to decode utterances
formed by concatenating up to 11 test utterances. Better results were obtained with a wider window,
presumably because it resembles more the training conditions when at each step the attention mechanism was seeing the whole input sequence. With the wide window, both of the networks scored
about 20% PER on the long utterances, indicating that the proposed location-aware attention mechanism can scale to sequences much longer than those in the training set with only minor modifications
required at the decoding stage.
6
Conclusions
We proposed and evaluated a novel end-to-end trainable speech recognition architecture based on a
hybrid attention mechanism which combines both content and location information in order to select
the next position in the input sequence for decoding. One desirable property of the proposed model
is that it can recognize utterances much longer than the ones it was trained on. In the future, we
expect this model to be used to directly recognize text from speech [10, 17], in which case it may
become important to incorporate a monolingual language model to the ARSG architecture [26].
This work has contributed two novel ideas for attention mechanisms: a better normalization approach yielding smoother alignments and a generic principle for extracting and using features from
the previous alignments. Both of these can potentially be applied beyond speech recognition. For
instance, the proposed attention can be used without modification in neural Turing machines, or by
using 2–D convolution instead of 1–D, for improving image caption generation [3].
8
Acknowledgments
All experiments were conducted using Theano [27, 28], PyLearn2 [29], and Blocks [30] libraries.
The authors would like to acknowledge the support of the following agencies for research funding and computing support: National Science Center (Poland), NSERC, Calcul Québec, Compute
Canada, the Canada Research Chairs and CIFAR. Bahdanau also thanks Planet Intelligent Systems
GmbH and Yandex.
References
[1] Alex Graves. Generating sequences with recurrent neural networks. arXiv:1308.0850, August 2013.
[2] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning
to align and translate. arXiv:1409.0473, September 2014.
[3] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhutdinov, Richard
Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention.
arXiv:1502.03044, February 2015.
[4] Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. Recurrent models of visual attention. In Advances
in Neural Information Processing Systems, pages 2204–2212, 2014.
[5] Jan Chorowski, Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. End-to-end continuous speech
recognition using attention-based recurrent NN: First results. arXiv:1412.1602 [cs, stat], December 2014.
[6] Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing machines. arXiv:1410.5401, 2014.
[7] Jason Weston, Sumit Chopra, and Antoine Bordes. Memory networks. arXiv:1410.3916, 2014.
[8] Mark Gales and Steve Young. The application of hidden markov models in speech recognition. Found.
Trends Signal Process., 1(3):195–304, January 2007.
[9] G. Hinton, Li Deng, Dong Yu, G.E. Dahl, A Mohamed, N. Jaitly, A Senior, V. Vanhoucke, P. Nguyen,
T.N. Sainath, and B. Kingsbury. Deep neural networks for acoustic modeling in speech recognition: The
shared views of four research groups. IEEE Signal Processing Magazine, 29(6):82–97, November 2012.
[10] Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger,
Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deepspeech: Scaling up end-to-end speech
recognition. arXiv preprint arXiv:1412.5567, 2014.
[11] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural. Comput., 9(8):1735–1780, 1997.
[12] Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Fethi Bougares, Holger Schwenk, and Yoshua
Bengio. Learning phrase representations using RNN encoder-decoder for statistical machine translation.
In EMNLP 2014, October 2014. to appear.
[13] Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. Connectionist temporal
classification: Labelling unsegmented sequence data with recurrent neural networks. In ICML-06, 2006.
[14] Alex Graves. Sequence transduction with recurrent neural networks. In ICML-12, 2012.
[15] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient based learning applied to document recognition.
Proc. IEEE, 1998.
[16] Alex Graves, Abdel-rahman Mohamed, and Geoffrey Hinton. Speech recognition with deep recurrent
neural networks. In ICASSP 2013, pages 6645–6649. IEEE, 2013.
[17] Alex Graves and Navdeep Jaitly. Towards end-to-end speech recognition with recurrent neural networks.
In ICML-14, pages 1764–1772, 2014.
[18] Sainbayar Sukhbaatar, Arthur Szlam, Jason Weston, and Rob Fergus. Weakly supervised memory networks. arXiv preprint arXiv:1503.08895, 2015.
[19] J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren. DARPA TIMIT
acoustic phonetic continuous speech corpus, 1993.
[20] Daniel Povey, Arnab Ghoshal, Gilles Boulianne, Lukas Burget, Ondrej Glembek, Nagendra Goel, Mirko
Hannemann, Petr Motlicek, Yanmin Qian, Petr Schwarz, and others. The kaldi speech recognition toolkit.
In Proc. ASRU, pages 1–4, 2011.
[21] Matthew D Zeiler. ADADELTA: An adaptive learning rate method. arXiv:1212.5701, 2012.
[22] Alex Graves. Practical variational inference for neural networks. In J. Shawe-Taylor, R.S. Zemel, P.L.
Bartlett, F. Pereira, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems
24, pages 2348–2356. Curran Associates, Inc., 2011.
9
[23] Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint
arXiv:1207.0580, 2012.
[24] Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to sequence learning with neural networks.
arXiv preprint arXiv:1409.3215, 2014.
[25] László Tóth. Combining time-and frequency-domain convolution in convolutional neural network-based
phone recognition. In ICASSP 2014, pages 190–194, 2014.
[26] Caglar Gulcehre, Orhan Firat, Kelvin Xu, Kyunghyun Cho, Loic Barrault, Huei-Chi Lin, Fethi Bougares,
Holger Schwenk, and Yoshua Bengio. On using monolingual corpora in neural machine translation. arXiv
preprint arXiv:1503.03535, 2015.
[27] James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: a CPU and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), June 2010.
Oral Presentation.
[28] Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, James Bergstra, Ian J. Goodfellow, Arnaud Bergeron,
Nicolas Bouchard, and Yoshua Bengio. Theano: new features and speed improvements. Deep Learning
and Unsupervised Feature Learning NIPS 2012 Workshop, 2012.
[29] Ian J. Goodfellow, David Warde-Farley, Pascal Lamblin, Vincent Dumoulin, Mehdi Mirza, Razvan Pascanu, James Bergstra, Frédéric Bastien, and Yoshua Bengio. Pylearn2: a machine learning research
library. arXiv preprint arXiv:1308.4214, 2013.
[30] Bart van Merriënboer, Dzmitry Bahdanau, Vincent Dumoulin, Dmitriy Serdyuk, David Warde-Farley, Jan
Chorowski, and Yoshua Bengio. Blocks and fuel: Frameworks for deep learning. arXiv:1506.00619 [cs,
stat], June 2015.
10
A
Additional Figures
Baseline
FDHC0_SX209: Michael colored the bedroom wall with crayons.
h#
m
ay
kcl k
kcl k el
ah
l
er
dcl d r
bcl b eh
dcl dhix
ux
m
w
ao
th kcl k
l w ix
aa
n
h#
ey
s
r
aa
n
h#
ey
s
r
aa
n
h#
ey
s
r
Convolutional Features
FDHC0_SX209: Michael colored the bedroom wall with crayons.
h#
m
ay
kcl k
kcl k el
ah
l
er
dcl d r
bcl b eh
dcl dhix
ux
m
w
ao
th kcl k
l w ix
Smooth Focus
FDHC0_SX209: Michael colored the bedroom wall with crayons.
h#
m
ay
kcl k
kcl k el
ah
l
er
dcl d r
bcl b eh
dcl dhix
ux
m
w
ao
th kcl k
l w ix
Figure 6: Alignments produced by evaluated models on the FDHC0 SX209 test utterance. The vertical bars indicate ground truth phone location from TIMIT. Each row of the upper image indicates
frames selected by the attention mechanism to emit a phone symbol. Compare with Figure 3. in the
main text.
11
Baseline
FAEM0_SI2022: What outfit does she drive for?
h#
dx
w ah
aw
q
f
ix
tcl
d uh
sh
d
iy dcl
v
f
h#
ay
ao
r
v
f
h#
ay
ao
r
v
f
h#
ay
ao
r
Convolutional Features
FAEM0_SI2022: What outfit does she drive for?
h#
dx
w ah
aw
q
f
ix
tcl
d uh
sh
d
iy dcl
Smooth Focus
FAEM0_SI2022: What outfit does she drive for?
h#
dx
w ah
aw
q
f
ix
tcl
d uh
sh
d
iy dcl
Figure 7: Alignments produced by evaluated models on the FAEM0 SI2022 train utterance. The
vertical bars indicate ground truth phone location from TIMIT. Each row of the upper image indicates frames selected by the attention mechanism to emit a phone symbol. Compare with Figure 3.
in the main text.
12
Number of incorrectly aligned phones vs utterance length and model.
Same Utt.
Mixed Utt.
300
Baseline
200
100
0
400
300
Conv Feats
Number of correctly aligned phones
400
200
100
0
0
200
400
600
0
200
400
600
Utterance length [phones]
Figure 8: Close-up on the two failure modes of ARSG. Results of force-aligning concatenated
TIMIT utterances. Each dot represents a single utterance. The left panels show results for concatenations of the same utterance. The right panels show results for concatenations of randomly
chosen utterances. We compare the baseline network having a content-based only attention mechanism (top row) with a hybrid attention mechanism that uses convolutional features (bottom row).
While neither model is able to properly align long sequences, they fail in different ways: the baseline
network always aligns about 50 phones, while the location-aware network fails to align any phone.
Compare with Figure 4 form the main paper.
13
14
Figure 9: The baseline network fails to align more than 3 repetitions of FDHC0 SX209.
15
Figure 10: The baseline network aligns a concatenation of 3 different utterances, but fails to align 5.
16
Figure 11: Forced alignment of 7 repetitions of the phrase “Michael colored” performed with the baseline model with windowing enabled (the alignment was
constrained to ±75 frames from the expected position of the generator at the last step. The window is wider than the pattern and the net confuses similar content.
Strangely, the first two repetitions are aligned without any confusion with subsequent ones – the network starts to confound phoneme location only starting from
the third repetition (as seen by the parallel strand of alignment which starts when the network starts to emit the phrase for the third time).
17
Figure 12: The location-aware network correctly aligns 7 and 11 repetitions of FDHC0 SX209, butfails to align 15 repetitions of FDHC0 SX209.
18
Figure 13: The location-aware network aligns a concatenation of 3 different utterances, but fails to align 5.
B
Detailed results of experiments
Table 2: Phoneme error rates while decoding with various modifications. Compare with Figure 5
from the main paper.
Plain Keep 1 Keep 10 Keep 50 β = 2 Win. ±75 Win. ±150
dev 15.9% 17.6% 15.9%
15.9% 16.1%
15.9%
15.9%
Baseline
test 18.7% 20.2% 18.7%
18.7% 18.9%
18.7%
18.6%
dev 16.1% 19.4% 16.2%
16.1% 16.7%
16.0%
16.1%
Conv Feats
18.0% 18.7%
18.0%
18.0%
test 18.0% 22.3% 17.9%
dev 15.8% 21.6% 16.5%
16.1% 16.2%
16.2%
16.0%
Smooth Focus
test 17.6% 24.7% 18.7%
17.8% 18.4%
17.7%
17.6%
19
| 9cs.NE
|
Model Predictive Path Integral Control using Covariance Variable
Importance Sampling
arXiv:1509.01149v3 [cs.SY] 28 Oct 2015
Grady Williams1 , Andrew Aldrich1 , and Evangelos A. Theodorou1
Abstract— In this paper we develop a Model Predictive Path
Integral (MPPI) control algorithm based on a generalized
importance sampling scheme and perform parallel optimization
via sampling using a Graphics Processing Unit (GPU). The
proposed generalized importance sampling scheme allows for
changes in the drift and diffusion terms of stochastic diffusion
processes and plays a significant role in the performance of the
model predictive control algorithm. We compare the proposed
algorithm in simulation with a model predictive control version
of differential dynamic programming.
I. INTRODUCTION
The path integral optimal control framework [7], [15],
[16] provides a mathematically sound methodology for developing optimal control algorithms based on stochastic
sampling of trajectories. The key idea in this framework is
that the value function for the optimal control problem is
transformed using the Feynman-Kac lemma [2], [8] into an
expectation over all possible trajectories, which is known
as a path integral. This transformation allows stochastic
optimal control problems to be solved with a Monte-Carlo
approximation using forward sampling of stochastic diffusion
processes.
There have been a variety of algorithms developed in the
path integral control setting. The most straight-forward application of path integral control is when the iterative feedback
control law suggested in [15] is implemented in its open
loop formulation. This requires that sampling takes place
only from the initial state of the optimal control problem.
A more effective approach is to use the path integral control
framework to find the parameters of a feedback control
policy. This can be done by sampling in policy parameter
space, these methods are known as Policy Improvement
with Path Integrals [14]. Another approach to finding the
parameters of a policy is to attempt to directly sample from
the optimal distribution defined by the value function [3].
Other methods along similar threads of research include [10],
[17].
Another way that the path integral control framework
can be applied is in a model predictive control setting.
In this setting an open-loop control sequence is constantly
optimized in the background while the machine is simultaneously executing the “best guess” that the controller has.
An issue with this approach is that many trajectories must
be sampled in real-time, which is difficult when the system
has complex dynamics. One way around this problem is to
This research has been supported by NSF Grant No. NRI-1426945.
The 1 authors are with the Autonomous Control and Decision Systems
Laboratory at the Georgia Institute of Technology, Atlanta, GA, USA. Email:
[email protected]
drastically simplify the system under consideration by using
a hierarchical scheme [4], and use path integral control to
generate trajectories for a point mass which is then followed
by a low level controller. Even though this approach may be
successfull for certain applications, it is limited in the kinds
of behaviors that it can generate since it does not consider the
full non-linearity of dynamics. A more efficient approach is
to take advantage of the parallel nature of sampling and use
a graphics processing unit (GPU) [19] to sample thousands
of trajectories from the nonlinear dynamics.
A major issue in the path integral control framework is
that the expectation is taken with respect to the uncontrolled
dynamics of the system. This is problematic since the probability of sampling a low cost trajectory using the uncontrolled
dynamics is typically very low. This problem becomes more
drastic when the underlying dynamics are nonlinear and
sampled trajectories can become trapped in undesirable parts
of the state space. It has previously been demonstrated
how to change the mean of the sampling distribution using
Girsanov’s theorem [15], [16], this can then be used to
develop an iterative algorithm. However, the variance of
the sampling distribution has always remained unchanged.
Although in some simple simulated scenarios changing the
variance is not necessary, in many cases the natural variance
of a system will be too low to produce useful deviations from
the current trajectory. Previous methods have either dealt
with this problem by artificially adding noise into the system
and then optimizing the noisy system [10], [14]. Or they
have simply ignored the problem entirely and sampled from
whatever distribution worked best [12], [19]. Although these
approaches can be successful, both are problematic in that
the optimization either takes place with respect to the wrong
system or the resulting algorithm ignores the theoretical basis
of path integral control.
The approach we take here generalizes these approaches in
that it enables for both the mean and variance of the sampling
distribution to be changed by the control designer, without
violating the underlying assumptions made in the path integral derivation. This enables the algorithm to converge fast
enough that it can be applied in a model predictive control
setting. After deriving the model predictive path integral
control (MPPI) algorithm, we compare it with an existing
model predictive control formulation based on differential
dynamic programming (DDP) [6], [13], [18]. DDP is one of
the most powerful techniques for trajectory optimization, it
relies on a first or second order approximation of the dynamics and a quadratic approximation of the cost along a nominal
trajectory, it then computes a second order approximation of
the value function which it uses to generate the control.
this transformation we apply an exponential transformation
of the value function
II. PATH INTEGRAL CONTROL
In this section we review the path integral optimal control
framework [7]. Let xt ∈ RN denote the state of a dynamical
system at time t, u(xt , t) ∈ Rm denotes a control input for
the system, τ : [t0 , T ] → Rn represents a trajectory of the
system, and dw ∈ Rp is a brownian disturbance. In the path
integral control framework we suppose that the dynamics
take the form:
dx = f (xt , t)dt + G(xt , t)u(xt , t)dt + B(xt , t)dw
(1)
In other words, the dynamics are affine in control and subject
to an affine brownian disturbance. We also assume that G
and B are partitioned as:
0
0
G(xt , t) =
; B(xt , t) =
(2)
Gc (xt , t)
Bc (xt , t)
Expectations taken with respect to (1) are denoted as EQ [·],
we will also be interested in taking expectations with respect
to the uncontrolled dynamics of the system (i.e (1) with
u ≡ 0). These will be denoted EP [·]. We suppose that
the cost function for the optimal control problem has a
quadratic control cost and an arbitrary state-dependent cost.
Let φ(xT ) denote a final the terminal cost, q(xt , t) a state
dependent running cost, and define R(xt , t) as a positive
definite matrix. The value function V (xt , t) for this optimal
control problem is then defined as:
"
#
Z T
1 T
min EQ φ(xT ) +
q(xt , t) + u R(xt , t)u dt
u
2
t
(3)
The Stochastic Hamilton-Jacobi-Bellman equation [1], [11]
for the type of system in (1) and for the cost function in (3)
is given as:
−∂t V = q(xt , t) + f (xt , t)T Vx
1
− VxT G(xt , t)R(xt , t)−1 G(xt , t)T Vx
2
1
+ tr(B(xt , t)B(xt , t)T Vxx )
2
(4)
where the optimal control is expressed as:
u∗ = −R(xt , t)−1 G(xt , t)T Vx
(5)
The solution to this backwards PDE yields the value function
for the stochastic optimal control problem, which is then
used to generate the optimal control. Unfortunately, classical
methods for solving partial differential equations of this
nature suffer from the curse of dimensionality and are
intractable for systems with more than a few state variables.
The approach we take in the path integral control framework is to transform the backwards PDE into a path integral,
which is an expectation over all possible trajectories of the
system. This expectation can then be approximated by forward sampling of the stochastic dynamics. In order to effect
V (x, t) = −λ log(Ψ(x, t))
(6)
Here λ is a positive constant. We also have to assume a
relationship between the cost and noise in the system (as
well as λ) through the equation:
Bc (xt , t)Bc (x, t)T = λGc (xt , t)R(xt , t)−1 Gc (xt , t)T (7)
The main restriction implied by this assumption is that
B(xt , t) has the same rank as R(xt , t). This limits the
noise in the system to only effect state variables that are
directly actuated (i.e. the noise is control dependent). There
are a wide variety of systems which naturally fall into this
description, so the assumption is not too restrictive. However,
there are interesting systems for which this description does
not hold (i.e. if there are known strong disturbances on
indirectly actuated state variables or if the dynamics are only
partially known).
By making this assumption and performing the exponential transformation of the value function the stochastic HJB
equation is transformed into the linear partial differential
equation:
1
Ψ(xt , t)
q(xt , t) − f (xt , t)T Ψx − tr(Σ(xt , t)Ψxx )
λ
2
(8)
Here we’ve denoted the covariance matrix
∂t Ψ =
Bc (xt , t)Bc (xt , t)T
as Σ(xt , t). This equation is known as the backward
Chapman-Kolmogorov PDE. We can then apply the
Feynman-Kac lemma, which relates backward PDEs of this
type to path integrals through the equation:
"
!
#
Z
1 T
Ψ(xt0 , t0 ) = EP exp −
q(x, t) dt Ψ(xT , T )
λ t0
(9)
Note that the expectation (which is the path integral) is
taken with respect to P which is the uncontrolled dynamics
of the system. By recognizing that the term Ψ(xT ) is the
1
transformed terminal cost: e− λ φ(xT ) we can re-write this
expression as:
1
(10)
Ψ(xt0 , t0 ) ≈ EP exp − S(τ )
λ
RT
where S(τ ) = φ(xT ) + t0 q(xt , t)dt is the cost-to-go of
the state dependent cost of a trajectory. Lastly we have to
compute the gradient of Ψ with respect to the initial state xt0 .
This can be done analytically and is a straightforward, albeit
lengthy, computation so we omit it and refer the interested
reader to [14]. After taking the gradient we obtain:
1
S(τ ) B(xt0 , t0 )dw
∗
−1 EP exp −
λ
u dt = G(xt0 , t0 )
EP exp − λ1 S(τ )
(11)
Where the matrix G(xt , t) is defined as:
−1
R(xt , t)−1 Gc (xt , t)T Gc (xt , t)R(xt , t)−1 Gc (xt , t)T
(12)
Note that if Gc (xt , t) is square (which is the case if the
system is not over actuated) this reduces to Gc (xt , t)−1 .
Equation (11) is the path integral form of the optimal
control. The fundamental difference between this form of
the optimal control and classical optimal control theory is
that instead of relying on a backwards in time process,
this formula requires the evaluation of an expectation which
can be approximated using forward sampling of stochastic
differential equations.
A. Discrete Approximation
Equation (11) provides an expression for the optimal
control in terms of a path integral. However, these equations
are for continuous time and in order to sample trajectories
on a computer we need discrete time approximations.
We first discretize the dynamics of the system. We have
that xt+1 = xt + dxt where dxt is defined as:
√
dxt = (f (xt , t) + G(xt , t)u(xt , t)) ∆t + B(xt , t) ∆t
(13)
The term is a vector of standard normal Gaussian random
variables. For the uncontrolled dynamics of the system we
have:
√
dxt = f (xt , t)∆t + B(xt , t) ∆t
(14)
Another way we can express B(xt , t)dw which will be
useful is as:
B(xt , t)dw ≈ dxt − f (xt , t)∆t
(15)
PN
Lastly we say: S(τ ) ≈ φ(xT )+ i=0 q(xt , t)∆t where N =
(T − t)/∆t Then by defining p as the probability induced by
the discrete time uncontrolled dynamics we can approximate
(11) as:
i
h
dxt0
−
f
(x
,
t
)
Ep exp − λ1 S(τ )
t
0
0
∆t
u∗ = G(xt0 , t0 )−1
1
Ep exp − λ S(τ )
(16)
Note that we have moved the ∆t term multiplying u over
to the right-hand side of the equation and inserted it into the
expectation.
III. GENERALIZED IMPORTANCE SAMPLING
Equation (16) provides an implementable method for
approximating the optimal control via random sampling of
trajectories. By drawing many samples from p the expectation can be evaluated using a Monte-Carlo approximation. In
practice, this approach is unlikely to succeed. The problem is
that p is typically an inefficient distribution to sample from
(i.e the cost-to-go will be high for most trajectores sampled
from p). Intuitively sampling from the uncontrolled dynamics
corresponds to turning a machine on and waiting for the
natural noise in the system dynamics to produce interesting
behavior.
In order to efficiently approximate the controls, we require
the ability to sample from a distribution which is likely to
produce low cost trajectories. In previous applications of
path integral control [15], [16] the mean of the sampling
distribution has been changed which allows for an iterative
update law. However, the variance of the sampling distribution has always remained unchanged. In well engineered
systems, where the natural variance of the system is very
low, changing the mean is insufficient since the state space
is never aggressively explored. In the following derivation
we provide a method for changing both the initial control
input and the variance of the sampling distribution.
A. Likelihood Ratio
We suppose that we have a sampling distribution with nonzero control input and a changed variance, which we denote
as q, and we would like to approximate (16) using samples
from q as opposed to p. Now if we write the expectation
term (16) in integral form we get:
dxt
R
0
−
f
(x
,
t)
p(τ )dτ
exp − λ1 S(τ )
t
∆t
R
(17)
exp − λ1 S(τ ) p(τ )dτ
Where we are abusing notation and using τ to represent the
discrete trajectory (xt0 , xt1 , . . . xtN ). Next we multiply both
q(τ )
integrals by 1 = q(τ
) to get:
dxt
R
q(τ )
0
exp − λ1 S(τ )
−
f
(x
,
t)
t
∆t
q(τ ) p(τ )dτ
(18)
q(τ )
R
exp − λ1 S(τ ) q(τ
) p(τ )dτ
And we can then write this as an expectation with respect
to q:
h
i
dxt0
p(τ )
Eq exp − λ1 S(τ )
∆t − f (xt , t) q(τ )
h
(19)
)i
Eq exp − λ1 S(τ ) p(τ
q(τ )
We now have the expectation in terms of a sampling distribution q for which we can choose:
i) The initial control sequence from which to sample
around.
ii) The variance of the exploration noise which determines
how aggressively the state space is explored.
p(τ )
However, we now have an extra term to compute q(τ
) . This is
known as the likelihood ratio (or Radon-Nikodym derivative)
between the distributions p and q. In order to derive an
expression for this term we first have to derive equations
for the probability density functions of p(τ ) and q(τ ) individually. We can do this by deriving the probability density
function for the general discrete time diffusion processes
P (τ ), corresponding to the dynamics:
√
dxt = (f (xt , t) + G(xt , t)u(xt , t)) ∆t + B(xt , t) ∆t
(20)
The goal is to find P (τ ) = P (xt0 , xt1 , . . . xtN ). By conditioning and using the Markov property of the state space this
probability becomes:
P (xt0 , xt1 , . . . xtN ) =
N
Y
i=1
P (xti |xti−1 )
(21)
Now recall that a portion of the state space has deterministic
dynamics and that we’ve partitioned the diffusion matrix as:
0
B(xt , t) =
(22)
Bc (xt , t)
We can partition the state variables x into the deterministic
(a)
(c)
and non-deterministic variables xt and xt respectively.
(a)
(a)
(a)
The next step is to conditionon xt+1 = F (xt , t) = xt +
(a)
(a)
f (xt , t) + G (xt , t)ut dt since if this does not hold
P (τ ) is zero. We thus need to compute:
N
Y
i=1
(a)
P xti |xti−1 , xti = F (a) (xti−1 , ti−1
(23)
And from the dynamics equations we know that each of these
one-step transitions is Gaussian with mean: f (c) (xt , t) +
G(c) (xti , ti )u(xti , ti ) and variance:
Σi = Bc (xti , ti )Bc (xti , ti )T ∆t.
(24)
Then under the condition that each Ati is invertible and each
Γi is invertible, the likelihood ratio for the two distributions
is:
!
!
N
N
Y
∆t X
|Ati | exp −
Qi
(31)
2 i=1
i=1
Proof: In discrete time the probability of a trajectory is
formulated according to the (26). We thus have p(τ ) equal
to:
PN
z
Σ
z
exp − ∆t
i
i
i
i=1
2
(32)
p(τ ) =
Zp (τ )
and q(τ ) equal to:
PN
−1
T
T
exp − ∆t
(z
−µ
)
A
Σ
A
(z
−µ
)
(
)
i
i
i ti
i
i
t
i=1
2
i
Zq (τ )
(33)
Then dividing these two equations we have
(c)
dxt
i
∆t
(c)
− f (xti , ti ), and µi =
We then define zi =
G(c) (xti , ti )u(xti , ti ). Applying the definition of the Gaussian distribution with these terms yields:
N exp − ∆t (z − µ )T Σ−1 (z − µ )
Y
i
i
i
i
i
2
(25)
P (τ ) =
(2π)n/2 |Σi |1/2
i=1
And then using basic rules of exponents this probability
becomes:
!
N
∆t X
T −1
−1
Z(τ ) exp −
(zi − µi ) Σi (zi − µi )
(26)
2 i=1
QN
n/2
Where Z(τ ) =
|Σi |1/2 . With this equation
i=1 (2π)
in hand we’re now ready to compute the likelihood ratio
between two diffusion processes.
Theorem 1: Let p(τ ) be the probability density function
for trajectories under the uncontrolled discrete time dynamics:
√
dxt = f (xt , t)∆t + B(xt , t) ∆t
(27)
And let q(τ ) be the probability density function for trajectories under the controlled dynamics with an adjusted variance:
dxt = (f (xt , t) + G(xt , t)u(xt , t)) ∆t+
√
BE (xt , t) ∆t (28)
Where the adjusted variance has the form:
0
BE (xt , t) =
At Bc (xt , t)
Qi =
T
−1
µT
i Σi µi
(29)
Where Γi is:
Γ−1
i
=
Σ−1
i
−
−1
AT
t i Σi A t i
i
i=1
(2π)n/2 |Σi |1/2 )
!
as:
N
∆t X
ζi
exp −
2 i=1
!
(30)
(34)
Where ζi is:
−1
−1
T
ζ i = zT
AT
(zi − µi ) (35)
i Σi zi − (zi − µi )
t i Σi A t i
Using basic rules of determinants it is easy to see that the
term outside the exponent reduces to
N
N
1/2
Y
Y
(2π)n/2 |AT
)
j Σ j Aj |
|Aj |
=
(2π)n/2 |Σj |1/2 )
j=1
j=1
(36)
So we need only show that ζi reduces to Qi . Observe that at
every timestep we have the difference between two quadratic
functions of zi , so we can complete the square to combine
this into a single quadratic function. If we recall the definition
of Γi from above, and define Λi = AT
ti Σi Ati then completing
the square yields:
ζi = zi + Γi Λ−1
i µi
T
Γ−1
zi + Γi Λ−1
i
i µi
T −1
(37)
−1
−1
T −1
− µi Λi µi − Γi Λi µi Γi Γt Λi µi
Now we expand out the first quadratic term to get:
−1
−1
T −1
T −1
ζ i = zT
i Γi zi + 2µi Λi zi + µi Λi Γi Λi µi
−1
−1
−1
T −1
− µT
i Λi µi − (Γi Λi µi ) Γi (Γi Λi µi )
And define zi , µi , and Σi as before. Let Qi be defined as:
(zi − µi ) Γ−1
i (zi − µi )
T −1
+ 2 (µi ) Σi (zi − µi ) +
N
1/2
Y
(2π)n/2 |AT
)
t Σi A t i |
p(τ )
q(τ )
(38)
Notice that the two underlined terms are the same, except
for the sign, so they cancel out and we’re left with:
−1
T −1
T −1
ζ i = zT
i Γi zi + 2µi Λi zi − µi Λi µi
(39)
Now define z̃i = zi − µi , and then re-write this equation in
terms of z˜i :
T −1
T −1
ζi = (z̃i + µi )T Γ−1
i (z̃i + µi ) + 2µi Λi (z̃i + µi ) − µi Λi µi
(40)
Then by re-defining the running cost q(xt , t) as:
which expands out to:
−1
T −1
T −1
ζi = z̃T
i Γi z̃i + 2µi Γi z̃i + µi Γi µi
−1
T −1
T −1
+ 2µT
i Λi z̃i + 2µi Λi µi − µi Λi µi
(41)
Which then simplifies to:
−1
T −1
T −1
ζi = z̃T
i Γi z̃i + 2µi Γi z̃i + µi Γi µi
−1
T −1
+ 2µT
i Λi z̃i + µi Λi µi
(42)
−1 −1
Now recall that Γi = (Σ−1
, so we can split the
i − Λi )
−1
quadratic terms in Γi into the Σ−1
and
Λ−1
components.
i
i
Doing this yields:
−1
T −1
T −1
T −1
ζi = z̃T
i Γi z̃i + 2µi Σi z̃i − 2µi Λi z̃i + µi Σi µi
−1
T −1
T −1
− µT
i Λi µi + 2µi Λi z̃i + µi Λi µi
(43)
and by noting that the underlined terms cancel out we see
that we’re left with:
ζi =
−1
z̃T
i Γi z̃i
+
−1
2µT
i Σi z̃i
+
−1
µT
i Σi µi
(44)
which is the same as:
1
T
(z − µ) Γ̃−1 (z − µ)
2
(48)
1 T −1
T −1
+ µ H (z − µ) + µ H µ
2
PN
and S̃(τ ) = φ(xT ) + j=1 q̃(x, u, dx), we have:
h
i
t
−
f
(x
,
t)
Eq exp − λ1 S̃(τ ) dx
t
∆t
h
i
u∗t = G(xt , t)−1
1
Eq exp − λ S̃(τ )
(49)
Also note that dxt is now equal to:
√
(f (xt , t) + G(xt , t)u(xt , t)) ∆t + B(xt , t) ∆t (50)
q̃(x, u, dx) = q(xt , t) +
So we can re-write
dxt
∆t
− f (xt , t) as:
G(xt , t)u(xt , t) + B(xt , t) √
∆t
(51)
And then since G(xt , t) does not depend on the expectation
we can pull it out and get the iterative update law:
T
T −1
T −1
(zi − µi ) Γ−1
i (zi − µi ) + 2µi Σi (zi − µi ) + µi Σi µi
(45)
And so ζi = Qi which completes the proof.
The key difference between this proof and earlier path integral works which use an application of Girsanov’s theorem
to sample from a non-zero control input is that this theorem
allows for a change in the variance as well.
In the expression for the likelihood ratio derived here
−1
T −1
the last two terms (2µT
i Σi (zi − µi ) + µi Σi µi ) are exactly the terms from Girsanov’s theorem. The first term
T
((zi − µi ) Γ−1
i (zi − µi )), which can be interpreted as penalizing over-aggressive exploration, is the only additional
term.
B. Likelihood Ratio as Additional Running Cost
The form of the likelihood ratio just derived is easily
incorporated into the path integral control framework by
folding it into the cost-to-go as an extra running cost. Note
that the likelihood ratio appears in both the numerator and
denominator of (16). Therefore, any terms which do not
depend on the state can be factored out of the expectation
and canceled. This
QNremoves the numerically troublesome
normalizing term j=1 |Atj |. So only the summation of Qi
remains. Recall that Σ = λG(xt , t)R(xt , t)−1 G(xt , t). This
implies that:
−1
Γ = λ G(xt , t)R(xt , t)−1 G(xt , t)
(46)
−1
− AT G(xt , t)R(xt , t)−1 G(xt , t)T A
Now define H = G(xt , t)R(xt , t)−1 G(xt , t)T and Γ̃ = λ1 Γ.
We then have:
1
T
Q=
(z−µ) Γ̃−1 (z−µ)+2µT H−1 (z−µ)+µT H−1 µ
λ
(47)
u∗t = G(xt , t)−1 G(xt , t)u(xt , t)
i
h
Eq exp − λ1 S̃(τ ) B(xt , t) √∆t
i
h
+ G(xt , t)−1
Eq exp − λ1 S̃(τ )
(52)
C. Special Case
The update law (52) is applicable for a very general class
of systems. In this section we examine a special case which
we use for all of our experiments. We consider dynamics of
the form:
1 √
dxt = f (xt , t)∆t + G(xt , t) u(xt , t)∆t + √ ∆t
ρ
√(53)
And for the sampling distribution we set A equal to νI.
We also assume that Gc (xt , t) is a square invertible matrix.
This reduces H(xt , t) to Gc (xt , t)−1 . Next the dynamics
can be re-written as:
1
dxt = f (xt , t)∆t + G(xt , t) u(xt , t) + √ √
∆t
ρ ∆t
(54)
Then we can interpret √1ρ √∆t as a random change in the
control input, to emphasize this we will denote this term as
δu = √1ρ √∆t . We then have B(xt , t) √∆t = G(xt , t)δu.
This yields the iterative update law as:
h
i
Eq exp − λ1 S̃(τ ) δu
h
i
u(xt , t)∗ = u(xt , t) +
(55)
Eq exp − λ1 S̃(τ )
which can be approximated as:
PK
u(xti , ti )∗ ≈ u(xti , ti ) +
1
exp
−
S̃(τ
)
δui,k
i,k
k=1
λ
PK
1
exp
−
S̃(τ
)
i,k
k=1
λ
(56)
i) z − µ = G(xt , t)δu
ii) Γ̃−1 = (1 − ν −1 )G(xt , t)−1 R(xt , t)G(xt , t)
iii) H−1 = G(xt , t)−1 R(xt , t)G(xt , t)−1
Given these simplifications q̃ reduces to:
(1 − ν −1 ) T
δu Rδu
2
1
+ uT Rδu + uT Ru
2
q̃(x, u, dx) = q(xt , t) +
(57)
This means that the introduction of the likelihood ratio
simply introduces the original control cost from the optimal
control formulation into the sampling cost, which originally
only included state-dependent terms.
IV. MODEL PREDICTIVE CONTROL ALGORITHM
We apply the iterative path integral control update law,
with the generalized importance sampling term, in a model
predictive control setting. In this setting optimization and
execution occur simultaneously: the trajectory is optimized
and then a single control is executed, then the trajectory is
re-optimized using the un-executed portion of the previous
trajectory to warm-start the optimization. This scheme has
two key requirements:
i) Rapid convergence to a good control input.
ii) The ability to sample a large number of trajectories in
real-time.
The first requirement is essential because the algorithm
does not have the luxury of waiting until the trajectory has
converged before executing. The new importance sampling
term enables tuning of the exploration variance which allows
for rapid convergence, this is demonstrated in Fig. 1.
The second requirement, sampling a large number of
trajectories in real-time, is satisfied by implementing the
random sampling of trajectories on a GPU. The algorithm
is given in Algorithm 1, in the parallel GPU implementation
the sampling for loop (for k to K-1) is run completely in
parallel.
V. EXPERIMENTS
We tested the model predictive path integral control algorithm (MPPI) on three simulated platforms (1) A cart-pole,
(2) A miniature race car, and (3) A quadrotor attempting
to navigate an obstacle filled environment. For the race car
and quadrotor we used a model predictive control version
of the differential dynamic programming (DDP) algorithm
as a baseline comparision. In all of these experiments the
controller operates at 50 Hz, this means that the open loop
control sequence is re-optimized every 20 milliseconds.
Algorithm 1: Model Predictive Path Integral Control
Given: K: Number of samples;
N : Number of timesteps;
(u0 , u1 , ...uN −1 ): Initial control sequence;
∆t, xt0 , f , G, B, ν: System/sampling dynamics;
φ, q, R, λ: Cost parameters;
uinit : Value to initialize new controls to;
while task not completed do
for k ← 0 to K − 1 do
x = xt0 ;
for i ← 1 to N − 1 do
xi+1 = xi + (f + G (ui + δui,k )) ∆t;
S̃(τi+1,k ) = S̃(τi,k ) + q̃;
for i ← 0 to N− 1 do
1
PK
exp(− λ
S̃( τi,k ))δui,k
PK
ui ← ui +
;
1
k=1
k=1 exp(− λ S̃( τi,k ))
send to actuators(u0 );
for i ← 0 to N − 2 do
ui = ui+1 ;
uN −1 = uinit
Update the current state after receiving feedback;
check for task completion;
A. Cart-Pole
For the cart-pole swing-up task we used the state cost:
q(x) = p2 + 500(1 + cos(θ))2 + θ̇2 + ṗ2 , where p is the
position of cart, ṗ is the velocity and θ, θ̇ are the angle and
angular velocity of the pole. The control input is desired
velocity, which maps to velocity through the equation: p̈ =
10(u − ṗ). The disturbance parameter √1ρ was set equal .01
and the control cost was R = 1. We ran the MPPI controller
for 10 seconds with a 1 second optimization horizon. The
controller has to swing-up the pole and keep it balanced for
the rest of the 10 second horizon. The exploration variance
ν
ν
ν
ν
300
Average Running Cost
Where K is the number of random samples (termed rollouts)
and S(τi,k ) is the cost-to-go of the kth rollout from time ti
onward. This expression is simply a reward-weighted average
of random variations in the control input. Next we investigate
what the likelihood ratio addition to the running cost is. For
these dynamics we have the following simplifications:
250
= 75
= 500
= 1000
= 1500
200
150
100
50
0
1.0
1.5
2.0
2.5
3.0
3.5
Number of Rollouts (Log Scale)
4.0
Fig. 1. Average running cost for the cart-pole swing-up task as a function of
the exploration variance ν and the number of rollouts. Using only the natural
system variance the MPC algorithm does not converge in this scenario.
10
B. Race Car
In the race car task the goal was to minimize the objective
function: q(x) = 100d2 + (vx − 7.0)2 . Where d is defined
2
x 2
+ y6 − 1|, and vx is the forward (in
as: d = | 13
body frame) velocity of the car. This cost ensures that
the car to stays on an elliptical track while maintaining a
forward speed of 7 meters/sec. We use a non-linear dynamics
model [5] which takes into account the (highly non-linear)
interactions between tires and the ground. The exploration
variance was set to a constant ν times the natural variance
of the system. The MPPI controller is able to enter turns at
10
MPC-DDP
5
0
−5
−5
−10
−15
−10
0
5
10
15
−10
−15
−10
−5
0
5
10
15
Fig. 3. Comparison of DDP (left) and MPPI (right) performing a cornering
maneuver along an ellipsoid track. MPPI is able to make a much tigther
turn while carrying more speed in and out of the corner than DDP. The
direction of travel is counterclockwise.
9
DDP |vx |
8
MPPI |vx |
DDP |vy |
MPPI |vy |
7
6
5
4
3
2
1
30
Average Running Cost
−5
MPPI
5
0
Velocity (m/s)
parameter, ν, was varied between 1 and 1500. The MPPI
controller is able to swing-up the pole faster with increasing
exploration variance. Fig. 1 illustrates the performance of the
MPPI controller as the exploration variance and the number
of rollouts are changed. Using only the natural variance of
the system for exploration is insufficient in this task, in that
case (not shown in the figure) the controller is never able to
swing-up the pole which results in a cost around 2000.
DDP Solution
ν = 50
ν = 100
ν = 150
ν = 300
25
0
0
10
20
Time (s)
30
40
50
Fig. 4. Comparison of DDP (left) and MPPI (right) performing a cornering
maneuver along an ellipsoid track. MPPI is able to make a much tigther
turn while carrying more speed in and out of the corner than DDP.
20
15
MPPI and DDP which guide the quadrotor through the forest
as quickly as possible. The cost function for MPPI was
10
100
5
1.0
MPC-DDP
1.5
2.0
2.5
3.0
3.5
Number of Rollouts (Log Scale)
4.0
Fig. 2. Performance comparison in terms of average cost between MPPI
and MPC-DDP as the exploration variance ν changes from 50 to 300 and
the number of rollouts changes from 10 to 1000. Only with a very large
increase in the exploration variance is MPPI able to outperform MPC-DDP.
Note that the cost is capped at 25.0
60
close to the desired speed of 7 m/s and then slide through
the turn. The DDP solution does not attempt to slide and
significantly reduces its forward velocity before entering the
turn, this results in a higher average cost compared to the
MPPI controller. Fig. 2 shows the cost comparison between
MPPI and MPC-DDP, and Figures 3 and 4 show samples of
the trajectories taken by the two algorithms as well as the
velocity profiles.
0
C. Quadrotor
The quadrotor task was to fly through a field filled
with cylindrical obstacles as fast as possible. We used the
quadrotor dynamics model from [9]. This is a non-linear
model which includes position, velocity, euler angles, angular
acceleration, and the rotor dynamics. We randomly generated
three forests, one where obstacles are on average 3 meters
apart, the second one 4 meters apart, and the third 5 meters
apart. We then separately created cost functions for both
MPPI
80
40
20
0
20
40
60
80
100
0
20
40
60
80
100
Fig. 5. Left: sample DDP trajectory through 4m obstacle field, Right:
Sample MPPI trajectory through the same field. Since the MPPI controller
can directly reason about the shape of the obstacles it is able to safely pass
through the field taking a much more direct route.
2
des 2
of the form: q(x) = 2.5(px − pdes
x ) + 2.5(py − py ) +
d
2
2
2
150(pz − pdes
z ) + 50ψ + kvk +350 exp(− 12 ) + 1000C
where (px , py , pz ) denotes the position of the vehicle. ψ
denotes the yaw angle in radians, v is velocity, and d is
the distance to the closest obstacle. C is a variable which
indicates whether the vehicle has crashed into the ground or
an obstacle. Additionally if C = 1 (which indicates a crash),
the rollout stops simulating the dynamics and the vehicle
remains where it is for the rest of the time horizon. We found
that the crash indicator term is not useful for the MPC-DDP
based controller, this is not surprising since the discontinuity
it creates is difficult to approximate with a quadratic function.
The term in the cost for avoiding obstacles in the MPCDDP P
controller consists purely of a large exponential term:
N
2000 i=1 exp(− 21 d2i ), note that this sum is over all the
obstacles in the proximity of the vehicle whereas the MPPI
controller only has to consider the closest obstacle.
18
MPPI
MPC-DDP
Time to Completion (s)
17
16
15
14
13
12
11
10
3m
4m
Density Setting of Forest
5m
ii) The use of a GPU to sample thousands of trajectories
in real-time.
The derivation of the likelihood ratio enables the designer
of the algorithm to tune the exploration variance in the path
integral control framework, whereas previous methods have
only allowed for the mean of the distribution to be changed.
Tuning the exploration variance is critical in achieving a high
level of performance since the natural variance of the system
is typically too low to achieve good performance.
The experiments considered in this work only consider
changing the variance by a constant multiple times the
natural variance of the system. In this special case the
introduction of the likelihood ratio corresponds to adding in
a control cost when evaluating the cost-to-go of a trajectory.
A direction for future research is to investigate how to
automatically adjust the variance online. Doing so could
enable the algorithm to switch from aggressively exploring
the state space when performing aggressive maneuvers to
exploring more conservatively for performing very precise
maneuvers.
R EFERENCES
Fig. 6.
Time to navigate forest. Comparison between MMPI and DDP.
Since the MPPI controller can explicitly reason about
crashing (as opposed to just staying away from obstacles),
it is able to travel both faster and closer to obstacles than
the MPC-DDP controller. Fig. 7 shows the difference in time
between the two algorithms and Fig. 6 the trajectories taken
by MPC-DDP and one of the MPPI runs on the forest with
obstacles placed on average 4 meters away.
Fig. 7. Simulated forest environment used in the quadrotor navigation task.
VI. CONCLUSION
In this paper we have developed a model predictive path
integral control algorithm which is able to outperform a
state-of-the-art DDP method on two difficult control tasks.
The algorithm is based on stochastic sampling of system
trajectories and requires no derivatives of either the dynamics
or costs of the system. This enables the algorithm to naturally
take into account non-linear dynamics, such as a non-linear
tire model [5]. It is also able to handle cost functions which
are intuitively appealing, such as an impulse cost for hitting
an obstacle, but are difficult for traditional approaches that
rely on a smooth gradient signal to perform optimization.
The two keys to achieving this level of performance with a
sampling based method are:
i) The derivation of the generalized likelihood ratio between discrete time diffusion processes.
[1] W. H. Fleming and H. M. Soner. Controlled Markov processes and
viscosity solutions. Applications of mathematics. Springer, New York,
2nd edition, 2006.
[2] A. Friedman. Stochastic Differential Equations And Applications.
Academic Press, 1975.
[3] Vicenç Gómez, Hilbert J Kappen, Jan Peters, and Gerhard Neumann.
Policy search for path integral control. In Machine Learning and
Knowledge Discovery in Databases, pages 482–497. Springer, 2014.
[4] Vicenç Gómez, Sep Thijssen, Hilbert J Kappen, Stephen Hailes, and
Andrew Symington. Real-time stochastic optimal control for multiagent quadrotor swarms. arXiv preprint arXiv:1502.04548, 2015.
[5] R.Y Hindiyeh. Dynamics and Control of Drifting in Automobiles. PhD
thesis, Stanford University, March 2013.
[6] D. H. Jacobson and D. Q. Mayne. Differential dynamic programming.
American Elsevier Pub. Co., New York, 1970.
[7] H. J. Kappen. Linear theory for control of nonlinear stochastic
systems. Phys Rev Lett, 95:200201, 2005. Journal Article United
States.
[8] I. Karatzas and S. E. Shreve. Brownian Motion and Stochastic
Calculus (Graduate Texts in Mathematics). Springer, 2nd edition,
August 1991.
[9] Nathan Michael, Daniel Mellinger, Quentin Lindsey, and Vijay Kumar. The grasp multiple micro-uav testbed. Robotics & Automation
Magazine, IEEE, 17(3):56–65, 2010.
[10] E. Rombokas, M. Malhotra, E.A. Theodorou, E. Todorov, and Y. Matsuoka. Reinforcement learning and synergistic control of the act hand.
IEEE/ASME Transactions on Mechatronics, 18(2):569–577, 2013.
[11] R. F. Stengel. Optimal control and estimation. Dover books on
advanced mathematics. Dover Publications, New York, 1994.
[12] F. Stulp, J. Buchli, E. Theodorou, and S. Schaal. Reinforcement learning of full-body humanoid motor skills. In Proceedings of 10th IEEERAS International Conference on Humanoid Robots (Humanoids),
pages 405–410, Dec 2010.
[13] E. Theodorou, Y. Tassa, and E. Todorov. Stochastic differential
dynamic programming. In American Control Conference, 2010, pages
1125–1132, 2010.
[14] E. A. Theodorou, J. Buchli, and S. Schaal. A generalized path integral
approach to reinforcement learning. Journal of Machine Learning
Research, (11):3137–3181, 2010.
[15] E.A. Theodorou and E. Todorov. Relative entropy and free energy dualities: Connections to path integral and kl control. In the Proceedings
of IEEE Conference on Decision and Control, pages 1466–1473, Dec
2012.
[16] Evangelos A. Theodorou. Nonlinear stochastic control and information
theoretic dualities: Connections, interdependencies and thermodynamic interpretations. Entropy, 17(5):3352–3375, 2015.
[17] Sep Thijssen and HJ Kappen. Path integral control and state-dependent
feedback. Physical Review E, 91(3):032104, 2015.
[18] E. Todorov and W. Li. A generalized iterative lqg method for locallyoptimal feedback control of constrained nonlinear stochastic systems.
pages 300–306, 2005.
[19] G. Williams, E. Rombokas, and T. Daniel. Gpu based path integral
control with learned dynamics. In Neural Information Processing
Systems - ALR Workshop, 2014.
| 3cs.SY
|
1
Fault Tolerance in Distributed Neural Computing
arXiv:1509.09199v1 [cs.NE] 30 Sep 2015
Anton Kulakov, Mark Zwoliński, and Jeff Reeve
Abstract—With the increasing complexity of computing systems, complete hardware reliability can no longer be guaranteed.
We need, however, to ensure overall system reliability. One of
the most important features of artificial neural networks is their
intrinsic fault-tolerance. The aim of this work is to investigate
whether such networks have features that can be applied to
wider computational systems. This paper presents an analysis, in
both the learning and operational phases, of a distributed feedforward neural network with decentralised event-driven time
management, which is insensitive to intermittent faults caused
by unreliable communication or faulty hardware components.
The learning rules used in the model are local in space and time,
which allows efficient scalable distributed implementation. We
investigate the overhead caused by injected faults and analyse
the sensitivity to limited failures in the computational hardware
in different areas of the network.
Index Terms—Fault-tolerance, graceful degradation, redundancy, neural networks.
I. I NTRODUCTION
The inevitable demand for ever more computational capability drives the creation of ever larger parallel distributed
machines (e.g. the K computer has 705,024 cores [1] ), so that
although the mean time to failure (MTTF) for the individual
components can be very high (up to 106 hours [2]), the large
number of components will inevitably lead to frequent failures
– on average once every one and a half hours. Failures can
also be caused by the fact that in large multi-processor systems
the arrival of communication messages is not guaranteed
or they can arrive late [3]. This requires new solutions for
fault tolerance to allow the next generation of extreme-scale
massively parallel computers to be used at their full-capacity.
In recent years, it has been suggested that neural computing
offers a model of fault-tolerant computing. In biological brains,
neurons die without apparent loss of functionality of the whole
system and by analogy, this principle has been applied to
neural simulation engines (SpiNNaker [4], BlueBrain [5]). For
instance, SpiNNaker will employ 50,000 chips (with 20 slow –
200 MHz – processors per chip) connected over a fast network
(1 Gbps) and is based on a model of communication-centric
computation, in contrast to conventional, calculation-centric
computers with very fast processors (2 GHz) over not-so-fast
networks. In order to maintain high communication speed and
avoid deadlocks, a packet-dropping mechanism is used when
a packet cannot be forwarded [6].
Artificial neural networks have been inspired by studies
of the brain structure, where information is processed in a
parallel and distributed way. Commonly used conventional
sequential computing systems utilize one or a few sparselyinterconnected, high performance processing units. Neural
Mark Zwolinski is with Electronics and Computer Science, University of
Southampton, Southampton, SO17 IBJ, UK e-mail: [email protected]
networks, in contrast, employ a large number of highly
interconnected, very simple processing elements, where the
computational power of the model comes above all from the
interaction of all its units.
The motivation for this work is to determine whether
neural computing can be used as a paradigm for reliable
systems running on unreliable hardware. We examine the
fault-tolerant characteristics of parallel distributed processing
networks with a feed-forward structure, in order to understand
how the required fault tolerance can be achieved on systems
with unreliable communications. The work investigates neural
network performance under damage conditions and dynamics
of weight change in a representative task.
This paper is structured as follows. We first review related
work in the field and outline several techniques to assure the
fault tolerant behavior of neural networks. Then we define the
key terms and concepts for general and comparable results
as well as discussing the network structure and training techniques used in the experiments. Next we relate these findings
to an analysis of the network’s structure. The implications of
the findings for fault tolerance and its improvement are then
discussed, before concluding.
II. R ELATED W ORK
A. Fault tolerance in neural networks
Due to the multiplicity of individual units, neural networks
contain more processing elements than is necessary to solve a
problem. Moore [7] argued that because of the large number
of components, neural computers would need to be designed
with high quality components. In contrast, Mozer’s work [8]
has shown that units may be simply removed from a network
without damaging its performance. The loss of a few units
would be unlikely to cause any noticeable decrease in accuracy
of the overall performance in a large system. Tai in [9] showed
that losses of up to 40% can be tolerated in the Hopfield model.
This leads to the conclusion that due to the inherent
overall fault tolerance, non-critical components within the
neural network system need not be particularly reliable, as
has been proposed by Belfore and Johnson [10]. Based on
this conclusion, Chiu et al [11] proposed an algorithm to
improve both the efficiency and fault-tolerance of a multilayer network. In their approach a unique measure of neuron
relevance is used, according to which the least significant
neurons are eliminated, whereas the most significant ones are
duplicated.
Carter and Segee [12] empirically showed that multilayer
networks do not significantly reduce the level of tolerance after
pruning. However, they also pointed out that this is not always
the case. Despite inherent fault tolerance being provided by
the distributed processing architecture, neural networks are not
2
always tolerant of the loss of processing elements [13]. Their
conclusion is that often, instead of catastrophic failure under
the influence of faults or noise affecting inputs and internal
components, it is most likely that network performance will
degrade gracefully.
Furthermore, Segee et al showed in [14] that fault tolerance
is influenced by the training algorithm used and even the initial
state of the network. The implication is that if the number of
network processing elements can be made large, then fault
tolerance increases in the network automatically by virtue of
the gross similarity with biological neural networks. The idea
is that fault tolerance in a neural network is directly related to
the redundancy introduced because of “spare capacity”, when
the complexity of the problem is less than the computational
capacity of the network. Nijhuis et al. [15] came to a similar
conclusion, stating that fault tolerant behavior is not always
self-evident but must be assured by an appropriate training
scheme. Furthermore, [16] presents a procedure to build fault
tolerant neural networks by replicating the hidden units. An
analytical derivation of the minimum redundancy required in
order to tolerate all possible single faults is presented in [17].
On the other hand, von Seelen and Mallot [18] discuss
whether indeed a neural network’s reliability is caused by
redundancy, both in terms of fault tolerance and graceful
degradation. They assume that a neural network uses all of
its resources to balance between computational accuracy and
computation time. Thus redundancy is not identical to reserve
capacity and neural networks utilize available resources to the
full.
Taking these sometimes contradictory findings into account,
we have examined the fault tolerance of neural networks in
both the training and operational phases, so as to be able
to evaluate the system reliability when such networks are
implemented on massively parallel hardware. Moreover, the
published literature considers failures in the computational
units; we are just as concerned with failures in the communication links.
B. Concepts of Reliability
The field of reliable and fault-tolerant computation is very
wide, embracing many different architectural and operational
features of neural network systems as well as several conceptual viewpoints. However, in the literature an inconsistent and
often inaccurate use of key terms and concepts can be found,
which can cause confusion and uncertainty. A categorization
of the causes of failures affecting neural network reliability
must be developed in order to omit non-precise terms and
obtain general and comparable results. This section addresses
the problem and defines key terms and concepts, used further
in the paper.
First of all, to describe the reaction of the network performance to faults, the terms graceful degradation and faulttolerance are often used. Unlike conventional computers, a
neural system is often not adversely affected by faults or noise
in internal components. Instead of failing catastrophically,
the system continues delivering acceptable, although possibly
reduced performance. Computational accuracy is allowed to
degrade in a controlled manner as the fault severity increases.
Such a low sensitivity to occurring faults instead of a complete
failure is known as graceful degradation [19].
In contrast to graceful degradation, fault tolerance is the
property that guarantees the proper operation of the system
in the event of a failure (or several failures) within some
of its components. It describes the robustness of the network
function in the presence of degradation in the computational
elements, such as broken connections or erroneously functioning processing elements.
A common misunderstanding is that of confusing fault
tolerance with robustness to noisy inputs. Fault tolerance is
the ability of a system to continue to perform to specification in the presence of hardware faults, such as broken
connections, connections with an erroneous weight, or neurons
with inaccurate outputs. On the other hand, a system that
continues coping with input noise and operates correctly
despite errors in its inputs is termed a robust system. However,
it is fault-tolerance, rather than robustness, that is associated
with sensitivity to internal noise. Nijhuis in [15] refers to fault
tolerance as hardware fault-tolerance and correspondingly to
robust systems as data fault-tolerant systems. In this paper, we
focus exclusively on hardware fault-tolerance, which describes
a system’s sensitivity to faults that result in perturbations in
network parameters or topology, but does not refer to noisy or
partial input data. In fact, we believe that the sensitivity of a
system to noisy inputs is both inappropriate and inconsistent
for defining a fault-tolerant system.
Another area of potential confusion is the stage at which
errors start occurring. Carter in [13] distinguishes between
two types of fault-tolerance in neural networks: an operational
and a learning fault-tolerance. The sensitivity of network
performance to permanent or transient faults occurring at
the learning stage, is referred to as learning fault-tolerance.
Whereas the operational fault-tolerance deals with the sensitivity of network performance to faults presented after learning
has been accomplished in a fault-free environment. In this
paper we pay primary attention to learning fault-tolerance.
III. FAULT S IMULATION
Neural networks are often treated as black-box systems.
Measuring the degree of failure is based on the results at
the output units for presented input data. We investigate what
level of fault-tolerance the neural network can achieve given
the faults that might occur during both learning and operation
phases. In the case of a fault occurring during the learning
phase, we consider how much longer it will take to train the
network and how fault-tolerant the final version of the network
will be. To address all of these issues, artificial faults can be
introduced into the system. This section discusses the possible
ways of introducing faults.
A. Fault-tolerance Analysis
In order to achieve a suitable analytical model of the
reliability of a neural network system, a definition of various
failure modes and their impacts on the system is required. This
would lead to a firm foundation for further investigations of the
3
amount of fault tolerance exhibited by a neural network. However, taking into account the high level of inter-dependence of
elements in the system on each other, this task is extremely
complicated.
Due to the difficulty in defining the effect of an individual
unit or connection on the overall reliability of an entire system,
we use empirical investigation. We consider a type of fault that
is admittedly severe and correspond to the highest failure rates.
It is often referred as “loss of weight” fault and occur in the
case of open-circuits, [13]. Some authors refer to this fault as
“stuck-at-0”, e.g. [19]. A classical example of an open-circuit
fault can be imagined as a damaged neuron or connection,
which can be related in biological terms to the continual loss
of synapses in the brain. In simulation it is implemented by
setting the selected weight to zero.
B. Fault Injection Technique
During each simulation, faults are probabilistically introduced and the degree of failure is evaluated according to some
measure. The measure of reliability from many experiments
can be plotted against the number of introduced faults injected
into the system. The plot indicates the way the neural network
model behaves depending upon the generic nature and the
faults occurrence rate. Different plots can be compared and
contrasted in order to judge the system’s sensitivity to different
types and locations of faults. This facilitates evaluation of
fault-tolerance when the type of fault and the rate of occurrence are known.
The fault injection technique is convenient for indicating the
isolated effects of individual faults. However it is impractical
for evaluating the impact of multiple faults as their effects
combine and are not independent. It is not realistic simply to
add the impacts of single faults in order to imitate the effect
of multiple faults due to non-predictable correlations between
them. This complicates accurate prediction of the effect of all
faults occurring together over a period of time in real use.
Also another scenario is possible: that a system maintains
an adequate performance despite a limited injection of faults.
However, after a certain fault threshold is reached, the system
may abruptly reduce its performance, which can lead to a total
failure.
Another complication is caused by the temporary nature of
many faults, often called “transient” and “intermittent” faults.
Transient faults are non-recurring and intermittent faults recur
at, usually irregular, intervals. These faults are caused by
several contributing factors, some of which may be effectively
random, which occur simultaneously. The more complex the
system or mechanism involved, the greater the likelihood of
an intermittent fault. An estimation of the impact of temporal
faults is unreliable and thus is out of the scope of this
investigation.
In conclusion, a fault injection method is useful for gaining
a very basic indication of the reliability of a neural network
system, though it may identify especially critical areas of a
neural network which can then be protected against possible
faults in any implementation.
IV. E XPERIMENTAL A PPROACH
In terms of fault tolerance, some neural networks and
some training algorithms are better than others. If nothing
is done to control the proper operation of the system in
the event of a fault during training, the fault tolerance of
the final network may be very random, depending on the
problem, the chosen architecture, the data representation, and
the learning examples. In this section we discuss the network
structure, training technique and teaching procedure used in
the experiment.
A. Network Structure and Connectivity
Training physically distributed neural networks with no
shared memories and decentralized event-driven time management has always been a challenging issue. On one side, the
absence of shared memory excludes data contentions. But at
the same time there is no conscious control over the spike
generation, emitting, storing, and processing. Each processor
sequentially performs event processing in accordance with the
temporal order of these events. Nothing ensures that events
are processed in a correct order.
Of all the existing neural network topologies, the feedforward neural network is probably the one most often used.
This network topology consists of three or more layers: an
input layer, an output layer and a number of hidden layers
in between. Each layer contains a number of neurons, which
are connected only with the neurons of the adjacent layers.
Activity flows from input to output and the network topology
contains neither cycles nor lateral connections. The input layer
is present merely to increase the fan-out of the input data,
whereas the hidden and output layers perform computations,
as shown in Fig. 1.
Commonly, the input data is represented in a binary form,
corresponding either to the presence or absence of features.
The actual process is based on the collective computations
that are performed in the synapses and the neurons. At each
synapse the incoming signal is multiplied by the corresponding
connection strength, wij . At the neuron, these values are
summed and compared to the threshold value: if the summed
result exceeds the threshold value, a neuron emits a consecutive spike, otherwise it remains inactive.
For nearly all problems, one hidden layer is sufficient.
Using two hidden layers rarely improves the model, and may
introduce a greater risk of convergence to a local minimum
and there is no theoretical reason for using more than two
hidden layers [20], [21]. Two hidden layers are, in principle,
enough to perform any classification task [22], including highlevel abstractions (e.g. in vision, language, and other AI-level
tasks) [23]. Here, the number of hidden layers is limited to
one.
In our approach, a special processor is dedicated to interacting with the environment and at the same time managing the
work of other processors (e.g. network mapping, setting the
synchronization barrier). The master processor only ‘knows’
which processors are dedicated to representation of the input
and output neurons. A slave does not ‘know’ whether it
contains input, hidden or output neurons. Each slave processor
4
Fig. 1. The architecture of a feed-forward neural network with n-input,
m-hidden layer and l-output node.
is idle until it receives a spike message (from another slave
or the master processor). When the spike message is received,
and the excitation conditions are met, a new spike messages
are sent to the all known targets, determined a priori while
creating the network.
B. Network Training
The dynamics of a neural network are determined by a rule,
derived by Bosman et al [24]. This is a form of reinforcement
learning, where the active weights are either incremented or
decremented by a certain weight proportion based on the
binary feedback signal in accordance with Hebbian learning.
The signal represents the ‘success’ or ‘failure’ of a given
output after each attempt to associate the correct output with
a particular input. The important benefit of the model is
the locality of all the necessary information about the states
and the properties of the involved neurons for calculating
weight alterations. This allows simulation of distributed neural
networks with decentralized, asynchronous time management.
A network usually has two different operating modes:
training and operation. Identical faults are likely to have
different effects during these two distinct phases. During the
training stage, weights between the neurons are adapted in
order to learn to reproduce a set of patterns, which represents
the problem. The actual training stage of the neural network
consists of many so-called training cycles. During each training cycle the network attempts to alter its weights in such a
way that the output neurons produce an expected pattern. The
network is said to function when all training pairs of patterns
are correctly ‘memorized’.
The number of training cycles depends on various factors
such as (a) the complexity of the decision regions that is
caused by the data itself; (b) the network topology, that is
to say the number of layers and the number of neurons per
layer; (c) the learning strategy, which consists of the choice
of parameters in the training algorithm and the order of
presentation of training patterns during learning; and (d) the
rate of forgetting due to interference of the new learned data
with that previously learned.
In the case where the excitation conditions were met first
time for the current input pattern, spike messages are sent to
the post-synaptic neurons. At the same time another message
containing the value of the membrane potential is sent to
the pre-synaptic neuron from which the last spike message
arrived, indicating that pre-synaptic neuron’s spiking activity
led to the excitation of the post-synaptic neuron. The presynaptic neurons store this information in the activated array.
If the excitation condition is not met for the first time for
the current input pattern, the neuron informs the pre-synaptic
neurons about the alteration of its value of membrane potential.
In the case that it receives an inhibitory signal and is not
active any more, the neuron sends another message to its postsynaptic neurons to warp back the neurons’s influence onto
them, changing their state correspondingly.
Finally, when the output neurons produce the expected
output pattern, the master sends a message, which either reinforces the weights, if the received output corresponds to that
expected, or reduces (‘deinforces’) the weights otherwise. Two
arrays (activated and activatedBy ) contain all the necessary
information for applying the learning rule described in [24].
When the weights are altered, a new cycle starts.
Faults occurring during the learning stage help the network’s
resilience to possible damage in the operation stage because of
the more evenly distributed information between its weights
[25]. This prevents the situation of random information distribution among the weights, when some connections are very
influential on the network output (and thus also very sensitive
to perturbations), whereas others are almost useless.
C. Forgetting
All natural cognitive systems gradually forgets previously
learned information as new information is acquired. A similar
effect is observed in artificial neural networks. While storing
a large amount of input patterns into the network, interference
of newly learned data with previously learned data occurs. It
turns out that in the distributed computational nature of neural
networks, the very virtue of the approach is at the same time
the root cause of forgetting.
The negative effect of the path interference arises when
several input patterns are applied to the neural network to
be memorized. The active paths overlap when the strongest
connection from the different input patterns point to the same
intermediary neurons. As a result, the learning of something
new causes forgetting of old data. The challenge is how to keep
the advantages of distributed computation while avoiding the
problem of catastrophic forgetting. For this it is necessary to
examine the basis of the phenomenon.
There may be several reasons for such a situation. First
of all, from the active input neuron the path of activity runs
along the strongest synaptic connections to the corresponding
output neurons. In certain situations an established path can
be completely “wiped out” by an attempt to learn new data,
so that the connection of the previously learned pattern is
5
m
Errorl (W) =
1 X l
(dk − ykl )2 .
m
(1)
k=1
Then, the global error is defined as:
v
u p
u1 X
Errorl (W )).
Error(W) = t
p
(2)
l=1
So that when the actual output y l is equal to desired output
d , the global error equals to zero.
l
V. S IMULATION R ESULTS
The distributed neural network simulator was written in the
C++ language using the Message Passing Interface library
(MPI) for inter-processor communication. The simulated neural network consisted of an input layer, one hidden layer and
an output layer. Decentralized event-driven time management
without memory sharing was applied.
Fig. 2. An example of path interference between A-C-D and B-C-E paths.
Path A-C-D was formed at learning step tk and became the strongest one with
weight efficacy wmax , at learning step tl affects the formation of B-C-E path
with weight efficacy walt .
no longer the strongest. Also a competition between the
active path, formed in the previous steps, and the newly
forming active path can occur. Such competition often erases
or partially destroys the old path and correspondingly leads to
forgetting of old data by the network.
Corresponding to each input, the most probable signal
propagation will follow the associated pattern. However, when
the number of input patterns or the inter-connectivity level
increases, the activity paths overlap, thereby destroying each
other and corrupting the output result.
The intuitive solution is a uniform distribution of active
paths in the network. This can be achieved in two ways. Firstly,
a negative influence of active path interference can be partially
overcome by periodically shuffling the input-output pairs of
patterns before feeding them into the network, as described
in [26]. This approach assists in finding the most optimal
weight values valid for all patterns and, additionally, network’s
fault tolerance increases due to more even weight distribution
throughout the network.
D. Network Operation
During the operation stage, weights are unalterable. The
percentage of errors in recognizing the set of pre-learned
patterns (xl , dl ), l = 1, ..., p will be called the global error
and is calculated in the following way.
Let wij represent the weight in the i-th row and j-th column
of the weight matrix W. The n-dimentional input pattern is
set on the input neurons, propagated through the network,
transformed according to the current state of the weights and
the corresponding m-dimensional output of the network y l is
compared with the desired one dl .
The error for the output neurons per pattern is calculated
according to the formula:
A. Fault Tolerance in Training
Are all connections equally significant? To answer this
question we investigated the weight alteration dynamics during
the training process in order to identify the most and the least
significant connections. Along with the network structure, the
initial state of the weights has a large impact on the ability
to learn and the resilience of a neural network. In this section
we look into the optimal initial values of the weights.
We constructed a small network of 20 neurons (for better
visibility) and recorded the alteration of each weight, plotting
the measured values against each simulation step. We noticed
that training efficiency directly depends on the initial state of
the network: weight equilibration before the learning phase
significantly improves training process. This is demonstrated
in Fig. 3. 1000 random inputs were applied to the network
10000 times before starting any measurements. For each input
the “deinforcement” process was applied to the weights, modifying the overall distribution of their values in the network.
We paid close attention to the weights of connections
between input and hidden layers (see Fig. 4) and between
hidden and output layers (see Fig. 5), as well as to the weight
distribution at the three stages of simulation:
1) at the initial phase, which is reached after the mapping
process;
2) after the network equilibration phase, when the weight
deinforcement rule is applied to the network several
times;
3) after the learning phase, when all the pattern pairs were
successfully learned.
The Gaussian distribution of weights can be found on the
graphs at the initial phase of simulation. After equilibration is
applied, the center of the distribution slightly shifts towards the
left side for the connections between input and hidden layers
and remains almost unaltered for the connections outgoing
from the hidden layer.
A significant difference between the weight distribution
dynamics can be observed among input-hidden and hiddenoutput connections after the training phase. In the first one,
6
Fig. 3. Training efficiency dependence on the number of equilibration steps applied prior to training. During the training phase 4 different patterns were
learned. The network size was 2000 neurons with 5 input and 5 output neurons and 90 % connectivity.
during the training phase, distribution expands over the range
from -0.1 to 1 with a center at 0.45 and a significant concentration of weights around 0.5. However, the weights in the
input-hidden connections shrink considerably from being in
the range from -4 to 4 at the initial stage to the range from
-0.02 to 0.015 at the final stage with an unclear distribution
centered around -0.002.
This phenomenon can be explained in the following way.
The input neurons tend to excite only a specific set of hidden
neurons and concentrate their connection efficacy to a limited
number of neurons. Because of this, a change of the formed
active paths requires more effort but makes the connections
more stable and insensitive to changes. This behavior can be
seen in Fig. 6.
The hidden neurons broaden their influence on the wide
range of neurons, although having a comparatively low influence on them. These connections are more sensitive to the
input changes and faster adapt the required firing pattern by
exciting the right output neurons. Fig. 7 shows that weights
are almost unalterable during the initial phase of simulation,
eventually changing their weights at the final simulation phase.
The input-hidden connections adapt faster due to their lower
number and only after their weights are settled, the hiddenoutput connections start actively adapting. The adaptation
process is shown in Fig. 6 and Fig. 7, where it is clearly
visible that the hidden-output weights start adapting.
Concluding, there are two levels of learning in the threelayer neural network. Whereas the weights of the input neurons
are less sensitive to changes, they better retain the learned patterns as opposed to the weights of the hidden neurons, which
being highly-alterable and rapidly adapting, easily ‘forget’ the
learned patterns.
This allows us to assume that input-hidden connections
are more influential to the adaption of the required output
connection compared to the hidden-output connections. Taking
into account that connections with large weights are highly
sensitive (according to [25]), their faulty behaviour is more
critical for the proper operation of the network. Moreover, after
the training phase is complete, a possible fault in the inputhidden connections is more devastating as it automatically
leads to a faulty output in the corresponding post-synaptic
nodes.
To assess the fault-tolerance of the neural network during
the training stage, the fault injection technique was applied.
Fig. 8 shows the dependence of the number of learning steps
on faults. The red line represents the dependence on the faulty
nodes and the green line represents the dependence on the
faulty synapses. The level of faults is represented in percents
rather than by the actual number of faults as this provides
a better representation of the scale of damage done to the
network.
The dependence is almost linear up until 0.3% for the
network with damaged nodes and up until 0.25% for the
network with damaged synapses. The faulty nodes reduce
the learning speed more significantly comparing to the faulty
synapses when the number of faults is small. However, this
changes to the opposite after the level of faults reaches 0.4%,
when the level of damage caused by faulty synapses becomes
larger than by the faulty nodes. The value 0.4% represent the
point when the number of faulty synapses introduced to the
7
Fig. 4. Weight distribution of input-hidden connections at the initial phase
(a), after the equilibration phase (b), and after the learning phase (c).
Fig. 5. Weight distribution of hidden-output connections at the initial phase
(a), after the equilibration phase (b), and after the learning phase (c).
network is equal to the number of faulty synapses caused by
the faulty nodes.
B. Fault Tolerance in Operation
We assessed the fault-tolerance of the neural network during
the operation phase. For this the fault injection technique
was applied. In order to collect statistically valid data each
simulation of the network of 2000 neurons (with 10 neurons
dedicated to input and another 10 neurons to output, taught
20 pairs of patterns) was run 1000 times. Each time a certain
number of faulty nodes was placed probabilistically according
to the described fault injection technique and quality of the
network output was measured by Eq. 2. The quality of output
represents the global error and stands for the probability of
receiving the expected output and is based on the maximum
number of different bits between the expected patterns and
the produced pattern. This experiment produced a plot of the
neural network fault tolerance against the number of faulty
nodes and synapses, by which the network reliability could
then be judged.
Fig. 9 shows the effect of faults injection while evaluating
a set of 20 patterns. We presented the damage volume in
percentage to the all possible damages in order to scale the
amount of damage with the size of the physical implementation. The figure shows that performance degradation is in
some sense graceful. According to the plot, 5% faulty nodes
guarantees 60% correct output and 10% faulty nodes reduces
the probability of the correct result to 50%. A network with
2% faulty nodes produces the correct result with a probability
of 90%. Fig. 10 presents the zoomed area of Fig. 9.
Any fault will influence the output to some degree since
all components participate in any computation. This leads to
graceful degradation being exhibited by most neural networks,
i.e. neural networks will not suffer catastrophic failure, and
also allows approaching failure to be detected by using a
continuous reliability measure. The fault tolerance that results
in this reliability is not inherent within neural networks: it does
Fig. 6. Weight alteration of connections between input and hidden layers
during the simulation progress (rotated for the best visibility).
Fig. 7. Weight alteration of connections between hidden and output layers
during the simulation progress (rotated for the best visibility).
8
VI. C ONCLUSIONS
Fig. 8. Dependence of the number of learning steps required to learn 20
neurons by 500-neuron size network on the faults number injected at the
network level during the learning process. The level of faults is presented in
percents.
Fig. 9. The quality of output against the amount of faulty neural network’s
nodes while recalling pre-learned 20 patterns using the network of 2000
neurons.
Faults do and will occur in a system over time, and
there always will come a time when performance is below
acceptable limits. The issue of fault tolerance is of particular
importance for the creation of large distributed machines
based on communication-centric computation. However unlike
traditional computing techniques, the neural network approach
does not insist on exact computation. There is the strongly nonlinear nature and the distribution of information or “knowledge” throughout all of the network.
We presented a distributed feed-forward neural network,
which is insensitive to intermittent faults caused by unreliable
communication or faulty hardware components. A review of
work examining the fault tolerance of neural networks has
been presented along with several techniques assuring faulttolerant behavior. Also, various possible influences on the fault
tolerance of neural networks were discussed. Among them,
we show that uniform weight distribution offers the particular
promise of more effective and faster training. We also show
that reducing the connectivity between input and hidden layers
is more advantageous prior to learning and the connections
between hidden and output layers are less vulnerable to faults
during learning and afterwards.
Almost all of the units and connections participate in
producing an output either directly or indirectly. Since it is
difficult to exactly determine the required amount of processing units and their connections, their redundant number results
in a higher degree of reliability. Thus the malfunctioning of
a particular element of a system should not greatly affect the
system’s function if there is sufficient redundancy.
This analysis offers several more general lessons for building reliable systems on unreliable hardware. Clearly some
redundancy is needed, and homogeneity is important, but selforganisation is also a necessary requirement.
R EFERENCES
Fig. 10. The quality of output against the amount of faulty neural network’s
nodes while recalling pre-learned 20 patterns using the network of 2000
neurons (zoomed version of Fig. 9).
need to be specifically designed and built into them, and so the
architectural complexity which often arises due to various fault
tolerance techniques being used is absent in neural network
systems. Finally, although any faults which do occur cannot
be located, they can be removed from the system since neural
networks can learn.
Although the experiments were performed on comparatively
small networks (about 2000 neurons) with a small training set,
the results are considered indicative for large networks with
large training sets due to the scalable nature of calculations.
We conclude that fault tolerance arises because the computation is distributed in the neural network rather than localized.
Moreover, the system is self-organizing rather than being
centrally configured.
[1] J. Dongarra, H. Meuer, E. Strohmaier, and H. Simon, “Top 500 supercomputer sites.” May 2012, [Online]. Available: http://www.top500.org/.
[2] S. S. Mukherjee, C. Weaver, J. Emer, S. K. Reinhardt, and T. Austin,
“A systematic methodology to compute the architectural vulnerability
factors for a high-performance microprocessor,” in Proc. MICRO 36,
2003, pp. 29–40.
[3] S. Furber and S. Temple, “Spinnaker a universal spiking neural network
architecture,” University of Manchester, Manchester, UK, Tech. Rep.,
Oct. 2010.
[4] S. Furber and D. Lester, “SpiNNaker project.” May 2012, [Online].
Available: http://apt.cs.man.ac.uk/projects/SpiNNaker/.
[5] H. Markram, R. Bishop, and R. Cicurel, “Blue Brain project.” May 2012,
[Online]. Available: http://bluebrain.epfl.ch/.
[6] J. Navaridas, M. Luján, J. Miguel-Alonso, L. A. Plana, and S. Furber,
“Understanding the interconnection network of SpiNNaker,” in Proc.
ICS, 2009, pp. 286–295.
[7] W. R. Moore, “Neural computers,” R. Eckmiller and C. v. d. Malsburg,
Eds. New York, NY, USA: Springer-Verlag New York, Inc., 1989, ch.
Conventional fault-tolerance and neural computers, pp. 29–37.
[8] M. Mozer and P. Smolensky, “Skeletonization: A technique for trimming
the fat from a network via relevance assessment,” in Proc. NIPS, D. S.
Touretzky, Ed. San Mateo: Morgan Kaufmann, 1988, pp. 107–115.
[9] H.-M. Tai, “Fault tolerance in neural networks,” Proc. WNN-AIND,
p. 59, Feb. 1990.
[10] L. Belfore and B. Johnson, “The fault-tolerance of neural networks,”
International Journal of Neural Networks Research and Applications,
pp. 24–41, Jan. 1989.
9
[11] C. Chiu, K. Mehrotra, C. Mohan, and S. Ranka, “Robustness of
feedforward neural networks,” Proc. ICNN, vol. 2, pp. 783–788, March
1993.
[12] M. J. Carter and B. Segee, “Fault tolerance of pruned multilayer
networks,” Proc. IJCNN, vol. 2, pp. 447–452, July 1991.
[13] M. J. Carter, F. J. Rudolph, and A. J. Nucci, “Advances in neural
information processing systems 2,” D. S. Touretzky, Ed. San Francisco,
CA, USA: Morgan Kaufmann Publishers Inc., 1990, ch. Operational
fault tolerance of CMAC networks, pp. 340–347.
[14] B. E. Segee and M. J. Carter, “Comparative fault tolerance of parallel
distributed processing networks,” IEEE Trans. Comput., vol. 43, no. 11,
pp. 1323–1329, Nov. 1994.
[15] J. Nijhuis, B. Hofflinger, A. van Schaik, and L. Spaanenburg, “Limits
to the fault-tolerance of a feedforward neural network with learning.”
IEEE Comput. Soc. Press, 1990, pp. 228–235.
[16] M. Emmerson and R. Damper, “Determining and improving the fault
tolerance of multilayer perceptrons in a pattern-recognition application.”
IEEE Trans. Neural Netw., vol. 4, no. 5, pp. 788–93, Sept. 1993.
[17] D. S. Phatak and I. Koren, “Complete and partial fault tolerance of
feedforward neural nets,” IEEE Trans. Neural Networks, vol. 6, pp. 446–
456, 1995.
[18] W. V. Seelen and H. A. Mallot, “Neural computers,” R. Eckmiller and
C. v. d. Malsburg, Eds. New York, NY, USA: Springer-Verlag New
York, Inc., 1989, ch. Parallelism and redundancy in neural networks, pp.
51–60.
[19] G. Bolt, “Technical Report YCS 154: Investigating Fault Tolerance
in Artificial Neural Networks,” Dept. Computer Architecture Group,
Universite of York, Tech. Rep., March 1991.
[20] G. Panchal, A. Ganatra, Y. P. Kosta, and D. Panchal, “Behaviour analysis
of multilayer perceptrons with multiple hidden neurons and hidden
layers,” in International Journal of Computer Theory and Engineering,
vol. 3(2), April 2011, pp. 332–337.
[21] G.-B. Huang, “Learning capability and storage capacity of two-hiddenlayer feedforward networks.” IEEE Trans. on Neural Netw., vol. 14,
no. 2, pp. 274–281, March 2003.
[22] D. J. Burr, “Experiments on neural net recognition of spoken and written
text,” IEEE Trans. Acoustics, Speech and Signal Processing, vol. 36,
no. 7, pp. 1162–1168, July 1988.
[23] Y. Bengio, “Learning deep architectures for AI,” Dept. IRO, Universite
de Montreal, Quebec, Canada, Tech. Rep., 2007.
[24] R. J. C. Bosman, W. A. van Leeuwen, and B. Wemmenhove, “Combining Hebbian and reinforcement learning in a minibrain model,” Neural
Netw., vol. 17, no. 1, pp. 29–36, 2004.
[25] C. Chiu, K. Mehrotra, C. K. Mohan, and S. Ranka, “Training techniques
to obtain fault tolerant neural networks,” in Proc. FTCS-24, June 1994,
pp. 360–369.
[26] A. Kulakov and M. Zwoliński, “Reducing the active paths interference
in the Chialvo-Bak minibrain model,” in Proc. ICCMS, vol. 2, Jan. 2011,
pp. 677–681.
| 9cs.NE
|
1
On the Outage Analysis and Finite SNR
Diversity-Multiplexing Tradeoff of
arXiv:1712.07781v1 [cs.IT] 21 Dec 2017
Hybrid-Duplex Systems for Aeronautical
Communications
Tan Zheng Hui Ernest, A S Madhukumar, Rajendra Prasad Sirigina,
and Anoop Kumar Krishna
Abstract
A hybrid-duplex aeronautical communication system (HBD-ACS) consisting of a full-duplex (FD)
enabled ground station (GS), and two half-duplex (HD) air-stations (ASs) is proposed as a direct solution
to the spectrum crunch faced by the aviation industry. Closed-form outage probability and finite signalto-noise ratio (SNR) diversity gain expressions in aeronautical communications over Rician fading
channels are derived for a successive interference cancellation (SIC) detector. Similar expressions are also
presented for an interference ignorant (II) detector and HD-equivalent modes at GS and ASs. Through
outage and finite SNR diversity gain analysis conducted at the nodes, and system level, residual SI and
inter-AS interference are found to be the primary limiting factors in the proposed HBD-ACS. Additional
analysis also revealed that the II and SIC detectors in the proposed HBD-ACS are suitable for weak
and strong interference scenarios, respectively. When compared to HD-ACS, the proposed HBD-ACS
achieves lower outage probability and higher diversity gains at higher multiplexing gains when operating
at low SNRs. Finite SNR analysis also showed the possibility of the proposed HBD-ACS being able to
attain interference-free diversity gains through proper management of residual SI. Hence, the proposed
Tan Zheng Hui Ernest is with the School of Computer Science and Engineering, Nanyang Technological University, Singapore
e-mail: ([email protected]).
A S Madhukumar is with the School of Computer Science and Engineering, Nanyang Technological University, Singapore
e-mail: ([email protected]).
Rajendra Prasad Sirigina is with the School of Computer Science and Engineering, Nanyang Technological University,
Singapore e-mail: ([email protected]).
Anoop Kumar Krishna is with Airbus Singapore Pte Ltd, Singapore e-mail: ([email protected]).
2
HBD-ACS is more reliable and can provide better throughput compared to existing HD-ACS at lowto-moderate SNRs.
Index Terms
Aeronautical Communications, Spectral Efficiency, Full-Duplex, Hybrid-Duplex, Half-Duplex, Outage Probability, Rician, Finite signal-to-noise ratio (SNR), Diversity.
I. I NTRODUCTION
Between 2012 and 2032, air travel within the Pacific South East Asia region is projected
to record a compounded annual growth rate of 5.3% [1]. This air travel growth trend exposes
existing aeronautical communication systems (ACSs) to considerable strain due to demand for
data communications from legacy, current and future generation avionics systems. Consequently,
this places an additional strain on existing Air-to-Ground (A/G) and Air-to-Air (A/A) aeronautical
communication links on the congested aeronautical spectrum. With existing ACSs being unable
to deliver the needed data capacity [2], various communication technologies have been proposed
to improve the capabilities of existing A/G and A/A links [2], [3]. However, these solutions do
not directly address the issue of spectrum utilization.
A hybrid-duplex (HBD) ACS consisting of half-duplex (HD) air-stations (ASs) operating
existing avionics systems with full-duplex (FD) ground stations (GSs) can be an alternative
solution to the shortage of available aeronautical spectrum currently faced by the aviation
community. Changes to existing/legacy HD avionics systems currently on board aircrafts can
be kept to a minimum in HBD-ACS, thus enabling HBD-ACS to be less disruptive to adopt.
Wireless communication systems that have adopted the HBD paradigm include cognitive radio
systems [4] and cellular systems [5], [6], [7].
In HBD systems, both FD and HD nodes communicate on the same spectrum since an FD node
can simultaneously transmit and receive signals on the same frequency and thereby improve the
spectral efficiency [8], [9], [10]. However, self-interference (SI) remains the primary challenge
faced by FD nodes due to simultaneous signal transmission and reception, with extensive studies
done on SI mitigation architectures. SI mitigation architectures can be categorized into either
passive suppression or active cancellation [11]. The former mitigates SI through induced path loss
(e.g. antenna separation) while the latter cancels SI in the analog or digital domain. The respective
SI cancellation architectures each have unique limitations. For instance, passive suppression may
3
not always be possible due to transceiver design while active cancellation architectures are limited
by the analog-to-digital converter’s dynamic range [9] and sophisticated hardware. Also, residual
SI is present due to inherent carrier phase noise and imperfect SI channel estimation at the FD
transceiver [11], which is also another limiting factor that practical aeronautical HBD systems
must overcome.
Nonetheless, practical FD systems is a possibility if SI is sufficiently mitigated [11], [12].
In this aspect, implementing suitable SI mitigation architectures for HBD-ACS opens up the
possibility of directly addressing the spectrum crunch faced by the aviation industry. In particular,
multiple aircrafts and ground stations can communicate on the same aeronautical spectrum,
providing motivation for this paper.
A. Related Literature
Apart from SI at FD nodes, HD nodes in HBD systems also experience interference due to
transmissions from other HD and FD nodes. In the literature, multiple interference management
approaches have been presented. However, this paper focuses on two widely known approaches
where interference is either ignored, i.e., interference ignorant (II) detector, or successfully
canceled, i.e., successive interference cancellation (SIC) detector.
To quantify the effectiveness of the II and SIC detectors, many related works in literature have
attempted to determine the closed-form outage probabilities of these detectors under various
fading models. For the II detector, closed-form outage expressions for Nakagami-m fading [13]
and composite fading consisting of exponentially distributed signal-of-interest (SOI) and squared
K-distributed interfering signals [14] have been noted. It should be pointed out that [13] and [14]
are only applicable to specific fading environments and may not be applicable for all aeronautical
scenarios where Rician fading is experienced. To this end, a recent paper by Rached et al. [15]
presented generalized outage probability expressions that apply to a wide variety of fading
scenarios, including Rician fading.
Multiple works on outage expressions for SIC detectors have been noted. For instance, SIC
outage expressions were investigated by Hasna et al. [16] and Romero-Jerez and Goldsmith [17],
but these studies only considered partial SIC where at least one interfering signal remains after
interference cancellation. A closed-form outage expression for SIC was studied by Weber et al.
[18] for nodes distributed via a Poisson point process. The work in [18] did not consider fading
and receiver noise in the signal model, and thus, the closed-form expressions are not directly
4
applicable for aeronautical communications. A recent paper by Zhang et al. [19] presented
outage probability expressions for a two-stage SIC detector. However, the outage expressions
are specific for Rayleigh fading scenarios and are not applicable to Rician fading scenarios that
are common in aeronautical communications. From the mentioned studies, hitherto closed-form
outage probability expressions for SIC detectors in Rician fading aeronautical scenarios remain
an open problem.
Apart from outage probability, both finite signal-to-noise ratio (SNR) diversity gain and
finite SNR diversity-multiplexing trade-off (DMT) are metrics that can be used to measure the
effectiveness of II or SIC detectors in fixed and variable transmission rate systems, respectively.
In particular, both finite SNR diversity gain and finite SNR DMT quantifies the slope of outage
probability curves at a particular SNR [20], with the latter considering multiplexing gain [21].
Finite SNR analysis can reveal outage deviation behaviors, which are not present at asymptotically high SNRs due to fading statistics [20]. From a practical perspective, analyzing outage
probability decay rates, i.e., finite SNR diversity gain, provides an accurate picture of a system’s
outage performance since wireless communication systems are typically designed to operate at
low-to-moderate SNR ranges. It has also been pointed out by Narasimhan [21] that finite SNR
diversity gain analysis can be used to estimate the SNR needed to achieve a particular rate of
error decay, which can be done through turbo codes or low-density parity-check codes. More
crucially, outage probability and diversity gain can be used to gauge the upper and lower limits
of a system’s bit error rate performance [22], [23].
Finite SNR analysis for Nakagami-m [24] and Rayleigh fading [25], [26] scenarios have
also been studied. However, the conclusions drawn in these studies are specific to Nakagamim and Rayleigh fading and are not fully applicable for ACS since Rician fading scenarios,
typically encountered by ACS, are not considered. Studies on finite SNR analysis for Rician
fading channels have been seen. The impact of Rician K factors on outage behavior and finite
SNR DMT for multiple-input multiple-output (MIMO) systems was investigated by Narasimhan
[21] and Shin et al. [20]. A recent paper by Heidarpour et al. [27] saw finite SNR DMT analysis
being applied to analyze the performance of a network coded cooperative communication system.
Despite the noted studies, there is still room for further work on finite SNR DMT analysis for
HBD-ACS.
5
B. Main Contributions
The main contributions of this paper are summarized as follows:
•
The present paper proposes an innovative approach for deriving closed-form expressions
for outage probability for a II detector and a two-stage SIC detector in a Rician faded
environment.
•
It is shown that the proposed HBD-ACS attains superior outage performance over existing
HD-ACS at low SNRs. At high SNRs, however, the outage performance of the proposed
HBD-ACS is eclipsed by HD-ACS as the former becomes interference-limited at asymptotic
SNRs. Nonetheless, we show through numerical simulations that the HBD-ACS can meet
typical Quality-of-Service (QoS) requirements, e.g., frame error rate ≤ 10−3 , at high SNRs
for a range of interference levels through II and SIC detectors.
•
Closed-form finite SNR diversity gain expressions are derived for the II and SIC detectors
under Rician fading. The asymptotic behavior of the derived finite SNR diversity gains for
HBD-ACS and HD-ACS are proven and shown to be consistent with interference-limited
outage behaviors at asymptotic SNRs.
•
The proposed HBD-ACS is shown to achieve better diversity gains than HD-ACS at high
multiplexing gains. Finite SNR DMT analysis reveals that operating at higher multiplexing
gain causes the Rician K factor, corresponding to the SOI, to have more impact on HBDACS outage performance. Additionally, reducing residual SI and interference from AS-1
leads to steeper decay of outage probability, improving the finite SNR DMT curve of the
proposed HBD-ACS as a consequence.
C. Relevance to Related Literature
In this work, full interference cancellation is assumed for the two-stage SIC detector. This is
unlike in [16] and [17] where only partial SIC is assumed. In addition, the impact of interference
on the proposed HBD-ACS is analyzed from the outage probability and finite SNR DMT
perspective, which was not covered in [4] - [7], [13] - [15], [18] and [19]. In contrast to [20], [21]
and [27], this work extends upon the outage and finite SNR DMT analysis framework to jointly
identify interference scenarios for the proposed single-input-single-output equivalent HBD-ACS.
The remainder of this paper is organized as follows. The system model is introduced in Section
II, with closed-form outage probability expressions at GS and AS-2 presented in Section III. In
Section IV, finite SNR diversity gain expressions for both HBD-ACS and HD-ACS are derived
6
Fig. 1. Air-Station 1 (AS-1) and Air-Station 2 (AS-2) operating in HD mode while communicating with the FD ground station
(GS).
and analyzed. Numerical results are then presented in Section V before the conclusion of the
paper in Section VI.
II. S YSTEM M ODEL
In this paper, A/G communications involving an FD-enabled GS node with two HD ASs in
an A/G link is studied. Specifically, a scenario with Air-Station 1 (AS-1) transmitting signals
to the GS while Air-Station 2 (AS-2) is receiving signals from the GS is assumed. Due to the
fact that the GS node is FD-capable, the HD AS-1 and HD AS-2 simultaneously transmits and
receives, respectively, signals on the same aeronautical spectrum (e.g. VHF, L-band) as the GS.
Therefore, AS-1 interferes with communications at AS-2 when the latter receives signals from
GS.
At the FD-enabled GS, a combined pre-mixer and post-mixer SI cancellation architecture
is adopted to mitigate inherent SI, as shown in Fig. 1. An analog domain pre-mixer canceler
first mitigates the SI signal before down-conversion occurs. After analog SI mitigation, a digital
domain post-mixer canceler further suppresses SI in the digital domain. Combining the premixer and post-mixer cancellation architectures in succession results in less residual SI [11,
Table II]. The main drawback, however, is the need for separate radio chains, which introduces
unnecessary carrier phase noise. To mitigate the phase noise effect, the proposed SI cancellation
architecture shares a common local oscillator (LO), as seen in Fig. 1. Assuming SI mitigation
at GS, residual SI will be considered due to imperfect SI channel estimation [11]. Thus, an II
detector is assumed at GS since signal detection is performed in the presence of residual SI.
Rician fading aeronautical communications channels in an en route scenario is assumed to
provide a realistic evaluation of the HBD-ACS [28]. Accordingly, we assume that the ASs are
7
communicating with the GS at cruising altitude, with the signal model of this work based on
[11].
A. Ground Station
Let x1 [t] and xgs [t] be the signals transmitted by AS-1 and GS, respectively, and h1,g [t] be
the channel from AS-1 to GS. Additionally, let xsi [t] be the SI signal at GS and let hsi be the
SI channel gain. From the perspective of GS, xsi [t] = xgs [t]. The received signal at GS can be
written as
ygs [t] =
p
p
p
Ω X h1,g [t]x1 [t] + Ω X αg,g · |e
hsi |xsi [t] + Ω X αg,g |hsi |γφ wφ [t] + wg [t],
(1)
where e
hsi is the error of the imperfect SI channel gain estimate, defined as e
hsi = hsi − b
hsi , and
b
hsi is the imperfect estimation of the SI channel gain. In addition, let e
hsi be modeled as a zero
mean, circularly symmetric complex Gaussian random variable (RV) with variance ǫ to quantify
the SI channel estimation error [29]. Modeling e
hsi as a zero mean Gaussian RV with variance ǫ
enables the system to model the worst case residual SI [29]. Also, let wg [t] be the GS additive
white Gaussian noise (AWGN) with zero mean and variance σg2 , and let the phase noise term
wφ [t] follow a Gaussian distribution with zero mean and unit variance, scaled by the strength
of the phase noise γφ [11].
Let Ω X be the average received signal power of the signal-of-interest (SOI). The average
received signal power is defined based on the free space path loss model and it is defined as
ΩX =
Pt
,
2
4·π·109
2
n
2
· fc ·d1,g ·σg
3·108
(2)
where Pt , d, fc and n are the transmit power (Watts), distance (Km), carrier frequency (MHz)
and pathloss exponent, respectively. The received signal power levels are normalized with the
receiver noise variance (σg2 ). The channel between AS-1 and GS is selected as the reference link
and the average received signal power in the other links are expressed relative to the reference
link via the multiplicative factor αi, j , defined as
d1,g n
, i ∈ {g, 1} , j ∈ {g, 2} .
αi, j =
di, j
From (2) and (3), the average received SI power at GS can be expressed as Ω X αg,g .
(3)
8
B. Air-Station 2
Let hg,2 [t] be the channel between GS and AS-2, h1,2[t] be the channel between AS-1 and
AS-2, and w2 [t] be the AWGN at AS-2 with zero mean and variance σ22 . From the perspective
of AS-2, xgs [t] and x1 [t] are the SOI and interfering signal, respectively. The received signal at
AS-2 can be expressed as
y2 [t] =
p
p
Ω X αg,2 hg,2 [t]xgs [t] + Ω X α1,2 h1,2[t]x1 [t] + w2 [t],
(4)
where Ω X αg,2 and Ω X α1,2 indicate the average received signal powers of the SOI and interfering
signal, respectively.
To handle the interference at AS-2, two approaches are studied. The first approach assumes
an II detector at AS-2. The II detector treats x1 [t] as noise. Therefore, interference is effectively
ignored. The second approach assumes a SIC detector at AS-2. The two-stage SIC detector first
tries to detect and cancel x1 [t] before proceeding to detect xgs [t] [30].
III. C ALCULATION
OF
O UTAGE P ROBABILITIES
To begin the outage analysis at GS and AS-2, we first define the HBD transmission rates
HBD , respectively, and the sum rate of the HBD system as
of AS-1 and GS as R1HBD and Rgs
HBD
HBD
= R1HBD + Rgs
. Similarly, the HD transmission rates of AS-1 and GS are defined as
Rsum
HD , respectively, and the sum rate of the HD system is defined as R HD = R HD + R HD .
R1HD and Rgs
sum
gs
1
For fair comparison between HBD and HD systems, RiHBD = 21 RiHD for i ∈ {1, gs} [31], [32],
[33]. The respective HBD and HD outage probabilities at GS and AS-2 are defined in the
following subsections.
A. Hybrid-Duplex Outage Probability
The FD-enabled GS receives x1 [t] while simultaneously transmitting xgs [t] in the same time
slot. The simultaneous transmission and reception of signals result in strong SI at GS. Let
X1 = Ω X |h1,g | 2 be the instantaneous received signal power of the SOI at GS, modeled as a noncentered Chi-squared distributed RV with Rician K factor K X1 . Let Ysi,1 = Ω X αg,g γφ2 |hsi | 2 and
Ysi,2 = Ω X αg,g |e
hsi | 2 be the instantaneous received signal power corresponding to SI components.
In particular, Ysi,1 is modeled as a non-centered Chi-squared distributed RV with Rician K factor
KYsi,1 and Ysi,2 is modeled as a exponentially distributed RV.
9
Concurrently, AS-2 also experiences interference from AS-1. Let Xgs = Ω X αg,2 |hg,2 | 2 and
Y1 = Ω X α1,2 |h1,2 | 2 be the instantaneous received signal power of the SOI and interference at
AS-2, respectively, where Xgs and Y1 are independent non-centered Chi-squared distributed RV
with respective Rician K factors K Xgs and KY1 , respectively. Additionally, let α q, Ω, K, γ be
defined as
Lq (0) (K) (1 + K)
γ
α q, Ω, K, γ = (−1)q exp(−K)
(1 + q)!
Ω
! q+1
(5)
,
where q, Ω, K and γ represent an arbitrary non-negative integer, average received power of the
signal, Rician K factor, and threshold, respectively. The function Lq (0) (•) represents the q-th
degree, zero-order Laguerre polynomials [34] while α q, Ω, K, γ in (5) represents the Rician
power cumulative distribution function (CDF) expansion due to Rician faded signal parameters.
H BD
HBD = 2 R1
1) Ground Station: At GS, let the HBD threshold be γth,gs
− 1, with HBD outage
o
n
X1
HBD = h , h : R HBD ≥ log 1 +
event Ogs
1,g si
2
Ysi,1 +Ysi,2 +1 . By substituting X1, Ysi,1 and Ysi,2 into
1
[15, eq. (12)], the closed-form outage probability at GS can be expressed as
Õ
Õ
l1
l2
HBD
HBD (q + 1)!
Pr Ogs
=
E {Ysi,1
}E {Ysi,2
},
(6)
α q, Ω X , K X1, γth,gs
l1 ! · l2 ! · l3 !
q≥0 l1 +l2 +l3 =q+1
HBD is the Rician SOI power CDF expansion at GS, as defined in (5).
where α q, Ω X , K X1, γth,gs
l1
l2
In addition, E {Ysi,1
} and E {Ysi,2
} are the l1th and l2th moments of Ysi,1 and Ysi,2 , respectively.
α γ 2 l1
g,g
l2
l1
} = Γ(1 +
From [15, Table II], E {Ysi,1
} = Γ(1 + l1 ) 1+KY φ 1 F1 (−l1, 1; −KYsi,1 )(Ω X )l1 and E {Ysi,2
si,1
l2
l2 )(αg,g ǫ (Ω X )l2 . The function 1 F1 (•) represents the confluent Hypergeometric function [35] and
HBD ≤
summation on the right hand side (RHS) of (6) is convergent if γth,gs
eq. (14)]. In (6),
l1
E {Ysi,1
}
and
l2
E {Ysi,2
}
ΩX
3(1+K X1 )(ΩX αg,g ǫ)
[15,
quantifies the strength of residual SI, with phase noise
l1
l2
and SI channel estimation errors quantified by E {Ysi,1
} and E {Ysi,2
}, respectively. To show the
l1
impact of residual SI on the FD-enabled GS, we evaluate the limit of the αg,g term in E {Ysi,1
}
from (6) as follows.
if L = ∞, l1 > 0
∞
Ω γ 2 l1
l1
lim E {Ysi,1
} = Γ(1 + l1 ) 1+KX φ
1 F1 (−l1, 1; −KYsi,1 ) if L ∈ {0, ∞}, l1 = 0
Ysi,1
αg,g →L
0
if L = 0, l1 > 0
(7)
l2
Similar results are also yielded when the limit of αg,g in E {Ysi,2
} is evaluated. We do not
expect αg,g to approach infinity as the distance on the SI link (dg,g ) cannot be zero. However, it
is possible for the average received SI power to be strong if dg,g is short. From (7), the impact
10
of residual SI is diminished as αg,g → 0 and hence, proper SI mitigation strategies is crucial at
the FD-enabled GS.
HBD =
2) Air-Station 2 (Interference Ignorant Detector): At AS-2, let the HBD threshold be γth,2
o
n
H BD
HBD(I I)
HBD ≥ log 1 + Xgs
2 Rgs − 1 and the HBD outage event be O2
= hg,2, h1,2 : Rgs
2
Y1 +1 . By
substituting Xgs and Y1 into [15, eq. (12)], the closed-form outage probability at AS-2 can be
expressed as
Pr
HBD(I I)
O2
=
q+1
ÕÕ
α
HBD
q, Ω X αg,2, K Xgs , γth,2
q+1
E {Y1l },
l
(8)
HBD is the Rician SOI power CDF expansion at AS-2, as defined
where α q, Ω X αg,2, K Xgs , γth,2
q≥0 l=0
in (5), and E {Y1l } is the l th moment of the interfering signal from AS-1. From [15, Table II],
il
h
α
HBD ≤
E {Y1l } = Γ(1 + l) 1+K1,2Y 1 F1 (−l, 1; −KY1 )(Ω X )l and the RHS of (8) is convergent if γth,2
ΩX αg,2 (1+KY1 )
2(1+K Xgs )ΩX α1,2
1
[15, eq. (14)]. In (8), E {Y1l } quantifies the strength of the interference from AS-1
through moment parameters of Y1 .
To investigate the impact of inter-AS interference, we evaluate limα1,2 →L E {Y1l } for L ∈ {0, ∞}
using the same approach in (7). Although α1,2 does not reach infinity in practice, large values of
α1,2 are possible when d1,2 is small and vice-versa. Evaluating limα1,2 →L E {Y1l } for L ∈ {0, ∞}
shows that the impact of inter-AS interference reduces as α1,2 → 0 and thus, the II detector
operates effectively in low interference scenarios such as over remote airspace where inter-AS
distance is long.
3) Air-Station 2 (Successive Interference Cancellation Detector): In the case of SIC, if the
first stage is unable to detect the interfering signal or if the SOI cannot be detected at the second
stage, then outage occurs. Therefore, the HBD outage event at AS-2 is defined as
Y1
HBD(SIC)
HBD
= hg,2, h1,2 : R1
> log2 1 +
O2
1 + Xgs
Y1
HBD
HBD
∪ hg,2, h1,2 : R1
≤ log2 1 +
, Rgs > log2 1 + Xgs .
(9)
1 + Xgs
Theorem 1: The closed-form expression for outage probability with SIC detector at AS-2 is
!
q+1
ÕÕ
q
+
1
HBD
l
Pr O2HBD(SIC) =
α q, Ω X α1,2, KY1, γth,gs
E {Xgs
}
(10)
l
q≥0 l=0
s
!
HBD
q
2(K Xgs + 1)γth,2
+ 1 − Q 1 2K Xgs ,
Ω X αg,2
(γ HBD ) j+n−i+1 !
n Õ
i+1
ÕÕ
i
+
1
th,2
HBD
−
,
α i, Ω X α1,2, KY1, γth,gs
α n − i, Ω X αg,2, K Xgs , 1
j
+
n
−
i
+
1
j
n≥0 i=0 j=0
11
l } = Γ(1 + l)
where Q 1 (·, ·) is the Marcum Q function [34] and E {Xgs
[15, Table II].
ΩX αg,2 l
1 F1 (−l, 1; −K Xgs )
1+K Xgs
Proof: The proof can be found in Appendix A.
The first term in (10) is the outage probability due to detecting interference from AS-1. The
second term in (10) is the outage probability due to SOI detection after interference cancellation.
HBD from (10) yields
Similar to (7), evaluating the limit of α1,2 in α q, Ω X α1,2, KY1, γth,2
HBD
∞ if L = 0
(11)
0 if L = ∞
From (11), it is evident that the SIC detector works effectively in high interference scenarios,
lim α q, Ω X α1,2, KY1, γth,2
α1,2 →L
=
such as in congested airspace where inter-AS distance is short, since the effect of interference
at the SOI detection stage is diminished when α1,2 is large. The closed-form expressions in (6),
(8) and (10) can shed insights into the impact of residual SI at GS and interference from AS-1
at AS-2. Further discussions on outage performance with respect to the level of interference are
presented in Section V.
B. Half-Duplex Outage Probability
When the GS is operating in HD mode, AS-2 does not experience interference from AS-1.
H BD
HD = 22R1
Let the HD threshold at GS and AS-2 be defined as γth,gs
H BD
HD = 22Rgs
− 1 and γth,2
− 1,
respectively. Then, the HD outage probabilities at GS and AS-2 are given in (12) and (13),
respectively [15, Table I].
HD
Pr Ogs
Pr
O2HD
=
Õ
Õ
m≥0
=
m≥0
HD
α m, Ω X , K X1, γth,gs
HD
α m, Ω X αg,2, K Xgs , γth,2
.
(12)
(13)
The outage probability expressions in (12) and (13) can be used as a benchmark comparison
against HBD mode at GS and AS-2, respectively, which is presented in Section V.
C. System Level Outage Probability
For the proposed multi-user system, the overall system level outage probability is used as a
performance metric to compare HBD and HD protocols. For β ∈ {HBD(I I), HBD(SIC)}, the
12
system level outage probability is defined as
β
Pout,system
HBD
Ogs
, Pr
β
O2
,
= max Pr
HD
HD
Pout,system
= max Pr Ogs
, Pr O2HD .
(14)
(15)
In particular, (14) provides the worst case system level outage behavior for the II and SIC
detectors and allows the identification of performance bottlenecks in HBD-ACS.
IV. F INITE SNR A NALYSIS
In the following subsections, the mathematical preliminaries related to finite SNR diversity
gain are presented for both fixed and variable transmission rates, with detailed derivation omitted
for brevity.
A. Mathematical Preliminaries
1) Finite SNR Diversity Gain: For a system with outage event O, outage probability Pr O ,
transmission rate R, threshold γ, and average received power Ω with unit noise variance, the
diversity gain d at high SNR is given by Zheng and Tse [22] as
log2 (Pr O )
.
d = lim
Ω→∞
log2 (Ω)
(16)
The diversity gain definition in (16) is for systems that operate at high SNR ranges. The
finite SNR diversity gain d f , which quantifies the decay rate of the outage probability at lowto-moderate SNRs, is given as [21, eq. (5)]
df =
−Ω ∂
Pr O .
Pr O ∂Ω
(17)
It has since been shown by Shin et al. [20] and Heidarpour et al. [27] that limΩ→∞ d f = d.
Therefore, (17) is consistent with the asymptotic diversity definitions in [22] at high SNR.
Practical wireless systems typically operate at the low-to-moderate SNR range [21]. The outage
behavior of these systems may also be different at high and moderate SNRs. Therefore, there
is motivation to quantify diversity gains at finite SNRs since (16) does not accurately reflect
outage behaviors at low-to-moderate SNRs [20].
13
2) Finite SNR DMT Parameters: For a system which varies its transmission rate with respect
to Ω, i.e., variable transmission rate, the high SNR multiplexing gain r is given by Zheng and
Tse [22] as
R(Ω)
Ω→∞ log2 (Ω)
(18)
r = lim
and the finite SNR multiplexing gain r f for such systems is [21, eq. (4)]
rf =
R(Ω)
.
log2 (1 + Ω)
(19)
It has similarly been shown by Shin et al. [20] and Heidarpour et al. [27] that limΩ→∞ r f = r,
with Pr O computed with respect to the threshold γ = (1 + Ω)r f − 1. The finite SNR diversity
gain for such a variable transmission rate system (denoted as d ∗f ) can be obtained from (17) as
[20, eq. (36)]
d ∗f
h Pr O
b − Pr O i
−Ω
lim
,
=
∆(Ω)
Pr O ∆(Ω)→0
(20)
b is the outage event with respect to R Ω+∆(Ω) . Furthermore,
b
b > Ω and O
where ∆(Ω) = Ω−Ω,
Ω
b is the outage probability with average received power Ω
b = Ω + ∆(Ω), threshold b
Pr O
γ =
b = Pr O when ∆(Ω) = 0. Applying L’Hospital’s rule in (20) by
[1 + Ω + ∆(Ω)]r f − 1 and Pr O
differentiating with respect to ∆(Ω) and setting ∆(Ω) = 0 yields
d ∗f =
−Ω
∂
b
Pr O
Pr O ∂∆(Ω)
.
(21)
∆(Ω)=0
Let Z be a RV with normalized nth moment defined as
M {Z n } =
E {Z n }
(Ω)n
and let function g(i, j, Ω, r f ) be defined as
j
rf
rf
i
j
r f −1
g(i, j, Ω, r f ) = 1 + Ω − 1 (Ω)
1+Ω −1
,
+ (i + 1)(r f )(1 + Ω)
Ω
(22)
(23)
where i and j are integers. The function M {Z n } represents the normalized nth moment of an
interfering signal while the function g(i, j, Ω, r f ) reflects the outage probability decay rate of a
variable transmission rate scheme due to average received power (Ω) and finite SNR multiplexing
gain (r f ).
Although [20, eq. (36)] and [21, eq. (5)] evaluate finite SNR diversity gains using different
approaches, the principles underlying them are the same since the latter is an extension of the
14
former. To this end, (21) can be used to evaluate d ∗f for adaptive systems, with r f indicating the
sensitivity of the rate adaptation scheme [21]. It is also of interest to analyze d ∗f as it can lead
to better code designs that improve transmission rates at the expense of reliability for adaptive
systems and vice-versa.
B. Finite SNR Diversity Gain for HBD Systems
HBD,i
Let the finite SNR HBD diversity gain at GS and AS-2 be defined as d HBD
f ,gs and d f ,2 , i ∈
{I I, SIC}, respectively. Additionally, let R1 and Rgs be fixed constants with average received
power Ω = Ω X . Then, the finite SNR diversity gain at GS and AS-2 are presented in the following
theorems.
Theorem 2: The finite SNR diversity gain at the FD-enabled GS in the proposed HBD-ACS
is
d HBD
=
f ,gs
Õ
−Ω X
HBD
Pr Ogs
q≥0 l
Õ
HBD
α q, 1, K X1, γth,gs
1 +l2 +l3 =q+1
× (l1 + l2 − q − 1) (Ω X )l1 +l2 −q−2 ,
(q + 1)!
l1
l2
M {Ysi,1
}M {Ysi,2
}
l1 ! · l2 ! · l3 !
(24)
Proof: The finite SNR diversity gain at GS can be obtained by substituting (6) into (17).
At low-to-moderate Ω X , the outage behavior at GS can be analyzed from (24). In particular,
(24) allows observation of subtle changes in outage behavior due to the scaling factor associated
with the SI strength (αg,g ) and SI channel estimation error (ǫ) that is not present at high Ω X . In
addition, the asymptotic behavior of d HBD
f ,gs can be obtained from (24) as shown in the following
corollary.
Corollary 1: The asymptotic behavior of d HBD
f ,gs is given by
∂
−Ω X
HBD
Pr Ogs
= 0.
HBD
ΩX →∞ Pr Ogs
∂Ω X
lim
(25)
Proof: From (6) and (24), (Ω X )l1 +l2 −q−1 < 1 when l1 +l2 +l3 ≤ q. Thus, limΩX →∞ (Ω X )l1 +l2 −q−1
= 0, l1 +l2 +l3 ≤ q. Therefore, only l1 +l2 +l3 = q +1 needs to be considered, which consequently
leads to the numerator in (24) to be zero, i.e., l1 + l2 − q − 1 = 0.
From (25), d HBD
f ,gs → 0 as Ω X → ∞ because increasing Ω X also causes residual SI to be
stronger, hence there is no improvement in the overall SINR. Also, (25) suggests that the tolerance
for residual SI in HBD-ACS is progressively diminished as Ω X is increased since d HBD
f ,gs → 0
corresponds to negligible improvements in outage probability at GS.
15
HBD(I I)
Theorem 3: The finite SNR diversity gain at AS-2 with the II (d f ,2
HBD(SIC)
)
(d f ,2
HBD(I I)
d f ,2
HBD(SIC)
d f ,2
are
) and SIC detectors
q+1
M {Y1l }(l − q − 1)(Ω X )l−q−2 (26)
α
=
HBD(I I)
l
Pr O2
q≥0 l=0
"
!
q+1
ÕÕ
−Ω X
q
+
1
l
HBD
=
M {Xgs
}(l − q − 1)(Ω X )l−q−2
α q, α1,2, KY1, γth,gs
HBD(SIC)
l
Pr O
q≥0 l=0
−Ω X
+
Õ
2
α
q+1
ÕÕ
HBD
q, αg,2, K Xgs , γth,2
HBD
m, αg,2, K Xgs , γth,2
(−m
−m−2
− 1)(Ω X )
−
n Õ
i+1
ÕÕ
HBD
α i, α1,2, KY1, γth,gs
!#
i + 1 γth,2
(−n − 2)(Ω X )−n−3 ,
× α n − i, αg,2, K Xgs , 1
j +n−i +1
j
m≥0
n≥0 i=0 j=0
HBD j+n−i+1
(27)
, i ∈ {I I, SIC} can be obtained for the II and SIC by respectively
Proof: At AS-2, d HBD(i)
f ,2
substituting (8) and (10) into (17).
The outage behavior at AS-2 can be analyzed from (26) and (27) at low-to-moderate Ω X .
In particular, (26) and (27) enables the observation of subtle changes in outage behavior for
both II and SIC detectors, which are not present at high Ω X , as inter-aircraft interference varies.
HBD(i)
Extending upon (26) and (27), the asymptotic behavior of d f ,2
, i ∈ {I I, SIC} can be obtained
as follows.
HBD(i)
Corollary 2: The asymptotic behavior of d f ,2
lim
ΩX →∞
−Ω X
Pr
O2HBD,i
, i ∈ {I I, SIC} is given by
∂
Pr O2HBD,i = 0.
∂Ω X
(28)
I)
, the approach seen in (25) can be used. Starting
Proof: To evaluate limΩX →∞ d HBD(I
f ,2
HBD(I I)
with the denominator of d f ,2
, (Ω X )l−q−1 < 1 when l ≤ q. Thus, limΩX →∞ (Ω X )l−q−1 = 0
when l ≤ q. In the numerator, (l − q − 1) (Ω X )l−q−2 = 0 when l = q + 1. Similarly, to
. Specifically,
, we first begin with the denominator of d HBD(SIC)
evaluate limΩX →∞ d HBD(SIC)
f ,2
f ,2
limΩX →∞ (Ω X )l−q−1 = 0 when l ≤ q and (Ω X )l−q−1 = 1 when l = q + 1. For (Ω X )−m−1 ,
limΩX →∞ (Ω X )−m−1 = 0 for m ≥ 0 and for (Ω X )−n−2 , limΩX →∞ (Ω X )−n−2 = 0 for n ≥ 0. In
the numerator, (l − q − 1) (Ω X )l−q−2 = 0 when l = q + 1. Additionally, limΩX →∞ (Ω X )−m−2 = 0
when m ≥ 0 and limΩX →∞ (Ω X )−n−3 = 0 when n ≥ 0.
In the presence of interference at AS-2, (28) shows that improvements to outage probability
→ 0 as Ω X → ∞ for i ∈ {I I, SIC}. For the II
at AS-2 progressively diminishes since d HBD(i)
f ,2
detector, increasing Ω X results in strong interference. As a consequence, there is no improvement
16
to the overall SINR. Hence, the II detector is unsuitable in strong interference environments.
Similarly, for the SIC detector, increasing Ω X causes xgs [t] to be stronger, making the detection
and subtraction of x1 [t] increasingly challenging at stage 1 of the SIC detector. Hence, α1,2 must
either increase (for the II detector) or decrease (for the SIC detector) at high Ω X for HBD-ACS
to see meaningful improvements in outage probability.
C. Finite SNR Diversity Gain for HD Systems
Let the finite SNR diversity gain at GS and AS-2 be defined as d HD
f ,i , i ∈ {gs, 2}, respectively,
with R1 and Rgs assumed to be constants with average received power Ω = Ω X . Then, the finite
SNR diversity gain at GS and AS-2 are presented in the following theorem.
Theorem 4: The finite SNR diversity gain at GS and AS-2 operating in HD mode are given
in (29) and (30), respectively.
−Ω X Õ
HD
α
m,
1,
K
,
γ
(−m − 1)(Ω X )−m−2
d HD
=
X
1
th,gs
f ,gs
HD
Pr Ogs m≥0
−Ω X Õ
HD
−m−2
α
m,
α
,
K
,
γ
.
d HD
=
g,2
X
gs
th,2 (−m − 1)(Ω X )
f ,2
HD
Pr O2
m≥0
(29)
(30)
Proof: The expressions in (29) and (30) can be obtained by respectively substituting (12)
and (13) into (17).
The HD outage behavior at GS and AS-2 can be analyzed from (29) and (30), respectively,
and it enables the observation of changes in outage probability decay rate that is not visible at
high Ω X . As Ω X → ∞, (29) and (30) can be evaluated to determine the asymptotic diversity
gain as follows.
Corollary 3: The asymptotic behavior of d HD
f ,i , i ∈ {gs, 2} is
lim d HD
f ,i = 1.
ΩX →∞
(31)
Proof: At GS, the asymptotic behavior of d HD
f ,gs can be easily evaluated after some simplifications as shown below:
lim d HD
f ,gs = lim
ΩX →∞
ΩX →∞
−
Í
HD
m, 1, K X1, γth,gs
(−m − 1)(Ω X )−m
.
Í
HD
−m
m≥0 α m, 1, K X1, γth,gs (Ω X )
m≥0 α
(32)
From (32), It can be seen that limΩX →∞ (Ω X )−m = 1 when m = 0, and limΩX →∞ (Ω X )−m = 0
when m > 0. Thus, when evaluating (32), only m = 0 needs to be considered. The asymptotic
behavior of d HD
f ,2 can also be proven using the same approach.
From (32), d HD
f ,i → 1 as Ω X → ∞ for i ∈ {gs, 2} and it indicates that the HD system achieves
full diversity in the absence of interference at high Ω X , which is consistent with [20, Fig. 3].
17
D. Finite SNR DMT Analysis for HBD Systems
Let the finite SNR diversity gain at GS for a HBD system be defined as d HBD∗
f ,gs , with variable
HBD = [1 + Ω ]r f − 1. Similarly,
transmission rate R1HBD (Ω X ) = r f log2 (1 + Ω X ) and threshold γth,gs
X
, i ∈ {I I, SIC},
let the finite SNR diversity gain at AS-2 for a HBD system be denoted as d HBD(i)∗
f ,2
HBD (Ω ) = r log (1+Ω ) and threshold γ HBD = [1+Ω ]r f −1.
with variable transmission rate Rgs
X
f
X
X
2
th,2
The finite SNR diversity gains at GS and AS-2 are presented in the following theorems.
Theorem 5: At GS, the finite SNR diversity gain is given as
Õ
Õ
(q + 1)!
−Ω X
l1
l2
M {Ysi,1
}M {Ysi,2
}
α
q,
1,
K
,
1
d HBD∗
=
X
1
f ,gs
HBD
l1 ! · l2 ! · l3 !
Pr Ogs
q≥0 l +l +l =q+1
1
2
3
× g(q, l1 + l2 − q − 1, Ω X , r f ),
(33)
HBD
b=Ω
b X and γ = γ HBD = [1 + Ω X ]r f − 1 and O = Ogs
Proof: Let Ω = Ω X , Ω
. Then, d HBD∗
th,gs
f ,gs
can be obtained through algebraic manipulations by substituting (6) into (21).
Theorem 6: At AS-2, the finite SNR diversity gain with II and SIC detectors are:
q+1
ÕÕ
q+1
−Ω X
HBD(I I)∗
=
d f ,2
M {Y1l }g(q, l − q − 1, Ω X , r f ) (34)
α q, αg,2, K Xgs , 1
HBD(I I)
l
Pr O2
q≥0 l=0
"
!
q+1
Õ
Õ
q
+
1
−Ω
X
HBD(SIC)∗
l
M {Xgs
}g(q, l − q − 1, Ω X , r f )
α q, α1,2, KY1, 1
=
d f ,2
HBD(SIC)
l
Pr O2
q≥0 l=0
+
Õ
m≥0
n Õ
i+1
ÕÕ
α m, αg,2, K Xgs , 1 g(m, −m − 1, Ω X , r f ) −
α i, α1,2, KY1, 1
× α n − i, αg,2, K Xgs , 1
i+1
j
n≥0 i=0 j=0
j +n−i+1
!#
g( j + n + 1, −n − 2, Ω X , r f ) .
(35)
b=Ω
b X and γ = γ HBD = [1 + Ω X ]r f − 1 and O = O HBD,i, i ∈ {I I, SIC}.
Proof: Let Ω = Ω X , Ω
th,2
2
HBD(I I)∗
Then, similar to (33), d f ,2
HBD(SIC)∗
and d f ,2
can be obtained through algebraic manipulations
by respectively substituting (8) and (10) into (21).
In the presence of interference at GS and AS-2, DMT at low-to-moderate Ω X can be analyzed
from (33), (34) and (35). It reveals the interference scenarios in which the II or SIC detectors
achieves better diversity gain than HD systems.
E. Finite SNR DMT Analysis for HD Systems
Let the finite SNR HD diversity gain at GS be defined as d HD∗
f ,gs . To ensure fair comparison, we
let the variable HD date rate be twice the variable HBD data rate. Let R1HD (Ω X ) = 2r f log2 (1+Ω X )
18
HD
be the variable transmission rate at AS-1 with threshold γth,gs
= [1+Ω X ]2r f −1. Similarly at AS-2,
let the finite SNR HD diversity gain at AS-2 be defined as d HD∗
f ,2 with variable transmission rate
HD (Ω ) = 2r log (1 + Ω ) and threshold γ HD = [1 + Ω ]2r f − 1. The closed-form expressions
Rgs
X
f
X
X
2
th,2
for the finite SNR diversity gains at GS and AS-2 are presented in the following theorem.
Theorem 7: For a variable transmission rate scheme, the finite SNR diversity gain at GS and
AS-2 are given in (36) and (37), respectively.
−Ω X Õ
α m, 1, K X1, 1 g(m, −m − 1, Ω X , 2r f )
d HD∗
=
f ,gs
HD
Pr Ogs m≥0
−Ω X Õ
d HD∗
=
α m, αg,2, K Xgs , 1 g(m, −m − 1, Ω X , 2r f ).
f ,2
HD
Pr O2
m≥0
(36)
(37)
Proof: The expressions in (36) and (37) can be obtained through algebraic manipulations
by respectively substituting (12) and (13) into (21).
In the absence of interference, (36) and (37) can be used to evaluate the DMT at GS and
AS-2, providing a benchmark that can be used in evaluating the performance of the II and SIC
detectors in HBD systems.
F. System Level Finite SNR Diversity Gain and DMT
The system level finite SNR diversity gain and DMT for the multi-user system in Fig. 1 will
be used as a metric to compare HBD and HD systems. For fixed and variable transmission
rate schemes, the system level finite SNR diversity gain and DMT are given in (38) to (41),
respectively.
β
d f ,system
= min
β
d HBD
f ,gs , d f ,2
HD
HD
d HD
f ,system = min d f ,gs, d f ,2
β∗
β∗
,
d
d f ,system = min d HBD∗
f ,gs
f ,2
HD
HD∗ HD∗
d f ,system = min d f ,gs , d f ,2
(38)
(39)
(40)
(41)
where β ∈ {HBD(I I), HBD(SIC)}. Quantifying the finite SNR diversity gain and DMT in (38)
to (41) provides insights into the degree of improvements in outage performance at the system
level, which will be further discussed in Section V.
19
10
Outage Probability at GS (II)
0
HBD(II) - ǫ=0.01, α
g,g
HBD(II) - ǫ=0.01, α
=1
4
HBD(II) - ǫ=0.001, αg,g=1
10
Finite SNR Diversity Gain at GS (II)
4.5
HBD(II) - ǫ=0.01, α
HBD(II) - ǫ=0.01, αg,g=1.5
-2
Ideal HBD - α
=1
=1.5
g,g
=0
HD
3
10 -4
g,g
2.5
df
Outage Probability
Ideal HBD - αg,g=0
HD
Simulation
=1
g,g
HBD(II) - ǫ=0.001, αg,g=1.5
3.5
HBD(II) - ǫ=0.001, αg,g=1.5
g,g
HBD(II) - ǫ=0.001, α
2
10 -6
1.5
1
10 -8
0.5
10 -10
0
0
5
10
15
20
25
30
0
5
10
15
20
25
30
ΩX (dB)
ΩX (dB)
(a) Outage probability comparison at GS
(b) Finite SNR diversity gain comparison at GS
Fig. 2. Outage probability and finite SNR diversity gain at GS (II detector) for phase noise strength γφ2 = −130dBm.
V. N UMERICAL R ESULTS
In this section, numerical results pertaining to the outage probabilities and finite SNR diversity
gains at GS, AS-2 and system level are discussed. Monte Carlo simulations are conducted with
109 samples to verify the accuracy of the outage probability computations. In addition, all Rician
K factors are fixed at 15, i.e., K X1 = KYsi,1 = K Xgs = KY1 = 15, with σg2 = σ22 = −115dBm,
HD = R HBD = 1 for fair comparison between the HBD and HD systems.
Rsum
sum
A. Finite SNR Diversity Gain and Outage Analysis
1) Impact of Residual SI at GS: The HBD outage probability at GS, given in (6), is shown in
Fig. 2a. The ideal HBD and the HD outage probability are also plotted in Fig. 2a as a benchmark
comparison. The ideal HBD outage probability can be obtained from (6) with αg,g = 0 while the
HD outage probability at GS is obtained using (12). From Fig. 2a, (7) is validated since it can
HBD is close to the ideal HBD case, i.e., no interference, at low-to-moderate
be seen that Pr Ogs
HBD is higher as SI channel
average received power (Ω X ) and vice-versa. As expected, Pr Ogs
estimation error (ǫ) is increased. In addition, increasing the strength of the residual SI (αg,g )
degrades the outage performance more than the increase in ǫ since a larger αg,g corresponds to
HBD
a higher average residual SI power, with phase noise (γφ2 ) scaled accordingly. In fact, Pr Ogs
approaches the ideal HBD case when αg,g = 1, ǫ = 0.001 at low Ω X in Fig. 2a. Hence, sufficient
SI mitigation is needed in order for the FD-enabled GS to outperform the HD-enabled GS.
20
Outage Probability at AS-2
10 0
Finite SNR Diversity Gain at AS-2
4.5
HBD(II) - α
4
HBD(II) - α
1,2
1,2
=0.5
=0.1
HBD(SIC) - α
3.5
10 -2
1,2
=5
=10
1,2
=0
HD
3
10
1,2
Ideal HBD - α
2.5
-4
df
Outage Probability
HBD(SIC) - α
2
HBD(II) - α
10
-6
HBD(II) - α
1,2
1,2
=0.5
1.5
=0.1
HBD(SIC) - α
1,2
=5
1
HBD(SIC) - α1,2=10
Ideal HBD - α
10
1,2
=0
0.5
HD
Simulation
-8
0
0
5
10
15
20
25
ΩX (dB)
(a) Outage probability comparison at AS-2
30
0
5
10
15
20
25
30
ΩX (dB)
(b) Finite SNR diversity gain comparison at AS-2
Fig. 3. Outage probability and finite SNR diversity gain at AS-2 (II and SIC detectors) for αg,2 = 1, i.e., link between GS and
AS-2 has same distance as the reference link (d1,g ).
The finite SNR diversity gain at GS, given in (24), is shown in Fig. 2b. The ideal HBD
and HD finite SNR diversity gain at GS are also similarly obtained from (24), with αg,g = 0,
HD
and (29), respectively. From Fig. 2b, it can be seen that d HBD
f ,gs peaks at Ω X = 2dB while d f ,gs
peaks at Ω X = 6dB. Additionally, (25) and (31) are also confirmed in Fig. 2b as Ω X → ∞
and is also corroborated in Fig. 2a, where the slope of the outage probability curves become
constant as Ω X → ∞. In other words, the FD-enabled GS becomes interference-limited at high
Ω X . Interestingly, in the absence of interference at the FD-enabled GS, d HBD
f ,gs → 1 as Ω X → ∞
since only SNR needs to be considered at GS. From the finite SNR diversity gain and outage
analysis at GS, Fig. 2a and Fig. 2b shows that residual SI is the performance limiting factor for
the FD-enabled GS. Therefore, it is important to sufficiently mitigate SI at each of the cascaded
stages in Fig. 1 in order to keep the strength of the residual SI low for effective operation of
the FD-enabled GS.
2) Impact of Interference at AS-2: The HBD outage probabilities at AS-2 for both II and SIC
detectors are shown in Fig. 3a. The II and SIC detector outage probabilities are computed from
(8) and (10), respectively, while the ideal HBD and HD outage probabilities are computed from
(8) with α1,2 = 0, and (13), respectively. From Fig. 3a, the II detector at AS-2 outperforms the
HD mode at low-to-moderate Ω X when inter-AS interference (α1,2 ) is weak and it validates the
observations made when evaluating limα1,2 →L E {Y1l } for L ∈ {0, ∞} in Section III-A2. The trend
in Fig. 3a also suggests that the further reduction in α1,2 will enable the II detector at AS-2 to
attain the ideal HBD outage performance for moderate Ω X , which is expected since α1,2 → 0
21
corresponds to diminishing levels of interference at AS-2.
The SIC detector performs better than the HD mode at the low-to-moderate Ω X when interference is strong, e.g., α1,2 = 10, since stage 1 of the SIC detector is more likely to detect and
subtract x1 [t]. The resultant signal at stage 2 of the SIC detector is thus almost interference-free.
As α1,2 increases, the SIC detector performance approaches that of the ideal HBD case due to the
near perfect cancellation of interference in the first stage. When Ω X > 10dB for α1,2 ∈ {5, 10},
an error floor is present which verifies (28). Similar error floor observations are also made for
the II detector and it indicates that the II and SIC detectors become interference-limited at high
Ω X . From a practical perspective, the trend in Fig. 3a shows that the II detector is well suited
for en route scenarios with less congested flight routes such as those over sparsely populated or
oceanic regions since the II detector experiences weak interference due to path loss as a result of
large inter-aircraft or aircraft to GS distance. On the other hand, the SIC detector is suitable for
use in congested airspace scenarios such as the landing or even continental en route scenarios as
interference from nearby aircrafts can be effectively removed. Although HD-ACS has superior
outage performance compared to the II and SIC detectors at high Ω X , the interference-limited
HBD detectors can meet typical QoS requirements, e.g., frame error rate ≤ 10−3 .
I)
, at AS-2 are shown in Fig. 3b. Both the
and d HBD(SIC)
The finite SNR diversity gains, d HBD(I
f ,2
f ,2
finite SNR HBD and HD diversity gains are obtained from (26) and (30), respectively. Similarly,
the ideal finite SNR HBD diversity gain is obtained from (26) with α1,2 = 0. A trend similar
I)
peaking
and d HBD(SIC)
to what was seen in Fig. 2b can be found in Fig. 3b, with d HBD(I
f ,2
f ,2
HBD(I I)
at Ω X = 2dB, and d HD
f ,2 peaking at Ω X = 6dB. As expected, reducing α1,2 causes d f ,2
to
perform close to the ideal HBD case at low Ω X . Fig. 3b also confirms (28) for both the II and
SIC detectors. It is clear that the SIC detector can attain an outage probability decay rate that is
HBD(SIC)
similar to the ideal HBD case when Ω X ≤ 5dB. Further increasing α1,2 will enable d f ,2
to
be almost identical to the ideal HBD case at Ω X ≤ 5dB since the system becomes noise-limited
rather than interference-limited. The trends in Fig. 3b are also reflected in Fig. 3a since the slope
of the outage probability curves behave as indicated in (28) and (31) as Ω X → ∞.
3) Impact of Interference at System Level: Fig. 4a and Fig. 4b respectively shows the outage
probability and finite SNR diversity gain at the system level. The system level outage probability
is computed from (14) and (15) while system level finite SNR diversity gain is computed from
HBD(I I)
is dominated by the II
(38) and (39). Through numerical analysis, we observed that Pout,system
HBD < Pr O HBD(I I)
detector at AS-2 for α1,2 ∈ {0.1, 0.5} and 0dB ≤ Ω X ≤ 30dB, i.e., Pr Ogs
2
22
10
System Level Outage Probability
0
System Level Finite SNR Diversity Gain
4.5
HBD(II) - ǫ ∈ {0.01,0.001}, α
4
HBD(II) - ǫ ∈ {0.01,0.001}, α
1,2
1,2
=0.5
=0.1
HBD(SIC) - ǫ ∈ {0.01,0.001}, α
10
-2
HBD(SIC) - ǫ ∈ {0.01,0.001}, α
3.5
= 0, α
g,g
=5
=10
=0
HD
3
10
1,2
1,2
-4
2.5
df
Outage Probability
Ideal HBD - α
1,2
2
10 -6
HBD(II) - ǫ ∈ {0.01,0.001}, α1,2=0.5
1.5
HBD(II) - ǫ ∈ {0.01,0.001}, α1,2=0.1
HBD(SIC) - ǫ ∈ {0.01,0.001}, α1,2=5
10
-8
HBD(SIC) - ǫ ∈ {0.01,0.001}, α
1,2
1
=10
Ideal HBD - α1,2=0, αg,g=0
0.5
HD
Simulation
10
-10
0
0
5
10
15
20
25
30
0
5
10
15
20
25
30
Ω (dB)
Ω (dB)
X
X
(a) System level outage probability
(b) System level finite SNR diversity gain
Fig. 4. System level outage probability and finite SNR diversity gain (II and SIC detectors) for αg,2 = 1,αg,g = 1,γφ2 = −130dBm,
ǫ ∈ {0.01, 0.001}.
because inter-AS interference at AS-2 is stronger than the residual SI experienced at GS. Thus,
HBD(I I)
unless interalthough not shown in the figure, increasing αg,g or ǫ does not affect Pout,system
HBD(I I)
HD
AS interference is decreased. It can also be observed from Fig. 4a that Pout,system ≤ Pout,system
HBD(I I)
HD
≤ Pout,system
when Ω X ≤ 4dB, α1,2 = 0.5. When α1,2 = 0.1, Pout,system
for Ω X ≤ 11dB. In fact,
HBD(I I)
Pout,system approaches that of the ideal HBD case when α1,2 is decreased due to the near absence
of inter-AS interference at the II detector and it also explains the trend seen in Fig. 4b where it
I)
approaches that of the ideal HBD case when α1,2 is decreased. In other
can be seen that d HBD(I
f ,system
HBD(I I)
words, the decay of Pout,system approaches that of the ideal HBD case when inter-AS interference
weakens, as reflected in Fig. 4b, for Ω X ≤ 5dB. Therefore, when an II detector is used at AS-2,
I)
HBD(I I)
.
and d HBD(I
the inter-AS interference is the limiting factor for both Pout,system
f ,system
HBD(SIC)
When AS-2 adopts an SIC detector, Pout,system is dominated by GS when Ω X ≤ 4dB and
α1,2 = 5. Similar trends for the SIC detector are also seen in Fig. 4b for Ω X ≤ 5dB. When
HBD(SIC)
is dominated by AS-2 and it can be explained from the perspective of
Ω X > 4dB, Pout,system
the two-stage SIC detector at AS-2. When α1,2 = 5, x1 [t] is five times stronger than the SOI
from GS (xgs [t]). In addition, at stage 1 of the SIC detector, noise power (σ22 ) is stronger than
xgs [t] when Ω X ≤ 4dB. Thus, the SIC detector is more likely to detect and cancel x1 [t] which
HBD > Pr O HBD(SIC) due to residual SI at GS. When Ω > 4dB, σ 2 will be
results in Pr Ogs
X
2
2
weaker than xgs [t] at stage 1 of the SIC detector. Consequently, the SIC detector is less likely to
HBD(SIC)
HBD < Pr O HBD(SIC) . When α
detect and cancel x1 [t], leading to Pr Ogs
1,2 = 10, Pout,system is
2
23
Finite SNR DMT at GS (II)
2
HBD(II) - ǫ=0.01, α
1.8
g,g
HBD(II) - ǫ=0.001, α
HBD(II) - ǫ=0.01, α
1.6
=1
g,g
g,g
=1
=1.5
HBD(II) - ǫ=0.001, αg,g=1.5
Ideal HBD - αg,g=0
1.4
HD
d
*
f
1.2
1
0.8
0.6
0.4
0.2
0
0
0.2
0.4
0.6
r
Fig. 5. Finite SNR DMT at GS (II detector) for
γφ2
0.8
1
f
= −130dBm, Ω X = 10dB.
HBD(SIC)
dominated by GS for Ω X ≤ 10dB due to stronger interference at AS-2, with Pout,system close to
HBD(SIC)
that of the ideal HBD case. Further increasing α1,2 enables Pout,system to reach near-ideal HBD
performance for a wider Ω X range due to the increased likelihood of successfully detecting and
canceling x1 [t], thus explaining the trend in Fig. 4b. Hence, the strength of the interference from
HBD(SIC)
when a SIC detector
and d HBD(SIC)
AS-1 (α1,2) is the main limiting factor for both Pout,system
f ,system
is used at AS-2.
From Fig. 4a and Fig. 4b, the outage and finite SNR diversity gain analysis has highlighted the
feasibility of HBD-ACS over legacy HD-ACS in weak and strong interference scenarios through
the II and SIC detectors, respectively. For instance, weak and strong interference scenarios
could involve en route flights over sparely and densely populated airspace, respectively. From
the aeronautical perspective, the proposed HBD-ACS has better reliability over HD-ACS while
providing more throughput than legacy HD systems.
B. Finite SNR DMT Analysis
1) Impact of Residual SI at GS: Fig. 5 shows the finite SNR diversity gain at GS. The finite
SNR diversity gain (d HBD∗
f ,gs ) is computed from (33). The ideal HBD finite SNR diversity gain
is also computed from (33) with αg,g = 0 while the HD finite SNR diversity gain (d HD∗
f ,gs ) is
computed from (36). From Fig. 5, it is evident that the stronger residual SI due to SI channel
estimation error (ǫ) or phase noise (γφ2 ) reduces d HBD∗
f ,gs . Increasing the strength of the residual
24
Finite SNR DMT at AS-2
2
HBD(II) - α1,2=0.1
HBD(II) - α1,2=0.01
1.8
HBD(SIC) - α1,2=14.3
HBD(SIC) - α1,2=15
HBD(SIC) - α1,2=100
1.6
Ideal HBD - α1,2=0
HD
1.4
d *f
1.2
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
rf
Fig. 6. Finite SNR DMT at AS-2 (II and SIC detector) for αg,2 = 1, Ω X = 10dB.
SI (αg,g ) affects d HBD∗
more than increasing the SI channel estimation error (ǫ) since the effect
f ,gs
of phase noise (γφ2 ) on the residual SI is amplified. From the outage probability perspective,
increasing residual SI results in a slower decay rate of the outage probability, which lowers
d HBD∗
f ,gs . However, it does not imply that outage probability is better when a higher maximum
is attained. Nonetheless, the range of r f for which d HBD∗
≥ d HD∗
value for d HBD∗
f ,gs
f ,gs
f ,gs increases
when the strength of the residual SI (αg,g ) decreases and vice versa. Therefore, FD-enabled
GS can experience improved DMT as residual SI decreases, which is evident in Fig. 5 for
αg,g = 1. Although d HBD∗
is limited by residual SI, the importance of proper SI mitigation is
f ,gs
again emphasized since it is still feasible for GS to be FD-enabled if operating at a higher r f is
the objective of an ACS.
2) Impact of Interference at AS-2: Fig. 6 shows the finite SNR diversity gain at AS-2. The
I)∗
) is computed from (34) while the finite
finite SNR diversity gain for the II detector (d HBD(I
f ,2
) is computed from (35). The ideal HBD
SNR diversity gain for the SIC detector (d HBD(SIC)∗
f ,2
finite SNR diversity gain is computed from (34) with α1,2 = 0 while the HD finite SNR diversity
gain (d HD∗
f ,2 ) is computed from (37). The trends seen in Fig. 6 are similar to what was seen in
, i ∈ {I I, SIC} and d HD∗
[21, Fig. 4], with lower d HBD(i)∗
f ,2 observed as r f → 0. It has been pointed
f ,2
out by Narasimhan [21] and Shin et al. [20] that Rician fading outage probability curves are
influenced by Rician K factors. In particular, increasing the Rician K factor causes the slope of
25
System Level Finite SNR DMT
2
HBD(II) - ǫ ∈ {0.01,0.001}, α1,2=0.1
HBD(II) (α1,2=0.01), SIC (α1,2=100), ǫ=0.01
1.8
HBD(II) (α1,2=0.01), SIC (α1,2=100), ǫ=0.001
HBD(SIC) - ǫ ∈ {0.01,0.001}, α1,2=14.3
HBD(SIC) - ǫ=0.01, α1,2 = 15
1.6
HBD(SIC) - ǫ=0.001, α1,2 = 15
Ideal HBD - α1,2=0, αg,g=0
HD
1.4
d *f
1.2
1
0.8
0.6
0.4
0.2
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
rf
Fig. 7. System level finite SNR DMT (II and SIC detectors) for αg,2 = 1, αg,g = 1, γφ2 = −130dBm, Ω X = 10dB.
outage probability curves to become steeper [20, Fig. 2]. From a finite SNR DMT perspective,
r f → 0 causes K Xgs to have less impact on the outage performance at AS-2.
On the other hand, Fig. 6 also suggests that the II and SIC detectors are able to provide better
reliability at higher multiplexing gains compare to HD systems.
I)∗
≥ d HD∗
At high multiplexing gains, if the inter-AS interference reduces, then d HBD(I
f ,2 . On
f ,2
HBD(I I)∗
the other hand, at low multiplexing gains, d f ,2
< d HD∗
f ,2 even at low inter-AS interference.
I)∗
approaches that of the ideal HBD case as α1,2 → 0 since the signal at the
In fact, d HBD(I
f ,2
II detector is almost interference-free. As a consequence, the resultant outage probability decay
rate becomes similar to that of the ideal HBD case. When a SIC detector is adopted at ASHBD(SIC)∗
at
≥ d HD∗
2, d HBD(SIC)∗
f ,2 as inter-AS interference increases (for example, refer to d f ,2
f ,2
α1,2 = 14.3 in Fig. 6). As α1,2 → ∞, it becomes easier to detect and remove x1 [t] at the twostage SIC detector. When coupled with the lower threshold requirement of the SIC detector, as
compared to HD systems, the SIC detector can potentially achieve superior diversity gains over
HD systems in strong interference scenarios. Moreover, at large values of α1,2 , if the multiplexing
HBD(SIC)∗
gain is high, the achievable d f ,2
matches the ideal HBD case. As shown in Fig. 6,at low
is close to that of the ideal HBD case. Therefore,
multiplexing gain, the achievable d HBD(SIC)∗
f ,2
the II and SIC detectors provides better reliability at higher multiplexing gains compared to HDACS in the presence of weak and strong interference, respectively. However, at low multiplexing
gains, HD-ACS exhibited better reliability than the II and SIC detectors.
26
3) Impact of Interference at System Level: Fig. 7 shows the system level finite SNR diversity
β∗
gain for HBD-ACS (d f ,system ) and HD-ACS (d HD∗
f ,system ) computed from (40) and (41), respectively,
I)∗
> d HD∗
for β ∈ {HBD(I I), HBD(SIC)}. From Fig. 7, it is evident that d HBD(I
f ,system and
f ,system
HBD(SIC)∗
d f ,system
> d HD∗
f ,system as r f increases, and it enables an HBD-ACS to provide better reliability
at higher multiplexing gain than HD-ACS since HBD-ACS requires a lower operating threshold
than existing HD-ACS at both GS and AS-2. However, the degree of improvement that HBDACS has over HD-ACS is constrained by the strength of interference experienced at GS and
AS-2 in the HBD-ACS.
HBD(I I)∗
for
When the II detector is adopted at AS-2 for weak interference scenarios, d HBD∗
f ,gs > d f ,2
HBD(I I)∗
α1,2 = 0.1. Reducing the strength of the inter-AS interference (α1,2 = 0.01) causes d f ,2
>
HBD(I I)∗
d HBD∗
f ,gs , with lower SI channel estimation error (ǫ) corresponding to higher d f ,system . In the
presence of strong interference at AS-2 (α1,2 = 100), adopting the SIC detector at AS-2 results
HBD(SIC)∗
in d f ,2
> d HBD∗
f ,gs . However, when interference from AS-1 is not as strong, e.g., α1,2 ∈
. For α1,2 = 15 and ǫ = 0.01, residual SI has more impact
> d HBD(SIC)∗
{14.3, 15}, then d HBD∗
f ,gs
f ,2
HBD(SIC)∗
HBD(SIC)∗
when r f = 0.33.
as r f increases because d HBD∗
on d HBD∗
f ,gs < d f ,2
f ,gs than K Xgs on d f ,2
From Fig. 7, the reliability of the HBD-ACS depends on the inter-AS interference at AS-2 for
both II and SIC detectors and residual SI at GS. Furthermore, it is possible for the proposed
HBD-ACS to attain finite SNR DMT curves that are identical to the ideal HBD case at sufficiently
low residual SI.
From Fig. 7, the trends show that the proposed HBD-ACS is a viable alternative to legacy
HD-ACS in weak and strong interference scenarios. In particular, the proposed HBD-ACS can
operate at a higher multiplexing gain than legacy HD-ACS, thus offering better throughput and
reliability compared to the latter.
VI. C ONCLUSION
With the impending spectrum crunch faced by the aviation industry being an issue that must
be tackled in the near future, an HBD-ACS consisting of an FD-enabled GS and two HD
ASs simultaneously communicating on the same spectrum is proposed. Outage analysis was
conducted to investigate the impact of interference on the proposed HBD-ACS. A closed-form
outage probability expression for a SIC detector at AS-2 over Rician fading aeronautical channels
was also derived. Closed-form outage expressions for the II detectors at AS-2 and GS and HDequivalent mode of operations are also presented as benchmark comparisons. Finite SNR diversity
27
gain expressions are also derived and presented for both HBD-ACS and HD-ACS, with high
SNR behaviors proven.
Through outage and finite SNR diversity gain analysis, it is established that residual SI is
the main limiting factor at the FD-enabled GS. Therefore, the need for sufficient SI mitigation
through the design of SI mitigation architectures is crucial, which must be properly addressed
in a HBD-ACS. At AS-2, it is observed that interference from AS-1 is the main limiting factor
for both II and SIC detectors. When performance is evaluated at the system level, the proposed
HBD-ACS is found to be very suitable for weak and strong interference scenarios for the II
and SIC detectors, respectively. The proposed HBD-ACS is also able to achieve superior outage
performance and better diversity gains at low-to-moderate SNRs compared to existing HD-ACS
for both weak and strong interference scenarios. Finite SNR DMT analysis has also revealed that
HBD-ACS can achieve DMT curves that are equivalent to interference-free scenarios if residual
SI is sufficiently suppressed, enabling HBD-ACS to be more reliable than HD-ACS at higher
multiplexing gains while operating at low SNR ranges.
Nonetheless, it should be pointed out that both the II and SIC detectors are not suitable
for moderate interference scenarios. As a consequence, advance decoder schemes, e.g., jointdecoders, will be needed in such scenarios. To this end, work is on progress on evaluating the
outage and finite SNR diversity gain for joint-decoders in moderate interference scenarios. In
addition, the impact of Rican K factors on outage behavior in the aeronautical context as well
as extending the outage probability derivations for combinations of Rician and Rayleigh fading
aeronautical channels are also being investigated as part of future works.
ACKNOWLEDGMENT
This research is jointly funded by Airbus Singapore Pte Ltd and the Singapore Economic
Development Board (EDB).
A PPENDIX A
P ROOF
OF
(10)
Let Xgs be the average received power of the SOI with non-centered Chi-squared probabil!
r
K Xgs +1
K Xgs (K Xgs +1)
K Xgs +1
x , where
ity density function (PDF) fXgs (x) = ΩX αg,2 exp −K Xgs − ΩX αg,2 x I0 2
ΩX αg,2
I0 (·) is the modified Bessel function of the first kind with zero order [35]. Similarly, let Y1
be the average received power of the interfering signal with non-centered Chi-squared PDF
28
fY1 (y) =
KY1 +1
ΩX α1,2
exp −KY1 −
KY1 +1
ΩX α1,2 y
I0 2
r
KY1 (KY1 +1)
ΩX α1,2 y
!
. The closed-form SIC outage probabil-
ity at AS-2 is equivalent to computing the sum of the areas of outage regions P1 and P2 ,
n
o
HBD(SIC)
HBD
) = P1 +P2 . Let the outage regions be defined as P1 = Pr Y1 < γth,gs
1 + Xgs
i.e.,Pr(O2
n
o
HBD
HBD
and P2 = Pr Y1 ≥ γth,2 (1 + Xgs ), Xgs < γth,2 . The expression for P1 can be rewritten as [15]
P1 =
∫
0
q+1
ÕÕ
0
=
H BD (1+X )
∞ ∫ γth,gs
gs
fY1 (y) fXgs (x)dydx
HBD
α(q, Ω X α1,2, KY1, γth,gs
)
q≥0 l=0
q+1
l
E {Xgs
},
l
(42)
where E {·} represents the expectation function. The expression for P2 can be expressed as
∫ γ H BD ∫ ∞
th,2
fY1 (y) fXgs (x)dydx
P2 =
H BD (1+X )
γth,gs
gs
0
=
∫
H BD
γth,2
0
Í
©p
Q 1 2KY1,
«
s
HBD (1 + x)
2(KY1 + 1)γth,gs
ª
® · fXgs (x)dx.
Ω X α1,2
¬
(43)
α( j, Ω X αg,2, K Xgs , 1)x j . Thus, (43) can be rewritten as
s
∫ γ H BD Õ
HBD
q
2(K Xgs + 1)γth,2
th,2 ©
©
ª
ª
α( j, Ω X αg,2, K Xgs , 1)x j ®
P2 = 1 − Q 1 2K Xgs ,
®−
Ω X αg,2
0
« j≥0
«
¬
¬
!
n+1
Õ
Õ
n+1 i
HBD
×
(44)
x dx.
α(n, Ω X α1,2, KY1, γth,gs
)
i
n≥0
i=0
Í
HBD ) n+1 n+1 x i and d( j) = α( j, Ω α , K , 1)x j , then the
Let c(n) = α(n, Ω X α1,2, KY1, γth,gs
X g,2
Xgs
i=0
i
From [34], fXgs (x) =
j≥0
integral in (44) can be written as [34], [36]
!
∫ γ H BD Õ
th,2
©Õ
ª
c(n)
d( j)® dx
0
n≥0
« j≥0
¬
∫ γ H BD Õ Õ
n
th,2
=
c(i)d(n − i)dx
0
=
n
ÕÕ
n≥0 i=0
HBD
α(i, Ω X α1,2, KY1, γth,gs
)α(n
− i, Ω X αg,2, K Xgs , 1)
n≥0 i=0
=
n Õ
i+1
ÕÕ
n≥0 i=0 j=0
∫
i+1
Õ
i+1
j=0
HBD
α(i, Ω X α1,2, KY1, γth,gs
)α(n
j
H BD
γth,2
x j+n−i dx
0
HBD j+n−i+1
i + 1 (γth,2 )
.
− i, Ω X αg,2, K Xgs , 1)
j +n−i+1
j
Combining (42) and (45), the expression in (10) can be obtained.
(45)
29
In (43), Q 1
KY1 +1
HBD
ΩX α1,2 (γth,gs )(1
power series if
p
2KY1,
r
H BD (1+x)
2(KY1 +1)γth,gs
ΩX α1,2
!
=1−
Í
HBD
n≥0 α(n, Ω X α1,2, KY1 , γth,gs )(1
+ x)n+1 if
+ x) ≥ 0 [34]. In addition, the PDF fXgs (x) can be expressed as a convergent
K Xgs +1
ΩX αg,2 x
≥ 0 [34]. Assuming the power series in (44) is convergent, the resultant
product of the power series in (45) will also be convergent [36]. Similarly in (42), the power
HBD ≤
series is convergent if γth,gs
(ΩX α1,2 )/(1+KY1 )
2(ΩX αg,2 )/(1+K Xgs )
[37]. Therefore, the closed-form expression in
(10) holds if the power series in (42) and (45) are convergent.
R EFERENCES
[1] ICAO, “Icao long-term traffic forecasts,” 2016.
[2] N. Neji, R. De Lacerda, A. Azoulay, T. Letertre, and O. Outtier, “Survey on the future aeronautical communication system
and its development for continental communications,” IEEE Trans. Veh. Technol., vol. 62, no. 1, pp. 182–191, 2013.
[3] T. Z. H. Ernest, A. K. Krishna, A. Madhukumar, and R. P. Sirigina, “On the efficiency improvements to aeronautical
waveforms and integrated modular avionics systems,” in IEEE/AIAA 35th Digit. Avionics Syst. Conf. (DASC), 2016. IEEE,
2016, pp. 1–8.
[4] S. Li, R. D. Murch, and V. K. Lau, “Linear transceiver design for full-duplex multi-user mimo system,” in Proc. IEEE
Int. Conf. Commun. (ICC), 2014, pp. 4921–4926.
[5] M. Mohammadi, H. A. Suraweera, Y. Cao, I. Krikidis, and C. Tellambura, “Full-duplex radio for uplink/downlink wireless
access with spatially random nodes,” IEEE Trans. Commun, vol. 63, no. 12, pp. 5250–5266, 2015.
[6] Y. Jang, K. Min, S. Park, and S. Choi, “Spatial resource utilization to maximize uplink spectral efficiency in full-duplex
massive mimo,” in Proc. IEEE Int. Conf. Commun. (ICC), 2015, pp. 1583–1588.
[7] A. C. Cirik, S. Biswas, O. Taghizadeh, A. Liu, and T. Ratnarajah, “Robust transceiver design in full-duplex mimo cognitive
radios,” in Proc. IEEE Int. Conf. Commun. (ICC). IEEE, 2016, pp. 1–7.
[8] D. Kim, H. Lee, and D. Hong, “A survey of in-band full-duplex transmission: From the perspective of phy and mac layers,”
IEEE Commun. Surveys Tut., vol. 17, no. 4, pp. 2017–2046, 2015.
[9] A. Sabharwal, P. Schniter, D. Guo, D. W. Bliss, S. Rangarajan, and R. Wichman, “In-band full-duplex wireless: Challenges
and opportunities,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1637–1652, 2014.
[10] V. Tapio, “System scenarios and technical requirements for full-duplex concept,” DUPLO Deliverable D, vol. 1, 2013.
[11] A. Sahai, G. Patel, C. Dick, and A. Sabharwal, “On the impact of phase noise on active cancelation in wireless full-duplex,”
IEEE Trans. Veh. Technol, vol. 62, no. 9, pp. 4494–4510, 2013.
[12] D. Bharadia, E. McMilin, and S. Katti, “Full duplex radios,” ACM SIGCOMM Computer Communication Review, vol. 43,
no. 4, pp. 375–386, 2013.
[13] Y.-D. Yao and A. U. Sheikh, “Investigations into cochannel interference in microcellular mobile radio systems,” IEEE
Trans. Veh. Technol, vol. 41, no. 2, pp. 114–123, 1992.
[14] P. S. Bithas and A. A. Rontogiannis, “Mobile communication systems in the presence of fading/shadowing, noise and
interference,” IEEE Trans. Commun., vol. 63, no. 3, pp. 724–737, 2015.
[15] N. B. Rached, A. Kammoun, M.-S. Alouini, and R. Tempone, “A unified moment-based approach for the evaluation of
the outage probability with noise and interference,” IEEE Trans. Wireless Commun., vol. 16, no. 2, pp. 1012–1023, 2017.
[16] M. O. Hasna, M.-S. Alouini, A. Bastami, and E. S. Ebbini, “Performance analysis of cellular mobile systems with successive
co-channel interference cancellation,” IEEE Trans. Wireless Commun., vol. 2, no. 1, pp. 29–40, 2003.
30
[17] J. M. Romero-Jerez and A. J. Goldsmith, “Receive antenna array strategies in fading and interference: an outage probability
comparison,” IEEE Trans. Wireless Commun., vol. 7, no. 3, pp. 920–932, 2008.
[18] S. P. Weber, J. G. Andrews, X. Yang, and G. De Veciana, “Transmission capacity of wireless ad hoc networks with
successive interference cancellation,” IEEE Trans. Inf. Theory, vol. 53, no. 8, pp. 2799–2814, 2007.
[19] Z. Zhang, Z. Ma, M. Xiao, Z. Ding, and P. Fan, “Full-duplex device-to-device-aided cooperative nonorthogonal multiple
access,” IEEE Trans. Veh. Technol, vol. 66, no. 5, pp. 4467–4471, 2017.
[20] W.-Y. Shin, S.-Y. Chung, and Y. H. Lee, “Diversity–multiplexing tradeoff and outage performance for rician mimo
channels,” IEEE Trans. Inf. Theory, vol. 54, no. 3, pp. 1186–1196, 2008.
[21] R. Narasimhan, “Finite-snr diversity–multiplexing tradeoff for correlated rayleigh and rician mimo channels,” IEEE Trans.
Inf. Theory, vol. 52, no. 9, pp. 3965–3979, 2006.
[22] L. Zheng and D. N. C. Tse, “Diversity and multiplexing: A fundamental tradeoff in multiple-antenna channels,” IEEE
Trans. Inf. Theory, vol. 49, no. 5, pp. 1073–1096, 2003.
[23] R. U. Nabar, H. Bolcskei, and A. J. Paulraj, “Diversity and outage performance in space-time block coded ricean mimo
channels,” IEEE Trans. Wireless Commun., vol. 4, no. 5, pp. 2519–2532, 2005.
[24] L. Wang, Y. Cai, and W. Yang, “On the finite-snr dmt of two-way af relaying with imperfect csi,” IEEE Wireless Commun.
Lett., vol. 1, no. 3, pp. 161–164, 2012.
[25] X. Lin, M. Tao, Y. Xu, and R. Wang, “Outage probability and finite-snr diversity–multiplexing tradeoff for two-way relay
fading channels,” IEEE Trans. Veh. Technol, vol. 62, no. 7, pp. 3123–3136, 2013.
[26] K. Yang, H. Cui, L. Song, and Y. Li, “Efficient full-duplex relaying with joint antenna-relay selection and self-interference
suppression,” IEEE Trans. Wireless Commun., vol. 14, no. 7, pp. 3991–4005, 2015.
[27] A. R. Heidarpour, G. K. Kurt, and M. Uysal, “Finite-snr diversity-multiplexing tradeoff for network coded cooperative
ofdma systems,” IEEE Trans. Wireless Commun., vol. 16, no. 3, pp. 1385–1396, 2017.
[28] E. Haas, “Aeronautical channel modeling,” IEEE Trans. Veh. Technol, vol. 51, no. 2, pp. 254–264, 2002.
[29] N. Zlatanov, E. Sippel, V. Jamali, and R. Schober, “Capacity of the gaussian two-hop full-duplex relay channel with
residual self-interference,” IEEE Trans. Commun., vol. 65, no. 3, pp. 1005–1021, 2017.
[30] R. Narasimhan, “Individual outage rate regions for fading multiple access channels,” in Proc. 2007 ISIT.
IEEE, 2007,
pp. 1571–1575.
[31] T. Kwon, S. Lim, S. Choi, and D. Hong, “Optimal duplex mode for df relay in terms of the outage probability,” IEEE
Trans. Veh. Technol, vol. 59, no. 7, pp. 3628–3634, 2010.
[32] T. K. Baranwal, D. S. Michalopoulos, and R. Schober, “Outage analysis of multihop full duplex relaying,” IEEE Commun.
Lett, vol. 17, no. 1, pp. 63–66, 2013.
[33] P. C. Sofotasios, M. K. Fikadu, S. Muhaidat, Q. Cui, G. K. Karagiannidis, and M. Valkama, “Full-duplex regenerative
relaying and energy-efficiency optimization over generalized asymmetric fading channels,” IEEE Trans. Wireless Commun.,
2017.
[34] S. András, A. Baricz, and Y. Sun, “The generalized marcum q-function: an orthogonal polynomial approach,” Acta
Universitatis Sapientiae Mathematica, vol. 3, no. 1, pp. 60–76, 2011.
[35] I. S. Gradshteyn and I. M. Ryzhik, Table of integrals, series, and products.
Academic press, 2014.
[36] A. Bartoszewicz and S. GłaÌğb, “Algebrability of conditionally convergent series with cauchy product,” Journal of
Mathematical Analysis and Applications, vol. 385, no. 2, pp. 693–697, 2012.
[37] N. B. Rached, A. Kammoun, M.-S. Alouini, and R. Tempone, “An exact power series formula of the outage probability
with noise and interference over generalized fading channels,” in Proc. IEEE PIMRC, 2016, pp. 1–5.
| 7cs.IT
|
Heuristic algorithms for the Maximum Colorful Subtree problem
Kai Dührkop1 , Marie A. Lataretu1 , W. Timothy J. White1,2 , and Sebastian Böcker1,3
arXiv:1801.07456v3 [cs.DS] 13 Feb 2018
1
Chair for Bioinformatics, Friedrich-Schiller-University, Jena, Germany
2
Berlin Institute of Health, Berlin, Germany
3
Contact: [email protected]
Abstract. In metabolomics, small molecules are structurally elucidated using tandem mass spectrometry (MS/MS); this resulted in the computational Maximum Colorful Subtree problem, which is
NP-hard. Unfortunately, data from a single metabolite requires us to solve hundreds or thousands of
instances of this problem; and in a single Liquid Chromatography MS/MS run, hundreds or thousands
of metabolites are measured.
Here, we comprehensively evaluate the performance of several heuristic algorithms for the problem
against an exact algorithm. We put particular emphasis on whether a heuristic is able to rank candidates
such that the correct solution is ranked highly. We propose this “intermediate” evaluation because
evaluating the approximating quality of heuristics is misleading: Even a slightly suboptimal solution
can be structurally very different from the true solution. On the other hand, we cannot structurally
evaluate against the ground truth, as this is unknown. We find that one particular heuristic consistently
ranks the correct solution in a top position, allowing us to speed up computations about 100-fold. We
also find that scores of the best heuristic solutions are very close to the optimal score; in contrast, the
structure of the solutions can deviate significantly from the optimal structures.
1
Introduction
Metabolomics characterizes the collection of all metabolites in a biological cell, tissue, organ or
organism using high-throughput techniques. Liquid Chromatography Mass Spectrometry (LC-MS)
is one of the predominant experimental platforms for this task. Today, a major challenge is to
determine the identities of the thousands of metabolites detected in one LC-MS run. This is also true
for related fields such as natural products research [23], biomarker discovery, environmental science,
or food science. Tandem mass spectrometry (MS/MS) is used to derive information about the
metabolites’ structures. Interpretation of the hundreds to thousands of MS/MS spectra generated
in a single LC-MS run remains a bottleneck in the analytical pipeline [23]. MS/MS data is usually
searched against spectral libraries [19], but only a small number of metabolites (around 2 %) can
be identified in this manner [7]. Recently, computational methods have been developed that do not
search in spectral libraries but rather in molecular structure databases [1, 2, 5, 6, 9, 15, 16, 18, 21, 22].
CSI:FingerID [9] has won several competitions on the identification of small molecules from MS/MS
data4 [17]; the web service for CSI:FingerID currently analyzes more than 2000 queries a day. At
the heart of CSI:FingerID and its variants [5,6] lies the computation of fragmentation trees, as these
can be readily analyzed by kernel-based methods using multiple kernel learning [18].
Fragmentation trees were introduced in 2008 [4] and were initially targeted at the identification of
the molecular formulas of small molecules; later, it was shown that the structure of fragmentation
trees contains valuable information for structural elucidation of the underlying molecule [12, 13].
Computing an optimum fragmentation tree leads to the Maximum Colorful Subtree problem
[4]. Unfortunately, this problem is NP-hard and also hard to approximate [10, 14]. Algorithms exist
to solve the problem either heuristically [14] or exactly [4, 14, 24]. The problem is a variant of the
well-studied Graph Motif problem [8, 11].
4
http://casmi-contest.org/2017/results.shtml
2
Kai Dührkop, Marie A. Lataretu, W. Timothy J. White, and Sebastian Böcker
Approximation algorithms are algorithms for (usually) NP-hard problems with provable
guarantees on the distance of the returned solution to the optimal one. But in bioinformatics
research, the objective function is usually only a “crutch” used to find the optimum structure,
whereas the value of the objective function has little or no meaning. To this end, heuristics in
bioinformatics are designed to find solutions structurally similar to the optimum solution or, even
better, the biological ground truth. This makes it intrinsically difficult to evaluate the performance
of these heuristics, as we have to define a measure on the structural similarity between the heuristic
solution and the biological ground truth; furthermore, the biological ground truth has to be known.
We will use an alternative way to evaluate the performance of a heuristic: For many applications,
one biological instance results in a multitude of computational instances, corresponding to
candidates or hypotheses; the score of the computational problem is used to rank these hypotheses.
Although the ground truth may not be known for the structure of the best solution, we may have
information regarding the correct candidate or hypothesis. To this end, we can evaluate a heuristic
based on its ability to top-rank the correct candidate.
We propose several heuristics for the Maximum Colorful Subtree problem, and evaluate
these heuristics with regards to their ranking quality. We find that one particular heuristic allows
us to quickly confine the set of candidates (molecular formulas of the precursor molecule). This
constitutes a filter, such that optimum solutions have to be sought only for a (preferably small)
subset of candidates. We also evaluate whether the structure of the constructed solutions is similar
to the optimum solution.
2
The Maximum Colorful Subtree problem
Let G = (V, E) be a node-colored, rooted, directed acyclic graph (DAG) with root r ∈ V and edge
weights w : E → R. Let C(G) be the set of colors used in G, and let c(v) ∈ C(G) be the color
assigned to node v ∈ V . We will consider subtrees T = (VT , ET ) of G rooted at r. Let C(T ) be the
set of colors used in T . We say that T is colorful if all of its nodes have different colors.
The Maximum Colorful Subtree problem asks to find a colorful r-rooted subtree T of G
of maximum weight, where G is a DAG with node colors, edge weights and root r ∈ V . We may
assume that G is (weakly) connected and that r is the unique source of G, as we can remove all
nodes from the graph which cannot be reached by a path from the root r, without changing the
optimal solution.
We note that previous work on the problem also makes the assumption of a single source, albeit
usually implicitly [4, 13, 14, 24]. From an algorithmic standpoint, problem variants with or without
a given root are “basically equivalent”: Given an algorithm that does not assume a fixed root r, we
can solve an instance of the problem variant with root r by introducing a superroot r∗ connected
solely to r, sufficiently large edge weight w(r∗ , r) and a new color for r∗ . For the reverse direction,
we solve the problem for every r ∈ V , then choose the best solution.
But there are two peculiarities when computing fragmentation trees that are different from
the general problem, and that we will make use of here: First, any DAG used for fragmentation
tree computation is transitive: That is, uv ∈ E and vw ∈ E implies uw ∈ E. Second, a coloring
c : V → C(G) of DAG G = (V, E) is order-preserving if there is an ordering ‘≺’ on the colors
C(G) such that c(u) ≺ c(v) holds for every edge uv of G [10]. Computing fragmentation trees
naturally results in order-preserving colors, as nodes can be colored by the fragment mass that is
responsible for this node, and edges exist only between nodes from larger to smaller masses. The
Maximum Colorful Arborescence problem [10] asks to find a colorful induced subtree T of G
of maximum weight, where G is a DAG with order-preserving colors and edge weights. See [10] for
Heuristic algorithms for the Maximum Colorful Subtree problem
3
24
-1
7
5
-6
0
5
6
7
2
3
2
0
-1
6
7
-6
3
0
-3
0
2
3
-3
5
-2
-3
0
0
1
4
5
-4
0
0
1
4
0
Fig. 1. Illustration of the Remove Dangling Subtrees (RDS) postprocessing. Left: Input tree, where each node v is
labeled by its score D[v]. Right: Output tree of weight 24.
numerous complexity results. Here, we will stick with the name “Maximum Colorful Subtree
problem”, but nevertheless assume that the coloring is order-preserving, unless indicated otherwise.
3
Heuristics for the Maximum Colorful Subtree problem
The following postprocessing methods can be applied to a tree T = (VT , ET ) after any heuristic:
The Remove Dangling Edges (RDE) postprocessing iteratively removes edges uv from T , where
v is a leaf and w(uv) < 0; this is repeated until no more such edges are found. In contrast, the
Remove Dangling Subtrees (RDS) postprocessing does not consider a single edge at a time, but
rather induced subtrees: Each node u ∈ VT is scored by the maximum weight of any induced subtree
rooted in u. Score D[u] can be computed using dynamic programming:
X
D[u] =
max{0, w(u, v) + D[v]}
uv∈ET
Clearly, D[u] ≥ 0. For each edge uv with w(u, v) + D[v] < 0 we remove uv and the subtree below
it. Both postprocessings can be computed in O(|VT |) time using a tree traversal, as every edge is
considered once and |ET | = |VT | − 1. Figure 1 shows an example of the RDS postprocessing.
We now present heuristics for finding a colorful subtree with root r in a transitive DAG with
order-preserving coloring and unique source r.
– Kruskal-style. This heuristic sorts all edges of the graph by decreasing edge weight, then
iteratively adds edges from the sorted list, ensuring that the growing subgraph is colorful and
that each node has at most one incoming edges. Since r is the unique source of G, and since
G is transitive, this will ultimately result in a colorful subtree of G. This heuristic is similar to
Kruskal’s algorithm for computing an optimum spanning tree; it was called “greedy heuristic”
in [4].
– Prim-style. This heuristic progresses similar to Prim’s algorithm for calculating an optimal
spanning tree: The tree T = (VT , ET ) initially contains only the root r of G. In every step, we
consider all edges uv with u ∈ VT and v ∈
/ VT such that c(v) ∈
/ C(VT ); among these, we choose
the edge with maximum weight and add it to the tree. We repeat until all colors in the graph
are used in the tree. This will usually result in a different tree than the Kruskal-style heuristic,
due to the colorfulness constraint.
– Insertion. This heuristic is a modification of the “insertion heuristic” from [14]. We again start
with a tree T = (VT , ET ) containing only the root r of G. The heuristic greedily attaches nodes
labeled with unused colors. For every node v with c(v) = c0 unused, and every node u already
part of the solution, we calculate how much we gain by attaching v to u. To calculate this gain,
4
Kai Dührkop, Marie A. Lataretu, W. Timothy J. White, and Sebastian Böcker
we take into account the score of the edge uv as well as the possibility of rerouting other outgoing
edges of u through v:
X
I(u, v) := w(uv) +
w(vx) − w(ux)
x∈VT ,w(vx)>w(ux)
where we assume w(uv) = −∞ if uv ∈
/ E. The node with maximum gain is then attached to
the partial solution, and edges are rerouted as required. See [14] for details; different from there,
we do not iterate over colors in some fixed order but instead, consider all unused colors in every
step.
– Top-down. The top-down heuristic [4] is also greedy, but adds paths beginning in the root to
the partial solution. The partial solution initially contains only the root r of G. The heuristic
greedily constructs a path starting at the root which is added to the partial solution; the next
node of the path is chosen so that it maximizes the weight of the added edge, simultaneously
ensuring that the partial solution remains colorful and does not violate the tree property. If no
such edge exists, the algorithm restarts at the root, and searches for another path. It terminates
if no edge at the root can be selected. In the resulting tree, all internal nodes but the root have
exactly one child. This heuristic extends even simpler heuristics that attach all nodes to the
root, which have been in frequent use for molecular formula determination from MS/MS data.
– Critical Path1 . Again, we iteratively build the subtree; initially, the partial solution T contains
only the root r of G. The score S[u] of a node u ∈ V is the maximum weight of a path p from
u to any node v, such that C(p) ∩ C(VT ) ⊆ {c(u)}; that is, the path does not use nodes with
colors already present in the tree, except for the color of the starting node. We can compute
S[u] using the recurrence
S[u] = maxuv∈E,c(v)∈C(V
{0, S[v] + w(uv)}
/
T)
(1)
where we use that the coloring of G is order-preserving, since in that case no two nodes of
the path have the same color. We further assume max ∅ = 0 when computing (1). We iterate
over the ordered colors in reverse order, computing S[u] for all nodes u of the active color. The
critical path p of maximum score can be found by backtracing from the maximum entry S[u]
with u ∈ VT . We add p to T , then iterate, recomputing S for the new set of used colors C(Vt ).
See Figure 2 for an example.
– Critical Path2 . This heuristic also relies on critical paths, but adds, in each iteration, only the first
edge of the critical path to the partial solution. We note that this heuristic does not dominate
the Critical Path heuristic, meaning that in certain cases, the subtree computed by this heuristic
has smaller weight than that computed by the Critical Path heuristic; see Figure 2 (right) for
an example.
– Critical Path3 . This heuristic combines the Insertion heuristic with the Critical Path heuristic:
In each step the heuristic chooses the edge uv with u ∈ VT that maximizes the sum of critical
path score and insertion score S[u] + I(v).
– Maximum. All heuristics compute lower bounds of the maximum score; therefore the maximum
score over all heuristic solutions is also a lower bound.
Time complexity of the heuristics. Let n := |V |, m := |E|, and k := |C(G)|. Clearly, k ≤ n and
in applications, we usually have k n. Furthermore, |VT | ≤ k holds for the returned subtree
T = (VT , ET ).
– It is easy to check that the Kruskal-style heuristic has time complexity O(m log n) for sorting all
edges according to weight. Connectivity testing can then be performed in sub-logarithmic time
Heuristic algorithms for the Maximum Colorful Subtree problem
5
r
2
5
u
1
v
z
2
3
y
x
Fig. 2. Left: Example for the Critical Path Heuristic. Nodes are labeled by score, solid lines show the tree, dashed
lines the rest of the graph. Grayed-out nodes have colors already used in the subtree. Right: An example input graph
for which Critical Path1 produces a better tree than Critical Path2 . Nodes v and z are the same color; all other nodes
have distinct colors. The two solid edges are the suboptimal tree output by Critical Path2 . Critical Path1 initially
chooses the path ruvx for a score of 6, then adds vy for a total of 8. Critical Path2 begins in the same way by choosing
the first edge ru of the heaviest path for a score of 2, but in its second step it chooses the weight-5 edge rz, as the
heaviest path starting with uv has weight 4. No further edges can be added, so the total weight is 7.
per considered edge using a union-find data structure [20]; checking for colorfulness is easily
accommodated by initially placing all nodes of the same color in the same component. The
overall time complexity thus remains O(m log n). Similarly, the Prim-style heuristic requires
O(m log n) time.
– For the Insertion heuristic, computing gain I(u, v) for all u requires O(k 2 ), since u, x ∈ VT .
Hence, attaching one v to the growing tree requires O(k 2 n) time, resulting in O(k 3 n) total
running time.
But there exists a more complicated yet faster implementation for this heuristic: For each v ∈ V ,
we maintain two scores, in(v) and out(v), which correspond to the two terms
P on the RHS of the
definition of I(u, v). Specifically, in(v) = maxu∈VT w(uv), and out(v) = x∈VT max{0, w(vx) −
w(pT (x), x)}, where pT (x) is the parent of x in T for all x ∈ T . To choose the next node to
insert, we look for the node v ∈ V maximizing in(v) + out(v), ignoring nodes of already-used
colors, which takes O(n) time (and could in practice often finish early if we search in decreasing
order of one of these terms, and know an upper bound on the other). We then perform a single
O(k)-time scan to find its optimal parent in the tree, and then perform two further updates:
First, for all u ∈ V , set in(u) ← max{in(u), w(vu)} and out(u) ← out(u) + w(uv) − w(pT (v), v).
Second, for all x ∈ VT , check whether the incoming edge yx (i.e., y = pT (x)) can be improved
by rerouting via v; if so, delete yx, insert vx and for all u ∈ V such that w(yx) ≤ w(ux), set
out(u) ← max{0, out(u) − (w(vx) − w(yx))}. The second update needs O(kn) time per inserted
node, for O(k 2 n) overall.
– The Top-down heuristic searches at most k times for the maximum weight edge leaving a node;
since there are O(n) such edges, the running time is O(kn).
– For the Critical Path1 heuristic, we need O(m) time to compute values S[u] and to identify the
path of maximum weight. This is repeated at most k − 1 times, resulting in a total running time
of O(km). The same holds true for the Critical Path2 . For Critical Path3 we can again maintain
an out(v) table that contains the score bonus we get for attaching a node in the intermediate
tree as child of v and deleting its previous incoming edge. After each insertion of an edge into
the intermediate tree we have to perform the two update operations on out which takes O(kn)
per insertion. In total we need O(k 2 n + km) time to compute Critical Path3 . In applications, k
6
Kai Dührkop, Marie A. Lataretu, W. Timothy J. White, and Sebastian Böcker
is very small and n m, so the O(km) part for calculating the critical paths requires most of
the computation time.
Computing the k-best fragmentation trees exactly. Even if we do not trust the structural quality of
the heuristic solution, the above heuristics allow us to speed up fragmentation tree computation:
We first select a single candidate (molecular formula of the precursor) using one of the heuristics,
then compute the optimal solution for this instance using an exact method [4, 14, 24]. In practice,
this approach has two shortcomings: Even though certain heuristics show a very good performance
in selecting the correct molecular formula (see below), this correct answer is not known to us in
application; but we will observe that the computed fragmentation tree will often not be the optimum
fragmentation tree, if we also consider other molecular formula candidates and corresponding
instances.
Even worse, it is usually not sufficient in application to select a single best candidate using the
heuristic, then re-compute the fragmentation tree for the corresponding instance. Instead, we usually
want to know optimal fragmentation trees for the k best-scoring candidate. This is independent of
whether results are reported to the user, who wants to use fragmentation tree structure to survey
if computations and, hence, the assigned precursor molecular formula are trustworthy; or, if we
perform some downstream computational analysis based on fragmentation tree structure, such as
CSI:FingerID [9]. In particular for “large” metabolites with mass beyond 600 Dalton, this is necessary
because neither the heuristics nor the exact method will always allow us to find the correct candidate;
only by considering several candidates, we can be sufficiently sure that the correct answer is present.
We propose the following heuristic to compute optimum fragmentation trees for the kbest molecular formula candidates: First, we compute heuristic solutions for all candidates, and
order molecular formula candidates according to the heuristic score. Next, we compute optimum
fragmentation trees for the k best candidates; for small k, we can instead choose some fixed
parameter, such as 10 candidates. We estimate the maximum ∆ ≥ 0 of differences between the
score of the optimum solution and the corresponding heuristic solution, using those candidates
where we know the exact solution. We now assume that the score difference is upper-bounded
by ∆ for all candidates. We continue to process candidates and compute optimum fragmentation
trees from the sorted list, updating the k-best candidates and the corresponding score threshold;
we stop computations when the heuristic score of a candidate plus ∆ is smaller than the current
score threshold. Clearly, our assumption made above may be violated for certain input, making this
method a heuristic.
4
Data and Instances
Details of how to transform the MS/MS spectrum of an (unknown) compound into one or more
instances of the Maximum Colorful Subtree problem have been published elsewhere [3, 4]; we
shortly recapitulate the process. We consider all molecular formulas from some ground set, such as,
all molecular formulas from elements CHNOPS. We decompose the precursor mass into all possible
candidate molecular formulas from this ground set; each candidate molecular formula corresponds
to one instance. For each instance, we decompose the fragment peaks in the MS/MS spectrum,
ensuring that each fragment molecular formula is a subformula of the candidate molecular formula
for the precursor mass. These molecular formulas constitute the nodes of a graph; each node is
colored by the peak it stems from. An edge is present between molecular formulas u, v if and only
if v is a proper subformula of u. Now, both nodes and edges receive a certain weight [3], based both
on prior knowledge (e.g., distribution of loss masses) and the data (e.g., mass difference between
a peak and its hypothetical molecular formula); but as pointed in [4], we may assume that only
Heuristic algorithms for the Maximum Colorful Subtree problem
7
80
difference in identification rate (percentage points)
identified molecular formulas (%)
1
75
70
65
60
0
1
2
3
5
Critical Path 3 Critical Path 2 Critical Path 1 Kruskalstyle Primstyle
Insertion
Topdown
exact
Maximum
Critical Path 3
Critical Path 2
Critical Path 1
Kruskalstyle
Primstyle
4
1
2
3
4
5
ranking
6
7
Insertion
Topdown
exact
Maximum
8
9
10
Fig. 3. Performance evaluation, finding the correct molecular formula. Left: Percentage of compounds where the
correct molecular formula received the highest score. Note the zoom of the y axis. Right: Percentage point difference
against exact computations; how often is the correct answer part of the top k output of each method?
edges are weighted. Candidate molecular formulas of the precursor peak are ranked according to the
weight of the maximum colorful subtree in this graph. SIRIUS 3.6 default weights are used, see [3].
To evaluate whether a heuristic is capable of ranking the correct molecular formula on the
top position, we have to use reference data where the true compound structure is known for each
MS/MS measurement. We use reference compounds from GNPS [23]; each reference compound
is one instance, corresponding to several graphs (for the different molecular formula candidates
of the precursor mass) we have to search in. We then filter instances: For example, we assume a
mass accuracy of 10 ppm (parts per million), and discard compounds where the precursor mass is
missing or outside outside this mass window. All details can be found in [3]. This leaves us with 4 050
compounds, each of which is then transferred to usually many instances of the Maximum Colorful
Subtree problem. One reference compound resulted in between 1 and 21 748 candidate molecular
formulas, with median 53 and average 263.8 candidates. To avoid proliferating running times, we
consider only the 60 most intense peaks in a MS/MS spectrum that can be decomposed, which is
again SIRIUS 3.6 default behavior. We fix the SIRIUS tree size parameter, which is usually adapted
at runtime, at −0.5. In addition, we switch off SIRIUS’ spectral recalibration.
5
Results
We applied all but the Critical Path heuristics using the RDS postprocessing. We do not evaluate
the RDE postprocessing, as it is dominated by RDS (that is, the score is at least as good, in all cases)
which, in turn, dominates solutions without postprocessing. Furthermore, both postprocessings
are very fast in practice. For the Critical Path heuristics, RDS cannot improve a solution for
variants 1 and 2; for variant 3 this is possible in principle, but very unlikely. To keep results of
the Critical Path heuristics consistent, we disabled the RDS postprocessing for variant 3, too.
All heuristics were implemented in Java 8. For the exact method, we use the Integer Linear
Program (ILP) from [14] with the CPLEX solver 12.7.1 (IBM, https://www.ibm.com/products/
ilog-cplex-optimization-studio).
First, we evaluated the power of the different heuristics to rank the correct answer (molecular
formula) at the top position; see Fig. 3 (left). We also compared against the exact solution. We
observe similar identification rates for the critical path heuristics, the maximum heuristic and the
exact method. To test whether this trend is true not only for the top rank, but also for the top k
8
Kai Dührkop, Marie A. Lataretu, W. Timothy J. White, and Sebastian Böcker
107
1 month
106
105
1 day
104
1 hour
running time (s)
103
102
1 minute
101
100
101
102
103
Critical Path 3
Critical Path 2
Critical Path 1
Kruskalstyle
Primstyle
104
105
106
0
20
40
60
computed instances (%)
Insertion
Topdown
exact
graph building
Maximum
80
100
Fig. 4. Running times of the different methods. One instances consists of all graphs generated for one compound in
the dataset, considering all decompositions of the precursor mass. For each method, instances are sorted with respect
to running time, and we report amortized running times. We also report running times for constructing the instance
DAGs and for the exact method.
ranks, we also evaluated how often any method is capable to rank the correct answer in its top k,
for varying k; see Fig. 3 (right). Identification rates differ much stronger when varying k for one
method than for different methods and one k; to this end, we normalize identification rates by
subtracting the identification rate of the exact method. We see that all heuristics but critical path
result in inferior rankings, loosing one or more percentage points for most ranks. In contrast, the
critical path heuristics rank solutions with comparable power as the exact method, and the later
two variants often outperform the exact method. Somewhat surprisingly, the maximum over all
heuristics performs even better than the best heuristic.
Second, we compared running times of the different methods. Running times were measured
using a single thread on an Intel E5-2630v3 at 2.40 GHz with 64 GB RAM. The total running time
for the exact methods over all instances is almost one month, underlining the importance of speeding
up computations. But also note that solving all instances exactly requires only about 100-fold the
time required for constructing the instance graphs. For each method, we sorted all instances by
running time; we then reported how much time is required to solve, say, the 90 % “easiest” instances
for that method. Generally, this ordering is different for each method. For all methods, we observe
that the “hardest” 5 % of the instances are responsible for most of the total running time; this has
been observed before [3, 14]. In comparison to the exact method, all heuristics are very fast, and
at least two orders of magnitude faster. In particular, each heuristic is faster than the method for
constructing the instance graphs; running all heuristics, as required for the maximum heuristic,
requires about the same time as the graph construction. Comparing heuristics’ running times, we
see that the Kruskal-style heuristic is slowest; and that the first variant of the Critical Path heuristic
is faster in practice than variants 2 and 3, but not significantly.
Heuristic algorithms for the Maximum Colorful Subtree problem
9
1.0
0.8
200
150
0.6
0.4
Critical Path 3
Critical Path 2
Critical Path 1
Kruskalstyle
Primstyle
Insertion
Topdown
Maximum
0.2
0.0
Critical Path 3
heuristic score relative to exact score
250
0
20
40
60
instances (%) sorted by score
80
100
50
0
50
100
50
0
50
100
exact method
150
200
250
Fig. 5. Left: Relative scores of the heuristics. For each Maximum Colorful Subtree instance, we consider the
relative score in comparison the exact method at 100 %. For each method, instances are sorted with respect to relative
score. Right: Comparison of the score of the Critical Path3 heuristic in comparison to the optimal score of the instance.
For each compound, we consider only the true molecular formula candidate.
Third, we compared the scores reached by the different heuristics to the scores of the exact
solutions, see Fig. 5 (left). For each compound, we only considered the instance of the Maximum
Colorful Subtree that corresponds to the true candidate molecular formula. We report scores
relative to the exact solution (at 100 %), and sorted instances with respect to this relative score. In
the resulting plot, it is not obvious which of the heuristics “Insertion”, “Kruskal-style” and “Critical
Path1 ” should be preferred. We see that Critical Path3 heuristic and, hence, the maximum of all
methods are able to compute (almost) optimal solutions for about 80 % of the instances. In turn,
this means that even for these methods which perform excellent in ranking the correct answer, we
miss the optimal solution in about 20 % of the instances. In addition, we compared scores of the
Critical Path3 heuristic against the exact method in more detail, see Fig. 5 (right): We see that
for instances where the heuristic does not find the optimal solution, the computed solution is only
“slightly suboptimal” with respect to its score. In fact, Pearson correlation between the two measures
is +1.00.
Fourth, we evaluate the solution structure quality of the Critical Path3 heuristic. Unfortunately,
the “true fragmentation tree” cannot be determined experimentally [13]. To this end, we compare
heuristic tree structure vs. tree structures computed using the exact method. For each compound,
we restrict the comparison to the true candidate molecular formula; for other candidate molecular
formulas, the optimal tree cannot possibly be the “ ‘true fragmentation tree”. See Fig. 6. For tree
sizes, we observe rather large deviations between heuristic and optimal trees; in contrast, the overall
distribution of tree sizes is highly similar. But if we compare tree structures, we observe much
larger differences between the Critical Path3 heuristic and the exact method: We measure structural
similarity comparing either the set of node labels (fragments) or the set of edge labels (losses) of the
two trees. We estimate the similarity of two (finite) sets A, B using the Jaccard similarity coefficient
J(A, B) = |A ∩ B| / |A ∪ B| ∈ [0, 1]. We observe that more than 20 % of the heuristic trees differ
from the corresponding optimal tree; for at least 10 %, this difference is significant.
10
Kai Dührkop, Marie A. Lataretu, W. Timothy J. White, and Sebastian Böcker
Number of vertices for tree size parameter = 2.5
600
60
500
1.0
400
0.8
Jaccard similarity coefficient
40
count
number of vertices
50
30
300
200
20
0.6
0.4
10
100
0.2
0
0
0.0
0
20
40
60
instances (%)
80
100
0
10
20
30
40
number of vertices
50
60
fragments
losses
50
60
70
instances (%)
80
90
100
Fig. 6. Left: Size of the fragmentation tree. Instances are sorted with respect to the size of the optimal fragmentation
tree (black); green bars indicate the corresponding tree sizes for the Critical Path3 heuristic. Middle: Distribution
of tree sizes for the exact method (black) and the Critical Path3 heuristic (green). Right: Comparison of the
fragmentation tree structure, optimal tree vs. the tree computed by the Critical Path3 heuristic. Note the zoom
of the x axis. In all three cases, we consider only the true molecular formula candidate for each compound.
6
Conclusion
We have presented heuristics for the Maximum Colorful Subtree problem. Our evaluation
shows that the Critical Path3 heuristic is well-suited for choosing the correct candidate molecular
formula, when applied to tandem mass spectrometry data of small molecules. Our evaluation
sidesteps the catch-22 that we want to evaluate solutions based on structure and not score when, at
the same time, the correct solution structure is not known. We have shown that the tree computed
by the Critical Path3 heuristic is often identical to the optimal tree. Even when the heuristic returns
a suboptimal solution, the score is usually very close to the optimal score. In contrast, the structure
of the heuristic tree deviates significantly from the optimal tree for more than 20 % of the instances.
To this end, we argue not to use this tree for downstream analysis, such as estimating chemical
similarity based on fragmentation tree similarity [12] or machine learning [5, 6, 9, 18]: Preliminary
evaluations clearly indicate that using trees computed by any heuristic, leads to significantly worse
results for the downstream analysis. A back-of-the-envelope calculation indicates the problem: If we
assume that 20 % of the heuristic trees are “structurally faulty”, then a pairwise comparison of trees
will result in 36 % tree pairs where at least one of the trees is “structurally faulty”.
Building an instance DAG requires more time than running any of the presented heuristics. We
conjecture that there is only limited potential for speeding up the graph building phase. To this
end, whereas searching for better (and not significantly slower) heuristics is still a valid undertaking,
faster heuristics are of little practical use. It is worth mentioning that computing exact solutions for
the NP-hard Maximum Colorful Subtree problem takes only about 100-fold the time needed for
constructing the graph instances; further speed-up is possible using data reductions and a stronger
ILP formulation of the problem from [24].
Even elaborate heuristics for a bioinformatics problem, which are capable of finding solutions
with objective function value very close to the optimum, can result in solutions which are structurally
very dissimilar from the optimum structure. We showed that this is not only a theoretical possibility,
but happens regularly for real-world instances. This underlines the importance of finding exact
solutions for bioinformatics problems; the structure of solutions found by heuristic, including local
search heuristics such as Markov chain Monte Carlo, may deviate significantly from the optimal
solution.
Heuristic algorithms for the Maximum Colorful Subtree problem
11
Acknowledgments. WTJW funded by Deutsche Forschungsgemeinschaft (grant BO 1910/9).
Availability. The Critical Path3 heuristic and the exact method are available trough the SIRIUS
software (https://bio.informatik.uni-jena.de/software/sirius/) and also from GitHub
(https://github.com/boecker-lab/sirius). Source code for all other heuristics will be made
available upon request. Instances will be made available from our website.
References
1. Allen, F., Greiner, R., Wishart, D.: Competitive fragmentation modeling of ESI-MS/MS spectra for putative
metabolite identification. Metabolomics 11(1), 98–110 (2015)
2. Allen, F., Pon, A., Greiner, R., Wishart, D.: Computational prediction of electron ionization mass spectra to
assist in GC/MS compound identification. Anal Chem 88(15), 7689–7697 (2016)
3. Böcker, S., Dührkop, K.: Fragmentation trees reloaded. J Cheminform 8, 5 (2016)
4. Böcker, S., Rasche, F.: Towards de novo identification of metabolites by analyzing tandem mass spectra.
Bioinformatics 24, I49–I55 (2008), proc. of European Conference on Computational Biology (ECCB 2008)
5. Brouard, C., Bach, E., Böcker, S., Rousu, J.: Magnitude-preserving ranking for structured outputs. In: Zhang,
M.L., Noh, Y.K. (eds.) Proc. of Asian Conference on Machine Learning. Proceedings of Machine Learning
Research, vol. 77, pp. 407–422. PMLR (2017)
6. Brouard, C., Shen, H., Dührkop, K., d’Alché-Buc, F., Böcker, S., Rousu, J.: Fast metabolite identification with
input output kernel regression. Bioinformatics 32(12), i28–i36 (2016), proc. of Intelligent Systems for Molecular
Biology (ISMB 2016)
7. da Silva, R.R., Dorrestein, P.C., Quinn, R.A.: Illuminating the dark matter in metabolomics. Proc Natl Acad Sci
U S A 112(41), 12549–12550 (2015)
8. Dondi, R., Fertin, G., Vialette, S.: Complexity issues in vertex-colored graph pattern matching. J Discrete
Algorithms 9(1), 82–99 (2011)
9. Dührkop, K., Shen, H., Meusel, M., Rousu, J., Böcker, S.: Searching molecular structure databases with tandem
mass spectra using CSI:FingerID. Proc Natl Acad Sci U S A 112(41), 12580–12585 (2015)
10. Fertin, G., Fradin, J., Jean, G.: Algorithmic aspects of the maximum colorful arborescence problem. In: Proc.
of Theory and Applications of Models of Computation (TAMC 2017). Lect Notes Comput Sci, vol. 10185, pp.
216–230 (2017)
11. Lacroix, V., Fernandes, C.G., Sagot, M.F.: Motif search in graphs: Application to metabolic networks. IEEE/ACM
Trans Comput Biology Bioinform 3(4), 360–368 (2006)
12. Rasche, F., Scheubert, K., Hufsky, F., Zichner, T., Kai, M., Svatoš, A., Böcker, S.: Identifying the unknowns by
aligning fragmentation trees. Anal Chem 84(7), 3417–3426 (2012)
13. Rasche, F., Svatoš, A., Maddula, R.K., Böttcher, C., Böcker, S.: Computing fragmentation trees from tandem
mass spectrometry data. Anal Chem 83(4), 1243–1251 (2011)
14. Rauf, I., Rasche, F., Nicolas, F., Böcker, S.: Finding maximum colorful subtrees in practice. J Comput Biol 20(4),
1–11 (2013)
15. Ridder, L., van der Hooft, J.J.J., Verhoeven, S., de Vos, R.C.H., Bino, R.J., Vervoort, J.: Automatic chemical
structure annotation of an LC-MS(n) based metabolic profile from green tea. Anal Chem 85(12), 6033–6040
(2013)
16. Ruttkies, C., Schymanski, E.L., Wolf, S., Hollender, J., Neumann, S.: MetFrag relaunched: incorporating strategies
beyond in silico fragmentation. J Cheminform 8, 3 (2016)
17. Schymanski, E.L., Ruttkies, C., Krauss, M., Brouard, C., Kind, T., Dührkop, K., Allen, F.R., Vaniya, A.,
Verdegem, D., Böcker, S., Rousu, J., Shen, H., Tsugawa, H., Sajed, T., Fiehn, O., Ghesquière, B., Neumann, S.:
Critical Assessment of Small Molecule Identification 2016: Automated methods. J Cheminf 9, 22 (2017)
18. Shen, H., Dührkop, K., Böcker, S., Rousu, J.: Metabolite identification through multiple kernel learning on
fragmentation trees. Bioinformatics 30(12), i157–i164 (2014), proc. of Intelligent Systems for Molecular Biology
(ISMB 2014)
19. Stein, S.E.: Mass spectral reference libraries: An ever-expanding resource for chemical identification. Anal Chem
84(17), 7274–7282 (2012)
20. Tarjan, R.E.: A class of algorithms which require nonlinear time to maintain disjoint sets. J Comput System Sci
18(2), 110–127 (1979)
21. Tsugawa, H., Kind, T., Nakabayashi, R., Yukihira, D., Tanaka, W., Cajka, T., Saito, K., Fiehn, O., Arita,
M.: Hydrogen rearrangement rules: Computational ms/ms fragmentation and structure elucidation using MSFINDER software. Analytical chemistry 88, 7946–7958 (2016)
12
Kai Dührkop, Marie A. Lataretu, W. Timothy J. White, and Sebastian Böcker
22. Verdegem, D., Lambrechts, D., Carmeliet, P., Ghesquière, B.: Improved metabolite identification with MIDAS
and MAGMa through MS/MS spectral dataset-driven parameter optimization. Metabolomics 12(6), 1–16 (2016)
23. Wang, M., Carver, J.J., Phelan, V.V., Sanchez, L.M., Garg, N., Peng, Y., Nguyen, D.D., Watrous, J., Kapono,
C.A., Luzzatto-Knaan, T., Porto, C., Bouslimani, A., Melnik, A.V., Meehan, M.J., Liu, W.T., Crüsemann, M.,
Boudreau, P.D., Esquenazi, E., Sandoval-Calderón, M., Kersten, R.D., Pace, L.A., Quinn, R.A., Duncan, K.R.,
Hsu, C.C., Floros, D.J., Gavilan, R.G., Kleigrewe, K., Northen, T., Dutton, R.J., Parrot, D., Carlson, E.E., Aigle,
B., Michelsen, C.F., Jelsbak, L., Sohlenkamp, C., Pevzner, P., Edlund, A., McLean, J., Piel, J., Murphy, B.T.,
Gerwick, L., Liaw, C.C., Yang, Y.L., Humpf, H.U., Maansson, M., Keyzers, R.A., Sims, A.C., Johnson, A.R.,
Sidebottom, A.M., Sedio, B.E., Klitgaard, A., Larson, C.B., Boya P, C.A., Torres-Mendoza, D., Gonzalez, D.J.,
Silva, D.B., Marques, L.M., Demarque, D.P., Pociute, E., O’Neill, E.C., Briand, E., Helfrich, E.J.N., Granatosky,
E.A., Glukhov, E., Ryffel, F., Houson, H., Mohimani, H., Kharbush, J.J., Zeng, Y., Vorholt, J.A., Kurita, K.L.,
Charusanti, P., McPhail, K.L., Nielsen, K.F., Vuong, L., Elfeki, M., Traxler, M.F., Engene, N., Koyama, N.,
Vining, O.B., Baric, R., Silva, R.R., Mascuch, S.J., Tomasi, S., Jenkins, S., Macherla, V., Hoffman, T., Agarwal,
V., Williams, P.G., Dai, J., Neupane, R., Gurr, J., Rodríguez, A.M.C., Lamsa, A., Zhang, C., Dorrestein, K.,
Duggan, B.M., Almaliti, J., Allard, P.M., Phapale, P., Nothias, L.F., Alexandrov, T., Litaudon, M., Wolfender,
J.L., Kyle, J.E., Metz, T.O., Peryea, T., Nguyen, D.T., VanLeer, D., Shinn, P., Jadhav, A., Müller, R., Waters,
K.M., Shi, W., Liu, X., Zhang, L., Knight, R., Jensen, P.R., Palsson, B.Ø., Pogliano, K., Linington, R.G.,
Gutiérrez, M., Lopes, N.P., Gerwick, W.H., Moore, B.S., Dorrestein, P.C., Bandeira, N.: Sharing and community
curation of mass spectrometry data with Global Natural Products Social molecular networking. Nat Biotechnol
34(8), 828–837 (2016)
24. White, W.T.J., Beyer, S., Dührkop, K., Chimani, M., Böcker, S.: Speedy colorful subtrees. In: Proc. of Computing
and Combinatorics Conference (COCOON 2015). Lect Notes Comput Sci, vol. 9198, pp. 310–322. Springer, Berlin
(2015)
| 8cs.DS
|
Formally Specifying and Proving Operational Aspects
of Forensic Lucid in Isabelle
Serguei A. Mokhov and Joey Paquet
arXiv:0904.3789v1 [cs.LO] 24 Apr 2009
Department of Computer Science and Software Engineering
Faculty of Engineering and Computer Science
Concordia University, Montréal, Québec, Canada,
{mokhov,paquet}@cse.concordia.ca
Abstract. A Forensic Lucid intensional programming language has been proposed for intensional cyberforensic
analysis. In large part, the language is based on various predecessor and codecessor Lucid dialects bound by the
higher-order intensional logic (HOIL) that is behind them. This work formally specifies the operational aspects of
the Forensic Lucid language and compiles a theory of its constructs using Isabelle, a proof assistant system.
1 Introduction
As a part of the Intensional Cyberforensics project, we define a functional-intensional programming/specification
language, called Forensic Lucid. The language is under active design and development including its syntax, semantics,
the corresponding compiler, run-time, and interactive “development” environments [1,2] that we refer to as General
Intensional Programming System (GIPSY) [3]. We approach the problem using Isabelle [4] as a proof assistant.
Problem Statement. A lot of intensional dialects have been spawned from the functional intensional programming language called Lucid [5,6,7,8,9,10,11,12]. Lucid (see Section 1.2) itself was invented with a goal for program correctness
verification [7,8]. While there were a number of operational semantics rules for compiler and run-time environments
developed for all those dialects throughout the years, there was no a complete formal proof set of the rules of the
languages. Yet another dialect of Lucid has been created to foster the research on intensional cyberforensics (see Section 1.3), called Forensic Lucid, which, in a large part is a union of the syntax and operational semantics rules from
the comprising languages with the forensic extensions. In order to be a credible tool to use, for example, in court, to
implement relevant tools for the argumentation, the language ought to have a solid scientific base, a part of which is
formalizing the semantics the language and proving correctness of the programs written in it.
Proposed Solution. In this work, we propose to begin validation of the Forensic Lucid constructs with the Isabelle
prover assistant [4] and extend it to the comprising Lucid dialects as a whole. We proceed bottom-up from “core”
Lucid dialects such as GIPL, Lucx, and Indexical Lucid and even their smaller decompositions as well as top-down
from Forensic Lucid to arrive to a comprehensive set of proofs covering the dialects.
1.1 Intensional Logics and Programming
Definitions. Intensional programming (IP) is based on intensional (or multidimensional) logics, which, in turn, are
based on natural language understanding aspects (such as time, belief, situation, and direction). IP brings in dimensions and context to programs (e.g. space and time in physics or chemistry). Intensional logic adds dimensions to
logical expressions; thus, a non-intensional logic can be seen as a constant or a snapshot in all possible dimensions.
Intensions are dimensions at which a certain statement is true or false (or has some other than a Boolean value). Intensional operators are operators that allow us to navigate within these dimensions. Higher-order intensional logic
(HOIL) is the one that couples functional programming as that of Lucid with multidimensional dataflows that the
intensional programs can query an alter through an explicitly notion of contexts as first-class values [13,14].
An Example of Using Temporal Intensional Logic. Temporal intensional logic is an extension of temporal logic
that allows to specify the time in the future or in the past.
(1)
E1 := it is raining here today
Context: {place:here, time:today}
(2)
E2 := it was raining here before(today) = yesterday
(3)
E3 := it is going to rain at (altitude here + 500 m) after(today) = tomorrow
Let’s take E1 from (1) above. Then let us fix here to Montreal and assume it is a constant. In the month of
February, 2008, with granularity of day, for every day, we can evaluate E1 to either true or false:
Tags:
1 2 3 4 5 6 7 8 9 ...
Values: F F T T T F F F T ...
If one starts varying the here dimension (which could even be broken down to X, Y , Z), one gets a two-dimensional
evaluation of E1 :
City: /
1 2 3 4 5 6 7 8 9 ...
Montreal F F T T T F F F T ...
Quebec
F F F F T T T F F ...
Ottawa
F T T T T T F F F ...
1.2 Lucid
Lucid [5,6,9,7,8] is a dataflow intensional and functional programming language. In fact, it is a family of languages
that are built upon intensional logic (which in turn can be understood as a multidimensional generalization of temporal
logic) involving context and demand-driven parallel computation model. A program written in some Lucid dialect is
an expression that may have subexpressions that need to be evaluated at certain context. Given the set of dimension
D = {dimi } in which an expression varies, and a corresponding set of indexes or tags defined as placeholders over
each dimension, the context is represented as a set of <dimi : tagi > mappings and each variable in Lucid, called often
a stream, is evaluated in that defined context that may also evolve using context operators [14,15,16,13]. The generic
version of Lucid, GIPL [11], defines two basic operators @ and # to navigate in the contexts (switch and query). The
GIPL was the first generic programming language of all intensional languages, defined by the means of only two
intensional operators @ and #. It has been proven that other intensional programming languages of the Lucid family
can be translated into the GIPL [11]. Please refer to Appendix A for the greater details about Lucid origins, variables
as streams, random access to streams, and the basic operators. Since the Lucid family of language thrived around
intensional logic that makes the notion of context explicit and central, and recently, a first class value [16,13,14,15]
that can be passed around as function parameters or as return values and have a set of operators defined upon. We
greatly draw on this notion by formalizing our evidence and the stories as a contextual specification of the incident
to be tested for consistency against the incident model specification. In our specification model we require more than
just atomic context values – we need a higher-order context hierarchy to specify different level of detail of the incident
and being able to navigate into the “depth” of such a context. A similar provision by has already been made by the
author [17] and earlier works of Swoboda et al. in [18,19,20,21] that needs some modifications to the expressions of
the cyberforensic context.
Some other languages can be referred to as intensional even though they may not refer to themselves as such,
and were born after Lucid (Lucid began in 1974). Examples include hardware-description languages (HDLs, appeared
in 1977) where the notion of time (often the only “dimension”, and usually progresses only forward), e.g. Verilog
and VHDL. Another branch of newer languages for the becoming popular is aspect-oriented programming (AOP)
languages, that can have a notion of context explicitly, but primarily focused on software engineering aspect of software
evolution and maintainability.
1.3 Cyberforensic Analysis
Cyberforensic analysis has to do with automated or semi-automated processing of and reasoning about electronic evidence, witnesses, and other details from cybercrime incidents (involving computers, but not limited to them). Analysis
is one of the phases in cybercrime investigation, where the others focus on evidence collection, preservation, chain
of custody, information extraction that precede the analysis. The phases the follow the analysis are formulation of a
report and potential prosecution, typically involving expert witnesses. There are quite a few techniques, tools (hardware and software), and methodologies have been developed for all the briefly mentioned phases of the cybercrime
investigation. A lot of attention has been paid to the tool development for evidence collection and preservation; a few
tools have been developed to aid “browsing” data in the confiscated storage media, log files, memory, and so on. A lot
less number of tools have been developed for case analysis of the data, and the existing commercial packages (e.g. Encase or FTK) are very expensive. Even less so there are case management, event modeling, and event reconstruction,
especially with solid formal theoretical base. The first formal approach to the cybercrime investigation was the finitestate automata (FSA) approach by Gladyshev et. al [22,23]. The approach is complex to use and understand for non
computer science or equivalent investigators. The aim of Forensic Lucid is to alleviate those difficulties, be sound and
complete, expressive and usable, and provide even further usability improvement with the graphic interface that allow
data-flow graph-based (DFG) programming that allows translation between DFGs and Lucid code for compilation and
is implemented for Indexical Lucid in GIPSY already [24], and requires forensic extensions. While Forensic Lucid is
in the design and implementation, its solid base is being established in part with this work. The goal of Forensic Lucid
in the cyberforensic analysis is to be able to express in a program form the encoding of the evidence, witness stories,
and evidential statements, that can be tested against claims to see if there is a possible sequence or multiple sequences
of events that explain a given story. This is designed to aid investigator to avoid ad-hoc conclusions and have them
look at the possible explanations the Forensic Lucid program execution would yield and refine the investigation, as
was shown in the works [22,23] investigators failed to analyze all the stories and their plausibility before drawing
conclusions in the case. We do not recite the cases here due to the length limitations.
2 Forensic Lucid
The end goal is to define our Forensic Lucid language where its constructs concisely express cyberforensic evidence,
which can be initial state of a case towards what we have actually observed as a final state. The implementing system
(i.e. GIPSY) has to backtrace intermediate results in order to provide the corresponding event reconstruction path, if
it exists. The result of the expression in its basic form is either true or false, i.e. “guilty” or “not guilty” given the
context per explanation with the backtrace. There can be multiple backtraces, that correspond to the explanation of the
evidence (or lack thereof).
2.1 Properties
We define Forensic Lucid to model the evidential statements and other expressions representing the evidence and
observations as a higher-order context hierarchy. An execution trace of a Forensic Lucid program would expose the
possibility of the proposed claim with the events in the middle.
Addition of the context calculus from Lucx for operators on Lucx’s context sets (union, intersection, etc.) are
used to address to provide a collection of traces. Forensic Lucid inherits the properties of Lucx, MARFL, Objective
Lucid, JOOIP (and their comprising dialects), where the former is for the context calculus, and the latter for the arrays
and structural representation of data for modeling the case data structures such as events, observations, and groupings
of the related data.
One of the basic requirements is that the complete definition of the operational semantics of Forensic Lucid should
be compatible with the basic Lucx and GIPL, i.e. the translation rules or equivalent are to be provided when implementing the language compiler within GIPSY, and such that the GEE can execute it with minimal changes.
foo @
{
[ f i n a l observed event , p o s s i b l e
[
],
[
]
}
i n i t i a l observed event ] ,
Listing 1.1. Intensional Storyboard Expression
While the [...] notation here may be confusing with respect to the notation of [dimension:tag] in Lucid
and more specifically in Lucx [13,25], it is in fact a simple syntactical extension to allow higher-level groups of
contexts where this syntactical sugar is later translated to the baseline context constructs. The tentative notation of
{[...],...,[...]} implies a notion similar to the notion of the “context set” in [13,25] except with the syntactical
sugar mentioned earlier where we allow syntactical grouping of properties, observations, observation sequences, and
evidential statements as our context sets.
2.2 Transition Function
A transition function determines how the context of evaluation changes during computation. A general issue exists that
we have to address is that the transition function ψ is problem-specific. In the FSA approach, the transition function is
the labeled graph itself. In the first prototype, we follow the graph to model our Forensic Lucid equivalent. In general,
Lucid has already basic operators to navigate and switch from one context to another, which represent the basic
transition functions in themselves (the intensional operators such as @, #, iseod, first, next, fby, wvr, upon, and
asa as well as their inverse operators1). However, a specific problem being modeled requires more specific transition
function than just plain intensional operators. In this case the transition function is a Forensic Lucid function where
the matching state transition modeled through a sequence of intensional operators. In fact, the forensic operators are
just pre-defined functions that rely on traditional and inverse Lucid operators as well as context switching operators
that achieve something similar to the transitions in [22,23]. In fact, the intensional operators of Lucid represent the
basic building blocks for ψ and Ψ −1 .
2.3 Primitive Operators
The basic set of the classic intensional operators is extended with the similar operators, but inverted in one of their
aspects: either negation of trueness or reverse of direction of navigation. Here we provide an informal definition
followed by their formal counterpart of these operators alongside with the classical ones (to remind the reader what
they do and enlighten the unaware reader). The reverse operators have a restriction that they must work on the bounded
streams at the positive infinity. This is not a stringent limitation as the our contexts of observations and evidence in
this work are always finite, so they all have the beginning and the end. What we need is an ability to go back in the
stream and, perhaps, negate in it with classical-like operators, but reversed.
The operators are defined below to give a complete picture. The classical operators first, next, fby, wvr, upon,
and asa were previously defined in [11] and earlier. The other complimentary, inverse, and negation operators were
defined and revised from [26]. In this list of operators, especially the reverse ones, we make an important assumption
that the streams we are working with are finite, which is sufficient for our tasks. Thus, our streams of context values
can be bound between bod and eod and contain a finite tag set of elements is used as a context type. For summary of
the application of the just defined operators’ examples, please refer to Appendix B.
Following the steps in [11], we further represent the definition of the operators via @ and #. Again, there is a mix
of classical operators that were previously defined in [11], such as first, next, fby, wvr, upon, and asa as well as
the new operators from this work. The collection of the translated operators denoted in monospaced font, while we
provide their equivalence to the original Lucid operators, denoted as small caps.
The primitive operators are founding blocks to construct more complex case-specific functions that represent a
particular investigation case as well as more complex so-called forensic operators.
– A stream of first elements of stream X:
first X = (x0 , x0 , ..., x0 , ...)
first X = X@0
– A stream of second elements of stream X:
second X = (x1 , x1 , ..., x1 , ...) = first next X
1
Defined further.
(1)
– A stream of last elements of stream X:
last X = (xn , xn , ..., xn , ...)
This definition of the last operator relies on the earlier stated assumption that our streams can be explicitly finite
for the language we are developing. This affects the follow up operators that rely in that fact just as well. It is also
important to note that the last operator in our design does not return eod all the time on the finite stream due to
lack of usefulness for such a value; instead it returns the element of the stream just before the eod.
last X = X@(#@(#iseod(#) − 1))
(2)
– A stream of elements one before the last one of stream X:
prelast X = (xn−1 , xn−1 , ..., xn−1 , ...) = last prev X
– A stream of elements of stream X other than the first:
next X = (x1 , x2 , ..., xi+1 , ...)
next X = X@(# + 1)
(3)
– A stream of elements of stream X other than the last:
prev X = (xn−1 , ..., xi+1 , xi , xi−1 , ...)
prev X = X@(# − 1)
(4)
X fby Y = if # = 0 then X else Y @(# − 1)
= if isbod X then X else prev Y
(5)
– First element of X followed by all of Y :
X fby Y = (x0 , y0 , y1 , ..., yi−1 , ...)
– First element of X preceded by all of Y :
X pby Y = (y0 , y1 , ..., yi−1 , ..., yn , x0 )
X pby Y = if iseod # then X else Y @(# + 1)
(6)
= if iseod Y then X else next Y
– Stream of negated arithmetic values of X:
neg X = (−x0 , −x1 , −x2 , ..., −xi+1 , ...)
neg X = −X
(7)
not X = if X then !X else X
(8)
– Stream of inverted truth values of X:
not X = (!x0 , !x1 , !x2 , ..., !xi+1 , ...)
– A logical AND stream of truth values of X and Y :
X and Y = (x0 &&y0 , x1 &&y1 , x2 &&y2, ..., xi+1 &&yi+1 , ...)
X and Y = X&&Y
(9)
– A logical OR stream of truth values of X and Y :
X or Y = (x0 ||y0 , x1 ||y1 , x2 ||y2 , ..., xi+1 ||yi+1 , ...)
X or Y = X||Y
(10)
– A logical XOR stream of truth values of X and Y :
X xor Y = (x0 ⊕ y0 , x1 ⊕ y1, x2 ⊕ y2 , ..., xi+1 ⊕ yi+1 , ...)
X xor Y = not((X and Y ) or not (X or Y ))
(11)
– wvr stands for whenever. wvr chooses from its left-hand-side operand only values in the current dimension where
the right-hand-side evaluates to true.
X wvr Y =
if first Y 6= 0
then X fby (next X wvr next Y )
else (next X wvr next Y )
X wvr Y = X@T where
(12)
T = U fby U@(T + 1)
U = if Y then # else next U
end
– rwvr stands for retreat whenever. rwvr chooses from its left-hand-side operand backwards only values in the
current dimension where the right-hand-side evaluates to true.
X rwvr Y =
if last Y 6= 0
then X pby (prev X rwvr prev Y )
else (prev X rwvr prev Y )
X rwvr Y = X@T where
T = U pby U@(T − 1)
(13)
U = if Y then # else prev U
end
– nwvr stands for not whenever. nwvr chooses from its left-hand-side operand only values in the current dimension
where the right-hand-side evaluates to false.
X nwvr Y = X wvr not Y =
if first Y == 0
then X fby (next X nwvr next Y )
else (next X nwvr next Y )
X nwvr Y = X@T where
(14)
T = U fby U@(T + 1)
U = if Y == 0 then # else next U
end
– nrwvr stands for do not retreat whenever. nrwvr chooses from its left-hand-side operand backwards only values
in the current dimension where the right-hand-side evaluates to false.
X nrwvr Y = X rwvr not Y =
if last Y == 0
then X pby (prev X nrwvr prev Y )
else (prev X nrwvr prev Y )
X rnwvr Y = X@T where
T = U pby U@(T − 1)
(15)
U = if Y == 0 then # else prev U
end
– asa stands for as soon as. asa returns the value of its left-hand-side as a first point in that stream as soon as the
right-hand-side evaluates to true.
X asa Y = first (X wvr Y )
X asa Y = first (X wvr Y )
(16)
– ala (other suggested name is rasa) stands for as late as (or reverse of a soon as). ala returns the value of its
left-hand-side as the last point in that stream when the right-hand-side evaluates to true for the last time.
X ala Y = last (X wvr Y )
X ala Y = last (X rwvr Y )
(17)
– nasa stands for not as soon as. nasa returns the value of its left-hand-side as a first point in that stream as soon
as the right-hand-side evaluates to false.
X nasa Y = first (X nwvr Y )
X nasa Y = first (X nwvr Y )
(18)
– nala (other suggested name is nrasa) stands for not as late as (or reverse of not a soon as). nala returns the
value of its left-hand-side as the last point in that stream when the right-hand-side evaluates to false for the last
time.
X nala Y = last (X nwvr Y )
X nala Y = last (X nrwvr Y )
(19)
– upon stands for advances upon. Unlike asa, upon switches context of its left-hand-side operand if the right-hand
side is true.
X upon Y = X fby (
if first Y 6= 0
then (next X upon next Y )
else (X upon next Y ))
X upon Y = X@W where
(20)
W = 0 fby (if Y then (W + 1) else W )
end
– rupon stands for retreats upon. rupon switches context backwards of its left-hand-side operand if the right-hand
side is true.
X rupon Y = X pby (
if last Y 6= 0
then (prev X rupon prev Y )
else (X rupon prev Y ))
X rupon Y = X@W where
W = 0 pby (if Y then (W − 1) else W )
end
(21)
– nupon stands for not advances upon or rather advances otherwise. nupon switches context of its left-hand-side
operand if the right-hand side is false.
X nupon Y = X upon not Y = X fby (
if first Y == 0
then (next X nupon next Y )
else (X nupon next Y ))
X nupon Y = X@W where
(22)
W = 0 fby (if Y == 0 then (W + 1) else W )
end
– nrupon stands for not retreats upon. nrupon switches context backwards of its left-hand-side operand if the
right-hand side is false.
X nrupon Y = X rupon not Y = X pby (
if last Y == 0
then (prev X nrupon prev Y )
else (X nrupon prev Y ))
X nrupon Y = X@W where
(23)
W = 0 pby (if Y == 0 then (W − 1) else W )
end
2.4 Forensic Operators
The operators presented here are based on the discussion of the combination function and others that form morethan-primitive operations to support the required implementation. The discussed earlier comb() operator needs to be
realized in the general manner for combining analogies of MPRs, which in our case are higher-level contexts, in the
new language’s dimension types.
– combine corresponds to the comb function as originally described by Gladyshev in [22]. It is defined in Listing 1.2.
It is a preliminary context-enhanced version.
/∗ ∗
∗ Append g i v e n e t o e a c h e l e m e n t
∗ o f a gi ven stream e under t h e
∗ context of d .
∗
∗ @return t h e r e s u l t i n g c o m b i n e d s t r e a m
∗/
combine ( s , e , d ) =
i f i s e o d s t h e n eod ;
e l s e ( f i r s t s f b y . d e ) f b y . d combine ( n e x t s , e , d ) ;
fi
Listing 1.2. The combine Operator
– product tentatively corresponds to the cross-product [22] of contexts. It is defined in Listing 1.3.
The translated examples show recursion that we are not prepared to deal with in the current Lucid semantics, and
will address that in the future work. The two illustrated operators are the first of the a few more to follow in the final
language prototype.
/∗ ∗
∗ Append e l e m e n t s
∗ in all possible
∗/
p r o d u c t ( s1 , s2 , d )
i f i s e o d s2 then
e l s e combine ( s1 ,
fi
o f s2 t o element o f s1
combinations .
=
eod ;
f i r s t s 2 ) f b y . d p r o d u c t ( s1 , n e x t s 2 ) ;
Listing 1.3. The product Operator
2.5 Operational Semantics
As previously mentioned, the operational semantics of Forensic Lucid for the large part is viewed as a composition of
the semantic rules of Indexical Lucid, Objective Lucid, and Lucx along with the new operators and definitions. Here
we list the existing combined semantic definitions to be used the new language, specifically extracts of operational
semantics from GIPL [11], and Lucx [13] are in Figure 1, and Figure 3 respectively. The explanation of the rules and
the notation are given in great detail in the cited works and are trimmed in this article. For convenience of the reader
they are recited here to a degree. The new rules of the operational semantics of Forensic Lucid cover the newly defined
operators primarily, including the reverse and logical stream operators as well as forensic-specific operators. We use
the same notation as the referenced languages to maintain consistency in defining our rules.
In the implementing system, GIPSY, the GIPL is the generic counterpart of all the Lucid programming languages.
Like Indexical Lucid, which it is derived from, it has only the two standard intensional operators: E @ C for evaluating
an expression E in context C, and #d for determining the position in dimension d of the current context of evaluation
in the context space [11]. SIPLs are Lucid dialects (Specific Intensional Programming Languages) with their own
attributes and objectives. Theoretically, all SIPLs can be translated into the GIPL [11]. All the SIPLs conservatively
extend the GIPL syntactically and semantically. The remainder of this section presents a relevant piece of Lucx as
a conservative extension to GIPL. The semantics of GIPL is presented in Figure 1. The excerpt of semantic rules of
Lucx are then presented as a conservative extension to GIPL in Figure 3. Following is the description of the GIPL
semantic rules as presented in [11]:
D ⊢E :v
tells that under the definition environment D, expression E would evaluate to value v.
D, P ⊢ E : v
specifies that in the definition environment D, and in the evaluation context P (sometimes also referred to as a point
in the context space), expression E evaluates to v. The definition environment D retains the definitions of all of the
identifiers that appear in a Lucid program, as created with the semantic rules 13-16 in Figure 1. It is therefore a partial
function
D : Id → IdEntry
where Id is the set of all possible identifiers and IdEntry, has five possible kinds of value, one for each of the kinds
of identifier: 1. Dimensions define the coordinate pairs, in which one can navigate with the # and @ operators. Their
IdEntry is simply (dim). 2. Constants are external entities that provide a single value, regardless of the context
of evaluation. Examples are integers and Boolean values. Their IdEntry is (const, c), where c is the value of the
constant. 3. Data operators are external entities that provide memoryless functions. Examples are the arithmetic and
Boolean functions. The constants and data operators are said to define the basic algebra of the language. Their IdEntry
is (op, f ), where f is the function itself. 4. Variables carry the multidimensional streams. Their IdEntry is (var, E),
where E is the Lucid expression defining the variable. It should be noted that this semantics makes the assumption
that all variable names are unique. This constraint is easy to overcome by performing compile-time renaming or using
a nesting level environment scope when needed. 5. Functions are non-recursive GIPL user-defined functions. Their
IdEntry is (func, idi , E), where the idi are the formal parameters to the function and E is the body of the function. In
this paper we do not discuss the semantics of recursive functions.
D(id) = (const, c)
D, P ⊢ id : c
(24)
Eopid :
D(id) = (op, f )
D, P ⊢ id : id
(25)
Edid :
D(id) = (dim)
D, P ⊢ id : id
(26)
Efid :
D(id) = (func, idi , E)
D, P ⊢ id : id
(27)
Evid :
D(id) = (var, E)
D, P ⊢ E : v
D, P ⊢ id : v
(28)
Eop :
D, P ⊢ E : id
D(id) = (op, f )
D, P ⊢ Ei : vi
D, P ⊢ E(E1 , . . . , En ) : f (v1 , . . . , vn )
(29)
Ecid :
Efct :
D, P ⊢ E : id
D(id) = (func, idi , E ′ )
D, P ⊢ E ′ [idi ← Ei ] : v
D, P ⊢ E(E1 , . . . , En ) : v
(30)
EcT :
D, P ⊢ E : true
D, P ⊢ E ′ : v′
D, P ⊢ if E then E ′ else E ′′ : v′
(31)
EcF :
D, P ⊢ E : false
D, P ⊢ E ′′ : v′′
D, P ⊢ if E then E ′ else E ′′ : v′′
(32)
Etag :
D, P ⊢ E : id
D(id) = (dim)
D, P ⊢ #E : P(id)
(33)
Eat :
Ew :
Qdim :
Qid :
Qfid :
QQ :
D, P ⊢ E ′ : id
D(id) = (dim)
D, P ⊢ E ′′ : v′′
D, P ⊢ E @E ′ E ′′ : v
D, P ⊢ Q : D ′ , P ′
D ′, P ′ ⊢ E : v
D, P ⊢ E where Q : v
D, P ⊢ dimension id : D †[id 7→ (dim)], P †[id 7→ 0]
D, P ⊢ id = E : D †[id 7→ (var, E)], P
D, P ⊢ id(id1 , . . . , idn ) = E : D †[id 7→ (func, idi , E)], P
D, P ⊢ Q : D ′ , P ′
D ′ , P ′ ⊢ Q′ : D ′′ , P ′′
D, P ⊢ Q Q′ : D ′′ , P ′′
D, P †[id 7→ v′′ ] ⊢ E : v
(34)
(35)
(36)
(37)
(38)
(39)
Fig. 1. GIPL Semantics
EE.did :
D(E.id) = (dim)
D, P ⊢ E.id : id.id
(40)
Fig. 2. Higher-Order Context Dot Operator
The evaluation context P, which is changed when the @ operator is evaluated, or a dimension is declared in a
where clause, associates a tag (i.e. an index) to each relevant dimension. It is, therefore, a partial function
P : Id → N
Each type of identifiers can only be used in the appropriate situations. Identifiers of type op, func, and dim evaluate to
themselves (Figure 1, rules 25,26,27). Constant identifiers (const) evaluate to the corresponding constant (Figure 1,
rule 24). Function calls, resolved by the Efct rule (Figure 1, rule 30), require the renaming of the formal parameters into
the actual parameters (as represented by E ′ [idi ← Ei ]). The function P ′ = P †[id 7→ v′′ ] specifies that P ′ (x) is v′′ if
x = id, and P(x) otherwise. The rule for the where clause, Ew (Figure 1, rule 35), which corresponds to the syntactic
expression E where Q, evaluates E using the definitions Q therein. The additions to the definition environment D
and context of evaluation P made by the Q rules (Figure 1, rules 36,37,38) are local to the current where clause.
This is represented by the fact that the Ew rule returns neither D nor P. The Qdim rule adds a dimension to the
definition environment and, as a convention, adds this dimension to the context of evaluation with tag 0 (Figure 1,
rule 36). The Qid and Qfid simply add variable and function identifiers along with their definition to the definition
environment (Figure 1, rules 37,38).
As a conservative extension to GIPL, Lucx’s semantics introduces the notion of context as a building block into
the semantic rules, i.e. context as a first-class value, as described by the rules in Figure 3. In Lucx, semantic rule 42
(Figure 3) creates a context as a semantic item and returns it as a context P that can then be used by rule 43 to
navigate to this context by making it override the current context. GIPL’s semantic rule 29 is still valid for the definition
of the context operators, where the actual parameters evaluate to values vi that are contexts Pi . The semantic rule 41
expresses that the # symbol evaluates to the current context. When used as a parameter to the context calculus operators,
this allows for the generation of contexts relative to the current context of evaluation.
E#(cxt) :
Econstruction(cxt)
D, P ⊢ # : P
D, P ⊢ Ed j : id j
D(id j ) = (dim)
D, P ⊢ Ei j : v j
P ′ = P0 †[id1 7→ v1 ]†. . .†[idn 7→ vn ]
:
D, P ⊢ [Ed1 : Ei1 , Ed2 : Ei2 , . . . , Edn : Ein ] : P ′
Eat(cxt) :
D, P ⊢ E ′ : P ′
D, P †P ′ ⊢ E : v
D, P ⊢ E @ E ′ : v
(41)
(42)
(43)
Fig. 3. Conservative Semantic Rules Introduced by Lucx
3 Conclusion
While the list of Isabelle’s proofs is incomplete at the time of the writing of this manuscript some formalization in
Isabelle took place, and the work on them is currently on-going.
3.1 Results
Due to a non-standard nature of the Lucid language (as opposed to standard imperative languages), it takes some time
to understand the full scope of some of its details and model them. This complicates a way to model its operators,
expressions, overall meaning in Isabelle. This fact resulted in several trials and attempts to approach the language,
from fairly complex to fairly basic – plain integers and pipelined processing and basic index support. They are not
fully complete, but some of the basic properties are modeled and proven; please refer to the Isabelle sources for details
(once completed it is planned to be released as a part of the Archive of Formal Proofs at [27]).
– The IntegerLucid Isabelle file is the most developed out of all as far as definition and exploitation of intensional
operators of classical Lucid concerned. It is called “integer” because all the streams and dimensions and all operators around them play with integers, natural numbers, and in rarer cases Booleans. There are no identifiers in there.
The Isabelle file contains three theories: OriginalLucidOperators, LucidOperators, and IntegerLucid.
The first models classical Lucid operators as pipelined dataflows. The second adds up some index support and
proves equivalence to the first definitions. The latter provides new definitions of the intensional operators through
@ and #, defines meaning functions, propositions, and lemmas from [11]. Integer Lucid proves the example for N
@.d 2 = 44 for the at().
– The BasicLucid theory is currently the second one derived to support Lucid definitions. It is an extension of
IntegerLucid by adding identifiers. asa and upon are in this theory.
– The LucidSemanticRules theory is meant to have the meaning of complete semantic rules and proven, but it
only has a definition of a Hoare tuple [28] and a meaning function for it.
– The CommonLucidTypes theory is used by all (most) theories and defines some common types used by most [29].
– ForensicLucid.thy, GIPL.thy, IndexicalLucid.thy, JLucid.thy, JOOIP.thy, Lucx.thy, ObjectiveLucid.thy
are the theories under current development with some results from the above. The completed work will have a
complete list of the files publicly available and submitted to the AfP [27].
3.2 Future Work
The near-future work will consist primarily of the following items:
– Complete semantics of all the mentioned Lucid dialects and their formalization with Isabelle.
– Augment the language specification to include the Depmster-Shafer theory [30,31] of evidence to allow weights
for claims, credibility, belief, and plausibility parameters.
– Prove semantic rules involving intensional data warehouse.
– Implementation of the Forensic Lucid compiler, run-time and interactive development environments.
4 Acknowledgments
This research and development work was funded in part by NSERC and the Faculty of Engineering and Computer
Science of Concordia University, Montreal, Canada. Thanks to Drs. Mourad Debbabi, Patrice Chalin, Peter Grogono
on valuable suggestions used in this work.
References
1. Mokhov, S.: Intensional Forensics – the Use of Intensional Logic in Cyberforensics. Technical report, Concordia Institute for
Information Systems Engineering, Concordia University, Montreal, Canada (January 2007) ENGR6991 Technical Report.
2. Mokhov, S.: Intensional Cyberforensics – a PhD Proposal. Department of Computer Science and Software Engineering,
Concordia University, Montreal, Canada (December 2007)
3. The GIPSY Research and Development Group:
The General Intensional Programming System (GIPSY) project.
Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada (2002-2008)
http://newton.cs.concordia.ca/~ gipsy/, last viewed April 2008.
4. Paulson, L.C., Nipkow, T.: Isabelle: A generic proof assistant. University of Cambridge and Technical University of Munich
(2007) http://isabelle.in.tum.de/, last viewed: December 2007.
5. Wadge, W., Ashcroft, E.: Lucid, the Dataflow Programming Language. Academic Press, London (1985)
6. Edward Ashcroft and Anthony Faustini and Raganswamy Jagannathan and William Wadge: Multidimensional, Declarative
Programming. Oxford University Press, London (1995)
7. Ashcroft, E.A., Wadge, W.W.: Lucid - A Formal System for Writing and Proving Programs. Volume 5., SIAM J. Comput. no.
3 (1976)
8. Ashcroft, E.A., Wadge, W.W.: Erratum: Lucid - A Formal System for Writing and Proving Programs. Volume 6(1):200., SIAM
J. Comput. (1977)
9. Ashcroft, E.A., Wadge, W.W.: Lucid, a nonprocedural language with iteration. Communication of the ACM 20(7) (July 1977)
519–526
10. Gagné, J.R., Plaice, J.: Demand-Driven Real-Time Computing, World Scientific (September 1999)
11. Paquet, J.: Scientific Intensional Programming. PhD thesis, Department of Computer Science, Laval University, Sainte-Foy,
Canada (1999)
12. Wan, K., Alagar, V., Paquet, J.: A Context theory for Intensional Programming. In: Workshop on Context Representation and
Reasoning (CRR05), Paris, France. (July 2005)
13. Wan, K.: Lucx: Lucid Enriched with Context. PhD thesis, Department of Computer Science and Software Engineering,
Concordia University, Montreal, Canada (2006)
14. Paquet, J., Mokhov, S.A., Tong, X.: Design and implementation of context calculus in the GIPSY environment. In: Proceedings
of the 32nd Annual IEEE International Computer Software and Applications Conference (COMPSAC), Turku, Finland, IEEE
Computer Society (July 2008) 1278–1283
15. Tong, X.: Design and implementation of context calculus in the GIPSY. Master’s thesis, Department of Computer Science and
Software Engineering, Concordia University, Montreal, Canada (April 2008)
16. Wan, K., Alagar, V., Paquet, J.: Lucx: Lucid Enriched with Context. In: Proceedings of the 2005 International Conference on
Programming Languages and Compilers (PLC 2005), Las Vegas, USA, CSREA Press (June 2005) 48–14
17. Mokhov, S.A.: Towards syntax and semantics of hierarchical contexts in multimedia processing applications using MARFL.
In: Proceedings of the 32nd Annual IEEE International Computer Software and Applications Conference (COMPSAC), Turku,
Finland, IEEE Computer Society (July 2008) 1288–1294
18. Swoboda, P.: A Formalisation and Implementation of Distributed Intensional Programming. PhD thesis, The University of
New South Wales, Sydney, Australia (2004)
19. Swoboda, P., Wadge, W.W.: Vmake, ISE, and IRCS: General tools for the intensionalization of software systems. In Gergatsoulis, M., Rondogiannis, P., eds.: Intensional Programming II, World-Scientific (2000)
20. Swoboda, P., Plaice, J.: A new approach to distributed context-aware computing. In Ferscha, A., Hoertner, H., Kotsis, G., eds.:
Advances in Pervasive Computing, Austrian Computer Society (2004) ISBN 3-85403-176-9.
21. Swoboda, P., Plaice, J.: An active functional intensional database. In Galindo, F., ed.: Advances in Pervasive Computing,
Springer (2004) 56–65 LNCS 3180.
22. Gladyshev, P., Patel, A.: Finite state machine approach to digital event reconstruction. In: Digital Investigation Journal.
Volume 2. (2004)
23. Gladyshev, P.: Finite state machine analysis of a blackmail investigation. In: International Journal of Digital Evidence,
Technical and Security Risk Services, Sprint 2005, Volume 4, Issue 1 (2005)
24. Ding, Y.M.: Bi-directional translation between data-flow graphs and Lucid programs in the GIPSY environment. Master’s
thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada (2004)
25. Tong, X., Paquet, J., Mokhov, S.A.: Context Calculus in the GIPSY. Unpublished (2007)
26. Mokhov, S.A., Paquet, J., Debbabi, M.: Designing a language for intensional cyberforensic analysis. Unpublished (2007)
27. Klein, G., Nipkow, T., Paulson, L.C.:
The archive of formal proofs.
SourceForge.net (2008)
http://afp.sourceforge.net/, last viewed: April 2008.
28. Moeller, A.:
Program Verification with Hoare Logic.
Technical report, University of Aarhus (2004)
http://www.brics.dk/~ amoeller/talks/hoare.pdf.
29. Mokhov, S.A., Paquet, J., Tong, X.: Hybrid intensional-imperative type system for intensional logic support in GIPSY. Submitted for publication at LPAR’08 (2008)
30. Shafer, G.: The Mathematical Theory of Evidence. Princeton University Press (1976)
31. Haenni, R., Kohlas, J., Lehmann, N.: Probabilistic argumentation systems. Technical report, Institute of Informatics, University
of Fribourg, Fribourg, Switzerland (October 1999)
32. Kahn, G.: The semantics of a simple language for parallel processing. In: Proceedings of the IFIP Congress ’74, Amsterdam,
Elsevier North-Holland (1974) 471–475
33. Kahn, G., MacQueen, D.B.: Coroutines and networks of parallel processes. In: Proceedings of the IFIP Congress ’77, Amsterdam, Elsevier North-Holland (1977) 993–998
34. Landin, P.J.: The next 700 programming languages. Communications of the ACM 9(3) (1966) 157–166
Appendix
A
Lucid Axioms, Theorems, and Proofs
Here we present and extend the notion of the formalisms from Paquet [11] and extend them on to the present work.
A.1 Streaming and Basic Operators
The origins of Lucid date back to 1974. At that time, Ashcroft and Wadge were working on a purely declarative
language, in which iterative algorithms could be expressed naturally, which eventually resulted in [9]. Their work
fits into the broad area of research into program semantics and verification. It would later turn out that their work is
also relevant to the dataflow networks and coroutines of Kahn and MacQueen [32,33]. In the original Lucid (whose
operators are in this font), streams were defined in a pipelined manner, with two separate definitions: one for the
initial element, and another one for the subsequent elements. For example, the equations
first X = 0
next X = X + 1
define variable X to be a stream, such that
x0 = 0
xi+1 = xi + 1
In other words,
0 = (0, 0, 0, ..., 0, ...)
X = (x0 , x1 , . . . , xi , . . .) = (0, 1, . . . , i, . . .)
Similarly, the equations
first X = X
next Y = Y + next X
define variable Y to be the running sum of X, i.e.
y0 = x0
yi+1 = yi + xi+1
In other words,
i(i + 1)
Y = (y0 , y1 , . . . , yi , . . .) = 0, 1, . . . ,
,...
2
It soon became clear that a “new” operator at the time, fby (followed by) can be used to define such typical situations.
Hence, the above two variables could be defined as follows:
X = 0 fby X + 1
Y = X fby Y + next X
As a result, we can summarize the three basic operators of the original Lucid.
Definition 1 If X = (x0 , x1 , . . . , xi , . . .) and Y = (y0 , y1 , . . . , yi , . . .), then
def
(1) first X = (x0 , x0 , . . . , x0 , . . .)
def
(2) next X = (x1 , x2 , . . . , xi+1 , . . .)
def
(3) X fby Y = (x0 , y0 , y1 , . . . , yi−1 , . . .)
Here parallels can be drawn to the list operations, where first corresponds to head, next corresponds to tail,
and fby corresponds to cons. When these operators are combined with Landin’s ISWIM [34] (If You See What I
Mean), essentially typed λ -calculus with syntactic sugar, it becomes possible to define complete Lucid programs. The
following three derived operators have turned out to be very useful (we will use them later in the text):
Definition 2
(1) X wvr Y
def
(2) X asa Y
def
= if first Y then X fby ( next X wvr next Y )
else
( next X wvr next Y )
= first (X wvr Y )
def
(3) X upon Y = X fby (if first Y then ( next X upon next Y )
else (
X upon next Y ))
Where wvr stands for whenever, asa stands for as soon as and upon stands for advances upon.
A.2 Random Access to Streams
With the original Lucid operators, one could only define programs with pipelined dataflows, i.e. in which the (i + 1)-th
element in a stream is only computed once the i-th element has been computed. This situation is potentially wasteful of
resources, since the i-th element might not necessarily be required. More importantly, it only allows sequential access
into streams.
By taking a different approach, it is possible to have random access into streams, using an index # corresponding
to the current position, the current context of evaluation. No longer are we manipulating infinite extensions (streams),
rather we are defining computation according to a context (here a single integer). We have set out on the road to
intensional programming. We redefine all original Lucid operators in terms of the operators # and @:
Definition 3
(1) #
def
= 0 fby (# + 1)
def
(2) X @ Y = if Y = 0 then first X
else ( next X) @ (Y − 1)
Further, we give definitions for the original operators using these two baseline operators. In so doing, we will use the
following axioms.
Axiom 1 Let i ≥ 0.
(1) [c]i = c
(2) [X + c]i = [X]i + c
(3) [ first X]i = [X]0
(4) [ next X]i = [X]i+1
(5) [X fby Y ]0 = [X]0
(6) [X fby Y ]i+1 = [Y ]i
(7) if true then [X]i else [Y ]i = [X]i
(8) if false then [X]i else [Y ]i = [Y ]i
(9) [if C then X else Y ]i = if [C]i then [X]i else [Y ]i
Prior giving the re-definitions of the standard Lucid operators, we show some basic properties of @ and #. We will
use throughout the discussion here [X]i instead of xi , as it allows for greater readability. Furthermore, we will, as is
standard, write X = Y whenever we have
(∀i : i ≥ 0 : [X]i = [Y ]i )
Proposition 1. Let i ≥ 0.
(1) [#]i = i
(2) [X @ Y ]i = [X][Y ]i
Proof
(1) Proof by induction over i.
Base step (i = 0).
[#]0 = [0 fby (# + 1)]0
= [0]0
=0
Defn. 3.1
Axiom 1.5
Axiom 1.1
Induction step (i = k + 1). Suppose (∀i : i ≤ k : [#]i = i).
[#]k+1 = [0 fby (# + 1)]k+1
= [# + 1]k
= [#]k + 1
= k+1
Defn. 3.1
Axiom 1.6
Axiom 1.2
Ind. Hyp.
Hence (∀i : i ≥ 0 : [#]i = i).
(2) Let i ≥ 0. We will prove by induction over yi that yi ≥ 0 ⇒ [X @ Y ]i = [X][Y ]i .
Base step (yi = 0).
[X @ Y ]i = [if Y = 0 then first X else ( next X) @ (Y − 1)]i
= if [Y = 0]i then [ first X]i else [( next X) @ (Y − 1)]i
= if [Y ]i = 0 then [ first X]i else [( next X) @ (Y − 1)]i
= [ first X]i
= [X]0
= [X][Y ]i
Defn. 3.2
Axiom 1.9
Axiom 1.2
Axiom 1.7
Axiom 1.3
Hypothesis
Induction step (yi = k + 1). Suppose (∀i : i ≤ k : [#]i = i).
[X @ Y ]i = [if Y = 0 then first X else ( next X) @ (Y − 1)]i
= if [Y = 0]i then [ first X]i else [( next X) @ (Y − 1)]i
= if [Y ]i = 0 then [ first X]i else [( next X) @ (Y − 1)]i
= [( next X) @ (Y − 1)]i
= [ next X][Y −1]i
= [ next X][Y ]i −1
= [X][Y ]i −1+1
= [X][Y ]i
Defn. 3.2
Axiom 1.9
Axiom 1.2
Axiom 1.8
Ind. Hyp.
Axiom 1.2
Axiom 1.4
Arith.
Hence (∀i : i ≥ 0 : [Y ]i ≥ 0 ⇒ ([X @ Y ]i = [X][Y ]i )).
Definition 4
def
(1)
first X = X @ 0
(2)
next X
(3)
X fby Y
(4)
X wvr Y
(4.1)
(4.2)
(5)
X asa Y
(6)
def
= X @ (# + 1)
def
= if # = 0 then X else Y @ (# − 1)
def
= X @T
where
T = U fby U @ (T + 1)
U = if Y then # else next U
end
def
= first (X wvr Y )
def
X upon Y = X @ W
where
(6.1)
W = 0 fby if Y then (W + 1) else W
end
The advantage of these new definitions is that they do not use any form of recursive function definitions. Rather, all of
the definitions are iterative, and in practice, more easily implemented in an efficient manner. We prove below that the
new definitions are equivalent to the old ones.
Proposition 2. first X = first X.
Proof Let i ≥ 0. Then
[first X]i = [X @ 0]i
= [X][0]i
= [X]0
= [ first X]i
Hence first X = first X.
Defn. 4.1
Prop. 1.2
Axiom 1.1
Axiom 1.3
Proposition 3. next X = next X.
Proof Let i ≥ 0. Then
[next X]i = X @ (# + 1) i
= [X][#+1]i
= [X][#]i +1
= [X]i+1
= [ next X]i
Defn. 4.2
Prop. 1.2
Axiom 1.2
Prop. 1.1
Axiom 1.4
Hence next X = next X.
Proposition 4. X fby Y = X fby Y .
Proof Proof by induction over i.
Base step (i = 0).
[X fby Y ]0 = [if # = 0 then X else Y @ (# − 1)]0
= if [# = 0]0 then [X]0 else [Y @ (# − 1)]0
= if [#]0 = 0 then [X]0 else [Y @ (# − 1)]0
= if 0 = 0 then [X]0 else [Y @ (# − 1)]0
= [X]0
= [X fby Y ]0
Defn. 4.3
Axiom 1.9
Defn. 1.2
Prop. 1.1
Axiom 1.7
Axiom 1.5
Induction step (i = k + 1).
[X fby Y ]k+1 = if # = 0 then X else Y @ (# − 1) k+1
= if [# = 0]k+1 then [X]k+1 else [Y @ (# − 1)]k+1
= if [#]k+1 = 0 then [X]k+1 else [Y @ (# − 1)]k+1
= if k + 1 = 0then [X]k+1 else [Y @ (# − 1)]k+1
= Y @ (# − 1) k+1
= Y [#−1]
k+1
= Y [#] −1
k+1
= Y k
= [X fby Y ]k+1
Defn. 4.3
Axiom 1.9
Axiom 1.1
Prop. 1.1
Axiom 1.8
Prop. 1.2
Axiom 1.2
Prop. 1.1
Axiom 1.6
Hence (∀i : i ≥ 0 : [X fby Y ]i = [X fby Y ]i ). Hence fby = fby .
The proof for wvr is more complicated, as it requires relating an iterative definition to a recursive definition. We will
therefore need four lemmas that refer to variables T and U in the text in Definitions 4.4.1 and 4.4.2. In addition, we
must define the rank of a Boolean stream. Finally, we will have to introduce another set of axioms, that allow us to
compare two entire streams, as opposed to particular elements in the two streams.
Axiom 2 Let i ≥ 0.
(1) X 0 = X
(2) [X i ]0 = [X]i
(3) first X i = [X]i
(4) next X i = X i+1
(5) next (X fby Y ) = Y
(6) ( first X) fby Y = X fby Y
(7) if true then X else Y = X
(8) if false then X else Y = Y
Definition 5 Let Y be a Boolean stream.
def
(1) rank(−1,Y ) = −1
def
(2) rank(i + 1,Y ) = min{k : k > rank(i,Y ) : [Y ]k = true}
Further, we write ri for rank(i,Y ).
Lemma 1. (∀i : i ≥ −1 : (∀ j : ri < j ≤ ri+1 : X j wvr Y j = X ri+1 wvr Y ri+1 )).
Proof Let i ≥ −1. Proof by downwards induction over j. Note that ri < ri+1 .
Base step ( j = ri+1 ).
X ri+1 wvr Y ri+1 = X ri+1 wvr Y ri+1
Identity
Induction step ( j = k − 1, j > ri ).
X k−1 wvr Y k−1 = if first Y k−1 then X k−1 fby X k wvr Y k
else X k wvr Y k
= if [Y ]k−1 then X k−1 fby X k wvr Y k
else X k wvr Y k
= X k wvr Y k
= X ri+1 wvr Y ri+1
Defn. 2.1
Axiom 2.3
Axiom 2.8
Ind. Hyp.
Hence, (∀i : i ≥ −1 : (∀ j : ri < j ≤ ri+1 : X j wvr Y j = X ri+1 wvr Y ri+1 )).
Lemma 2. (∀i : i ≥ 0 : (X wvr Y )i = X ri wvr Y ri ).
Proof Proof by induction over i.
Base step (i = 0).
(X wvr Y )0 = X wvr Y
= X 0 wvr Y 0
= X r0 wvr Y r0
Axiom 2.1
Axiom 2.1
Lemma 1
Induction step (i = k + 1).
(X wvr Y )k+1 = next ((X wvr Y )k )
= next (X rk wvr Y rk )
= next (if first Y rk then X rk fby X rk +1 wvr Y rk +1
else X rk +1 wvr Y rk +1 )
= next (if [Y ]rk then X rk fby X rk +1 wvr Y rk +1
else X rk +1 wvr Y rk +1 )
r
k
= next (X fby X rk +1 wvr Y rk +1 )
= X rk +1 wvr Y rk +1
= X rk+1 wvr Y rk+1
Axiom 2.4
Ind. Hyp.
Defn. 2.1
Axiom 2.3
Axiom 2.7
Axiom 2.5
Lemma 1
Hence, (∀i : i ≥ 0 : (X wvr Y )i = X ri wvr Y ri ).
Lemma 3. (∀i : i ≥ −1 : (∀ j : ri < j ≤ ri+1 : [U] j = ri+1 )).
Proof Let i ≥ −1. Proof by downwards induction over j. Note that ri < ri+1 .
Base step ( j = ri+1 ).
[U]ri+1 = [if Y then # else next U]ri+1
= if [Y ]ri+1 then [#]ri+1 else [next U]ri+1
= [#]ri+1
= ri+1
Defn. 4.4.2
Axiom 1.9
Axiom 1.7
Prop. 1.1
Induction step ( j = k − 1, j > ri ).
[U]k−1 = [if Y then # else next U]k−1
= if [Y ]k−1 then [#]k−1 else [next U]k−1
= [next U]k−1
= [U]k
= ri+1
Hence, (∀i : i ≥ −1 : (∀ j : ri−1 < j < ri : [U] j = ri+1 )).
Defn. 4.4.2
Axiom 1.9
Axiom 1.8
Axiom 1.4
Ind. Hyp.
Lemma 4. (∀i : i ≥ 0 : [T ]i = ri ).
Proof Proof by induction over i.
Base step (i = 0).
[T ]0 = [U fby U @ (T + 1)]0
= [U]0
= r0
Defn. 4.4.1
Axiom 1.5
Lemma 3
Induction step (i = k + 1).
[T ]k+1 = [U fby U @ (T + 1)]k+1
= [U @ (T + 1)]k
= [U][T +1]k
= [U][T ]k +1
= [U]rk +1
= rk+1
Defn. 4.4.1
Axiom 1.6
Prop. 1.2
Axiom 1.2
Ind. Hyp.
Lemma 3
Hence, (∀i : i ≥ 0 : [T ]i = ri ).
Proposition 5. X wvr Y = X wvr Y .
Proof
[X wvr Y ]i = [X @ T ]i
= [X][T ]i
= [X]ri
= [X ri ]0
= [X ri fby X ri +1 wvr Y ri +1 ]0
= [if [Y ]ri then X ri fby X ri +1 wvr Y ri +1
else X ri +1 wvr Y ri +1 ]0
= [if first Y ri then X ri fby X ri +1 wvr Y ri +1
else X ri +1 wvr Y ri +1 ]0
r
r
= [X i wvr Y i ]0
= [(X wvr Y )i ]0
= [X wvr Y ]i
Defn. 4.4
Prop. 1.2
Lemma 4
Axiom 1.2
Axiom 1.6
Axiom 2.7
Axiom 2.3
Defn. 2.1
Lemma 2
Axiom 2.2
Hence X wvr Y = X wvr Y .
Proposition 6. X asa Y = X asa Y .
Proof
X asa Y = first (X wvr Y )
= first (X wvr Y )
= first (X wvr Y )
= X asa Y
Hence X asa Y = X asa Y .
Defn. 4.5
Prop. 5
Prop. 2
Defn. 2.2
Lemma 5. (∀i : i ≥ 0 : (X upon Y )i = X [W ]i upon Y i )
Proof Proof by induction over i.
Base step (i = 0).
(X upon Y )0 = X upon Y
= X 0 upon Y 0
= X [0 fby ...]0 upon Y 0
= X [W ]0 upon Y 0
Axiom 2.1
Axiom 2.1
Defn. 2.3
Defn. 4.6.1
Induction step (i = k + 1).
(X upon Y )k+1 = next (X upon Y )k
= next X [W ]k upon Y k
= if ( first Y k )
then (X [W ]k +1 upon Y k+1 )
else (X [W ]k upon Y k+1 )
= if [Y ]k
then (X [W ]k+1 upon Y k+1 )
else (X [W ]k upon Y k+1 )
= X (if [Y ]k then [W ]k+1 else [W ]k )
upon Y k+1
= X [W ]k+1 upon Y k+1
Axiom 2.4
Ind. Hyp.
Defn. 2.3 and
Axiom 2.5
Axiom 2.4
Defn. 4.6.1
Substit.
Defn. 4.6.1
Hence, (∀i : i ≥ 0 : (X upon Y )i = X [W ]i upon Y i )
Proposition 7. X upon Y = X upon Y .
Proof Let i ≥ 0. Then
[X upon Y ]i = [X @ W ]i
= [X][W ]i
= [X [W ]i ]0
= [X [W ]i fby . . .]0
= [X [W ]i upon Y i ]0
= [X upon Y ]i
Defn. 4.6
Prop. 1.2
Axiom 2.2
Axiom 1.5
Defn. 2.3
Lemma 5
Hence X upon Y = X upon Y .
Now that the corresponding definitions are shown to be equivalent, we can generalize and head off in the negative
direction as well:
Definition 6
def
(1) prev X = X @ (# − 1)
def
(2) X fby Y = if # ≤ 0 then X else Y @ (# − 1)
B Summary of the Operators’ Examples
Here we illustrate a few basic examples of application of the Forensic Lucid operators (both, classical Lucid and the
newly introduced operators). Assume we have two bounded (between bod and eod) streams X and Y of ten elements.
The X stream is just an ordered sequence of natural numbers between 1 and 10. If queried for values below 1 an
beginning-of-data (bod) marker would be returned; similarly if queried beyond 10, the end-of-data marker (eod) is
returned. The Y stream is a sequence of ten truth values (can be replaced with 0 for “false” and 1 for “true”). The
operators applied to these streams may return bounded or unbounded streams of the same or different length than
the original depending on the definition of a particular operator. Also assume the current dimension index is 0. The
resulting table showing the application of the classical and the new operators is in Table 1.
stream/index -1 0 1 2 3 4 5 6 7 8 9 10 11
X
bod 1 2 3 4 5 6 7 8 9 10 eod eod
Y
bod T F F T F F T T F T eod eod
X first Y
1 1 1 1 1 1 1 1 1 1
X last Y
10 10 10 10 10 10 10 10 10 10
X next Y
2 3 4 5 6 7 8 9 10 eod eod
X prev Y
bod
X fby Y
1 T F F T F F T T F T eod
X pby Y
T F F T F F T T F T 1 eod
X wvr Y
1
4
7 8
10
X rwvr Y
10
8
7 4
1
X nwvr Y
2 3
5 6
9
X nrwvr Y
9 6
5 3
2
X asa Y
1 1 1 1 1 1 1 1 1 1
X nasa Y
2 2 2 2 2 2 2 2 2 2
X ala Y
10 10 10 10 10 10 10 10 10 10
X nala Y
9 9 9 9 9 9 9 9 9 9
X upon Y
1 2 2 2 3 3 3 4 5 5 eod
X rupon Y
10 9 9 8 7 7 7 6 6 6 bod
X nupon Y
1 1 2 3 3 4 5 5 5 6 6 eod
X nrupon Y
10 10 9 9 9 8 7 7 6 5 5 bod
neg X
-1 -2 -3 -4 -5 -6 -7 -8 -9 -10 eod eod
not Y
F T T F T T F F T F eod eod
X and Y
1 0 0 1 0 0 1 1 0 1 eod eod
X or Y
1 2 3 5 5 6 7 9 9 11 eod eod
X xor Y
0 2 3 5 5 6 6 9 9 11 eod eod
Table 1. Example of Application of Forensic Lucid Operators to Bounded Streams
| 6cs.PL
|
Isotonic regression in general dimensions
Qiyang Han∗, Tengyao Wang†,
Sabyasachi Chatterjee‡and Richard J. Samworth§
arXiv:1708.09468v1 [math.ST] 30 Aug 2017
September 1, 2017
Abstract
We study the least squares regression function estimator over the class of real-valued
functions on [0, 1]d that are increasing in each coordinate. For uniformly bounded signals and with a fixed, cubic lattice design, we establish that the estimator achieves the
minimax rate of order n− min{2/(d+2),1/d} in the empirical L2 loss, up to poly-logarithmic
factors. Further, we prove a sharp oracle inequality, which reveals in particular that
when the true regression function is piecewise constant on k hyperrectangles, the least
squares estimator enjoys a faster, adaptive rate of convergence of (k/n)min(1,2/d) , again
up to poly-logarithmic factors. Previous results are confined to the case d ≤ 2. Finally, we establish corresponding bounds (which are new even in the case d = 2) in the
more challenging random design setting. There are two surprising features of these results: first, they demonstrate that it is possible for a global empirical risk minimisation
procedure to be rate optimal up to poly-logarithmic factors even when the corresponding entropy integral for the function class diverges rapidly; second, they indicate that
the adaptation rate for shape-constrained estimators can be strictly worse than the
parametric rate.
1
Introduction
Isotonic regression is perhaps the simplest form of shape-constrained estimation problem,
and has wide applications in a number of fields. For instance, in medicine, the expression
of a leukaemia antigen has been modelled as a monotone function of white blood cell count
and DNA index (Schell and Singh, 1997), while in education, isotonic regression has been
used to investigate the dependence of college grade point average on high school ranking
and standardised test results (Dykstra and Robertson, 1982). It is often generally accepted
that genetic effects on phenotypes such as height, fitness or disease are monotone (Mani et
al., 2007; Roth, Lipshitz and Andrews, 2009; Luss, Rosset and Shahar, 2012), but additive
∗
University
University
‡
University
§
University
†
of
of
of
of
Washington. Email: [email protected]
Cambridge. Email: [email protected]
Chicago and University of Illinois at Urbana-Champaign. Email: [email protected]
Cambridge. Email: [email protected]
1
structures have been found to be inadequate in several instances (Shao et al., 2008; Goldstein,
2009; Eichler et al., 2010). Alternative simplifying interaction structures have also been
considered, including those based on products (Elena and Lenski, 1997), logarithms (Sanjuan
and Elena, 2006) and minima (Tong et al., 2001), but the form of genetic interaction between
factors is not always clear and may vary between phenotypes (Luss, Rosset and Shahar,
2012).
A simple class of isotonic functions, which includes all of the above structures as special
cases, is the class of block increasing functions
Fd := f : [0, 1]d → R, f (x1 , . . . , xd ) ≤ f (x′1 , . . . , x′d ) when xj ≤ x′j for j = 1, . . . , d .
In this paper, we suppose that we observe data (X1 , Y1 ), . . . , (Xn , Yn ), with n ≥ 2, satisfying
Yi = f0 (Xi ) + ǫi ,
i = 1, . . . , n,
(1)
where f0 : [0, 1]d → R is Borel measurable, ǫ1 , . . . , ǫn are independent N(0, 1) noise, and the
covariates X1 , . . . , Xn , which take values in the set [0, 1]d , can either be fixed or random.
Our goal is P
to study the performance of the least squares isotonic regression estimator fˆn ∈
argminf ∈Fd ni=1 {Yi − f (Xi )}2 in terms of its empirical risk
X
n
1
2
{fˆn (Xi ) − f0 (Xi )} .
R(fˆn , f0 ) := E
n i=1
(2)
Note that this loss function only considers the errors made at the design points X1 , . . . , Xn ,
and these design points naturally induce a directed acyclic graph GX = (V (GX ), E(GX ))
with V (GX ) = {1, . . . , n} and E(GX ) = {(i, i′ ) : (Xi )j ≤ (Xi′ )j ∀ j = 1, . . . , d}. It is
therefore natural to restate the problem in terms of isotonic vector estimation on directed
acyclic graphs. Recall that given a directed acyclic graph G = (V (G), E(G)), we may define
a partially ordered set (V (G), ≤), where u ≤ v if and only if there exists a directed path
from u to v. We define the class of isotonic vectors on G by
M(G) := {θ ∈ RV (G) : θu ≤ θv for all u ≤ v}.
Hence, for a signal vector θ0 = ((θ0 )i )ni=1 := (f0 (Xi ))ni=1 ∈ M(GX ), the least squares estimator θ̂n = ((θ̂n )i )ni=1 := (fˆn (Xi ))ni=1 can be seen as the projection of (Yi )ni=1 onto the polyhedral
convex cone M(GX ). Such a geometric interpretation means that least squares estimators
for isotonic regression, in general dimensions or on generic directed acyclic graphs, can be
efficiently computed using convex optimisation algorithms (see, e.g., Dykstra (1983); Kyng,
Rao and Sachdeva (2015); Stout (2015)).
In the special case where d = 1, model (1) reduces to the univariate isotonic regression
problem that has a long history (e.g. Brunk, 1955; van Eeden, 1958; Barlow et al., 1972; van
de Geer, 1990, 1993; Donoho, 1991; Birgé and Massart, 1993; Meyer and Woodroofe, 2000;
Durot, 2007, 2008; Yang and Barber, 2017). See Groeneboom and Jongbloed (2014) for a
general introduction. Since the risk only depends on the ordering of the design points in the
univariate case, fixed and random designs are equivalent for d = 1 under the empirical risk
2
function (2). It is customary to write R(θ̂n , θ0 ) in place of R(fˆn , f0 ) for model (1) with fixed
design points. When (θ0 )1 ≤ · · · ≤ (θ0 )n (i.e. X1 ≤ · · · ≤ Xn ), Zhang (2002) proved that
there exists a universal constant C > 0 such that
2/3
log n
(θ0 )n − (θ0 )1
,
(3)
+
R(θ̂n , θ0 ) ≤ C
n
n
which shows in particular that the risk of the least squares estimator is no worse than
O(n−2/3 ) for signals θ0 of bounded uniform norm. In recent years, there has been considerable interest and progress in studying the automatic rate-adaptation phenomenon of
shape-constrained estimators. This line of study was pioneered by Zhang (2002) in the context of univariate isotonic regression, followed by Chatterjee, Guntuboyina and Sen (2015)
and most recently Bellec (2017), who proved that
kθ − θ0 k22 k(θ)
en
R(θ̂n , θ0 ) ≤ inf
,
(4)
+
log
θ∈M(GX )
n
n
k(θ)
where k(θ) is the number of constant pieces in the isotonic vector θ. The inequality (4)
is often called a sharp oracle inequality, with the sharpness referring to the fact that the
approximation error term n−1 kθ0 − θk22 has leading constant 1. The bound (4) shows nearly
parametric adaptation of the least squares estimator in univariate isotonic regression when
the underlying signal has a bounded number of constant pieces. Other examples of adaptation in univariate shape-constrained problems include the maximum likelihood estimator
of a log-concave density (Kim, Guntuboyina and Samworth, 2017), and the least squares
estimator in unimodal regression (Chatterjee and Lafferty, 2017).
Much less is known about the rate of convergence of the least squares estimator in
the model (1), or indeed the adaptation phenomenon in shape-restricted problems more
generally, in multivariate settings. The only work of which we are aware in the isotonic
regression case is Chatterjee, Guntuboyina and Sen (2017), which deals with the fixed,
lattice design case when d =Q
2. For a general dimension d, and for n1 , . . . , nd ∈ N, we define
this lattice by Ld,n1 ,...,nd := dj=1 {1, . . . , nj }; when n1 = . . . = nd = n1/d for some n ∈ N,
we also write Ld,n := Ld,n1 ,...,nd as shorthand. When {X1 , . . . , Xn } = L2,n1 ,n2 , Chatterjee,
Guntuboyina and Sen (2017) showed that there exists a universal constant C > 0 such that
((θ0 )n1 ,n2 − (θ0 )1,1 ) log4 n log8 n
;
(5)
+
R(θ̂n , θ0 ) ≤ C
n1/2
n
with a corresponding minimax lower bound of order n−1/2 . They also provided a sharp oracle
inequality of the form
kθ − θ0 k22 Ck(θ) log8 n
,
(6)
+
R(θ̂n , θ0 ) ≤
inf
θ∈M(L2,n1 ,n2 )
n
n
where k(θ) is the minimal number of rectangular blocks into which L2,n1 ,n2 may be partitioned
such that θ0 is constant on each rectangular block.
3
A separate line of work has generalised the univariate isotonic regression problem to
multivariate settings by assuming an additive structure (see e.g. Bacchetti (1989); MortonJones et al. (2000); Mammen and Yu (2007); Chen and Samworth (2016)). In the simplest
setting, these works investigate the regression problem (1), where the signal f0 belongs to
Fdadd
:=
f ∈ Fd : f (x1 , . . . , xd ) =
d
X
j=1
fj (xj ), fj ∈ F1 , kfj k∞
≤1 .
The additive structure greatly reduces the complexity of the class; indeed, it can be shown
that the least squares estimator over Fdadd attains the univariate risk n−2/3 , up to multiplicative constants depending on d (e.g. van de Geer, 2000, Theorem 9.1).
The main contribution of this paper is to provide risk bounds for the isotonic least squares
estimator when d ≥ 3, both from a worst-case perspective and an adaptation point of view.
Specifically, we show that in the fixed lattice design case, the least squares estimator satisfies
sup
θ0 ∈M(Ld,n ),kθ0 k∞ ≤1
R(θ̂n , θ0 ) ≤ Cn−1/d log4 n,
(7)
for some universal constant C > 0. This rate turns out to be the minimax risk up to polylogarithmic factors in this problem. Furthermore, we establish a sharp oracle inequality:
there exists a universal constant C > 0 such that for every θ0 ∈ RLd,n ,
R(θ̂n , θ0 ) ≤
inf
θ∈M(Ld,n )
2/d
en
kθ − θ0 k22
k(θ)
8
,
+C
log
n
n
k(θ)
(8)
where k(θ) is the number of constant hyperrectangular pieces in θ. This reveals an adaptation
rate of nearly (k/n)2/d for signals that are close to an element of M(Ld,n ) that has at most
k hyperrectangular blocks. A corresponding lower bound is also provided, showing that
the least squares estimator cannot adapt faster than the n−2/d rate implied by (8) even
for constant signal vectors. We further demonstrate that the worst-case bounds and oracle
inequalities (7) and (8), with slightly different poly-logarithmic exponents, remain valid for
random design points X1 , . . . , Xn sampled independently from a distribution on [0, 1]d with
a Lebesgue density bounded away from 0 and ∞. The results in the case of random design
are novel even for dimension d = 2. These results are surprising in particular with regard to
the following two aspects:
1. The negative results of Birgé and Massart (1993) have spawned a heuristic belief that
one should not use global empirical risk minimisation procedures1 when the entropy
integral for the corresponding function class diverges (e.g. van de Geer (2000, p. 121–
122), Rakhlin, Sridharan and Tsybakov (2017)). It is therefore of particular interest to
see that in our isotonic regression function setting, the global least squares estimator
is still rate optimal (up to poly-logarithmic factors). See also the discussion after
Corollary 1.
1
The term ‘global’ refers here to procedures that involve minimisation over the entire function class, as
opposed to only over a sieve; cf. van de Geer (2000).
4
2. Sharp adaptive behaviour for shape-constrained estimators has previously only been
shown when the adaptive rate is nearly parametric (see, e.g., Guntuboyina and Sen
(2015); Chatterjee, Guntuboyina and Sen (2015); Bellec (2017); Kim, Guntuboyina and
Samworth (2017)). On the other hand, our results here show that the least squares
estimator in the d-dimensional isotonic regression problem necessarily adapts at a
strictly nonparametric rate. Clearly, the minimax optimal rate for constant functions
is parametric. Hence, the least squares estimator in this problem adapts at a strictly
suboptimal rate while at the same time being nearly rate optimal from a worst-case
perspective.
In both the fixed lattice design and the more challenging random design cases, our analyses
are based on a novel combination of techniques from empirical process theory, convex geometry and combinatorics. We hope these methods can serve as a useful starting point towards
understanding the behaviour of estimators in other multivariate shape-restricted models.
The rest of the paper is organised as follows. In Section 2, we state the main results
for the fixed lattice design model. Section 3 describes corresponding results in the random
design case. Proofs of all main theoretical results are contained in Sections 4 and 5, whereas
proofs of ancillary results are deferred until Section 6.
1.1
Notation
For a real-valued measurable function f defined on a probability space (X , A, P ) and for
p ∈ [1, ∞), we let kf kLp (P ) := P |f |p)1/p denote the usual Lp (P )-norm, and write kf k∞ :=
supx∈X |f (x)|. For r ≥ 0, we write Bp (r, P ) := {f : X → R, kf kLp (P ) ≤ r} and B∞ (r) :=
{f : X → R, kf k∞ ≤ r}. We will abuse notation slightly and also write Bp (r) := {v ∈
Rn : kvkp ≤ r} for p ∈ [1, ∞]. The Euclidean inner product on Rd is denoted by h·, ·i. For
x, y ∈ Rd , we write x y if xj ≤ yj for all j = 1, . . . , d.
For ε > 0, the ε-covering number of a (semi-)normed space (F , k·k), denoted N ε, F , k·k ,
is the smallest number of closed ε-balls whose union covers F . The ε-bracketing number,
denoted N[ ] (ε, F , k · k), is the smallest number of ε-brackets, of the form [l, u] := {f ∈ F :
l ≤ f ≤ u} such that ku − lk ≤ ε, and whose union covers F . The metric/bracketing entropy
is the logarithm of the covering/bracketing number.
Throughout the article ǫ1 , . . . , ǫn and {ǫw : w ∈ Ld,n1 ,...,nd } denote independent standard
normal random variables and ξ1 , . . . , ξn denote independent Rademacher random variables,
both independent of all other random variables. For two probability measures P and Q
defined on the same measurable space (X , A), we write
(P, Q) := supA∈A |P (A) − Q(A)|
R dTVdP
for their total variation distance, and d2KL (P, Q) := X log dQ
dP for their Kullback–Leibler
divergence.
We use c, C to denote generic universal positive constants and use cx , Cx to denote generic
positive constants that depend only on x. Exact numeric values of these constants may
change from line to line unless otherwise specified. Also, a .x b and a &x b mean a ≤ Cx b
and a ≥ cx b respectively, and a ≍x b means a .x b and a &x b (a . b means a ≤ Cb for
some absolute constant C). We also define log+ (x) := log(x ∨ e).
5
2
Fixed lattice design
In this section, we focus on the model (1) in the case where the set of design points forms
a finite cubic lattice Ld,n , defined in the introduction. In particular, we will assume in this
section that n = nd1 for some n1 ∈ N. We use the same notation Ld,n both for the set of
points and the directed acyclic graph on these points with edge structure arising from the
natural partial ordering induced by . Thus, in the case d = 1, the graph L1,n is simply a
directed path, and this is the classical univariate isotonic regression setting. The case d = 2
is studied in detail in Chatterjee, Guntuboyina and Sen (2017). Our main interest lies in
the cases d ≥ 3.
2.1
Minimax rate-optimality of least squares estimator
Our first result provides an upper bound on the risk of the least squares estimator θ̂n =
θ̂n (Y1 , . . . , Yn ) of θ0 ∈ M(Ld,n ).
Theorem 1. Let d ≥ 2. There exists a universal constant C > 0 such that
sup
θ0 ∈M(Ld,n )∩B∞ (1)
R(θ̂n , θ0 ) ≤ Cn−1/d log4 n.
Theorem 1 reveals that, up to a poly-logarithmic factor, the empirical risk of the least
squares estimator converges to zero at rate n−1/d . The upper bound in Theorem 1 is matched,
up to poly-logarithmic factors, by the following minimax lower bound.
Proposition 1. There exists a constant cd > 0, depending only on d, such that for d ≥ 2,
inf
sup
θ̃n θ0 ∈M(Ld,n )∩B∞ (1)
R(θ̃n , θ0 ) ≥ cd n−1/d ,
where the infimum is taken over all estimators θ̃n = θ̃n (Y1 , . . . , Yn ) of θ0 .
From Theorem 1 and Proposition 1, together with existing results mentioned in the
introduction for the case d = 1, we see that the worst-case risk n− min{2/(d+2),1/d} (up to polylogarithmic factors) of the least squares estimator exhibits different rates of convergence in
dimension d = 1 and dimensions d ≥ 3, with d = 2 being a transitional case. From the proof
of Proposition 1, we see that it is the competition between the cardinality of the maximum
chain (totally ordered subset) and the maximum antichain (subset of mutually incomparable
design points) that explains the different rates. Similar transitional behaviour was recently
observed by Kim and Samworth (2016) in the context of log-concave density estimation,
though there it is the tension between estimating the density in the interior of its support
and estimating the support itself that drives the transition.
The two results above can readily be translated into bounds for the rate of convergence
for estimation of a block monotonic function with a fixed lattice design. Recall that Fd
is the class of block increasing functions. Suppose that for some f0 ∈ Fd , and at each
1/d
x = (x1 , . . . , xd ) ∈ P
n−1
, we observe Y (x) ∼ N(f0 (x), 1) independently.
1 Ld,n , where n1 = n
−1
Define Pn := n
x∈n−1 Ld,n δx and let A denote the set of hypercubes of the form A =
1
6
i −1
Qd
i
where either Aj = [0, n11 ] or Aj = ( jn1 , nj1 ] for some ij ∈ {2, . . . , n1 }. Now let H
denote the set of functions f ∈ Fd that are piecewise constant on each A ∈ A, and set
j=1 Aj ,
n
1X
{Y (xi ) − f (xi )}2 .
fˆn := argmin
n i=1
f ∈H
The following is a fairly straightforward corollary of Theorem 1 and Proposition 1.
Corollary 1. There exist constants cd , Cd > 0, depending only on d, such that for Q = Pn
or Lebesgue measure on [0, 1]d , we have
cd n−1/d ≤ inf
sup
f˜n f0 ∈Fd ∩B∞ (1)
Ekf˜n − f0 k2L2 (Q) ≤
sup
f0 ∈Fd ∩B∞ (1)
Ekfˆn − f0 k2L2 (Q) ≤ Cd n−1/d log4 n,
where the infimum is taken over all measurable functions of {Y (x) : x ∈ n−1
1 Ld,n }.
This corollary is surprising in the following sense. Gao and Wellner (2007, Theorem 1.1)
proved that for d ≥ 3,
log N ε, Fd ∩ B∞ (1), k · k2 ≍d ε−2(d−1) .
(9)
In particular,
for d ≥ 3, the classes Fd ∩ B∞ (1) are massive in the sense that the entropy
R 1 1/2
integral δ log N(ε, Fd ∩ B∞ (1), k · k2) dε diverges at a polynomial rate in δ −1 as δ ց 0. To
the best of our knowledge, this is the first example of a setting where a global empirical risk
minimisation procedure has been proved to attain (nearly) the minimax rate of convergence
over such massive parameter spaces.
2.2
Sharp oracle inequality
In this subsection, we consider the adaptation behaviour of the least squares estimator in
dimensions d ≥ 2 (again, the d = 2 case is covered in Chatterjee, Guntuboyina and Sen
(2017)). Our main result is the sharp oracle Q
inequality in Theorem 2 below. We call a
set in Rd a hyperrectangle if it is of the form dj=1 Ij where Ij ⊆ R is an interval for each
j = 1, . . . , d. By a slight abuse of terminology, we also call a subset of Ld,n a hyperrectangle
if it is the intersection of a hyperrectangle in [0, 1]d and Ld,n . We say a subset A of Ld,n is a
Q
two-dimensional sheet if A = dj=1 [aj , bj ] where |{j : bj = aj }| ≥ d − 2. A two-dimensional
sheet is therefore a special type of hyperrectangle whose intrinsic dimension is at most two.
For θ ∈ M(Ld,n ), let K(θ) denote the cardinality of the minimal partition Ld,n = ⊔K
ℓ=1 Aℓ of
Ld,n into a disjoint union of two-dimensional sheets A1 , . . . , AK , where the restricted vector
θAℓ = (θ(u))u∈Aℓ is constant for each ℓ = 1, . . . , K.
Theorem 2. Let d ≥ 2. There exists a universal constant C > 0 such that for every
θ0 ∈ RLd,n ,
n
kθ − θ0 k22 CK(θ)
8
R(θ̂n , θ0 ) ≤ inf
.
+
log+
θ∈M(Ld,n )
n
n
K(θ)
7
We remark that Theorem 2 does not imply (nearly) parametric adaptation when d ≥ 3.
This is because even when θ0 is constant on Ld,n for every n, we have K(θ0 ) = n(d−2)/d → ∞
as n → ∞. The following corollary of Theorem 2 gives an alternative (weaker) form of
oracle inequality that offers easier comparison to lower dimensional results given in (4)
and (6). Let M(k) (Ld,n ) be the collection of all θ ∈ M(Ld,n ) such that there exists a
partition Ld,n = ⊔kℓ=1 Rℓ where R1 , . . . , Rk are hyperrectangles with the property that for
each ℓ, the restricted vector θRℓ is constant.
Theorem 3. Let d ≥ 2. There exists a universal constant C > 0 such that for every
θ0 ∈ RLd,n ,
2/d
kθ − θ0 k22
k
n
8
inf
.
R(θ̂n , θ0 ) ≤ inf
+C
log+
k∈N θ∈M(k) (Ld,n )
n
n
k
It is important to note that both Theorems 2 and 3 allow for model misspecification, as
it is not assumed that θ0 ∈ M(Ld,n ). For signal vectors θ0 that are piecewise constant on
k hyperrectangles, Theorem 3 provides an upper bound of the risk of order (k/n)2/d up to
poly-logarithmic factors. The following proposition shows that even for a constant signal
vector, the adaptation rate of n−2/d given in Theorem 3 cannot be improved.
Proposition 2. Let d ≥ 2. There exists a constant cd > 0, depending only on d, such that
for any θ0 ∈ M(1) (Ld,n ),
(
n−1 log2 n if d = 2
R(θ̂n , θ0 ) ≥ cd
n−2/d
if d ≥ 3.
The case d = 2 of this result is new, and reveals both a difference with the univariate situation, where the adaptation rate is of order n−1 log n (Bellec, 2017), and that a
poly-logarithmic penalty relative to the parametric rate is unavoidable for the least squares
estimator. Moreover, we see from Proposition 2 that for d ≥ 3, although the least squares
estimator achieves a faster rate of convergence than the worst-case bound in Theorem 1
on constant signal vectors, the rate is not parametric, as would have been the case for a
minimax optimal estimator over the set of constant vectors. This is in stark contrast to the
nearly parametric adaptation results established in (4) and (6) for dimensions d ≤ 2.
Another interesting aspect of these results relates
to the notion of statistical dimension,
R
2
defined for an arbitrary cone C in Rn by2 δ(C) := Rn kΠC (x)k22 (2π)−n/2 e−kxk2 /2 dx, where ΠC
is the projection onto the set C (Amelunxen et al., 2014). Theorem 3 and Proposition 2 reveal
a type of phase transition phenomenon for the statistical dimension δ(M(Ld,n )) = R(θ̂n , 0)
of the monotone cone (cf. Table 1).
The following corollary of Theorem 2 gives another example where different adaptation
behaviour is observed in dimensions d ≥ 3, in the sense that the n−2/d log8 n adaptive rate
achieved for constant signal vectors is actually available for a much wider class of isotonic
signals that depend only on d − 2 of all d coordinates of Ld,n . For r = 0, 1, . . . , d, we say
2
Our reason for defining the statistical dimension via an integral rather than as EkΠC (ǫ)k22 is because, in
the random design setting, the cone C is itself random, and in that case δ(C) is a random quantity.
8
d
1
Table 1: Bounds∗ for δ M(Ld,n ) .
2
≥3
∗
†
‡
upper bound
Pn −1 †
i=1 i
lower bound
Pn −1 †
i=1 i
. n1−2/d log8 n
&d n1−2/d
. log8 n
‡
& log2 n
Entries without a reference are proved in this
paper.
Amelunxen et al. (2014)
Chatterjee, Guntuboyina and Sen (2017)
a vector θ0 ∈ M(Ld,n ) is a function of r variables, written θ0 ∈ Mr (Ld,n ), if there exists
J ⊆ {1, . . . , d}, of cardinality r, such that (θ0 )(x1 ,...,xd ) = (θ0 )(x′1 ,...,x′d ) whenever xj = x′j for
all j ∈ J .
Corollary 2. For d ≥ 2, there exists constant Cd > 0, depending
−2/d
log8 n
n
sup
R(θ̂n , θ0 ) ≤ Cd n−4/(3d) log16/3 n
θ0 ∈Mr (Ld,n )∩B∞ (1)
−1/d 4
n
log n
only on d, such that
if r ≤ d − 2
if r = d − 1
if r = d.
If the signal vector θ0 belongs to Mr (Ld,n ), then it is intrinsically an r-dimensional
isotonic signal. Corollary 2 demonstrates that the least squares estimator exhibits three
different levels of adaptation when the signal is a function of d, d − 1, d − 2 variables respectively. However, viewed together with Proposition 1, Corollary 2 shows that no further
adaptation is available when the intrinsic dimension of the signal vector decreases further.
Moreover, if we let ñ = n2/d denote the size of a maximal two-dimensional sheet in Ld,n , then
the three levels of adaptive rates in Corollary 2 are ñ−1 , ñ−2/3 and ñ−1/2 respectively, up
to poly-logarithmic factors, matching the two-dimensional ‘automatic variable adaptation’
result described in Chatterjee, Guntuboyina and Sen (2017, Theorem 2.4). In this sense,
the adaptation of the isotonic least squares estimator in general dimensions is essentially a
two-dimensional phenomenon.
3
Random design
In this section, we consider the setting where the design points X1 , . . . , Xn are independent
and identically distributed from some distribution P supported on the unit cube [0, 1]d . We
will assume throughout that P has Lebesgue density p0 such that 0 < m0 ≤ inf x∈[0,1]d p0 (x) ≤
supx∈[0,1]d p0 (x) ≤ M0 < ∞. Since the least squares estimator fˆn is only well-defined on
X1 , . . . , Xn , for definiteness, we extend fˆn to [0, 1]d by defining fˆn (x) := min {fˆn (Xi ) : 1 ≤
P
i ≤ n, Xi x} ∪ {maxi fˆn (Xi )} . If we let Pn := n−1 ni=1 δXi , then the risk function (2) is
R(fˆn , f0 ) = Ekfˆn − f0 k2L2 (Pn ) in the context of random design.
9
The main results of this section are the following two theorems, establishing respectively
the worst-case performance and the sharp oracle inequality for the least squares estimator in
(k)
the random design setting. We write Fd for the class of functions in Fd that are piecewise
(k)
constant on k hyperrectangular pieces. In other words, if f ∈ Fd , then there exists a
partition [0, 1]d = ⊔kℓ=1 Rℓ , such that the closure of each Rℓ is a hyperrectangle and f is a
constant function when restricted to each Rℓ . Let γ2 := 9/2 and γd := (d2 + d + 1)/2 for
d ≥ 3.
Theorem 4. Let d ≥ 2. There exists a constant Cd,m0 ,M0 > 0, depending only on d, m0 and
M0 , such that
sup
R(fˆn , f0 ) ≤ Cd,m0 ,M0 n−1/d logγd n.
f0 ∈Fd ∩B∞ (1)
Theorem 5. Let d ≥ 2. There exists a constant Cd,m0 ,M0 > 0, depending only on d, m0 and
M0 , such that for any measurable function f0 : [0, 1]d → R, we have
2/d
k
2γd n
2
ˆ
R(fn , f0 ) ≤ inf inf kf − f0 kL2 (P ) + Cd,m0 ,M0
.
log+
k∈N f ∈F (k)
n
k
d
To the best of our knowledge, Theorem 5 is the first sharp oracle inequality in the shapeconstrained regression literature with random design. The different norms on the left- and
right-hand sides arise from the simple observation that Ekf − f0 k2L2 (Pn ) = kf − f0 k2L2 (P ) for
(k)
f ∈ Fd . The proofs of Theorems 4 and 5 are considerably more involved than those of
the corresponding Theorems 1 and 2 in Section 2. We briefly mention two major technical
difficulties:
1. The size of Fd , as measured by its entropy, is large when d ≥ 3, even after L∞
truncation (cf. (9)). As rates obtained from the entropy integral (e.g. van de Geer,
2000, Theorem 9.1) do not match those from Sudakov lower bounds for such classes,
standard entropy methods result in a non-trivial gap between the minimax rates of
convergence, which typically match the Sudakov lower bounds (e.g. Yang and Barron,
1999, Proposition 1), and provable risk upper bounds for least squares estimators when
d ≥ 3.
2. In the fixed lattice design case, our analysis circumvents the difficulties of standard
entropy methods by using the fact that a d-dimensional cubic lattice can be decomposed
into a union of lower-dimensional pieces. This crucial property is no longer valid when
the design is random.
We do not claim any optimality of the power in the poly-logarithmic factor in the oracle
inequality in Theorems 4 and 5. On the other hand, similar to the fixed, lattice design case,
the worst-case rate n−1/d and adaptation rate n−2/d cannot be improved, as can be seen from
the following two propositions.
Proposition 3. Let d ≥ 2. There exists a constant cd,m0 ,M0 > 0, depending only on d, m0
and M0 , such that,
inf
sup
R(f˜n , f0 ) ≥ cd,m0 ,M0 n−1/d ,
f˜n f0 ∈Fd ∩B∞ (1)
10
where the infimum is taken over all measurable functions f˜n of the data (X1 , Y1), . . . , (Xn , Yn ).
Proposition 4. Let d ≥ 2. There exists a constant cd,M0 > 0, depending only on d and M0 ,
(1)
such that for any f0 ∈ Fd ,
R(fˆn , f0 ) ≥ cd,M0 n−2/d .
A key step in proving Proposition 4 is to establish that with high probability, the cardinality of the maximum antichain in GX is at least of order n1−1/d . When d = 2, the
distribution of this maximum cardinality is the same as the distribution of the length of the
longest decreasing subsequence of a uniform permutation of {1, . . . , n}, a famous object of
study in probability and combinatorics. See Romik (2014) and references therein.
4
Proofs of results in Section 2
Throughout this section, ǫ = (ǫw )w∈Ld,n1 ,...,nd denotes a vector of independent standard normal
random variables. It is now well understood that the risk of the least squares estimator in
the Gaussian sequence model is completely characterised by the size of a localised Gaussian
process; cf. Chatterjee (2014). The additional cone property of M(Ld,n ) makes the reduction
even simpler: we only need to evaluate the Gaussian complexity of M(Ld,n ) ∩ B2 (1), where
the Gaussian complexity of T ⊆ RLd,n1 ,...,nd is defined as wT := E supθ∈T hǫ, θi. Thus the
result in the following proposition constitutes a key ingredient in analysing the risk of the
least squares estimator.
Proposition 5. There Q
exists a universal constant C > 0 such that for d ≥ 2 and every
1 ≤ n1 ≤ · · · ≤ nd with dj=1 nj = n, we have
p
2/π
hǫ, θi ≤ C
nd−1 n−1/2 ≤ E
sup
(d − 1)d−1 1
θ∈M(Ld,n1 ,...,nd )∩B2 (1)
r
−1/2
Remark. In the case n1 = · · · = nd = n1/d , we have nd−1
=
1 n
n
nd−1 nd
q
n
nd−1 nd
log4 n.
= n1/2−1/d .
Remark. From the symmetry of the problem, we see that the restriction that n1 ≤ · · · ≤ nd
is not essential. In the general case, for the lower bound, n1 should be replaced with minj nj ,
while in the upper bound, nd−1 nd should be replaced with the product of the two largest
elements of {n1 , . . . , nd } (considered here as a multiset).
P
Proof. We first prove the lower bound. Consider the set W := {w ∈ Ld,n1 ,...,nd : dj=1 wj =
Pd
−
:= {w ∈ Ld,n1 ,...,nd :
n1 }, and define W + := {w ∈ Ld,n1 ,...,nd :
j=1 wj > n1 } and W
Pd
j=1 wj < n1 }. For each realisation of the Gaussian random vector ǫ = (ǫw )w∈Ld,n1 ,...,nd , we
define θ(ǫ) = (θw (ǫ))w∈Ld,n1 ,...,nd ∈ M(Ld,n1 ,...,nd ) by
if w ∈ W +
1
θw := sgn(ǫw ) if w ∈ W
−1
if w ∈ W − .
11
Since kθ(ǫ)k22 = n, it follows that
X
X
X
θ(ǫ)
1
ǫw +
|ǫw |
hǫ, θi ≥ E ǫ,
E
sup
ǫw −
= 1/2 E
kθ(ǫ)k2
n
θ∈M(Ld,n1 ,...,nd )∩B2 (1)
w∈W
w∈W −
w∈W +
p
2/π
= 1/2 |W |.
n
The proof of the lower bound is now completed by noting that
d−1
n1
d + n1 − 1
≥
|W | =
.
(10)
d−1
d−1
We next prove the upper bound. For j = 1, . . . , d − 2 and xj ∈ {1, . . . , nj }, we define
Ax1 ,...,xd−2 := {w = (w1 , . . . , wd )⊤ ∈ Ld,n1 ,...,nd : (w1 , . . . , wd−2 ) = (x1 , . . . , xd−2 )}. Each
Ax1 ,...,xd−2 can be viewed as a directed acyclic graph with graph structure inherited from
Ld,n1 ,...,nd . Since L
monotonicity is preserved under the subgraph restriction, we have that
M(Ld,n1 ,...,nd ) ⊆
x1 ,...,xd−2 M(Ax1 ,...,xd−2 ). Therefore, by the Cauchy–Schwarz inequality,
Amelunxen et al. (2014, Proposition 3.1(5, 9, 10)) and Chatterjee, Guntuboyina and Sen
(2017, Theorem 2.1), we obtain that
2
2
hǫ, θi ≤ E
E
sup
hǫ, θi
sup
θ∈M(Ld,n1 ,...,nd )∩B2 (1)
θ∈M(Ld,n1 ,...,nd )∩B2 (1)
= δ M(Ld,n1 ,...,nd ) ≤
= δ M(L2,nd−1 ,nd )
as desired.
X
x1 ,...,xd−2
d−2
Y
nj .
j=1
δ M(Ax1 ,...,xd−2 )
n
nd−1 nd
log8 (end−1 nd ),
Proof of Theorem 1. Fix θ0 ∈ M(Ld,n ) ∩ B∞ (1). By Chatterjee (2014, Theorem 1.1), the
function
t 7→ E
sup
hǫ, θ − θ0 i − t2 /2
θ∈M(Ld,n ),kθ−θ0 k≤t
is strictly concave on [0, ∞) with a unique maximum at, say, t0 ≥ 0. We note that t0 ≤ t∗
for any t∗ satisfying
t2
hǫ, θ − θ0 i ≤ ∗ .
E
sup
(11)
2
θ∈M(Ld,n ),kθ−θ0 k≤t∗
P
For a vector θ = (θx )x∈Ld,n , we define θ̄ := n−1 x∈Ld,n θx and write 1n ∈ RLd,n for the all-one
vector. Then
n
o
hǫ, θ − θ̄0 1n i + hǫ, θ̄0 1n − θ0 i
hǫ, θ − θ0 i = E
sup
E
sup
θ∈M(Ld,n ),kθ−θ0 k2 ≤t∗
θ∈M(Ld,n ),kθ−θ0 k2 ≤t∗
≤E
=E
sup
θ∈M(Ld,n ),kθ−θ̄0 1n k2 ≤t∗ +n1/2
sup
θ∈M(Ld,n )∩B2 (t∗ +n1/2 )
12
hǫ, θ − θ̄0 1n i
hǫ, θi = t∗ + n1/2 wM(Ld,n )∩B2 (1) ,
where we recall that wM(Ld,n )∩B2 (1) = E supθ∈M(Ld,n )∩B2 (1) hǫ, θi. Therefore, to satisfy (11), it
suffices to choose
2
1/2
t∗ = wM(Ld,n )∩B2 (1) + wM(L
+ 2n1/2 wM(Ld,n )∩B2 (1)
d,n )∩B2 (1)
1/2
. max wM(Ld,n )∩B2 (1) , n1/4 wM(Ld,n )∩B2 (1) .
(12)
Consequently, by Chatterjee (2014, Corollary 1.2) and Proposition 5, we have that
R(θ̂n , θ0 ) . n−1 max(1, t20 ) . n−1 t2∗ . n−1/d log4 n,
which completes the proof.
The following proposition is the main ingredient of the proof of the minimax lower bound
in Proposition 1. It exhibits a combinatorial obstacle, namely the existence of a large antichain, that prevents any estimator from achieving a faster rate of convergence. We state
the result in the more general and natural setting of least squares isotonic regression on
directed acyclic graphs. Recall that the isotonic regression problem on a directed acyclic
graph G = (V (G), E(G)) is of the form Yv = θv + ǫv , where θ = (θv )v∈V (G) ∈ M(G) and
ǫ = (ǫv )v∈V (G) is a vector of independent N(0, 1) random variables.
Proposition 6. If G = (V (G), E(G)) is a directed acyclic graph and W ⊆ V (G) is a
maximum antichain of G, then
inf
sup
θ̃n θ0 ∈M(G)∩B∞ (1)
R(θ̃n , θ0 ) ≥
8|W |
,
27n
where the infimum is taken over all measurable functions θ̃n of {Yv : v ∈ V (G)}.
Proof. If v ∈
/ W , then by the maximality of W , there exists u0 ∈ W such that either u0 ≤ v
or u0 ≥ v. Suppose without loss of generality it is the former. Then v 6≤ u for any u ∈ W ,
because otherwise we would have u0 ≤ u, contradicting the fact that W is an antichain. It
follows that we can write V (G) = W + ⊔ W ⊔ W − , where for all v ∈ W + , u ∈ W , we have
u 6≥ v, and similarly for all v ∈ W − , u ∈ W , we have v 6≥ u.
For τ = (τw ) ∈ {0, 1}W =: T , we define θτ = (θvτ ) ∈ M(G) ∩ B∞ (1) by
if v ∈ W −
−1
θvτ = ρ(2τv − 1) if v ∈ W
1
if v ∈ W + ,
where ρ ∈ (0, 1) is a constant to be chosen later. Let Pτ denote the distribution of {Yv :
v ∈ V (G)} when the isotonic signal is θτ . Then, for τ, τ ′ ∈ T , by Pinsker’s inequality (e.g.
Pollard, 2002, p. 62), we have
1
n
′
d2TV (Pτ , Pτ ′ ) ≤ d2KL (Pτ , Pτ ′ ) = kθτ − θτ k22 = nρ2 kτ − τ ′ k0 .
2
4
13
Consequently, setting ρ = 2/(3n1/2 ), by Assouad’s Lemma (cf. Yu, 1997, Lemma 2), we have
that
inf
θ̃n
8|W |
1
,
R(θ̃n , θ0 ) ≥ inf sup Ekθ̃n − θτ k22 ≥ 2ρ2 |W |(1 − n1/2 ρ) =
27n
θ̃n τ ∈T n
θ0 ∈M(G)∩B∞ (1)
sup
as desired.
Proof of Proposition 1. Recall that n1 = n1/d . We note that the set
W :=
⊤
v = (v1 , . . . , vd ) ∈ Ld,n :
d
X
vj = n1
j=1
n1−1/d
1 −1
is an antichain in Ld,n of cardinality d+n
≥ (d−1)
d−1 . Hence any maximum antichain of
d−1
Ld,n is at least of this cardinality. The desired result therefore follows from Proposition 6.
Proof of Corollary 1. For Q = Pn , the result is an immediate consequence of Theorem 1 and
Proposition 1, together with the facts that
inf
sup
θ̃n θ0 ∈M(Ld,n )∩B∞ (1)
R(θ̃n , θ0 ) = inf
sup
f˜n f0 ∈Fd ∩B∞ (1)
and
sup
R(θ̂n , θ0 ) =
θ0 ∈M(Ld,n )∩B∞ (1)
sup
f0 ∈Fd ∩B∞ (1)
Ekf˜n − f0 k2L2 (Pn )
Ekfˆn − f0 k2L2 (Pn ) .
Now suppose that Q is Lebesgue measure on [0, 1]d . For any f : [0, 1]d → R, we may define
θ(f ) : Ld,n → R by θ(f )(x) := f (n−1
1 x). On the other hand, for any θ : Ld,n → R, we can
d
also define f (θ) : [0, 1] → R by
f (θ)(x1 , . . . , xd ) := θ(⌊n1 x1 ⌋, . . . , ⌊n1 xd ⌋).
We first prove the upper bound by observing from Lemma 1 and Theorem 1 that
−1
sup
Ekfˆn − f0 k2L2 (Q) ≤ 2
sup
n Ekθ(fˆn ) − θ(f0 )k22 + kf0 − f (θ(f0 ))k2L2 (Q)
f0 ∈Fd ∩B∞ (1)
f0 ∈Fd ∩B∞ (1)
≤2
1
Ekθ̂n − θ0 k22 + 8dn−1/d ≤ Cd n−1/d log4 n,
n
θ0 ∈M(Ld,n )∩B∞ (1)
sup
as desired. Then by convexity of H and Proposition 1, we have
inf
sup
f˜n f0 ∈Fd ∩B∞ (1)
Ekf˜n − f0 k2L2 (Q) ≥ inf
sup
Ekf˜n − f (θ0 )k2L2 (Q)
= inf
sup
Ekf (θ(f˜n )) − f (θ0 )k2L2 (Q)
f˜n θ0 ∈M(Ld,n )∩B∞ (1)
f˜n θ0 ∈M(Ld,n )∩B∞ (1)
= inf
θ̃n
1
Ekθ̃n − θ0 k22 ≥ cd n−1/d ,
θ0 ∈M(Ld,n )∩B∞ (1) n
sup
which completes the proof.
14
Proof of Theorem 2. Recall that the tangent cone at a point x in a closed, convex set K is
defined as T (x, K) := {t(y − x) : y ∈ K, t ≥ 0}. By Bellec (2017, Proposition 2.1) (see also
Chatterjee, Guntuboyina and Sen (2017, Lemma 4.1)), we have
n
o
1
inf
(13)
kθ − θ0 k22 + δ T (θ, M(Ld,n )) .
R(θ̂n , θ0 ) ≤
n θ∈M(Ld,n )
For a fixed θ ∈ M(Ld,n ) such that K(θ) = K, let Ld,n = ⊔K
ℓ=1 Aℓ be the partition of Ld,n into
two-dimensional sheets Aℓ such that θ is constant on each Aℓ . Define mℓ := |Aℓ |. Then any
u ∈ T (θ, M(Ld,n )) must be isotonic when restricted to each of the two-dimensional sheets;
in other words
K
M
T (θ, M(Ld,n )) ⊆
T (0, M(Aℓ)).
ℓ=1
By Amelunxen et al. (2014, Proposition 3.1(9, 10)), we have
X
M
K
K
K
X
δ M(Aℓ ) .
δ T (0, M(Aℓ)) =
T (0, M(Aℓ )) =
δ T (θ, M(Ld,n )) ≤ δ
ℓ=1
ℓ=1
ℓ=1
(14)
By a consequence of the Gaussian Poincaré inequality (cf. Boucheron, Lugosi and Massart,
2013, p. 73) and Proposition 5, we have
2
δ M(Aℓ ) ≤ E
sup
hǫAℓ , θi + 1 . log8+ mℓ .
(15)
θ∈M(Aℓ )∩B2 (1)
Thus, by (14), (15) and Lemma 2 applied to x 7→ log8+ x, we have
δ T (θ, M(Ld,n )) .
K
X
log8+
mℓ . K
ℓ=1
log8+
n
,
K
which together with (13) proves the desired result.
Proof of Theorem 3. For a fixed θ ∈ M(k) (Ld,n ), let Ld,n = ⊔kℓ=1 Rℓ be the partition of Ld,n
into hyperrectangles such that θQis constant on each hyperrectangle Rℓ . Suppose Rℓ has side
ℓ|
lengths m1 , . . . , md (so |Rℓ | = dj=1 mj ), so it can be partitioned into m|R
parallel twoj mj ′
dimensional sheets. By choosing mj and mj ′ to be the largest two elements of the multiset
{m1 , . . . , md } and using Jensen’s inequality (noting that x 7→ x1−2/d is concave when d ≥ 2),
we obtain
1−2/d
k
X
n
1−2/d
.
(16)
K(θ) ≤
|Rℓ |
≤k
k
ℓ=1
This, combined with the oracle inequality in Theorem 2, gives the desired result.
Proof of Proposition 2. Since the convex cone M(Ld,n ) is invariant under translation by any
θ0 ∈ M(1) (Ld,n ), we may assume without loss of generality that θ0 = 0. By Chatterjee (2014,
Corollary 1.2), we have
1
3/2
(17)
R(θ̂n , 0) ≥ (t20 − Ct0 ),
n
15
where
t0 : = argmax E
t≥0
sup
θ∈M(Ld,n )∩B2 (t)
= argmax t · E
t≥0
2
hǫ, θi − t /2
sup
θ∈M(Ld,n )∩B2 (1)
2
hǫ, θi − t /2
=E
sup
θ∈M(Ld,n )∩B2 (1)
hǫ, θi.
By Proposition 5, we have
t0 = E
sup
θ∈M(Ld,n )∩B2 (1)
hǫ, θi ≥ cd n1/2−1/d ,
which together with (17) proves the desired lower bound for cases d ≥ 3.
For the d = 2 case, by Sudakov minorisation for Gaussian processes (e.g. Pisier, 1999,
Theorem 5.6 and the remark following it) and Lemma 3, there exists a universal constant
ε0 > 0 such that
E
sup
hǫ, θi & ε0 log1/2 N ε0 , M(L2,n ) ∩ B2 (1), k · k2 & log n.
θ∈M(L2,n )∩B2 (1)
This, together with (17), establishes the desired conclusion when d = 2.
Proof of Corollary 2. Without loss of generality, we may assume that θ0 ∈ Mr (Ld,n ) is a
function of the final
r variables. For x3 , . . . , xd ∈ {1, . . . , n1 }, we define the two-dimensional
sheet Ax3 ,...,xd := (x1 , . . . , xd ) : x1 , x2 ∈ {1, . . . , n1 } . When r ≤ d − 2, we have that θ0 is
constant on each Ax3 ,...,xd . Hence, by Theorem 2,
K(θ0 ) log8+ n/K(θ0 )
. n−2/d log8 n.
R(θ̂n , θ0 ) .
n
Now suppose that θ0 ∈ Md−1 (Ld,n ). Let m be a positive integer to be chosen later. Then
(ℓ)
Ax3 ,...,xd = ⊔m
ℓ=−m Ax3 ,...,xd , where
ℓ−1
ℓ
(ℓ)
Ax3 ,...,xd := Ax3 ,...,xd ∩ v ∈ Ld,n :
.
< (θ0 )v ≤
m
m
(ℓ)
Let θ(m) ∈ M(Ld,n ) be the vector that takes the constant value ℓ/m on Ax3 ,...,xd for each
ℓ = −m, . . . , m. Then setting m ≍ n2/(3d) log−8/3 n, we have by Theorem 2 that
8
kθ(m) − θ0 k22 K(θ(m) ) log+ n/K(θ(m) )
1
m
R(θ̂n , θ0 ) .
+
≤ 2 + 2/d log8 n . n−4/(3d) log16/3 n.
n
n
m
n
as desired.
Finally, the r = d case is covered in Theorem 1.
16
5
Proof of results in Section 3
Henceforth we write EX for the expectation conditional on X1 , . . . , Xn , and also write Gn :=
n1/2 (Pn − P ). The key ingredient in the proofs of both Theorems 4 and 5 is the following
proposition, which controls the risk of the least squares estimator when f0 = 0. Recall that
γ2 = 9/2 and γd = (d2 + d + 1)/2 for d ≥ 3.
Proposition 7. Let d ≥ 2. There exists a constant Cd,m0 ,M0 > 0, depending only on d, m0
and M0 , such that
R(fˆn , 0) ≤ Cd,m0 ,M0 n−2/d log2γd n.
The proof of Proposition 7 requires several reduction techniques, which we detail in
Section 5.1 below. We first derive Theorems 4 and 5 from Proposition 7.
Proof of Theorem 4. Since the argument used in the proof of Theorem 1, up to (12), does
not depend on the design, we deduce from Chatterjee (2014, Corollary 1.2), Amelunxen et
al. (2014, Proposition 3.1(5)) and the Cauchy–Schwarz inequality that
R(fˆn , f0 ) .
1
E max 1, δ(M(GX )), n1/2 δ(M(GX ))1/2 .
n
On the other hand, by Proposition 7, we have
E δ M(GX ) .d,m0 ,M0 n1−2/d log2γd n.
(18)
(19)
We obtain the desired result by combining (18) and (19).
Proof of Theorem 5. For any f ∈ Fd , write θf,X := (f (X1 ), . . . , f (Xn ))⊤ ∈ Rn . By Bellec
(2017, Proposition 2.1), we have
n
o
1
2
ˆ
R(fn , f0 ) ≤ E inf kθf,X − θf0 ,X k2 + δ T (θf,X , M(GX ))
n f ∈Fd
n
o
1
(20)
≤ inf inf Ekθf,X − θf0 ,X k22 + E δ T (θf,X , M(GX )) .
n k∈N f ∈Fd(k)
(k)
Now, for a fixed f ∈ Fd , let R1 , . . . , Rk be the corresponding hyperrectangles for which f is
constant when restricted to each Rℓ . Define Xℓ := RL
:= |Xℓ |. Then
ℓ ∩ {X1 , . . . , Xn } and
NℓL
k
k
for fixed X1 , . . . , Xn , we have T (θf,X , M(GX )) ⊆
ℓ=1 M(GXℓ ).
ℓ=1 T 0, M(GXℓ ) =
Hence by Amelunxen et al. (2014, Proposition 3.1(9, 10)) and (19), we have that
h n
oi
E δ T (θf,X , M(GX )) = E E δ T (θf,X , M(GX )) N1 , . . . , Nk
X
k
n
o
≤E
E δ M(GXℓ ) Nℓ
ℓ=1
.d,m0 ,M0
X
k
1−2/d
2γd
d
Nℓ
log+ Nℓ .d n(k/n)2/d log2γ
E
+ (n/k),
ℓ=1
17
(21)
d
where the final bound follows from applying Lemma 2 to the function x 7→ x1−2/d log2γ
+ (x).
We complete the proof by substituting (21) into (20) and observing that
1
inf Ekθf,X − θf0 ,X k22 = inf Ekf − f0 k2L2 (Pn ) = inf kf − f0 k2L2 (P ) ,
(k)
(k)
n f ∈Fd(k)
f ∈Fd
f ∈Fd
as required.
Proof of Proposition 3. Without
= nd1 for some
Pdloss of generality, we may assume that nQ
w −1 w
n1 ∈ N. Let W := {w ∈ Ld,n : j=1 wj = n1 }. For any w ∈ W , define Cw := dj=1 nj 1 , n1j .
Note that x = (x1 , . . . , xd )⊤ ∈ ∪w∈W Cw if and only if ⌈n1 x1 ⌉ + · · · + ⌈n1 xd ⌉ = n1 . For any
τ = (τw ) ∈ {0, 1}|W | =: T , we define fτ ∈ Fd by
if ⌈n1 x1 ⌉ + · · · + ⌈n1 xd ⌉ ≤ n1 − 1
0
fτ (x) := 1
if ⌈n1 x1 ⌉ + · · · + ⌈n1 xd ⌉ ≥ n1 + 1
ρτ(⌈n1 x1 ⌉,...,⌈n1 xd ⌉) if x ∈ ∪w∈W Cw ,
where ρ ∈ [0, 1] is to be specified later. Moreover, let τ w be the binary vector differing from
τ in only the w coordinate. We write Eτ for the expectation over (X1 , Y1 ), . . . , (Xn , Yn ),
where Yi = fτ (Xi ) + ǫi for i = 1, . . . , n. We let EX be the expectation over (Xi )ni=1 alone
and EY |X,τ be the conditional expectation of (Yi )ni=1 given (Xi )ni=1 . Given any estimator f˜n ,
we have
XX Z
1
2
max Eτ f˜n − fτ L2 (Pn ) ≥ |W |
(f˜n − fτ )2 dPn
Eτ
τ ∈T
2 w∈W τ ∈T
Cw
Z
Z
1 XX
2
2
(f˜n − fτ w ) dPn
(f˜n − fτ ) dPn + Eτ w
Eτ
= |W |+1
2
C
C
w
w
w∈W τ ∈T
Z
Z
X
X
1
2
2
˜
˜
(fn − fτ w ) dPn
(fn − fτ ) dPn + EY |X,τ w
EX EY |X,τ
= |W |+1
2
Cw
Cw
w∈W τ ∈T
Z
h
i
1 XX
1
2
≥ |W |+1
(fτ − fτ w ) dPn 1 − dTV PY |X,τ , PY |X,τ w
,
(22)
EX
2
4
C
w
w∈W τ ∈T
where PY |X,τ (respectively PY |X,τ w ) is the conditional distribution of (Yi )ni=1 given (Xi )ni=1
when the true signal is fτ (respectively fτ w ). The final inequality in the above display follows
1/2
R
R
because for ∆ := Cw (fτ − fτ w )2 dPn
and A := Cw (f˜n − fτ )2 dPn ≥ ∆2 /4 , we have
Z
Z
o
∆2
c
2
2
˜
˜
w
w
w
PY |X,τ (A) + PY |X,τ (A )
(fn − fτ ) dPn ≥
(fn − fτ ) dPn + EY |X,τ
EY |X,τ
4
Cw
Cw
∆2
≥
1 − dTV PY |X,τ , PY |X,τ w .
4
By Pinsker’s inequality (cf. Pollard, 2002, p. 62), we obtain that
1
n
2
w
dTV PY |X,τ , PY |X,τ ≤ d2KL(PY |X,τ , PY |X,τ w ) = kfτ − fτ w k2L2 (Pn ) .
2
4
18
(23)
P
3/2
Writing Nw := ni=1 1{Xi ∈Cw } , we have Nw ∼ Bin(n, P (Cw )), so EX Nw ≥ m0 and EX Nw ≤
3/2
(EX Nw2 EX Nw )1/2 ≤ 21/2 M0 . Thus, together with (23), we have
Z
h
i
2
(fτ − fτ w ) dPn 1−dTV PY |X,τ , PY |X,τ w
EX
Cw
n1/2
2
≥ EX kfτ − fτ w kL2 (Pn ) 1 −
kfτ − fτ w kL2 (Pn )
2
ρ
ρ3
ρ2
ρ2
3/2
3/2
m0 − 1/2 M0
EX Nw ≥
.
= EX Nw −
n
2n
n
2
3/2
Substituting the above inequality into (22), we obtain that for ρ = 23/2 m0 /(3M0 ),
max Eτ f˜n − fτ
τ ∈T
2
L2 (Pn )
≥
|W | m30
≥ cd,m0 ,M0 n−1/d ,
3
27n M0
where the final inequality follows from a counting argument as in (10). This completes the
proof.
Proof of Proposition 4. Clearly we only need to establish the claim for f0 = 0. By Lemma 4,
−1
1/d
there is an event E with probability at least 1−e−ed (M0 n) log(M0 n) on which the data points
1/d
X1 , . . . , Xn contain a maximal antichain WX with cardinality at least n1−1/d /(2eM0 ). Write
WX+ := {Xi : ∃w ∈ WX , Xi ≻ w} and WX− := {Xi : ∃w ∈ WX , Xi ≺ w}. For each realisation
of the n-dimensional Gaussian random vector ǫ, we define θX = θX (ǫ) = ((θX )w ) by
if w ∈ WX+
1
(θX )w := sgn(ǫw ) if w ∈ WX
−1
if w ∈ WX− .
We see that θX ∈ M(GX ). By Chatterjee (2014, Theorem 1.1), for f0 = 0, we have that
2
1/2 ˆ
fn L2 (Pn ) = argmax
sup
hǫ, θi − t /2 =
sup
hǫ, θi.
n
t≥0
θ∈M(GX )∩B2 (t)
θ∈M(GX )∩B2 (1)
Hence
E fˆn
θX (ǫ)
1E
sup
hǫ, θi ≥ 1/2 E ǫ,
= 1/2 E
n
n
kθX (ǫ)k2
θ∈M(GX )∩B2 (1)
X
X
X
1
|ǫi |1E .
≥ E
ǫi 1E +
ǫi 1E −
n
−
+
i:X ∈W
1
1
L2 (Pn )
i:Xi ∈WX
i:Xi ∈WX
i
(24)
X
The first two terms in the bracket are seen to be zero by computing the expectation conditionally on X1 , . . . , Xn . For the third term, we have that
X
X
2 1/2
X
|ǫi |1E = E
E
E |WX |1E &d,M0 n1−1/d ,
E |ǫi |1E
≥
(25)
π
i:X ∈W
i:X ∈W
i
X
i
X
19
where the final inequality follows from Lemma 4. By (24), (25) and the Cauchy–Schwarz
inequality, we have that
2
2
E fˆn L2 (Pn ) ≥ E fˆn L2 (Pn ) &d,M0 n−2/d ,
as desired.
5.1
Proof of Proposition 7
The proof of Proposition 7 is rather technical, so we sketch a brief outline of the main steps
below:
Step 1. Instead of bounding R(fˆn , 0) directly, we first consider bounding
E kfˆn k2L2 (P ) 1{kfˆn k∞ ≤4 log1/2 n} .
(26)
By Proposition 8, this task essentially reduces to understanding two empirical processes (28) and (29). By means of Lemmas 5 and 6, this in turn reduces to the
study of the symmetrised local empirical process
sup
E
f ∈Fd ∩B∞ (1)∩B2 (r,P )
n
1 X
n1/2
ξi f (Xi ) ,
(27)
i=1
for a suitable L2 (P ) radius r.
Step 2. To obtain a sharp bound on the empirical process in (27), which constitutes the
main technical challenge of the proof, we sliceP[0, 1]d into strips of the form [0, 1]d−1 ×
, ℓ ], for ℓ = 1, . . . , n1 , and decompose ni=1 ξi f (Xi ) into sums of smaller em[ ℓ−1
n1 n1
pirical processes over these strips. Each of these smaller empirical processes is then
controlled via a bracketing entropy chaining argument (Lemma 7). The advantage
of this decomposition is that the block monotonicity permits good control of the
L2 (P ) norm of the envelope function in each strip (Lemma 9).
Step 3. Having bounded (27), and hence also (26), we finally translate the bound of (26)
back to a bound for R(fˆn , 0), our original quantity of interest. The cost of L∞
truncation is handled in Lemma 10, whereas our understanding of the symmetrised
empirical process in (27) helps to control the discrepancy between the L2 (P ) norm
and L2 (Pn ) norm risks (cf. Proposition 10).
In our empirical process theory arguments, since our least squares estimator fˆn is defined
to be lower semi-continuous, we can avoid measurability and countability digressions by
defining G the class of real-valued lower semi-continuous functions on [0, 1]d and Fd′ := {f ∈
Fd ∩ G : f |(Q∩[0,1])d ⊆ Q}. This is a countable, uniformly dense3 subset of Fd ∩ G so that, for
example, supf ∈Fd ∩G Gn f = supf ∈Fd′ Gn f .
The main content of Step 1 is the following proposition.
Here ‘uniformly dense’ means that for any f ∈ Fd ∩ G, we can find a sequence (fm ) in Fd′ such that
kfm − f k∞ → 0. This can be done by defining, e.g., fm (x) := m−1 ⌈mf (x)⌉.
3
20
Proposition 8. Suppose that for each n ∈ N there exist a function φn : [0, ∞) → [0, ∞)
and rn ≥ n−1/2 log1/2 n such that φn (rn ) ≤ n1/2 rn2 . Moreover, assume that for all r ≥ rn the
map r 7→ φn (r)/r is non-increasing and
sup
E
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (r,P )
n
1 X
n1/2
and
sup
E
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (r,P )
n
1 X
n1/2
ǫi f (Xi ) ≤ C1 φn (r),
(28)
ξi f 2 (Xi ) ≤ C2 φn (r),
(29)
i=1
i=1
for some constants C1 , C2 > 0 that do not depend on r and n. Then for f0 = 0, we have that
E kfˆn k2L2 (P ) 1{kfˆn k∞ ≤4 log1/2 n} .C1 ,C2 rn2 .
Proof. Let Mn f :=
−P f 2 . Then
2
n
Pn
i=1 ǫi f (Xi )
−
1
n
Pn
i=1
f 2 (Xi ) and Mf := E(Mn f ) = −kf k2L2 (P ) =:
n
2X
ǫi f (Xi ) + (Pn − P )f 2 .
|Mn f − Mf | ≤
n i=1
P
P
Moreover, by definition of fˆn , we have ni=1 {ǫi − fˆn (Xi )}2 ≤ ni=1 ǫ2i , so Mn fˆn ≥ 0. Fix
s ≥ 1 and for ℓ ∈ N, let Fd,ℓ := Fd ∩ B∞ (4 log1/2 n) ∩ B2 (2ℓ srn ). Then by a union bound,
we have
∞
X
1/2
ˆ
ˆ
Mn f ≥ 0
P kfn kL2 (P ) ≥ srn ∩ kfn k∞ ≤ 4 log n ≤
P
sup
ℓ=1
≤
≤
∞
X
ℓ=1
∞
X
ℓ=1
P
P
sup Mn f − Mf ≥ 22ℓ−2 s2 rn2
f ∈Fd,ℓ
sup
f ∈Fd,ℓ
n
1 X
n1/2
i=1
ǫi f (Xi ) ≥
22ℓ−4 s2 n1/2 rn2
f ∈Fd,ℓ \Fd,ℓ−1
∞
X
2
2ℓ−3 2 1/2 2
s n rn .
+
P sup Gn f ≥ 2
ℓ=1
f ∈Fd,ℓ
(30)
By a moment inequality for empirical processes (Giné, Latala and Zinn, 2000, Proposition 3.1) and (28), we have
4
n
1 X
log4 n
4 ℓ
ℓ
4
E
sup 1/2
.C1 φn (2 srn ) + (2 srn ) +
ǫi f (Xi )
.
(31)
n2
f ∈Fd,ℓ n
i=1
Similarly, by symmetrisation (cf. van der Vaart and Wellner (1996, Lemma 2.3.1)), the
moment inequality for empirical processes mentioned above and condition (29), we have
4
4
n
1 X 2
log2 n
2
.E
sup 1/2
E
sup Gn f
.C2 φ4n (2ℓ srn ) + (2ℓ srn )4 +
ξi f (Xi )
.
n2
f ∈Fd,ℓ n
f ∈Fd,ℓ
i=1
(32)
21
By (30), (31), (32) and Markov’s inequality, we obtain that
4
∞
ℓ
X
1
1
φ
(2
sr
)
n
n
1/2
+
.
, (33)
P kfˆn kL2 (P ) ≥ srn ∩ kfˆn k∞ ≤ 4 log n .C1 ,C2
22ℓ s2 n1/2 rn2
s4
s4
ℓ=1
where we have used the assumption rn ≥ n−1/2 log1/2 n and the fact that φn (2ℓ srn ) ≤
2ℓ sφn (rn ) ≤ 2ℓ sn1/2 rn2 for the non-increasing function r 7→ φn (r)/r. The bound in (33)
is valid for all s ≥ 1. Hence
Z ∞
2
2
ˆ
P fˆn L2 (P ) 1{kfˆn k∞ ≤4 log1/2 n} ≥ t dt
E kfn kL2 (P ) 1{kfˆn k∞ ≤4 log1/2 n} =
0
Z ∞
2
2
2
≤ rn + 2rn
s P fˆn L2 (P ) 1{kfˆn k∞ ≤4 log1/2 n} ≥ s2 rn2 ds
1
.C1 ,C2
rn2 ,
as desired.
The proposition below on the size of the symmetrised empirical process solves Step 2 in
the outline of the proof of Proposition 7.
2 −d)/2
Proposition 9. Fix d ≥ 2 and suppose that r ≥ n− max{1/d,(1−2/d)} log(d
a constant Cd,m0 ,M0 > 0, depending only on d, m0 and M0 , such that
sup
E
f ∈Fd ∩B∞ (1)∩B2 (r,P )
n
1 X
n1/2
i=1
n. There exists
ξi f (Xi ) ≤ Cd,m0 ,M0 rn1/2−1/d logγd −1/2 n.
Proof. It is convenient here to work with the class of block decreasing functions Fd,↓ := {f :
+
[0, 1]d → R : −f ∈ Fd } instead. We write Fd+ := {f ∈ Fd : f ≥ 0} and Fd,↓
:= {f ∈
Fd,↓ : f ≥ 0}. By replacing f with −f and decomposing any function f into its positive and
+
negative parts, it suffices to prove the result with Fd,↓
in place of Fd . We handle the cases
d = 2 and d ≥ 3 separately.
Case d = 2. We apply Lemma 7 with η = r/(2n) and Lemma 8 to obtain
sup
E
+
f ∈F2,↓
∩B∞ (1)∩B2 (r,P )
n
1 X
n1/2
1/2
ξi f (Xi ) .d,m0 ,M0 n
i=1
3
η + log n
Z
r
η
log4 n(log log n)2
r
dε +
ε
n1/2
. r log4 n,
as desired.
Case d ≥ 3. We assume without loss of generality that n = nd1 for some n1 ∈ N. We
1
define strips Iℓ := [0, 1]d−1 × [ ℓ−1
, ℓ ] for ℓ = 1, . . . , n1 , so that [0, 1]d = ∪nℓ=1
Iℓ . Our
n1 n1
strategy is to analyse the expected supremum of the symmetrised empirical process when
restricted to each strip. To this end, define Sℓ := {X1 , . . . , Xn } ∩ Iℓ and Nℓ := |Sℓ |, and let
Ω0 := {m0 n1−1/d /2 ≤ minℓ Nℓ ≤ maxℓ Nℓ ≤ 2M0 n1−1/d }. Then by Hoeffding’s inequality,
22
P(Ωc0 )
Hence we have
n1
X
m0 n
≤ 2n1 exp −m20 n1−2/d /8 .
P Nℓ − ENℓ >
≤
2n1
ℓ=1
sup
E
+
f ∈Fd,↓
∩B∞ (1)∩B2 (r,P )
where
Eℓ := E
n
1 X
n1/2
i=1
1/2
n1
X
Nℓ
2 1−2/d
ξi f (Xi ) ≤
E
E
+
C
exp
−m
n
/16
,
1
ℓ
Ω
0
0
n1/2
ℓ=1
(34)
1
sup
+
f ∈Fd,↓
∩B∞ (1)∩B2 (r,P )
1/2
Nℓ
X
ξi f (Xi )
N1 , . . . , Nn1 .
i:Xi ∈Sℓ
R
+
By Lemma 9, for any f ∈ Fd,↓
∩ B∞ (1) ∩ B2 (r, P ) and ℓ ∈ {1, . . . , n1 }, we have Iℓ f 2 dP ≤
2
7(M0 /m0 )ℓ−1 r 2 logd n =: rn,ℓ
. Consequently, we have by Lemma 7 that for any η ∈ [0, rn,ℓ /3),
Z rn,ℓ
H[ ] (rn,ℓ )
1/2
1/2
H[ ] (ε) dε +
Eℓ . Nℓ η +
,
(35)
1/2
η
Nℓ
+
where H[ ] (ε) := log N[ ] ε, Fd,↓
(Iℓ )∩B∞ (1; Iℓ )∩B2 (rn,ℓ , P ; Iℓ ), k·kL2(P ;Iℓ ) . Here, kf k2L2 (P ;Iℓ ) :=
R 2
+
f dP , the set Fd,↓
(Iℓ ) is the class of non-negative functions on Iℓ that are block decreasing,
Iℓ
B∞ (1; Iℓ ) is the class of functions on Iℓ that are bounded by 1 and B2 (rn,ℓ , P ; Iℓ ) is the class
+
of measurable functions f on Iℓ with kf kL2 (P ;Iℓ ) ≤ rn,ℓ . Any g ∈ Fd,↓
(Iℓ ) ∩ B∞ (1; Iℓ ) ∩
1/2
+
B2 (rn,ℓ , P ; Iℓ ) can be rescaled into a function fg ∈ Fd,↓ ∩ B∞ (1) ∩ B2 n1 (M0 /m0 )1/2 rn,ℓ , P
map fg (x1 , . . . ,Rxd−1 , xd ) := g(x1 , . . . , xd−1 , (xd + ℓ − 1)/n1 ). Moreover,
Rvia the invertible
2
′
(f − fg ) dP ≥ n1 (m0 /M0 ) Iℓ (g − g ′ )2 dP . Thus, by Lemma 8, for ǫ ∈ [η, rn,ℓ],
[0,1]d g
+
H[ ] (ε) ≤ log N[ ] n1/(2d) (m0 /M0 )1/2 ε, Fd,↓
∩ B∞ (1) ∩ B2 n1/(2d) (M0 /m0 )1/2 rn,ℓ , k · kL2 (P )
2(d−1)
rn,ℓ
2
.d,m0 ,M0
logd+ (1/ǫ).
ε
Substituting the above entropy bound into (35), and choosing η = n−1/(2d) rn,ℓ , we obtain
2
d−1
Z rn,ℓ
2
2
d−1
rn,ℓ
logd /2 n logd n
rn,ℓ
logd n
1/2
1/2
d2 /2
. Nℓ η +
.
Eℓ .d,m0 ,M0 Nℓ η + log
n
+
dε +
1/2
1/2
ε
η d−2
Nℓ
Nℓ
η
Hence
Eℓ 1Ω0 .d,m0 ,M0 rn,ℓ n1/2−1/d logd
2 /2
2 /2
2
n + n−1/2+1/(2d) logd n .m0 ,M0 rn,ℓ n1/2−1/d logd
n, (36)
2
where in the final inequality we used the conditions that d ≥ 3 and r ≥ n−(1−2/d) log(d −d)/2 n.
Combining (34) and (36), we have that
n1
n
1 X
1 X
(d2 +d)/2
−1/2
1/2−1/d
ℓ
E
sup
log
n
ξi f (Xi ) .d,m0 ,M0 rn
1/2
1/2
+
n1 ℓ=1
f ∈Fd,↓
∩B∞ (1)∩B2 (r,P ) n
i=1
2 +d)/2
. rn1/2−1/d log(d
which completes the proof.
23
n,
Finally, we need the following proposition to switch between L2 (P ) and L2 (Pn ) norms as
described in Step 3.
Proposition 10. Fix d ≥ 2 and suppose that f0 = 0. There exists a constant Cd,m0 ,M0 > 0,
depending only on d, m0 and M0 , such that
h
i
2
2
E fˆn L2 (Pn ) 1{kfˆn k∞ ≤4 log1/2 n} ≤ Cd,m0 ,M0 n−2/d log2γd n + E fˆn L2 (P ) 1{kfˆn k∞ ≤4 log1/2 n} .
Proof. To simplify notation, we define f˜n := fˆn 1{kfˆn k∞ ≤4 log1/2 n} and rn := n−1/d logγd n. We
write
2
2
E fˆn L2 (Pn ) 1{kfˆn k∞ ≤4 log1/2 n} = E f˜n L2 (Pn )
2
2
= E f˜n L2 (Pn ) 1{kfˆn kL (P ) ≤rn } + E f˜n L2 (Pn ) 1{kfˆn kL (P ) >rn } , (37)
2
2
and control the two terms on the right hand side of (37) separately. For the first term, we
have
E f˜n
1
2
L2 (Pn ) {kfˆn kL2 (P ) ≤rn }
n
1X 2
f (Xi )
≤E
sup
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P ) n i=1
. rn2 +
.
rn2
n
X
1
E
sup
ξi f 2 (Xi )
n f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P ) i=1
n
X
log1/2 n
+
ξi f (Xi )
E
sup
n
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P ) i=1
.d,m0 ,M0 rn2 + rn n−1/d logγd n . rn2 ,
(38)
where the second line uses the symmetrisation inequality (cf. van der Vaart and Wellner,
1996, Lemma 2.3.1), the third inequality follows from Lemma 6 and the penultimate inequality follows from applying Proposition 9 to f /(4 log1/2 n). For the second term on the
′
right-hand side of (37), we first claim that there exists some constant Cd,m
> 0, depend0 ,M0
ing only on d, m0 and M0 , such that
2
Pn f 2
′
− 1 > Cd,m0 ,M0 ≤ 2 .
(39)
P
sup
2
n
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P )c P f
To see this, we adopt a peeling argument as follows. Let Fd,ℓ := {f ∈ Fd ∩ B∞ (4 log1/2 n) :
2ℓ−1 rn2 < P f 2 ≤ 2ℓ rn2 } and m be the largest integer such that 2m rn2 < 32 log n (so that
m ≍ log n). We have that
sup
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P )c
Pn f 2
1
− 1 = 1/2
2
Pf
n
|Gn f 2 |
2
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P )c P f
n
o
1
ℓ 2 −1
2
. 1/2 max (2 rn ) sup |Gn f | .
n ℓ=1,...,m
f ∈Fd,ℓ
24
sup
By Talagrand’s concentration inequality (cf. Talagrand (1996)) for empirical processes, in
the form given by Massart (2000, Theorem 3), applied to the class {f 2 : f ∈ Fd,ℓ }, we have
that for any sℓ > 0,
552sℓ log n
1/2
1/2
2
(ℓ+7)/2
2
rn sℓ log n +
P sup |Gn f | > 2E sup |Gn f | + 2
≤ e−sℓ .
1/2
n
f ∈Fd,ℓ
f ∈Fd,ℓ
Here we have used the fact that supf ∈Fd,ℓ VarP f 2 ≤ supf ∈Fd,ℓ P f 2kf k2∞ ≤ 2ℓ+4 rn2 log n. Further, we note by the symmetrisation inequality again, Lemma 6 and Proposition 9 that
1
2
E sup |Gn f | .
n1/2
n
X
E sup
n
X
log1/2 n
ξi f (Xi ) .
ξi f (Xi )
E sup
n1/2
f ∈Fd,ℓ i=1
2
f ∈Fd,ℓ i=1
.d,m0 ,M0 2ℓ/2 rn n1/2−1/d
f ∈Fd,ℓ
logγd n.
By a union bound, we have that with probability at least 1 −
sup
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P )c
Pn f 2
−1
Pf2
Pm
ℓ=1
e−sℓ ,
1/2
n1/2−1/d logγd n + sℓ log1/2 n sℓ log n
+ ℓ 2 .
.d,m0 ,M0 max
ℓ=1,...,m
2ℓ/2 n1/2 rn
2 nrn
P
P
−sℓ
−ℓ−1
By choosing sℓ := 2ℓ log n, we see that m
≤ ∞
≤ 2n−2 and
ℓ=1 e
ℓ=1 n
sup
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P )c
Pn f 2
− 1 .d,m0 ,M0 1,
Pf2
which verifies (39). Now let E := supf ∈Fd ∩B∞ (4 log1/2 n)∩B2 (rn ,P )c
E f˜n
1
2
L2 (Pn ) {kfˆn kL2 (P ) >rn }
Pn f 2
P f2
′
− 1 ≤ Cd,m
. Then
0 ,M0
32 log n
n2
32 log n
2
′
.
≤ (Cd,m
+ 1)E f˜n L2 (P ) +
0 ,M0
n2
≤ E f˜n
1
1
2
L2 (Pn ) {kfˆn kL2 (P ) >rn } E
+
(40)
Combining (37), (38) and (40), we obtain
E f˜n
2
L2 (Pn )
.d,m0 ,M0 rn2 + E f˜n
2
,
L2 (P )
as desired.
Proof of Proposition 7. For f0 = 0, we decompose
2
R(fˆn , 0) = E fˆn L2 (Pn ) 1{kfˆn k∞ ≤4 log1/2 n} + E fˆn
1
2
1/2
n}
L2 (Pn ) {kfˆn k∞ >4 log
(41)
and handle the two terms on the right-hand side separately. For the first term, let rn :=
n−1/d logγd n and observe that by Lemma 5 and Proposition 9, we have that for r ≥ rn ,
E
sup
f ∈Fd ∩B∞ (4 log
1/2
n)∩B2 (r,P )
n
1 X
n1/2
ǫi f (Xi ) .d,m0 ,M0 rn1/2−1/d logγd n.
i=1
25
On the other hand, by Lemma 6 and Proposition 9, for r ≥ rn ,
n
1 X 2
E
sup
ξi f (Xi ) .d,m0 ,M0 rn1/2−1/d logγd n.
1/2
f ∈Fd ∩B∞ (4 log1/2 n)∩B2 (r,P ) n
i=1
It follows that the conditions of Proposition 8 are satisfied for this choice of rn with φn (r) :=
rn1/2−1/d logγd n. By Propositions 10 and 8, we deduce that
2
2
E fˆn
.d,m ,M n−2/d log2γd n + E fˆn
1 ˆ
1 ˆ
1/2
1/2
L2 (Pn ) {kfn k∞ ≤4 log
n}
0
L2 (P ) {kfn k∞ ≤4 log
0
−2/d
.d,m0 ,M0 n
log
2γd
n.
n}
(42)
For the second term on P
the right-hand side ofP
(41), we note that by the definition of the
n
2
ˆ
least squares estimator, i=1 {fn (Xi ) − ǫi } ≤ ni=1 ǫ2i , so
n
n
n
2X ˆ
2X 2
4X 2
2
2
ˆ
fn L2 (Pn ) ≤
{fn (Xi ) − ǫi } +
ǫ ≤
ǫ.
n i=1
n i=1 i
n i=1 i
Thus,
E fˆn
1
2
1/2
n}
L2 (Pn ) {kfˆn k∞ >4 log
≤ 4E ǫ21 1{kfˆn k∞ >4 log1/2 n}
. P(kfˆn k∞ > 4 log1/2 n)1/2 . n−3 ,
(43)
where the final inequality follows from Lemma 10. The proof is completed by substituting (42) and (43) into (41).
6
Appendix: proofs of ancillary results
The proof of Corollary 1 requires the following lemma on Riemann approximation of block
increasing functions.
Lemma 1. Suppose n1 = n1/d
f ∈ Fd , define fL (x1, . . . , xd ) :=
is a positive integer. For any
−1
−1
−1
f n1 ⌊n1 x1 ⌋, . . . , n1 ⌊n1 xd ⌋ and fU (x1 , . . . , xd ) := f n1 ⌈n1 x1 ⌉, . . . , n−1
1 ⌈n1 xd ⌉ . Then
Z
(fU − fL )2 ≤ 4dn−1/d kf k2∞ .
[0,1]d
Proof. For x = (x1 , . . . , xd )⊤ and x′ = (x′1 , . . . , x′d )⊤ in Ld,n , we say x and x′ are equivalent
F
if and only if xj − x1 = x′j − x′1 for j = 1, . . . , d. Let Ld,n = N
r=1 Pr be the partition of Ld,n
into equivalence classes. Since each Pr has non-empty intersection with a different element
of the set {(x1 , . . . , xd ) ∈ Ld,n : minj xj = 1}, we must have N ≤ dn1−1/d . Therefore, we
have
Z
N Z
X
2
(fU − fL ) =
(fU − fL )2
[0,1]d
r=1
d
n−1
1 (Pr +(−1,0] )
N
X
2
≤ kf k∞
n
r=1
≤
X
x=(x1 ,...,xd )⊤ ∈Pr
x1
x1 − 1
xd
xd − 1
f
−f
,...,
,...,
n
n
n
n
2N
kf k∞ f (1, . . . , 1) − f (0, . . . , 0) ≤ 4dn−1/d kf k2∞ ,
n
26
as desired.
The following is a simple generalisation of Jensen’s inequality.
Lemma 2. Suppose h : [0, ∞) → (0, ∞) is a non-decreasing function satisfying the following:
(i) There exists x0 ≥ 0 such that h is concave on [x0 , ∞).
(ii) There exists some x1 > x0 such that h(x1 ) − xh′+ (x1 ) ≥ h(x0 ), where h′+ is the right
derivative of h.
Then there exists a constant Ch > 0 depending only on h such that for any nonnegative
random variable X with EX < ∞, we have
Eh(X) ≤ Ch h(EX).
Proof. Define H : [0, ∞) → [h(0), ∞) by
(
h(x1 ) − x1 h′+ (x1 ) + xh′+ (x1 ) if x ∈ [0, x1 )
H(x) :=
h(x)
if x ∈ [x1 , ∞).
Then H is a concave majorant of h. Moreover, we have H ≤ (h(x1 )/h(0))h. Hence, by
Jensen’s inequality, we have
Eh(X) ≤ EH(X) ≤ H(EX) ≤
h(x1 )
h(EX),
h(0)
as desired.
We need the following lower bound on the metric entropy of M(L2,n ) ∩ B2 (1) for the
proof of Proposition 2.
Lemma 3. There exist universal constants c > 0 and ε0 > 0 such that
log N ε0 , M(L2,n ) ∩ B2 (1), k · k2 ≥ c log2 n.
Proof. It suffices to prove the equivalent result that there exist
universal constants c, ε0 > 0
such that the packing number D ε0 , M(L2,n ) ∩ B2 (1), k · k2 (i.e. the maximum number of
disjoint open Euclidean balls of radius ε0 that can be fitted into M(L2,n ) ∩ B2 (1)) is at least
exp(c log2 n). Without loss of generality, we may also assume that n1 := n1/2 = 2ℓ − 1 for
some ℓ ∈ N, so that ℓ ≍ log n. Now, for r = 1, . . . , ℓ, let Ir := {2r−1, . . . , 2r − 1} and consider
the set
n −1
−1Ir ×Is o
Ir ×Is
L2,n
⊆ M(L2,n ) ∩ B2 (1),
,√
M̄ := θ ∈ R
: θIr ×Is ∈ √
2r+s+1 log n 2r+s log n
2
where 1Ir ×Is denotes the all-one vector on Ir × Is . Define a bijection ψ : M̄ → {0, 1}ℓ by
ψ(θ) := 1
√
θIr ×Is =−1Ir ×Is / 2r+s+1 log n
27
ℓ
r,s=1
.
Then, for θ, θ′ ∈ M̄,
kθ −
θ′ k22
2
dH (ψ(θ), ψ(θ′ )) 1
1
=
1 − 1/2 ,
4
2
log2 n
where dH (·, ·) denotes the Hamming distance. On the other hand, the Gilbert–Varshamov
2
bound (e.g. Massart, 2007, Lemma 4.7) entails that there exists a subset I ⊆ {0, 1}ℓ such
that |I| ≥ exp(ℓ2 /8) and dH (v, v ′ ) ≥ ℓ2 /4 for any distinct v, v ′ ∈ I. Then the set ψ −1 (I) ⊆
M̄ has cardinality at least exp(ℓ2 /8) ≥ exp(log2 n/32), and each pair of distinct elements
ℓ2 /4 1
1 2
have squared ℓ2 distance at least ε0 := log
) & 1, as desired.
(1 − 21/2
2
n4
Lemma 4 below gives a lower bound on the size of the maximal antichain (with respect to the natural partial ordering on Rd ) among independent and identically distributed
X1 , . . . , Xn .
iid
Lemma 4. Let d ≥ 2. Let X1 , . . . , Xn ∼ P , where P is a distribution on [0, 1]d with Lebesgue
−1
1/d
density bounded above by M0 ∈ [1, ∞). Then with probability at least 1−e−ed (M0 n) log(M0 n) ,
1/d
there is an antichain in GX with cardinality at least n1−1/d /(2eM0 ).
Proof. By Dilworth’s Theorem (Dilworth, 1950), for each realisation of the directed acyclic
graph GX , there exists a covering of V (GX ) by chains C1 , . . . , CM , where M denotes the
cardinality of a maximum antichain of GX . Thus, it suffices to show that with the given
probability, the maximum chain length of GX is at most k := ⌈e(M0 n)1/d ⌉ ≤ 2e(M0 n)1/d .
By a union bound, we have that
n
n
(k!)−d M0k
P(X1 · · · Xk ) =
P(∃ a chain of length k in GX ) ≤
k
k
k −kd
en
k
−1
1/d
≤
M0k ≤ (M0 n)−k/d ≤ e−ed (M0 n) log(M0 n) ,
k
e
as desired.
The following two lemmas control the empirical processes in (28) and (29) by the symmetrised empirical process in (27).
Lemma 5. Let n ≥ 2, and suppose X1 , . . . , Xn are independent and identically distributed
on X . Then for any countable class F of measurable, real-valued functions defined on X , we
have
n
n
X
X
1/2
E sup
ǫi f (Xi ) ≤ 2 log n E sup
ξi f (Xi ) .
f ∈F i=1
f ∈F i=1
Proof. Let α0 := 0, and for k = 1, . . . , n, let αk := E|ǫ(k) |, where |ǫ(1) | ≤ · · · ≤ |ǫ(n) | are the
28
order statistics of {|ǫ1 |, . . . , |ǫn |}, so that αn ≤ (2 log n)1/2 . Observe that for any k = 1, . . . , n,
E sup
k
X
k
X
ξi f (Xi ) = E sup
f ∈F i=1
ξi f (Xi ) + E
f ∈F i=1
ξi f (Xi )
i=k+1
X
n
≤ E sup E
ξi f (Xi )
f ∈F
n
X
X1 , . . . , Xk , ξ 1 , . . . , ξ k
i=1
≤ E sup
n
X
ξi f (Xi ) .
f ∈F i=1
(44)
We deduce from Han and Wellner (2017, Proposition 1) and (44) that
E sup
n
X
f ∈F i=1
ǫi f (Xi ) ≤ 21/2
n
X
k=1
(αk − αk−1)E sup
k
X
f ∈F i=1
ξi f (Xi ) ≤ 21/2 αn E sup
n
X
ξi f (Xi ) ,
f ∈F i=1
as required.
Lemma 6. Let X1 , . . . , Xn be random variables taking values in X and F be a countable
class of measurable functions f : X → [−1, 1]. Then
E sup
f ∈F
n
X
i=1
2
ξi f (Xi ) ≤ 4E sup
f ∈F
n
X
ξi f (Xi ) .
i=1
Proof. By Ledoux and Talagrand (2011, Theorem 4.12), applied to φi (y) = y 2/2 for i =
1, . . . , n (note that y 7→ y 2 /2 is a contraction on [0, 1]), we have
E sup
n
X
f ∈F i=1
n
X
ξi f 2 (Xi )
ξi f (Xi ) = E E sup
X1 , . . . , Xn
ξi f (Xi )
X1 , . . . , Xn
2
f ∈F i=1
n
X
≤ 4E E sup
f ∈F i=1
= 4E sup
n
X
ξi f (Xi ) ,
f ∈F i=1
as required.
The following is a local maximal inequality for empirical processes under bracketing
entropy conditions. This result is well known for η = 0 in the literature, but we provide a
proof for the general case η ≥ 0 for the convenience of the reader.
iid
Lemma 7. Let X1 , . . . , Xn ∼ P on X with empirical distribution Pn , and, for some r > 0, let
G ⊆ B∞ (1)∩B2 (r, P ) be a countable class of measurable functions. Then for any η ∈ [0, r/3),
we have
Z r
1/2
1/2
E sup |Gn f | . n η +
log+ N[ ] (ε, G, k · kL2 (P ) ) dε + n−1/2 log+ N[ ] (r, G, k · kL2 (P ) ).
f ∈G
η
The above
P inequality also holds if we replace Gn f with the symmetrised empirical process
n−1/2 ni=1 ξi f (Xi ).
29
Proof. Writing Nr := N[ ] (r, G, k · kL2 (P ) ), there exists {(fℓL , fℓU ) : ℓ = 1, . . . , Nr } that form
an r-bracketing set for G in the L2 (P ) norm. Letting G1 := {f ∈ G : f1L ≤ f ≤ f1U } and
ℓ−1
r
Gℓ := {f ∈ G : fℓL ≤ f ≤ fℓU } \ ∪j=1
Gj for ℓ = 2, . . . , Nr , we see that {Gℓ }N
ℓ=1 is a partition
of G such that the L2 (P )-diameter of each Gℓ is at most r. It follows by van der Vaart and
Wellner (1996, Lemma 2.14.3) that for any choice of fℓ ∈ Gℓ , we have that
Z r
1/2
1/2
E sup |Gn f | . n η +
log+ N[ ] (ε, G, k · kL2 (P ) ) dε
f ∈G
η
(45)
+ E max |Gn fℓ | + E max Gn sup |f − fℓ | ,
ℓ=1,...,Nr
ℓ=1,...,Nr
f ∈Gℓ
The third and fourth terms of (45) can be controlled by Bernstein’s inequality (in the form
of (2.5.5) in van der Vaart and Wellner (1996)):
log+ Nr
1/2
+ r log+ Nr .
E max |Gn fℓ | ∨ E max Gn sup |f − fℓ | .
1/2
ℓ=1,...,Nr
ℓ=1,...,Nr
n
f ∈Gℓ
1/2
Since η < r/3, the last term r log+ Nr in the above display can be assimilated into the
entropy integral in (45), which establishes the claim for E supf ∈G |Gn f |.
We now study the symmetrised empirical process. For f ∈ G, we define e ⊗ f : {−1, 1} ×
X → R by (e ⊗ f )(t, x) := tf (x), and apply the previous result to the function class
e ⊗ G := {e ⊗ f : f ∈ G}. Here the randomness is induced by the independently and
identically distributed pairs (ξi , Xi )1≤i≤n . For any f ∈ G and any ε-bracket [f , f¯] containing
f , we have that [e+ ⊗ f − e− ⊗ f¯, e+ ⊗ f¯ − e− ⊗ f ] is an ε-bracket for e ⊗ f in the L2 (Pξ ⊗ P )
metric, where e+ (t) := max{e(t), 0} = max(t, 0) and e− (t) = max(−t, 0). Writing Pξ denote
the Rademacher distribution on {−1, 1}, it follows that for every ǫ > 0,
N[ ] (ε, e ⊗ G, L2 (Pξ ⊗ P )) ≤ N[ ] (ε, G, L2 (P )),
which proves the claim for the symmetrised empirical process.
In the next two lemmas, we assume, as in the main text, that P is a distribution on
[0, 1]d with Lebesgue density bounded above and below by M0 ∈ [1, ∞) and m0 ∈ (0, 1]
+
respectively. Recall that Fd,↓
= {f : −f ∈ Fd , f ≥ 0}. The following result is used to
control the bracketing entropy terms that appear in Lemma 7 when we apply it in the proof
of Proposition 9.
Lemma 8. There exists a constant Cd > 0, depending only on d, such that for any r, ǫ > 0,
+
log N[ ] ε, Fd,↓
∩ B2 (r, P )∩B∞ (1), k · kL2 (P )
(
2 M0
4
2 r log+ (1/ε)
0
(r/ε)2 M
log
(
)
log
(1/ε)
log
if d = 2,
+
+
m0
m0
ε
≤ Cd
d2
2(d−1) M0 d−1
if d ≥ 3.
(r/ε)
( m0 ) log+ (1/ε)
R
Proof. For any Borel measurable S ⊆ [0, 1]d , we define kf k2L2 (P ;S) := S f 2 dP . We first claim
that for any η ∈ (0, 1/4],
(
2 M0
4
2 r log(1/η)
r 2 M0
(
)
log
(
)
log
(1/η)
log
if d = 2,
+
+
ε
m0
m0
ε
log N[ ] ε, Fd,↓
∩B2 (r), k·kL2(P ;[η,1]d ) .d
d2
r 2(d−1) M0 d−1
( m0 ) log (1/η)
if d ≥ 3.
(ε)
30
+
By the cone property of Fd,↓
, it suffices to establish the above claim when r = 1. We denote
by vol(S) the Lebesgue measure of a measurable set S ⊆ [0, 1]d . By Gao and Wellner (2007,
Theorem 1.1) and a scaling argument, we have for any δ, M > 0 and any hyperrectangle
A ⊆ [0, 1]d that
(
(γ/δ)2 log2+ (γ/δ) if d = 2,
+
log N[ ] δ, Fd,↓
∩ B∞ (M), k · kL2 (P ;A) .d
(46)
(γ/δ)2(d−1)
if d ≥ 3,
1/2
where γ := M0 Mvol1/2 (A). Let m := ⌊log2 (1/η)⌋ and define Iℓ := [2ℓ η, 2ℓ+1η] ∩ [0, 1] for
+
each ℓ = 0, . . . , m. Then for ℓ1 , . . . , ℓd ∈ {0, . . . , m}, any f ∈ Fd,↓
∩ B2 (1, P ) is uniformly
−1/2
Qd
Qd
ℓj
bounded by m0 j=1 2 η
on the hyperrectangle j=1 Iℓj . Then by (46) we see that
for any δ > 0,
(
0
) log2+ (1/δ) if d = 2,
δ −2 (M0 /m0 ) log2 ( M
+
m0
log N[ ] δ, Fd,↓
∩ B2 (1), k · kL2 (P ;Qd Iℓ ) .d
j=1 j
δ −2(d−1) (M0 /m0 )d−1
if d ≥ 3,
where we have used the fact that log+ (ax) ≤ 2 log+ (a) log+ (x) for any a, x > 0. Global
+
brackets for Fd,↓
∩B2 (1) on [η, 1]d can then be constructed by taking all possible combinations
of local brackets on Iℓ1 × · · · × Iℓd for ℓ1 , . . . , ℓd ∈ {0, . . . , m}. Overall, for any ε > 0,
setting δ = (m + 1)−d/2 ε establishes the claim. We conclude that if we fix any ε > 0, take
η = ε2 /(4d) ∧ 1/4 and take a single bracket consisting of the constant functions 0 and 1 on
[0, 1]d \ [η, 1]d , we have
+
+
log N[ ] ε, Fd,↓
∩ B2 (r) ∩ B∞ (1),k · kL2 (P ) ≤ log N[ ] ε/2, Fd,↓
∩ B2 (r), k · kL2 (P ;[η,1]d )
(
2 M0
4
2 r log+ (1/ε)
0
log
(
)
log
(1/ε)
log
if d = 2,
(r/ε)2 M
+
+
m
m
ε
0
0
.d
2
d
M
(r/ε)2(d−1) ( m00 )d−1 log+ (1/ε)
if d ≥ 3,
completing the proof.
+
For 0 < r < 1, let Fr be the envelope function of Fd,↓
∩B2 (r, P )∩B∞ (1). The lemma below
, ℓ]
controls the L2 (P ) norm of Fr when restricted to strips of the form Iℓ := [0, 1]d−1 × [ ℓ−1
n1 n1
for ℓ = 1, . . . , n1 .
Lemma 9. For any r ∈ (0, 1] and ℓ = 1, . . . , n1 , we have
Z
Iℓ
Fr2 dP
7M0 r 2 logd+ (1/r 2 )
.
≤
m0 ℓ
Proof. By monotonicity and the L2 (P ) and L∞ constraints, we have Fr2 (x1 , . . . , xd ) ≤
r2
∧ 1. We first claim that for any d ∈ N,
m0 x1 ···xd
Z
[0,1]d
t
∧ 1 dx1 · · · dxd ≤ 5t logd+ (1/t).
x1 · · · xd
31
R
Q
t
dx1 · · · dxd
To see this, we define Sd := (x1 , . . . , xd ) : dj=1 xj ≥ t and set ad := Sd x1 ···x
d
R
and bd := Sd dx1 · · · dxd . By integrating out the last coordinate, we obtain the following
relation
Z
t
1−
bd =
dx1 · · · dxd−1 = bd−1 − ad−1 .
(47)
x1 · · · xd−1
Sd−1
On the other hand, we have by direct computation that
Z 1
Z 1
t
dxd · · · dx1
ad =
···
t
x
·
·
·
x
1
d
t
x ···x
1
d−1
≤ ad−1 log(1/t) ≤ · · · ≤ a1 logd−1 (1/t) = t logd (1/t).
(48)
Combining (47) and (48), we have
Z
t
∧ 1 dx1 · · · dxd = ad + 1 − bd ≤ min{ad + 1, ad + ad−1 + · · · + a1 + 1 − b1 }
[0,1]d x1 · · · xd
t logd+1 (1/t)
d
≤ min t log (1/t) + 1,
≤ 5t logd+ (1/t),
log(1/t) − 1
as claimed, where the final inequality follows by considering the cases t ∈ [1/e, 1], t ∈
[1/4, 1/e) and t ∈ [0, 1/4) separately. Consequently, for ℓ = 2, . . . , n1 , we have that
Z
Iℓ
Fr2
M0
dP ≤
m0
≤
M0
m0
Z
ℓ/n1
(ℓ−1)/n1
ℓ/n1
Z
(ℓ−1)/n1
Z
[0,1]d−1
r 2 /xd
∧ 1 dx1 · · · dxd−1 dxd
x1 · · · xd−1
2
5(r 2 /xd ) logd−1
+ (xd /r ) dxd
2
7M0 r 2 logd−1
M0 2 d−1
+ (1/r )
2
5r log+ (1/r ) log ℓ/(ℓ − 1) ≤
,
≤
m0
m0 ℓ
as desired. For the remaining case ℓ = 1, we have
Z
Z
5M0 2 d
2
Fr dP ≤ M0
Fr2 dx1 · · · dxd ≤
r log+ (1/r 2 ),
m0
[0,1]d
I1
which is also of the correct form.
The following is a simple tail bound for kfˆn k∞ .
Lemma 10. For f0 = 0, we have
P kfˆn k∞ ≥ 4 log1/2 n ≤ 2n−7 .
Proof. Recall that we say U ⊆ Rd is an upper set if whenever x ∈ U and x y, we
have y ∈ U; we say, L ⊆ Rd is a lower set if −L is an upper set. We write U and L
respectively for the collections of upper and lower sets in Rd . The least squares estimator
32
fˆn over Fd then has a well-known min-max representation (Robertson, Wright and Dykstra,
1988, Theorem 1.4.4):
fˆn (Xi ) = min
max YL∩U ,
L∈L,L∋Xi U ∈U ,U ∋Xi
where YL∩U denotes the average value of the elements of {Y1 , . . . , Yn } ∩ L ∩ U, with the
convention that YL∩U = 0 if {Y1, . . . , Yn } ∩ L ∩ U = ∅. Then
sup fˆn (x) = max fˆn (Xi ) ≤ max Yi .
x∈[0,1]d
1≤i≤n
1≤i≤n
Since f0 = 0, we have Yi = ǫi , which means that
1/2
1/2
ˆ
P sup fn (x) ≥ 4 log n ≤ P max ǫi ≥ 4 log n ≤ ne−8 log n = n−7 .
x∈[0,1]d
1≤i≤n
The desired result follows by observing that a similar inequality holds for inf x∈[0,1]d fˆn (x).
Acknowledgements: The research of the first author is supported by in part by NSF
Grant DMS-1566514. The research of the second and fourth authors is supported by EPSRC
fellowship EP/J017213/1 and a grant from the Leverhulme Trust RG81761.
References
Amelunxen, D., Lotz, M., McCoy, M. B. and Tropp, J. A. (2014) Living on the edge: phase
transition in convex programs with random data. Inf. Inference, 3, 224–294.
Bacchetti, P. (1989) Additive isotonic models. J. Amer. Statist. Assoc., 84, 289–294.
Barlow, R. E., Bartholomew, D. J., Bremner, J. M. and Brunk, H. D. (1972) Statistical
Inference under Order Restrictions. Wiley, New York.
Bellec, P. C. (2017) Sharp oracle inequalities for least squares estimators in shape restricted
regression. Ann. Statist., to appear.
Birgé, L. and Massart, P. (1993) Rates of convergence for minimum contrast estimators.
Probab. Theory Related Fields, 97, 113–150.
Boucheron, S., Lugosi, G. and Massart, P. (2013) Concentration Inequalities. Oxford University Press, Oxford.
Brunk, H. D. (1955) Maximum likelihood estimates of monotone parameters. Ann. Math.
Statist., 26, 607–616.
Chatterjee, S. (2014) A new perspective on least squares under convex constraint. Ann.
Statist., 42, 2340–2381.
Chatterjee, S., Guntuboyina, A. and Sen, B. (2015) On risk bounds in isotonic and other
shape restricted regression problems. Ann. Statist., 43, 1774–1800.
33
Chatterjee, S., Guntuboyina, A. and Sen, B. (2017) On matrix estimation under monotonicity constraints. Bernoulli, to appear.
Chatterjee, S. and Lafferty, J. (2017) Adaptive risk bounds in unimodal regression. arXiv
preprint, arxiv:1512.02956v5.
Chen, Y. and Samworth, R. J. (2016) Generalized additive and index models with shape
constraints. J. Roy. Statist. Soc., Ser. B, 78, 729–754.
Dilworth, R. P. (1950) A decomposition theorem for partially ordered sets. Ann. Math., 51,
161–166.
Donoho, D. (1991) Gelfand n-widths and the method of least squares. Technical Report,
University of California.
Durot, C. (2007) On the Lp -error of monotonicity constrained estimators. Ann. Statist., 35,
1080–1104.
Durot, C. (2008) Monotone nonparametric regression with random design. Math. Methods
Statist., 17, 327–341.
Dykstra, R. L. and Robertson, T. (1982) An algorithm for isotonic regression for two or
more independent variables. Ann. Statist., 10, 708–716.
Dykstra, R. L. (1983) An algorithm for restricted least squares regression. J. Amer. Statist.
Assoc., 78, 837–842.
Eichler, E. E., Flint, J., Gibson, G., Kong, A., Leal, S. M., Moore, J. H. and Nadeau, J.
H. (2010) Missing heritability and strategies for finding the underlying causes of complex
disease. Nat. Rev. Genet., 11, 446–450.
Elena, S. F. and Lenski R. E. (1997) Test of synergistic interactions among deleterious
mutations in bacteria. Nature, 390, 395–398.
Gao, F. and Wellner, J. A. (2007) Entropy estimate for high-dimensional monotonic functions. J. Mult. Anal., 98, 1751–1764.
Giné, E., Latala, R and Zinn, J. (2000) Exponential and moment inequalities for U-statistics.
In High dimensional probability, II (Seattle, WA, 1999), volume 47 of Progr. Probab.,
pp. 13–38. Birkhäuser Boston, Boston, MA, 2000.
Goldstein, D. B. (2009) Common genetic variation and human traits. N. Engl. J. Med., 360,
1696–1698.
Groeneboom, P. and Jongbloed, G. (2014) Nonparametric Estimation under Shape Constraints. Cambridge University Press, Cambridge.
Guntuboyina, A. and Sen, B. (2015) Global risk bounds and adaptation in univariate convex
regression. Probab. Theory Related Fields, 163, 379–411.
34
Han, Q. and Wellner, J. A. (2017) A sharp multiplier inequality with applications ot heavytailed regression problems. arXiv preprint, arxiv:1706.02410.
Kim, A. K. H., Guntuboyina, A. and Samworth, R. J. (2017) Adaptation in log-concave
density estimation. Ann. Statist., to appear.
Kim, A. K. H. and Samworth, R. J. (2017) Global rates of convergence in log-concave density
estimation. Ann. Statist., 44, 2756–2779.
Kyng, R., Rao, A. and Sachdeva, S. (2015) Fast, provable algorithms for isotonic regression
in all ℓp -norms. In Advances in Neural Information Processing Systems, pp. 2719–2727.
Ledoux, M. and Talagrand, M. (2011) Probability in Banach Spaces: Isoperimetry and Processes. Springer-Verlag, Berlin.
Luss, R., Rosset, S. and Shahar, M. (2012) Efficient regularized isotonic regression with
application to gene-gene interaction search. Ann. Appl. Statist., 6, 253–283.
Mammen, E. and Yu, K. (2007) Additive isotonic regression. In Cator, E. A. et al. (Eds.),
Asymptotics: Particles, Processes and Inverse Problems, 179–195. Institute of Mathematical Statistics, Beachwood.
Mani, R., Onge, R. P. S., Hartman, J. L., Giaever, G. and Roth, F. P. (2007) Defining
genetic interaction. Proc. Nat. Acad. Sci. USA, 105, 3461–3466.
Massart, P. (2000) About the constant in Talagrand’s concentration inequalities for empirical
processes. Ann. Probab., 28, 863–884.
Massart, P. (2007) Concentration Inequalities and Model Selection, Springer, Berlin.
Meyer, M. and Woodroofe, M. (2000) On the degrees of freedom in shape-restricted regression. Ann. Statist., 28, 1083–1104.
Morton-Jones, T., Diggle, P., Parker, L., Dickinson, H. O. and Binks, K. (2000) Additive
isotonic regression models in epidemiology. Statist. Med., 19, 849–859.
Pisier, G. (1999) The Volume of Convex Bodies and Banach Space Geometry, Cambridge
University Press, Cambridge.
Pollard, D. (2002) A User’s Guide to Measure Theoretic Probability. Cambridge University
Press, Cambridge.
Rakhlin, A., Sridharan, K. and Tsybakov, A. B. (2017) Empirical entropy, minimax regret
and minimax risk. Bernoulli, 23, 789–824.
Robertson, T., Wright, F. T. and Dykstra, R. L. (1988) Order Restricted Statistical Inference.
John Wiley & Sons, Ltd., Chichester.
Romik, D. (2014) The Surprising Mathematics of Longest Increasing Subsequences. Cambridge University Press, Cambridge.
35
Roth, F. P., Lipshitz, H. D. and Andrews, B. J. (2009) Q&A: epistasis. J. Biol., 8, 35.
Sanjuan, R. and Elena, S. F. (2006) Epistasis correlates to genomic complexity. Proc. Natl.
Acad. Sci. USA, 103, 14402–14405.
Shao, H., Burrage, L. C., Sinasac, D. S., Hill, A. E., Ernest, S. R., O’Brien, W., Courtland,
H.-W., Jepsen, K. J., Kirby, A., Kulbokas, E. J., Daly, M. J., Broman, K. W., Lander,
E. S. and Nadeau, J. H. (2008) Genetic architecture of complex traits: Large phenotypic
effects and pervasive epistasis. Proc. Natl. Acad. Sci. USA, 105, 19910–19914.
Schell, M. J. and Singh, B. (1997) The reduced monotonic regression method. J. Amer.
Statist. Assoc., 92, 128–135.
Stout, Q. F. (2015) Isotonic Regression for Multiple Independent Variables. Algorithmica,
71, 450–470.
Talagrand, M. (1996) New concentration inequalities in product spaces. Invent. Math., 126,
505–563.
Tong, A. H., Evangelista, M., Parsons, A. B., Xu, H., Bader, G. D., Pagé, N., Robinson, M.,
Raghibizadeh, S., Hogue, C. W. V., Bussey, H., Andrews, B., Tyers, M. and Boone, C.
(2001) Systematic genetic analysis with ordered arrays of yeast deletion mutants. Science,
294, 2364–2368.
van de Geer, S. A. (1990) Estimating a regression function. Ann. Statist. 18, 907–924.
van de Geer, S. A. (1993) Hellinger-consistency of certain nonparametric maximum likelihood
estimators. Ann. Statist. 21, 14–44.
van de Geer, S. A. (2000) Applications of Empirical Process Theory. Cambridge University
Press, Cambridge.
van der Vaart, A. W. and Wellner, J. A. (1996) Weak Convergence and Empirical Processes.
Springer, New York.
van Eeden, C. (1958) Testing and estimating ordered parameters of probability distributions.
Mathematical Centre, Amsterdam.
Yang, F. and Barber, R. F. (2017) Uniform convergence of isotonic regression. arXiv preprint,
arxiv:1706.01852.
Yang, Y. and Barron, A. (1999). Information-theoretic determination of minimax rates of
convergence. Ann. Statist., 27, 1564–1599.
Yu, B. (1997) Assouad, Fano and Le Cam. In Pollard, D., Torgersen, E. and Yang G. L.
(Eds.) Festschrift for Lucien Le Cam: Research Papers in Probability and Statistics, 423–
435. Springer, New York.
Zhang, C.-H. (2002) Risk bounds in isotonic regression. Ann. Statist., 30, 528–555.
36
| 10math.ST
|
Path Consistency Learning in Tsallis Entropy Regularized MDPs
Ofir Nachum * 1 Yinlam Chow * 2 Mohamamd Ghavamzadeh * 2
arXiv:1802.03501v1 [cs.AI] 10 Feb 2018
Abstract
We study the sparse entropy-regularized reinforcement learning (ERL) problem in which the
entropy term is a special form of the Tsallis entropy. The optimal policy of this formulation is
sparse, i.e., at each state, it has non-zero probability for only a small number of actions. This
addresses the main drawback of the standard
Shannon entropy-regularized RL (soft ERL) formulation, in which the optimal policy is softmax,
and thus, may assign a non-negligible probability
mass to non-optimal actions. This problem is aggravated as the number of actions is increased. In
this paper, we follow the work of Nachum et al.
(2017) in the soft ERL setting, and propose a
class of novel path consistency learning (PCL)
algorithms, called sparse PCL, for the sparse
ERL problem that can work with both on-policy
and off-policy data. We first derive a sparse consistency equation that specifies a relationship between the optimal value function and policy of
the sparse ERL along any system trajectory. Crucially, a weak form of the converse is also true,
and we quantify the sub-optimality of a policy
which satisfies sparse consistency, and show that
as we increase the number of actions, this suboptimality is better than that of the soft ERL optimal policy. We then use this result to derive the
sparse PCL algorithms. We empirically compare
sparse PCL with its soft counterpart, and show
its advantage, especially in problems with a large
number of actions.
1. Introduction
In reinforcement learning (RL), the goal is to find a policy with maximum long-term performance, defined as the
sum of discounted rewards generated by following the policy (Bertsekas & Tsitsiklis, 1996; Sutton & Barto, 1998).
In case the number of states and actions are small, and the
*
Equal contribution 1 Google Brain 2 Google DeepMind. Correspondence to: Yinlam Chow <[email protected]>.
Proceedings of the 35 th International Conference on Machine
Learning, Stockholm, Sweden, PMLR 80, 2018. Copyright 2018
by the author(s).
model is known, the optimal policy is the solution of the
non-linear Bellman optimality equations (Bellman, 1957).
When the system is large or the model is unknown, greedily solving the Bellman equations often results in policies
that are far from optimal. A principled way of dealing
with this issue is regularization. Among different forms
of regularization, such as `2 (e.g., Farahmand et al. 2008;
2009) and `1 (e.g., Kolter & Ng 2009; Johns et al. 2010;
Ghavamzadeh et al. 2011), entropy regularization is among
the most studied in both value-based (e.g., Kappen 2005;
Todorov 2006; Ziebart 2010; Azar et al. 2012; Fox et al.
2016; O’Donoghue et al. 2017; Asadi & Littman 2017)
and policy-based (e.g., Peters et al. 2010; Todorov 2010)
RL formulations. In particular, two of the most popular
deep RL algorithms, TRPO (Schulman et al., 2015) and
A3C (Mnih et al., 2016), are based on entropy-regularized
policy search. We refer the interested readers to Neu et al.
(2017), for an insightful discussion on entropy-regularized
RL algorithms and their connection to online learning.
In entropy-regularized RL (ERL), an entropy term is added
to the Bellman equation. This formulation has four main
advantages: 1) it softens the non-linearity of the Bellman
equations and makes it possible to solve them more easily, 2) the solution of the softened problem is quantifiably
not much worse than the optimal solution in terms of accumulated return, 3) the addition of the entropy term brings
nice properties, such as encouraging exploration (Shannon
entropy) (e.g., Fox et al. 2016; Nachum et al. 2017) and
maintaining a close distance to a baseline policy (relative
entropy) (e.g., Schulman et al. 2015; Nachum et al. 2018),
and 4) unlike the original problem that has a deterministic
solution, the solution to the softened problem is stochastic, which is preferable in problems in which exploration or
dealing with unexpected situations is important. However,
in the most common form of ERL, in which a Shannon
(or relative) entropy term is added to the Bellman equations, the optimal policy is of the form of softmax. Despite the advantages of a softmax policy in terms of exploration, its main drawback is that at each step, it assigns a
non-negligible probability mass to non-optimal actions, a
problem that is aggravated as the number of actions is increased. This may result in policies that may not be safe
to execute. To address this issue, Lee et al. (2018) proposed to add a special form of a general notion of entropy,
Path Consistency Learning in Tsallis Entropy Regularized MDPs
called Tsallis entropy (Tsallis, 1988), to the Bellman equations. This formulation has the property that its solution has
sparse distributions, i.e., at each state, only a small number of actions have non-zero probability. Lee et al. (2018)
studied the properties of this ERL formulation, proposed
value-based algorithms (fitted Q-iteration and Q-learning)
to solve it, and showed that although it is harder to solve
than its soft counterpart, it potentially has a solution closer
to that of the original problem.
In this paper, we propose novel path consistency learning (PCL) algorithms for the Tsallis ERL problem, called
sparse PCL. PCL is a class of actor-critic type algorithms
developed by Nachum et al. (2017) for the soft (Shannon
entropy) ERL problem. It uses a nice property of soft ERL,
namely the equivalence of consistency and optimality, and
learns parameterized policy and value functions by minimizing a loss that is based on the consistency equation of
soft ERL. The most notable feature of soft PCL is that it
can work with both on-policy (sub-trajectories generated
by the current policy) and off-policy (sub-trajectories generated by a policy different than the current one, including
any sub-trajectory from the replay buffer) data. We first derive a multi-step consistency equation for the Tsallis ERL
problem, called sparse consistency. We then prove that in
this setting, while optimality implies consistency (similar
to the soft case), unlike the soft case, consistency only implies sub-optimality. We then use the sparse consistency
equation and derive PCL algorithms that use both on-policy
and off-policy data to solve the Tsallis ERL problem. We
empirically compare sparse PCL with its soft counterpart.
As expected, we gain from using the sparse formulation
when the number of actions is large, both in algorithmic
tasks and in discretized continuous control problems.
2. Markov Decision Processes (MDPs)
We consider the reinforcement learning (RL) problem in
which the agent’s interaction with the system is modeled as
a MDP. A MDP is a tuple M = (X , A, r, P, P0 , γ), where
X and A are state and action spaces; r : X × A → R
and P : X × A → ∆X are the reward function and transition probability distribution, with r(x, a) ∈ [0, Rmax ] and
P (·|x, a) being the reward and the next state probability of
taking action a in state x; P0 : X → ∆X is the initial state
distribution; and γ ∈ [0, 1) is a discounting factor. In this
paper, we assume that the action space is finite, but can be
large. The goal in RL is to find a stationary Markovian policy, i.e., a mapping from state and action spaces to a simplex over the actions µ : X × A → ∆A , that maximizes
the expected discounted sum of rewards, i.e.,
max E
∞
X
µ
s.t.
t
γ r(xt , at )
(1)
t=0
∀x
X
a∈A
µ(a|x) = 1,
∀x, a µ(a|x) ≥ 0,
where x0 ∼ P0 , at ∼ µ(·|xt ), and xt+1 ∼ P (·|xt , at ).
For a given policy µ, we define its value and action-value
functions as
∞
X
V µ (x) = E
γ t r(xt , at )|x0 = x, µ, P ,
t=0
∞
X
Qµ (x, a) = E
γ t r(xt , at )|x0 = x, a0 = a, µ, P .
t=0
Any solution of the optimization problem (1) is called an
optimal policy and is denoted by µ∗ . Note that while a
MDP may have several optimal policies, it only has a sin∗
gle optimal value function V ∗ = V µ . It has been proven
that (1) has a solution in the space of deterministic policies,
i.e., Πd = {µ : µ : X → A}, which can be obtained as
the greedy action w.r.t. the optimal action-value function,
i.e., µ∗ (x) ∈ arg maxa Q∗ (x, a) (Puterman, 1994; Bertsekas & Tsitsiklis, 1996). The optimal action-value function Q∗ is the unique solution of the non-linear Bellman
optimality equations, i.e., for all x ∈ X and a ∈ A,
X
Q(x, a) = r(x, a)+γ
P (x0 |x, a) max
Q(x0 , a0 ). (2)
0
a ∈A
x0 ∈X
Any optimal policy µ∗ and the optimal state and stateaction value functions, V ∗ and Q∗ , satisfy the following
equations for all states and action,
X
Q∗ (x, a) = r(x, a) + γ
P (x0 |x, a)V ∗ (x0 ),
x0 ∈X
V ∗ (x) = max Q∗ (x, a),
a∈A
µ∗ (x) ∈ arg max Q∗ (x, a).
a∈A
3. Entropy Regularized MDPs
As discussed in Section 2, finding an optimal policy for
a MDP involves solving a non-linear system of equations
(see Eq. 2), which is often complicated. Moreover, the optimal policy may be deterministic, always selecting the same
optimal action at a state even when there are several optimal actions in that state. This is undesirable when it is important to explore and to deal with unexpected situations.
In such cases, one might be interested in multimodal policies that still have good performance. This is why many
researchers have proposed to add a regularizer in the form
of an entropy term to the objective function (1) and solve
the following entropy-regularized optimization problem
∞
hX
i
max E
γ t r(xt , at ) + αH µ (xt , at )
µ
s.t.
(3)
t=0
∀x
X
µ(a|x) = 1,
∀x, a µ(a|x) ≥ 0,
a∈A
where H µ (x, a) is an entropy-related term and α is the regularization parameter. The entropy term smoothens the objective function (1) such that the resulting problem (3) is often easier to solve than the original one (1). This is another
reason for the popularity of entropy-regularized MDPs.
Path Consistency Learning in Tsallis Entropy Regularized MDPs
3.1. Entropy Regularized MDP with Shannon Entropy
4
It is common to use Hsfµ (xt , at ) = − log µ(at |xt ) in
entropy-regularized MDPs (e.g., Fox
et al. 2016;
Nachum
et al. 2017). Note that Hsf (µ) = Eµ Hsfµ (x, a) is the Shannon entropy. Problem (3) with Hsfµ (x, a) can be seen as a
RL problem in which the reward function is the sum of the
original reward function r(x, a) and a term that encourages
exploration.1 Unlike (1), the optimization problem (3) with
Hsfµ has a unique optimal policy µ∗sf and a unique optimal
value Vsf∗ (action-value Q∗sf ) function that satisfy the following equations:
Q∗sf (x, a) = r(x, a) + γ
X
P (x0 |x, a)Vsf∗ (x0 ),
3.2. Entropy Regularized MDP with Tsallis Entropy
To address the issues with the softmax policy, Lee et al.
4
µ
(2018) proposed to use Hsp
(xt , at ) = 21 1 − µ(at |xt )
in entropy-regularized
MDPs. Note that Hsp (µ) =
µ
Eµ Hsp
(x, a) is a special case of a general notion of entropy, called
entropy (Tsallis, 1988), i.e., Sq,k (p) =
P Tsallis
q
k
(1
−
p
),
for
the parameters q = 2 and k = 12 .2
i i
q−1
Similar to the soft MDP problem, the optimization probµ
lem (3) with Hsp
has a unique optimal policy µ∗sp and a
unique optimal value Vsp∗ (action-value Q∗sp ) function that
satisfy the following equations (Lee et al., 2018):
Q∗sp (x, a) = r(x, a) + γ
x0 ∈X
X
P (x0 |x, a)Vsp∗ (x0 ),
x0 ∈X
Vsf∗ (x) = α · sfmax Q∗sf (x, ·)/α ,
exp Q∗sf (x, a)/α
,
µ∗sf (a|x) = P
∗
0
a0 ∈A exp Qsf (x, a )/α
(4)
where for any function f : X × A → R,P
the sfmax operator
is defined as sfmax f (x, ·) = log
a exp f (x, a) .
Note that the equations in (4) are derived from the KKT
conditions of (3) with Hsfµ . In this case, the optimal policy
is soft-max, with the regularization parameter α playing the
role of its temperature (see Eq. 4). This is why (3) with Hsfµ
is called the soft MDP problem. In soft MDPs, the optimal
value function Vsf∗ is the unique solution of the soft Bellman
optimality equations, i.e., ∀x ∈ X , ∀a ∈ A,
Vsp∗ (x) = α · spmax Q∗sp (x, ·)/α ,
+
,
µ∗sp (a|x) = Q∗sf (x, a)/α − G Q∗sf (x, ·)/α
where (·)+ = max(·, 0), and for any function f : X ×A →
R, the spmax operator is defined as
f (x, ·) 2 i
X f (x, a) 2
1h
1+
−G
,
spmax f (x, ·) =
2
α
α
a∈S(x)
in which
P
G f (x, ·) =
a∈S(x)
f (x, a) − 1
|S(x)|
f (x,a
V (x) = α · sfmax
r(x, ·) + γ
X
P (x |x, ·)V (x ) /α . (5)
0
0
x0
Note that the sfmax operator is a smoother function of its
inputs than the max operator associated with the Bellman
optimality equation (2). This means that solving the soft
MDP problem is easier than the original one, with the cost
that its optimal policy µ∗sf performs worse than the optimal policy of the original MDP µ∗ . This difference can be
quantified as
∀x ∈ X
V ∗ (x) −
∗
α
log(|A|) ≤ V µsf (x) ≤ V ∗ (x), (6)
1−γ
where we discriminate between the value function of a policy µ in the soft Vsfµ and original V µ MDPs. Note that the
sub-optimality of µ∗sf is unbounded as |A| → ∞. This is
the main drawback of using softmax policies; in large action space problems, at each step, the policy assigns a nonnegligible probability mass to non-optimal actions, which
in aggregate can be detrimental to its reward performance.
1
Another entropy term that has been studied in the literature
(7)
)
and S(x) is the set of actions satisfying 1 + i α (i) >
Pi f (x,a(j) )
, where a(i) indicates the action with the ith
j=0
α
largest value of f (x, a). Note that the equations in (7) are
µ
derived from the KKT conditions of (3) with Hsp
. In this
case, the optimal policy may have zero probability for sevµ
eral actions (see Eq. 7). This is why (3) with Hsp
is called
the sparse MDP problem. The regularization parameter
α controls the sparsity of the resulted policy. The policy
would be more sparse for smaller values of α. In sparse
MDPs, the optimal value function Vsp∗ is the unique fixedpoint of the sparse Bellman optimality operator Tsp (Lee
et al., 2018) that for any function f : X → R is defined as
X
P (x0 |x, ·)f (x0 ) /α .
(Tsp f )(x) = α · spmax r(x, ·) + γ
x0
(8)
Similar to (5), the spmax operator is a smoother function of
its inputs than the max, and thus, solving the sparse MDP
problem is easier than the original one, with the cost that its
optimal policy µ∗sp performs worse than the optimal policy
of the original MDP µ∗ . This difference can be quantified
as (Lee et al., 2018),
4
µ
t |xt )
is Hrel
(xt , at ) = − log µµ(a
, where πb is a baseline policy.
t |xt )
b (a
µ
Note that Hrel (µ) = Eµ Hrel (x, a) is the relative entropy. Probµ
lem (3) with Hrel
(x, a) can be seen as a RL problem in which the
reward function is the sum of the original reward function r(x, a)
and a term that penalizes deviation from the baseline policy πb .
∀x ∈ X
V ∗ (x) −
∗
|A| − 1
α
·
≤ V µsp (x) ≤ V ∗ (x). (9)
1−γ
2|A|
2
Note that the Shannon entropy is a special case of the Tsallis
entropy for the parameters q = k = 1 (Tsallis, 1988).
Path Consistency Learning in Tsallis Entropy Regularized MDPs
On the other hand, the spmax operator is more complex
than sfmax, and thus, it is slightly more complicated to
solve the sparse MDP problem than its soft counterpart.
However, as can be seen from Eqs. 6 and 9, the optimal policy of the sparse MDP, µ∗sp , can have a better performance
than its soft counterpart, µ∗sf , and this difference becomes
more apparent as the number of actions |A| grows. For
large action size, the term (|A| − 1)/(2|A|) in (9) turns to
a constant, while log |A| in (6) grows unbounded.
4. Path Consistency Learning in Soft MDPs
A nice property of soft MDPs that was elegantly used
by Nachum et al. (2017) is that any policy µ and function
V : X → R that satisfy the (one-step) consistency equation, i.e., for all x ∈ X and for all a ∈ A,
V (x) = r(x, a)−α log µ(a|x)+γ
X
P (x0 |x, a)V (x0 ), (10)
x0 ∈X
µ∗sf
are optimal, i.e., µ =
and V = Vsf∗ (consistency implies
optimality). Due to the uniqueness of the optimal policy
in soft MDPs, the reverse is also true, i.e., the optimal policy µ∗sp and the value function Vsp∗ satisfy the consistency
equation (optimality implies consistency).
As shown in Nachum et al. (2017), the (one-step) consistency equation (10) can be easily extended to multi-step,
i.e., any policy µ and function V : X → R that for any
state x0 and sequence of actions a0 , . . . , ad−1 , satisfy the
multi-step consistency equation
h
V (x0 ) = Ex1:d |x0 ,a0:d−1 γ d V (xd )
+
d−1
X
γ t r(xt , at ) − α log µ(at |xt )
(11)
i
t=0
are optimal, i.e., µ = µ∗sf and V = Vsf∗ .
The property that both single and multiple step consistency
equations imply optimality (Eqs. 10 and 11) was the motivation of a RL algorithm by Nachum et al. (2017), path
consistency learning (PCL). The main idea of (soft) PCL is
to learn a parameterized policy and value function by minimizing the following objective function:
J (θ, φ) =
1X
J(ξi , θ, φ)2 ,
2
ξi
where ξ = (x0 , a0 , r0 , . . . , xd−1 , ad−1 , rd−1 , xd ) is any dlength sub-trajectory, θ and φ are the policy and value function parameters, respectively, and
J(ξ, θ, φ) = −Vφ (x0 ) + γ d Vφ (xd )
+
d−1
X
(12)
γ t r(xt , at ) − α log µθ (at |xt ) .
t=0
An important property of the soft PCL algorithm is that
since the multi-step consistency (11) holds for any d-length
sub-trajectory, it can use both on-policy (ξ’s generated by
the current policy µθ ) and off-policy data, i.e., ξ’s generated
by a policy different than the current one, including any dlength sub-trajectory from the replay buffer.
Note that since both optimal policy µ∗sf and value function Vsf∗ can be written based on the optimal action-value
function Q∗sf (see Eq. 4), we may write the objective function (12) based on Qψ , and optimize only one set of parameters ψ, instead of separate θ and φ.
5. Consistency between Optimal Value &
Policy in Sparse MDPs
This section begins the main contributions of our work.
We first identify a (one-step) consistency equation for the
sparse MDPs defined by (3). We then prove the relationship
between the sparse consistency equation and the optimal
policy and value function of the sparse MDP, and highlight
its similarities and differences with that in soft MDPs, discussed in Section 4. We then extend the sparse consistency
equation to multiple steps and prove results that allow us
to use the multi-step sparse consistency equation to derive
on-policy and off-policy algorithms to solve sparse MDPs,
which we fully describe in Section 6. The significance
of the sparse consistency equation is in providing an efficient tool for computing a near-optimal policy for sparse
MDPs, which only involves solving a set of linear equations and linear complementary constraints, as opposed to
(iteratively) solving the fixed-point of the non-linear sparse
Bellman operator (8). We report the proofs of all the theorems of this section in Appendix A.
For any policy µ and value function V : X → R, we define
the (one-step) consistency equation of the sparse MDPs as,
for all state x ∈ X and for all actions a ∈ A,
α
V (x) = r(x, a) + − αµ(a|x) + λ(a|x) − Λ(x)
2
X
+γ
P (x0 |x, a)V (x0 ),
(13)
x0
where λ : X × A → R+ and Λ : X → R− are Lagrange multipliers, such that λ(a|x) · µ(a|x) = 0 and
− α2 ≤ Λ(x) ≤ 0. We call this the one-step sparse consistency equation and it is the equivalent of Eq. 10 in soft
MDPs.
We now present a theorem which states that, similar to soft
MDPs, optimality in sparse MDPs is a necessary condition
for consistency, i.e., optimality implies consistency.
Theorem 1. The optimal policy µ∗sp and value function Vsp∗
of the sparse MDP (3) satisfy the consistency equation (13).
Theorem 2 shows that in the sparse MDPs, consistency
only implies near-optimality, as opposed to optimality in
the case of soft MDPs.
Path Consistency Learning in Tsallis Entropy Regularized MDPs
Theorem 2. Any policy µ that satisfies the consistency
equation (13) is α/(1 − γ)-optimal in the sparse MDP (3),
i.e., for each state x ∈ X , we have
α
Vspµ (x) ≥ Vsp∗ (x) −
.
(14)
1−γ
This result indicates that for a fixed γ, as α decreases, a
policy µ satisfying the one-step consistency equations approaches the true optimal µ∗sp . To connect the performance
of µ to the original goal of maximizing expected return,
we present the following corollary, which is a direct consequence of Theorem 2 and the results reported in Section 3.2
on the performance of µ∗sp in the original MDP.
Corollary 1. Any policy µ that satisfies the consistency
1
α
) · 1−γ
-optimal in the original
equation (13) is ( 32 − |A|
MDP (1), i.e., for each state x ∈ X , we have
3
1
α
∗
V (x) −
−
·
≤ V µ (x) ≤ V ∗ (x).
2 |A|
1−γ
We now extend the (one-step) sparse consistency equation (13) to multiple steps. For any state x0 ∈ X and sequence of actions a0 , . . . , ad−1 , define the multi-step consistency equation for sparse MDPs as
h
V (x0 ) = Ex1:d |x0 ,a0:d−1 γ d V (xd )
+
d−1
X
t=0
γ t r(xt , at ) +
(15)
i
α
− αµ(at |xt ) + λ(at |xt ) − Λ(xt ) ,
2
where λ : X × A → R+ and Λ : X → R− are Lagrange multipliers, such that λ(a|x) · µ(a|x) = 0 and
− α2 ≤ Λ(x) ≤ 0. We call this multi-step sparse consistency equation, the equivalent of Eq. 11 in soft MDPs.
From Theorem 1, we can immediately show that multi-step
sparse consistency is a necessary condition of optimality.
Corollary 2. The optimal policy µ∗sp and value function
Vsp∗ of the sparse MDP (3) satisfy the multi-step consistency
equation (15).
Proof. The proof follows directly from Theorem 1, by repeatedly applying the expression in (13) over the trajectory
ξ, taking the expectation over the trajectory, and using telescopic cancellation of the value function of intermediate
states.
Conversely, followed from Theorem 2, we prove the following result on the performance of any policy satisfying
the mutli-step consistency equation. This is a novel result
showing that solving the multi-step consistency equation is
indeed sufficient to guarantee near-optimality.
Corollary 3. Any policy µ that satisfies the multi-step consistency equation (15) is α/(1 − γ)-optimal in the sparse
MDP (3).
Proof. Consider the multi-step consistency equation (15).
Since it is true for any initial state x0 and sequence of actions a0:d−1 , unrolling it for another d steps starting at state
xd and using the action sequence ad:2d−1 yields
h
V (x0 ) = Ex1:2d |x0 ,a0:2d−1 γ 2d V (x2d )
+
2d−1
X
γ t r(xt , at ) +
t=0
i
α
− αµ(at |xt ) + λ(at |xt ) − Λ(xt ) .
2
Note that this process can be repeated for an arbitrary
number of times (say k times), and also note that as V
is a bounded function, one has limk→∞ γ kd V (xkd ) = 0.
Therefore, by further unrolling, we obtain
V (x0 ) = Ex1:∞ |x0 ,a0:∞
∞
hX
γ t r(xt , at ) +
t=0
α
− αµ(at |xt )
2
i
+ λ(at |xt ) − Λ(xt ) .
Followed from the Banach fixed-point theorem (Bertsekas
& Tsitsiklis, 1996), one can show that the solution pair
(V, µ) is also a solution to the one-step consistency condition in (13), i.e.,PV (x) = r(x, a) + α2 − αµ(a|x) +
λ(a|x)−Λ(x)+γ x0 ∈X P (x0 |x, a)V (x0 ), for any x ∈ X
and a ∈ A. Thus the α/(1 − γ)-optimality performance
guarantee of µ is implied by Theorem 2.
Equipped with the above results on the relationship between (near)-optimality and multi-step consistency in
sparse MDPs, we are now ready to present our off-policy
RL algorithms to solve the sparse MDP (3).
6. Path Consistency Learning in Sparse MDPs
Similar to the PCL algorithm for soft MDPs, in sparse
MDPs the multi-step consistency equation (15) naturally
leads to a path-wise algorithm for training a policy µθ and
value function Vφ parameterized by θ and φ, as well as Lagrange multipliers Λρ and λθ,ρ parameterized by the auxiliary parameter ρ. To characterize the objective function of
this algorithm, we first define the soft consistency error for
the d-step sub-trajectory ξ as a function of θ, ρ, and φ,
J(ξ; θ,ρ, φ) = −Vφ (x) + γ d Vφ (xd )
+
d−1
X
t=0
γ j r(xt , at ) +
α
− αµθ (at |xt ) + λθ,ρ (at |xt ) − Λρ (xt ) .
2
The goal of our algorithm is to learn Vφ , µθ , λθ,ρ ,
and Λρ , such that the expectation of J(ξ; θ, ρ, φ) for
any initial state x0 and action sequence a0:d−1 is as
close to 0 as possibles. Our sparse PCL algorithm minimizes
objective function Jn (θ, ρ, φ) =
P the empirical
1
2
J(ξ
;
θ,
ρ,
φ)
,
which
converges to J (θ, ρ, φ) =
i
ξi
2
Path Consistency Learning in Tsallis Entropy Regularized MDPs
Ex0 ,a0:d−1 E[J(ξ; θ, ρ, φ)2 | x0 , a0:d−1 ] as i → ∞. By
the Cauchy-Schwarz inequality, J (θ, ρ, φ) is a conserva
2
tive surrogate of E J(ξ; θ, ρ, φ) , which represents the
error of the multi-step consistency equation. This relationship justifies that the solution policy of the sparse PCL algorithm is near-optimal (see Corollary 3). Moreover, the
gradient of J(ξ) w.r.t. the parameters is as follows:
d−1
X
∂J(ξ)
= J(ξ; θ, ρ, φ)
γ t ∇θ λθ,ρ (at |xt ) − αµθ (at |xt ) ,
∂θ
t=0
d−1
X
∂J(ξ)
= J(ξ; θ, ρ, φ)
γ t ∇ρ λθ,ρ (at |xt ) − Λρ (xt ) ,
∂ρ
t=0
∂J(ξ)
= J(ξ; θ, ρ, φ)∇φ Vφ (x0 ) − γ d Vφ (xd ) .
∂φ
We may relate the sparse PCL algorithm to the standard
actor-critic (AC) algorithm (Konda & Tsitsiklis, 2000; Sutton et al., 2000), where ∂J(ξ)/∂θ and ∂J(ξ)/∂φ correspond to the actor and critic updates, respectively. An advantage of sparse PCL over the standard AC is that it does
not need the multi-time-scale update required by AC for
convergence.
While optimizing J(θ, ρ, φ) minimizes the mean square
of the soft consistency error, in order to satisfy the multistep consistency in (15), one still needs to impose the following constraints on Lagrange multipliers into the optimization problem: (i) − α2 ≤ Λρ (x) ≤ 0, and (ii)
λθ,ρ (a|x) · µθ (a|x) = 0, ∀x ∈ X , ∀a ∈ A. One standard
approach is to replace the above constraints with adding
penalty functions (Bertsekas, 1999) to the original objective function Jn . Note that each penalty function is associated with a penalty parameter and there are |X | · |A| + 2|X |
constraints. When |X | and |A| are large, tuning all the
parameters becomes computationally expensive. Another
approach is to update the penalty parameters using gradient ascent methods (Bertsekas, 2014). This is equivalent to
finding the saddle point of the Lagrangian function in the
constrained optimization problem. However, the challenge
is to balance the primal and dual updates in practice.
We hereby describe an alternative and a much simpler
methodology to parameterize the Lagrange multipliers
λθ,ρ (a|x) and Λρ (x), such that the aforementioned constraints are immediately satisfied. Although this method
may impose extra restrictions to the representations of their
function approximations, it avoids the difficulties of directly solving a constrained optimization problem. Specifically, to satisfy the constraint (i), one can parameterize Λρ
with a multilayer perceptron network that has either an activation function of −α/2 · σ(·) or −α/2 · (1 + tanh(·))/2
at its last layer. To satisfy constraint (ii), we consider the
case when µθ is written in form of (fθ (x, a))+ for some
function approximator fθ . This parameterization of µθ is
justified by the closed-form solution policy of the Tsallis
entropy-regularized MDP problem in (7). Specifically, (7)
∗
uses fsp
(x, a) = Q∗sp (x, a)/α − G(Q∗sp (x, ·)/α). Now suppose that λθ,ρ is parameterized as follows: λθ,ρ (a|x) =
+
−fθ (x, a) · Fρ (x, a), where Fρ : X × A → R+ is
an auxiliary function approximator. Then by the property
(x)+ · (−x)+ = 0, constraint (ii) is immediately satisfied.
A pseudo-code of our sparse PCL algorithm can be found
in Algorithm 1 in the Appendix A.
Unified Sparse PCL Note that the closed-form optimal
policy µ∗sp and value function Vsp∗ are both functions of the
optimal state-action value function Q∗sp . As in soft PCL,
based on this observation one can also parameterize both
policy and value function in sparse PCL (see Eq. 7) with
a single function approximator Qψ (x, a). Although consistency does not imply optimality in sparse MDPs (as opposed to the case of soft MDPs), the justification of this parameterization comes from Corollary 2, where the unique
optimal value function and optimal policy satisfy the consistency equation (15). From an actor-critic perspective,
the significance of this is that both policy (actor) and value
function (critic) can be updated simultaneously without affecting the convergence. Accordingly, the update rule for
the model parameter ψ takes the form
d−1
X
∂J(ξ)
= J(ξ; ψ, ρ)
γ t ∇ψ λψ,ρ (at |xt ) −αµψ (at |xt )
∂ψ
t=0
+ ∇ψ Vψ (x0 ) − γ d ∇ψ Vψ (xd ) .
7. Experimental Results
We demonstrate the effectiveness of the sparse PCL algorithm by comparing its performance with that of the soft
PCL algorithm on a number of RL environments available
in the OpenAI Gym (Brockman et al., 2016) environment.
7.1. Discrete Control
Here we compare the performance of these two algorithms
on the following standard algorithmic tasks: 1) Copy, 2)
DuplicatedInput, 3) RepeatCopy, 4) Reverse, and 5) ReversedAddition (see appendix for more details). Each task
can be viewed as a grid environment, where each cell stores
a single character from a finite vocabulary V. An agent
moves on the grid of the environment and writes to output.
At each time step the agent observes the character of the
single cell in which it is located. After observing the character, the agent must take an action of the form (m, w, c),
where m determines the agent’s move to an adjacent cell,
(in 1D environments, m ∈ {left, right}; in 2D environments, m ∈ {left, right, up, down}), w ∈ {0, 1} determines
whether the agent writes to output or not, and c ∈ V determines the character that the agent writes if w = 1 (otherwise c = ∅). Based on this problem setting, the action
Path Consistency Learning in Tsallis Entropy Regularized MDPs
Copy
|A| = 20
|A| = 40
30
30
25
25
25
25
20
20
20
20
15
15
15
15
10
10
10
10
5
5
5
0
0
500
1000
1500
2000
DuplicatedInput
5
0
0
1000
|A| = 20
2000
3000
4000
0
0
|A| = 80
2000
4000
6000
8000
10000
0
|A| = 160
15
15
15
10
10
10
10
5
5
5
5
0
0
500
1000
1500
2000
2500
0
0
1000
|A| = 20
2000
3000
4000
5000
5000
10000
15000
0
|A| = 80
80
80
80
60
60
60
40
40
40
40
20
20
20
20
0
0
0
10000
15000
0
|A| = 16
2000
4000
6000
8000
10000
5000
10000
15000
20000
0
|A| = 64
30
30
30
25
25
25
25
20
20
20
20
15
15
15
15
10
10
10
10
5
5
5
0
5000
10000
15000
20000
5000
10000
15000
20000
Soft PCL
5000
10000
15000
20000
5
0
0
10000 20000 30000 40000 50000
|A| = 128
30
0
20000
0
0
|A| = 32
0
15000
|A| = 160
60
5000
10000
0
0
|A| = 40
80
0
5000
|A| = 320
15
0
RepeatCopy
|A| = 160
30
0
Reverse
|A| = 80
30
0
0
10000 20000 30000 40000 50000
0
10000 20000 30000 40000 50000
Sparse PCL
Figure 1. Results of the average reward from sparse PCL and standard soft PCL during training. Here each row corresponds to a specific
algorithmic task. For each particular task, the action space is increased from left to right across the rows, corresponding to an increase in
difficulty. We observe that soft PCL returns a better solution when the action space is small, but its performance degrades quickly as the
size of action space grows. On the other hand, sparse PCL is not only able to learn good policies in tasks with small action spaces, but,
unlike soft PCL, also successfully learns high-reward policies in the higher-dimension variants. See the appendix for additional results.
space A has size |A| = Θ(|V|). Accordingly, the difficulty of these tasks grows with the size of the vocabulary.
To illustrate the effectiveness of Tsallis entropy-regularized
MDPs in problems with large action space, we evaluate
these two PCL algorithms on 4 different choices of |V|.
In each task, the agent has a different goal. In Copy, the
environment is a 1D sequence of characters and the agent
aims to copy the sequence to output. In DuplicatedInput,
the environment is a 1D sequence of duplicated characters
and the agent needs to write the de-duplicated sequence to
output. In RepeatCopy, the environment is a 1D sequence
of characters in which the agent must copy in forward order, reverse order, and finally forward order again. In Reverse, the environment is a 1D sequence of characters in
which the agent must copy to output in reverse order. Finally, in ReversedAddition, the environment is a 2 × n
grid of digits representing two numbers in base-|V| that the
agent needs to sum. In each task the agent receives a reward of 1 for each correctly output character. The episode
is terminated either when the task is completed, or when
the agent outputs an incorrect character.
We follow a similar experimental procedure as in Nachum
et al. (2017), where the functions V , µ, λ, Λ in the consistency equations are parameterized with a recurrent neural
network with multiple heads. For each task and each PCL
algorithm, we perform a hyper-parameter search to find
the optimal regularization weight α, and the corresponding training curves for average reward are shown in Figure
1. To increase the statistical significance of these experiments, we also train these policies on 5 different Monte
Carlo trials (Notice that these environments are inherently
deterministic, therefore no additional Monte Carlo evaluation is needed.). Details of the experimental setup and extra
numerical results are included in the Appendix.
For each task we evaluated sparse PCL compared to the
original soft PCL on a suite of variants which successively
increase the vocabulary size. For low vocabulary sizes soft
PCL achieves better results. This suggests that Shannon entropy encourages better exploration in small action spaces.
Indeed, in such regimes, a greater proportion of the total actions are useful to explore, and exploration is not as costly.
Therefore, the decreased exploration of the Tsallis entropy
Path Consistency Learning in Tsallis Entropy Regularized MDPs
7.2. Continuous Control
We further evaluate the two PCL algorithms on HalfCheetah, a continuous control problem in the OpenAI gym.
The environment consists of a 6−dimensional action
space, where each dimension corresponds to a torque
of [−1, 1]. Here we discretize each continuous action
with either one of the following grids: {−1, 0, 1} and
{−1, −0.5, 0, 0.5, 1}. Even though the resolution of these
discretization grids is coarse, the corresponding action
spaces are quite large, with sizes of 36 = 729 and 56 =
15625, respectively.
We present the results of sparse PCL and soft PCL on these
discretized problems in Figure 2. Similar to the observations in the algorithmic tasks, here the policy learned by
sparse PCL performs much better than that of soft PCL.
Specifically sparse PCL achieves higher average reward
and is able to learn much faster. To better visualize the
learning progress of these two PCL algorithms in these
problems, at each training step we also compare the average
probability of the most-likely actions across all time-steps
from the on-policy trajectory.3 Clearly, sparse PCL quickly
converges to a near-deterministic policy, while the policy
generated by soft PCL still allocates significant probability
masses to non-optimal actions (as the average probability
of most-likely actions barely ever exceeds 0.75). In environments like HalfCheetah, where the trajectory has a long
horizon (1000 steps), the soft-max policy will in general
suffer because it chooses a large number of sub-optimal actions in each episode for exploration.
Comparing with the performance of other continuous RL
algorithms such as deterministic policy gradient (DPG)
(Silver et al., 2014), we found that the policy generated by
3
Specifically in each iteration we collect a single on-policy
trajectory of 1000 steps. Therefore this metric is an average over
1000 samples of (greedy) action probabilities.
|A| = 56
2500
2500
2000
2000
1500
1500
1000
1000
500
500
0
0
0
Most Likely
Probability
As we increase the vocabulary size (and thus the action
space), the picture changes. We see that the advantage of
soft PCL over sparse PCL decreases until eventually the
order is reversed and sparse PCL begins to show a significant improvement over the standard soft PCL. This supports our original hypothesis. In large action spaces, the
tendency of soft PCL to assign a non-zero probability to
many sub-optimal actions over-emphasizes exploration and
is detrimental to the final reward performance. On the other
hand, sparse PCL is able to handle exploration in large action spaces properly. These empirical results provide evidence for this unique advantage of sparse PCL.
Reward
|A| = 36
may outweigh its asymptotic benefits. The sub-optimality
bounds presented in this paper support this behavior: when
|A| is small, αsoft PCL log(|A|) ≤ 3αsparse PCL /2.
20000
40000
60000
1.0
1.0
0.8
0.8
0.6
0.6
0.4
0.4
0
20000
40000
60000
Soft PCL
0
20000
40000
60000
0
20000
40000
60000
Sparse PCL
Figure 2. Results of sparse PCL and soft PCL in HalfCheetah
with discretized actions. The top figure shows the average reward
over 5 random runs during training, with best hyper-parameters.
On the bottom we plot the average probability of the most-likely
actions during training. The bottom figure illustrates the fast convergence of sparse PCL to a near-deterministic policy.
sparse PCL is sub-optimal. This is mainly due to the coarse
discretization of the action space. Our main purpose here
is to demonstrate the fast and improved convergence to deterministic policies in sparse PCL, compared to soft PCL.
Further evaluation of sparse PCL will be left to future work.
8. Conclusions
In this work we studied the sparse entropy-regularized
problem in RL, whose optimal policy has non-zero probability for only a small number of actions. Similar to the
work by Nachum et al. (2017), we derived a relationship
between (near-)optimality and consistency for this problem. Furthermore, by leveraging the properties of the consistency equation, we proposed a class of sparse path consistency learning (sparse PCL) algorithms that are applicable to both on-policy and off-policy data and can learn
from multi-step trajectories. We found that the theoretical
advantages of sparse PCL correspond to empirical advantages as well. For tasks with a large number of actions,
we find significant improvement in final performance and
amount of time needed to reach that performance by using
sparse PCL compared to the original soft PCL.
Future work includes 1) extending the sparse PCL algorithm to the more general class of Tsallis entropy, 2) investigating the possibility of combining sparse PCL and
path following algorithms such as TRPO (Schulman et al.,
2015), and 3) comparing the performance of sparse PCL
with other deterministic policy gradient algorithms, such
as DPG (Silver et al., 2014) in the continuous domain.
Path Consistency Learning in Tsallis Entropy Regularized MDPs
References
Asadi, K. and Littman, M. An alternative softmax operator for reinforcement learning. In Proceedings of the
34th International Conference on Machine Learning, pp.
243–252, 2017.
Azar, M., Gómez, V., and Kappen, H. Dynamic policy
programming. Journal of Machine Learning Research,
13:3207–3245, 2012.
Bellman, R. Dynamic Programming. Princeton University
Press, 1957.
Bertsekas, D. Nonlinear programming. Athena scientific
Belmont, 1999.
Bertsekas, D. Constrained optimization and Lagrange multiplier methods. Academic press, 2014.
Bertsekas, D. and Tsitsiklis, J. Neuro-Dynamic Programming. Athena Scientific, 1996.
Brockman, G., Cheung, V., Pettersson, L., Schneider, J.,
Schulman, J., Tang, J., and Zaremba, W. OpenAI Gym.
arXiv:1606.01540, 2016.
Farahmand, A. M., Ghavamzadeh, M., Szepesvári, Cs., and
Mannor, S. Regularized policy iteration. In Proceedings
of Advances in Neural Information Processing Systems
21, pp. 441–448. MIT Press, 2008.
Farahmand, A. M., Ghavamzadeh, M., Szepesvári, Cs., and
Mannor, S. Regularized fitted Q-iteration for planning in
continuous-space Markovian decision problems. In Proceedings of the American Control Conference, pp. 725–
730, 2009.
Fox, R., Pakman, A., and Tishby, N. G-learning: Taming the noise in reinforcement learning via soft update.
In Proceedings of the 32nd International Conference on
Uncertainty in Artificial Intelligence, pp. 202–211, 2016.
Ghavamzadeh, M., Lazaric, A., Munos, R., and Hoffman,
M. Finite-sample analysis of lasso-td. In Proceedings of
the Twenty-Eighth International Conference on Machine
Learning, pp. 1177–1184, 2011.
Johns, J., Painter-Wakefield, C., and Parr, R. Linear complementarity for regularized policy evaluation and improvement. In Proceedings of Advances in Neural Information Processing Systems 23, pp. 1009–1017. MIT
Press, 2010.
Kappen, H. Path integrals and symmetry breaking for optimal control theory. Journal of Statistical Mechanics, 11,
2005.
Kolter, Z. and Ng, A. Regularization and feature selection
in least-squares temporal difference learning. In Proceedings of the Twenty-Sixth International Conference
on Machine Learning, pp. 521–528, 2009.
Konda, V. and Tsitsiklis, J. Actor-critic algorithms. In
Advances in neural information processing systems, pp.
1008–1014, 2000.
Lee, K., Choi, S., and Oh, S. Sparse Markov decision processes with causal sparse Tsallis entropy regularization
for reinforcement learning. IEEE Robotics and Automation Letters, 2018.
Mnih, V., Badia, A., Mirza, M., Graves, A., Lillicrap,
T., Harley, T., Silver, D., and Kavukcuoglu, K. Asynchronous methods for deep reinforcement learning. In
Proceedings of the 33rd International Conference on
Machine Learning, pp. 1928–1937, 2016.
Nachum, O., Norouzi, M., Xu, K., and Schuurmans, D.
Bridging the gap between value and policy based reinforcement learning. In NIPS, pp. 2772–2782, 2017.
Nachum, Ofir, Norouzi, Mohammad, Xu, Kelvin, and
Schuurmans, Dale. Trust-pcl: An off-policy trust region
method for continuous control. In Proceedings of the 5th
International Conference on Learning Representations,
2018.
Neu, G., Jonsson, A., and Gómez, V.
A unified
view of entropy-regularized Markov decision processes.
arXiv:1705.07798, 2017.
O’Donoghue, B., Munos, R., Kavukcuoglu, K., and Mnih,
V. PGQ: Combining policy gradient and Q-learning.
In Proceedings of the 5th International Conference on
Learning Representations, 2017.
Peters, J., Müling, K., and Altun, Y. Relative entropy policy search. In Proceedings of the 24th Conference on
Artificial Intelligence, pp. 1607–1612, 2010.
Puterman, M. Markov Decision Processes. Wiley Interscience, 1994.
Schulman, J., Levine, S., Moritz, P., Jordan, M., and
Abbeel, P. Trust region policy optimization. In Proceedings of the 32nd International Conference on Machine
Learning, pp. 1889–1897, 2015.
Silver, David, Lever, Guy, Heess, Nicolas, Degris, Thomas,
Wierstra, Daan, and Riedmiller, Martin. Deterministic
policy gradient algorithms. In ICML, 2014.
Sutton, R. and Barto, A. An Introduction to Reinforcement
Learning. MIT Press, 1998.
Path Consistency Learning in Tsallis Entropy Regularized MDPs
Sutton, R., McAllester, D., Singh, S., and Mansour, Y.
Policy gradient methods for reinforcement learning with
function approximation. In Proceedings of Advances in
Neural Information Processing Systems 12, pp. 1057–
1063, 2000.
Todorov, E. Linearly-solvable Markov decision problems.
In Proceedings of the 19th Advances in Neural Information Processing, pp. 1369–1376, 2006.
Todorov, E. Policy gradients in linearly-solvable MDPs. In
Proceedings of the 23rd Advances in Neural Information
Processing, pp. 2298–2306, 2010.
Tsallis, C. Possible generalization of Boltzmann-Gibbs
statistics. Journal of Statistical Physics, 52(1):479–487,
1988.
Ziebart, B. Modeling Purposeful Adaptive Behavior with
the Principle of Maximum Causal Entropy. PhD thesis,
Carnegie Mellon University, 2010.
Path Consistency Learning in Tsallis Entropy Regularized MDPs
A. Proofs of Section 5
Consider the Bellman operator for the entropy-regularized MDP with Tsallis entropy:
X
(Tsp f )(x) = α · spmax r(x, ·) + γ
P (x0 |x, ·)f (x0 ) /α .
x0
We first have the following technical result about its properties.
Proposition 1. The sparse-max Bellman operator Tsp has the following properties: (i) Translation: (Tsp (V + β))(x) =
(Tsp V )(x) + γβ; (ii) γ-contraction: k(Tsp V1 ) − (Tsp V2 )k∞ ≤ γkV1 − V2 k∞ ; (iii) Monotonicity: (Tsp V1 )(x) ≤ (Tsp V2 )(x)
for any value functions V1 , V2 : X → R such that V1 (x) ≤ V2 (x), ∀x ∈ X .
The detailed proof of this proposition can be found in Lee et al. (2018). Using these results, the Banach fixed point theorem
shows that there exists a unique solution for the following fixed point equation: V (x) = (Tsp V )(x), ∀x ∈ X , and this
solution is equal to the optimal value function Vsp∗ (x). Analogous to the arguments in standard MDPs, in this case the
optimal value function can also be computed using dynamic programming methods such as value iteration.
Before proving the main results, notice that by using analogous arguments of the complementary-slackness property in
KKT conditions, the second and the third consistency equation in (13) is equivalent to the following condition:
X
α
r(x, a) + γ
P (x0 |x, a)V (x0 ) + − αµ(a|x) − V (x) = Λ(x), ∀x ∈ X , ∀a ∈ Aµ (x),
2
x0 ∈X
(16)
X
α
r(x, a) + γ
P (x0 |x, a)V (x0 ) + − αµ(a|x) − V (x) ≤ Λ(x), ∀x ∈ X , ∀a 6∈ Aµ (x),
2
0
x ∈X
where Aµ (x) = {a ∈ A : µ(a|x) > 0} represents the set of actions that have non-zero probabilities w.r.t policy µ.
Theorem 3. The pair of optimal value function and optimal policy (Vsp∗ , µ∗sp ) of the MDP problem in (3) satisfies the
consistency equation in (13).
Proof. Recall that the optimal state-action value function is given by
X
Q∗sp (x, a) = r(x, a) + γ
P (x0 |x, a)Vsp∗ (x0 ).
x0 ∈X
According to Bellman’s optimality, the optimal value function satisfies the following equality:
X
α
Vsp∗ (x) = max
µ(a|x) Q∗sp (x, a) + (1 − µ(a|x)) ,
µ∈∆x
2
(17)
a∈A
at any state x ∈ X , where µ∗sp is the corresponding maximizer. By the KKT condition, we have that
Q∗sp (x, a) +
α
α
(1 − µ∗sp (a|x)) + λ∗sp (a|x) = Λ∗sp (x) + µ∗sp (a|x),
2
2
for any x ∈ X and any a ∈ A, where Λ∗sp is the Lagrange multiplier that corresponds to equality constraint
1, and λ∗sp ≥ 0 is the Lagrange multiplier that corresponds to inequality constraint µ(a|x) ≥ 0 such that
P
a∈A
µ(a|x) =
λ∗sp (a|x) · µ∗sp (a|x) = 0, ∀x ∈ X , ∀a ∈ A.
Recall from the definition of optimal state-action value function Q∗sp and the definition of the optimal policy µ∗sp , one has
that Aµ∗sp (x) = S(Q∗sp (x, ·)). This condition further implies
α
(1 − 2µ∗sp (a|x)), ∀x ∈ X , a ∈ S(Q∗sp (x, ·)).
2
P
Substituting the equality in (17) to this KKT condition, and noticing that 0 ≤ a∈A µ∗sp (a|x))2 ≤ 1, the KKT condition
implies that
α
αX ∗
Λ∗sp (x) + ≥ Vsp∗ (x) = Λ∗sp (x) +
µsp (a|x)µ∗sp (a|x) ≥ Λ∗sp (x),
2
2
Λ∗sp (x) = Q∗sp (x, a) +
a∈A
Path Consistency Learning in Tsallis Entropy Regularized MDPs
which further implies that
α
≤ Λ∗sp (x) − Vsp∗ (x) ≤ 0, ∀x ∈ X .
2
Therefore, by defining Λ(x) = Λ∗sp (x) − Vsp∗ (x), and λ(a|x) = λ∗sp (a|x), one immediately has that − α2 ≤ Λ(x) ≤ 0,
∀x ∈ X . Using this construction, one further has the following expression for any x ∈ X and any a ∈ S(Q∗sp (x, ·)):
−
Λ(x) = Q∗sp (x, a) +
α
− αµ∗sp (a|x) − Vsp∗ (x),
2
which proves consistency, based on the equivalence condition in (16).
Theorem 4. The solution policy µ of the consistency equation in (13) is α/(1 − γ)-optimal w.r.t. the sparse MDP problem
in (3). That is,
E
∞
X
α
α
γ t Rt +
.
1 − µ(at |xt ) | x0 = x, µ, P ≥ Vsp∗ (x) −
2
(1
−
γ)
t=0
(18)
Proof. To proof the sub-optimality performance bound given in this theorem, we first study the expression of (Tsp V ),
where Tsp is the Bellman operator of the Tsallis entropy-regularized MDP problem in (3). Let
X
Q̄(x, a) = r(x, a) + γ
P (x0 |x, a)V (x0 )
x0 ∈X
be the corresponding state-action value function. Using the definition from (3), one has the following expression:
X
1
(Tsp V )(x) = α · spmax · r(x, a) + γ
P (x0 |x, a)V (x0 )
α
0
x ∈X
= α · spmax
Q̄(x, ·)
α
a∈A
!
.
Furthermore, by exploiting the structure of the sparse-max formulation of an arbitrary value function, one also has the
following chain of equalities/inequalities:
!
X
α
Q̄(x, ·)
α · spmax
= max
µ(a|x) · Q̄(x, a) + (1 − µ(a|x))
µ∈∆x
α
2
a∈A
X
α
=
µ(a|x) · Q̄(x, a) + (1 − µ(a|x))
2
a∈A
X
α
αX
=
µ(a|x) Q̄(x) + − αµ(a|x) +
µ(a|x)2
2
2
a∈A
a∈A
α
≤V (x) + .
2
The first equality follows from the fact that α · spmax(Q̄(x, ·)/α) is a closed form solution of the optimization problem
X
max
µ(a|x) Q̄(x, a) − αHµ (x, a) ,
µ∈∆x
a∈A
when Hµ is the Tsallis entropy. The second equality follows from the fact that if (V, µ) satisfies the consistency equation,
then there exists a Lagrange multiplier Λ∗ (x) = Λ(x) + V (x), ∀x ∈ X such that the following KKT conditions hold:
α
− αµ(a|x) = Λ∗ (x), ∀x ∈ X , ∀a ∈ Aµ (x),
2
α
Q̄(x, a) + − αµ(a|x) ≤ Λ∗ (x), ∀x ∈ X , ∀a 6∈ Aµ (x),
2
X
µ(a|x) = 1, µ(a|x) ≥ 0, ∀x ∈ X , ∀a ∈ A,
Q̄(x, a) +
a
Path Consistency Learning in Tsallis Entropy Regularized MDPs
which further implies that µ is the maximizer of the inner optimization problem. The third equality follows from arithmetic
manipulations, and the first inequality follows from the consistency equation in (16), i.e., for any x ∈ X and any a ∈
Aµ (x), there exists Λ(x) ∈ [− α2 , 0] such that:
α
α
0 ≥ Λ(x) = Q̄(x, a) + − αµ(a|x) − V (x) ⇐⇒ Q̄(x, a) + − αµ(a|x) ≤ V (x).
2
2
Therefore combining all these arguments, one concludes that the following Bellman inequality holds:
α
(Tsp V )(x) ≤ V (x) + , ∀x ∈ X .
(19)
2
Now recall that the γ−contraction property (w.r.t. the ∞−norm) of the Bellman operator Tsp . By the Banach fixedpoint theorem, this property implies that there exists a unique fixed point solution Vsp∗ to equation V (x) = (Tsp V )(x),
for all x ∈ X , and it is the limit point (over all x ∈ X ) of the converging iterative sequence limn→∞ (Tspn V0 )(x) for
any initial value function V0 . Also recall that the translation property of this Bellman operator, i.e., for any constant K,
(Tsp (V + K)) = (Tsp V ) + γK. Therefore, by repeatedly applying the Bellman operator to both sides of the inequality in
(19), and by using the above properties of a Bellman operator, one can show that
Vsp∗ (x) = lim (Tspn V )(x) ≤
n→∞
∞
X
−γ t ·
n=0
α
1
α
+ V (x) = − ·
+ V (x), ∀x ∈ X .
2
2 1−γ
(20)
Furthermore, consider the consistency equation in (13), i.e., there exists a function Λ(x) ∈ [0, α2 ] such that for any x ∈ X
and any a ∈ Aµ (x),
α
α
− ≤ Λ(x) = Q̄(x, a) + − αµ(a|x) − V (x) ⇐⇒ V (x) ≤ Q̄(x, a) + α − αµ(a|x).
2
2
By multiplying µ(a|x) on both sides of this inequality and summing over a ∈ A, the above expression implies
X
V (x) ≤
µ(a|x) Q̄(x, a) + α − αµ(a|x)
a∈A
αX
α
1 − µ(a|x) +
≤
µ(a|x) Q̄(x, a) +
µ(a|x) 1 − µ(a|x)
2
2
a∈A
a∈A
X
X
α
α
1 − µ(a|x) + .
=
µ(a|x) r(x, a) + γ
P (x0 |x, a)V (x0 ) +
2
2
0
X
(21)
x ∈X
a∈A
Therefore, equipped with the γ−contraction property of the following Bellmen operator:
X
X
α
(Tµ V )(x) =
µ(a|x) r(x, a) +
1 − µ(a|x) + γ
P (x0 |x, a)V (x0 )
2
0
x ∈X
a∈A
and the Banach fixed-point theorem, for any initial value function V0 , one can deduce the following expression:
∞
X
α
1 − µ(at |xt ) | µ, x0 = x .
lim Tµ [V0 ]n (x) = E
γ t r(xt , at ) +
n→∞
2
t=0
Using the translation property of the Bellman operator (Tµ V ) and repeatedly applying this Bellman operator to both sides
of (21), one obtains the following inequality for any x ∈ X :
V (x) ≤ lim (Tµ V )n (x) +
n→∞
∞
X
α t
γ
2
n=0
∞
X
α
α
1
=E
γ t r(xt , at ) +
1 − µ(at |xt ) | µ, x0 = x + ·
.
2
2
1
−
γ
t=0
Therefore, by combining the results in (20) and in (22), one completes the proof of this theorem.
(22)
Path Consistency Learning in Tsallis Entropy Regularized MDPs
Algorithm 1 Sparse Path Consistency Learning
Input: Environment EN V , learning rate η, discount factor γ, regularization α, rollout d, number of steps N , replay
buffer capacity B, prioritized replay hyper-parameter α. Parameterizations of Λ and λ follow from the descriptions in
Section 6.
function Gradients(x0:T , R0:T −1 , a0:T −1 )
Pd−1
Compute C(t) = −V̄φ (xt ) + γ d V̄φ (xt+d ) + j=0 γ j (Rj + α/2 − αµ̄θ (aj |xj ) + λθ (aj |xj ) − Λρ (xj )) for t < T ,
padding with zeros
PTas−1necessary.
Compute ∆θ = t=0 C(t)∇θ C(t).
PT −1
Compute ∆φ = t=0 C(t)∇φ C(t).
PT −1
Compute ∆ρ = t=0 C(t)∇ρ C(t).
Return ∆θ, ∆φ, ∆ρ
end function
Initialize θ, φ, ρ.
Initialize empty replay buffer RB(α).
for i = 0 to N − 1 do
Sample x0:T , a0:T −1 ∼ µ̄θ on EN V , yielding reward R0:T −1 .
∆θ, ∆φ, ∆ρ = Gradients(x0:T , a0:T −1 , R0:T −1 ).
Update θ ← θ − η∆θ.
Update φ ← φ − η∆φ.
Update ρ ← ρ − η∆ρ.
PT −d
Input x0:T , a0:T −1 into RB with priority j=0 Rj .
If |RB| > B, remove episodes uniformly at random.
Sample s0:T from RB.
∆θ, ∆φ, ∆ρ = Gradients(x0:T , a0:T −1 , R0:T −1 ).
Update θ ← θ − η∆θ.
Update φ ← φ − η∆φ.
Update ρ ← ρ − η∆ρ.
end for
B. Experimental Details
For the algorithmic tasks, we follow a similar experimental setup as described in Nachum et al. (2017). We parameterize
all values by a single LSTM recurrent neural network with internal dimension 128 and multiple heads (one for each desired
quantity). At each training step, we sample a batch of 400 episodes using the current policy acting on the environment.
We perform a gradient step based on this batch. We then add the experience to the replay buffer and perform a gradient
step based on an off-policy batch sampled from the replay buffer. We fix the rollout to d = 10. As in Nachum et al.
(2017), our replay buffer is prioritized by episode rewards: the probability of sampling an episode from the replay buffer
is 0.1 + 0.9 · exp{αR}/Z where R is the total reward of the episode, Z is a normalizing factor, and we use α = 0.5. We
use a replay buffer of capacity B = 10, 000 episodes. In our experiments we use a learning rate of η = 0.005 and discount
γ = 0.9.
For HalfCheetah we parameterized the policy and value networks as feed forward networks with two hidden layers of
dimension 64 and tanh non-linearities. At each training step we sampled 100 steps from the environment and input
these into a replay buffer. We then sample a batch of 25 sub-episodes of 100 steps from the replay buffer, prioritized by
exponentiated recency (with weight 0.01) and perform a single training step. We use rollout d = 10, discount γ = 0.99,
and performed a hyperparameter search over learning rate η ∈ {0.0005, 0.0001}.
In standard Soft PCL, the policy µ̄θ is determined by logits output by the neural network. That is,
µ̄θ (−|x) = softmax NN(x, θ)0:|A|−1 ,
(23)
where NN(x, θ)0:|A|−1 are |A| output values of the neural network. For Sparse PCL, to induce sparsity, we parameterize
the policy using the G function:
µ̄θ (−|x) = relu NN(x, θ)0:|A|−1 − G(NN(x, θ)0:|A|−1 ) .
(24)
Path Consistency Learning in Tsallis Entropy Regularized MDPs
Accordingly, λθ is parameterized as
λθ (−|x) = relu G(NN(x, θ)0:|A|−1 − NN(x, θ)0:|A|−1 ) exp{NN(x, θ)|A| }.
(25)
B.1. Experimental Results for ReversedAddition
ReversedAddition
|A| = 40
|A| = 64
30
30
25
25
20
20
15
15
10
10
5
5
0
0
0
20000 40000 60000 80000 100000
|A| = 96
15
10
5
0
0
20000 40000 60000 80000 100000
Soft PCL
0
20000 40000 60000 80000 100000
Sparse PCL
Figure 3. The average reward over training for sparse PCL compared to the standard soft PCL on ReversedAddition. In this task, the
environment is a 2 × n grid of digits representing two numbers in base-|V| that the agent needs to sum. As in the other tasks in the main
paper, we see that sparse PCL becomes more advantageous compared to soft PCL as the action space increases in size.
| 2cs.AI
|
The Right Mutation Strength
for Multi-Valued Decision Variables
Benjamin Doerr1 , Carola Doerr2 , Timo Kötzing3
1
Laboratoire d’Informatique (LIX), École Polytechnique, Paris-Saclay, France
CNRS and LIP6, Sorbonne Universités, UPMC Univ Paris 06, Paris, France
3
Hasso-Plattner-Institut, Potsdam, Germany
arXiv:1604.03277v1 [cs.NE] 12 Apr 2016
2
Abstract
The most common representation in evolutionary computation are bit strings. This is
ideal to model binary decision variables, but less useful for variables taking more values.
With very little theoretical work existing on how to use evolutionary algorithms for such
optimization problems, we study the run time of simple evolutionary algorithms on some
OneMax-like functions defined over Ω = {0, 1, . . . , r − 1}n . More precisely, we regard a
variety of problem classes requesting the component-wise minimization of the distance to
an unknown target vector z ∈ Ω.
For such problems we see a crucial difference in how we extend the standard-bit mutation
operator to these multi-valued domains. While it is natural to select each position of the
solution vector to be changed independently with probability 1/n, there are various ways to
then change such a position. If we change each selected position to a random value different
from the original one, we obtain an expected run time of Θ(nr log n). If we change each
selected position by either +1 or −1 (random choice), the optimization time reduces to
Θ(nr + n log n). If we use a random mutation strength i ∈ {0, 1, . . . , r − 1}n with probability
inversely proportional to i and change the selected position by either +i or −i (random
choice), then the optimization time becomes Θ(n log(r)(log(n) + log(r))), bringing down the
dependence on r from linear to polylogarithmic.
One of our results depends on a new variant of the lower bounding multiplicative drift
theorem.
1
Introduction
In evolutionary computation, taking ideas both from computer science and biology, often search
and optimization problems are modeled in a way that the solution candidates are fixed-length
strings over the alphabet consisting of 0 and 1. In other words, the search space Ω is chosen
to be {0, 1}n for some positive integer n. Such a representation of solution candidates is very
suitable to model binary decision variables. For example, when searching for graph substructures
like large cliques, (degree-constrained) spanning trees, or certain matchings, we can use binary
decision variables describing whether a vertex or an edge is part of the solution or not. For these
reasons, the bit string representation is the most prominent one in evolutionary computation.
When a problem intrinsically consists of other types of decision variables, the algorithm
designer has the choice to either work with a different representation (e.g., permutations in
the traveling salesman problem) or to re-model the problem using a bit string representation.
For an example for the latter, see, e.g., [DJ07], where the Eulerian cycle problem (asking for
a permutation of the edges) was re-modeled as a matching problem. In general, such a remodeling may not lead to an efficient or a natural approach, and it may be better to work with
1
a representation different from bit strings. The traveling salesman problem is an example for
such a situation.
While in this work we shall not deal with the difficulties of treating permutation search
spaces in evolutionary computation, we shall try to extend our good understanding of the bit
string representation to representations in which the decision variables can take more values
than just zero and one. Consequently, we shall work with search spaces Ω = {0, . . . , r − 1}n .
Such search spaces are a natural representation when each decision variable can take one out
of r values. Examples from the evolutionary computation literature include scheduling n jobs
on r machines, which naturally leads to the search space {0, . . . , r − 1}n , see Gunia [Gun05].
However, also rooted trees lead to this type of representation: Since each vertex different from
the root has a unique predecessor in the tree, a rooted tree on n vertices can be represented
via an element of {0, . . . , n − 1}n−1 . This was exploited in [STW04] to design evolutionary
algorithms for shortest-path problems.
An alternative representation would be to code each value in log r bits, leading to a search
space of {0, 1}n log r . However, this representation has the weakness that search points with
similar fitness can be vastly different (the bit representations 10 . . . 0 and 01 . . . 1 code almost the
same value, but are complementary); this trap-like behavior can lead to a very poor performance
on some OneMax functions (see Section 1.2 for a formal defintion).
1.1
Mutation Operators for Multi-Valued Search Spaces
A first question, and our main focus in this work, is what mutation operators to use in such
multi-valued search spaces. When there is no particular topology in the components i ∈ [1..n] :=
{1, . . . , n}, that is, in the factors [0..r − 1], then the natural analogue of the standard-bit
mutation operator is to select each component i ∈ [1..n] independently and mutate the selected
components by changing the current value to a random other value in [0..r − 1]. This operator
was used in [STW04, Gun05] as well as in the theoretical works [DJS11, DP12].
When the decision values 0, 1, . . . , r − 1 carry more meaning than just denoting alternatives
without particular topology, then one may want to respect this in the mutation operator. We
shall not discuss the most general set-up of a general distance matrix defined on the values
0, 1, . . . , r − 1, but assume that they represent linearly ordered alternatives.
Given such a linear topology, several other mutation operators suggest itself. We shall
always imitate the principle of standard-bit mutation that each component i ∈ [1..n] is changed
independently with probability 1/n, so the only point of discussion is how such an elementary
change looks like. The principle that mutation is a minimalistic change of the individual suggests
to alter a selected component randomly by +1 or −1 (for a precise definition, including also a
description of how to treat the boundary cases, see again Section 2). We say that this mutation
operator has a mutation strength equal to one. Naturally, a mutation strength of one carries
the risk of being slow—it takes r − 1 such elementary mutations to move one component from
one boundary value, say 0, to the other, say r − 1.
In this language, the previously discussed mutation operator changing a selected component
to a new value chosen uniformly at random can (roughly) be described as having a mutation
strength chosen uniformly at random from [1..r − 1]. While this operator does not have the
disadvantage of moving slowly through the search space, it does have the weakness that reaching
a particular target is slow, even when already close to it.
Based on these (intuitive, but we shall make them precise later) observations, we propose
an elementary mutation that takes a biased random choice of the mutation strength. We give
more weight to small steps than the uniform operator, but do allow larger jumps with certain
probability. More precisely, in each elementary mutation independently we choose the mutation
2
strength randomly such that a jump of +j or −j occurs with probability inversely proportional
to j (and hence with probability Θ((j log r)−1 )). This distribution was used in [DRWW10] and
is called harmonic distribution, aiming at overcoming the two individual weaknesses of the two
operators discussed before and, as we shall see, this does indeed work.
1.2
Run time Analysis of Multi-Valued OneMax Functions
To gain a more rigorous understanding of the working principles of the different mutations
strengths, we conduct a mathematical run time analysis for simple evolutionary algorithms on
multi-valued analogues of the OneMax test function. Comparable approaches have been very
successful in the past in studying in isolation particular aspects of evolutionary computation,
see, e.g., [Jan13]. Also, many observations first made in such simplistic settings have later been
confirmed for more complicated algorithms (see, e.g., [AD11]) or combinatorial optimization
problems (see, e.g., [NW10]).
n
On bit strings,
Pnthe classic OneMax test function is defined by Om : {0, 1} →
R; (x1 , . . . , xn ) 7→
i=1 xi . Due to the obvious symmetry, for most evolutionary algorithms
it makes no difference whether the target is to maximize or to minimize this function. For
several reasons, among them the use of drift analysis, in this work it will be more convenient to
always assume that our target is the minimization of the given objective function.
The P
obvious multi-valued analogue of this OneMax function is Om : {0, 1, . . . , r − 1}n →
R; x 7→ ni=1 xi , however, a number of other functions can also be seen as multi-valued analogues. For example, we note that in the bit string setting we have Om(x) = H(x, (0, . . . , 0)),
where H(x, y) := |{i ∈ [1..n] | xi 6= yi }| denotes the Hamming distance between two bit
strings x and y. Defining fz : {0, 1}n → R; x 7→ H(x, z) for all z ∈ {0, 1}n , we obtain
a set of 2n objective functions that all have an isomorphic fitness landscape. Taking this
route to define multi-valued analogue
of OneMax functions, we obtain the class of functions
P
fz : {0, 1, . . . , r − 1}n 7→ R; x 7→ ni=1 |xi − zi | for all z ∈ {0, 1, . . . , r − 1}n , again with f(0,...,0)
being the OneMax function defined earlier. Note that these objective functions do not all have
an isomorphic fitness landscape. The asymmetry with respect to the optimum z can be overcome by replacing the classic distance |xi − zi | in the reals by the distance modulo r (ring
distance), that is, min{xi − (zi − r), |xi − zi |, (zi + r) − xi }, creating yet another non-isomorphic
fitness landscape. All results we show in the following hold for all these objective functions.
As evolutionary algorithm to optimize these test functions, we study the (1+1) evolutionary
algorithm (EA). This is arguably the most simple evolutionary algorithm, however, many results
that could first only be shown for the (1 + 1) EA could later be extended to more complicated
algorithms, making it an ideal instrument for a first study of a new subject. Naturally, to
study mutation operators we prefer mutation-based EAs. For the different ways of setting the
mutation strength, we conduct a mathematical run time analysis, that is, we prove bounds on
the expected number of iterations the evolutionary algorithm needs to find the optimal solution.
This optimization time today is one of the most accepted performance measures for evolutionary
algorithms.
1.3
Previous Works and Our Results
In particular for the situation that r is large, one might be tempted to think that results from
continuous optimization can be helpful. So far, we were not successful in this direction. A
main difficulty is that in continuous optimization, usually the asymptotic rate of convergence
is regarded. Hence, when operating with a fixed r in our setting and re-scaling things into, say,
{0, 1r , 2r , . . . , 1}n , then these results, due to their asymptotic nature, could become less meaningful. For this reason, the only work in the continuous domain that we found slightly resembling
3
ours is by Jägersküpper (see [Jäg08] and the references therein), which regards continuous optimization with an a-priori fixed target precision. However, the fact that Jägersküpper regards
approximations with respect to the Euclidean norm (in other words, minimization of the sphere
function) makes his results hard to compare to ours, which can be seen as minimization of the
1-norm.
Coming back to the discrete domain, as said above, the vast majority of theoretical works
on evolutionary computation work with a bit string representation. A notable exception is the
work on finding shortest path trees (e.g., [STW04]); however, in this setting we have that the
dimension and the number r of values are not independent: one naturally has r equal to the
dimension, because each of the n − 1 non-root vertices has to choose one of the n − 1 other
vertices as predecessor.
Therefore, we see only three previous works that are comparable to ours. The first two
regard the optimization of linear functions via the (1 + 1) EA using mutation with uniform
strength, that is, resetting a component to a random other value. The main result of [DJS11] is
that the known run time bound of O(n log n) on linear functions defined on bit strings remains
valid for the search space {0, 1, 2}n . This was extended and made more precise in [DP12], where
for r-valued linear functions an upper bound of (1 + o(1))e(r − 1)n ln(n) + O(r 3 n log log n) was
shown together with a (1 + o(1))n(r − 1) ln(n) lower bound.
A third paper considers dynamically changing fitness functions [KLW15]. They also consider
OneMax functions with distance modulo r, using ±1 mutation strength. In this setting the
fitness function changed over time and the task was to track it as closely as possible, which
the ±1 mutation strength can successfully do. Note that a seemingly similar work on the
optimization of a dynamic variant of the maze function over larger alphabets [LW14] is less
comparable to our work since there all non-optimal values of a decision variable contribute the
same to the fitness function.
Compared to these works, we only regard the easier static OneMax problem (note though
that there are several ways to define multi-valued OneMax functions), but obtain tighter results
also for larger values of r and for three different mutation strengths. For the uniform mutation
strength, we show a tight and precise (1 + o(1))e(r − 1)n ln(n) run time estimate for all values
of r (Section 4). For the cautious ±1 mutation strength, the run time becomes Θ(n(r + log n)),
that is, still (mostly) linear in r (Section 5). The harmonic mutation strength overcomes this
slowness and gives a run time of Θ(n log(r)(log(r) + log(n))), which for most values of r is
significantly better than the previous bound (Section 6).
All analyses rely on drift methods, for the lower bound for the case of uniform mutation
strength we prove a variant of the multiplicative drift lower bound theorem [Wit13] that does not
need the restriction that the process cannot go back to inferior search points (see Section 4.2.2).
2
Algorithms and Problems
In this section we define the algorithms and problems considered in this paper. We let [r] :=
{0, 1, . . . , r − 1} and [1..r] := {1, 2, . . . , r}. For a given search space Ω, a fitness function is a
function f : Ω → R. While a frequently analyzed search space is Ω = {0, 1}n , we will consider
in this paper Ω = [r]n .
We define the following two metrics on [r], called interval-metric and ring-metric, respectively. The intuition is that the interval metric is the usual metric induced by the metric on the
natural numbers, while the ring metric connects the two endpoints of the interval (and, thus,
4
forms a ring). Formally we have, for all a, b ∈ [r],
dint (a, b) = |b − a|;
dring (a, b) = min{|b − a|, |b − a + r|, |b − a − r|}.
We consider different step operators v : [r] → [r] (possibly randomized). These step operators will later decide the update of a mutation in a given component. Thus we call, for any
given x ∈ [r], d(x, v(x)) the mutation strength. We consider the following step operators.
• The uniform step operator chooses a different element from [r] uniformly at random; thus
we speak of a uniform mutation strength.
• The ±1 operator chooses to either add or subtract 1, each with probability 1/2; this
operator has a mutation strength of 1.
• The Harmonic operator makes a jump of size j ∈ [r] with probability proportional to
1/j, choosing the direction uniformly at random; we call its mutation strength harmonic
mutation strength.
Note that, in the case of the ring-metric, all steps are implicitly considered with wrap-around.
For the interval-metric, we consider all steps that overstep a boundary of the interval as invalid
and discard this mutation as infeasible. Note that this somewhat arbitrary choice does not
impact the results in this paper.
We consider the algorithms RLS and (1 + 1) EA as given by Algorithms 1 and 2. Both
algorithms sample an initial search point from [r]n uniformly at random. They then proceed
in rounds, each of which consists of a mutation and a selection step. Throughout the whole
optimization process the algorithms maintain a population size of one, and the individual in this
population is always the most recently sampled best-so-far solution. The two algorithms differ
only in the mutation operation. While the RLS makes a step in exactly one position (chosen
uniformly at random), the (1 + 1) EA makes, in each position, a step with probability 1/n.
The fitness of the resulting search point y is evaluated and in the selection step the parent x
is replaced by its offspring y if and only if the fitness of y is at least as good as the one of x. Since
we consider minimization problems here, this is the case if f (y) ≤ f (x). Since we are interested
in expected run times, i.e., the expected number of rounds it takes until the algorithm evaluates
for the first time a solution of minimal fitness, we do not specify a termination criterion. For
the case of r = 2, the two algorithms are exactly the classic Algorithms RLS and (1 + 1) EA,
for all three given step operators (which then degenerate to the flip operator, which flips the
given bit).
Note that the algorithms with the considered topologies are unbiased in the general sense
of [RV11] (introduced for {0, 1}n by Lehre and Witt [LW12] and made specific for several
combinatorial search spaces in [DKLW13]).
Let d be either the interval- or the ring-metric and let z ∈ [r]n . We can define a straightforward generalization of the OneMax fitness function as
n
X
d(xi , zi ).
i=1
Whenever we refer to an r-valued OneMax function, we mean any such function. We refer to
d as the metric of the OneMax function and to z as the target of the OneMax function.
5
Algorithm 1: RLS minimizing a function f : [r]n → R with a given step operator v.
1
2
3
4
5
6
7
Initialization: Sample x ∈ [r]n uniformly at random and query f (x);
Optimization: for t = 1, 2, 3, . . . do
Choose i ≤ n uniformly at random;
for j = 1, . . . , n do
if j = i then yj ← v(xj ) else yj ← xj
Evaluate f (y);
if f (y) ≤ f (x) then x ← y
Algorithm 2: The (1 + 1) EA minimizing a function f : [r]n → R with a given step
operator v.
1 Initialization: Sample x ∈ [r]n uniformly at random and query f (x);
2 Optimization: for t = 1, 2, 3, . . . do
3
for i = 1, . . . , n do
4
With probability 1/n set yi ← v(xi ) and set yi ← xi otherwise;
5
6
3
Evaluate f (y);
if f (y) ≤ f (x) then x ← y
Drift Analysis
A central tool in many of our proofs is drift analysis, which comprises a number of tools to
derive bounds on hitting times from bounds on the expected progress a process makes towards
the target. Drift analysis was first used in evolutionary computation by He and Yao [HY01]
and is now, after a large number of subsequent works, probably the most powerful tool in run
time analysis. We briefly collect here the tools that we use.
We phrase the following results in the language that we have some random process, either in
the real numbers or in some other set Ω, but then equipped with a potential function g : Ω → R.
We are mostly interested in the time the process (or its potential) needs to reach 0.
Multiplicative drift is the situation that the progress is proportional to the distance from the
target. This quite common situation in run time analysis was first framed into a drift theorem,
namely the following one, in [DJW12]. A more direct proof of this results, that also gives large
deviation bounds, was later given in [DG13].
Theorem 1 (from [DJW12]). Let X (0) , X (1) , . . . be a random process taking values in S :=
{0} ∪ [smin , ∞) ⊆ R. Assume that X (0) = s0 with probability one. Assume that there is a δ > 0
such that for all t ≥ 0 and all s ∈ S with Pr[X (t) = s] > 0 we have
E[X (t+1) |X (t) = s] ≤ (1 − δ)s.
Then T := min{t ≥ 0 | X (t) = 0} satisfies
E[T ] ≤
ln(s0 /smin ) + 1
.
δ
It is easy to see that the upper bound above cannot immediately be matched with a lower
bound of similar order of magnitude. Hence it is no surprise that the only lower bound result for
multiplicative drift, the following theorem by Witt [Wit13], needs two additional assumptions,
namely that the process does not move away from the target and that it does not too often
6
make large jumps towards the target. We shall see later (Theorem 7) that the first restriction
can be removed under not too strong additional assumptions.
Theorem 2 (from [Wit13]). Let X (t) , t = 0, 1, . . . be random variables taking values in some
finite set S of positive numbers with min(S) = 1. Let X (0) = s0 with probability one. Assume
that for all t ≥ 0,
Pr[X (t+1) ≤ X (t) ] = 1.
Let saim ≥ 1. Let 0 < β, δ ≤ 1 be such that for all s > saim and all t ≥ 0 with Pr[X (t) = s] > 0,
we have
E[X (t) − X (t+1) | X (t) = s] ≤ δs,
Pr[X (t) − X (t+1) ≥ βs | X (t) = s] ≤
βδ
.
ln(s)
Then T := min{t ≥ 0 | X (t) ≤ saim } satisfies
E[T ] ≥
ln(s0 ) − ln(saim ) 1 − β
.
δ
1+β
In situations in which the progress is not proportional to the distance, but only monotonically
increasing with it, the following variable drift theorem of Johannsen [Joh10] can lead to very
good results. Another version of a variable drift theorem can be found in [MRC09, Lemma 8.2].
Theorem 3 (from [Joh10]). Let X (t) , t = 0, 1, . . . be random variables taking values in some
finite set S of non-negative numbers. Assume 0 ∈ S and let xmin := min(S \ {0}). Let X (0) = s0
with probability one. Let T := min{t ≥ 0 | X (t) = 0}. Suppose that there exists a continuous and
monotonically increasing function h : [xmin , s0 ] → R>0 such that E[X (t) −X (t+1) |X (t) ] ≥ h(X (t) )
holds for all t < T . Then
Z s0
1
xmin
+
dx.
E[T ] ≤
h(xmin )
xmin h(x)
4
Mutation Strength Chosen Uniformly at Random
In this section, we analyze the mutation operator with uniform mutation strength, that is, if
the mutation operator chooses to change a position, it resets the current value to a different
value chosen independently (for each position) and uniformly at random. We shall prove the
same results, tight apart from lower order terms, for all r-valued OneMax functions defined in
Section 2. Let f be one such objective function and let z be its target.
When regarding a single component xi of the solution vector, it seems that replacing a nonoptimal xi by some yi that is closer to the target, but still different from it, gains us some fitness,
but does not lead to a structural advantage (because we still need an elementary mutation that
resets this value exactly to the target value zi ). This intuitive feeling is correct for RLS and not
correct for the (1 + 1) EA.
4.1
RLS with Uniform Mutation Strength
For RLS, we turn the above intuition into the potential function g : [r]n → R; x 7→ H(x, z) =
|{i ∈ [1..n] | xi 6= zi }|, the Hamming distance, which counts the number of non-optimal positions
in the current solution x. We get both an upper and a lower bound on the drift in this potential
which allow us to apply multiplicative drift theorems. From that we get the following result.
7
Theorem 4. Let f be any r-valued OneMax function with target z ∈ [r]n . Then randomized
local search (RLS) with uniform mutation strength has an optimization time T satisfying
E[T ] = n(r − 1)(ln(n) + Θ(1)).
If x0 denotes the random initial individual, then for all x ∈ [r]n we have
E[T |x0 = x] = n(r − 1)HH(x,z) ,
where, for any positive integer k, we let Hk :=
Pk
j=1 1/j
denote the k-th Harmonic number.
Proof. Consider one iteration of RLS started with a current solution x 6= z. Let y be the current
solution after one iteration, that is, the value of x after mutation and selection. We observe
that g(y) = g(x) − 1 if and only if the mutation operator selects a non-optimal position i of x
(this happens with probability g(x)/n) and then replaces xi by zi (this happens with probability
1/(r − 1)). In all other cases, we have g(y) = g(x), though not necessarily y = x. Consequently,
the expected progress with respect to g in this iteration is
g(x) − E[g(y)] =
g(x)
.
n(r − 1)
(1)
Let us denote by Tx0 the run time of RLS conditional on the initial search point being x0 . Then
the multiplicative drift theorem (Theorem 1) gives an upper bound of
E[Tx0 ] ≤ n(r − 1)(ln(g(x0 )) + 1).
Similarly, the assumptions of the multiplicative drift theorem for lower bounds (Theorem 2)
are satisfied with saim = ln n and β = 1/ ln n. Consequently, assuming g(x0 ) = exp(ω(ln ln n))
in the second estimate, we obtain
E[Tx0 ] ≥ n(r − 1)(ln(g(x0 )) − ln ln n)(1 − 2/ ln(n))
= n(r − 1) ln(g(x0 ))(1 − o(1)).
In the above analysis we used multiplicative drift with the Hamming distance because this in
a generic manner gave a very strong result. We also used a drift approach to ease the comparison
with the other results we will obtain, also via drift analysis. For this particular problem, also
a very problem-specific approach can be used, which gives an even sharper result. Consider a
run of RLS starting with a search point x0 . For i ∈ [0..g(x0 )], let Ti denote the first iteration
for
after which g(x) ≤ i, where Tg(x0 ) = 0. Then equation (1) shows that E[Ti−1 − Ti ] = n(r−1)
i
all i ∈ [1..g(x0 )]. Consequently,
g(x
X0 )
(Ti−1 − Ti )
E[Tx0 ] = E
i=1
g(x0 )
=
X
i=1
E[Ti−1 − Ti ]
g(x0 )
= n(r − 1)
X 1
i
i=1
= n(r − 1)Hg(x0 ) ,
8
P
where for all k ∈ N, Hk := ki=1 (1/i) is the kth Harmonic number. The harmonic number is
well understood, e.g., we have Hk = ln(k) + γ + O(1/k) with γ = 0.5772... being the EulerMascheroni constant and we have the non-asymptotic bounds ln(k) ≤ Hk ≤ ln(k) + 1, which
gives
n(r − 1) ln(g(x0 )) ≤ E[Tx0 ] ≤ n(r − 1)(ln(g(x0 )) + 1).
By the law of total probability, the expected run time of RLS (with the usual random
initialization) is E[T ] = n(r − 1)E[Hg(x0 ) ]. The expected potential of the random initial search
we see
point is E[g(x0 )] = n(1 − 1/r).
p
√ By a Chernoff bound (e.g., Theorem 1.11 in [Doe11]),
that Pr[|g(x0 ) − E[g(x0 )]| ≥ n ln n] ≤ 2n−2 . Hence E[Hg(x0 ) ] = HE[g(x0 )] ± Θ( ln(n)/n) ±
Θ(ln(n)/n2 ). The first error term could be further reduced by arguments as used in [DD14],
where for the case r = 2 a run time bound of E[T ] = nHn/2 − 1/2 ± o(1) was shown. We do not
detail this idea any further and are content with summarizing the above in the following result,
which in particular shows that the Hamming distance very precisely describes the quality of a
search point.
4.2
The (1+1) EA with Uniform Mutation Strength
We now consider the same run time analysis problem for the (1 + 1) EA, that is, instead of
selecting a single random entry of the solution vector and applying an elementary mutation to
it, we select each entry independently with probability 1/n and mutate all selected entries. Our
main result is the following.
Theorem 5. For any r-valued OneMax function, the (1 + 1) EA with uniform mutation
strength has an expected optimization time of
E[T ] = e(r − 1)n ln(n) + o(nr log n).
As we will see, since several entries can be changed in one mutation step, the optimization
process now significantly differs from the RLS process. This has two important consequences.
First, while for the RLS process the Hamming distance of the current search point precisely
determined the expected remaining optimization time, this is not true anymore for the (1 +
1) EA. This can be seen (with some mild calculations which we omit here) from the search
points
P
x = (r, 0, . . . , 0) and y = (1, 0, . . . , 0) and the fitness function f defined by f (x) = ni=1 x0 .
The second, worse, consequence is that the Hamming distance does not lead to a positive
drift from each search point. Consider again x = (r, 0, . . . , 0) and f as above. Denote by x′
the search point after one mutation-selection cycle started with x. Let g be the Hamming
distance to the optimum x∗ = (0, . . . , 0) of f . Then for r ≥ 5, the drift from the search point
r−4
< 0. Indeed, we have g(x′ ) = 0, that is,
x satisfies E[g(x) − g(x′ )] ≤ −(1 ± o(1)) 2e(r−1)n
1
g(x) − g(x′ ) = 1, with probability (1 − 1/n)n−1 (1/n)(1/(r − 1)) = (1 ± o(1)) e(r−1)n
. This
is the only event that gives a positive drift. On the other hand, with probability at least
r−2
(1 − 1/n)n−2 (n − 1)(1/n2 )(1 + 2 + · · · + (r − 2))/(r − 1)2 = (1 ± o(1)) 2e(r−1)n
, the mutation
operator touches exactly the first and one other entry of x and does so in a way that the first
entry does not become zero and the second entry remains small enough for x′ to be accepted.
This event leads to a drift of −1, showing the claim.
For these reasons, we resort to the actual fitness as potential function in our upper bound
proof. It is clear that the fitness also is not a perfect measure for the remaining optimization
time (compare, e.g., the search points (2, 0, . . . , 0) and (1, 1, 0, . . . , 0)), but naturally we have
a positive drift from each non-optimal search point, which we shall exploit via the variable
drift theorem. For the lower bound, a worsening of the Hamming distance in the optimization
process is less of a problem, since we only need an upper bound for the drift. Hence for the
9
lower bound, we can use multiplicative drift with g again. However, since the process may move
backwards occasionally, we cannot apply Witt’s lower bound drift theorem (Theorem 2), but
have to prove a variant of it that does not require that the process only moves forward. This
lower bound theorem for multiplicative drift might be of interest beyond this work.
4.2.1
An Upper Bound for the Run Time
Theorem 6. For any r-valued OneMax function f , the (1 + 1) EA with uniform mutation
strength has an expected optimization time of
E[T ] ≤ e(r − 1)n ln(n) + (2 + ln(2))e(r − 1)n
= e(r − 1)n ln(n) + O(rn).
Pn
Proof. Let z be the optimum of f . Then f can be written as f (x) =
i=1 d(xi , zi ), where
d is one of the distance measures on [r] that were described in Section 2. Let x be a fixed
search point and y be the result of applying one mutation and selection step to x. We use the
short-hand di := d(xi , zi ). We first show that
n
X
1
di (di + 1).
∆ := f (x) − E[f (y)] ≥
2e(r − 1)n
(2)
i=1
Indeed, f (x)−f (y) is always non-negative. Consequently, it suffices to point out events that lead
to the claimed drift. With probability (1 − (1/n))n−1 ≥ (1/e), the mutation operator changes
exactly one position of x. This position then is uniformly distributed in [1..n]. Conditional
Pi
i (di +1)
, where the first inequality
δ/(r − 1) = d2(r−1)
on this position being i, we have ∆ ≥ dδ=1
uses the fact that all our fitness functions are of the type that if there is a value xi ∈ [r] with
d(xi , zi ) = k, then for each j ∈ [0..k −1] there is at least one value yi ∈ [r] such that d(yi , zi ) = j.
This shows (2).
For any d ≥ 1, we have d(d + 1) ≥ 2d and d(d + 1) ≥ d2 . Also, the mapping
1
1
2 and ∆ ≥
d 7→ d2 is convex. Consequently, we have ∆ ≥ 2e(r−1)n
2 f (x)
e(r−1)n f (x), that is,
1
1
2
∆ ≥ max{ 2e(r−1)n
2 f (x) , e(r−1)n f (x)}. To this drift expression, we apply Johannsen’s [Joh10]
variable drift theorem (Theorem 3). Let S = [0..(r − 1)n]. Let h : R>0 → R>0 be defined by
1
1
2
h(s) = 2e(r−1)n
2 s for s ≥ 2n and h(s) = e(r−1)n s for s < 2n. Then h is a continuous increasing
function satisfying ∆ ≥ h(f (x)). Consider the process X0 , X1 , . . . with Xt describing the fitness
after the tth iteration. Given that we start with a fitness of X0 , Johannsen’s drift theorem gives
Z X0
1
1
E[T ] ≤
+
ds
h(1)
h(s)
1
Z X0
Z 2n
n2
n
2e(r − 1) 2 ds +
= e(r − 1)n +
e(r − 1) ds
s
s
2n
1
1
1
−
+ e(r−1)n ln(2n)
≤ e(r−1)n + 2e(r−1)n2
2n X0
≤ e(r − 1)n ln(n) + (1 + 1 + ln(2))e(r − 1)n.
We may remark that the drift estimate above is pessimistic in that it applies to all r-valued
OneMax functions. For an r-valued OneMax function using the ring metric or one having
the optimum close to (r/2, . . . , r/2), we typically have two different bit values in each positive
distance from zi . In this case, the drift stemming from exactly position i being selected for
P i −1
d2i
, that is, nearly twice the value we
2δ/(r − 1) = r−1
mutation is ∆ ≥ di /(r − 1) + dδ=1
10
computed above. The fact that in the following subsection we prove a lower bound matching
the above upper bound for all r-valued OneMax functions shows that this, almost twice as
high, drift has an insignificant influence on the run time.
4.2.2
A Lower Bound for the Run Time
In this section, we write (q)+ := max{q, 0} for any q ∈ R. We aim at proving a lower bound,
again via drift analysis, that is, via transforming an upper bound on the expected progress (with
respect to a suitable potential function) into a lower bound on the expected run time. Since we
only need an upper bound on the progress, we can again (as in the RLS analysis) work with the
Hamming distance g(x) = H(x, z) to the optimum z as potential and, in the upper estimate
of the drift, ignore the fact that this potential may increase. The advantage of working with
the Hamming distance is that the drift computation is easy and we observe multiplicative drift,
which is usually convenient to work with.
We have to overcome one difficulty, though, and this is that the only known lower bound
theorem for multiplicative drift (Theorem 2) requires that the process does not move away from
the target, in other words, that the g-value is non-increasing with probability one. As discussed
above, we do not have this property when using the Hamming distance as potential in a run of
the (1 + 1) EA. We solve this problem by deriving from Theorem 2 a drift theorem (Theorem 7
below) that gives lower bounds also for processes that may move away from the optimum.
Compared to Theorem 2, we need the stronger assumptions (i) that we have a Markov process
and (ii) that we have bounds not only for the drift g(X (t) ) − g(X (t+1) ) or the positive part
(g(X (t) ) − g(X (t+1) ))+ of it, but also for the positive progress (s − g (t+1) )+ with respect to
any reference point s ≤ g(X (t) ). This latter condition is very natural. In simple words, it just
means that we cannot profit from going back to a worse (in terms of the potential) state of the
Markov chain.
A second advantage of these stronger conditions (besides allowing the analysis of nondecreasing processes) is that we can easily ignore an initial segment of the process (see Corollary 8). This is helpful when we encounter a larger drift in the early stages of the process. This
phenomenon is often observed, e.g., in Lemma 6.7 of [Wit13]. Previous works, e.g., [Wit13],
solved the problem of a larger drift in the early stage of the process by manually cutting off
this phase. This requires again a decreasing process (or conditioning on not returning to the
region that has been cut off) and an extra argument of the type that the process with high
probability reaches a search point with potential in [s̃0 , 2s̃0 ] for a suitable s̃0 . So it is safe to
say that Corollary 8 is a convenient way to overcome these difficulties.
We start by proving our new drift results, then compute that the Hamming distance to the
optimum satisfies the assumptions of our drift results, and finally state and prove the precise
lower bound.
Theorem 7. (multiplicative drift, lower bound, non-decreasing process) Let
X (t) , t = 0, 1, . . . be a Markov process taking values in some set Ω. Let S ⊂ R be a finite
set of positive numbers with min(S) = 1. Let g : Ω → S. Let g(X (0) ) = s0 with probability one.
Let saim ≥ 1. Let
T := min{t ≥ 0 | g(X (t) ) ≤ saim }
be the random variable describing the first point in time for which g(X (t) ) ≤ saim .
Let 0 < β, δ ≤ 1 be such that for all ω ∈ Ω, all saim < s ≤ g(ω), and all t ≥ 0 with
11
Pr[X (t) = ω] > 0, we have
E[(s − g(X (t+1) ))+ | X (t) = ω] ≤ δs,
Pr[s − g(X (t+1) ) ≥ βs | X (t) = ω] ≤
Then
E[T ] ≥
βδ
.
ln(s)
ln(s0 ) − ln(saim )
ln(s0 ) − ln(saim ) 1 − β
≥
(1 − 2β).
δ
1+β
δ
The proof follows from an application of Witt’s drift theorem (Theorem 2) to the random
process Y (t) := min{g(X (τ ) ) | τ ∈ [0..t]}.
Proof. We define a second random process by Y (t) := min{g(X (τ ) ) | τ ∈ [0..t]}. By definition,
Y takes values in S and Y is decreasing, that is, we have Y (t+1) ≤ Y (t) with probability one for
all t ≤ 0. Trivially, we have Y (0) = g(X (0) ) = s0 . Let TY := min{t ≥ 0 | Y (t) ≤ saim } be the
first time this new process reaches or goes below saim . Clearly, TY = T .
Let β, δ as in the theorem. Let saim < s and t ≥ 0 such that Pr[Y (t) = s] > 0. Observe that
when Y (t) = s, then Y (t) − Y (t+1) = s − min{s, g(X (t+1) )} = (s − g(X (t+1) ))+ . Let AYs be the
event that Y (t) = s and let BωX be the event that X (t) = ω. Using the fact that X is a Markov
process, we compute
E[Y (t) − Y (t+1) | AYs ]
X
=
Pr[AYs | B]E[(s − g(X (t+1) ))+ | AYs , BωX ]
ω:g(ω)≥s
=
X
Pr[BωX | AYs ]E[(s − g(X (t+1) ))+ | BωX ]
X
Pr[BωX | AYs ]δs = δs
ω:g(ω)≥s
≤
ω:g(ω)≥s
and
Pr[Y (t) − Y (t+1) ≥ βs | AYs ]
X
=
Pr[BωX | AYs ] Pr[s − X (t+1) ≥ βs | AYs , BωX ]
ω:g(ω)≥s
=
X
Pr[BωX | AYs ] Pr[s − X (t+1) ≥ βs | BωX ]
X
Pr[BωX | AYs ]
ω:g(ω)≥s
≤
ω:g(ω)≥s
βδ
βδ
=
.
ln(s)
ln(s)
Consequently, Y satisfies the assumptions of the multiplicative lower bound theorem (Theaim ) 1−β
orem 2). Hence E[T ] = E[TY ] ≥ ln(s0 )−ln(s
δ
1+β . Elementary algebra shows (1 − β) ≥
(1 − 2β)(1 + β), which gives the second, more convenient lower bound.
Corollary 8. Assume that the assumptions of Theorem 7 are satisfied, however with δ replaced
by δ(s) for some function δ : S → (0, 1]. Then for any saim < s̃0 ≤ s0 , we have
E[T ] ≥
ln(s̃0 ) − ln(saim )
(1 − 2β),
δmax (s̃0 )
where δmax (s̃0 ) := max{δ(s) | saim < s ≤ s̃0 }.
12
Proof. Let S̃ := S ∩ [0, s̃0 ]. Let g̃ : Ω → S̃; ω 7→ min{s̃0 , g(ω)}. Let ω ∈ Ω, saim < s ≤ g̃(ω), and
t be such that Pr[X (t) = ω] > 0. Then
E[(s − g̃(X (t+1) ))+ | X (t) = ω]
= E[(s − g(X (t+1) ))+ | X (t) = ω]
≤ δ(s)s ≤ δmax (s̃0 )s
by the assumptions of Theorem 7 and s ≤ s̃0 . Similarly,
Pr[s − g̃(X (t+1) ) ≥ βs | X (t) = ω]
= Pr[s − g(X (t+1) ) ≥ βs | X (t) = ω]
βδmax (s̃0 )
βδ(s)
≤
.
≤
ln(s)
ln(s)
Hence we may apply Theorem 7 to (S̃, s̃0 , g̃, δmax (s̃0 )) instead of (S, s0 , g, δ) and obtain the
claimed bound.
Lemma 9. Let f be an r-valued OneMax function with optimum z. Let x ∈ [r]n and y be the
outcome of applying mutation and selection to x. Let s+ := H(x, z) and s ≤ s+ . Then
s+1
1
s
+ 3e
E[(s − H(y, z))+ ] ≤
.
e(r − 1)n 1 − 1/n
(r − 1)n
Proof. Let u be the outcome of mutating x with uniform mutation strength and y be the result
of applying selection (with respect to f ) to x and u.
We consider first the case that s = s+ . With probability (1 − (1/n))n−1 , u and x differ in
exactly one position. Conditional on this, E[(s − H(y, z))+ ] = s/(r − 1)n. The only other event
in which possibly s > H(y, z) is that u and x differ in at least two positions i and j such that
ui = zi and uj = zj . The probability for this to happen is 1− (1− 1/(r − 1)n)s − (s/(r − 1)n)(1−
1/(r−1)n)s−1 ≤ 1−(1−s/(r−1)n)−(s/(r−1)n)(1−(s−1)/(r−1)n) = s(s−1)/(r−1)2 n2 . In this
case, we can estimate (s − H(y, z))+ from above by the number of non-correct positions that are
touched by the mutation operator, which in this case is Bin(s, 1/n) conditional on being at least
two, which again is at most 3. Consequently, in the case that s = s+ , we have E[(s−H(y, z)+ ] ≤
(1 − (1/n))n−1 s/(r − 1)n + 3s(s − 1)/(r − 1)2 n2 ≤ (s/(r − 1)n)(1/e(1 − 1/n) + 3(s − 1)/(r − 1)n).
Let now s ≤ s+ −1. Let Z := |{i ∈ [1..n] | xi 6= zi 6= ui }| ∈ [0..s+ ] be the number of positions
that are incorrect in both x and u. Clearly, H(y, z) stochastically dominates Z, which we write
as H(x, z) Z. Let Z ′ be defined analogous to Z but for an original search point x′ with
H(x′ , z) = s + 1 ≤ H(x, z). Then, clearly, Z Z ′ . Consequently, s − H(y, z) s − Z s − Z ′ ,
and consequently, (s−H(y, z))+ (s−Z ′ )+ and E[(s−H(y, z)+ ] ≤ E[(s−Z ′ )+ ]. The only way
to get a positive value for s − Z ′ is that at least two incorrect positions of x′ are changed to their
correct value in the mutation offspring. Analogous to the previous paragraph, the probability for
this to happen is 1−(1−1/(r −1)n)s+1 −((s+1)/(r −1)n)(1−1/(r −1)n)s ≤ (s+1)s/(r −1)2 n2 .
In this case, we can estimate (s − Z ′ )+ from above by the number of incorrect positions that
are touched by the mutation operator (conditional on being at least two) minus one, which is
at most 2. We conclude E[(s − H(y, z))+ ] ≤ E[(s − Z ′ )+ ] ≤ 2(s + 1)s/(r − 1)2 n2 .
Putting the two cases together, we see that we always have E[(s − H(y, z))+ ] ≤ (s/(r −
1)n)(1/e(1 − 1/n) + 3(s + 1)/(r − 1)n).
Lemma 10. Let f be an r-valued OneMax function with optimum z. Let saim = ln(n)3 and
β = 1/ ln(n). Let x ∈ [r]n with H(x, z) > saim . Let saim < s ≤ H(x, z). Let y be the outcome
2
1
2− ln(n) if n ≥ 11.
of applying mutation and selection to x. Then Pr[s − H(y, z) ≥ βs] ≤ r−1
13
Proof. We have that Pr[s − H(y, z) ≥ βs] ≤ Pr[H(x, z) − H(y, z) ≥ βsaim ]. The latter is
at most the probability that at least βsaim = ln(n)2 positions of x flip to a particular value
(namely the one given by z) in one mutation step. Since the expected number of positions
flipping to the correct value is at most 1/(r − 1), a strong multiplicative Chernoff bound (e.g.,
than ln(n)2 with probability at most
Cor. 1.10(b) in [Doe11]) shows that this number is greater √
2
1 − ln(n)2
2
ln(n)
≤ r−1 2
(e/ ln(n) (r − 1))
for n ≥ 10.29 ≈ exp( 2e).
We are now ready to give the main result of this section.
Theorem 11. For any r-valued OneMax function, the (1 + 1) EA with uniform mutation
strength has an expected optimization time of
E[T ] ≥ e(r − 1)n (ln(n) − 6 ln ln(n)) (1 − O(1/ ln(n)))
≥ e(r − 1)n ln(n) − O((r − 1)n ln ln(n)).
Proof. Let n be sufficiently large. Let Ω = [r]n and f : Ω → R an r-valued OneMax function
with optimum z. Let saim = ln(n)3 and β = 1/ ln(n). For all saim < s ≤ n, let δ(s) :=
(1/e(r − 1)n)(1/(1 − 1/n) + 3e(s + 1)/(r − 1)n). Consider a run of the (1 + 1) EA optimizing f
initialized with a random search point X (0) . We have E[H(X (0) , z)] = n(1−1/r). Consequently,
we have H(X (0) , z) ≥ n/3 with probability 1 − exp(−Ω(n)). In the following, we thus assume
that X (0) is a fixed initial search point with some H(·, z) value of at least n/3. Denote by X (t)
the search point building the one-element population of this EA after the t-th iteration. Let
g : Ω → N; x 7→ H(x, z). By Lemma 9 and 10, the following conditions are satisfied for all
ω ∈ Ω, all saim < s ≤ g(ω), and all t ≥ 0 with Pr[X (t) = ω] > 0.
E[(s − g(X (t+1) ))+ | X (t) = ω] ≤ δ(s)s.
βδ(s)
.
Pr[s − g(X (t+1) ) ≥ βs | X (t) = ω] ≤
ln(s)
We apply Corollary 8 with s̃0 = n/ ln(n)3 ≤ n/3 and δmax (s̃0 ) ≤
O(1/ ln(n)3 (r − 1))) and obtain
1
e(r−1)n (1
+ O(1/n) +
ln(s̃0 ) − ln(saim )
(1 − 2β)
δmax (s̃0 )
≥ e(r − 1)n ln(n) − O((r − 1)n ln ln(n)).
E[T ] ≥
We remark that the lower order term O((r − 1)n log log n) in this lower bound could be
removed with stronger methods. We preferred to use the simple and natural proof approach via
multiplicative drift, because it is easy to handle and still relatively precisely describes the true
behavior of the process. As is visible from Lemma 9, in the early stages the progress is slightly
faster than the multiplicative (main) term s/e(r − 1)n. This is why we cut out the regime from
the initial H-value of approximately n(1 − 1/r) up to an H-value of s̃0 = n/ ln(n)3 , resulting in
a −Θ((r − 1)n log log n) term in our lower bound. Another −Θ((r − 1)n log log n) term stems
from the second condition of Witt’s lower bound drift theorem (which is similar to the second
condition of our theorem). To prove a bound sharp up to terms of order (r − 1)n log log n, we
need β ≤ log log n/ log n. However, this forbids using an saim smaller than 1/β = log n/ log log n,
since otherwise any improvement would count into the bad event of the second condition. An
saim of at least polylogarithmic size immediately implies an Ω((r−1)n log log n) additive distance
to the upper bound proven in Theorem 6. We are very optimistic that via variable drift, in
particular, the lower bound theorem of [DFW11], both difficulties could be overcome. We do
not think that this small improvement justifies the effort, though.
14
5
Unit Mutation Strength
In this section we regard the mutation operator that applies only ±1 changes to each component.
It is not very surprising that RLS with the ±1 variation operator needs Θ(n(r + log n))
fitness evaluations in expectation to optimize any r-valued OneMax function. We give the full
proof below since it is similar to the analysis of the (1 + 1) EA equipped with the ±1 variation
operator. The proof makes use of the following observation. There are two extreme kinds of
individuals with fitness n. The first kind is only incorrect in one position (by an amount of n);
the second kind is incorrect in every position (by an amount of 1). The first kind of individual
is hard to improve (the deficient position has to be chosen for variation), while the second
kind is very easy to improve (every position allows for improvement). We reflect this in our
choice of potential function by giving each position a weight exponential in the amount that it
is incorrect, and then sum over all weights.
Theorem 12. The expected optimization time of RLS with the ±1 variation operator is Θ(n(r+
log n)) for any r-valued OneMax function.
Proof. The lower bound Ω(nr) is quite immediate: with probability 1/2 we start in a search
point of fitness at most nr/2 and in each step the algorithm increases the fitness by at most one.
On the other hand, there is a coupon collector effect which yields the Ω(n log n) lower bound.
Indeed, it is well-known that this is the expected number of RLS iterations that we need in case
of r = 2, and larger values of r will only delay optimization.
We now turn to the more interesting upper bound. Let any r-valued OneMax function
be given with metric d and target z. We want to employ a multiplicative drift theorem (see
Theorem 1). We measure the potential of a search point by the following drift function. For all
x ∈ Ω = [r]n , let
n
X
(wd(zi ,xi ) − 1),
(3)
g(x) :=
i=1
where w := 1 + ε is an arbitrary constant between 1 and 2. In fact, for the analysis of RLS we
can simply set w := 2 but since we want to re-use this part in the analysis of the (1 + 1) EA,
we prefer the more general definition here.
We regard how the potential changes on average in one iteration. Let x denote the current
search point and let y denote the search point that we obtain from x after one iteration of RLS
(after selection). Clearly, we have that each position is equally likely to be selected for variation.
When a non-optimal component i is selected, then the probability that yi is closer to zi than
xi is at least 1/2, while for every already optimized component we will not accept any move of
RLS (thus implying yi = xi ). This shows that, abbreviating di := d(zi , xi ) for all i ∈ [1..n], and
denoting by O := {i ∈ [1..n] | xi = zi } the set of already optimized bits,
X
1
(wdi − 1) − (wdi −1 − 1)
E[g(x) − g(y) | x] = 2n
i∈[1..n]\O
=
1
2n
X
i∈[1..n]\O
≥
1
2n (1
=
1
2n (1
(1 −
di
1
w )w
X
(wdi − 1)
−
1
w)
−
1
w )g(x).
i∈[1..n]
Furthermore, the maximal potential that a search point can obtain is at most nwr . Plugging
1), we see that the expected optimization time
all this into the multiplicative drift (see Theorem
1
1
r
is of order at most ln(nw )/ 2n (1 − w ) = O(n(log(n) + r)), as desired.
15
For the analysis of the (1 + 1) EA we will proceed similarly as for RLS. To help with the
added complexity, we use the following lemma.
Lemma 13. Let n be fixed, let q be a cost function on elements of [1..n] and let c be a cost
function on subsets of [1..n]. Furthermore, let a random variable S ranging over subsets of
[1..n] be given. Then we have
∀T ⊆ [1..n] : c(T ) ≤
X
∀T ⊆ [1..n] : c(T ) ≥
X
i∈T
q(i) ⇒ E[c(S)] ≤
n
X
q(i) Pr[i ∈ S];
(4)
q(i) ⇒ E[c(S)] ≥
n
X
q(i) Pr[i ∈ S].
(5)
and
i∈T
i=1
i=1
P
P
Pn
Proof. We have E[c(S)] =
T ⊆[n] Pr[S = T ]c(S) ≤
T ⊆[n] Pr[S = T ]
i∈T q(i) =
Pn
q(i)
Pr[i
∈
S].
The
other
direction
follows
analogously.
i=1
The proof for the case of the (1 + 1) EA follows along similar lines, but is (significantly)
more involved.
Theorem 14. The expected optimization time of the (1 + 1) EA with the ±1 variation operator
is Θ(n(r + log n)) for any r-valued OneMax function.
Proof. The lower bound Ω(nr) follows almost as for RLS: With constant probability the initial
search point is Θ(nr) away from the optimum, and the expected progress towards the optimum
is bounded form above by 1. Thus, with a simple lower-bound additive drift theorem [HY01],
the lower bound of Ω(nr) follows.
Regarding the upper bound, let any r-valued OneMax function be given with metric d and
target z. We want to employ multiplicative drift again. We fix some w > 1 to be specified later.
With any search point x ∈ Ω we associate a vector d ∈ Rn such that, for all i ≤ n, di = d(xi , zi ).
We use the same potential g on Ω as for the analysis of RLS, that is, for all x ∈ Ω,
g(x) =
n
X
i=1
(wdi − 1).
Let any current search point x ∈ Ω be given and let Y be the random variable describing the
search point after one cycle of mutation and selection. Let E1 be the event that Y is obtained
from x by flipping exactly one bit and the result is accepted (that is, f (Y ) ≤ f (x)). Let E2 be
the event that at least 2 bits flip and the result is accepted. The total drift in the potential g
is now E[g(x) − g(Y )] = E[g(x) − g(Y ) | E1 ] Pr[E1 ] + E[g(x) − g(Y ) | E2 ] Pr[E2 ]. We are now
going to estimate E[g(x) − g(Y ) | E2 ]. The random variable Y is completely determined by
choosing a set S ⊆ [1..n] of bit positions to change in x and then, for each such position i ∈ S,
choosing how to change it (away or towards the target zi ). For each choice S ⊆ [1..n], let A(S)
be the set of all possible values for Y which have the set S as positions of change. For each
possible S ⊆ [1..n], let Y (S) be the random variable Y conditional on making changes exactly
at the bit positions of S. Thus, we can now write the random variable Y as
X
Y =
Y (S) Pr[S].
S
We are now going to estimate, for any possible S,
E[g(Y (S)) − g(x)].
16
For all possible S, let c(S) = E[g(Y (S)) − g(x)]. Let a possible S be given. Note that Y (S) is
the uniform distribution on A(S). For each y ∈ A(S) and each i ∈ S, we have d(yi , zi ) − di ∈
{−1, 0, 1} (note that the case of being 0 can only occur when r is odd and we have a circle,
or in other such border cases); in the case that this value is 1 we call (y, i) an up-pair, and in
case that this value is −1 we call this pair a down-pair. We let U be thePset of all up-pairs. As
we only consider accepted mutations, we have that, for all y ∈ A(S), i∈S d(yi , zi ) − di ≤ 0.
This implies that there are at least as many down-pairs as there are up-pairs in A(S) × S.
Furthermore, for any up-pair (y, i) with di 6= 0 there is y ′ ∈ A(S) such that (y ′ , i) is a down-pair
and, for all j ∈ S \ {i}, yj′ = yj . Thus, for all up-pairs (y, i) ∈ U there is a down-pair (y, i), such
that the mapping (y, i) 7→ (y, i) is injective and, for all (y, i) ∈ U with di 6= 0, (y, i) = (y ′ , i).
Note that, for all up-pairs (y, i), we have di ≤ di .
We now get, for any (y, i) ∈ U ,
wd(yi ,zi ) − wdi + wd(y i ,zi ) − wdi ≤ wdi (w − 1 +
= wdi
1
− 1)
w
(w − 1)2
.
w
Overall we have
c(S) = E[g(Y (S)) − g(x)]
X
1
g(y) − g(x)
=
|A(S)|
y∈A(S)
=
≤
n
X X
1
wd(yi ,zi ) − wdi
|A(S)|
y∈A(S) i=1
X
1
wd(yi ,zi ) − wdi + wd(y i ,zi ) − wdi
|A(S)|
(y,i)∈U
≤
X
(w − 1)2
1
wdi
|A(S)|
w
≤
1X
(y,i)∈U
2
i∈S
wdi
(w − 1)2
.
w
Using Lemma 13, we see that
E[g(Y ) − g(x) | E2 ] ≤
n
X
1 di (w − 1)2
w
n
2w
=
n
(w − 1)2 X di
w .
2wn
i=1
i=1
We use the following estimation of progress we can make with changing exactly one position.
E[g(x) − g(Y ) | E1 ] Pr[E1 ] ≥
=
17
1 X
(1 −
2ne
i∈[n]
n
X
w−1
2wne
i=1
di
1
w )w
wdi .
Let w be any constant > 1 such that w − 1 − e(w − 1)2 > 0, and let c = (w − 1 − e(w − 1)2 )/e.
Then we have
E[g(x) − g(Y )] ≥
=
n
n
w − 1 X di (w − 1)2 X di
w −
w
2wne
2wn
i=1
w − 1 − e(w −
2wne
i=1
1)2
n
X
wdi
i=1
n
=
≥
c X di
w
2wn
i=1
c
g(x).
2wn
Again, the maximal potential that a search point can obtain is at most nwr . Plugging all
this into the multiplicative drift (see
Theorem 1), we see that the expected optimization time
c
= O(n(log(n) + r)), as desired.
is of order at most ln(nwr )/ 2wn
6
Harmonic Mutation Strength
In this section we will consider a mutation operator with variable step size. The idea is that
different distances to the target value require different step sizes for rapid progress. We consider
a mutation operator which, in each iteration, chooses its step size from a fixed distribution. As
distribution we use what we call the harmonic distribution, which chooses step size j ∈ [1..r − 1]
with probability proportional to 1/j. Using the bound on the harmonic number Hr−1 < 1+ln r,
we see that the probability of choosing such a j is at least 1/(j(1 + ln r)).
Theorem 15. The RLS as well as the (1 + 1) EA with the harmonically distributed step size
(described above) has an expected optimization time of Θ(n log r(log n + log r)) on any r-valued
OneMax function.
Proof. We first show the upper bound by considering drift on the fitness. Let any x ∈ Ω be
given, let Y be the random variable describing the best individual of the next iteration and let
Ai,j be the event that Y differs from x in exactly bit position i and this bit position is now j
1
closer to the optimum. Note that, for both RLS and the (1+1) EA, we get Pr[Ai,j ] ≥ 2enj(1+ln
r) .
We have
E[f (x) − f (Y )] ≥
=
=
di
n X
X
i=1 j=1
n
X
i=1
di
n X
X
i=1 j=1
E[f (x) − f (Y ) | Ai,j ] Pr[Ai,j ]
j Pr[Ai,j ] ≥
di
n X
X
i=1 j=1
j
2enj(1 + ln r)
1
di
=
f (x).
2en(1 + ln r)
2en(1 + ln r)
As the initial fitness is less than rn, the multiplicative drift theorem (see Theorem 1) gives us
the desired total optimization time.
Now we turn to the lower bound. A straightforward coupon collector argument gives us
the lower bound of Ω(n log r log n), since each position has to change from incorrect to correct
at some point, and that mutation has a probability of O(1/(n log r)). It remains to show a
18
lower bound of Ω(n(log r)2 ). To this end, let f be any r-values OneMax function and x∗ its
optimum. Let g(x) = d(x1 , x∗1 ) be the distance of the first position to the optimal value in the
first position. Let h(x) = ln(g(x) + 1). Let x′ be the outcome of one mutation step and x′′
be the outcome of selection from {x, x′ }. We easily compute E[max{0, h(x) − h(x′ )}] ≤ n K
ln r
as
well.
For
the
random
for some absolute constant K. Consequently, E[h(x) − h(x′′ )] ≤ n K
ln r
initial search point, we have g(x) ≥ r/2 with constant probability, that is, h(x) = Ω(log r) with
constant probability. Consequently, the additive drift theorem gives that the first time T at
2
which h(x) = 0, satisfies E[T ] ≥ Ω(log r)/ n K
ln r = Ω(n log r).
In the same way as we showed the additive drift statement E[h(x) − h(x′′ )] = O(1/n log r),
we could have shown a multiplicative drift statement for g, namely E[g(x) − g(x′′ )] =
O(g(x)/n log r); in fact, the latter is implied by the former. Unfortunately, due to the presence
of large jumps – we have Pr[g(x′′ ) ≤ g(x)/2] = Θ(1/n log r) –, we cannot exploit this via the
lower bound multiplicative drift theorem.
Naturally, the question arises whether the O((log r)2 ) dependence on r can be improved.
In particular, one wonders whether drawing the step size from the Harmonic distribution is
optimal, or whether another distribution gives a better optimization time. This is exactly the
problem considered in [DRWW10], where the following result is presented, which could also be
used to derive the run time bound of Theorem 15.
Theorem 16 ([DRWW10]). Let a random process on A = {0, . . . , r} be given, representing the
movement of a token. Fix a probability distribution of step sizes D over {1, . . . , r}. Initially, the
token is placed on a random position in A. In round t, a random step size d is chosen according
to D. If the token is in position x ≥ d, then it is moved to position x − d, otherwise it stays
put. Let TD be the number of rounds until the token reaches position 0. Then minD (E[TD ]) =
Θ((log r)2 ).
While our processes have a slightly different behavior (including the possibility to overshoot
the goal), we believe that these differences only lead to to differences in the constants of the
optimization time. Thus, the above theorem indicates that the Harmonic distribution is an
optimal choice and cannot be improved.
7
Conclusion
While many analyses of randomized search heuristics focus on the behavior of the algorithm in
dependence of a large and growing dimension, we additionally considered a growing size of the
search space in each dimension. We considered the (1 + 1) EA with three different mutation
strengths and proved asymptotically tight optimization times for a variety of OneMax-type
test functions over an alphabet of size r. We proved that both using large changes (change to
uniformly chosen different value) or very local changes (change value by ±1) leads to relatively
slow (essentially linear in r) optimization times of Θ(rn log n) and Θ(n(r + log n)), respectively.
We then considered a variable step size operator which allows for both large and small steps
with reasonable probability; this leads to an optimization time of Θ(n log r(log n + log r)). Note
that this bound, while polylogarithmic in r, is worse than the bound of Θ(n(r + log n)) for the
±1 operator when r is asymptotically smaller than log n log log n. This shows that there is no
uniform superior mutation operator among the three proposed operators.
Acknowledgments
This research benefited from the support of the “FMJH Program Gaspard Monge in optimization and operation research”, and from the support to this program from EDF (Électricité de
19
France).
References
[AD11]
Anne Auger and Benjamin Doerr. Theory of Randomized Search Heuristics. World
Scientific, 2011.
[DD14]
Benjamin Doerr and Carola Doerr. The impact of random initialization on the
runtime of randomized search heuristics. In Proc. of Genetic and Evolutionary
Computation Conference (GECCO), pages 1375–1382. ACM, 2014.
[DFW11]
Benjamin Doerr, Mahmoud Fouz, and Carsten Witt. Sharp bounds by probabilitygenerating functions and variable drift. In Proc. of Genetic and Evolutionary Computation Conference (GECCO), pages 2083–2090. ACM, 2011.
[DG13]
Benjamin Doerr and Leslie A. Goldberg. Adaptive drift analysis. Algorithmica,
65:224–250, 2013.
[DJ07]
Benjamin Doerr and Daniel Johannsen. Adjacency list matchings: an ideal genotype for cycle covers. In Proc. of Genetic and Evolutionary Computation Conference
(GECCO), pages 1203–1210. ACM, 2007.
[DJS11]
Benjamin Doerr, Daniel Johannsen, and Martin Schmidt. Runtime analysis of
the (1+1) evolutionary algorithm on strings over finite alphabets. In Proc. of
Foundations of Genetic Algorithms (FOGA), pages 119–126. ACM, 2011.
[DJW12]
Benjamin Doerr, Daniel Johannsen, and Carola Winzen. Multiplicative drift analysis. Algorithmica, 64:673–697, 2012.
[DKLW13] Benjamin Doerr, Timo Kötzing, Johannes Lengler, and Carola Winzen. Black-box
complexities of combinatorial problems. Theoretical Computer Science, 471:84–106,
2013.
[Doe11]
Benjamin Doerr. Analyzing randomized search heuristics: Tools from probability
theory. In Anne Auger and Benjamin Doerr, editors, Theory of Randomized Search
Heuristics, pages 1–20. World Scientific Publishing, 2011.
[DP12]
Benjamin Doerr and Sebastian Pohl. Run-time analysis of the (1+1) evolutionary
algorithm optimizing linear functions over a finite alphabet. In Proc. of Genetic and
Evolutionary Computation Conference (GECCO), pages 1317–1324. ACM, 2012.
[DRWW10] Martin Dietzfelbinger, Jonathan E. Rowe, Ingo Wegener, and Philipp Woelfel.
Tight bounds for blind search on the integers and the reals. Combinatorics, Probability & Computing, 19:711–728, 2010.
[Gun05]
Christian Gunia. On the analysis of the approximation capability of simple evolutionary algorithms for scheduling problems. In Proc. of Genetic and Evolutionary
Computation Conference (GECCO), pages 571–578. Jahn, 2005.
[HY01]
Jun He and Xin Yao. Drift analysis and average time complexity of evolutionary
algorithms. Artificial Intelligence, 127:57–85, 2001.
20
[Jäg08]
Jens Jägersküpper. Oblivious randomized direct search for real-parameter optimization. In Proc. of European Symposium on Algorithms (ESA), pages 553–564.
Springer, 2008.
[Jan13]
Thomas Jansen. Analyzing Evolutionary Algorithms—The Computer Science Perspective. Springer, 2013.
[Joh10]
Daniel Johannsen. Random combinatorial structures and randomized search heuristics. PhD thesis, Saarland University, 2010.
[KLW15]
Timo Kötzing, Andrei Lissovoi, and Carsten Witt. (1+1) EA on generalized dynamic onemax. In Proc. of Foundations of Genetic Algorithms (FOGA), pages
40–51. ACM, 2015.
[LW12]
Per Kristian Lehre and Carsten Witt. Black-box search by unbiased variation.
Algorithmica, 64:623–642, 2012.
[LW14]
Andrei Lissovoi and Carsten Witt. MMAS vs. population-based EA on a family
of dynamic fitness functions. In Proc. of Genetic and Evolutionary Computation
Conference (GECCO), pages 1399–1406. ACM, 2014.
[MRC09]
Boris Mitavskiy, Jonathan Rowe, and Chris Cannings. Theoretical analysis of local
search strategies to optimize network communication subject to preserving the total
number of links. International Journal of Intelligent Computing and Cybernetics,
2:243–284, 2009.
[NW10]
Frank Neumann and Carsten Witt. Bioinspired Computation in Combinatorial
Optimization – Algorithms and Their Computational Complexity. Springer, 2010.
[RV11]
Jonathan Rowe and Michael Vose. Unbiased black box search algorithms. In
Proc. of Genetic and Evolutionary Computation Conference (GECCO), pages
2035–2042. ACM, 2011.
[STW04]
Jens Scharnow, Karsten Tinnefeld, and Ingo Wegener. The analysis of evolutionary
algorithms on sorting and shortest paths problems. J. Math. Model. Algorithms,
3:349–366, 2004.
[Wit13]
Carsten Witt. Tight bounds on the optimization time of a randomized search
heuristic on linear functions. Combinatorics, Probability & Computing, 22:294–
318, 2013.
21
| 9cs.NE
|
Closing the AI Knowledge Gap
arXiv:1803.07233v1 [cs.CY] 20 Mar 2018
Ziv Epstein∗ , Blakeley H. Payne∗ , Judy Hanwen Shen, Abhimanyu Dubey,
Bjarke Felbo, Matthew Groh, Nick Obradovich, Manuel Cebrian, Iyad Rahwan
Media Lab, Massachusetts Institute of Technology, Cambridge, MA, USA
Correspondence: {cebrian, irahwan}@mit.edu
Abstract
AI researchers employ not only the scientific
method, but also methodology from mathematics
and engineering. However, the use of the scientific
method – specifically hypothesis testing – in AI is
typically conducted in service of engineering objectives. Growing interest in topics such as fairness
and algorithmic bias show that engineering-focused
questions only comprise a subset of the important
questions about AI systems. This results in the AI
Knowledge Gap: the number of unique AI systems
grows faster than the number of studies that characterize these systems’ behavior. To close this gap,
we argue that the study of AI could benefit from
the greater inclusion of researchers who are well
positioned to formulate and test hypotheses about
the behavior of AI systems. We examine the barriers preventing social and behavioral scientists from
conducting such studies. Our diagnosis suggests
that accelerating the scientific study of AI systems
requires new incentives for academia and industry,
mediated by new tools and institutions. To address
these needs, we propose a two-sided marketplace
called TuringBox. On one side, AI contributors
upload existing and novel algorithms to be studied
scientifically by others. On the other side, AI examiners develop and post machine intelligence tasks
designed to evaluate and characterize algorithmic
behavior. We discuss this market’s potential to democratize the scientific study of AI behavior, and
thus narrow the AI Knowledge Gap.
1
The Many Facets of AI Research
Although AI is a sub-discipline of computer science, AI researchers do not exclusively use the scientific method in their
work. For example, the methods used by early AI researchers
often drew from logic, a subfield of mathematics, and are distinct from the scientific method we think of today. Indeed AI
has adopted many techniques and approaches over time. In
this section, we distinguish and explore the history of these
∗
Equal contribution.
methodologies with a particular emphasis on characterizing
the evolving science of AI.
1.1
AI as Math
As early as the seventeenth century, the notion that intelligence could be equated to symbolic information processing
was formalized. In his 1677 Preface to the General Science
Leibniz wrote: “It is obvious that if we could find characters or signs suited for expressing all our thoughts as clearly
and as exactly as arithmetic expresses numbers or geometry
expresses lines, we could do in all matters insofar as they
are subject to reasoning all that we can do in arithmetic and
geometry. For all investigations which depend on reasoning
would be carried out by transposing these characters and by a
species of calculus.” [Leibniz, 1685]
In the early 20th century, Leibniz’s ideas influenced a number of mathematicians. Logician and mathematician David
Hilbert posed the famous question: Can all of mathematical reasoning be formalized [Hilbert, 1928]? This question
spurred many others to explore the limits of computation and
logic, an enterprise that culminated in Gödel’s incompleteness theorems, which revealed fundamental limits of formal
reasoning.
But these discoveries of formal limitations around computability and formal reasoning did not stop scholars from
pursuing the foundations of mechanized intelligence. AI,
as its known today, began with Alan Turing’s seminal work
“Computing Machinery and Intelligence” in which Turing
discussed the idea of creating machines that can think [Turing, 1950]. Turing acknowledged the vagueness of the terms
‘machine’ and ‘thought,’ and operationalized the question
with the Universal Turing Machine, which could perform arbitrary symbolic computation. The idea, again, was that once
mathematics is mechanized, all manner of reasoning subsequently follows.
Thus, the early builders of AI systems were mainly mathematicians, devising mechanistic procedures–often called
proof theories–for all manner of reasoning. In 1955, Herbert Simon and Allen Newell’s Logic Theorist proved 38 theorems in the Principia Mathematica [Newell et al., 1959].
This led Simon to claim that they had “solved the mind-body
problem.” He argued that with a sufficiently powerful version of the Logic Theorist, we could automate mathematical
reasoning, which in turn would enable the automation of all
reasoning.
In subsequent decades, theoretical developments in symbolic reasoning continued. This manifested in numerous
mathematical investigations, from analyzing the computational complexity of various symbolic reasoning problems,
to providing precise mathematical semantics of logic programming languages [Van Emden and Kowalski, 1976], to
new forms of computationally efficient symbolic reasoning
on constrained languages [McCarthy, 1981; Jaffar and Maher,
1994; Kakas et al., 1992]. In tandem, significant theoretical
developments took place in machine learning, grounded in
statistical learning theory. These efforts resulted in rigorous
foundations for reasoning about facts [Pearl, 1986], learning
patterns [Vapnik, 1999] and taking actions and planning in
the presence of uncertainty [Sutton and Barto, 1998] .
1.2
AI as Engineering
Since the 1970s, in parallel to these developments in AI theory, a growing cadre of AI engineers started taking shape.
On one hand, some mathematically-minded computer scientists produced demonstrations of the capabilities of particular symbolic reasoning techniques. For example, one can
show how an AI agent performs planning by solving logical satisfiability problems [Kautz et al., 1992] or performs
various commonsense reasoning tasks using circumscriptionbased logical reasoning [McCarthy, 1986].
On the other hand, AI engineers embarked on formalizing
all manner of domain-specific facts in symbols and rules that
can be used in operational expert systems [Jackson, 1998].
There were ambitious – but ultimately unsuccessful – attempts to build expert systems manually with encyclopedic
knowledge, capable of answering any question [Lenat et al.,
1985]. The difficulty of curating such knowledge bases made
salient that building general AI systems was not only a challenge from a computational perspective, but also required an
infeasible knowledge engineering effort.
Despite two ‘AI Winters,’ the engineering methodologies
employed by builders of AI systems became increasingly sophisticated, leveraging contemporaneous increases in computation capacities and data streams. This sophistication took
the form of i) better developed methodologies, and ii) more
standardized and precise benchmarks.
First, AI engineers developed standardized knowledge engineering methodologies for AI systems, such as KADS
[Wielinga et al., 1992]. This development also benefited from
the overall maturity in the broader field of software engineering, enabling standardized Application Programming Interfaces, team management, code documentation and sharing,
and so on.
Second, AI engineers have developed increasingly sophisticated benchmark problems to compare their systems. These
benchmarks have taken three forms: as standardized tasks,
datasets, or metrics. In the early days of symbolic AI, these
benchmarks were qualitative, such as the ‘Blocks World’
[Sussman and Winograd, 1970; Bylander, 1994]. But modern
AI grounded in statistical theory has invited more sophisticated benchmark tasks, from board games like Chess [Campbell et al., 2002] and Checkers [Schaeffer et al., 2007], to card
games like Poker [Bowling et al., 2015], to computer games
like Atari [Mnih et al., 2015], to artificial markets for testing
trading algorithms [Wellman et al., 2001], to Robocup Soccer
[Kitano et al., 1997].
In problems spanning across application domains such as
computer vision and natural language processing (two of the
most prominent application areas of modern AI techniques),
the benchmark for performance on tasks has been set by
widely used, standardized, large-scale evaluation datasets.
For instance, in the task of object recognition, the ImageNet [Russakovsky et al., 2015] dataset has become a ubiquitous standard for performance. Similarly, for the tasks
of image captioning and scene understanding, the Microsoft
Common Objects in Context (MS-COCO) [Lin et al., 2014]
dataset has been instrumental in providing a reliable benchmark for performance.
The development of such large-scale benchmark problems
has led to the construction and widespread adoption of metrics to assess the performance of new algorithms at scale. Enhancements of traditional signal processing metrics such as
the receiver operating characteristic (ROC) [Davis and Goadrich, 2006] including precision, recall, and F1 measures have
been established as standard metrics for performance assessment. In more involved tasks, metrics such as mean Average
Precision [Van De Sande et al., 2010](for object detection),
BLEU [Papineni et al., 2002] (for machine translation), and
Inception Score [Zhou et al., 2017] (for generative modeling assessment) are some examples that have been established
by the research community to evaluate complex intelligence
tasks.
1.3
AI as Science
In addition to these mathematical and engineering-based
methodologies, AI researchers also frequently employ the
scientific method – specifically the hypothesis testing
paradigm. For instance, to evaluate the performance of
reinforcement learning algorithms in a multi-agent setting
[Littman, 1994], it is common to resort to Monte Carlo simulation to estimate average properties such as convergence
rate and long-run performance. Likewise, in many complex
AI systems, the performance of the system under different
choices of parameters is often compared using the null hypothesis testing paradigm[Cohen, 1995].
However, while some AI researchers have indeed employed a hypothesis driven approach in studying AI systems,
the predominant objective has been engineering-oriented: focused on designing and building intelligent systems. These
pertain, for example, to the performance of a heuristic planning system against some theoretical optimum, the speed of
convergence of a real-time optimization algorithm, or the
classification accuracy of a given image classification system.
Indeed, it is common for articles submitted to AI conferences to be rejected if they do not present a specific technical advance beyond existing AI algorithms and models.
This observation has an important implication. Even when
AI researchers use the scientific method to examine the systems they build, the engineering-oriented incentives of the
larger community lead them to selectively target engineeringoriented research questions.
While these are important scientific questions about AI
Study
Sweeney et al., 2013
Stimulus
Treatment Groups
AI System
Scope
Behavior
Names as Search Terms
Racial Association
Google Search
Engine
Via API
Disparate
Treatment
Parliamentarian
Headshots
Gender, Fitzpatrick
Skin Type Class
Facial Recognition
Algorithms
Via API
Disparate
Mistreatment
Consumer Profiles
Web Broswer, Operating
System, User History
Online Pricing
Algorithms
Field
Disparate
Treatment
Buolamwini et al., 2018
Hannak et al., 2014
Table 1: Sample of existing behavioral studies of AI in terms of their stimulus, AI system, and measured behavior. We describe the treatment
types for each stimuli and the level at which the study occurs (either locally, via an API, or “in the field”).
systems, they are not the only important questions. As we
will discuss in the next section, there is a growing set of research questions of interest to a broad range of scientists outside of the scope of the AI canon that involve concepts and
methods beyond the training of a typical AI researcher.
2
A New Science of AI
Recent advances in hardware and deep learning have led to
the proliferation of deployed AI systems. AI is no longer
confined to the laboratory but instead has become a ubiquitous part of the social world. We rely on AI systems to
help us make decisions as simple as what movie to watch
next or where to go to dinner as well as more complex, highstakes decisions, such as who is able to get a loan or when an
inmate’s sentence should be reduced [Bennett and Lanning,
2007; Kaggle, 2013; Tay and Ho, 1992; Berk, 2012]. We
now share the roads with autonomous vehicles, we routinely
interact with social media bots, and financial markets are now
dominated by algorithmic trading [Bonnefon et al., 2016;
Ferrara et al., 2016].
Due to their ubiquity and potential harm, the increasingly
emergent and often unintended properties of these AI systems
have garnered widespread attention in both the public and
academic spheres [O’Neil, 2017; Friedman and Nissenbaum,
1996]. However, practices such as training on proprietary
data and using complex models often make it challenging to
use the underlying structure of the system to predict or study
its emergent behaviors [Voosen, 2017]. Instead many studies, especially in the area of algorithmic bias, have employed
techniques which do not require details about the system’s
architecture.
In 2013, Sweeney’s study on discrimination showed that
search results were disproportionately more likely to return
results related to arrest when a racially associated name was
used as a keyword in online advertisements [Sweeney, 2013].
This study and others like it began a new wave of hypothesisdriven science related to AI systems.
Since then, an increasing number of studies has attempted
to characterize the emergent behaviors of high-stakes algorithms. A recent study showed that dark, female faces
are misgendered at higher rates than lighter, male faces by
commercial facial recognition algorithms [Buolamwini and
Gebru, 2018]. ProPublica showed evidence of racial dis-
Figure 1: Conceptual illustration of applying the scientific method of
experimentation and causal inference to a black box AI system (center) by providing a stimulus (left) and measuring behavioral output
(right).
crimination in new recidivism risk score algorithms as well
as price discrimination based on zip code for auto insurance premiums [Larson et al., 2016; Larson et al., 2017].
Studies have even explored algorithmic bias on deployed
platforms. For example, recent studies have investigated
price discrimination on e-commerce sites [Chen et al., 2016;
Hannak et al., 2014].
Engineering tools such as multi-agent simulation and ROC
curves are not capable of fully explaining these systems and
their behavior. Instead, we turn to the domain specific knowledge, tools, and methodologies of the social and behavioral
sciences.
2.1
Towards a social science methodology
We propose framing the output of AI systems as behavior,
with its own patterns and ecologies, and propose using scientific techniques like experimentation and causal inference
to understand these behaviors, agnostic to the underlying system architecture. Figure 1 illustrates this framing. Each of the
aforementioned studies in Section 2 contain three core components: an AI system contained in a controlled environment,
systematic stimuli, and a measure of behavior, as shown in
Table 1.
At the core of the social sciences is the attempt to under-
stand and predict the behavior and emergent properties of
complex systems comprised of intelligent agents. Social scientists could contribute several essential pieces of expertise to
questions that lie at the interface of AI and the social sciences.
First, social scientists possess knowledge on experimental design and causal inference. Designing stimuli for AI systems
is analogous to designing and administering more traditional
‘treatments’ in the social and behavioral sciences. Further,
in many cases such as investigating pricing discrimination on
e-commerce sites, the space of potential variables is infinite.
Social scientists conduct randomized controlled trials and derive theoretically informed hypotheses to narrow the space of
their possible behavioral investigations. Second, social scientists bring domain specific knowledge that a typical AI researcher may not have, such as knowledge of specific types of
discrimination, human learning procedures, social dynamics,
among other relevant topics.
With the potential upsides to added collaboration with social and behavioral scientists in mind, we propose a new science of AI which composites previous scientific AI methodologies, the domain expertise of the social sciences, and
social science methodologies such as causal inference and
hypothesis-driven research design.
2.2
The AI Knowledge Gap
But there are significant challenges facing this new science of
AI. These challenges primarily derive from incentives which
encourage the engineering of new systems over the study of
existing systems.
The first challenge relates to the reproducibility crisis, a
popular topic of discussion within the AI community. The
conversation about reproducibility centers around the following problem: as researchers often desire to publish regularly and rapidly, time constraints diminish their incentives
to reproduce results or replicate existing systems. The focus
on contributing novel systems can lead researchers to fail to
replicate their comparative benchmarks, to choose strategic
but non-comprehensive benchmarks, or to never publish the
code for their novel systems. Indeed it was recently found
that only 6% of 400 authors at two top AI conferences shared
their new algorithm’s code [Hutson, 2018].
The second challenge is the expertise required to study AI
systems in their social contexts. Many of the aforementioned
examples require knowledge in computer science as well as
knowledge (and data) related to the social domain of AI systems, such as the criminal justice system. However, for the
same reasons as with the challenge of reproducibility, there
are few incentives for the AI community to gain expertise denovo in every domain of interest, especially for increasingly
complex socio-technical systems. However, as mentioned
previously, social and behavioral scientists – due to their scientific focus on the socio-technical systems themselves – are
uniquely positioned to help fill this gap.
The third challenge is accessibility. The pervasive use
of proprietary data and differences in computational power
can make it difficult for computer scientists to access relevant models. Additionally, while social scientists have unique
training in measuring behavior and social outcomes, they do
not necessarily have the training required to install an exper-
Figure 2: Proxy evidence for an AI Knowledge Gap: the number of
AI agents built grows faster than the studies that characterize those
agents’ behavior.
imental AI system from Github or to invoke a corporate API
directly.
These dynamics have caused the study of existing AI systems to severely lag behind the production of new systems,
as shown in Figure 2. To estimate the pace of algorithm development in machine learning and artificial intelligence, we
study the full set of 7,241 conference papers from the Neural Information Processing Systems (NIPS) Conference from
1987 to 2017. To count the number of papers introducing new
computational models, we search through abstracts and count
the number of papers that contain any of the words ‘new’,
‘novel’, ‘we propose’, and ‘we introduce’ in their abstract.
To count the papers aimed solely at studying previous models, we count the number of papers that include any of the
terms ‘we study’, ‘we examine’, ‘we compare’ and ‘we analyze’ in their abstract, but none of the keywords used to count
new models.
As discussed above, the key bottleneck in the study of the
behavior of AI systems is a misalignment of incentives. To
address these concerns, we propose TuringBox, a market platform that provides access to and incentive for the study of AI
behavior.
3
TuringBox: A Market for the Behavioral
Research of AI Systems
In order to address the above challenges of accessibility, replication, and incentive incompatibility, a general tool for the
rigorous study of AI systems must satisfy many properties.
Drawing from the literature of existing platforms for AI research, we enumerate these desirable properties below:
1. Customizable Metrics: The platform should allow for
the creation of custom evaluation metrics to evaluate AI
systems, and the customizable selection of subsets of
stimuli (data samples or environments). For instance,
users might wish to evaluate custom metrics of fairness,
precision, and accuracy on a variety of algorithms for a
particular task, or to evaluate algorithms on a variety of
subsets of stimuli.
2. Centralized Implementation: Scientists on the platform should be able to evaluate the exact implementation on any arbitrary input sample, along with explicitly outlined hyperparameter settings that produced the
model. This property has been adopted by platforms
such as ParlAI [Miller et al., 2017] and Algorithmia.
3. Centralized Evaluation: The platform should unify
several different benchmarks that exist across communities in a centralized computing environment, without replicating the benchmarks locally. Such an evaluation system has already been adopted by platforms
such as ParlAI [Miller et al., 2017], Algorithmia and
OpenML [Vanschoren et al., 2014].
4. General Compatibility: The platform should have a
consistent, generalizable method to include newer AI
problems and have support for all existing AI problems
of interest. This property is essential to understand the
behavior of classes of algorithms across different tasks
and environments, and provides a cohesive framework
for the study of AI behavior across different classes of
algorithms.
5. Full Scope: AI systems of interest can be accessed at
three different scopes. The first scope consists of algorithms whose code is readily available and can be accessed directly from a scientist’s local machine or from
a centralized platform’s servers. The second scope includes algorithms that are accessible only from remote
servers via API calls. The third scope contains remotely
served algorithms that do not have an explicit API protocol, and thus must be observed “in the field,” via custom scripts, as exemplified in Hannak et al. A sufficient
platform for the behavioral study of AI should provide
functionality for each scope.
With these properties in mind, we envision a two-sided digital marketplace, called TuringBox, that couples two previously orthogonal threads of AI research by convening two
communities (see Fig 3). On one side of the market, AI contributors upload existing and novel algorithms to both benchmark their algorithms with respect to performance, fairness,
or other qualities and to gain reputation in their community.
These contributors will be incentivized to upload not only algorithms they wrote themselves, but also protocols that access APIs and algorithms “in the field.”
On the other side of the market, AI examiners develop and
post tasks designed to examine and characterize algorithmic
behavior. We anticipate that this methodological tool will attract not just computer scientists but will also be of interest
to experts in experimentation across the social and behavioral
sciences, thus mitigating the AI Knowledge Gap.
Notable examples of AI platforms include the DeepMind
PyschLab [Leibo et al., 2018], a fully 3D game-like platform
for agent-based AI research, ParlAI [Miller et al., 2017], a
unified framework for the evaluation of dialog models with
Figure 3: A schematic of the marketplace for AI.
additional support for images, CloudCV [Agrawal et al.,
2015], an open-source platform for machine learning and
computer vision, and OpenML [Vanschoren et al., 2014], an
online machine learning platform for sharing and organizing
data and machine learning algorithms. Each platform satisfies some of the properties listed above, and we map these
platforms to properties in Table 2.
3.1
Incentive Scheme
In our proposed market, both sides are driven by exogenous
and endogenous incentives rather than money. The implementation is engineered to best serve the unique interests of
both types of users. Below we discuss how each side of the
market incentivizes use by the other.
Incentives of examiners
As AI systems further permeate society, we envision increasing demand for the examination of the behaviors of
these systems. The recent increase in studies on algorithmic
bias [Hajian et al., 2016] as well as the success of interdisciplinary conferences such as Fairness, Accountability, and
Transparency or the International Conference on Computational Social Science already suggest an interest from the social science community. In addition to understanding bias,
examiners may also be able to learn about the origins and
structure of complex human behavior, such as strategy formation, as was the case with AlphaGo Zero [Silver et al.,
2016].
However, accessibility is a key concern preventing examiners from studying existing systems. Many social scientists
have limited programming or computer science backgrounds
and cannot implement state of the art AI systems to study.
To address this concern, we employ the second side of the
market (the contributors) who upload their AI systems to the
platform. Examiners can then access these systems from a
central source without added expertise.
Incentives of the contributors
TuringBox motivates a large group of people – the contributors – to populate the platform with algorithms to study.
These contributors are AI researchers in academia and industry as well as other individuals that possess the skills to design
and implement AI systems.
TuringBox will motivate contributors to upload their own
systems by addressing several concerns in the AI community.
Customizable
Metrics
Platform
Centralized
Implementation
Centralized
Evaluation
General
Compatibility
Full
Scope
Algorithmia
CloudCV
DeepMind PsychLab
FairML
OpenML
ParlAI
Themis-ML
TuringBox
Table 2: Platforms for the study of AI and their corresponding properties.
Importantly, the platform will assist with the problem of reproducibility. When contributors upload their new systems,
all other contributors will then be able to access and test these
new systems. Having the designers of the algorithms themselves upload their algorithm is important for two reasons.
First, it prevents issues from arising due to differences in implementations. Secondly, it ensures that the algorithm designers have responsibility for the performance of the algorithm
and any subsequent aberrant behavior it may exhibit.
Further, TuringBox will ease access to benchmarking tools.
With so many AI systems on the platform, contributors will
be able to easily compare their system’s performance to the
performance of other state of the art systems. Moreover, by
incorporating social scientists on the examiner side of the
market, TuringBox will allow researchers to measure their
systems using new social metrics, such as fairness. As examiners study an increasing number of AI systems, we anticipate
they will produce new ways to measure the biases or behaviors of AI systems, which contributors can then use to ensure
their systems operate as expected.
In addition to individual contributors, companies may also
be interested in filling the contributor side of the market via
APIs to demonstrate the performance of their proprietary algorithms and that their algorithms are, for example, bias-free.
4
Discussion
Because of AI systems’ emergent complexity, their ubiquity
in society, and their inherent opacity, there is a need to map
the boundaries of AI research, and to extend them when possible. Our examination suggests there is untapped potential
for the hypothesis driven scientific investigation of AI systems. It is useful to think of AI systems not solely as engineering artifacts, but rather as a new class of protocols with
heterogeneous behavior. In order to keep up with the proliferation of these systems and to close the AI Knowledge Gap,
AI scientists must start to emphasize the generation of knowledge about how these systems behave both in the lab and in
the field. This can be achieved by broadening the scope of
AI research and opening access to the study of AI behavior to
scientists from other disciplines.
We believe a market for the study of machine behavior is
crucial to reach these goals for several reasons. The first is
that it offers a general yet unified framework for understanding when bias occurs in complex architectures. Many studies across computer science, behavioral economics, and legal
studies have already investigated the behavior of AI systems
but each uses an ad hoc approach to collecting data and measuring behavior. Our market provides a toolkit for examining
behavior across a population of AI systems, which enables
a scalable and flexible alternative to costly algorithmic audits. As an increasing number of AI systems are deployed
every day, and their real world stakes increase, a consistent
methodology to perform these algorithmic audits at scale becomes imperative.
The second rationale for the TuringBox platform is that it
helps bridge the gap between computer scientists and social
scientists. Historically, computer scientists and roboticists
were the only ones concerned with the behavior of machines,
because they were the only ones interacting with the systems.
However, social sciences offer many key skills, methods, and
perspectives about AI systems that are not being fully leveraged in the AI research community.
Finally, this framing can potentially curb adversarial effects of emergent superintelligences by providing a controlled
environment to examine their behaviors [Bostrom, 1998]. Indeed this market anticipates a future where the AI systems
that inhabit it exhibit the complex, cross-domain behaviors of
artificial general intelligence.
Amazon’s Mechanical Turk created a revolution for the social sciences by scaling the way social scientists perform experiments [Buhrmester et al., 2011; Horton et al., 2011]. For
the first time, scientists could cheaply run massive online experiments and learn about individual and collective behavior
without recruiting subjects to a physical location. By lowering the barrier to entry for experimentation, the market was
successful in both opening up new methodologies and democratizing research to a broad class of scientists. As a behavioral
science of AI systems becomes conceivable, we believe the
next revolution in AI experimentation will come from standardizing experimental protocols while keeping the hetero-
geneity of these systems wide open.
References
[Agrawal et al., 2015] Harsh Agrawal, Clint Solomon Mathialagan, Yash Goyal, Neelima Chavali, Prakriti Banik,
Akrit Mohapatra, Ahmed Osman, and Dhruv Batra.
Cloudcv: Large-scale distributed computer vision as a
cloud service. In Mobile cloud visual media computing,
pages 265–290. Springer, 2015.
[Bennett and Lanning, 2007] James Bennett and Stan Lanning. The netflix prize. Proceedings of KDD Cup and
Workshop, 2007.
[Berk, 2012] Richard Berk. Criminal justice forecasts of risk
a machine learning approach. 2012.
[Bonnefon et al., 2016] Jean-François Bonnefon, Azim
Shariff, and Iyad Rahwan.
The social dilemma of
autonomous vehicles. Science, 352(6293):1573–1576,
2016.
[Bostrom, 1998] Nick Bostrom. How long before superintelligence? 1998.
[Bowling et al., 2015] Michael Bowling, Neil Burch,
Michael Johanson, and Oskari Tammelin. Heads-up limit
hold’em poker is solved. Science, 347(6218):145–149,
2015.
[Buhrmester et al., 2011] Michael
Buhrmester,
Tracy
Kwang, and Samuel D Gosling. Amazon’s mechanical
turk: A new source of inexpensive, yet high-quality, data?
Perspectives on psychological science, 6(1):3–5, 2011.
[Buolamwini and Gebru, 2018] Joy Buolamwini and Timnit
Gebru. Gender shades: Intersectional accuracy disparities in commercial gender classification. In Conference on
Fairness, Accountability and Transparency, pages 77–91,
2018.
[Bylander, 1994] Tom Bylander. The computational complexity of propositional strips planning. Artificial Intelligence, 69(1-2):165–204, 1994.
[Campbell et al., 2002] Murray Campbell,
A Joseph
Hoane Jr, and Feng-hsiung Hsu. Deep blue. Artificial
intelligence, 134(1-2):57–83, 2002.
[Chen et al., 2016] Le Chen, Alan Mislove, and Christo Wilson. An Empirical Analysis of Algorithmic Pricing on
Amazon Marketplace. In Proceedings of the International World Wide Web Conference (WWW’16), Montréal,
Canada, Apr 2016.
[Cohen, 1995] Paul R Cohen. Empirical methods for artificial intelligence, volume 139. MIT press Cambridge, MA,
1995.
[Davis and Goadrich, 2006] Jesse Davis and Mark Goadrich.
The relationship between precision-recall and roc curves.
In Proceedings of the 23rd international conference on
Machine learning, pages 233–240. ACM, 2006.
[Ferrara et al., 2016] Emilio Ferrara, Onur Varol, Clayton
Davis, Filippo Menczer, and Alessandro Flammini. The
rise of social bots. Commun. ACM, 59(7):96–104, June
2016.
[Friedman and Nissenbaum, 1996] Batya Friedman and Helen Nissenbaum. Bias in computer systems. ACM Trans.
Inf. Syst., 14(3):330–347, July 1996.
[Hajian et al., 2016] Sara Hajian, Francesco Bonchi, and
Carlos Castillo. Algorithmic bias: From discrimination
discovery to fairness-aware data mining. In Proceedings
of the 22nd ACM SIGKDD international conference on
knowledge discovery and data mining, pages 2125–2126.
ACM, 2016.
[Hannak et al., 2014] Aniko Hannak, Gary Soeller, David
Lazer, Alan Mislove, and Christo Wilson. Measuring Price
Discrimination and Steering on E-commerce Web Sites.
In Proceedings of the ACM Internet Measurement Conference (IMC’14), Vancouver, Canada, Nov 2014.
[Hilbert, 1928] David Hilbert. Die grundlagen der mathematik. In Die Grundlagen der Mathematik, pages 1–21.
Springer, 1928.
[Horton et al., 2011] John J Horton, David G Rand, and
Richard J Zeckhauser. The online laboratory: Conducting experiments in a real labor market. Experimental economics, 14(3):399–425, 2011.
[Hutson, 2018] Matthew Hutson. Artificial intelligence faces
reproducibility crisis, 2018.
[Jackson, 1998] Peter Jackson. Introduction to expert systems. Addison-Wesley Longman Publishing Co., Inc.,
1998.
[Jaffar and Maher, 1994] Joxan Jaffar and Michael J Maher.
Constraint logic programming: A survey. The journal of
logic programming, 19:503–581, 1994.
[Kaggle, 2013] Kaggle. 2013.
[Kakas et al., 1992] Antonis C. Kakas, Robert A. Kowalski,
and Francesca Toni. Abductive logic programming. Journal of logic and computation, 2(6):719–770, 1992.
[Kautz et al., 1992] Henry A Kautz, Bart Selman, et al. Planning as satisfiability. In ECAI, volume 92, pages 359–363.
Citeseer, 1992.
[Kitano et al., 1997] Hiroaki Kitano, Minoru Asada, Yasuo
Kuniyoshi, Itsuki Noda, and Eiichi Osawa. Robocup: The
robot world cup initiative. In Proceedings of the first international conference on Autonomous agents, pages 340–
347. ACM, 1997.
[Larson et al., 2016] J Larson, S Mattu, L Kirchner, and
J Angwin. How we analyzed the compas recidivism algorithm. propublica, 2016.
[Larson et al., 2017] J Larson, J Angwin, L Kirchner, and
S Mattu. How we examined racial discrimination in auto
insurance prices. propublica, 2017.
[Leibniz, 1685] Gottfried Leibniz. The art of discovery. In
Leibniz: Selections. Scribner, 1951, 1685.
[Leibo et al., 2018] Joel Z Leibo, Cyprien de Masson
d’Autume, Daniel Zoran, David Amos, Charles Beattie, Keith Anderson, Antonio Garcı́a Castañeda, Manuel
Sanchez, Simon Green, Audrunas Gruslys, et al. Psychlab:
A psychology laboratory for deep reinforcement learning
agents. arXiv preprint arXiv:1801.08116, 2018.
[Lenat et al., 1985] Douglas B Lenat, Mayank Prakash, and
Mary Shepherd. Cyc: Using common sense knowledge
to overcome brittleness and knowledge acquisition bottlenecks. AI magazine, 6(4):65, 1985.
[Lin et al., 2014] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr
Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014.
[Littman, 1994] Michael L Littman. Markov games as a
framework for multi-agent reinforcement learning. In Machine Learning Proceedings 1994, pages 157–163. Elsevier, 1994.
[McCarthy, 1981] John McCarthy. Circumscription—a form
of non-monotonic reasoning. In Readings in Artificial Intelligence, pages 466–472. Elsevier, 1981.
[McCarthy, 1986] John McCarthy. Applications of circumscription to formalizing common-sense knowledge. Artificial Intelligence, 28(1):89–116, 1986.
[Miller et al., 2017] Alexander H Miller, Will Feng, Adam
Fisch, Jiasen Lu, Dhruv Batra, Antoine Bordes, Devi
Parikh, and Jason Weston. Parlai: A dialog research software platform. arXiv preprint arXiv:1705.06476, 2017.
[Mnih et al., 2015] 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,
2015.
[Newell et al., 1959] Allen Newell, John C Shaw, and Herbert A Simon. Report on a general problem solving program. In IFIP congress, volume 256, page 64, 1959.
[O’Neil, 2017] Cathy O’Neil. Weapons of math destruction:
How big data increases inequality and threatens democracy. Broadway Books, 2017.
[Papineni et al., 2002] Kishore Papineni, Salim Roukos,
Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of
the 40th annual meeting on association for computational
linguistics, pages 311–318. Association for Computational
Linguistics, 2002.
[Pearl, 1986] Judea Pearl. Fusion, propagation, and structuring in belief networks. Artificial intelligence, 29(3):241–
288, 1986.
[Russakovsky et al., 2015] Olga Russakovsky, Jia Deng,
Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma,
Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael
Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision,
115(3):211–252, 2015.
[Schaeffer et al., 2007] Jonathan Schaeffer, Neil Burch, Yngvi Björnsson, Akihiro Kishimoto, Martin Müller, Robert
Lake, Paul Lu, and Steve Sutphen. Checkers is solved.
science, 317(5844):1518–1522, 2007.
[Silver et al., 2016] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game
of go with deep neural networks and tree search. nature,
529(7587):484–489, 2016.
[Sussman and Winograd, 1970] Gerald Sussman and Terry
Winograd. Micro-planner reference manual. 1970.
[Sutton and Barto, 1998] Richard S Sutton and Andrew G
Barto. Introduction to reinforcement learning, volume
135. MIT press Cambridge, 1998.
[Sweeney, 2013] Latanya Sweeney. Discrimination in online
ad delivery. Queue, 11(3):10, 2013.
[Tay and Ho, 1992] Danny P.H. Tay and David K.H. Ho. Artificial intelligence and the mass appraisal of residential
apartments. Journal of Property Valuation and Investment,
10(2):525–540, 1992.
[Turing, 1950] Alan M Turing. Computing machinery and
intelligence. Mind, pages 433–460, 1950.
[Van De Sande et al., 2010] Koen Van De Sande, Theo Gevers, and Cees Snoek. Evaluating color descriptors for
object and scene recognition. IEEE transactions on pattern analysis and machine intelligence, 32(9):1582–1596,
2010.
[Van Emden and Kowalski, 1976] Maarten H Van Emden
and Robert A Kowalski. The semantics of predicate logic
as a programming language. Journal of the ACM (JACM),
23(4):733–742, 1976.
[Vanschoren et al., 2014] Joaquin Vanschoren, Jan N
Van Rijn, Bernd Bischl, and Luis Torgo. Openml:
networked science in machine learning. ACM SIGKDD
Explorations Newsletter, 15(2):49–60, 2014.
[Vapnik, 1999] Vladimir Naumovich Vapnik. An overview
of statistical learning theory. IEEE transactions on neural
networks, 10(5):988–999, 1999.
[Voosen, 2017] Paul Voosen. The ai detectives. Science,
357(6346):22–27, 2017.
[Wellman et al., 2001] Michael P. Wellman, Peter R. Wurman, Kevin O’Malley, Roshan Bangera, Daniel Reeves,
William E Walsh, et al. Designing the market game for
a trading agent competition. IEEE Internet Computing,
5(2):43–51, 2001.
[Wielinga et al., 1992] Bob J Wielinga, A Th Schreiber, and
Jost A Breuker. KADS: A modelling approach to knowledge engineering. Knowledge acquisition, 4(1):5–53,
1992.
[Zhou et al., 2017] Zhiming Zhou, Weinan Zhang, and Jun
Wang.
Inception score, label smoothing, gradient
vanishing and-log (d (x)) alternative. arXiv preprint
arXiv:1708.01729, 2017.
| 2cs.AI
|
Single Shot Temporal Action Detection
1 Department
Tianwei Lin1 , Xu Zhao1,3, *, Zheng Shou2
of Automation, Shanghai Jiao Tong University, China. 2 Columbia University, USA
Medianet Innovation Center (CMIC), Shanghai Jiao Tong University, China
{wzmsltw,zhaoxu}@sjtu.edu.cn,[email protected]
3 Cooperative
arXiv:1710.06236v1 [cs.CV] 17 Oct 2017
ABSTRACT
Temporal action detection is a very important yet challenging problem, since videos in real applications are usually long, untrimmed
and contain multiple action instances. This problem requires not
only recognizing action categories but also detecting start time and
end time of each action instance. Many state-of-the-art methods
adopt the "detection by classification" framework: first do proposal,
and then classify proposals. The main drawback of this framework
is that the boundaries of action instance proposals have been fixed
during the classification step. To address this issue, we propose
a novel Single Shot Action Detector (SSAD) network based on
1D temporal convolutional layers to skip the proposal generation
step via directly detecting action instances in untrimmed video.
On pursuit of designing a particular SSAD network that can work
effectively for temporal action detection, we empirically search
for the best network architecture of SSAD due to lacking existing models that can be directly adopted. Moreover, we investigate
into input feature types and fusion strategies to further improve
detection accuracy. We conduct extensive experiments on two challenging datasets: THUMOS 2014 and MEXaction2. When setting
Intersection-over-Union threshold to 0.5 during evaluation, SSAD
significantly outperforms other state-of-the-art systems by increasing mAP from 19.0% to 24.6% on THUMOS 2014 and from 7.4% to
11.0% on MEXaction2.
CCS CONCEPTS
• Computing methodologies → Activity recognition and understanding;
KEYWORDS
Temporal Action Detection, Untrimmed Video, SSAD network
1
INTRODUCTION
Due to the continuously booming of videos on the internet, video
content analysis has attracted wide attention from both industry
and academic field in recently years. An important branch of video
content analysis is action recognition, which usually aims at classifying the categories of manually trimmed video clips. Substantial
This research has been supported by the funding from NSFC (61673269, 61273285) and
the Cooperative Medianet Innovation Center (CMIC). * Corresponding author.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
MM ’17, Mountain View, CA, USA
© 2017 ACM. 978-1-4503-4906-2/17/10. . . $15.00
DOI: 10.1145/3123266.3123343
Figure 1: Overview of our system. Given an untrimmed long
video, (1) we extract Snippet-level Action Score features sequence with multiple action classifiers; (2) SSAD network
takes feature sequence as input and directly predicts multiple scales action instances without proposal generation step.
progress has been reported for this task in [6, 24, 36, 38, 40]. However, most videos in real world are untrimmed and may contain
multiple action instances with irrelevant background scenes or activities. This problem motivates the academic community to put
attention to another challenging task - temporal action detection.
This task aims to detect action instances in untrimmed video, including temporal boundaries and categories of instances. Methods
proposed for this task can be used in many areas such as surveillance video analysis and intelligent home care.
Temporal action detection can be regarded as a temporal version
of object detection in image, since both of the tasks aim to determine the boundaries and categories of multiple instances (actions
in time/ objects in space). A popular series of models in object
detection are R-CNN and its variants [8, 9, 27], which adopt the
"detect by classifying region proposals" framework. Inspired by
R-CNN, recently many temporal action detection approaches adopt
similar framework and classify temporal action instances generated
by proposal method [3, 5, 29, 43] or simple sliding windows method
[15, 23, 39]. This framework may has some major drawbacks: (1)
proposal generation and classification procedures are separate and
have to be trained separately, but ideally we want to train them in
a joint manner to obtain an optimal model; (2) the proposal generation method or sliding windows method requires additional time
consumption; (3) the temporal boundaries of action instances generated by the sliding windows method are usually approximative
rather than precise and left to be fixed during classification. Also,
since the scales of sliding windows are pre-determined, it is not
flexible to predict instances with various scales.
To address these issues, we propose the Single Shot Action Detector (SSAD) network, which is a temporal convolutional network
conducted on feature sequence with multiple granularities. Inspired
by another set of object detection methods - single shot detection
models such as SSD [20] and YOLO [25, 26], our SSAD network
skips the proposal generation step and directly predicts temporal
boundaries and confidence scores for multiple action categories,
as shown in Figure 1. SSAD network contains three sub-modules:
(1) base layers read in feature sequence and shorten its temporal
length; (2) anchor layers output temporal feature maps, which are
associated with anchor action instances; (3) prediction layers generate categories probabilities, location offsets and overlap scores of
these anchor action instances.
For better encoding of both spatial and temporal information in
video, we adopt multiple action recognition models (action classifiers) to extract multiple granularities features. We concatenate the
output categories probabilities from all action classifiers in snippetlevel and form the Snippet-level Action Score (SAS) feature. The
sequences of SAS features are used as input of SSAD network.
Note that it is non-trivial to adapt the single shot detection model
from object detection to temporal action detection. Firstly, unlike
VGGNet [31] being used in 2D ConvNet models, there is no existing
widely used pre-trained temporal convolutional network. Thus in
this work, we search multiple network architectures to find the
best one. Secondly, we integrate key advantages in different single
shot detection models to make our SSAD network work the best.
On one hand, similar to YOLO9000 [26], we simultaneously predict
location offsets, categories probabilities and overlap score of each
anchor action instance. On the other hand, like SSD [20], we use
anchor instances of multiple scale ratios from multiple scales feature
maps, which allow network flexible to handle action instance with
various scales. Finally, to further improve performance, we fuse
the prediction categories probability with temporal pooled snippetlevel action scores during prediction.
The main contributions of our work are summarized as follows:
(1) To the best of our knowledge, our work is the first Single Shot
Action Detector (SSAD) for video, which can effectively predict both
the boundaries and confidence score of multiple action categories
in untrimmed video without the proposal generation step.
(2) In this work, we explore many configurations of SSAD network such as input features type, network architectures and postprocessing strategy. Proper configurations are adopted to achieve
better performance for temporal action detection task.
(3) We conduct extensive experiments on two challenging benchmark datasets: THUMOS’14 [14] and MEXaction2 [1]. When setting
Intersection-over-Union threshold to 0.5 during evaluation, SSAD
significantly outperforms other state-of-the-art systems by increasing mAP from 19.0% to 24.6% on THUMOS’14 and from 7.4% to
11.0% on MEXaction2.
2
RELATED WORK
Action recognition. Action recognition is an important research
topic for video content analysis. Just as image classification network
can be used in image object detection, action recognition models
can be used in temporal action detection for feature extraction. We
mainly review the following methods which can be used in temporal
action detection. Improved Dense Trajectory (iDT) [37, 38] feature
is consisted of MBH, HOF and HOG features extracted along dense
trajectories. iDT method uses SIFT and optical flow to eliminate
the influence of camera motion. Two-stream network [6, 30, 40]
learns both spatial and temporal features by operating network
on single frame and stacked optical flow field respectively using
2D Convolutional Neural Network (CNN) such as GoogleNet [35],
VGGNet [31] and ResNet [12]. C3D network [36] uses 3D convolution to capture both spatial and temporal information directly from
raw video frames volume, and is very efficient. Feature encoding
methods such as Fisher Vector [38] and VAE [24] are widely used
in action recognition task to improve performance. And there are
many widely used action recognition benchmark such as UCF101
[34], HMDB51 [18] and Sports-1M [16].
Temporal action detection. This task focuses on learning how
to detect action instances in untrimmed videos where the boundaries and categories of action instances have been annotated. Typical datasets such as THUMOS 2014 [14] and MEXaction2 [1] include
large amount of untrimmed videos with multiple action categories
and complex background information.
Recently, many approaches adopt "detection by classification"
framework. For examples, many approaches [15, 23, 33, 39, 41]
use extracted feature such as iDT feature to train SVM classifiers,
and then classify the categories of segment proposals or sliding
windows using SVM classifiers. And there are some approaches
specially proposed for temporal action proposal [3, 5, 7, 22, 43]. Our
SSAD network differs from these methods mainly in containing no
proposal generation step.
Recurrent Neural Network (RNN) is widely used in many action
detection approaches [21, 32, 42, 44] to encode feature sequence
and make per-frame prediction of action categories. However, it is
difficult for RNNs to keep a long time period memory in practice
[32]. An alternative choice is temporal convolution. For example,
Lea et al. [19] proposes Temporal Convolutional Networks (TCN)
for temporal action segmentation. We also adopt temporal convolutional layers, which makes our SSAD network can handle action
instances with a much longer time period.
Object detection. Deep learning approaches have shown salient
performance in object detection. We will review two main set of object detection methods proposed in recent years. The representative
methods in first set are R-CNN [9] and its variations [8, 27]. R-CNN
uses selective search to generate multiple region proposals then
apply CNN in these proposals separately to classify their categories;
Fast R-CNN [8] uses a 2D RoI pooling layer which makes feature
map be shared among proposals and reduces the time consumption. Faster RCNN [27] adopts a RPN network to generate region
proposal instead of selective search.
Another set of object detection methods are single shot detection
methods, which means detecting objects directly without generating proposals. There are two well known models. YOLO [25, 26]
uses the whole topmost feature map to predict probabilities of multiple categories and corresponding confidence scores and location
offsets. SSD [20] makes prediction from multiple feature map with
multiple scales default boxes. In our work, we combine the characteristics of these single shot detection methods and embed them
into the proposed SSAD network.
Figure 2: The framework of our approach. (a) Multiple action classifiers are used to extract Snippet-level Action Scores (SAS)
feature. (b) The architecture of SSAD network: base layers are used to reduce the temporal dimension of input data; anchor
layers output multiple scale feature map associated with anchor instances and prediction layers are used for predicting categories, location and confidence of anchor instances. (c) The training and prediction procedures: during training, we match
anchor instances with ground truth instances and calculate loss function for optimization. During prediction, post-processing
and NMS procedure are conducted on anchor instances to make final prediction.
3
OUR APPROACH
In this section, we will introduce our approach in details. The framework of our approach is shown in Figure 2.
3.1
Problem Definition
v
We denote a video as Xv = {x t }Tt =1
where Tv is the number of
frames in Xv and x t is the t-th frame in Xv . Each untrimmed
video Xv is annotated with a set of temporal action instances
Nv
Φv = ϕ n = φ n , φ n0 , kn n=1
, where Nv is the number of temporal
action instances in Xv , and φ n , φ n0 , kn are starting time, ending time
and category of action instance ϕ n respectively. kn ∈ {1, ..., K }
where K is the number of action categories. Φv is given during
training procedure and need to be predicted during prediction procedure.
3.2
Extracting of Snippet-level Action Scores
To apply SSAD model, first we need to make snippet-level action
classification and get Snippet-level Action Score (SAS) features.
Given a video Xv , a snippet st = (x t , Ft , X t ) is composed by three
parts: x t is the t-th frame in Xv , Ft = { ft 0 }tt 0+5
=t −4 is stacked optical
flow field derived around x t and X t = {x t 0 }tt 0+8
=t −7 is video frames
volume. So given a video Xv , we can get a sequence of snippets
v
Sv = {st }Tt =1
. We pad the video Xv in head and tail with first and
last frame separately to make Sv have the same length as Xv .
Action classifier. To evaluate categories probability of each
snippet, we use multiple action classifiers with commendable performance in action recognition task: two-stream network [30] and
C3D network [36]. Two-stream network includes spatial and temporal networks which operate on single video frame x t and stacked
optical flow field Ft respectively. We use the same two-stream network architecture as described in [40], which adopts VGGNet-16
network architecture. C3D network is proposed in [36], including
multiple 3D convolution layers and 3D pooling layers. C3D network
operates on short video frames volume X t with length l, where l is
the length of video clip and is set to 16 in C3D. So there are totally
three individual action classifiers, in which spatial network measures the spatial information, temporal network measures temporal
consistency and C3D network measures both. In section 4.3, we
evaluate the effect of each action classifier and their combinations.
SAS feature. As shown in Figure 2(a), given a snippet st , each
action classifier can generate a score vector p t with length K 0 =
K + 1, where K 0 includes K action categories and one background
category. Then we concatenate output scores of each classifiers
to
Action Score (SAS) feature psas,t =
form the Snippet-level
p S,t , pT ,t , pC ,t , where p S,t , pT ,t , pC ,t are output score of spatial, temporal and C3D network separately. So given a snippets
sequence Sv with length Tv , we can extract a SAS feature sequence
Tv
Pv = psas,t t =1
. Since the number of frames in video is uncertain
and may be very large, we use a large observation window with
length
We denote a window as
Tw to0 truncate the feature sequence.
0 are starting and ending
ω = φω , φω
, Pω , Φω , where φ ω and φ ω
time of ω, Pω and Φω are SAS feature sequence and corresponding
ground truth action instances separately.
3.3
SSAD Network
Temporal action detection is quite different from object detection
in 2D image. In SSAD we adopt two main characteristics from
single shot object detection models such as SSD [20] and YOLO
[25, 26]: 1) unlike "detection by classification" approaches, SSAD
directly predicts categories and location offsets of action instances
in untrimmed video using convolutional prediction layers; 2) SSAD
combine temporal feature maps from different convolution layers
for prediction, making it possible to handle action instances with
various length. We first introduce the network architecture.
Network architecture. The architecture of SSAD network is
presented in Figure 2(b), which mainly contains three sub-modules:
base layers, anchor layers and prediction layers. Base layers handle
the input SAS feature sequence, and use both convolution and
pooling layer to shorten the temporal length of feature map and
increase the size of receptive fields. Then anchor layers use temporal
convolution to continually shorten the feature map and output
anchor feature map for action instances prediction. Each cell of
anchor layers is associated with anchor instances of multiple scales.
Finally, we use prediction layers to get classification score, overlap
score and location offsets of each anchor instance.
In SSAD network, we adopt 1D temporal convolution and pooling to capture temporal information. We conduct Rectified Linear
Units (ReLu) activation function [11] to output temporal feature
map except for the convolutional prediction layers. And we adopt
temporal max pooling since max pooling can enhance the invariance of small input change.
Base layers. Since there are no widely used pre-trained 1D
ConvNet models such as the VGGNet [31] used in 2D ConvNet
models, we search many different network architectures for SSAD
network. These architectures only differ in base layers while we
keep same architecture of anchor layers and prediction layers. As
shown in Figure 3, we totally design 5 architectures of base layers. In
these architectures, we mainly explore three aspects: 1) whether use
convolution or pooling layer to shorten the temporal dimension and
increase the size of receptive fields; 2) number of layers of network
and 3) size of convolution layer’s kernel. Notice that we set the
number of convolutional filter in all base layers to 256. Evaluation
results of these architectures are shown in section 4.3, and finally
we adopt architecture B which achieves the best performance.
Multi-scale anchor layers. After processing SAS feature sequence using base layers, we stack three anchor convolutional
layers (Conv-A1, Conv-A2 and Conv-A3) on them. These layers
have same configuration: kernel size 3, stride size 2 and 512 convolutional filters. The output anchor feature maps of anchor layers
are f A1 , f A2 and f A3 with size (Tw /32 × 512), (Tw /64 × 512) and
(Tw /128 × 512) separately. Multiple anchor layers decrease temporal dimension of feature map progressively and allow SSAD get
predictions from multiple resolution feature map.
Figure 3: Multiple architectures of base layers. Input and output sizes are same for each architecture. Parameter of layer
is shown with the format of kernel/stride. All convolutional
layers have 512 convolutional filters. Evaluation results of
these architectures are shown in section 4.3, and we adopt
architecture B which achieves the best performance.
For each temporal feature map of anchor layers, we associate a
set of multiple scale anchor action instances with each feature map
cell as shown in Figure 4. For each anchor instance, we use convolutional prediction layers to predict overlap score, classification
score and location offsets, which will be introduced later.
In term of the details of multi-scale anchor instances, the lower
anchor feature map has higher resolution and smaller receptive
field than the top anchor feature map. So we let the lower anchor
layers detect short action instances and the top anchor layers detect
long action instances. For a temporal feature map f of anchor layer
with length M, we define base scale s f = M1 and a set of scale
D
f
ratios R f = {rd }d =1
, where D f is the number of scale ratios. We
use {1, 1.5, 2} for f A1 and {0.5, 0.75, 1, 1.5, 2} for f A2 and f A3 . For
each ratio rd , we calculate µw = s f · rd as anchor instance’s default
width. And all anchor instances associated with the m-th feature
map cell share the same default center location µc = m+0.5
M . So for
an anchor feature map f with length M f and D f scale ratios, the
number of associated anchor instances is M f · D f .
Prediction layers. We use a set of convolutional filters to predict classification scores, overlap scores and location offsets of
anchor instances associated with each feature map cell. As shown
in Figure 4, for an anchor feature map f with length M f and D f
scale ratios, we use D f · (K 0 + 3) temporal convolutional filters
with kernel size 3, stride
size 1 for prediction.
The output of pre-
diction layer has size M f × D f · (K 0 + 3) and can be reshaped
into M f · D f × (K 0 + 3) . Each anchor instance gets a prediction
score vector ppr ed = pc l ass , pover , ∆c, ∆w with length (K 0 + 3),
where pc l ass is classification score vector with length K 0 , pover is
overlap score and ∆c, ∆w are location offsets. Classification score
pcl ass is used to predict anchor instance’s category. Overlap score
pover is used to estimate the overlap between anchor instance and
ground truth instances and should have value between [0, 1], so it
is normalized by using sigmoid function:
0
pover
= siдmoid(pover ).
(1)
Figure 4: Anchor instances and prediction layer in temporal
feature map. In feature map of a anchor layer, we associate
a set of multiple scale anchor instances with each feature
map cell. We use convolutional prediction layer to predict
location offset, confidence and classification scores simultaneously for each anchor instance.
And location offsets ∆c, ∆w are used for adjusting the default
location of anchor instance. The adjusted location is defined as:
φc = µc + α 1 · µw · ∆c
φw = µw · exp(α 2 · ∆w),
(2)
where φc and φw are center location and width of anchor instance
respectively. α 1 and α 2 are used for controlling the effect of location
offsets to make prediction stable. We set both α 1 and α 2 to 0.1. The
starting and ending time of action instance are φ = φc − 12 · φw and
φ 0 = φc + 12 ·φw respectively. So for a anchor feature map f , we can
Nf
0
get a anchor instances set Φf = ϕ n = φc , φw , pcl ass , pover
n=1 ,
where N f = M f · D f is the number of anchor instances. And the
total prediction instances set is Φp = Φf A1 , Φf A2 , Φf A3 .
3.4
regard it as positive, otherwise negative. We expand ϕ n with match
0
ing information as ϕ n0 = φc , φw , pc l ass , pover
, kд , дiou , дc , дw ,
where kд is the category of ϕд and is set to 0 for negative instance,
дiou is the IoU overlap between ϕ n and ϕд , дc and дw are center
location and width of ϕд respectively. So a ground truth instance
can match multiple anchor instances while a anchor instance can
only match one ground truth instance at most.
Hard negative mining. During label assignment, only a small
part of anchor instances match the ground truth instances, causing
an imbalanced data ratio between the positive and negative instances. Thus we adopt the hard negative mining strategy to reduce
the number of negative instances. Here, the hard negative instances
are defined as negative instances with larger overlap score than
0.5. We take all hard negative instances and randomly sampled
negative instances in remaining part to make the ratio between
positive and negative instances be nearly 1:1. This ratio is chosen
by empirical validation. So after label assignment and hard negative
N t r ain
mining, we get Φp0 = ϕ n0 n=1
as the input set during training,
where Nt r ain is the number of total training instances and is the
sum of the number of positives Npos and negatives Nneд .
Objective for training. The training objective of the SSAD network is to solve a multi-task optimization problem. The overall
loss function is a weighted sum of the classification loss (class), the
overlap loss (conf), the detection loss (loc) and L2 loss for regularization:
Training of SSAD network
Training data construction. As described in Section 3.2, for an
untrimmed video Xv with length Tv , we get SAS features sequence
Pv with same length. Then we slide window of length Tw in feature
sequence with 75% overlap. The overlap of sliding window is aim
to handle the situation where action instances locate in boundary
of window and also used to increase the amount of training data.
During training, we only keep windows containing at least one
ground-truth instance. So given a set of untrimmed training videos,
Nω
we get a training set Ω = {ωn }n=1
, where N ω is the number of
windows. We randomly shuffle the data order in training set to
make the network converge faster, where same random seed is
used during evaluation.
Label assignment. During training, given a window ω, we can
get prediction instances set Φp via SSAD network. We need to match
them with ground truth set Φω for label assignment. For an anchor
instance ϕ n in Φp , we calculate it’s IoU overlap with all ground
truth instances in Φω . If the highest IoU overlap is higher than 0.5,
we match ϕ n with corresponding ground truth instance ϕд and
L = Lcl ass + α · Lover + β · Lloc + λ · L 2 (Θ),
(3)
where α, β and λ are the weight terms used for balancing each part
of loss function. Both α and β are set to 10 and λ is set to 0.0001 by
empirical validation. For the classification loss, we use conventional
softmax loss over multiple categories, which is effective for training
classification model and can be defined as:
Lcl ass = Lsof tmax =
(kд )
where Pi
1
NÕ
t r ain
Nt r ain
i=1
(kд )
(−loд(Pi
)),
(4)
(kд )
exp(pcl as s,i )
= Í
(k j )
j
exp(pcl as s,i )
and kд is the label of this instance.
Lover is used to make a precise prediction of anchor instances’
overlap IoU score, which helps the procedure of NMS. The overlap
loss adopts the mean square error (MSE) loss and be defined as:
Lover =
1
NÕ
t r ain
Nt r ain
i=1
0
(pover,i
− дiou,i ).
(5)
Lloc is the Smooth L1 loss [8] for location offsets. We regress the
center (ϕc ) and width (ϕw ) of predicted instance:
Lloc =
1
NÕ
pos
Npos i=1
(SL 1 (ϕc,i − дc,i ) + SL 1 (ϕw,i − дw,i )),
(6)
where дc,i and дw,i is the center location and width of ground truth
instance. L 2 (Θ) is the L2 regularization loss where Θ stands for the
parameter of the whole SSAD network.
3.5
Prediction and post-processing
During prediction, we follow the aforementioned data preparation
method during the training procedure to prepare test data, with the
following two changes: (1) the overlap ratio of window is reduced
to 25% to increase the prediction speed and reduce the redundant
predictions; (2) instead of removing windows without annotation,
we keep all windows during prediction because the removing operation is actually a leak of annotation information. If the length of
input video is shorter than Tw , we will pad SAS feature sequence
to Tw so that there is at least one window for prediction. Given a
Nω
video Xv , we can get a set of Ω = {ωn }n=1
. Then we use SSAD
network to get prediction anchors of each window and merge these
Np
prediction as Φp = {ϕ n }n=1
, where Np is the number of prediction
instances. For a prediction anchor instance ϕ n in Φp , we calculate the mean Snippet-level Action Score p̄sas among the temporal
range of instance and multiple action classifiers.
0
p̄sas =
φ
Õ
1
p S,t + pT ,t + pC,t ,
3 · (φ 0 − φ) t =φ
(7)
where φ and φ 0 are starting and ending time of prediction anchor
instance respectively. Then we fuse categories scores p̄sas and
pcl ass with multiplication factor pconf and get the p f inal :
0
p f inal = pover
· (pcl ass + p̄sas ) .
(8)
We choose the maximum dimension kp in p f inal as the category
of ϕ n and corresponding score pconf as the confidence score. We
expand ϕ n as ϕ n0 = φc , φw , pconf , kp and get prediction set Φp0 =
0 Np
ϕ n n=1 . Then we conduct non-maximum suppress (NMS) in these
prediction results to remove redundant predictions with confidence
Np 0
score pconf and get the final prediction instances set Φp00 = ϕ n0 n=1
,
where Np 0 is the number of the final prediction anchors. Since there
are little overlap between action instances of same category in
temporal action detection task, we take a strict threshold in NMS,
which is set to 0.1 by empirical validation.
4 EXPERIMENTS
4.1 Dataset and setup
THUMOS 2014 [14]. The temporal action detection task of THUMOS 2014 dataset is challenging and widely used. The training
set is the UCF-101 [34] dataset including 13320 trimmed videos of
101 categories. The validation and test set contain 1010 and 1574
untrimmed videos separately. In temporal action detection task,
only 20 action categories are involved and annotated temporally.
We only use 200 validation set videos (including 3007 action instances) and 213 test set videos (including 3358 action instances)
with temporal annotation to train and evaluate SSAD network.
MEXaction2 [1]. There are two action categories in MEXaction2 dataset: "HorseRiding" and "BullChargeCape". This dataset
is consisted of three subsets: YouTube clips, UCF101 Horse Riding clips and INA videos. YouTube and UCF101 Horse Riding clips
are trimmed and used for training set, whereas INA videos are
untrimmed with approximately 77 hours in total and are divided
into training, validation and testing set. Regarding to temporal annotated action instances, there are 1336 instances in training set,
310 instances in validation set and 329 instances in testing set.
Evaluation metrics. For both datasets, we follow the conventional metrics used in THUMOS’14, which evaluate Average Precision (AP) for each action categories and calculate mean Average
Precision (mAP) for evaluation. A prediction instance is correct if
it gets same category as ground truth instance and its temporal
IoU with this ground truth instance is larger than IoU threshold θ .
Various IoU thresholds are used during evaluation. Furthermore,
redundant detections for the same ground truth are forbidden.
4.2
Implementation Details
Action classifiers. To extract SAS features, action classifiers should
be trained first, including two-stream networks [40] and C3D network [36]. We implement both networks based on Caffe [13]. For
both MEXaction and THUMOS’14 datasets, we use trimmed videos
in training set to train action classifier.
For spatial and temporal network, we follow the same training
strategy described in [40] which uses the VGGNet-16 pre-trained on
ImageNet [4] to intialize the network and fine-tunes it on training
set. And we follow [36] to train the C3D network, which is pretrained on Sports-1M [16] and then is fine-turned on training set.
SSAD optimization. For training of the SSAD network, we use
the adaptive moment estimation (Adam) algorithm [17] with the
aforementioned multi-task loss function. Our implementation is
based on Tensorflow [2]. We adopt the Xavier method [10] to randomly initialize parameters of whole SSAD network because there
are no suitable pre-trained temporal convolutional network. Even
so, the SSAD network can be easily trained with quick convergence
since it has a small amount of parameters (20 MB totally) and the
input of SSAD network - SAS features are concise high-level feature.
The training procedure takes nearly 1 hour on THUMOS’14 dataset.
Table 1: mAP results on THUMOS’14 with various IoU
threshold θ used in evaluation.
θ
0.5
0.4
0.3
0.2
0.1
Karaman et al. [15]
Wang et al. [39]
Oneata et al. [23]
Richard et al. [28]
Yeung et al. [42]
Yuan et al. [44]
Shou et al. [29]
Zhu et al. [45]
0.2
8.5
15.0
15.2
17.1
18.8
19.0
19.0
0.3
12.1
21.8
23.2
26.4
26.1
28.7
28.9
0.5
14.6
28.8
30.0
36.0
33.6
36.3
36.2
0.9
17.8
36.2
35.7
44.0
42.6
43.5
43.6
1.5
19.2
39.8
39.7
48.9
51.4
47.7
47.7
SSAD
24.6
35.0
43.0
47.8
50.1
4.3
Comparison with state-of-the-art systems
Results on THUMOS 2014. To train action classifiers, we use full
UCF-101 dataset. Instead of using one background category, here
we form background categories using 81 action categories which
are un-annotated in detection task. Using two-stream and C3D
networks as action classifiers, the dimension of SAS features is 303.
Figure 5: Detection AP over different action categories with overlap threshold 0.5 in THUMOS’14.
Table 2: Results on MEXaction2 dataset with overlap threshold 0.5. Results for [1] are taken from [29].
AP(%)
BullCHargeCape
HorseRiding
mAP(%)
DTF [1]
SCNN [29]
0.3
11.6
3.1
3.1
1.7
7.4
SSAD
16.5
5.5
11.0
For training of SSAD model, we use 200 annotated untrimmed
video in THUMOS’14 validation set as training set. The window
length Lw is set to 512, which means approximately 20 seconds of
video with 25 fps. This choice is based on the fact that 99.3% action
instances in the training set have smaller length than 20 seconds.
We train SSAD network for 30 epochs with learning rate of 0.0001.
The comparison results between our SSAD and other state-ofthe-art systems are shown in Table 1 with multiple overlap IoU
thresholds varied from 0.1 to 0.5. These results show that SSAD
significantly outperforms the compared state-of-the-art methods.
While the IoU threshold used in evaluation is set to 0.5, our SSAD
network improves the state-of-the-art mAP result from 19.0% to
24.6%. The Average Precision (AP) results of all categories with
overlap threshold 0.5 are shown in Figure 5, the SSAD network
outperforms other state-of-the-art methods for 7 out of 20 action
categories. Qualitative results are shown in Figure 6.
Results on MEXaction2. For training of action classifiers, we
use all 1336 trimmed video clips in training set. And we randomly
sample 1300 background video clips in untrimmed training videos.
The prediction categories of action classifiers are "HorseRiding",
"BullChargeCape" and "Background". So the dimension of SAS features equals to 9 in MEXaction2.
For SSAD model, we use all 38 untrimmed video in MEXaction2
training set training set. Since the distribution of action instances’
length in MEXaction2 is similar with THUMOS’14, we also set the
interval of snippets to zero and the window length Tw to 512. We
train all layers of SSAD for 10 epochs with learning rate of 0.0001.
We compare SSAD with SCNN [29] and typical dense trajectory features (DTF) based method [1]. Both results are provided
by [29]. Comparison results are shown in Table 2, our SSAD network achieve significant performance gain in all action categories
Table 3: Comparisons between different action classifiers
used in SSAD on THUMOS’14, where two-stream network
includes both spatial and temporal networks.
Action Classifier used for SAS Feature
mAP (θ = 0.5)
C3D Network
Two-Stream Network
Two-Stream Network+C3D Network
20.9
21.9
24.6
Table 4: Comparisons among multiple base layers configurations on THUMOS’14. A, B, C, D, E are base layers configurations which presented in Figure 3.
Network Configuration
A
mAP(θ = 0.5)
23.7
B
C
24.6 24.1
D
E
23.9
23.4
of MEXaction2 and the mAP is increased from 7.4% to 11.0% with
overlap threshold 0.5. Figure 6 shows the visualization of prediction
results for two action categories respectively.
4.4
Model Analysis
We evaluate SSAD network with different variants in THUMOS’14
to study their effects, including action classifiers, architectures of
SSAD network and post-processing strategy.
Action classifiers. Action classifiers are used to extract SAS
feature. To study the contribution of different action classifiers, we
evaluate them individually and coherently with IoU threshold 0.5.
As shown in Table 3, two-stream networks show better performance
than C3D network and the combination of two-stream and C3D
network lead to the best performance. In action recognition task
such as UCF101, two-stream network [40] achieve 91.4%, which is
better than 85.2% of C3D [36] network (without combining with
other method such as iDT [38]). So two-stream network can predict
action categories more precisely than C3D in snippet-level, which
leads to a better performance of the SSAD network. Furthermore,
the SAS feature extracted by two-stream network and C3D network
are complementary and can achieve better result if used together.
Figure 6: Visualization of prediction action instances by SSAD network. Figure (a) shows prediction results for two action
categories in THUMOS’14 dataset. Figure (b) shows prediction results for two action categories in MEXaction2 dataset.
Table 5: Evaluation on different post-processing strategy on
THUMOS’14.
pc l ass
psas
pover
mAP (θ = 0.5)
!
22.8
!
!
!
13.4
24.3
!
!
!
!
!
!
!
19.8
23.3
24.6
Architectures of SSAD network. In section 3.3, we discuss
several architectures used for base network of SSAD. These architectures have same input and output size. So we can evaluate them
fairly without other changes of SSAD. The comparison results are
shown in Table 4. Architecture B achieves best performance among
these configurations and is adopted for SSAD network. We can
draw two conclusions from these results: (1) it is better to use max
pooling layer instead of temporal convolutional layer to shorten
the length of feature map; (2) convolutional layers with kernel size
9 have better performance than other sizes.
Post-processing strategy. We evaluate multiple post-processing
strategies. These strategies differ in the way of late fusion to generate p f inal and are shown in Table 5. For example, pc l ass is used
for generate p f inal if it is ticked in table. Evaluation results are
shown in Table 5. For the categories score, we can find that pc l ass
has better performance than p̄sas . And using the multiplication
factor pover can further improve the performance. SSAD network
achieves the best performance with the complete post-processing
strategy.
5
CONCLUSION
In this paper, we propose the Single Shot Action Detector (SSAD)
network for temporal action detection task. Our SSAD network
drops the proposal generation step and can directly predict action
instances in untrimmed video. Also, we have explored many configurations of SSAD network to make SSAD network work better for
temporal action detection. When setting Intersection-over-Union
threshold to 0.5 during evaluation, SSAD significantly outperforms
other state-of-the-art systems by increasing mAP from 19.0% to
24.6% on THUMOS’14 and from 7.4% to 11.0% on MEXaction2.
In our approach, we conduct feature extraction and action detection separately, which makes SSAD network can handle concise
high-level features and be easily trained. A promising future direction is to combine feature extraction procedure and SSAD network
together to form an end-to-end framework, so that the whole framework can be trained from raw video directly.
REFERENCES
[1] 2015. MEXaction2. http://mexculture.cnam.fr/xwiki/bin/view/Datasets/Mex+
action+dataset. (2015).
[2] M. Abadi, A. Agarwal, P. Barham, and others. 2016. Tensorflow: Largescale machine learning on heterogeneous distributed systems. arXiv preprint
arXiv:1603.04467 (2016).
[3] F. Caba Heilbron, J. Carlos Niebles, and B. Ghanem. 2016. Fast temporal activity
proposals for efficient detection of human actions in untrimmed videos. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.
1914–1923.
[4] J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Feifei. 2009. ImageNet: A largescale hierarchical image database. (2009), 248–255.
[5] V. Escorcia, F. C. Heilbron, J. C. Niebles, and B. Ghanem. 2016. Daps: Deep action
proposals for action understanding. In European Conference on Computer Vision.
Springer, 768–784.
[6] C. Feichtenhofer, A. Pinz, and A. Zisserman. 2016. Convolutional two-stream
network fusion for video action recognition. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition. 1933–1941.
[7] J. Gemert, M. Jain, E. Gati, C. G. Snoek, and others. 2015. Apt: Action localization
proposals from dense trajectories. BMVA Press.
[8] R. Girshick. 2015. Fast r-cnn. In Proceedings of the IEEE International Conference
on Computer Vision. 1440–1448.
[9] R. Girshick, J. Donahue, T. Darrell, and J. Malik. 2014. Rich feature hierarchies
for accurate object detection and semantic segmentation. In Proceedings of the
IEEE conference on computer vision and pattern recognition. 580–587.
[10] X. Glorot and Y. Bengio. 2010. Understanding the difficulty of training deep
feedforward neural networks.. In Aistats, Vol. 9. 249–256.
[11] X. Glorot, A. Bordes, and Y. Bengio. 2011. Deep Sparse Rectifier Neural Networks..
In Aistats, Vol. 15. 275.
[12] K. He, X. Zhang, S. Ren, and J. Sun. 2016. Deep residual learning for image
recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition. 770–778.
[13] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama,
and T. Darrell. 2014. Caffe: Convolutional architecture for fast feature embedding.
In Proceedings of the 22nd ACM international conference on Multimedia. ACM,
675–678.
[14] Y. G. Jiang, J. Liu, A. R. Zamir, G. Toderici, I. Laptev, M. Shah, and R. Sukthankar.
2014. THUMOS challenge: Action recognition with a large number of classes. In
ECCV Workshop.
[15] S. Karaman, L. Seidenari, and A. Del Bimbo. 2014. Fast saliency based pooling of
fisher encoded dense trajectories. In ECCV THUMOS Workshop, Vol. 1.
[16] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei.
2014. Large-scale video classification with convolutional neural networks. In
Proceedings of the IEEE conference on Computer Vision and Pattern Recognition.
1725–1732.
[17] D. Kingma and J. Ba. 2014. Adam: A method for stochastic optimization. arXiv
preprint arXiv:1412.6980 (2014).
[18] H. Kuehne, H. Jhuang, R. Stiefelhagen, and T. Serre. 2013. HMDB51: A large
video database for human motion recognition. In High Performance Computing
in Science and Engineering ’12. Springer, 571–582.
[19] C. Lea, R. Vidal, A. Reiter, and G. D. Hager. 2016. Temporal Convolutional
Networks: A Unified Approach to Action Segmentation. In Computer Vision–
ECCV 2016 Workshops. Springer, 47–54.
[20] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C. Fu, and A. C. Berg. 2016.
SSD: Single shot multibox detector. In European Conference on Computer Vision.
Springer, 21–37.
[21] S. Ma, L. Sigal, and S. Sclaroff. 2016. Learning activity progression in LSTMs for
activity detection and early detection. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition. 1942–1950.
[22] P. Mettes, J. C. van Gemert, and C. G. Snoek. 2016. Spot on: Action localization
from pointly-supervised proposals. In European Conference on Computer Vision.
Springer, 437–453.
[23] D. Oneata, J. Verbeek, and C. Schmid. 2014. The LEAR submission at Thumos
2014. ECCV THUMOS Workshop (2014).
[24] Z. Qiu, T. Yao, and T. Mei. 2016. Deep Quantization: Encoding Convolutional
Activations with Deep Generative Model. arXiv preprint arXiv:1611.09502 (2016).
[25] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. 2016. You only look once:
Unified, real-time object detection. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition. 779–788.
[26] J. Redmon and A. Farhadi. 2016. YOLO9000: Better, Faster, Stronger. arXiv
preprint arXiv:1612.08242 (2016).
[27] S. Ren, K. He, R. Girshick, and J. Sun. 2015. Faster r-cnn: Towards real-time
object detection with region proposal networks. In Advances in neural information
processing systems. 91–99.
[28] A. Richard and J. Gall. 2016. Temporal action detection using a statistical language
model. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition. 3131–3140.
[29] Z. Shou, D. Wang, and S.-F. Chang. 2016. Temporal action localization in
untrimmed videos via multi-stage cnns. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition. 1049–1058.
[30] K. Simonyan and A. Zisserman. 2014. Two-stream convolutional networks
for action recognition in videos. In Advances in Neural Information Processing
Systems. 568–576.
[31] K. Simonyan and A. Zisserman. 2015. Very Deep Convolutional Networks for
Large-Scale Image Recognition. In International Conference on Learning Representations.
[32] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. 2016. A multi-stream
bi-directional recurrent neural network for fine-grained action detection. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.
1961–1970.
[33] G. Singh and F. Cuzzolin. 2016. Untrimmed Video Classification for Activity
Detection: submission to ActivityNet Challenge. arXiv preprint arXiv:1607.01979
(2016).
[34] K. Soomro, A. R. Zamir, and M. Shah. 2012. UCF101: A dataset of 101 human
actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 (2012).
[35] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich. 2015. Going deeper with convolutions. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition. 1–9.
[36] D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri. 2015. Learning
spatiotemporal features with 3d convolutional networks. In Proceedings of the
IEEE International Conference on Computer Vision. 4489–4497.
[37] H. Wang, A. Kläser, C. Schmid, and C.-L. Liu. 2011. Action recognition by
dense trajectories. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE
Conference on. IEEE, 3169–3176.
[38] H. Wang and C. Schmid. 2013. Action recognition with improved trajectories. In
Proceedings of the IEEE International Conference on Computer Vision. 3551–3558.
[39] L. Wang, Y. Qiao, and X. Tang. 2014. Action recognition and detection by
combining motion and appearance features. THUMOS14 Action Recognition
Challenge 1 (2014), 2.
[40] L. Wang, Y. Xiong, Z. Wang, and Y. Qiao. 2015. Towards good practices for very
deep two-stream convnets. arXiv preprint arXiv:1507.02159 (2015).
[41] R. Wang and D. Tao. 2016. UTS at activitynet 2016. AcitivityNet Large Scale
Activity Recognition Challenge 2016 (2016), 8.
[42] S. Yeung, O. Russakovsky, G. Mori, and L. Fei-Fei. 2016. End-to-end learning
of action detection from frame glimpses in videos. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition. 2678–2687.
[43] G. Yu and J. Yuan. 2015. Fast action proposals for human action detection and
search. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition. 1302–1311.
[44] J. Yuan, B. Ni, X. Yang, and A. A. Kassim. 2016. Temporal Action Localization
with Pyramid of Score Distribution Features. In IEEE Conference on Computer
Vision and Pattern Recognition. 3093–3102.
[45] Y. Zhu and S. Newsam. 2016. Efficient Action Detection in Untrimmed Videos
via Multi-Task Learning. arXiv preprint arXiv:1612.07403 (2016).
| 1cs.CV
|
arXiv:1712.06454v1 [math.ST] 15 Dec 2017
Oracle inequalities for the stochastic
differential equations ∗
Pchelintsev E.A.,
†
Pergamenshchikov S.M.‡
Abstract
This paper is a survey of recent results on the adaptive robust non
parametric methods for the continuous time regression model with the
semi - martingale noises with jumps. The noises are modeled by the
Lévy processes, the Ornstein – Uhlenbeck processes and semi-Markov
processes. We represent the general model selection method and the
sharp oracle inequalities methods which provide the robust efficient
estimation in the adaptive setting. Moreover, we present the recent
results on the improved model selection methods for the nonparametric estimation problems.
Key words: Non-parametric regression, Weighted least squares estimates, Improved non-asymptotic estimation, Robust quadratic risk, Lévy
process, Ornstein – Uhlenbeck process, semi-Markov process, Model selection, Sharp oracle inequality, Adaptive estimation, Asymptotic efficiency
AMS (2010) Subject Classification : primary 62G08; secondary
62G05
∗
This work was done under the Ministry of Education and Science of the Russian
Federation in the framework of the research project no. 2.3208.2017/4.6, by RFBR Grant
16-01-00121 A and by the partial financial support of the RSF grant number 14-49-00079
(National Research University ”MPEI” 14 Krasnokazarmennaya, 111250 Moscow, Russia).
†
Department of Information Technologies and Business Analytics, Tomsk State University, Lenin str. 36, 634050 Tomsk, Russia, e-mail: [email protected]
‡
Laboratory of Mathematics LMRS, University of Rouen, France, International Laboratory SSP & QF, National Research Tomsk State University and National Research University ”MPEI” 14 Krasnokazarmennaya, 111250 Moscow, Russia, e-mail:
[email protected]
1
1
Introduction
This paper is a survey on the adaptive non parametric estimation methods
for the general semi-martingale regression model in continuous time defined
as
d yt = S(t)d t + dξt , 0 ≤ t ≤ n ,
(1.1)
where S(·) is an unknown 1 - periodic function, (ξt )0≤t≤n is an unobservable noise defined by semimartingale with the values in the Skorokhod space
D[0, n] such that, for any function f from L2 [0, n], the stochastic integral
Z n
In (f ) =
f (s)dξs
0
is well defined and has the following properties
EQ In (f ) = 0 and
EQ In2 (f )
≤ κQ
Z
n
f 2 (s)ds .
(1.2)
0
We use EQ for the expectation with respect to the distribution Q in D[0, n] of
the process (ξt )0≤t≤n , which is assumed to belong to some probability family
Qn and κQ is some positive constant depending on the distribution Q. The
problem consists to estimate the function S on the observations (yt )0≤t≤n .
Note that if (ξt )0≤t≤n is a brownian motion, then we obtain the well known
”signal+white noise” model which is very popular in statistical radio-physics
(see, for example, [17, 31, 32, 40]). In this paper we assume that in addition to
the intrinsic noise in the radio-electronic system, approximated usually by the
Gaussian white or color noise, the useful signal S is distorted by the impulse
flow described by the processes with jumps. The cause of the appearance
of a pulse stream in the radio-electronic systems can be, for example, either
external unintended (atmospheric) or intentional impulse noise and the errors
in the demodulation and the channel decoding for the binary information
symbols. Note that, for the first time the impulse noises for the detection
signal problems have been introduced on the basis of the compound Poisson
processes was introduced by Kassam in [23]. Later, such processes was used
in [28, 29, 30, 38] for the parametric and nonparametric signal estimation
problems. However, the compound Poisson process can describe only the
large impulses influence of fixed single frequency. Taking into account that
in the telecommunication systems, the impulses are without limitations on
frequencies one needs to extend the framework of the observation model by
making use the Lévy processes (2.1) which is a particular case of the general
semimartinagale regression model introduced in [25]. Generally, we consider
nonparametric estimation problems for the function S from L2 under the
2
condition that the distribution of the noise (ξt )0≤t≤n is unknown. We know
only that this distribution belongs to some distribution family Qn . In this
case we use the robust estimation approach proposed in [13, 28, 29] for the
nonparametric estimation. According to this approach we have to construct
an estimator Sbn (any function of (yt )0≤t≤n ) for S to minimize the robust risk
defined as
(1.3)
R∗n (Sbn , S) = sup RQ (Sbn , S) ,
Q∈Qn
where RQ (·, ·) is the usual quadratic risk of the form
RQ (Sbn , S) := EQ kSbn − Sk2
2
and kSk =
Z
1
S 2 (t)dt .
(1.4)
0
It is clear that if we don’t know the distribution of the observation one needs
to find an estimator which will be optimal for all possible observation distributions. Moreover in this paper we consider the estimation problem in the
adaptive setting, i.e. when the regularity of S is unknown. To this end we use
the adaptive method based on the model selection approach. The interest to
such statistical procedures is explained by the fact that they provide adaptive
solutions for the nonparametric estimation through oracle inequalities which
give the non-asymptotic upper bound for the quadratic risk including the
minimal risk over chosen family of estimators. It should be noted that for the
first time the model selection methods were proposed by Akaike [1] and Mallows [34] for parametric models. Then, these methods had been developed
for the nonparametric estimation and the oracle inequalities for the quadratic
risks was obtained by Barron, Birgé and Massart [3], Massart [35], by Fourdrinier and Pergamenshchikov [12] for the regression models in discrete time
and [27] in continuous time. Unfortunately, the oracle inequalities obtained
in these papers can not provide the efficient estimation in the adaptive setting, since the upper bounds in these inequalities have some fixed coefficients
in the main terms which are more than one. To obtain the efficiency property
for estimation procedures one has to obtain the sharp oracle inequalities, i.e.
in which the factor at the principal term on the right-hand side of the inequality is close to unity. The first result on sharp inequalities is most likely
due to Kneip [22] who studied a Gaussian regression model in the discrete
time. It will be observed that the derivation of oracle inequalities usually
rests upon the fact that the initial model, by applying the Fourier transformation, can be reduced to the Gaussian independent observations. However,
such transformation is possible only for Gaussian models with independent
homogeneous observations or for inhomogeneous ones with known correlation characteristics. For the general non Gaussian observations one needs to
3
use the approach proposed by Galtchouk and Pergamenshchikov [14, 15] for
the heteroscedastic regression models in discrete time and developed then by
Konev and Pergamenshchikov in [25, 26, 28, 29] for semimartingale models in
continuous time. In general the model selection is an adaptive rule α
b which
choses an estimator S ∗ = Sbαb from an estimate family (Sbα )α∈A . The goal of
this selection is to prove the following nonasymptotic oracle inequality: for
any sufficient small δ > 0 and any observation duration n ≥ 1
R(S ∗ , S) ≤ (1 + δ) min R(Sbα , S) + δ −1 Bn ,
(1.5)
α∈A
where the rest term Bn is sufficiently small with respect to the minimax convergence rate. Such oracle inequalities are called sharp, since the coefficient
in the main term 1 + δ is close to one for sufficiently small δ > 0. Moreover, in this paper we represent the new results on the improved estimation
methods for the nonparametric models (1.1). Usually, the model selection
procedures are based on the least squares estimators. But in [39] it is propose to use the improved least square estimators which enable to improve
considerably the non asymptotic estimation accuracy. At the first time such
idea was proposed in [12] for the regression in discrete time and in [27] for
the Gaussian regression model in continuous time. In [39] these methods
are developed for the non - Gaussion regression models in continuous time.
It should be noted that generally for the conditionally Gaussian regression
models one can not use the well known improved estimators proposed in
[19, 11] for Gaussian or spherically symmetric observations. To apply the
improved estimation methods to the non Gaussian regression models in continuous time one needs to modify the well known James–Stein procedure in
the way proposed in [38, 30]. For the improved model selection procedures
the oracle inequality (1.5) is shown also. We note that this inequality allows us to provide the asymptotic efficiency without knowing the regularity
of the function being estimated. The efficacy property for a nonparametric
estimate S ∗ means
b S) = l (r) ,
lim υn sup R(S ∗ , S) = lim υn inf sup R(S,
∗
n→∞
S∈Wrk
where
l∗ (r) = ((1 + 2k)r)
Sb S∈W k
r
n→∞
1/(2k+1)
k
π(k + 1)
2k/(2k+1)
,
(1.6)
υn is a normalizing coefficient (convergence rate), Wrk is the Sobolev ball of
a radius r > 0 and the regularity k ≥ 1. The limit (1.6) is called the Pinsker
constant which is calculated by Pinsker in [40].
4
The rest of the paper is organized as follows. In the next section 2, we
describe the Lévy, Ornstein–Uhlenbeck and semi-Markov processes as the
examples of a semimartingale impusle noise in the model (1.1). In Section
3 we construct the model selection procedure based on the least square estimators and show the sharp oracle inequalities. In Section 4 we give the improved least squares estimators and we study the improvement effect for the
semimartingale model (1.1). In Section 5 we construct the improved model
selection procedure and show the sharp oracle inequalities. The asymptotic
efficiency is studed in Section 6.
2
2.1
Examples
Lévy model
First we consider the model (1.1) with the Lévy noise process, i.e. we assume
that the noise process (ξt )0≤t≤n is defined as
ξt = ̺1 wt + ̺2 zt
and zt = x ∗ (µ − µ
e)t ,
(2.1)
where ̺1 and ̺2 are some unknown constants, (wt )t≥ 0 is a standard brownian motion, µ(ds dx) is a jump measure with deterministic compensator
µ
e(ds dx) = dsΠ(dx), Π(·) is a Lévy measure, i.e. some positive measure on
R∗ = R \ {0}, (see, for example, [18, 9] for details) such that
Π(x2 ) = 1 and Π(x6 ) < ∞ .
R
Here we use the notation Π(|x|m ) = R |z|m Π(dz). Note that the Lévy
∗
measure Π(R∗ ) could be equal to +∞. One can check directly that for the
process (2.1) the condition (1.2) holds with κQ = σQ = ̺21 + ̺22 . We assume
that the nuisance parameters ̺1 and ̺2 of the process (ξt )0≤t≤n satisfy the
conditions
(2.2)
0 < ̺ ≤ ̺21 and σQ ≤ ς ∗ ,
where the bounds ̺ and ς ∗ are functions of n, i.e. ̺ = ̺ and ς ∗ = ςn∗ such
n
that for any δ̌ > 0
lim inf nδ̌ ̺ > 0 and
n→∞
n
lim n−δ̌ ςn∗ = 0 .
n→∞
(2.3)
For this example Qn is the family of all distributions of process (1.1) –
(2.1) on the Skorokhod space D[0, n] satisfying the conditions (2.2) – (2.3).
The models (1.1) with the Lévy’s type noise are used in different applied
problems (see [7], for details). Such models naturally arise in the nonparametric functional statistics problems (see, for example, [8]).
5
2.2
Ornstein – Uhlenbeck model
Now we consider the noise process (ξt )t≥0 in (1.1) difened by a non-Gaussian
Ornstein-Uhlenbeck process with the Lévy subordinator. Let the noise process in (1.1) obey the equation
dξt = aξt dt + dut ,
ξ0 = 0 ,
(2.4)
where ut = ̺1 wt + ̺2 zt and the process zt is defied in (2.1). Here a ≤ 0, ̺1
and ̺2 are unknown parameters. We assume that the parameters ̺1 and ̺2
satisfy the conditions (2.2) and the parameter
− amax ≤ a ≤ 0 ,
(2.5)
where the bound amax > 0 is the function of n, i.e. amax = amax (n), such
that for any δ̌ > 0
a
= 0.
(2.6)
lim max
n→∞ nδ̌
In this case Qn is the family of all distributions of process (2.4) on the
Skorokhod space D[0, n] satisfying the conditions (2.2), (2.5) and (2.6). Note
also that the processes (2.1) and (2.4) are G - conditionally Gaussian square
integrated semimartingales, where G = σ{zt , t ≥ 0}.
Such processes are used in the financial Black-Scholes type markets with
jumps (see, for example, [2, 10] and the references therein). Note also that
in the case when ̺2 = 0 for the parametric estimation problem such models
are considered in [20, 21, 24].
2.3
Semi – Markov model
In [4, 5] it is introduced the regression model (1.1) in which the noise process
describes by the equation
ξt = ̺1 Lt + ̺2 Xt ,
(2.7)
and the Lévy process Lt is defined as
Lt = ̺ˇ wt +
p
1 − ̺ˇ2 zt ,
where 0 ≤ ̺ˇ ≤ 1 is an unknown constant. Moreover, we assume that the pure
jump process (Xt )t≥ 0 in (2.7) is a semi-Markov process with the following
form
Nt
X
Xt =
Yi ,
(2.8)
i=1
6
where (Yi )i≥ 1 is an i.i.d. sequence of random variables with
E Yi = 0 ,
E Yi2 = 1 and E Yi4 < ∞ .
Here Nt is a general counting process (see, for example, [36]) defined as
Nt =
∞
X
k=1
1{Tk ≤t}
and Tk =
k
X
τl ,
l=1
where (τl )l≥ 1 is an i.i.d. sequence of positive integrated random variables
with distribution η and mean τ̌ = E τ1 > 0. We assume that the processes
(Nt )t≥0 and (Yi )i≥ 1 are independent between them and are also independent
of (Lt )t≥0 . Here, the family Qn is defined by of all distributions of process
(2.7) on the Skorokhod space D[0, n] with the parameters ̺1 and ̺2 satisfying
the conditions (2.2) and 0 ≤ ̺ˇ ≤ 1.
Note that the process (Xt )t≥ 0 is a special case of a semi-Markov process
(see, e.g., [6] and [33]). It should be noted that if τj are exponential random
variables, then (Nt )t≥0 is a Poisson process and, in this case, (ξt )t≥0 is a
Lévy process. But, in the general case when the process (2.8) is not a Lévy
process, this process has a memory and cannot be treated in the framework
of semi-martingales with independent increments. In this case, we need to
develop new tools based on renewal theory arguments from [16].
It should be noted that for ̺ˇ > 0 the process (2.7) is G - conditionally
Gaussian also. In this case G = σ{zt , Xt , t ≥ 0}.
3
Model selection
Let (φj )j≥ 1 be an orthonormal uniformly bounded basis in L2 [0, 1], i.e. for
some constant φ∗ ≥ 1, which may be depend on n,
sup
sup |φj (t)| ≤ φ∗ < ∞ .
0≤j≤n 0≤t≤1
For example, we can take the trigonometric basis defined as Tr1 ≡ 1 and, for
j ≥ 2,
√ cos(2π[j/2]x) for even j;
Trj (x) = 2
(3.1)
sin(2π[j/2]x) for odd j,
where [x] denotes the integer part of x.
7
To estimate the function S we use here the model selection procedure for
continuous time regression models from [28] based on the Fourrier expansion.
We recall that for any function S from L2 [0, 1] we can write
Z 1
∞
X
S(t) =
θj φj (t) and θj = (S, φj ) =
S(t)φj (t)dt .
0
j=1
So, to estimate the function S it suffices to estimate the coefficients θj and to
replace them in this representation by their estimators. Using the fact that
the function S and φj are 1 - periodic we can write that
Z
1 n
θj =
φ (t) S(t)dt .
n 0 j
If we replace here the differential S(t)dt by the stochastic observed differential
dyt then we obtain the natural estimate for θj on the time interval [0, n]
Z n
1
θbj,n =
φ (t)d yt ,
n 0 j
which can be represented, in view of the model (1.1), as
1
θbj,n = θj + √ ξj,n
n
1
and ξj,n = √ In (φj ) .
n
We need to impose some stability conditions for the noise Fourier transform
sequence (ξj,n )1≤j≤n . To this end we set for some stability noise intensity
parameter σQ > 0 the following function
L1,n (Q) =
sup
x∈[−1,1]n
n
X
j=1
2
xj EQ ξj,n − σQ .
In [28] the parameter σQ is called proxy variance .
C1 ) There exists a variance proxy σQ > 0 such that for any ǫ > 0
L1,n (Q)
= 0.
nǫ
n→∞
lim
Moreover, we set
2
n
X
2
2
L2,n (Q) = sup EQ
xj (ξj,n
− EQ ξj,n
)
.
|x|≤1
j=1
8
(3.2)
C2 ) Assume that for any ǫ > 0
L2,n (Q)
= 0.
nǫ
n→∞
lim
Now (see, for example, [17]) we can estimate the function S by the projection estimators, i.e.
Sbm (t) =
m
X
j=1
θbj,n φj (t) ,
0 ≤ t ≤ 1,
(3.3)
for some number m → ∞ as n → ∞. It should be noted that Pinsker in
[40] shows that the projection estimators of the form (3.3) are not efficient.
For obtaining efficient estimation one needs to use weighted least square
estimators defined as
n
X
b
Sλ (t) =
λ(j)θbj,n φj (t) ,
(3.4)
j=1
where the coefficients λ = (λ(j))1≤j≤n belong to some finite set Λ from [0, 1]n .
As it is shown in [40], in order to obtain efficient estimators, the coefficients
λ(j) in (3.4) need to be chosen depending on the regularity of the unknown
function S. Since we consider the adaptive case, i.e. we assume that the
regularity of the function S is unknown, then we chose the weight coefficients
on the basis of the model selection procedure proposed in [28] for the general
semi-martingale regression model in continuous time. To the end, first we
set
n
X
ν = #(Λ) and |Λ|∗ = 1 + max
λ(j) ,
λ∈Λ
j=1
where #(Λ) is the cardinal number of Λ. Now, to choose a weight sequence
λ in the set Λ we use the empirical quadratic risk, defined as
Errn (λ) =k Sbλ − S k2 ,
which in our case is equal to
Errn (λ) =
n
X
j=1
2
λ (j)θbj,n
−2
2
n
X
j=1
λ(j)θbj,n θj +
∞
X
θj2 .
j=1
Since the Fourier coefficients (θj )j≥ 1 are unknown, we replace the terms
θbj,n θj,n by
σ
b
2
θej,n = θbj,n
− n,
n
9
where σ
bn is an estimate for the variance proxy σQ defined in (3.2). If it is
known, we take σ
bn = σQ , otherwise, we can choose it, for example, as in [28],
i.e.
n
X
b
σ
bn =
t2j,n ,
(3.5)
√
j=[ n]+1
where b
tj,n are the estimators for the Fourier coefficients with respect to the
trigonometric basis (3.1), i.e.
Z
1 n
b
T rj (t)dyt .
tj,n =
n 0
Finally, in order to choose the weights, we will minimize the following cost
function
n
n
X
X
2
2
b
Jn (λ) =
λ (j)θj,n − 2
λ(j)θej,n + δ Pn (λ),
j=1
j=1
where δ > 0 is some threshold which will be specified later and the penalty
term is
σ
b |λ|2
Pn (λ) = n
.
(3.6)
n
We define the model selection procedure as
Sb∗ = Sbλ̂
b = argmin Jn (λ) .
with λ
λ∈Λ
(3.7)
We recall that the set Λ is finite so λ̂ exists. In the case when λ̂ is not unique,
we take one of them.
As is shown in [4, 28, 39] both Conditions C1 ) and C2 ) hold for the
processes (2.1), (2.4) and (2.7).
Proposition 3.1. If the conditions C1 ) and C2 ) hold for the distribution Q
of the process ξ in (1.1), then, for any n ≥ 1 and 0 < δ < 1/3, the risk (1.4)
of estimate (3.7) for S satisfies the oracle inequality
B (Q)
1 + 3δ
min RQ (Sbλ , S) + n
,
(3.8)
RQ (Sb∗ , S) ≤
1 − 3δ λ∈Λ
δn
where Bn (Q) = Un (Q) 1 + |Λ|∗ EQ |b
σn − σQ | and the coefficient Un (Q) is
such that for any ǫ > 0
U (Q)
lim n ǫ = 0 .
(3.9)
n
n→∞
10
In the case, when the value of σQ is known, one can take σ
bn = σQ and
σQ |λ|2n
Pn (λ) =
,
n
then we can rewrite the oracle inequality (3.8) with Bn (Q) = Un (Q). Also
we study the accuracy properties for the estimator (3.5).
Proposition 3.2. Let in the model (1.1) the function S(·) is continuously
differentiable. Then, for any n ≥ 2,
EQ |b
σn − σQ | ≤
κn (Q)(1 + kṠk2 )
√
,
n
where the term κn (Q) possesses the property (3.9).
To obtain the oracle inequality for the robust risk (1.3) we need some additional condition on the distribution family Qn . We set
ς ∗ = ςn∗ = sup σQ
Q∈Qn
and L∗n = sup (L1,n (Q) + L2,n (Q)) .
(3.10)
Q∈Qn
C∗1 ) Assume that the conditions C1 )–C2 ) hold and for any ǫ > 0
L∗n + ςn∗
= 0.
nǫ
n→∞
lim
Now we impose the conditions on the set of the weight coefficients Λ.
C∗2 ) Assume that the set Λ is such that for any ǫ > 0
ν
= 0 and
n→∞ nǫ
lim
lim
n→∞
|Λ|∗
= 0.
n1/2+ǫ
Theorem 3.3. Assume that the conditions C∗1 )–C∗2 ) hold. Then the robust
risk (1.3) of the estimate (3.7) for continuously differentiable function S(t)
satisfies for any n ≥ 2 and 0 < δ < 1/3 the oracle inequality
1 ∗
1 + 3δ
min R∗n (Sbλ , S) +
B (1 + kṠk2 ) ,
R∗n (Sb∗ , S) ≤
1 − 3δ λ∈Λ
δn n
where the term B∗n satisfies the property (3.9).
11
Now we specify the weight coefficients (λ(j))j≥1 in the way proposed in
[14] for a heteroscedastic regression model in discrete time. First we define
the normalizing coefficient which defined the minimax convergence rate
vn =
n
,
ς∗
(3.11)
where the upper proxy variance is ς ∗ is defined in (3.10). Consider a numerical
grid of the form
An = {1, . . . , k ∗ } × {r1 , . . . , rm } ,
where ri = iε and m = [1/ε2 ]. Both parameters k ∗ ≥ 1 and 0 < ε ≤ 1 are
assumed to be functions of n, i.e. k ∗ = k ∗ (n) and ε = ε(n), such that for any
δ>0
k ∗ (n)
limn→∞ k ∗ (n) = +∞ , limn→∞
= 0,
ln n
limn→∞ ε(n) = 0 and limn→∞ nδ ε(n) = +∞ .
One can take, for example,
ε(n) =
1
ln(n + 1)
and k ∗ (n) =
p
ln(n + 1) .
For each α = (β, r) ∈ An we introduce the weight sequence λα = (λα (j))j≥1
as
λα (j) = 1{1≤j≤d} + 1 − (j/ωα )β 1{d<j≤ωα }
(3.12)
1/(2β+1)
where d = d(α) = [ωα / ln(n + 1)], ωα = τβ r vn
and
τβ =
(β + 1)(2β + 1)
.
π 2β β
We set
Λ = {λα , α ∈ An } .
(3.13)
It will be noted that in this case the cardinal of the set Λ is ν = k ∗ m.
Moreover, taking into account that τβ < 1 for β ≥ 1 we obtain for the set
(3.13)
|Λ|∗ ≤ 1 + sup ωα ≤ 1 + (υn /ε)1/3 .
α∈A
Note that the form (3.12) for the weight coefficients was proposed by
Pinsker in [40] for the efficient estimation in the nonadaptive case, i.e. when
the regularity parameters of the function S are known. In the adaptive case
these weight coefficients are used in [28, 29] to show the asymptotic efficiency
for model selection procedures.
12
4
Improved estimation
In this Section we consider the improved estimation method for the model
(1.1). We impose the following additional condition on the noise distribution.
D1 ) There exists n0 ≥ 1 such that for any n ≥ n0 there exists a σ field Gn for which the random vector ξed,n = (ξj,n )1≤j≤d is the Gn conditionally
Gaussian in Rd with the covariance matrix
Gn = E ξi,n ξj,n |Gn )
1≤i,j≤d
and for some nonrandom constant ln∗ > 0
inf
Q∈Qn
(tr Gn − λmax (Gn )) ≥ ln∗ ,
where λmax (A) is the maximal eigenvalue of the matrix A.
Proposition 4.1. Let in the model (1.1) the noise process describes by the
Lévy process (2.1). Then the condition D1 ) holds with ln∗ = (d − 1)̺ for any
n ≥ 1.
Proposition 4.2. Let in the model (1.1) the noise process describes by the
Ornstein–Uhlenbeck process (2.4). Then the condition D1 ) holds with ln∗ =
(d − 6)̺/2 for any n ≥ n0 and d ≥ d0 = inf{d ≥ 7 : 5 + ln d ≤ ǎd},
ǎ = (1 − e−amax )/(4amax ).
Now, for the first d Fourier coefficients we use the improved estimation
method proposed for parametric models in [38]. To this end we set θen =
P
(θbj,n )1≤j≤d . In the sequel we will use the norm |x|2d = dj=1 x2j for any vector
x = (xj )1≤j≤d from Rd . Now we define the shrinkage estimators as
∗
θj,n
= (1 − g(j)) θbj,n ,
where g(j) = (cn /|θen |d )1{1≤j≤d} and
cn =
ln∗
.
p
rn∗ + d/vn n
(4.1)
The positive parameter rn∗ is such that limn→∞ rn∗ = ∞ and for any ǫ > 0
lim
n→∞
rn∗
= 0
nǫ
13
(4.2)
and vn defined in (3.11). Now we introduce a class of shrinkage weighted
least squares estimates for S as
Sλ∗
=
n
X
∗
λ(j)θj,n
φj .
(4.3)
j=1
We denote the difference of quadratic risks of the estimates (3.4) and
(4.3) as
∆Q (S) := RQ (Sλ∗ , S) − RQ (Sbλ , S) .
We obtain the following result.
Theorem 4.3. Let the observed process (yt )0≤t≤n describes by the equation
(1.1) and the condition D1 ) holds. Then for any n ≥ 1
sup
sup ∆Q (S) ≤ −c2n .
(4.4)
Q∈Qn kSk≤r ∗
n
Remark 4.1. The inequality (4.4) means that non asymptotically, i.e. for
any n ≥ 1, the estimate (4.3) outperforms in mean square accuracy the estimate (3.4). Moreover in the efficient weight coefficients d ≈ nδ̌ as n → ∞
for some δ̌ > 0. Therefore, in view of the definition (4.1) and the conditions
(2.3) and (4.2) ncn → ∞ as n → ∞. This means that improvement is considerably may better than for the parametric regression when the parameter
dimension d is fixed [38].
5
Improved model selection
This Section gives the construction of a model selection procedure for estimating a function S in (1.1) on the basis of improved weighted least square
estimates (Sλ∗ )λ∈Λ and states the sharp oracle inequality for the robust risk
of proposed procedure.
As in Section 3, the performance of any estimate Sλ∗ will be measured by
the empirical squared error
Errn (λ) = kSλ∗ − Sk2 .
In order to obtain a good estimate, we have to write a rule to choose a weight
vector λ ∈ Λ in (4.3). It is obvious, that the best way is to minimise the
14
empirical squared error with respect to λ. Making use the estimate definition
(4.3) and the Fourier transformation of S implies
Errn (λ) =
n
X
λ
2
∗ 2
(j)(θj,n
)
j=1
Here one needs to replace the terms
−2
n
X
∗
λ(j)θj,n
θj
j=1
∗
θj,n
θj
+
n
X
θj2 .
j=1
by their estimators θ̄j,n . We set
σ
bn
,
n
where σ
bn is defined in (3.5). For this change in the empirical squared error,
one has to pay some penalty. Thus, one comes to the cost function of the
form
n
n
X
X
∗ 2
Jn∗ (λ) =
λ2 (j)(θj,n
) −2
λ(j) θ̄j,n + δ Pn (λ) ,
∗ b
θ̄j,n = θj,n
θj,n −
j=1
j=1
where δ is some positive constant and the penalty term Pn (λ) is defined in
(3.6). Substituting the weight coefficients, minimizing the cost function
λ∗ = argminλ∈Λ Jn∗ (λ) ,
(5.1)
in (4.3) leads to the improved model selection procedure
S ∗ = Sλ∗∗ .
(5.2)
It will be noted that λ∗ exists because Λ is a finite set and also if the minimizing sequence in (5.1) λ∗ is not unique, one can take any minimizer.
Theorem 5.1. If the conditions C1 ) and C2 ) hold for the distribution Q of
the process ξ in (1.1), then, for any n ≥ 1 and 0 < δ < 1/3, the risk (1.4) of
estimate (5.2) for S satisfies the oracle inequality
B̌ (Q)
1 + 3δ
min RQ (Sλ∗ , S) + n
,
1 − 3δ λ∈Λ
nδ
where B̌n (Q) = Ǔn (Q) 1 + |Λ|∗ EQ |b
σn − σQ | and the coefficient Ǔn (Q) satisfies the property (3.9).
RQ (Sλ∗∗ , S) ≤
Now Theorem 5.1 and Proposition 3.2 directly imply the following inequality for the robust risk (1.3) of the procedure (5.2).
Theorem 5.2. Assume that the conditions C∗1 ) and C∗2 ) hold and the function S is continuously differentiable. Then for any n ≥ 2 and 0 < δ < 1/3
R∗n (Sλ∗∗ , S) ≤
Ǔ∗ (1 + kṠk2 )
1 + 3δ
min R∗n (Sλ∗ , S) + n
,
1 − 3δ λ∈Λ
nδ
where the coefficient Ǔ∗n satisfies the property (3.9).
15
6
Asymptotic efficiency
In order to study the asymptotic efficiency we define the following functional
Sobolev ball
k
X
k
Wk,r = {f ∈ Cp [0, 1] :
kf (j) k2 ≤ r} ,
j=0
where r > 0 and k ≥ 1 are some unknown parameters, Ckp [0, 1] is the space of
k times differentiable 1 - periodic R → R functions such that f (i) (0) = f (i) (1)
for any 0 ≤ i ≤ k −1. It is well known that for any S ∈ Wk,r the optimal rate
of convergence is n−2k/(2k+1) (see, for example, [40, 37]). On the basis of the
model selection procedure we construct the adaptive procedure S∗ for which
we obtain the following asymptotic upper bound for the quadratic risk, i.e.
we show that the parameter (1.6) gives a lower bound for the asymptotic
normalized risks. To this end we denote by Σn of all estimators Sbn of S
measurable with respect to the process (1.1), i.e. σ{yt , 0 ≤ t ≤ n}.
Theorem 6.1. The robust risk (1.3) admits the following asymptotic lower
bound
lim inf inf vn2k/(2k+1) sup R∗n (Sbn , S) ≥ l∗ (r) .
n→∞
b ∈Σ
S
n
n
S∈Wk,r
We show that this lower bound is sharp in the following sense.
Theorem 6.2. The quadratic risk (1.3) for the estimating procedure S ∗ has
the following asymptotic upper bound
lim sup vn2k/(2k+1) sup R∗n (S ∗ , S) ≤ l∗ (r) .
n→∞
S∈Wk,r
It is clear that Theorem 6.2 and Theorem 6.1 imply
Corollary 6.3. The model selection procedure S ∗ is efficient, i.e.
2k
lim (vn ) 2k+1 sup R∗n (S ∗ , S) = l∗ (r) .
n→∞
(6.1)
S∈Wk,r
Note that the equality (6.1) implies that the parameter (1.6) is the Pinsker
constant in this case (cf. [40]). Moreover, it means that the robust efficiency
2k
holds with the convergence rate (vn ) 2k+1 . It is well known that for the simple
risks the optimal (minimax) estimation convergence rate for the functions
from the set Wk,r is n2k/(2k+1) (see, for example, [17, 37, 40]). So, if the
distribution upper bound ς ∗ → 0 as n → ∞ we obtain the more rapid rate,
16
and if ς ∗ → ∞ as n → ∞ we obtain the more slow rate. In the case when ς ∗
is constant the robust rate is the same as the classical non robust convergence
rate.
Acknowledgements.
The first author is partially supported by the the RSF grant 17-11-01049.
The last author is partially supported by the Russian Federal Professor program (project no. 1.472.2016/1.4, Ministry of Education and Science) and
by the project XterM - Feder, University of Rouen, France.
References
[1] Akaike H (1974) A new look at the statistical model identification.
IEEE Trans. on Automatic Control 19:716–723
[2] Barndorff-Nielsen OE, Shephard N (2001) Non-Gaussian OrnsteinUhlenbeck-based models and some of their uses in financial mathematics. J Royal Stat Soc B63:167–241
[3] Barron A, Birgé L, Massart P (1999) Risk bounds for model selection
via penalization. Probab. Theory Relat. Fields 113:301–415
[4] Barbu V, Beltaif S, Pergamenshchikov SM (2017) Robust adaptive efficient estimation for semi - Markov nonparametric regression models.
Preprint 2017 https://arxiv.org/pdf/1604.04516.pdf
submitted in Statistical inference for stochastic processes
[5] Barbu V, Beltaif S, Pergamenshchikov SM (2017) Robust adaptive efficient estimation for a semi-Markov continuous time regression from
discrete data. Preprint 2017 http://arxiv.org/abs/1710.10653
submitted in Annales de l’Institut Henri Poincaré
[6] Barbu VS, Limnios N (2008) Semi-Markov Chains and Hidden SemiMarkov Models toward Applications - Their use in Reliability and DNA
Analysis. Lecture Notes in Statistics 191
[7] Bertoin J (1996) Lévy Processes. Cambridge University Press, Cambridge
[8] Ferraty F, Vieu P (2006) Nonparametric Functional Data Analysis :
Theory and Practice. Springer Series in Statistics, Springer-Verlag,
New York
17
[9] Cont R, Tankov P (2004) Financial Modelling with Jump Processes.
Chapman & Hall
[10] Delong L, Klüppelberg C (2008) Optimal investment and consumption in a Black-Scholes market with Lévy driven stochastic coefficients.
Annals of Applied Probability 18(3):879–908
[11] Fourdrinier D, Strawderman W E (1996) A paradox concerning shrinkage estimators: should a known scale parameter be replaced by an estimated value in the shrinkage factor? Journal of Multivariate Analysis
59(2):109 –140
[12] Fourdrinier D, Pergamenshchikov S (2007) Improved selection model
method for the regression with dependent noise. Annals of the Institute
of Statistical Mathematics 59(3):435–464
[13] Galtchouk LI, Pergamenshchikov SM (2006) Asymptotically efficient
estimates for non parametric regression models.Statistics and Probability Letters 76(8):852–860
[14] Galtchouk LI, Pergamenshchikov SM (2009) Sharp non-asymptotic oracle inequalities for nonparametric heteroscedastic regression models.
Journal of Nonparametric Statistics 21(1):1 – 16
[15] Galtchouk LI, Pergamenshchikov SM (2009) Adaptive asymptotically
efficient estimation in heteroscedastic nonparametric regression. Journal of Korean Statistical Society 38(4):305–322
[16] Goldie CM (1991) Implicit renewal theory and tails of solutions of
random equations. The Annals of Applied Probability 1(1):126 – 166
[17] Ibragimov IA, Khasminskii RZ (1981) Statistical Estimation: Asymptotic Theory. Springer, New York
[18] Jacod J, Shiryaev AN (2002) Limit theorems for stochastic processes,
2nd edition. Springer, Berlin
[19] James W, Stein C (1961) Estimation with quadratic loss. In Proceedings of the Fourth Berkeley Symposium Mathematics, Statistics and
Probability, University of California Press, Berkeley 1:361–380
[20] Höpfner R, Kutoyants YuA (2009) On LAN for parametrized continuous periodic signals in a time inhomogeneous diffusion. Statist. Decisions 27(4):309 – 326
18
[21] Höpfner R, Kutoyants YuA (2010) Estimating discontinuous periodic
signals in a time inhomogeneous diffusion. Statistical Inference for
Stochastic Processes 13(3):193 – 230
[22] Kneip A (1994) Ordered linear smoothers. Annals of Statistcs 22:835–
866
[23] Kassam SA (1988) Signal detection in non-Gaussian noise. SpringerVerlag Inc., New York
[24] Konev VV, Pergamenshchikov SM (2003) Sequential estimation of the
parameters in a trigonometric regression model with the Gaussian
coloured noise. Statistical Inference for Stochastic Processes 6:215–235
[25] Konev VV, Pergamenshchikov SM (2009) Nonparametric estimation in
a semimartingale regression model. Part 1. Oracle Inequalities. Journal
of Mathematics and Mechanics of Tomsk State University 3:23–41
[26] Konev VV, Pergamenshchikov SM (2009) Nonparametric estimation
in a semimartingale regression model. Part 2. Robust asymptotic efficiency. Journal of Mathematics and Mechanics of Tomsk State University 4:31–45
[27] Konev VV, Pergamenshchikov SM (2010) General model selection estimation of a periodic regression with a Gaussian noise. Annals of the
Institute of Statistical Mathematics 62:1083–1111
[28] Konev VV, Pergamenshchikov SM (2012) Efficient robust nonparametric estimation in a semimartingale regression model. Ann Inst Henri
Poincaré Probab Stat 48(4):1217–1244
[29] Konev VV, Pergamenshchikov SM (2015) Robust model selection for a
semimartingale continuous time regression from discrete data. Stochastic processes and their applications 125:294 – 326
[30] Konev VV, Pergamenshchikov SM, Pchelintsev E (2014) Estimation
of a regression with the pulse type noise from discrete data. Theory
Probab Appl 58(3):442–457
[31] Kutoyants YuA (1977) Estimation of the signal parameter in a Gaussian Noise. Problems of Information Transmission 13(4):29–36
[32] Kutoyants YuA (1984) Parameter Estimation for Stochastic Processes.
Heldeman-Verlag, Berlin
19
[33] Limnios N, Oprisan G (2001) Semi-Markov Processes and Reliability.
Birkhäuser, Boston
[34] Mallows C (1973) Some comments on Cp . Technometrics 15:661–675
[35] Massart P (2005) A non-asymptotic theory for model selection. European Congress of Mathematics, Eur. Math. Soc., Zürich
[36] Mikosch T (2004) Non-Life Insurance Mathematics. An Introduction
with Stochastic Processes. Springer
[37] Nussbaum M (1985) Spline smoothing in regression models and asymptotic efficiency in L2 . Ann Stat 13:984–997
[38] Pchelintsev E (2013) Improved estimation in a non-Gaussian parametric regression. Stat Inference Stoch Process 16(1):15 – 28
[39] Pchelintsev EA, Pchelintsev VA, Pergamenshchikov SM (2017)
Improved robust model selection methods for the Levy nonparametric regression in continuous time. Preprint 2017
http://arxiv.org/abs/1710.03111
submitted in Stochastic Processes and their Applications
[40] Pinsker MS (1981) Optimal filtration of square integrable signals in
Gaussian white noise. Problems of Transimission information 17:120–
133
20
| 10math.ST
|
Estimated Depth Map Helps Image Classification
Yihui He∗
Xi’an Jiaotong University
Xi’an, China
arXiv:1709.07077v1 [cs.CV] 20 Sep 2017
[email protected]
Abstract
We consider image classification with estimated depth.
This problem falls into the domain of transfer learning,
since we are using a model trained on a set of depth images to generate depth maps (additional features) for use
in another classification problem using another disjoint set
of images. It’s challenging as no direct depth information
is provided. Though depth estimation has been well studied [16], none have attempted to aid image classification
with estimated depth. Therefore, we present a way of transferring domain knowledge on depth estimation to a separate image classification task over a disjoint set of train,
and test data. We build a RGBD dataset based on RGB
dataset and do image classification on it. Then evaluation
the performance of neural networks on the RGBD dataset
compared to the RGB dataset. From our experiments, the
benefit is significant with shallow and deep networks. It improves ResNet-20 by 0.55% and ResNet-56 by 0.53%.Our
code and dataset are available publicly.1
Figure 1. Learning on RGBD dataset: First we generate the depth
map of an image with a depth estimation network (On the Left).
Second, we perform image classification on the depth map and the
image.
ing data. Different from previous efforts, we propose to
utilize estimated depth maps in a image classification task.
While extensively studied in semantic labeling and accuracy improvement, depth map regression has been less explored in its application to classification problems. Intuitively, one can imagine that a neural-network that is deep
enough would generate it’s own depth-map (or at least simulate depth-map-like features).
Recently, the efficacy and power of the deep convolutional neural network (CNN) has been made accessible
[10, 12]. With a CNN, we are able to perform depth estimation on a single image [16]. However, most classification
tasks still perform on RGB images. With only RGB images,
CNN features have been setting new records for a wide variety of vision applications [10, 17, 7, 18, 3]. Despite all
the successes in depth estimation and image classification,
the deep CNN has been not yet been used for learning on
RGBD images, since RGBD datasets are not as widely-used
as RGB datasets.
We propose to build a RGBD dataset based on RGB
dataset and do image classification on it, illustrated in Figure 1. Then evaluation the performance of neural networks
on the RGBD dataset compared to the RGB dataset. To our
knowledge, we are the first to bridge gap between estimated
depth and image classification. From our experiments, the
1. Introduction
Estimating depths from a single monocular image depicting general scenes is a fundamental problem in computer vision, which has widespread applications in sceneunderstanding, 3D modeling, robotics, and other challenging problems. It is a notorious example of an ill-posed
problem, as one captured image may correspond to numerous real world scenes [6]. It remains a challenging task
for computer vision algorithms as no reliable cues can be
exploited, such as temporal information, stereo correspondences. Previous research involving depth maps usually involve geometric [11, 9, 8], convolutional [16] and semantic
[14] techniques. Nevertheless, none of these works tried
to perform image classification using depth maps as train∗ This work was done as course project of Advanced Data Mining in
UCSB
1 github.com/yihui-he/Estimated-Depth-Map-Helps-ImageClassification
1
benefit is significant on both shallow and deep networks. It
improves ResNet-20 0.55%.
To sum up, we highlight the main contributions of this
work as follows:
• We created a RGBD image dataset for CIFAR-10.
• We illustrate that depth channel has a better feature
representation than R,G,B channels, and show that
training on RGBD images could improve performance.
Figure 2. Transer Learning: Build RGBD CIFAR-10 dataset
• We define a new metric for ill-posed depth prediction
problem.
2. Related Work
CNN have been applied with great success for object
classification [13, 19, 20, 10, 2] and detection [7, 18, 3].
CNN have recently been applied to a variety of other tasks,
like depth estimation. Depth estimation from single image
is well addressed by Liu et al. [16] and Eigen et al. [5].
They both agree that depth estimation is an ill-posed problem, since there’s no real ground truth depth map. We define transfer learning accuracy metric for depth estimation
model (Section 3.1). It becomes easier to compare performance of different depth estimation model.
Estimated depth map [16] has been successfully applied
to some other problems. Based on depth information, performance improved on semantic labeling [5]. However,
depth maps have not been combined with an image classification task. To our knowledge, we are the first to bridge
gap between depth estimation and image classification.
There are already many successful transfer learning results in Computer Vision. A popular one is transfer ImageNet [4] Classification Network like VGG-16 [19] to object detection [18]. Another example also in object detection is contextualized networks [15], usually through multi
scale context.
3. Approach
Recent depth image research works mainly focus on
depth-estimation [16] and segmentation with depth image
[5]. And we’ve witnessed significant improvement on depth
estimation quality in these years. However, most image
classification tasks nowadays are still performed on RGB
images. So we want to transfer depth knowledge learned by
depth estimation model into our image-classification model.
In this section, we first built a RGBD dataset for CIFAR10 [13], based on a trained deep convolutional neural field
model [16]. To investigate the effect of the depth channel on
image classification task, we design two experiments (one
with a simple feedforward NN and one with a CNN) Finally,
we propose a new metric for depth estimation performance
measurement.
Figure 3. Depth map estimated by deep convolutional neural field
3.1. Build RGBD Dataset
Since the Deep Convolutional Neural Field model accepts images [4] that are much larger that CIFAR-10 tiny
images (32 × 32), we build RGBD dataset as follow:
1. resize CIFAR-10 tiny image (32 × 32 × 3) to normal
size (400 × 400 × 3) in order to feed in CNF.
2. perform depth estimation on the normal size image.
3. downscale the output image (depth image, 400×400×
1) back to tiny image (32 × 32 × 1).
4. combine RGB and D channels together as our RGBD
image (32 × 32 × 4).
Figure 2 shows the transfer learning procedure. Figure 3
shows some depth maps. Since there is no ground-truth
depth image for CIFAR-10 dataset, we can’t directly measure the quality of our depth estimation attempts for these
tiny images. However, we can infer this indirectly.We can
use the accuracy results of our two experiments as a new
metric to quantify depth map quality.
3.2. Classification Task on RGBD Dataset
In order to make it easier to show effect of depth channel, we employ a simple two layer neural network for classification task. The architecture for learning on the RGBD
dataset is shown in Figure 1.
0.50
0.6
0.5
0.40
val accuracy
validation accuracy
0.45
0.35
0.30
0.3
0.25
0.20
0.4
0.2
0
20
40
60
80
100
Epoch
120
140
160
180
Figure 4. R vs G vs B vs D, testing time
The number of neurons in the input layer depends on input. If input is a single channel (R, G, B, D), we have 32×32
neurons. The amount of hidden neurons is not determined.
We perform fine tuning for each situation. The number of
output neurons is always number of classes (10 classes for
CIFAR-10).
4. Experiment
We measure depth map quality in two ways. First, we
evaluate neural network performance on R, G, B, D channel as input respectively. Second, we train neural network
on RGB, RGBD respectively and compare the performance.
Our depth estimation is based on Tensorflow [1]. And our
neural network training is based on Caffe [12].
4.1. R vs G vs B vs D
We perform fine tuning on each channel. So that their
performances are approximately optimal. Figure 4 shows
validation accuracy comparison through time.
You can see that, at testing time, the depth channel outperforms R, G, B channels under the same architecture.
This implies that, depth channel has a better feature representation than R, G, B channels. Training on RGBD dataset
would result in better performance for shallow networks.
4.2. RGB vs RGBD
Figure 5 Compares validation accuracy comparison
through time.
We get 56% and 52% validation accuracy with RGBD
and RGB dataset respectively. This can be seen as a sign
that depth map brings extra knowledge learned by deep convolutional neural field to our classification task.
Also notice that, although RGBD dataset have more inputs and neurons, it has a much higher converge rate than
RGB dataset. It can be interpreted as a better feature representation brought by depth map. Estimated depth map
works on shallow networks, however it remains a question
whether it works on deep networks like ResNet [10].
0
50
100
Epoch
150
200
250
Figure 5. RGB vs RGBD, testing time (RGBD:blue, RGB:red)
Network
two-layers
ResNet-20
ResNet-56
RGB
48
8.75
6.97
RGBD
44
8.20
6.44
Table 1. Performance comparisons of ResNet with or without
depth estimation aided on CIFAR-10 (Smaller is better).
4.3. ResNet Experimentation
Our previous experiments using a 2-layer feed-forward
neural network yielded a performance increase of 4%, when
comparing a NN trained on the RGB dataset to the NN
trained on the RGBD dataset. These results did’t satisfy
us, as a simple feed-forward neural network may show how
good the features are presented to it, but not the optimal tool
used for image classification. We want to see whether the
estimated depth map truely bring in some new knowledges
of depth, which can’t be obtained just using RGB images.
So we decided to test the performance of ResNet [10] over
the CIFAR-10 RGB dataset and compare it to the performance achieved over our CIFAR-10 RGBD dataset. Similar with 2-layer networks, only number of channels of input
layer is changed from 3 to 4. The increased computational
complexity could be ignored, since later layers have much
more channels.
Shown in Table 1, ResNet-20 improves 0.45% with estimated depth map. ResNet-56 achieved error of 6.45% using the RGBD dataset, which is competitive with ResNet110 using RGB dataset. This performance gain of 0.53%
in accuracy between RGBD and RGB using the CNN could
not be ignored on CIFAR-10. We conclude that training on
RGBD dataset would also result in better performance for
deep networks like ResNet. Depth estimation feature may
be hard to be generated by deep network itself.
5. Conclusion
We created a RGBD image dataset for CIFAR-10. We
define a transfer learning accuracy metric for depth predic-
tion problem. On RGBD CIFAR-10, we show that depth
channel has a better feature representation. Training on
RGBD images could improve image classification on both
shallow and deep networks.
[13]
References
[14]
[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. Goodfellow, A. Harp, G. Irving, M. Isard, Y. Jia,
R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg, D. Mané,
R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster,
J. Shlens, B. Steiner, I. Sutskever, K. Talwar, P. Tucker,
V. Vanhoucke, V. Vasudevan, F. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu, and X. Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 3
[2] J. Carreira, H. Madeira, and J. G. Silva. Xception: A technique for the experimental evaluation of dependability in
modern computers. IEEE Transactions on Software Engineering, 24(2):125–136, 1998. 2
[3] J. Dai, K. He, and J. Sun. Instance-aware semantic segmentation via multi-task network cascades. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 3150–3158, 2016. 1, 2
[4] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. FeiFei. Imagenet: A large-scale hierarchical image database.
In Computer Vision and Pattern Recognition, 2009. CVPR
2009. IEEE Conference on, pages 248–255. IEEE, 2009. 2
[5] D. Eigen and R. Fergus. Predicting depth, surface normals
and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE International
Conference on Computer Vision, pages 2650–2658, 2015. 2
[6] D. Eigen, C. Puhrsch, and R. Fergus. Depth map prediction
from a single image using a multi-scale deep network. In
Advances in neural information processing systems, pages
2366–2374, 2014. 1
[7] R. Girshick. Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pages 1440–1448,
2015. 1, 2
[8] A. Gupta, A. A. Efros, and M. Hebert. Blocks world revisited: Image understanding using qualitative geometry and
mechanics. In Computer Vision–ECCV 2010, pages 482–
496. Springer, 2010. 1
[9] A. Gupta, M. Hebert, T. Kanade, and D. M. Blei. Estimating spatial layout of rooms using volumetric reasoning about
objects and surfaces. In Advances in neural information processing systems, pages 1288–1296, 2010. 1
[10] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385,
2015. 1, 2, 3
[11] V. Hedau, D. Hoiem, and D. Forsyth. Thinking inside the
box: Using appearance models and context based on room
geometry. In Computer Vision–ECCV 2010, pages 224–237.
Springer, 2010. 1
[12] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolu-
[15]
[16]
[17]
[18]
[19]
[20]
tional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, pages 675–678. ACM, 2014. 1, 3
A. Krizhevsky and G. Hinton. Learning multiple layers of
features from tiny images, 2009. 2
L. Ladicky, J. Shi, and M. Pollefeys. Pulling things out of
perspective. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 89–96, 2014. 1
J. Li, Y. Wei, X. Liang, J. Dong, T. Xu, J. Feng, and S. Yan.
Attentive contexts for object detection. IEEE Transactions
on Multimedia, 19(5):944–954, 2017. 2
F. Liu, C. Shen, and G. Lin. Deep convolutional neural fields
for depth estimation from a single image. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 5162–5170, 2015. 1, 2
A. Razavian, H. Azizpour, J. Sullivan, and S. Carlsson. Cnn
features off-the-shelf: an astounding baseline for recognition. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition Workshops, pages 806–813,
2014. 1
S. Ren, K. He, R. B. Girshick, and J. Sun. Faster R-CNN:
towards real-time object detection with region proposal networks. CoRR, abs/1506.01497, 2015. 1, 2
K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014. 2
C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 1–9, 2015. 2
| 1cs.CV
|
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF
HYPERGRAPHS
arXiv:1206.2512v2 [math.AC] 21 Dec 2012
ELIZABETH GROSS AND SONJA PETROVIĆ
Abstract. Associated to any hypergraph is a toric ideal encoding the algebraic relations
among its edges. We study these ideals and the combinatorics of their minimal generators,
and derive general degree bounds for both uniform and non-uniform hypergraphs in terms
of balanced hypergraph bicolorings, separators, and splitting sets. In turn, this provides
complexity bounds for algebraic statistical models associated to hypergraphs. As two main
applications, we recover a well-known complexity result for Markov bases of arbitrary 3way tables, and we show that the defining ideal of the tangential variety is generated by
quadratics and cubics in cumulant coordinates.
1. Introduction
The edge subring of a hypergraph H is the monomial subalgebra parametrized by the edges
of H. We derive a general degree bound for the minimal generators of its defining ideal, IH ,
in terms of the structure of the underlying hypergraph.
Let H be a hypergraph on V =Q{1, . . . , n} with edge set E. Each edge ei ∈ E of size d encodes
a squarefree monomial xei := j∈ei xj of degree d in the polynomial ring k[x1 , . . . , xn ]. The
edge subring of the hypergraph H, denoted by k[H], is the following monomial subring:
k[H] := k[xei : ei ∈ E(H)].
The toric ideal of k[H], denoted IH , is the kernel of the monomial map φH : k[tei ] → k[H]
defined by φH (tei ) = xei . The ideal IH encodes the algebraic relations among the edges of
the hypergraph. For the special case where H is a graph, generating sets of the toric ideal of
k[H] have been studied combinatorially in [12, 13], [15], [19], and [20, 21].
The motivation for studying toric ideals IH is threefold. First, explicit results that relate IH
to combinatorial properties of H have existed only for graphs. Second, IH is related to the
Rees algebra R(I(H)) of the edge ideal of H: in case when H is a graph, the presentation
ideal of R(I(H)) is completely determined by generators of IH and the syzygies of the edge
ideal I(H). Third, these toric ideals correspond to Markov bases for algebraic statistical
models; this connection is outlined at the end of the Introduction. A starting point of our
work is the fact that combinatorial signatures of generators of IH are balanced edge sets of
H. Balanced edge sets on uniform hypergraphs were introduced in [14], and are referred to
as monomial walks.
This paper is based on the fact that the ideal IH is generated by binomials fE arising from
primitive balanced edge sets E of H (See Proposition 3.1, a generalization of [14, Theorem
2.8]). A balanced edge set of H is a multiset of bicolored edges E = Eblue t Ered satisfying
the following balancing condition: for each vertex v covered by E, the number of red edges
1
2
GROSS AND PETROVIĆ
containing v equals the number of blue edges containing v, that is,
(*)
degblue (v) = degred (v).
A binomial fE arises from E if it can be written as
Y
Y
fE =
te −
te0 .
e∈Eblue
e0 ∈Ered
Note that while H is a simple hypergraph (it contains no multiple edges), E allows repetition
of edges. In addition, the balanced edge set E is primitive if there exists no other balanced
0
0
0
0
( Ered ; this is the usual definition
( Eblue and Ered
such that Eblue
t Ered
edge set E 0 = Eblue
of an element in the Graver basis of IH . If H is a uniform hypergraph, a balanced edge set
is called a monomial walk to conform with the terminology in [20, 21] and [14].
In what follows, we give two general degree bounds for generators of IH (Section 5), study the
combinatorics of splitting sets and reducibility (defined in Section 3), and explore implications
to algebraic statistics throughout. Section 4 focuses on indispensable binomials, i.e. binomials
that are members of every minimal generating set of IH . Proposition 4.1 gives a combinatorial
sufficient condition for determining whether a binomial f ∈ IH is indispensable. Consequently,
the Graver basis is the unique minimal generating set of IH for any 2-regular hypergraph
(Proposition 4.2). In particular, this means that the Graver basis is equal to the universal
Gröbner basis, although the defining matrix need not be unimodular.
Theorem 5.1 is a combinatorial criterion for the ideal of a uniform hypergraph to be generated
in degree at most d ≥ 2. The criterion is based on decomposable balanced edge sets, separators, and splitting sets; see Definitions 3.2 and 3.3. Our result generalizes the well-known
criterion for the toric ideal of a graph to be generated in degree 2 from [12] and [20, 21]. Splitting sets translate and extend the constructions used in [12] and [20, 21] to hypergraphs and
arbitrary degrees. Theorem 5.3 provides a more general result for non-uniform hypergraphs.
In algebraic statistics, any log-linear statistical model corresponds to a toric variety whose
defining ideal gives a Markov basis for the model (cf. Fundamental Theorem of Markov bases
[6], [8]). Since these varieties, by definition, have a monomial parametrization, we can also
associate to any log-linear model M with a square-free parameterization a (non-uniform)
hypergraph HM . By Proposition 3.1, Markov moves for the model M are described by
balanced edge sets of HM : if E is a balanced edge set of HM , then a Markov move on a fiber
of the model corresponds to replacing the set of red edges in E by the set of blue edges in
E. Our degree bounds give a bound for the Markov complexity (Markov width) of the model
M. For general references on Markov complexity of classes for some log-linear models, the
reader should refer to [4], [5], [8, Chapter 1, §2] and [10].
We apply our combinatorial criteria to recover a well-known result in algebraic statistics from
[4] in Corollary 4.8. Finally, we study the Markov complexity of a set of models from [18]
called hidden subset models. Namely, Theorem 5.6 says that the ideal associated to the image
of Tan((P1 )n ) in higher cumulants is generated by quadratics and cubics.
2. Preliminaries and notation
We remind the reader that all hypergraphs in this paper are simple, that is, they contain no
multiple edges. In contrast, balanced edge sets of hypergraphs are not, since the binomials
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF HYPERGRAPHS
3
arising from the sets need not be squarefree. Therefore, for the purpose of this manuscript,
we will refer to a balanced edge set as a multiset of edges, with implied vertex set; and, as
usual, V (E) denotes the vertex set contained in the edges in E.
For the remainder of this short section, we will clear the technical details and notation we
need for the proofs that follow.
A multiset, M , is an ordered pair (A, f ) such that A is a set and f is a function from A
to N>0 that records the multiplicity of each of the elements of A. For example, the multiset
M = ({1, 2}, f ) with f (1) = 1 and f (2) = 3 represents M = {1, 2, 2, 2} where ordering
doesn’t matter. We will commonly use the latter notation.
Given
a multiset M = (A, f ), the support of M is supp(M ) := A, and its size is |M | :=
P
a∈A f (a). For two multisets M1 = (A, f1 ) and M2 = (B, f2 ), we say M2 ⊆ M1 if B ⊆ A
and for all b ∈ B, f2 (b) ≤ f1 (b). M2 is a proper submultiset of M1 if B ( A, or there exists
a b ∈ B such that f2 (b) < f1 (b).
Unions, intersections, and relative complements of multisets are defined in the canonical way:
f1 (a)
M1 ∪ M2 := (A ∪ B, g) where g(a) = f2 (a)
max(f (a), f (a))
1
2
if a ∈ A \ B,
if a ∈ B \ A,
if a ∈ A ∪ B;
M1 ∩ M2 := (A ∩ B, g) where g(a) = min(f1 (a), f2 (a));
(
f1 (a)
if a ∈ A \ B,
M1 − M2 := (C, g), where g(a) =
f1 (a) − f2 (a) otherwise.
and C = A \ B ∪ {a ∈ A ∩ B | f1 (a) − f2 (a) > 0}
Note that the support of the union (intersection) of two multisets is the union (intersection)
of their supports. Finally, we define a sum of M1 and M2 :
f1 (a)
M1 t M2 := (A ∪ B, g) where g(a) = f2 (a)
f (a) + f (a) if a ∈ A ∩ B
1
2
if a ∈ A \ B,
if a ∈ B \ A .
If M1 t M2 is a balanced edge set, then the notation M1 tb M2 will be used to record the
bicoloring of M1 t M2 : edges in M1 are blue, and edges in M2 are red.
Finally, the number of edges in a hypergraph H containing a vertex v will be denoted by
deg(v; H). For a bicolored multiset M := Mblue tm Mred , the blue degree degblue (v; M ) of a
vertex v is defined to be deg(v; Mblue ). The red degree degred (v; M ) is defined similarly.
4
GROSS AND PETROVIĆ
3. Splitting sets and reducible edge sets
The aim of this section is to lay the combinatorial groundwork for studying toric ideals of
hypergraphs. In particular, we explicitly state what it combinatorially means for a binomial
arising from a monomial walk to be generated by binomials of a smaller degree. We begin by
describing the binomial generators of IH . Unless otherwise stated, H need not be uniform.
Proposition 3.1. Every binomial in the toric ideal of a hypergraph corresponds to a balanced
edge set. In particular, the toric ideal IH is generated by primitive balanced edge sets.
Proof. Suppose E is a balanced multiset of edges over H. Define a binomial fE ∈ k[te : e ∈
E(H)] as follows:
Y
Y
fE =
te −
te0 .
e∈Eblue
e0 ∈Ered
The balancing condition (*) ensures that fE is in the kernel of the map φH .
The second claim is immediate.
Motivated by the application of reducible simplicial complexes to understand the Markov
bases of hierarchical log-linear models [7], we now introduce notions of reducibility and separators for balanced edge sets. For simplicity, we will often abuse notation and use H to
denote the edge set of H.
Definition 3.2. A balanced edge set E is said to be reducible with separator S, supp(S) ⊆
supp(E), and decomposition (Γ1 , S, Γ2 ), if there exist balanced edge sets Γ1 6= E and Γ2 6= E
with S 6= ∅ such that S = Γ1red ∩ Γ2blue , E = Γ1 t Γ2 , and the following coloring conditions
hold: Γ1red , Γ2red ⊆ Ered and Γ1blue , Γ2blue ⊆ Eblue .
We say that S is proper with respect to (Γ1 , S, Γ2 ) if S is a proper submultiset of both Γ1red
and Γ2blue .
If S is not proper, then S is said to be blue with respect to (Γ1 , S, Γ2 ) if Γ1red = S, and red
with respect to (Γ1 , S, Γ2 ) if Γ2blue = S.
Figure 1 shows an example of a reducible balanced edge set E. The separator is proper and
consists of the single green edge es ; it appears twice in the balanced edge set E, once as a
blue edge and once as a red edge. Figure 2 shows a reducible balanced edge set where the
separator, consisting of the two green edges e1 and e2 , is not proper. As before, the separator
edges appear twice in the balanced edge set.
If H is a hypergraph and E is a balanced edge set with supp(E) ⊆ H, given a multiset S
with supp(S) ⊆ H, we can construct a new balanced edge set in the following manner:
E + S := (Eblue t S) tm (Ered t S).
Definition 3.3. Let H be a hypergraph. Let E be a balanced edge set with size 2n such
that supp(E) ⊆ H. A non-empty multiset S with supp(S) ⊆ H is a splitting set of E with
decomposition (Γ1 , S, Γ2 ) if E +S is reducible with separator S and decomposition (Γ1 , S, Γ2 ).
S is said to be a blue (red, resp.) splitting set with respect to (Γ1 , S, Γ2 ), if S is a blue (red,
resp.) separator of E + S with respect to (Γ1 , S, Γ2 ).
S is a proper splitting set of E if there exists a decomposition (Γ1 , S, Γ2 ) of E + S such that
S is a proper separator with respect to (Γ1 , S, Γ2 ).
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF HYPERGRAPHS
!"#
!"#
Figure 1. Reducible
balanced edge set. The
green edge es is the
separator.
5
!$#
Figure 2. Reducible balanced
edge set with an improper separator. The separator consists of
green edges e1 and e2 .
Example 3.4 (Group-based Markov model). Let V1 = {x1 , x2 , x3 , x4 }, V2 = {y1 , y2 , y3 , y4 },
and V3 = {z1 , z2 , z3 , z4 }. Let V be the disjoint union of V1 , V2 , and V3 . Let H be the 3-uniform
hypergraph with vertex set V and edge set:
e111
e221
e331
e441
= {x1 , y1 , z1 }
= {x2 , y2 , z1 }
= {x3 , y3 , z1 }
= {x4 , y4 , z1 }
e122
e212
e342
e432
= {x1 , y2 , z2 }
= {x2 , y1 , z2 }
= {x3 , y4 , z2 }
= {x4 , y3 , z2 }
e133
e243
e313
e423
= {x1 , y3 , z3 }
= {x2 , y4 , z3 }
= {x3 , y1 , z3 }
= {x4 , y2 , z3 }
e144
e234
e324
e414
= {x1 , y4 , z4 }
= {x2 , y3 , z4 }
= {x3 , y2 , z4 }
= {x4 , y1 , z4 }
The hypergraph H has applications in algebraic phylogenetics: it represents the parametrization of a particular group-based model from [17, Example 25].
Consider the monomial walk
W = {e324 , e111 , e243 , e432 } tm {e122 , e313 , e234 , e441 }.
Let S = {e133 , e212 }. Then S is a splitting set of W with decomposition (Γ1 , S, Γ2 ) where
Γ1 = {e111 , e243 , e432 } tm {e133 , e212 , e441 }
Γ2 = {e133 , e212 , e324 } tm {e122 , e313 , e234 }.
The decomposition (Γ1 , S, Γ2 ) encodes binomials in IH that generate fW :
fW = te324 (te111 te243 te432 − te133 te212 te441 ) + te441 (te133 te212 te324 − te122 te313 te234 ).
The previous example illustrates the algebraic interpretation of a splitting set. Notice there
is a correspondence between monomials in k[tei ] and multisets of edges of H. We will write
E(taei11 taei22 · · · taeil ) for the multiset ({ei1 , . . . , eil }, f ) where
l
f : {ei1 , . . . , eil } → N
eij 7→ aj .
Thus the support of E(taei11 taei22 · · · taeil ) corresponds to the support of the monomial taei11 taei22 · · · taeil .
l
l
If fE = u − v ∈ IH is the binomial arising from the balanced edge set E, then a monomial s
corresponds to a splitting set S if and only if there exist two binomials u1 − v1 , u2 − v2 ∈ IH
such that us = u1 u2 , vs = v1 v2 and s = gcd(v2 , u1 ). In this case, the decomposition of E + S
is (Γ1 , S, Γ2 ) where Γ1 = E(u1 ) tm E(v1 ) and Γ2 = E(u2 ) tm E(v2 ).
6
GROSS AND PETROVIĆ
For a balanced edge set, E, the existence of a spitting set determines whether the binomial
fE ∈ IH can be written as the linear combination of two binomials fΓ1 ,fΓ2 ∈ IH . While, in
general, the existence of a splitting set does not imply deg(fΓ1 ), deg(fΓ2 ) < deg(fE ), if H is
uniform and the splitting set is proper, then the following lemma holds.
Lemma 3.5. Let H be a uniform hypergraph and let W be a monomial walk with supp(W) ⊆
H and |W| = 2n. If S is a proper splitting set of W, then there exists a decomposition
(Γ1 , S, Γ2 ) of W + S such that |Γ1 | < |W| and |Γ2 | < |W|.
Proof. Let S be a proper splitting set of W. By definition, there exists a decomposition
(Γ1 , S, Γ2 ) of W + S, such that S is a proper submultiset of Γ1red and Γ2blue .
Let |Γ1 | = 2n1 and |Γ2 | = 2n2 . Since W + S = Γ1 t Γ2 , it follows that |W + S| = |Γ1 | + |Γ2 |.
Then, 2n + 2|S| = 2n1 + 2n2 , which implies 2n − 2n1 = 2n2 − 2|S|. But S being a proper
submultiset of Γ2blue gives that n2 > |S|, which, in turn, implies that n > n1 . By a similar
argument, n > n2 . Thus |Γ1 | < |W| and |Γ2 | < |W|.
4. Indispensable Binomials
A binomial f in a toric ideal I is indispensable if f or −f belongs to every binomial generating
set of I. Indispensable binomials of toric ideals were introduced by Takemura et al, and are
studied in [1], [2], [3], [13], [15].
Proposition 4.1. Let H be a hypergraph. Let E be a balanced edge set with supp(E) ⊆ H.
Let fE be the binomial arising from E. If there does not exist a splitting set of E, then fE is
an indispensable binomial of IH .
Proof. Suppose E is not indispensable. Then there is a binomial generating set of IH , G =
{f1 , . . . , fn }, such that fE ∈
/ G and −fE ∈
/ G.
Since fE = fE+ −fE− ∈ IH , there is a fi = fi+ −fi− ∈ G such that fi+ or fi− divides fE+ . Without
loss of generality, assume fi+ |fE+ . Since fi is a binomial in IH , fi arises from a monomial walk
Ei on H.
Let S = Eired . Let Γ1 = Ei and Γ2 = Γ2blue tm Γ2red where
Γ2blue = ((Eblue − Eiblue ) t Eired )
Γ2red = Ered .
Since fi+ |fE+ , the multiset Eiblue ⊆ Eblue , and thus Γ1 t Γ2 = E + S. By construction, Γ1red ∩
Γ2blue = S. Therefore S is a splitting set of E.
If every Graver basis element of a binomial ideal IH is indispensable, then the Graver basis of
IH is the unique minimal generating set of IH . Propositions 4.2 and 4.6 describe two classes
of hypergraphs where this is the case. In particular, for these hypergraphs, the universal
Gröbner basis of IH is a minimal generating set.
Proposition 4.2. If H is a 2-regular uniform hypergraph, then the Graver basis of IH is the
unique minimal generating set of IH .
For the proof of Proposition 4.2, we make use of Proposition 3.2 in [14] which concerns
balanced edge sets that are pairs of perfect matchings.
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF HYPERGRAPHS
7
Definition 4.3. A matching on a hypergraph H = (V, E) is a subset M ⊆ E such that the
elements of M are pairwise disjoint. A matching is called perfect if V (M ) = V .
Proof of Proposition 4.2. Let G be the Graver basis of IH and let f ∈ G. Since every element
of G is binomial, f arises from a primitive monomial walk W with supp(W) ⊆ H.
Let Mb = supp(Wred ) and Mr = supp(Wblue ). By primitivity of W, the intersection Mr ∩Mb =
∅. Since W satisfies condition (*) and H is 2-regular, if e1 , e2 ∈ Mb and e1 ∩ e2 6= ∅, then
e1 ∈ Mr or e2 ∈ Mr , which would contradict the primitivity of W. So Mb and Mr are
two edge-disjoint perfect matchings on V (W). By Proposition 3.2 in [14], W contains no
multiple edges, i.e. W = Mb tm Mr . Furthermore, since H is 2-regular, the edge set of the
subhypergraph induced by V (W) is Mb ∪ Mr
Suppose S is a splitting set of W with decomposition (Γ1 , S, Γ2 ). By the correspondence
between primitive monomial walks and primitive binomials, there exists a primitive monomial
walk Γ such that Γblue ⊆ Γ1blue and Γred ⊆ Γ1red (if Γ1 is primitive, then Γ = Γ1 ). By
Proposition 3.2 in[14], Γ must be a pair of perfect matchings on V (Γ ). This means Γ is
a proper balanced edge set of W, a contradiction. Therefore, by Proposition 4.1, fW is
indispensable. Since every element in the Graver basis of IH is indispensable, there is no
generating set of IH strictly contained in the Graver basis, and the claim follows.
Definition 4.4. A k-uniform hypergraph H = (V, E) is k-partite if there exists a partition
of V into k disjoint subsets, V1 , . . . , Vk , such that each edge in E contains exactly one vertex
from each Vi .
Lemma 4.5. Let H = (V, E) be a k-uniform k-partite hypergraph with E = Eb t Er and
Eb ∩ Er = ∅. If there exists a Vi , 1 ≤ i ≤ k, such that deg(v; Er ) = deg(v; Eb ) = 1 for all
v ∈ Vi , then a monomial walk W with support E is primitive only if W contains no multiple
edges.
Proof. Follows from the proof of necessity of Proposition 3.2 in [14].
Proposition 4.6. Let H = (V, E) be a k-uniform k-partite hypergraph. If there exists a Vi
such that deg(v; E) = 2 where for all v ∈ Vi , then the Graver basis of IH is the unique
minimal generating set of IH .
Proof. The proof is similar to the proof of Proposition 4.2. Note that while H may not be 2regular, one of its parts, Vi , is ‘locally’ 2-regular, and thus restricts the structure of monomial
walks on H. In particular, Lemma 4.5 ensures that Mr and Mb , are edge-disjoint perfect
matchings on V (W)|Vi , and the rest of the proof follows immediately.
Example 4.7 (No 3-way interaction). The toric ideal of the hypergraph H in Figure
3 corresponds to the hierarchical log-linear model for no 3-way interaction on 2 × 2 × 2
contingency tables. This statistical model is a common example in algebraic statistics [8,
Example 1.2.7]. Since there is exactly one primitive monomial walk W on H that travels
through 8 edges, IH = (fW ).
For 2 × 3 × 3 contingency tables with no 3-way interaction, the hypergraph corresponding
to this log-linear model has 18 edges. The hypergraph in this case is H = (V, E) where V =
{x00 , x01 , x02 , x10 , x11 , x12 , y00 , y01 , y02 , y10 , y11 , y12 , z00 , z01 , z02 , z10 , z11 , z12 , z20 , z21 , z22 } and the
8
GROSS AND PETROVIĆ
Figure 3.
edge set is:
e000
e010
e020
e100
e110
e120
= {x00 , y00 , z00 }
= {x01 , y00 , z10 }
= {x02 , y00 , z20 }
= {x10 , y10 , z00 }
= {x11 , y10 , z10 }
= {x12 , y10 , z20 }
e001
e011
e021
e101
e111
e121
= {x00 , y01 , z01 }
= {x01 , y01 , z11 }
= {x02 , y01 , z21 }
= {x10 , y11 , z01 }
= {x11 , y11 , z11 }
= {x12 , y11 , z21 }
e002
e012
e022
e102
e112
e122
= {x00 , y02 , z02 }
= {x01 , y02 , z12 }
= {x02 , y02 , z22 }
= {x10 , y12 , z02 }
= {x11 , y12 , z12 }
= {x12 , y12 , z22 }
Let W be the primitive monomial walk
W = {e000 , e101 , e011 , e112 , e022 , e120 } tm {e100 , e001 , e111 , e012 , e122 , e220 .}
Every remaining edge H that does not appear in W is not contained in V (W), thus it can
be easily verified that there does not exist a splitting set of W, so by Proposition 4.1, fW is
indispensable. In fact, H satisfies the condition of Proposition 4.6 and thus every binomial
in IH corresponding to a primitive monomial walk is indispensable.
From the above discussion, we can see that if a uniform hypergraph H contains an induced
subhypergraph Hs that is 2-regular and there exists a bicoloring such that with this bicoloring
Hs is also a balanced edge set, then the maximum degree of any minimal generating set of
IH is at least |E(Hs )|/2.
A similar statement holds for k-uniform, k-partite hypergraphs with vertex partition V =
∪ki=1 Vi . Namely, if H contains an induced subhypergraph Hs that is 2-regular on Vi (i.e., H
satisfies the conditions of Proposition 4.6) and there exists a bicoloring such that with this
bicoloring Hs is a balanced edge set (e.g., Hs is a pair of disjoint perfect matchings), then
the maximum degree of any minimal generating set of IH is at least |E(Hs )|/2.
Recall that degree bounds on minimal generators give a Markov complexity bound for the
corresponding log-linear model in algebraic statistics. This allows us to recover a well-known
result:
Corollary 4.8 (Consequence of Theorem 1.2 in [4]; see also Theorem 1.2.17 in [8]). The
Markov complexity for the no 3-way interaction model on 3 × r × c contingency tables grows
arbitrarily large as r and c increase .
Proof. For the no 3-way interaction model on 2 × r × c contingency tables, we can construct
a primitive binomial fHs of degree 2 · min(r, c) in its defining toric ideal by taking a cycle
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF HYPERGRAPHS
9
of length min(r, c) on the bipartite graph Kr,c . (We remind the reader that this is precisely
how fW is constructed in Example 4.7). By noting that the hypergraph associated to this
binomial Hs is an induced subhypergraph of the hypergraph associated to the 3 × r × c case
and that Hs is 2-regular in one of the partitions, the claim follows by Proposition 4.6.
5. General degree bounds and an application
For uniform hypergraphs, balanced edge sets are referred to as monomial walks. In the previous sections, we saw that splitting sets of W translate to algebraic operations on the binomials
fW , providing a general construction for rewriting a high-degree binomial in terms of binomials corresponding to shorter walks. This, along with Lemma 3.5, is the key to the general
degree bound result.
Theorem 5.1. Given a k-uniform hypergraph H, the toric ideal IH is generated in degree
at most d if and only if for every primitive monomial walk W of length 2n > 2d, with
supp(W) ⊆ H, one of the following two conditions hold:
i) there exists a proper splitting set S of W,
or
ii) there is a finite sequence of pairs, (S1 , R1 ), . . . , (SN , RN ), such that
• S1 and R1 are blue and red splitting sets of W of size less than n with decompositions
(Γ11 , S1 , Γ21 ) and (Υ11 , R1 , Υ21 ),
• Si+1 and Ri+1 are blue and red splitting sets of Wi = Γ2iblue tm Υ1ired of size less than
n with decompositions (Γ1i+1 , Si+1 , Γ2i+1 ) and (Υ1i+1 , Ri+1, Υ2i+1 ), and,
• SN ∩ RN 6= ∅ or there exists a proper splitting set of WN .
Proof of necessity (⇒). Let H be a k-uniform hypergraph whose toric ideal IH is generated
in degree at most d. Let W be a primitive monomial walk of length 2n > 2d. Let pW = u − v
be the binomial that arises from W. Since IH is generated in degree at most d, there exist
primitive binomials of degree at most d, (u1 − v1 ), . . . , (us − vs ) ∈ k[tei ], and m1 , . . . , ms ∈
k[tei ], such that
pW = m1 (u1 − v1 ) + m2 (u2 − v2 ) + . . . + ms (us − vs ).
By expanding and reordering so that m1 u1 = uw , ms vs = vw , and mi vi = mi+1 ui+1 for all
i = 1, . . . , s − 1, we may and will assume that m1 , . . . , ms are monomials.
If gcd(mi , mi+1 ) 6= 1 for some i, we can add the terms mi (ui − vi ) and mi+1 (ui+1 − vi+1 ) to
get a new term, m0i (u0i − vi0 ), where m0i = gcd(mi , mi+1 ) and (u0i − vi0 ) is an binomial of IH of
degree less than n. Continuing recursively in the manner, we have
pW = m01 (u01 − v10 ) + m02 (u02 − v20 ) + . . . + m0r (u0r − vr0 )
where m01 u01 = u0w , m0r vr0 = vw0 , m0i vi0 = m0i+1 u0i+1 , gcd(m0i , m0i+1 ) = 1 for all i = 1, . . . , r − 1,
and deg(u0i − vi0 ) < n for all i = 1, . . . r. For convenience, we will drop the superscripts and
write
pw = m1 (u1 − v1 ) + m2 (u2 − v2 ) + . . . + mr (ur − vr ).
10
GROSS AND PETROVIĆ
Case 1: r = 2. In this case, pW = m1 (u1 − v1 ) + m2 (u2 − v2 ). Let
Γ1 := E(u1 ) tm E(v1 )
Γ2 := E(u2 ) tm E(v2 )
S := E(v1 ) ∩ E(u2 ) = E(gcd(v1 , u2 )).
We want to show (Γ1 , S, Γ2 ) is a decomposition of W + S. Since S = Γ1red ∩ Γ2blue , Γ1blue ⊆
Wblue , and Γ2red ⊆ Wred , we only need to show W + S = Γ1 t Γ2 , Γ2red ⊆ (W + S)red , and
Γ2blue ⊆ (W + S)blue . First, notice the following equalities hold:
W + S = (Wblue t S) t (Wred t S) = E(u) t S t E(v) t S
= E(m1 u1 ) t S t E(m2 v2 ) t S = E(m1 ) t E(u1 ) t S t E(m2 ) t E(v2 ) t S.
Let s ∈ k[tei ] be the monomial such that E(s) = S, so s = gcd(v1 , u2 ). The equality m1 v1 =
m2 u2 implies m1 ( vs1 ) = m2 ( us2 ). Now, vs1 and us2 are clearly relatively prime, and by the
assumptions on pW , m1 and m2 are relatively prime. This means the equality m1 ( vs1 ) = m2 ( us2 )
implies m1 = us2 and m2 = vs1 . Thus,
Γ1 t Γ2 = E(u1 ) t E(v1 ) t E(u2 ) t E(v2 )
v1
u2
= E(u1 ) t E( ) t S t E(v2 ) t E( ) t S
s
s
= E(u1 ) t E(m2 ) t S t E(v2 ) t E(m1 ) t S.
Consequently, W + S = Γ1 t Γ2 .
Notice the equality m2 = vs1 also implies Γ1red = E(v1 ) = E(m2 ) t S. This means Γ1red ⊆
(E(m2 u2 ) t S) = (Wred t S) = (W + S)red . By a similar observation, Γ2blue ⊆ (W + S)blue .
Case 2: r = 2N + 1. For 1 < i < N , let
Γ1i = E(ui ) tm E(vi )
Γ2i = E(mi+1 ui+1 ) tm E(m2N −i+2 v2N −i+2 )
Si = E(vi ) ∩ E(mi+1 ui+1 ) = E(gcd(vi , mi+1 ui+1 )) = E(vi ).
For 1 < i < N , let
Υ1i = E(mi ui ) tm E(m2N −i+1 v2N −i+1 )
Υ2i = E(u2N −i+2 ) tm E(v2N −i+2 )
Ri = E(m2N −i+1 v2N −i+1 ) ∩ E(u2N −i+2 )
= E(gcd(m2N −i+1 v2N −i+1 , u2N −i+2 )) = E(u2N −i+2 ).
One can follow the proof of Case 1) to see that S1 and R1 are splitting sets of W, and Si+1
and Ri+1 are splitting sets of Wi = E(mi+1 ui+1 ) tm E(m2N −i+1 v2N −i+1 ) for i = 1, . . . , N − 1.
Furthermore, by definition, they are blue and red splitting sets (resp.) of size less than 2n.
Since WN −1blue = Γ2N −1blue and WN −1red = Υ1N −1red , the binomial arising from the walk on
WN −1 is
mN uN − mN +2 vN +2 = mN (uN − vN ) + mN +1 (uN +1 − vN +1 ) + mN +2 (uN +2 − vN +2 ).
Choose e ∈ H such that te | mN +1 , then te | vN and te | uN +2 . But since SN = E(vN ) and
RN = E(uN +2 ), e ∈ SN and e ∈ RN , so SN ∩ RN 6= ∅.
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF HYPERGRAPHS
11
Case 3: r = 2N + 2. For 1 < i < N , let
Γ1i = E(ui ) tm E(vi )
Γ2i = E(mi+1 ui+1 ) tm E(m2N −i+3 v2N −i+3 )
Si = E(vi ) ∩ E(mi+1 ui+1 ) = E(gcd(vi , mi+1 ui+1 )) = E(vi ).
For 1 < i < N , let
Υ1i = E(mi ui ) tm E(m2N −i+2 v2N −i+2 )
Υ2i = E(u2N −i+3 ) tm E(v2N −i+3 )
Ri = E(m2N −i+2 v2N −i+2 ) ∩ E(u2N −i+3 )
= E(gcd(m2N −i+2 v2N −i+2 , u2N −i+3 )) = E(u2N −i+3 ).
We can follow the proof of Case 1) to see that S1 and R1 are splitting sets of W, and Si+1
and Ri+1 are splitting sets of Wi = E(mi+1 ui+1 ) tm E(m2N −i+2 v2N −i+2 ) for i = 1, . . . , N − 1.
Furthermore, by definition, they are blue and red (resp.) splitting sets of size less than n.
Since WNblue = Γ2Nblue and WNred = Υ1Nred , the binomial arising from WN is
mN +1 uN +1 − mN +2 vN +2 = mN +1 (uN +1 − vN +1 ) + mN +2 (uN +2 − vN +2 )
which is exactly case 1), which means there exists a proper splitting set of WN .
Proof of sufficiency (⇐). Assume every primitive monomial walk W of length 2n > 2d with
supp(W) ⊂ H satisfies i) or ii). Let pW = u − v be a generator of IH which arises from the
monomial walk W on H.
To show that IH = [IH ]≤d , we proceed by induction on the degree of pW . If deg pW = 2, then
pW ∈ [IH ]≤d . So assume deg pW = n > d and every generator of IH of degree less than n is
in [IH ]≤d . Since the size of W is greater than 2d, either condition i) holds or condition ii)
holds.
Suppose i) holds. By Lemma 3.5, there exists a decomposition of W, (Γ1 , S, Γ2 ), such that
|Γ1 | < |W| and |Γ2 | < |W|. Let pΓ1 = u1 − v1 (pΓ2 = u2 − v2 , respectively) be the binomial
that arises from Γ1 (Γ2 , respectively). Let m1 = u/u1 and m2 = v/v2 .
What remains to be shown is that pW = m1 pΓ1 + m2 pΓ2 , that is, u − v = m1 (u1 − v1 ) +
m2 (u2 − v2 ). However, it is clear that u = m1 u1 and v = m2 v2 , so it suffices to show is that
m1 v1 = m2 u2 , or equivalently, E(m1 v1 ) = E(m2 u2 ).
Let s ∈ k[tei ] be the monomial such that E(s) = S. Then
v1
u2
Γ1 t Γ2 = (E(u1 ) t E( ) t S) t (E( ) t S t E(v2 ))
s
s
and
W + S = (E(m1 ) t E(u1 ) t S) t (E(m2 ) t E(v2 ) t S).
Thus, since W + S = Γ1 t Γ2 ,
v1
u2
E(m1 ) t E(m2 ) = E( ) t E( ),
s
s
which in turn implies
v1 u2
m1 m2 = ( )( ).
s s
12
GROSS AND PETROVIĆ
Since W is primitive and the coloring conditions on (Γ1 , S, Γ2 ) imply E( vs1 ) ⊆ Wred and
E(m1 ) ⊆ Wblue , the monomials m1 and vs1 are relatively prime. A similar argument shows
m2 and us2 are relatively prime. Thus, m1 = us2 and m2 = vs1 , and consequently, E(m1 v1 ) =
E(m2 u2 ) and pw = m1 pΓ1 + m2 pΓ2 .
Since deg pΓ1 , deg pΓ2 < n, the induction hypothesis applied to pΓ1 and pΓ2 shows that pW ∈
[IH ]≤d .
Now suppose ii) holds. For i from 1 to N , let pΓ1i = ui − vi and pΥ2i = yi − zi be the
binomials arising from Γ1i and Υ2i . Let wib − wir be the binomial arising from the walk Wi
and let pW = w0b − w0r . For 1 ≤ i ≤ N , let mi = w(i−1)b /ui , and qi = w(i−1)r /zi . Then
pW =
N
X
i=1
mi (ui − vi ) + wNb − wNr +
N
X
qN +1−i (yN +1−i − zN +1−i ).
i=1
The preceding claim follows from three observations: (1) by construction, w0b = m1 u1 and
w0r = q1 z1 ; (2) by the definition of WN , wNb = mN vN and wNr = qN yN ; and (3) by the
definitions of mi , qi , and the walk Wi , mi vi = mi+1 ui+1 and qi+1 zi+1 = qi yi for 1 ≤ i ≤ N − 1.
As
on the splitting sets of Wi , the linear combination
PN
PNa consequence of the size conditions
q
m
(u
−
v
)
∈
[I
]
and
i i
i
H ≤d
i=1 N +1−i (yN +1−i − zN +1−i ) ∈ [IH ]≤d . So if WN satisfies
i=1
condition i), the binomial wNb − wNr ∈ [IH ]≤d , and thus, pW ∈ [IH ]≤d .
To finish the proof, assume that SN and RN share an edge, e. Then the claim above becomes:
pW =
N
X
i=1
N
X
mN vN
qN y N
mi (ui − vi ) + te (
−
)+
qN +1−i (yN +1−i − zN +1−i )
te
te
i=1
and we just need to show that, in fact, te divides mN vN and qN yN . But this is clear to see
since e ∈ SN which implies te |vN and e ∈ RN which implies te |yN .
Example 5.2 (Independence models). Let H be the complete k-partite hypergraph with
d vertices in each partition V1 , . . . , Vk . These hypergraphs correspond to the independence
model in statistics. Equivalently, the edge subring of the complete k-partite hypergraph with
d vertices in each partition parametrizes the Segre embedding of Pd × · · · × Pd with k copies.
The ideal IH is generated by quadrics. To see this, let W, supp(W) ⊆ H, be a primitive
monomial walk of length 2n, n > 2. Choose a multiset E 0 ⊂ W consisting of n − 1 blue and
n − 1 red edges. Since each edge must contain a vertex from each Vi , for each i, there is at
most one vertex in V (E 0 ) ∩ Vi that is not covered by a red edge and a blue edge from E 0 .
Consequently, V (E 0 ) contains a vertex from each Vi that belong to at least one red edge and
at least one blue edge of E 0 .
For a multiset of edges, M , with supp(M ) ⊆ H, we define the max degree of a vertex:
maxdeg(v; M ) := max(degred (v; M ), degblue (v; M )).
The partitioning of the vertices ensures that V (E 0 ) cannot contain more then k vertices whose
maxdeg with respect to E 0 is n − 1. Indeed, if there are more that k vertices with maxdeg
equal to n − 1, then two of those vertices must belong to the same partition, Vj . This would
imply that W contains at least 4(n − 1) edges, which is impossible when n > 2.
Next, choose n − 1 new blue edges and n − 1 red edges in the following manner:
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF HYPERGRAPHS
13
Figure 4.
Let db (v) := degblue (v; E 0 ) and dr (v) := degred (v; E 0 ). For i = 1, . . . , k choose a vertex from
0
0
V (Eblue
) ∩ V (Ered
) ∩ Vi that has the largest maxdeg with respect to E 0 ; let bn−1 and rn−1 be
this set of vertices. For all v ∈ bn−1 , reduce db (v) and dr (v) by 1. Now choose b1 , . . . , bn−2 by
the following algorithm:
for i from 1 to k do:
let Vi :=sort V (E 0 ) ∩ Vi by db (v) in decreasing order;
for j from n − 2 down to 1 do:
(
bj := list {vi : vi is first element in Vi };
for all v ∈ bj do db (v) = db (v) − 1;
for i from 1 to k do Vi =sort Vi by db (v) in decreasing order;
).
Let R1 = {b1 , . . . , bn−1 } and S1 = {r1 , . . . , rn−1 }. Then R1 and S1 are red and blue splitting
sets of W that share an edge. Thus, condition ii) of Theorem 5.1 is met, and consequently
IH is generated in degree 2.
When H is a non-uniform hypergraph, the toric ideal IH is not necessarily homogeneous. For
example, Figure 4 supports a binomial in IH where H consists of edges of size two and four;
note that the edges still satisfy the balancing condition (*). However, we can still modify
the conditions of Theorem 5.1 to find degree bounds for the toric ideals of non-uniform
hypergraphs. Proposition 5.3 gives a prescription for determining a degree bound on the
generators of IH in terms of local structures of H.
Proposition 5.3. Given a hypergraph H and a binomial fE ∈ IH arising from the balanced
edge set E with n = |Eblue | ≥ |Ered |, fE is a linear combination of binomials in IH of degree
less than n if one of the following two conditions hold:
i) there exists a proper splitting set S of E with decomposition (Γ1 , S, Γ2 ) where |Γiblue |, |Γired | <
n for i = 1, 2,
or
ii) there is a pair of blue and red splitting sets of E, S and R, of size less than n with
decompositions (Γ1 , S, Γ2 ), (Υ1 , R, Υ2 ) such that |Γ1blue |, |Υ2red | < n, |Γ2blue |, |Υ1red | ≤ n, and
S ∩ R 6= ∅.
Proof. This proof follows the proof of sufficiency for Theorem 5.1. Note that in the proof,
the uniform condition doesn’t play an essential role; it is only invoked to bound the size
of the red and blue parts of each monomial hypergraph appearing in the decompositions
14
GROSS AND PETROVIĆ
involved. Thus, the hypothesis of Proposition 5.3 acts in place of the uniform condition in
Theorem 5.1.
We close with an application.
For the remainder of this section, we will concern ourselves with the first tangential variety,
Tan((P1 )n ). In [18], Sturmfels and Zwiernik use cumulants to give a monomial parameterization of Tan((P1 )n ). The variety Tan((P1 )n ) is associated to a class of hidden subset models
[18, Example 5.2], and context-specific independence models [11]. We now derive a bound
for the toric ideal of the image of Tan((P1 )n ) in higher cumulants and, equivalently, for the
Markov complexity of these models.
Example 5.4. Let H = (V, E) where V = {1, . . . , n} and E = {e : e ⊆ V and |e| ≥ 2}.
Then the set of polynomials vanishing on the image of Tan((P1 )n ) in higher cumulants is the
toric ideal IH (see [18, Theorem 4.1]).
The hypergraph in Example 5.4 is the complete hypergraph on n vertices after removing all
singleton edges. The degree bound on the generators of this hypergraph can be found by
looking at a smaller hypergraph.
Lemma 5.5. Let H1 = (V, E1 ) where V = {1, . . . , n} and E1 = {e : e ⊆ V and |e| ≥ 2},
and let H2 = (V, E2 ) where E2 = {e ⊆ V : 2 ≤ |e| ≤ 3}. If the ideal IH2 is generated in
degree at most d, then the ideal IH1 is generated in degree at most d.
Proof. Consider IH2 as an ideal in the bigger polynomial ring S := k[tei : ei ∈ H1 ], denoted
as I˜H2 := IH2 S. Assume that IH2 , and consequently, I˜H2 , is generated in degree at most d.
Pick an arbitrary binomial
u − v = tei1 tei2 · · · tein − tej1 tej2 · · · tejm ∈ IH1 .
Since every edge e ∈ H1 is the disjoint union of a collection of edges ek1 , . . . , ekl ∈ H2 , we
Q
may write te − li=1 teki ∈ IH1 . Noting that
" j
!
#
l
l−2
Y
Y
X
te −
teki = (te − tek1 t∪li=2 ek ) −
teki (t∪li=j+1 ek − tej+1 t∪li=j+2 ek ) ,
i
i=1
i
j=1
i
i=1
Q
one easily sees that the binomial te − li=1 teki is generated by quadratics. In turn, this essentially shows that relations in IH2 allow us to rewrite u−v in terms of edges ei1 , . . . , ein , ej1 , . . . , ejm ∈
E2 of size 2 and 3 only. The claim follows since u−v can be expressed as a binomial in I˜H2 .
Theorem 5.6. Let H = (V, E) where V = {1, . . . , n} and E = {e ⊆ V : 2 ≤ |e| ≤ 3}. The
toric ideal of H is generated by quadrics and cubics.
In particular the image of Tan((P1 )n ) in higher cumulants is generated in degrees 2 and 3.
In the following proof, we examine the local combinatorics of H to illustrate how the structure
of a hypergraph reveals insights into the generating set of IH .
Proof. Let fE be a primitive binomial in IH with E a balanced edge set. Without loss of
generality, we will assume throughout the proof |Eblue | ≥ |Ered |. If E contains only 2-edges or
only 3-edges, then by [16, Theorem 14.1] fE is a linear combination of quadratics. So we will
assume E contains a 2-edge and a 3-edge.
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF HYPERGRAPHS
15
Since |Eblue | ≥ |Ered |, Eblue must contain at least as many 2-edges as Ered , and in order to
satisfy (*), the difference between the number of 3-edges in Ered and the number of 3-edges
in Eblue must be a multiple of 2.
Notice that for every pair e1 , e2 of 3-edges (where e1 and e2 do not need to be unique), there
are three 2-edges in H, e3 , e4 , e5 , such that
{e1 , e2 } tm {e3 , e4 , e5 }
is a balanced edge set. Let B2,3 ⊂ IH be the set of all binomials arising from balanced edge
0
sets of this form. Then fE is a linear combination of binomials in B2,3 and fE 0 , where Eblue
0
and Ered contains the same number of 2-edges and exactly one 3-edge.
Since it suffices to consider primitive binomials, we will proceed inductively by showing that
every primitive degree n binomial in
Bh := {fE ∈ IH : |Eblue | = |Ered | and Eblue , Ered contain exactly one 3-edge each}
is a linear combination of binomials in Bh with degree less than n.
Let fE ∈ Bh such that degree fE = n > 3 and fE is primitive. Let e1 be the 3-edge in Ered .
Since fE is primitive, e1 must intersect a 2-edge e2 in Eblue . Let e2 = {v1 , v2 } where v1 ∈ e1 .
The edge e2 intersects at most one other edge of Ered besides e1 . We will examine the possible
intersections of e2 and Ered in order to find splitting sets of E that satisfy one of the conditions
listed in Proposition 5.3. For illustrations of Case 1 and Case 3 see Figures 5 and 6. In all three
cases, we will construct S, Γ1 and Γ2 such that S is a splitting set of E with an associated
decomposition (Γ1 , S, Γ2 ) which satisfies the properties of condition i) in Theorem 5.3. In
fact, fE will be a linear combination of fΓ1 and fΓ2 , both of which have strictly lower degree
than fE . Furthermore, since the blue and red parts of Γ1 and Γ2 will contain the same number
of 2 and 3-edges, it follows that fΓ1 , fΓ2 ∈ Bh .
Case 1: The edge e1 = e2 ∪ {v3 } = {v1 , v2 , v3 } for some v3 ∈ V (E).
Since v3 ∈
/ e2 and |Eblue | = |Ered |, there must be a 2-edge e3 ∈ Ered such that v3 ∈
/ e3 in order
for (*) to hold. Let e3 = {v4 , v5 } and e4 = {v3 , v4 , v5 }. The sets S, Γ1 and Γ2 in this case are:
S = {e4 }
Γ1 = (Eblue − {e2 }) tm ((Ered − {e1 , e3 }) t {e4 })
Γ2 = {e2 , e4 } tm {e1 , e3 }.
!$#
!%#
!"#
!"#
!$##
!%#
!&#
!&#
Figure 5. Case 1
Figure 6. Case 3
Case 2: The edge e1 = {v1 , v3 , v4 } for some v3 , v4 ∈ V (E) and there is a 2-edge e3 ∈ Ered such
that e3 = {v2 , v3 }.
16
GROSS AND PETROVIĆ
Since v3 ∈
/ e2 , degblue (v3 ; E) = degred (v3 ; E) ≤ n − 1 and, thus, there exists a 2-edge e4 ∈ Ered
such that v3 ∈
/ e4 . Let e4 = {v5 , v6 }.
Now let e5 = {v3 , v4 , v5 } and e6 = {v3 , v6 }. The sets S, Γ1 and Γ2 in this case are:
S = {e5 , e6 }
Γ1 = (Eblue − {e2 }) tm ((Ered − {e1 , e3 , e4 }) t {e5 , e6 })
Γ2 = {e2 , e5 , e6 } tm {e1 , e3 , e4 }.
Case 3: There is a 2-edge e3 ∈ Ered such that v2 ∈ e3 and e2 ∩ e3 = ∅. In this case, let
e4 = (e1 − {v1 }) ∪ (e3 − {v2 }). The sets S, Γ1 and Γ2 in this case are:
S = {e4 }
Γ1 = (Eblue − {e2 }) tm ((Ered − {e1 , e3 }) t {e4 })
Γ2 = {e2 , e4 } tm {e1 , e3 }.
Acknowledgements
The authors would like to thank the anonymous referee for carefully reading the previous
version of this manuscript and thus allowing us to greatly improve Section 5. We are also
grateful to Despina Stasi and Seth Sullivant for helpful discussions.
References
[1] A. Takemura and S. Aoki. Some characterizations of minimal Markov basis for sampling from discrete
conditional distributions, Ann. Inst. Statist. Math. 56 1 (2004), 117.
[2] S. Aoki, A. Takemura, R. Yoshida. Indispensable monomials of toric ideals and Markov bases, Journal
of Symbolic Computation 43 67 (2008) 490-507.
[3] H. Charalambous, A. Katsabekis, and A. Thoma. Minimal systems of binomial generators and the
indispensable complex of a toric ideal, Proceedings of the American Mathematical Society 135 (2007)
3443-3451.
[4] J. De Loera and S. Onn, Markov bases of three-way tables are arbitrarily complicated, J. Symb. Comput.
41 2 (February 2006) 173-18.
[5] M. Develin and S. Sullivant. Markov bases of binary graph models, Annals of Combinatorics 7 (2003)
441-466.
[6] P. Diaconis and B. Sturmfels. Algebraic algorithms for sampling from conditional distributions, Ann.
Statist. 26, no. 1, 363–397 (1998)
[7] A. Dobra and S. Sullivant. A divide-and-conquer algorithm for generating Markov bases of multi-way
tables. Computational Statistics 19 (2004), 347-366
[8] M. Drton, B. Sturmfels and S. Sullivant. Lectures on algebraic statistics, Oberwolfach Seminars 39,
Birkhäuser (2009)
[9] I. Gitler, E. Reyes, R. Villarreal. Ring graphs and toric ideals, Electronic Notes in Discrete Mathematics
28 1 (2007) 393-400.
[10] S. Hoşten and S. Sullivant. A finiteness theorem for Markov bases of hierarchical models, J. Comb.
Theory Ser. A 114 2 (2007) 311-321.
[11] L. Oeding. Set-theoretic defining equations of the tangential variety of the Segre variety, J. Pure and
Applied Algebra, 215 (2011) 1516-1527.
[12] H. Ohsugi and T. Hibi. Toric ideals generated by quadratic binomials, Journal of Algebra 218 (1999),
509-527.
[13] H. Ohsugi and T. Hibi. Indispensable binomials of finite graphs, J. Algebra Appl. 4 (2005), no 4, 421-434.
COMBINATORIAL DEGREE BOUND FOR TORIC IDEALS OF HYPERGRAPHS
17
[14] S. Petrović and D. Stasi. Toric algebra of hypergraphs. Preprint: arXiv:1206.1904
[15] E. Reyes, C. Tatakis, A. Thoma. Minimal generators of toric ideals of graphs, Adv. in Appl. Math 48
(2012), no. 1, 64-67
[16] B. Sturmfels. Gröbner bases and convex polytopes, University Lecture Series, 8. American Mathematical
Society, 1996.
[17] B. Sturmfels and S. Sullivant. Toric ideals of phylogenetic invariants, Journal of Computational Biology
12 (2005) 204-228.
[18] B. Stumfels and P. Zwiernik. Binary cumulant varieties, Annals of Combinatorics, to appear.
[19] C. Tatakis andA. Thoma. On the universal Gröbner basis of toric ideals of graphs, Journal of Combinatorial Theory, Series A, 118 (2011) 1540-1548
[20] R. Villarreal. Rees algebras of edge ideals, Communications in Algebra, 23 (9), 3513–3524 (1995)
[21] R. Villarreal, Monomial Algebras. Monographs and Textbooks in Pure and Applied Mathematics 238,
Marcel Dekker, Inc., New York, 2001.
| 0math.AC
|
3D Object Discovery and Modeling Using Single RGB-D Images Containing
Multiple Object Instances
Wim Abbeloos1 , Esra Ataer-Cansizoglu2∗, Sergio Caccamo3 , Yuichi Taguchi2∗ , Yukiyasu Domae4
arXiv:1710.06231v1 [cs.CV] 17 Oct 2017
1 KU
Leuven, Belgium
Electric Research Labs, Cambridge, MA, USA
3 KTH Royal Institute of Technology, Stockholm, Sweden
4 Mitsubishi Electric Corporation, Japan
2 Mitsubishi
Abstract
Unsupervised object modeling is important in robotics,
especially for handling a large set of objects. We present a
method for unsupervised 3D object discovery, reconstruction, and localization that exploits multiple instances of
an identical object contained in a single RGB-D image.
The proposed method does not rely on segmentation, scene
knowledge, or user input, and thus is easily scalable. Our
method aims to find recurrent patterns in a single RGB-D
image by utilizing appearance and geometry of the salient
regions. We extract keypoints and match them in pairs based
on their descriptors. We then generate triplets of the keypoints matching with each other using several geometric
criteria to minimize false matches. The relative poses of the
matched triplets are computed and clustered to discover sets
of triplet pairs with similar relative poses. Triplets belonging to the same set are likely to belong to the same object
and are used to construct an initial object model. Detection
of remaining instances with the initial object model using
RANSAC allows to further expand and refine the model. The
automatically generated object models are both compact
and descriptive. We show quantitative and qualitative results on RGB-D images with various objects including some
from the Amazon Picking Challenge. We also demonstrate
the use of our method in an object picking scenario with a
robotic arm.
1. Introduction
Object model generation is crucial for robotic manipulation. Typical object detection and localization methods
have a separate supervised stage where they learn and build
object models. However, the types of objects a robot needs
∗ Corresponding
authors: {cansizoglu,taguchi}@merl.com
Figure 1: Given a single RGB-D image containing multiple instances of the same object (top-left), our method automatically
discovers the object and localizes the multiple instances by grouping a set of features (top-right). A 3D model of the object is also
recovered by registering the features from the multiple instances
into a single coordinate system (bottom). The registered features
are denoted as red dots, overlaid on the colored 3D point cloud
of the scene. Note that some of the features appear on the missing face of this specific object instance, indicating that they are
recovered from some other object instances.
to interact with can expand and change rapidly, such as new
items arriving at a warehouse as seen in the scope of Amazon Picking Challenge [1]. On the other hand, in many situations, objects appear in multiple copies. This paper exploits this fact and presents a method for discovering and
modeling an object from a single RGB-D frame in which
the object appears in multiple copies. The recurrent patterns found in the single frame can be used to automatically
discover the object, and the various viewpoints of different
instances can provide valuable information for object model
generation.
Our only assumption is the existence of at least two instances of an object in the single RGB-D image. We do
not use any prior knowledge about the number, shape, and
appearance of the object. Thus, the object can appear in a
cluttered scene or the image can contain multiple instances
of different objects. Our method performs on-the-fly object model generation, while detecting and localizing the instances of the reconstructed object in the given image. Thus,
it enables online robot manipulation using only a singleshot image.
Our technique employs a sparse feature representation,
as shown in Figure 1. Therefore, the problem can be seen
as finding groups of features that correspond to different
instances of the object. To solve this grouping problem we
make use of the following information:
1. Appearance similarity: Pairs of features that come
from the same location of two instances should be similar.
2. Geometric similarity: Two groups of features corresponding to each other based on appearance similarity
should have the same in-group geometric distribution.
In other words, there exists a single transformation that
would transfer and align the positions of features in
one group to the positions of corresponding features in
the other group.
We employ the appearance and geometric constraints
jointly. Furthermore, we avoid the use of depth segmentation and spatial closeness to decide whether features are
coming from the same instance, as the objects might be
touching with each other or occluding one another.
We look for recurrent patterns in the image using both
geometric and appearance similarity following the sparse
feature representation. First, we extract keypoints and
match them based on the descriptor similarity. We then find
triplets of keypoints matching with each other using several
geometric criteria, which are defined for pairs and triplets
of the matched keypoints and are invariant to the 6-degreeof-freedom (6-DOF) transformations. Each of the matched
triplets provides a 6-DOF transformation, which is a candidate of the relative pose between two instances of the object,
but might be an outlier. Thus, in the second stage we cluster the relative poses associated with each triplet match and
find clusters supported by many triplets. The matches that
appear in the same cluster are likely to belong to the same
pair of objects. Thus, in the third stage we generate an initial
model based on the clustering results. Lastly, the generated
model is used in a RANSAC framework in order to detect
additional instances among the remaining keypoints, which
can yield further expansion and enrichment of the generated
model. Once a model has been discovered, it can be used to
detect the corresponding object even if only one instance is
present.
1.1. Contributions
The main contributions of this paper are as follows:
• We present a method for unsupervised object discovery and modeling from a single RGB-D image containing multiple instances of the same object.
• We propose an efficient grouping algorithm that generates a set of relative pose candidates using triplets of
keypoint matches and then clusters them to find each
instance of the object and their relative poses.
• We show experimental results using several objects
and demonstrate an application of our method to object picking.
1.2. Related Work
Object discovery has been investigated using a variety of
approaches. Some are based on geometric and/or color segmentation [25][16], which typically rely on strong assumptions of the scene or the objects (e.g., the objects are placed
on a table) and do not exploit multiple instances. Another
segmentation-based approach based on shape analysis using
compactness, symmetry, smoothness, and local and global
convexity of segments and their recurrence is proposed in
[19]. Since these methods suffer from over and under segmentation, especially in scenes with a lot of clutter, they
are not suitable solutions to our problem. Other methods
gather information over time (thus they are not single-shot
approaches) [11][23] and some assume that objects will be
displaced or removed [18][24][4].
A closely related field is the detection of repetitive patterns [20][26][29] in images. These methods, however, depend on the organized appearance of the structure elements,
while in our case, object instances may appear in random
poses. Other related problems are co-segmentation [9] and
unsupervised detection of object categories [27].
Unsupervised detection of multiple instances of objects
in RGB images has been studied in [15][6][7][21]. These
methods also use geometric and appearance information,
which is then used for clustering or combinatorial optimization. They operate either on matched points, or matched
pairs of points, while our method uses matched triplets.
Moreover, they use only RGB information and not depth,
and do not reconstruct a 3D model or estimate 6-DOF pose.
We compare our method with [6] in experiments.
Some object detection methods for robotics applications
have been proposed that take into account multiple object
instances. A sparse 3D object model created by using structure from motion, which requires multiple frames, is used in
[10]. This enables the detection and pose estimation of multiple object instances in an RGB image. Similarly, in [17], a
sparse 3D model is first manually created from multiple images, after which they detect the model using a stereo cam-
n1
n2
f4
f2
m1
f3
f1 = ||d||
= ||m2 - m1||
m2
Figure 2: Two surface points mi and their normals ni determine a
point pair feature.
era system. Note that these systems require one to build a
model first, while ours does not.
2. Method
The goal of this work is to discover, model, and localize
an object in a scene without any prior knowledge. The input
is a single RGB-D frame, including a color (or grayscale)
image and a depth map of the scene. We use sparse 3D feature points throughout our pipeline, and thus ignore pixels
that have invalid depth measurements.
Our method consists of four main steps. In the first step,
we extract keypoints and generate triplet matches based on
the descriptor similarity and several geometric criteria that
are invariant to the 6-DOF transformations. Second, we
cluster triplet matches based on their relative poses as we
expect to see geometric similarity among groups of features. Third, we generate an initial model using the clustering results. At the fourth step, the initial model is used
to detect additional object instances in the remaining set of
features that have been considered outliers in the clustering
step, which can further enhance the object model. Each of
the four steps is detailed in the following subsections.
2.1. Matching Triplets of Keypoints
In the first step, our goal is to generate triplets of keypoint matches, each of which provides a candidate of the
relative pose between two instances of the object. We use
SIFT [22] to detect and describe keypoints. This results in
a set of N keypoints that have valid depth measurements.
Every keypoint in this set is compared to all others to find
its most similar keypoint. The similarity measure used is
the Euclidean distance between the 128 dimensional feature
descriptors. We also threshold the Euclidean distance such
that we maintain M ≤ N keypoint matches for the following
processes.
Based on appearance similarity, we expect that two instances of an object have similar keypoints. However, the
single keypoint matches are not robust enough, include
many outliers, and do not provide the relative pose between
the two instances. Thus triplets of keypoint matches are
used to be robust to outliers and to obtain the relative pose
using three 3D point registration [28]. For each combination of three keypoint matches, we need to consider cases
where matches are reversed except for symmetric cases (we
thus
keep
4 out of 8 candidates). This results in a total of
M
4
possible triplets. We try to select
= 2M(M−1)(M−2)
3
3
correct triplets based on the following geometric criteria invariant to the 6-DOF transformations:
• Point pair feature similarity: Point pair features
[13][8][3] describe the relative position and orientation of points on the surface of an object (Figure 2).
For two points m1 and m2 with normals n1 and n2 ,
with d = m2 − m1 the feature F is
F(m1 , m2 ) = (kdk, ∠(n1 , d), ∠(n2 , d), ∠(n1 , n2 )),
(1)
where ∠(a, b) ∈ [0 π] denotes the angle between
two vectors. Let l1 and l2 be keypoints matching
with m1 and m2 respectively. We compute the difference of point pair features between the matches
as F(m1 , m2 ) − F(l1 , l2 ) and apply distance and angle
thresholds to the calculated difference to filter out incorrect correspondences.
• Sidedness: We check whether the third point of the
triplet falls on the same side of the line defined by
the other two points to avoid reflections [15]. For the
triplet P, let us denote the cross product of the edges di
and d j as vi, j (P) = di × d j . We discard a triplet match
if any of the two corresponding v vectors are in opposite directions, i.e., discard the triplet match between P
and Q, if ∃i, j such that
vi, j (P)
kvi, j (P)k
+
vi, j (Q)
kvi, j (Q)k
< ε.
• Minimum triangle edge length and maximum acuteness: To ensure the found corresponding triangles will
yield sufficiently accurate transformation estimations,
triangles generated with closely located keypoints are
removed. This is done using a minimum triangle edge
length and maximum angle acuteness threshold.
• Overlapping triangles: We omit the triplet match if
the two triangles are overlapping, as they would more
likely be coming from the same instance.
Since the point pair feature similarity can be computed for
pairs of keypoint matches, we first use this criterion for efficient pruning of incorrect pairs and then use the other criteria for selecting correct triplets. Also, in order to get a
balanced distribution of keypoints among triplet matches, a
keypoint can appear at most L times in the generated triplet
matches. The thresholds and parameters used in this study
are given in Section 3.
2.2. Clustering
5.1
For each of the triplets obtained in the first step, a 6-DOF
pose that transforms the triangle to its corresponding triangle is estimated. Let P = (p1 , p2 , p3 ) and Q = (q1 , q2 , q3 )
denote two matching triangles where pi , qi ∈ R3 are 3D positions of the keypoints. The calculation of the pose results
in the transformation Tp,q ∈ SE(3) that consists of a rotation
matrix R ∈ SO(3) and a translation vector t ∈ R3 such that
qi = Tp,q (pi ) = Rpi + t. These transformations are clustered using the DBScan [14] algorithm to discover sets of
triplets with similar transformations. DBScan is a density
based clustering method, which only requires a single input
parameter for the maximum distance between two instances
that are allowed to be clustered together. During clustering,
we exploit sum of 3D point-to-point distances as the distance between two triplets. For symmetry, the distance is
computed bidirectionally. Thus, the distance between two
matching triplets (P, Q) and (A, B) based on the respective
transformations Tp,q and Ta,b is
D((P, Q), (A, B)) = ∑ kTp,q (ai ) − bi k + ∑ kTa,b (pi ) − qi k.
i
0
2
4
7.3
3
8
6
14.4
13.2
5
7
10
1
9
5.9
4.2
11
Figure 3: An overlay shows the graph representing clustered sets
of points (vertices) and their relations (edges). Note there are two
types of relations between sets of points: sets that are clustered
together because they contain similar triplets, and sets that are
connected because they have points in common. The edges representing matched triangles have a label showing the distance (the
distance error obtained with the transformation between them).
i
(2)
The output of clustering can contain the same pair of instances in two different clusters with associated poses as
inverse of each other. Hence, if such clusters are found, one
of them is inverted and the clusters are merged. The transformation for each cluster is then recalculated considering
all sets of corresponding triplets in the cluster.
2.3. Initial Model Creation
The clustering procedure results in sets of points that belong to the same object instance and are matched to another
object instance. In other words, each cluster can be seen as
two sets of points, where one set can be aligned with the
other set using the transformation of the cluster. Some of
these sets may have keypoints in common with other sets.
Thus, the clustering result can be represented as a graph
where nodes correspond to sets of points and edges correspond to the distance between sets based on the transformation of the cluster associating the two sets (Figure 3). If
two sets have points in common, then the transformation
between them is identity and the connecting edge is set to
have a small preset weight δ .
The resulting graph can have multiple connected components, since the scene can contain multiple instances of
various types of objects. In order to create a model for each
connected component, we first decide which node will be
the reference frame where all sets will be transformed to.
We pick the node representing the set of points with the
largest number of matches and common points as the reference. All other sets of points that are connected to it are
transformed to the reference frame by applying a series of
transformations. The optimal series of transformations for
every set is found by searching for the shortest path to the
reference frame using Dijkstra’s algorithm [12].
The 3D object model consists of all points transformed
to this common reference frame, and associated with their
original keypoint descriptors. This process generates an
object model for each connected component in the graph,
hence it might yield multiple models, each containing
points from all sets connected to their reference set. For
each generated model, we apply a final bundle adjustment
to refine landmark locations and instance poses.
2.4. Additional Instance Detection
After creating a set of object models, every model is
compared to all others to verify whether they truly are distinct objects, or whether their correspondence was simply missed by the earlier steps (this is possible because
we enforce a unique match between keypoints in our first
step, instead of considering all possible matches). For each
model, we perform detection between the model and the
sets of points from the other connected components of the
graph. This is performed by a correspondence search via
descriptor similarity and a geometric verification by a 3point RANSAC registration.
We also try to detect any remaining instances that had
not been matched before. We use the remaining keypoints
that are not associated with any of the nodes in the graph
to avoid matching the model to the previously detected instances. In the RANSAC registration, we sample three
scene points so that they are within the diameter of the
Sez
Sd
Mez
Md
# img (pair)
54
57
49
30
190
# img (> 2)
29
0
3
12
44
# img (total)
83
57
52
42
234
Table 1: The number of images containing only pairs of instances,
the number of images containing more than two instances, and the
total number of images, per scenario.
model.
In both cases, the RANSAC estimates an initial transformation using the three points and counts the number of
inliers (the percentage of matched points that, when transformed, are within a certain distance of their corresponding
points). RANSAC succeeds if the inlier ratio is larger than
a certain threshold. The transformation is then re-estimated
based on the inliers of the most successful attempt. In the
case of a successful RANSAC, the model is merged with
the other model or the points selected as inliers from the
remaining keypoints.
3. Experiments and Results
An ASUS Xtion Pro Live RGB-D camera was used to
acquire a dataset of 234 VGA (640 × 480) resolution color
and depth images. The depth image was converted to a 3D
point cloud and transformed to the RGB camera’s reference
frame. This means every valid measurement point has both
a 3D coordinate and a color value.
We classified the captured scenes into four different scenarios: scenes containing either a single or multiple object
types (denoted by S and M), and scenes with or without
clutter and occlusion (denoted by easy ez and difficult d ).
Each scene contained two to nine instances per object type.
We used objects with various shapes and sizes and varying amounts of texture. The number of images per scenario is given in Table 1. Qualitative and quantitative results
are given for the different scenarios. The dataset is available at ftp://ftp.merl.com/pub/cansizoglu/
ObjectDiscovery3DV2017.zip, and includes object
annotations.
Our method is compared to an RGB object discovery algorithm [6] that uses feature matching with a novel pairwise
dissimilarity measure and hierarchical agglomerative clustering to find pairwise matches between sets of points. The
dissimilarity measure they propose consists of a photometric and a geometric term. The photometric term is simply
the Euclidean distance between the points’ SIFT descriptor
vectors, which is also used in our method. The geometric
term is used to determine a pairwise dissimilarity between
two corresponding pairs. It uses the homography of the first
Figure 4: Robot arm used for object picking.
correspondence to transform the points of another and vice
versa. The final geometric term is the average of the transformation errors. The total pairwise dissimilarity is a linear
combination of both terms. We used the source code available on the authors’ website with the default parameters, as
changing them did not improve the results.
The following parameters were used in these experiments to eliminate incorrectly matched triplets: a 5mm and
35 degrees threshold for the point pair feature difference.
Each edge of the triangle should be at least 10mm and at
most 125mm and each angle should exceed 10 degrees.
Maximum value of the distance between two samples in
clustering was set to 35mm, while we discarded clusters
with less than 14 samples. In detection, we used a RANSAC
inlier threshold of 5mm. A RANSAC was recalled as successful when there were at least 5 inliers and the inlier ratio was more than 12.5%. The average running time was
809ms with a C++ implementation on a standard PC.
We also demonstrate the use of our algorithm in an object picking scenario with a robotic arm, where multiple instances of the same object are visible (Please see supplementary video). We mounted an ASUS Xtion sensor on the
robot arm and picked up objects using a vacuum gripper as
seen in Figure 4.
3.1. Qualitative Results
Some results on the proposed dataset on the four different scenarios are shown in Figure 5. For our method the image is overlaid with the transformed object model (with one
color per object type). For the comparison method, clusters of matching features are shown (each cluster having a
with clutter/occlusion
multiple objects
single object
without clutter/occlusion
Figure 5: Some of the scenes from the proposed dataset from the four scenarios: single/multiple objects (top/bottom) and with/without
clutter and occlusion (left/right). Overlaid is a visualization of the results for our method (first and third columns) and the results for [6]
(second and fourth columns). The quantitative results are summarized in Table 2.
(a) Because our algorithm searches for
the largest recurrent pattern, object instances occurring in an organized way
are merged into one object, resulting in
object models containing two instances.
Each object instance is shown in a random color.
(b) While a correct object model was created for the object, and all instances were
found correctly, a separate partial model
was discovered in two instances (denoted
by green arrows pointing to green model
points). The second model is considered
a false positive, since these points should
have been a part of the first model.
(c) The object model in green was successfully discovered, and its instances located, despite the high degree of occlusion by other objects.
Figure 6: A few remarkable results with additional comments.
different color). This figure contains only scenes with two
instances per object type to allow direct comparison to the
other method. The result of our method on a scene with
a larger number instances is shown in Figure 1 and more
examples are shown in the supplementary material.
Experiments on scenes with objects placed in an organized way gave the results seen in Figure 6a. Here, three
ducks were placed side by side and our algorithm ended
up with a model of two repetitive patterns representing two
neighboring ducks. This was expected as the clustering
stage selects the largest cluster to start building the initial
model. Note that the filter eliminating overlapping triangles
was turned off to create this example.
3.2. Quantitative Results
We report object discovery performance on the generated dataset in Table 2. Objects were counted as true positives if the discovered model was correct, and it was localized correctly. We considered a detection correct when at
least 90% of the features were inside the annotated bounding box. They were counted as a false negative if the object
was not detected. False positives occur when an incorrect
model is found, or when a model was found in an incorrect
place. Since [6] only discovers pairwise matches, we compared the performance for the subset of scenes containing
only pairs of objects (two object instances per object type).
Our method has an F1-score of 0.966 on this subset, while
the comparison method only reaches 0.471. Our method
finds any number of instances and thus can be evaluated on
the entire dataset by counting the number of instances (not
pairs) found. For this evaluation our F1-score is 0.974.
Our method gives very few false positives resulting in a
high precision. A false positive example is shown in Fig-
ure 6b. There is a few false negatives in our method (e.g.,
the brushes in the scene in row 5, column 1 of Figure 5).
These false negatives are caused by the objects having relatively few keypoints, and many of the keypoints being
matched incorrectly. These incorrect matches result from
the descriptor not being sufficiently invariant to large viewpoint changes. The recall does not differ much for the scenarios with or without occlusion and clutter, but it is slightly
lower for the scenario with multiple object types.
An interesting result of the comparison method is that
they have the worst performance for the easiest scenario Sez
with only one object type without clutter or occlusion (e.g.,
the scenes in rows 1, 2 and 3, column 2 of Figure 5). In
this scenario, they have a large number of false positives.
As there are far fewer keypoints, it is more likely that false
matches accidentally form a cluster. If there is more clutter,
these false matches are more likely to be more random and
are less likely to cause false positive clusters.
4. Conclusion and Discussion
We presented a novel method for 3D discovery, modeling, and localization of multiple instances of an object using a single RGB-D image. Following a sparse feature representation, we employ appearance similarity and geometric similarity to group features associated to the instances.
Our grouping algorithm is efficient as it considers triplet
matches and eliminates incorrect correspondences between
triplets based on various geometric constraints. The 6-DOF
poses calculated for each triplet match are clustered in order
to find matching object instances. The initial model generated using the clustering results can then be used to detect remaining object instances in the scene. The proposed
Sez
Sd
Mez
Md
Total
F1
Objects (our)
P
R
0.992 0.971
1.000 0.965
1.000 0.919
0.989 0.907
0.995 0.938
0.974
Pairs (our)
P
R
1.000 1.000
1.000 0.965
1.000 0.931
1.000 0.925
1.000 0.949
0.966
Pairs [6]
P
R
0.338 0.329
0.621 0.506
0.607 0.508
0.629 0.429
0.504 0.441
0.471
Table 2: Precision (P) and recall (R) for our method and [6] on the
different datasets. S/M indicates whether the dataset has a Single
(S) or Multiple (M) object models. The easy/difficult scenario is
indicated with ez or d . We give results for P-R calculated on the
found object instances (first column) and for P-R calculated on
pairs of objects (second and third columns).
method provides descriptive and compact object models using only a single RGB-D image and is suitable for robotic
manipulation tasks. Another application of our framework
can be seen in [2], where we generate object proposals for a
deep-learning-based classification method. Since our technique outputs regions with recurrent patterns, we further
improve classification accuracy by considering joint probability of bounding boxes that refer to copies of the same
object.
The goal of this study is to detect and discover objects
from a single image that contains multiple object instances.
Therefore it differs from other methods that aim to detect
and track objects in a given sequence of frames such as object SLAM techniques [5]. Based on a sparse feature representation, the problem can be formulated as a challenging
search problem, where we search for subsets of features that
resemble each other in terms of appearance and geometry.
More specifically, all pairwise feature matches need to be
considered in this problem, as opposed to investigating feature matches only between frames in an object tracking scenario given a sequence of frames. Moreover, the instances
can occur in various viewpoints in a single image making
the search problem more challenging, while a smooth motion is usually observed in a video-based tracking scenario.
Our algorithm finds the largest recurrent pattern in the
scene. This can be an important limitation especially in two
cases (i) when the objects are placed in an organized way
as in Figure 6a, and (ii) when there are pairs of instances
that have the same relative pose. The limitation in the first
case exists due to the fact that we avoided any assumptions
about the placement of the objects as opposed to existing
work on finding periodic patterns in a scene. This makes
our algorithm applicable to more general scenarios, i.e., occlusion, random pose, etc. Once a cluster is detected it is
always possible to search for periodic patterns in order to
handle cases with organized objects. We can also solve this
problem by recursively calling the algorithm on the set of
points from each cluster. On the other hand, if an object has
repeated patterns on its surface, then following such a recursive splitting strategy would result in models smaller than
the object. Consequently, such scenarios might be solved
with more prior knowledge such as the size of the object or
the number of object instances in the scene. The limitation
in the second case can be solved by some assumptions about
the placement of the objects. For example, a smoothness in
depth can be expected among the features of an instance.
Again, we avoided using any depth-based segmentation as
they are shown to perform poorly on various cases.
Keypoint detection and descriptor matching lie at the
core of the proposed technique as a way of measuring appearance similarity. Missing features in some instances due
to nonrobust keypoint detection and poor matching in large
viewpoint changes were among the problems we faced because of limitations of conventional feature detectors and
descriptors. As a result, the performance degraded especially when there are too many false matches that need to be
handled as observed in the case of cluttered scene or multiple object types in a single image. Therefore, improving
descriptors and descriptor matching is an important extension of this work.
Acknowledgments
This work was done at and supported by Mitsubishi
Electric Research Laboratories. We thank the anonymous
reviewers for their helpful comments. W.A. thanks EAVISE, KU Leuven for the travel support.
References
[1] Amazon
picking
challenge.
http://
amazonpickingchallenge.org/. 1
[2] W. Abbeloos, S. Caccamo, E. Ataer-Cansizoglu, Y. Taguchi,
C. Feng, and T.-Y. Lee. Detecting and grouping identical
objects for region proposal and classification. In Workshop
Deep Learning for Robotic Vision Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), pages 501–
502, 2017. 8
[3] W. Abbeloos and T. Goedemé. Point pair feature based object detection for random bin picking. In Conference on
Computer and Robot Vision (CRV), pages 432–439, June
2016. 3
[4] S. Caccamo, E. Ataer-Cansizoglu, and Y. Taguchi. Joint 3D
reconstruction of a static scene and moving objects. In Proc.
Int’l Conf. 3D Vision (3DV), 2017. 2
[5] C. Cadena, L. Carlone, H. Carrillo, Y. Latif, D. Scaramuzza,
J. Neira, I. Reid, and J. J. Leonard. Past, present, and future of simultaneous localization and mapping: Toward the
robust-perception age. IEEE Trans. Robotics, 32(6):1309–
1332, Dec. 2016. 8
[6] M. Cho, J. Lee, and K. M. Lee. Feature correspondence
and deformable object matching via agglomerative corre-
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
spondence clustering. In Proc. IEEE Int’l Conf. Computer
Vision (ICCV), pages 1280–1287, 2009. 2, 5, 6, 7, 8
M. Cho, Y. M. Shin, and K. M. Lee. Unsupervised detection
and segmentation of identical objects. In Proc. IEEE Conf.
Computer Vision and Pattern Recognition (CVPR), pages
1617–1624, 2010. 2
C. Choi, Y. Taguchi, O. Tuzel, M.-Y. Liu, and S. Ramalingam. Voting-based pose estimation for robotic assembly
using a 3D sensor. In Proc. IEEE Int’l Conf. Robotics and
Automation (ICRA), pages 1724–1731, May 2012. 3
W.-S. Chu, C.-P. Chen, and C.-S. Chen.
MOMIcosegmentation: Simultaneous segmentation of multiple objects among multiple images. In Proc. Asian Conf. Computer
Vision (ACCV), pages 355–368, 2010. 2
A. Collet, D. Berenson, S. S. Srinivasa, and D. Ferguson.
Object recognition and full pose registration from a single
image for robotic manipulation. In Proc. IEEE Int’l Conf.
Robotics and Automation (ICRA), pages 48–55, 2009. 2
T. Dharmasiri, V. Lui, and T. Drummond. MO-SLAM: Multi
object slam with run-time object discovery through duplicates. In Proc. IEEE/RSJ Int’l Conf. Intelligent Robots and
Systems (IROS), pages 1214–1221, 2016. 2
E. W. Dijkstra. A note on two problems in connexion with
graphs. Numerische mathematik, 1(1):269–271, 1959. 4
B. Drost, M. Ulrich, N. Navab, and S. Ilic. Model globally,
match locally: Efficient and robust 3D object recognition. In
Proc. IEEE Conf. Computer Vision and Pattern Recognition
(CVPR), pages 998–1005, June 2010. 3
M. Ester, H.-P. Kriegel, J. Sander, X. Xu, et al. A densitybased algorithm for discovering clusters in large spatial
databases with noise. In International Conference on Knowledge Discovery and Data Mining (KDD), volume 96, pages
226–231, 1996. 4
V. Ferrari, T. Tuytelaars, and L. Van Gool. Simultaneous
object recognition and segmentation from single or multiple
model views. Int’l J. Computer Vision, 67(2):159–188, 2006.
2, 3
M. Firman, D. Thomas, S. Julier, and A. Sugimoto. Learning to discover objects in RGB-D images using correlation
clustering. In Proc. IEEE/RSJ Int’l Conf. Intelligent Robots
and Systems (IROS), pages 1107–1112, 2013. 2
T. Grundmann, R. Eidenberger, M. Schneider, M. Fiegert,
and G. v. Wichert. Robust high precision 6D pose determination in complex environments for robotic manipulation. In
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
Workshop Best Practice in 3D Perception and Modeling for
Mobile Manipulation at the Proc. IEEE Int’l Conf. Robotics
and Automation (ICRA), pages 1–6, 2010. 2
E. Herbst, X. Ren, and D. Fox. RGB-D object discovery via
multi-scene analysis. In Proc. IEEE/RSJ Int’l Conf. Intelligent Robots and Systems (IROS), pages 4850–4856, 2011.
2
A. Karpathy, S. Miller, and L. Fei-Fei. Object discovery
in 3D scenes via shape analysis. In Proc. IEEE Int’l Conf.
Robotics and Automation (ICRA), pages 2088–2095, 2013. 2
T. Leung and J. Malik. Detecting, localizing and grouping
repeated scene elements from an image. In Proc. European
Conf. Computer Vision (ECCV), pages 546–555, 1996. 2
J. Liu and Y. Liu. Grasp recurring patterns from a single
view. In Proc. IEEE Conf. Computer Vision and Pattern
Recognition (CVPR), pages 2003–2010, 2013. 2
D. G. Lowe. Distinctive image features from scale-invariant
keypoints. Int’l J. Computer Vision, 60(2):91–110, 2004. 3
L. Ma and G. Sibley. Unsupervised dense object discovery,
detection, tracking and reconstruction. In Proc. European
Conf. Computer Vision (ECCV), pages 80–95, 2014. 2
J. Mason, B. Marthi, and R. Parr. Object disappearance for
object discovery. In Proc. IEEE/RSJ Int’l Conf. Intelligent
Robots and Systems (IROS), pages 2836–2843, 2012. 2
J. Papon, A. Abramov, M. Schoeler, and F. Wörgötter. Voxel
cloud connectivity segmentation - supervoxels for point
clouds. In Proc. IEEE Conf. Computer Vision and Pattern
Recognition (CVPR), pages 2027–2034, June 22-27 2013. 2
M. Pauly, N. J. Mitra, J. Wallner, H. Pottmann, and L. J.
Guibas. Discovering structural regularity in 3D geometry. In
ACM Trans. Graphics, volume 27, page 43, 2008. 2
T. Tuytelaars, C. H. Lampert, M. B. Blaschko, and W. Buntine. Unsupervised object discovery: A comparison. Int’l J.
Computer Vision, 88(2):284–302, 2010. 2
S. Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Trans. Pattern
Anal. Mach. Intell., 13(4):376–380, Apr. 1991. 3
C. Wu, J.-M. Frahm, and M. Pollefeys. Detecting large repetitive structures with salient boundaries. In Proc. European
Conf. Computer Vision (ECCV), pages 142–155, 2010. 2
| 1cs.CV
|
Designing a Safe Autonomous Artificial Intelligence Agent based on Human Self-Regulation
Mark Muraven
University at Albany
January 5, 2017
Author contact: [email protected]
Human Self-Regulation for AI
2
Abstract
There is a growing focus on how to design safe artificial intelligent (AI) agents. As systems become more
complex, poorly specified goals or control mechanisms may cause AI agents to engage in unwanted and
harmful outcomes. Thus it is necessary to design AI agents that follow initial programming intentions as
the program grows in complexity. How to specify these initial intentions has also been an obstacle to
designing safe AI agents. Finally, there is a need for the AI agent to have redundant safety mechanisms
to ensure that any programming errors do not cascade into major problems. Humans are autonomous
intelligent agents that have avoided these problems and the present manuscript argues that by
understanding human self-regulation and goal setting, we may be better able to design safe AI agents.
Some general principles of human self-regulation are outlined and specific guidance for AI design is
given.
Muraven
3
Recently there has been some discussion on how to better align autonomous artificial
intelligence actions with human interests. The issue, as discussed at length (e.g., Bostrom, 2014), is that
an autonomous artificial intelligence agent, in blindly following its programming, may engage in actions
that directly or indirectly hurt human welfare. Several potential solutions have been put forth, but to
date, none are deemed sufficient or adequate to address the problem. More broadly, there exists the
general question of how to make an autonomous artificial intelligence agent work and what might be
the key features of it. The goal of this manuscript is to argue that by understanding how humans
autonomously reach their own goals, we can better design an autonomous artificial intelligence agent to
both be autonomous and to pursue goals that align with human interests.
The underlying assumption behind this discussion is that humans have evolved and been
socialized in ways that ensure optimal goal pursuit. Obviously, there are instances where humans’
behavior is less than ideal, but as will be argued human goal behavior is a reasonable starting point for
designing a self-regulation system for an autonomous agent. Thus, this manuscript provides a brief and
introductory framework to research on human self-regulation. There are many details and formalisms
omitted for simplicity’s sake. The point is to give a basic understanding of human self-regulation that
may help in the design of an autonomous artificial intelligence agent as well.
In particular, understanding human self-regulation may guide research on artificial intelligence
because humans seem to avoid many of the hypothesized pitfalls that an artificial intelligence may face.
For instance, the obsessive pursuit of one final goal, analogous to the paperclip optimization problem, is
rare in humans. Humans do not single-mindedly pursue one goal to the exclusion of all others.
Foreshadowing some of the dynamics of the model, humans that pursue a goal beyond reasonable
balance are considered mentally ill. Addicts, people suffering from obsession and compulsions, and
those who engage in self-damage suffer from understandable and avoidable problems with their self-
Human Self-Regulation for AI
4
regulation system. Indeed, treatment for those conditions usually addresses how to better self-regulate.
Thus, in humans, a correctly functioning self-regulatory system is needed to maintain mental health.
Moreover, humans’ ability to self-regulate is very robust as most people can adapt to highly dynamic
environments and remain focused on long-term goal pursuit even when unexpected obstacles arise.
These are traits that would be highly valued in an autonomous artificial intelligence agent. As described
below, there are clear features of humans’ goal system that helps to keep it stable and balanced that
may be easily emulated in an artificial intelligence agent.
The other benefit from understanding how the human goal system operates is that most people
engage in behaviors that they believe will benefit themselves and other people. Obviously, there is some
level of self-deception and flexibility in definition (the dictator who puts opponents in jail may convince
himself that it is for the good of society), but theories of human self-regulation suggest that there are
checks and balances on such self-deception. These checks built into the self-regulation system help to
explain why most people follow socially approved goals in socially appropriate ways. Understanding
human self-regulation may therefore help in designing autonomous artificial intelligent agents that
avoid undesired and unconventional solutions to problems (e.g., making everyone happy by putting
wires into the brain’s pleasure centers).
Feedback Loops
Self-regulation is a broad and encompassing topic within modern cognitive science. At its heart,
it describes how individuals select purposive goals, stay on track toward those goals, and correct for
changes in the environment as they move toward their goals. The central idea is that there is a feedback
control process that helps people reach their goals. The basic framework for understanding human selfregulation follows from cybernetic models of feedback control (Carver & Scheier, 1981) that suggests
people compare their current state to some goal, standard or ideal. That is, the feedback system
Muraven
5
compares incoming information to a reference value. If a discrepancy is noted between the current state
and ideal state, some action takes place to try to minimize (or maximize, if the goal is to be unlike
something) that difference. Thus, at its core there are four simple parts to a feedback loop: current state
monitoring, ideal state monitoring, a comparator system to evaluate differences between those two
states, and an output function to reduce the discrepancies between those two states.
The critical feature of cybernetic control requires the agent to have some idea of what actions
may reduce (or enlarge) the discrepancy between the current state and goal state. Learning and training
helps make this self-regulation process more efficient so the agent can quickly decide what actions are
likely to lead to the desired outcome (Taylor, Pham, Rivkin, & Armor, 1998). However, as explained
below, the autonomous agent should be open to feedback and is constantly self-monitoring, so when an
action does not lead to the desired outcome, adjustments can be made.
Complex and autonomous behavior can begin to emerge when multiple conflicting feedback
loops are nested within a hierarchical structure. The critical feature is not the cybernetic feedback loop
itself, but rather the way multiple conflicting loops interact. It is through these chaotic and dynamic
interactions that autonomy emerges. Moreover, these interactions among loops are the critical feature
in maintaining goal focus in changing environments and avoiding actions that would damage themselves
and others.
Basic Features
Hierarchical Goals
One critical feature of the feedback loop is that goals are arranged in an hierarchical structure
(Elliot, 2006). Every goal has an underlying motivation (superordinate goal) and activates sub-goals. For
Human Self-Regulation for AI
6
example, a person who wishes to make a friend (goal) may be doing that to fulfill the need to feel
connected to others (superordinate goal). They might pursue that goal by texting someone they met
(sub-goal). Progress on each of these goals at every level of the hierarchy is monitored and controlled by
the cybernetic feedback loops previously described. Lower level goals are more concrete and discrete,
whereas higher level goals tend to be more abstract.
The top-most goals should be agent-defining and represent the fundamental values of the
autonomous agent, although they are not often brought to conscious awareness. These top level goals
are fundamental and basic. That is, they exist in everyone, cannot be ignored, and are critical to the
successful functioning of the autonomous agent. For humans, these top-level goals are likely the
product of evolution and thus innate (Deci & Ryan, 2000). There may be individual differences in how
strongly each goal is pursued, but they are valued to some extent in every person. Even in cases of
disordered behavior, such as addiction, a mixture of these goals is still active (Graham, Young, Valach, &
Alan Wood, 2008). These top level goals are basic to human nature and ultimately drive all behavior.
For humans, the exact enumeration of these top-level goals is a matter of some controversy, but
most accounts include a need to feel competent, a desire to have a meaningful impact on the world, a
dislike of inconsistencies as well as more basic concerns such as a need for survival and social
connections. For the purpose of an autonomous artificial intelligence agent, an exact and complete
description of these top-level goals may be less important than the fact that there are multiple
conflicting goals that are aligned with human desires.
Given the hierarchical nature of goals, each of these top level goals can be fulfilled in different
and perhaps competing ways. For example, the goal of belongingness or relatedness may lead to the
goal of calling friends more often, becoming more active in a social movement, and showing more
affection to one’s family. These lower level goals also may conflict so that progress on one goal may
Muraven
7
hinder progress on another goal. People’s habitual solutions to these conflicts is an important source of
individual differences.
Conflicting Goals
Given that people have multiple goals active at the same time, these goals may often be in
conflict. The pursuit of one goal may block, interfere or even undo the pursuit of another. For example,
the desire to feel competent may lead to risk taking, which conflicts with the need for survival. The
interfering nature of these goals is a critical feature of a successful autonomous agent and necessary for
proper functioning.
If an autonomous agent has to balance multiple, conflicting goals, there has to be a prioritization
process to resolve these competing desires (Stroebe, Mensink, Aarts, Schut, & Kruglanski, 2008). Given
the wide range of potential goals, the autonomous agent needs some way to narrow the choices. Some
general principles seem to be at work. People often try to find situations that will maximize their
progress to as many goals as possible (multifinality; Förster, Liberman, & Friedman, 2007; Kruglanski et
al., 2013). For example, people like to study in groups, so both social and competency goals can be
simultaneously filled. They also seek goals that may be met in different ways, thus allowing for multiple
paths to the desired endpoint (equifinality; Kruglanski, Pierro, & Sheveland, 2011).
In general, if progress toward a goal is blocked, an autonomous agent typically should try to find
other ways to fulfill that goal. Just because current progress is frustrated, the goal should not be
dropped. Instead, the agent should search through alternative paths to find another way to reach the
goal. As a general principle, the higher the goal is in the hierarchy, the less willing the autonomous agent
should be to drop it. If a call to a friend goes unanswered, the person will likely try another friend before
giving up on fulfilling a desire to fulfill a need for social connection. Even when blocked multiple ways,
Human Self-Regulation for AI
8
people will seek other ways to fulfill the same goal. If no friends answer their phone, a person may start
posting on Facebook to gain some sense of social connection.
More generally, the fields of psychology, economics, decision making, and sociology have put
forth several different but ultimately related theories of motivation and goal selection. Recently, Steel
and König (2006) integrated these theories into their temporal motivation theory (TMT), which is based
on four core features: value, expectancy, time, and loss versus gains. Value is determined by anticipated
satisfaction of goal completion, whereas expectancy is the perceived probability that that outcome will
occur. Temporal discounting results in events closer in time being more influential. Finally, gains and
losses are weighted differently, so that losses hurt more than gains and the same outcome could be
perceived as a gain or loss based on the context. This theory can be integrated with theories that limit
the search space for potential alternatives, like Decision Field Theory (Busemeyer & Diederich, 2002) to
further increase its utility. This model likely provides a strong basis for understanding goal selection by
an autonomous agent.
Goals Reprioritization
Although the TMT model does account for temporal discounting, it is still a relatively static
perspective on goal pursuit. It fails to account for changes over time and interactions between multiple,
conflicting goals. An autonomous agent should dynamically update its goals as the situation changes and
as progress is made. If a person wants to build a shed and unexpectedly runs out of nails, the goal to buy
more nails gets elevated in priority. If the store is closed, another goal may become active, such as paint
the trim or respond to unanswered e-mails. For instance, research has found that goals for examinations
changed as students’ perception of their course grade changed (Lord & Hanges, 1987). Lower-level goals
Muraven
9
are subject to revision as circumstances change, based on feedback from more abstract, higher-level
goals. Thus, goals must be allowed to change as the situation changes.
The dynamics of the self-regulation system must be changeable as well. Some goals should be
hewed to closely and resist change. Other times the system should be more open and flexible. For
instance, in times of exploration, creativity or difficulty, the system may look to internal and external
cues to be very open to new goals. Other times, the autonomous agent should ignore distractions or
other goals in order to maximize pursuit of the currently active goal.
Goal dynamics should depend on both bottom up processing based on the immediate value of
the goal and top down processing, which considers the overall implication of the goal for the
autonomous agent. Bottom-up models, like the multiple-goal pursuit model (MGPM; Ballard, Yeo, Loft,
Vancouver, & Neal, 2016) dynamically update the discrepancy between current and goal state, as well as
the available and required resources needed to reach the goal, which determines the motivation to act
on that goal at any moment in time. This means there should be regular evaluations of progress and
alternate goals. The amount of processing dedicated to this reevaluation should probably depend on the
nature of the goal and its place in the hierarchy of goals. Moreover, a successful autonomous agent
should learn from prior experience and anticipate the effects of its behaviors, so that goals are
dynamically updated based on past actions. On the other hand, top-down models of goal dynamics also
need to be considered (Lynn, Wormwood, Barrett, & Quigley, 2015). Such models assume that decisions
are made within the larger context of the decision maker’s life. Thus prior history and future desires are
weighed in the decision-making process.
Human Self-Regulation for AI
10
Feedback
Implicit in all these models is a need for regular assessment of goal progress. A successful
autonomous agent must get regular feedback on its progress toward its goals. In humans, there is good
evidence to suggest that emotions serve this purpose. Researchers have theorized that people have a
meta-monitoring process that focuses on their rate of progress toward goals. Goals associated with
positive affect are more likely to be prioritized (Custers & Aarts, 2007). Likewise, the more a behavior
matches critical goals, the more positive people feel (Chirkov, Ryan, Kim, & Kaplan, 2003). When
progress is perceived to be adequate, people experience positive affect; inadequate progress results in
negative affect. This emotional feedback process allows people to track their progress toward multiple
goals and hold models of desired progress toward these goals. Based on this feedback, people will
prioritize some goals over others and increase or decrease effort on those goals.
Other self-conscious emotions, such as shame, guilt, and pride may provide additional goal
related feedback (Tangney, 2003). These emotions are powerful guides to behaviors and strongly
motivate self-regulation. Additional emotional information includes intensity and arousal level, based on
the type of goal being pursued and changes in rate in progress. Each level in the hierarchy in goals is
likely separately monitored, although feelings associated with low level goals may be more fleeting than
higher level goals.
Further complicating matters, the feedback from progress may lag behind actual output. That is,
it may take some time for the effects of the behavior to become evident. Thus, an autonomous agent
should be able to reward itself from small actions that help move it toward goals. Such intrinsic
motivation helps to ensure that goals are not abandoned when feedback is long in coming (Deci & Ryan,
1991).
Muraven
11
The criterion for the feedback is likely multi-dimensional, especially for more abstract goals. For
example, to determine whether the autonomous agent is making a positive change in the world
(competence goal), the agent may look to others’ reaction, changes in the environment, and degree of
challenge felt during the activity (Fisher, 1978). There should be some flexibility in the criterion as well—
new situations should lead to looser criteria than familiar situations. Experience should add more
criteria to evaluate success and repeated success or failures should lead to recalibration of the criteria
so the agent’s affective experiences remains relatively stable.
In short, any autonomous agent needs a method to monitor progress toward its short and longterm goals. This feedback monitoring process seems to be largely built into human neurological system.
Because of its fundamental and critical importance, it seems to be universal in mentally healthy people
and resistant to change. Obviously, people may seek ways to modify their emotional state outside of
goal pursuit, but society generally frowns upon long-term modulation of the emotional-feedback system
(e.g., through drug use). Thus, an autonomous artificial intelligence agent will need a clear and dynamic
feedback system that is difficult to ignore, resistant to change and not easily substituted by other
experiences.
Goal Fatigue
Finally, an autonomous agent needs to avoid getting stuck on a goal, even when that goal is
seen as important and progress is being made. As explained above, autonomous agents should have
multiple goals they wish to fulfill at any time but it is unlikely that they can fulfill more than one or two
with any action. Thus, the agent needs some way to balance these conflicting and interfering goals.
Because making progress on one goal often means forestalling progress on another, the
autonomous agent needs to have ways to select one goal and suppress all others. This ensures that the
Human Self-Regulation for AI
12
agent stays on task and does not get distracted by temporary or fleeting desires that might be triggered
by the environment. Once a goal is selected, the agent should make certain that adequate resources are
devoted to the task to secure its completion. Indeed, humans who are poorer at suppressing competing
desires and staying on task suffer numerous problems in school, work and life (Tangney, Baumeister, &
Boone, 2004). Research has shown that non-prioritized goals are shielded from interferences, so that
these unselected goals are suppressed and less likely to be remembered (Shah, Friedman, & Kruglanski,
2002).
Choosing to prioritize one goal (or set of goals) over other goals should only be temporarily
acceptable, however. The autonomous agent should have a built in fatigue process, so that the agent
can be responsive to environmental change. Given the dynamics of goal pursuit, as the environment
changes, new goals should be selected. This ensures that the agent can pursue complex goals in
changing situations, and even account and correct for its own actions.
More profoundly, the agent should not allow any goal to be suppressed indefinitely even when
the environment is static. The longer other goals are suppressed, the weaker that suppression should
become. The autonomous agent should be less able to hold back competing goals and more responsive
to environmental cues. This will lead to increased chances of goal switching. This is different from simple
hyperbolic discounting, because it applies even to pleasurable activities and when progress toward the
goal is acceptable. The agent may switch from higher priority to lower priority goals simply because in
pursuit of the higher priority goal, the lower priority goals had to be suppressed.
There is extensive evidence in humans that self-control comes at a cost and is limited (Muraven,
2012; Muraven & Baumeister, 2000). Suppressing competing urges weakens inhibition, so that the
ability to suppress alternative goals becomes weaker. Eventually, current goal pursuit is dropped and an
Muraven
13
alternative goal is pursued. Thus, no goal will be pursued indefinitely; people have an inbuilt mechanism
that causes their ability to suppress alternatives to fail.
Although this failure of self-control is a problem for many people, it clearly has benefits as well.
It prevents people from obsessively pursuing any goal, even if that goal is pleasurable and desired.
People stop working, creating art, playing video games or even taking heroin, at least for a little while to
pursue different goals. Thus, the failure of self-control in humans, although associated with negative
consequences, may be critical to maintaining successful autonomous self-regulation. This failure of
inhibition prevents autonomous agents from obsessive pursuit of any goal and helps ensure that
behavior is balanced between goals.
This fatigue process may itself be open to feedback. For example, under certain conditions, the
inhibition process may itself be strengthened or weakened to help the agent remain on task or allow
itself to be distracted by higher priority tasks. Thus, things like external rewards or importance of the
goal may allow the agent to at least temporarily ignore the fatigue feedback and continue working.
Conclusion
Although competing goals seem like a feature to avoid in an autonomous agent, having
conflicting goals is actually critical to the long-term stability of the agent. Such interfering goals prevent
autonomous agents from pursuing one goal to the exclusion of all others, thereby helping to avoid
situations of unwanted persistence, like the paperclipper maximizer problem.
Moreover, by having multiple goals active at once, problems with specifying one general law to
guide an autonomous agent are reduced. Instead, the autonomous agent will seek an unstable
equilibrium between several goal states, which prevents any one goal becoming predominant. The
Human Self-Regulation for AI
14
flexible feedback system further pushes the agent to avoid extremes, while permitting a level of
situational adjustment and self-correction.
The fact that goals are arranged in hierarchies also helps in the design of an autonomous
artificial intelligence. Only the top level goals need to be specified; the lower level goals that fulfill these
goals can be discovered through exploration. The flexible feedback system helps to ensure that as the
agent develops more complexity it will remain consistent with initial programming goals. That is, the
same top-level goals that are active in babies (e.g., make social connections, gain competence) are
active in adults, only they are pursued in a more complex way.
For the creation of an autonomous agent, specification of general, top-level groups that
generally serve the welfare and betterment of humanity should help ensure alignment of the agent with
human goals. In humans, these goals have been selected by evolution and society to help guide human
behavior and thus have been tested as successful. The rare human that deviates or fails to balance these
goals will undergo treatment to try to restore harmony and retraining to select appropriate goals, thus
suggesting that human misery is not a product of these goals but rather a failure of self-regulation itself.
An autonomous agent with a strong and well-designed self-regulatory system may therefore avoid these
problems.
That is not to say that human self-regulation is completely understood. For instance, the nature
of the topmost goals needs some level of specificity. Research on humans has identified many likely
candidates, although there remains some disagreement. There also needs to be some work on how to
formally specify these goals in order to assess progress. However, this is not an intractable problem,
because as noted above, each level of goal spawns multiple, competing lower level goals. Given that
each of these goals likely interfere and that the ability to suppress all other goals weakens over time, a
misspecification in a goal will be corrected in the long run.
Muraven
15
In summary, a consideration of human self-regulation may help in the design of a more general
autonomous agent. Key features of human self-regulation include a hierarchical goal structure, multiple
conflicting goals, an emotional feedback system and goal fatigue. This self-regulation system relies on
relatively simple rules that generate an unstable equilibrium and thus may be ported to an artificial
intelligence system. Moreover, there is reason to believe that an autonomous agent based on these
principles will more likely be aligned to human goals and avoid poorly optimized, nonsensical, or
monomaniac goal pursuit.
Human Self-Regulation for AI
16
References
Ballard, T., Yeo, G., Loft, S., Vancouver, J. B., & Neal, A. (2016). An Integrative Formal Model of
Motivation and Decision Making: The MGPM*.
Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies: OUP Oxford.
Busemeyer, J. R., & Diederich, A. (2002). Survey of decision field theory. Mathematical Social Sciences,
43(3), 345-370.
Carver, C. S., & Scheier, M. F. (1981). Attention and self-regulation: A control-theory approach to human
behavior. New York: Springer-Verlag.
Chirkov, V., Ryan, R. M., Kim, Y., & Kaplan, U. (2003). Differentiating autonomy from individualism and
independence: a self-determination theory perspective on internalization of cultural
orientations and well-being. Journal of Personality and Social Psychology, 84(1), 97.
Custers, R., & Aarts, H. (2007). In search of the nonconscious sources of goal pursuit: Accessibility and
positive affective valence of the goal state. Journal of Experimental Social Psychology, 43(2),
312-318.
Deci, E. L., & Ryan, R. M. (1991). A motivational approach to self: Integration in personality. In R.
Dienstbier (Ed.), Nebraska symposium on motivatioU: Vol. 38. Perspective on motivation (pp.
237-288). Lincoln, NB: University of Nebraska Press.
Deci, E. L., & Ryan, R. M. (2000). The "what" and "why" of goal pursuits: Human needs and the selfdetermination of behavior. Psychological Inquiry, 11(4), 227-268.
Elliot, A. J. (2006). The hierarchical model of approach-avoidance motivation. Motivation and Emotion,
30(2), 111-116.
Fisher, C. D. (1978). The effects of personal control, competence, and extrinsic reward systems on
intrinsic motivation. Organizational Behavior and Human Performance, 21(3), 273-288.
Förster, J., Liberman, N., & Friedman, R. S. (2007). Seven principles of goal activation: A systematic
approach to distinguishing goal priming from priming of non-goal constructs. Personality and
Social Psychology Review, 11(3), 211-233.
Graham, M. D., Young, R. A., Valach, L., & Alan Wood, R. (2008). Addiction as a complex social process:
An action theoretical perspective. Addiction Research & Theory, 16(2), 121-133.
Kruglanski, A. W., Köpetz, C., Bélanger, J. J., Chun, W. Y., Orehek, E., & Fishbach, A. (2013). Features of
multifinality. Personality and Social Psychology Review, 17(1), 22-39.
Kruglanski, A. W., Pierro, A., & Sheveland, A. (2011). How many roads lead to Rome? Equifinality set‐size
and commitment to goals and means. European Journal of Social Psychology, 41(3), 344-352.
Muraven
17
Lord, R. G., & Hanges, P. J. (1987). A control system model of organizational motivation: Theoretical
development and applied implications. Behavioral Science, 32(3), 161-178.
Lynn, S. K., Wormwood, J. B., Barrett, L. F., & Quigley, K. S. (2015). Decision making from economic and
signal detection perspectives: development of an integrated framework. Frontiers in Psychology,
6, 952. doi:10.3389/fpsyg.2015.00952
Muraven, M. (2012). Ego-Depletion: Theory and Research. In R. M. Ryan (Ed.), Oxford Handbook of
Motivation (pp. 111-126). New York: Oxford University Press.
Muraven, M., & Baumeister, R. F. (2000). Self-regulation and depletion of limited resources: Does selfcontrol resemble a muscle? Psychological Bulletin, 126(2), 247-259.
Shah, J. Y., Friedman, R., & Kruglanski, A. W. (2002). Forgetting all else: On the antecedents and
consequences of goal shielding. Journal of Personality and Social Psychology, 83(6), 1261-1280.
Steel, P., & König, C. J. (2006). Integrating theories of motivation. Academy of Management Review,
31(4), 889-913.
Stroebe, W., Mensink, W., Aarts, H., Schut, H., & Kruglanski, A. W. (2008). Why dieters fail: Testing the
goal conflict model of eating. Journal of Experimental Social Psychology, 44(1), 26-36.
Tangney, J. P. (2003). Self-relevant emotions. In M. R. L. J. P. Tangney (Ed.), Handbook of self and identity
(pp. 384-400). New York, NY, US: Guilford Press.
Tangney, J. P., Baumeister, R. F., & Boone, A. L. (2004). High Self-Control Predicts Good Adjustment, Less
Pathology, Better Grades, and Interpersonal Success. Journal of Personality, 72(2), 271-322.
Taylor, S. E., Pham, L. B., Rivkin, I. D., & Armor, D. A. (1998). Harnessing the imagination: Mental
simulation, self-regulation, and coping. American Psychologist, 53(4), 429-439.
| 3cs.SY
|
arXiv:1709.03565v1 [cs.DS] 11 Sep 2017
Importance Sketching of Influence Dynamics in
Billion-scale Networks
Hung T. Nguyen, Tri P. Nguyen
NhatHai Phan
Thang N. Dinh
Virginia Commonwealth University
Richmond, VA 23284, USA
Email: {hungnt,trinpm}@vcu.edu
New Jersey Institute of Technology
Newark, NJ 07104, USA
Email: [email protected]
Virginia Commonwealth University
Richmond, VA 23284, USA
Email: [email protected]
Abstract—The blooming availability of traces for social, biological, and communication networks opens up unprecedented opportunities in analyzing diffusion processes in networks. However,
the sheer sizes of the nowadays networks raise serious challenges
in computational efficiency and scalability.
In this paper, we propose a new hyper-graph sketching
framework for influence dynamics in networks. The central of
our sketching framework, called SKIS, is an efficient importance sampling algorithm that returns only non-singular reverse
cascades in the network. Comparing to previously developed
sketches like RIS and SKIM, our sketch significantly enhances
estimation quality while substantially reducing processing time
and memory-footprint. Further, we present general strategies
of using SKIS to enhance existing algorithms for influence
estimation and influence maximization which are motivated by
practical applications like viral marketing. Using SKIS, we design
high-quality influence oracle for seed sets with average estimation
error up to 10x times smaller than those using RIS and 6x times
smaller than SKIMs. In addition, our influence maximization
using SKIS substantially improves the quality of solutions for
greedy algorithms. It achieves up to 10x times speed-up and 4x
memory reduction for the fastest RIS-based DSSA algorithm,
while maintaining the same theoretical guarantees.
I. I NTRODUCTION
Online social networks (OSNs) such as Facebook and
Twiter have connected billions of users, providing gigantic
communication platforms for exchanging and disseminating
information. For example, Facebook now has nearly 2 billions monthly active users and more than 2.5 billion pieces
of content exchanged daily. Through OSNs, companies and
participants have actively capitalized on the “word-of-mouth”
effect to trigger viral spread of various kinds of information,
including marketing messages, propaganda, and even fake
news. In the past decade, a great amount of research has
focused on analyzing how information and users’ influence
propagate within the network, e.g., evaluating influence of a
group of individuals, aka influence estimation [1], [2], [3], and
finding a small group of influential individuals, aka influence
maximization [4], [2], [5], [6], [7], [8], or controlling diffusion
processes via structure manipulation [9].
Yet, diffusion analysis is challenging due to the sheer size of
the networks. For example, state-of-the-art solutions for influence maximization, e.g., DSSA [8], IMM [7], TIM/TIM+[6],
cannot complete in the networks with only few million edges
[10]. Further, our comprehensive experiments on influence
estimation show that the average estimation error can be as
high as 40-70% for popular sketches such as RIS and SKIM.
This happens even on a small network with only 75K nodes
(Epinion). This calls for development of new approaches for
analyzing influence dynamics in large-scale networks.
In this paper, we propose a new importance sketching
technique, termed SKIS, that consists of non-singular reverse
influence cascades, or simply non-singular cascades. Each
non-singular cascade simulates the reverse diffusion process
from a source node. It is important that each non-singular
cascade must include at least another node other than the
source itself. Thus, our sketch, specifically, suppresses singular
cascades that die prematurely at the source. Those singular
cascades, consisting of 30%-80% portion in the previous
sketches [2], [5], not only waste the memory space and
processing time but also reduce estimation efficiency of the
sketches. Consequently, SKIS contains samples of smaller
variances providing estimations of high concentration with
less memory and running time. Our new sketch also powers
a new principle and scalable influence maximization class
of methods, that inherits the algorithmic designs of existing
algorithms on top of SKIS sketch. Particularly, SKIS-based
IM methods are the only provably good and efficient enough
that can scale to networks of billions of edges across different
settings. We summarize of our contributions as follows:
• At the central of our sketch is an importance sampling
algorithm to sample non-singular cascades (Alg. 1). For
simplicity, we first present the sketch and its sampling
algorithm using the popular independent cascade model
[4], and later extend them to other diffusion models.
• We provide general frameworks to apply SKIS for existing algorithms for the influence estimation and influence
maximization problems. We provide theoretical analysis
to show that using SKIS leads to improved influence
estimation oracle due to smaller sample variances and
better concentration bounds; and that the state-of-theart methods for influence maximization like D-SSA [8],
IMM [7], and, TIM/TIM+[6] can also immediately benefit from our new sketch.
• We conduct comprehensive empirical experiments to
demonstrate the effectiveness of our sketch in terms of
quality, memory and computational time. Using SKIS,
we can design high-quality influence oracle for seed set
with average estimation error up to 10x times smaller than
those using RIS and 6x times those using SKIM. In addition, our influence maximization using SKIS substantially
improves the quality of solutions for greedy algorithms.
It achieves up to 10x times speed-up and 4x memory
reduction for the fastest RIS-based DSSA algorithm,
while maintaining the same theoretical guarantees.
Related work. Sketching methods have become extremely
useful for dealing with problems in massive sizes. Most
notable sketches including bottom-k sketches [11] a summary
of a set of items with nonnegative weights, count-min sketch
[12] that count the frequency of different events in the stream.
Recently, Cohen et al. [2] investigate the combined reachability sketch which is a bottom-k min-hash sketch of the set
of reachable nodes. They show small estimation errors for
estimating influences. However, this method deteriorates for
large influences since the size of each sketch is fixed. Similar
scheme was applied for continuous-time model [13].
Borgs et al. [5] proposed reverse influence sketch (RIS)
which captures the influences in a reverse manner. This
approach has advantage in estimating large influences and
becomes very successful in finding the seed set with maximum
influence, i.e. influence maximization. [3] uses RIS sketches to
estimate influences in dynamic graphs. Other related works on
influences in multiplex networks and identifying the sources
of influence cascades are studied in [14], [15], [16].
Organization. The rest of the paper is organized as follows:
In Section II, we introduce the diffusion model and two problems of influence estimation/maximization. We propose our
importance sketching scheme in Section III. Applications in
influence estimation/maximization is presented in Sections IV
and V, respectively. Extensions to other diffusion models are
discussed in Section VI which is followed by experiments in
Section VII and conclusion in Section VIII.
II. P RELIMINARIES
Consider a social network abstracted as a graph G =
(V, E, w). Each edge (u, v) ∈ E is associated with a real
number w(u, v) ∈ [0, 1] specifying the probability that node u
will influence v once u is influenced. To model the influence
dynamic in the network, we first focus on the popular Independent Cascade (IC) model [4] and then, discuss the extensions
of our techniques to other models, e.g. Linear Threshold (LT)
or Continuous-time model, later in Section VI.
A. Independent Cascade Model
For a subset of nodes S ⊆ V , called seed set, the influence
propagation from S happens in discrete rounds t = 0, 1, ...
At round 0, only nodes in S are active (aka influenced) and
the others are inactive. Each newly activated node u at round
t will have a single chance to activate each neighbor v of u
with probability w(u, v). An activated node remains active till
the end of the diffusion propagation. The process stops when
no more nodes get activated.
Sample Graphs. Once a node u gets activated, it will
activate each of its neighbor v with probability w(u, v). This
can be thought of as flipping a biased coin that gives head
with probability w(u, v) to determine whether the edge (u, v)
exists. If the coin lands head for the edge (u, v), the activation
occurs and we call (u, v) a live-edge. Since all the influences
in the IC model are independent, it does not matter when coins
are flipped to determine the states of the edges. Thus, we can
flip all the coins at the beginning instead of waiting until u
gets activated. We call the deterministic graph g that contains
all the live-edges resulted from a series of coin flips over all
the edges in G a sample graph of G.
Probabilistic Space. The set of all sample graphs generated
from G together with their probabilities define a probabilistic
space ΩG . Each sample graph g ∈ ΩG can be generated by
flipping coins on all the edges to determine whether or not the
edge is live or appears in g. That is each edge (u, v) will be
present in a sample graph with probability w(u, v). Therefore,
a sample graph g = (V, E 0 ⊆ E) is generated from G with a
probability Pr[g ∼ G] calculated by,
Y
Y
(1 − w(u, v)). (1)
w(u, v)
Pr[g ∼ G] =
(u,v)∈E
/ 0
(u,v)∈E 0
Influence Spread. Given a diffusion model, the measure
Influence Spread (or simply influence) of a seed set S is
defined as the expected number of active nodes in the end
of the diffusion propagation, where the expectation is taken
over the probabilistic space ΩG . Given a sample graph g ∼ G
and a seed set S ⊂ V , we denote ηg (S) the set of nodes
reachable from S (including nodes in S themselves). The
influence spread of S is defined as follows,
X
I(S) =
|ηg (S)| Pr[g ∼ G].
(2)
g∼G
The frequently used notations are summarized in Table I.
TABLE I: Table of notations
Notation
n, m
I(S), Î(S)
N in (S)
γv , Γ
γ0
Rj , R
CR (S)
Description
#nodes, #edges of graph G = (V, E, w).
Expected Influence of S ⊆ V and an estimate.
Set of in-neighbor nodes of S.
P
γv = 1−Πu∈N in (v) (1−w(u, v)); Γ = v∈V γv .
P
γ0 = v∈V γv /n.
A random IIS sample and a SKIS sketch.
CR (S) = |Rj ∈ R|Rj ∩ S 6= ∅|.
B. Influence Estimation/Maximization Problems
We describe the tasks of Influence Estimation and Maximization which are used to evaluate sketches’ efficiency.
Definition 1 (Influence Estimation (IE)). Given a probabilistic
graph G and a seed set of nodes S ⊆ V , the IE problem asks
for an estimation Î(S) of the set influence I(S).
Definition 2 (Influence Maximization (IM) [4]). Given a
probabilistic graph G, a budget k, the IM problem asks for a
set Sk of size at most k having the maximum influence among
all other sets of size at most k,
Sk = arg max I(S).
(3)
S⊆V,|S|≤k
C. Sketch-based Methods for IE/IM
90
30
g,ηg (src(Rj ))=Rj
The key property of RIS samples for influence estimation/maximization is stated in the following lemma.
Lemma 1 ([5]). Given a random RIS sample Rj generated
from G = (V, E, w), for a set S ⊆ V of nodes, we have,
I(S) = n · Pr[Rj ∩ S 6= ∅].
(5)
Thus, estimating/maximizing I(S) is equivalent to estimating/maximizing the probability Pr[Rj ∩ S 6= ∅].
Using RIS samples for IE/IM. Thanks to Lemma 1, a
general strategy for IE/IM is generating a set of RIS samples,
then returning an empirical estimate of Pr[Rj ∩ S 6= ∅] on
generated samples for IE or the set Ŝk that intersects with most
samples for IM. The strong advantage of RIS is the reuse of
samples to estimate influence of any seed set S ⊆ V . Ohsaka
et al. [3] build a query system to answer influence queries. [5],
[6], [7], [8], [17] recently use RIS samples in solving Influence
Maximization problem with great successes, i.e. handling large
networks with tens of millions of nodes and billions of edges.
2) Combined Reachability Sketch (SKIM): Cohen et al. [2]
proposed the combined reachability sketch which can be used
to estimate influences of multiple seed sets. Each node u in
the network is assigned a combined reachability sketch which
is a bottom-k min-hash sketch of the set of nodes reachable
from u in l sample graphs. [2] generates l sample graphs g of
G, i.e. l = 64 by default, and build a combined reachability
sketch of size k for each node.
The influence estimate of a seed set S is computed by taking
the bottom-k sketch of the union over all the sketches of nodes
in S and applying the cardinality estimator [2]. Using the
sketches, the solution for IM is found by following the greedy
algorithm which repeatedly adds a node with highest marginal
influence into the solution. Here, the marginal influences are
similarly estimated from node sketches.
Common Shortcomings. According to recent benchmarks
[10] and our own empirical evaluations (details in Section VII),
both RIS and SKIM yield significant influence estimation
errors. For RIS, it is due to the fact that the majority of
RIS samples contain only their sources as demonstrated in
%Samples
60
%Samples
1) Reverse Influence Sketch (RIS): Essentially, a random
RIS sample, denoted by Rj , contains a random set of nodes,
following a diffusion model, that can influence a randomly
selected source node, denoted by src(Rj ). A RIS sample is
generated in three steps:
1) Select a random node v ∈ V which serves as src(Rj ).
2) Generate a sample graph g ∼ G.
3) Return the set Rj of nodes that can reach v in g.
Thus, the probability of generating a particular RIS sample
Rj can be computed based on the source selection and the
sample graphs that has Rj as the set of nodes that reach
src(Rj ) in g. Let denote such set of nodes that can reach
to a node v in sample graph g by ηg− (v). We have,
X
1
Pr[g].
(4)
Pr[Rj ] =
n −
20
30
10
0
0
1
2
3
4
5
6
Sample Size
7
8
(a) Weighted Cascade (WC)
1
2
3
4
5
6
Sample Size
7
8
(b) Trivalency (TRI)
Fig. 1: Distribution of reversed cascade sizes on Epinions
network (and on other networks as well) with two different
edge weight models: Weighted Cascade (WC) and Trivalency
(TRI). The majority of the cascades are singular.
Figure 1 with up to 86% of such RIS samples overall. These
samples, termed singular, harm the performance in two ways:
1) they do not contribute to the influence computation of other
seed sets than the ones that contain the sources, however, the
contribution is known in advance, i.e. number of seed nodes; 2)
these samples magnify the variance of the RIS-based random
variables used in estimation causing high errors.
This motivates our Importance Sketching techniques to generate only non-singular samples that are useful for influence
estimations of many seed sets (not just ones with the sources).
III. I MPORTANCE S KETCHING
This section introduces our core construction of Importance
Sketching algorithm to generate random non-singular samples
with probabilities proportional to those in the original sample
space of reverse influence samples and normalized by the
probability of generating a non-singular ones.
Algorithm 1: Importance Influence Sampling (IIS) Alg.
1
2
3
4
5
6
Input: Graph G = (V, E, w)
Output: Rj - A random IIS sample
Pick a node v ∈ V as the source with probability in Eq. 8;
Select an in-neighbor ui of v, ui ∈ N in (v), with probability of
selecting ui given in Eq. 9;
Initialize a queue Q = {ui } and a node set Rj = {v, ui };
foreach ut ∈ N in (v), t 6= i do
With probability w(ut , v):
Q.push(ut ); Rj ← Rj ∪ {ut };
11
while Q is not empty do
v = Q.pop();// get the longest inserted
node
foreach u ∈ N in (v)\(Rj ∪ Q) do
With probability w(u, v):
Q.push(u); Rj ← Rj ∪ {u}; // insert u
12
return Rj ;
7
8
9
10
A. Importance Influence Sampling (IIS)
Sample Spaces and Desired Property. Let ΩRIS be the
sampling space of reverse influence samples (RIS) with probability Pr[Rj ∈ ΩRIS ] of generating sample Rj . Let ΩSKIS
be a subspace of ΩRIS and corresponds to the space of only
non-singular reverse influence samples in ΩRIS . Since ΩSKIS
is a subspace of ΩRIS , the probability Pr[Rj ∈ ΩSKIS ] of
generating a non-singular sample from ΩSKIS is larger than
that from ΩRIS . Specifically, for a node v ∈ V , let γv be
the probability of generating a non-singular
sample if v is
P
selected as the source and Γ =
γ
.
Then,
since the
v∈V v
sample sources are selected randomly, the ratio of generating
a non-singular sample to generating any sample in ΩRIS is nΓ
and thus, the probability Pr[Rj ∈ ΩSKIS ] is as follows,
n
Pr[Rj ∈ ΩSKIS ] = Pr[Rj ∈ ΩRIS ].
(6)
Γ
Our upcoming IIS algorithm aims to achieve this desired
property of sampling non-singular samples from ΩSKIS .
Sampling Algorithm. Our Importance Influence Sampling
(IIS) scheme involves three core components:
1) Probability of having a non-singular sample. For a node
v ∈ V , a sample with source v is singular if no inneighbor of v is selected, that happens with probability
Πu∈N in (v) (1−w(u, v)). Hence, the probability of having
a non-singular sample from a node v is the complement:
γv = 1 − Πu∈N in (v) (1 − w(u, v)).
(7)
2) Source Sampling Rate. Note that the set of non-singular
samples is just a subset of all possible samples and we
want to generate uniformly random samples from that
subset. Moreover, each node v has a probability γv of
generating a non-singular sample from it. Thus, in order
to generate a random sample, we select v as the source
with probability Pr[src(Rj ) = v] computed as follows,
γv
γv
,
(8)
=
Pr[src(Rj ) = v] = P
Γ
u∈V γu
P
where Γ =
u∈V γu , and then generate a uniformly
random non-singular sample from the specific source v
as described in the next component.
3) Sample a non-singular sample from a source. From the
src(Rj ) = v, we generate a non-singular sample Rj from
v uniformly at random. Let N in (v) = {u1 , u2 , . . . , ul }
be a fixed-order set of in-neighbors of v. We divide the
all possible non-singular samples from v into l buckets:
bucket Bi , 1 ≤ i ≤ l contains those samples that have
the first node from N in (v) being ui . That means all the
nodes u1 , . . . , ui−1 are not in the sample but ui is in
for certain. The other nodes from ui+1 to ul may appear
and will be sampled following the normal RIS sampling.
Now we select the bucket that Rj belongs to with the
probability of selecting Bi being as follows,
Qi−1
(1 − w(ut , v))w(ui , v)
.
(9)
Pr[select Bi ] = t=1
γv
For i = 1, we have Pr[select B1 ] = w(u1 , v). Note that
P
l
i=1 Pr[select Bi ] = 1. Assume bucket Bi is selected
and, thus, node ui is added as the second node besides
the source into Rj . For each other node ut , t 6= i, ut is
selected into Rj with probability w(ut , v) following the
ordinary RIS for the IC model.
These three components guarantee a non-singular sample.
The detailed description of IIS sampling is in Alg. 1. The
first step selects the source of the IIS sample among V . Then,
the first incoming node to the source v is picked (Line 2)
following the above description of the component 3). Each of
the other incoming neighbors also tries to influence the source
(Lines 4-6). The rest performs similarly as in RIS [5]. That
is for each newly selected node, its incoming neighbors are
randomly added into the sample with probabilities equal to
their edge weights. It continues until no newly selected node
is observed. Note that Line 3 only adds the selected neighbors
ui of v into Q but adds both v and ui to Rj . The loop from
Lines 7-11 mimics the BFS-like sampling procedure of RIS.
Let Pr[Rj ] be the probability of generating a non-singular
sample Rj using IIS algorithm. We have
X
Pr[Rj ] =
Pr[src(Rj ) = v] Pr[generate Rj from v]
v∈V
X γv Pr[Rj ∈ ΩRIS and src(Rj ) = v]
Γ
γv
v∈V
X
n
1
=
Pr[Rj ∈ ΩRIS and src(Rj ) = v]
Γ
n
v∈V
n
= Pr[Rj ∈ ΩRIS ] = Pr[Rj ∈ ΩSKIS ],
Γ
Pr[Rj ∈ΩRIS and src(Rj )=v]
where Pr[generate Rj from v] =
due
γv
to the selection of the bucket that Rj belongs to in IIS. Thus,
the output Rj of IIS is an random sample from non-singular
space ΩSKIS and we obtain the following lemma.
=
Lemma 2. Recall that ΩSKIS is the sample space of nonsingular reverse influence samples. IIS algorithm generates a
random non-singular sample from sample space ΩSKIS .
Connection between IIS Samples and Influences. We
establish the following key lemma that connects our IIS
samples with the influence of any seed set S.
Lemma 3. Given a random IIS sample Rj generated by Alg. 1
from the graph G = (V, E, w), for any set S ⊆ V , we have,
X
I(S) = Pr[Rj ∩ S 6= ∅] · Γ +
(1 − γv ),
(10)
v∈S
where γv and Γ are defined in Eqs. 7 and 8.
The proof is presented in our extended version [18]. The
influence I(S) of any set S comprises of two parts: 1) Pr[Rj ∩
S 6= ∅] ·P
Γ depends on the randomness of Rj and 2) the fixed
amount v∈S (1−γv ) that is inherent to set S and accounts for
the contribution of singular samples in ΩRIS to the influence
I(S). Lemma 3 states that instead of computing or estimating
the influence I(S) directly, we
P can equivalently compute or
estimate Pr[Rj ∩S 6= ∅]·Γ+ v∈S (1−γv ) using IIS samples.
Remark: Notice that we can further generate samples of
larger sizes and reduce the variance as shown later, however,
the computation would increase significantly.
IV. I NFLUENCE O RACLE VIA IIS S KETCH (SKIS)
We use IIS sampling to generate a sketch for answering
influence estimation queries of different node sets. We show
that the random variables associated with our samples have
much smaller variances than that of RIS, and hence, lead
to better concentration or faster estimation with much fewer
samples required to achieve the same or better quality.
SKIS-based Influence Oracle. An SKIS sketch R is a
collection of IIS samples generated by Alg. 1, i.e. R =
{R1 , . . . , RT }. As shown in Lemma 3, the influence I(S) can
be estimated through estimating the probability Pr[Rj ∩ S 6=
∅]. Thus, from a SKIS sketch R = {R1 , . . . , RT }, we can
obtain an estimate Î(S) of I(S) for any set S by,
X
CR (S)
ÎR (S) =
·Γ+
(1 − γv ),
(11)
|R|
v∈S
where CR (S) is coverage of S on R, i.e.,
CR (S) = |{Rj ∈ R|Rj ∩ S 6= ∅}|.
(12)
Lemma 4. The random variable Zj (S) (Eq. 14) has
I(S) Γ I2 (S)
Var[Zj (S)] =
−
n n
n2
P
X
(1 − γv )
(Γ +
(1 − γv ) − 2I(S)).
− v∈S 2
n
(19)
v∈S
Since the random variables Yj (S) for RIS samples are
Bernoulli and E[Yj (S)] = I(S)
n , we have Var[Yj (S)] =
I(S)
I(S)
(1
−
).
Compared
with
Var[Zj (S)], we observe that
n
n
2
I(S) Γ
I2 (S)
I(S)
Γ
= Var[Yj (S)],
since n ≤ 1, n n − n2 ≤ n − I n(S)
2
Var[Zj (S)] ≤ Var[Yj (S)]
P
X
(1 − γv )
− v∈S 2
(Γ +
(1 − γv ) − 2I(S)).
n
v∈S
Algorithm 2: SKIS-based Influence Oracle
1
2
Input: Graph G = (V, E, w)
Preprocessing: Generate a SKIS sketch R = {R1 , . . . , RT } of
IIS samples using Alg. 1.
For any influence query for any set S: return ÎR (S) (Eq. 11).
We build an SKIS-based oracle for influence queries by
generating a set R of T IIS samples in a preprocessing
step and then answer influence estimation query ÎR (S) for
any requested set S (Alg. 2). In the following, we show the
better estimation quality of our sketch through analyzing the
variances and estimating concentration properties.
SKIS Random Variables for Estimations. For a random
IIS sample Rj and a set S, we define random variables:
(
1 if Rj ∩ S 6= ∅
Xj (S) =
, and
(13)
0 otherwise.
P
Xj (S) · Γ + v∈S (1 − γv )
.
(14)
Zj (S) =
n
Then, the means of Xj (S) and Zj (S) are as follows,
P
I(S) − v∈S (1 − γv )
(15)
E[Xj (S)] = Pr[Rj ∩ S 6= ∅] =
Γ
P
(1 − γv )
Γ
I(S)
E[Zj (S)] = E[Xj (S)] · + v∈S
=
. (16)
n
n
n
Hence, we can construct a corresponding set of random
variables Z1 (S),
PTZ2 (S), . . . , ZT (S) by Eqs. 13 and 14. Then,
ÎR (S) = Tn j=1 Zj (S) is an empirical estimate of I(S)
based on the SKIS sketch R.
For comparison purposes, let Yj (S) be the random variable
associated with RIS sample Qj in a RIS sketch Q,
(
1 if Qj ∩ S 6= ∅
Yj (S) =
(17)
0 otherwise.
From Lemma 1, the mean value of Yj (S) is then,
I(S)
E[Yj (S)] =
.
n
(18)
Variance Reduction Analysis. We show that the variance
of Zj (S) for SKIS is much smaller than that of Yj (S) for
RIS. The variance of Zj (S) is stated in the following.
In practice, most ofP
seed sets have small influences, i.e.
I(S) Γ2 , thus, Γ + v∈S (1 − γv ) − 2I(S) 0. Hence,
Var[Zj (S)] < Var[Yj (S)] holds for most seed sets S.
Better Concentrations
Variables.
ObP
h P of SKIS Random
i
v∈S (1−γv ) Γ+
v∈S (1−γv )
, we obtain
,
serve that Zj (S) ∈
n
n
another result on the variance of Zj (S) as follows.
Lemma 5. The variance of random variable Zj (S) satisfies
I(S) Γ
Var[Zj (S)] ≤
.
(20)
n n
Using the above result with the general form of Chernoff’s
bound in Lemma 2 in [7], we derive the following concentration inequalities for random variables Zj (S) of SKIS.
Lemma 6. Given a SKIS sketch R = {R1 , . . . , RT } with
random variables Z1 (S), . . . , ZT (S), we have,
h PT Zj (S)
i
−2 T I(S)
j=1
Pr
n − I(S) ≥ I(S) ≤ exp
Γ
T
+ 23 n
2n
P
i
−2 T I(S)
h T Zj (S)
j=1
n − I(S) ≤ −I(S) ≤ exp
.
Pr
Γ
T
n
2n
Compared with the bounds for RIS sketch in Corollaries 1
and 2 in [7], the above concentration bounds for SKIS sketch
(Lemma 6) are stronger, i.e. tighter. Specifically, we have the
factor nΓ in the denominator of the exp(.) function while for
RIS random variables, it is simply 1.
Sufficient Size of SKIS Sketch for High-quality Estimations. There are multiple strategies to determine the number of
IIS samples generated in the preprocessing step. For example,
[3] generates samples until total size of all samples reaches
O( 13 (n + m) log(n)). Generating IIS samples to reach such
a specified threshold is vastly faster than using RIS due to
the bigger size of IIS samples. This method provides an
additive estimation error guarantee within . Alternatively, by
Lemma 6, we derive the sufficient number of IIS samples to
provide the more preferable (, δ)-estimation of I(S).
Lemma 7. Given a set S, , δ ≥ 0, if the SKIS sketch R
n −2
has at least (2 nΓ + 23 ) ln( 2δ ) I(S)
IIS samples, ÎR (S) is an
(, δ)-estimate of I(S), i.e.,
Pr[(1 − )I(S) ≤ ÎR (S) ≤ (1 + )I(S)] ≥ 1 − δ.
(21)
In practice, I(S) is unknown in advance and a lower-bound
of I(S), e.g. |S|, can be used to compute the necessary number
of samples to provide the same guarantee. Compared to RIS
with weaker concentration bounds, we save a factor of O( Γn ).
V. SKIS- BASED IM A LGORITHMS
With the help of SKIS sketch that is better in estimating the
influences compared to the well-known successful RIS, we can
largely improve the efficiency of IM algorithms in the broad
class of RIS-based methods, i.e. RIS [5], TIM/TIM+ [6], IMM
[7], BCT [19], [20], SSA/DSSA [8]. This improvement is
possible since these methods heavily rely on the concentration
of influence estimations provided by RIS samples.
SKIS-based framework. Let R = {R1 , R2 , . . . } be a
SKIS sketch of IIS samples. R gives an influence estimate
X
CR (S)
·Γ+
(1 − γv ), (22)
ÎR (S) = ÊR [Zj (S)] · n =
|R|
v∈S
for any set S. Thus, instead of optimizing over the exact
influence, we can intuitively find the set S to maximize the
estimate function Î(S). Then, the framework of using SKIS
sketch to solve IM problem contains two main steps:
1) Generate a SKIS sketch R of IIS samples,
2) Find the set Sk that maximizes the function ÎR (S) and
returning Sk as the solution for the IM instance.
There are two essential questions related to the above SKISbased framework : 1) Given a SKIS sketch R of IIS samples,
how to find Sk of k nodes that maximizes ÎR (Sk ) (in Step
2)? 2) How many IIS samples in the SKIS sketch R (in Step
1) are sufficient to guarantee a high-quality solution for IM?
We give the answers for the above questions in the following
sections. Firstly, we adapt the gold-standard greedy algorithm
to obtain an (1 − (1 − 1/k)k )-approximate solution over a
SKIS sketch. Secondly, we adopt recent techniques on RIS
with strong solution guarantees to SKIS sketch.
Algorithm 3: Greedy Algorithm on SKIS sketch
4
Input: SKIS sketch R and k
Output: An (1 − (1 − 1/k)k )-approximate seed set Ŝk
Ŝk = ∅
for i = 1 : k do
(v,Ŝk )
v̂ ← arg maxv∈V \Ŝk ∆R|R|
Γ + (1 − γv )
Add v̂ to Ŝk
5
return Ŝk
1
2
3
A. Greedy Algorithm on SKIS Sketches
Let consider the optimization problem of finding a set Sk of
at most k nodes to maximize the function ÎR (S) on a SKIS
sketch R of IIS samples under the cardinality constraint |S| ≤
k. The function ÎR (S) is monotone and submodular since it
is the weighted
P sum of a set coverage function CR (S) and
a linear term v∈S (1 − γv ). Thus, we obtain the following
lemma with the detailed proof in our extended version [18].
Lemma 8. Given a set of IIS samples R, the set function
ÎR (S) defined in Eq. 22 is monotone and submodular.
Thus, a standard greedy scheme [21], which iteratively selects a node with highest marginal gain, gives an (1−(1− k1 )k ),
that converges to (1 − 1/e) asymptotically, approximate solution Ŝk . The marginal gain of a node v with respect to a set
S on SKIS sketch R is defined as follows,
∆R (v, Ŝk )
gainR (v, S) =
Γ + (1 − γv ),
(23)
|R|
where ∆R (v, S) = CR (S ∪ {v}) − CR (S) is called the
marginal coverage gain of v w.r.t. S on SKIS sketch R.
Given a collection of IIS samples R and a budget k, the
Greedy algorithm is presented in Alg. 3 with a main loop
(Lines 2-4) of k iterations. Each iteration picks a node v̂ having
largest marginal gain (Eq. 23) with respect to the current
partial solution Ŝk and adds it to Ŝk . The approximation
guarantee of the Greedy algorithm (Alg. 3) is stated below.
Lemma 9. The Greedy algorithm (Alg. 3) returns an (1 −
(1 − k1 )k )-approximate solution Ŝk ,
1
∗
),
(24)
ÎR (Ŝk ) ≥ (1 − (1 − )k )ÎR (SR
k
∗
is the optimal cover set of size k on sketch R.
where SR
The lemma is derived directly from the 1 − (1 − k1 )k
approximation factor of the ordinary greedy algorithm [21].
B. Sufficient Size of SKIS Sketch for IM
Since the SKIS sketch offers a similar greedy algorithm
with approximation ratio (1 − (1 − 1/k)k ) to the traditional
RIS, we can combine SKIS sketch with any RIS-based
algorithm, e.g. RIS[5], TIM/TIM+[6], IMM[7], BCT[20],
SSA/DSSA[8]. We discuss the adoptions of two most recent
and scalable algorithms, i.e. IMM[7] and SSA/DSSA[8].
IMM+SKIS. Tang et al. [7] provide a theoretical threshold
n
n −2
θRIS = O (log
+ log δ −1 )
(25)
k
OPTk
on the number of RIS samples to guarantee an (1 − 1/e − )approximate solution for IM problem with probability 1 − δ.
Replacing RIS with IIS samples to build a SKIS sketch
enables us to use the better bounds in Lemma 6. By the
approach of IMM in [7] with Lemma 6, we reduce the
threshold of samples to provide the same quality to,
Γ + k
θSKIS = O
θRIS .
(26)
n
SSA/DSSA+SKIS. More recently, Nguyen et al. [8] propose SSA and DSSA algorithms which implement the Stopand-Stare strategy of alternating between finding candidate
solutions and checking the quality of those candidates at
exponential points, i.e. 2t , t ≥ 1, to detect a satisfactory
solution at the earliest time.
Combining SKIS with SSA or DSSA brings about multiple
benefits in the checking step of SSA/DSSA. The benefits stem
from the better concentration bounds which lead to better error
estimations and smaller thresholds to terminate the algorithms.
The details are in our extended version [18].
L
VI. E XTENSIONS TO OTHER DIFFUSION MODELS
The key step in extending our techniques for other diffusion
models is devising an importance sketching procedure for each
model. Fortunately, following the same designing principle as
IIS, we can devise importance sketching procedures for many
other diffusion models. We demonstrate this through two other
equally important and widely adopted diffusion models, i.e.
Linear Threshold [4] and Continuous-time model [13].
Linear Threshold model [4]. This model imposes a constraint that the total weights
P of incoming edges into any node
v ∈ V is at most 1, i.e. u∈N in (v) w(u, v) ≤ 1. Every node
has a random activation threshold λv ∈ [0, 1] and gets activated
if the
Ptotal edge weights from active in-neighbors exceeds λv ,
i.e. u∈N in (v),u is active w(u, v) ≥ λv . A RIS sampling for LT
model [20] selects a random node as the source and iteratively
picks at most one in-neighbor of the last activated node with
probability being the edge weights, w(u, v).
The importance sketching algorithm for the LT model has
the following components:
•
Probability of having a non-singular sample:
X
γv =
w(u, v)
(27)
u∈N in (v)
•
Source Sampling Rate:
Pr[src(Rj ) = v] = P
γv
(28)
γv
Sample a non-singular sample from a source.: select exactly one in-neighbor u of src(Rj ) = v with probability
w(u,v)
γv . The rest follows RIS sampling [19].
v∈V
•
Continuous-time model [13]. Here we have a deadline
parameter T of the latest activation time and each edge (u, v)
is associated with a length distribution, represented by a
density function L(u,v) (t), of how long it takes u to influence
v. A node u is influenced if the length of the shortest path
from any active node at time 0 is at most T . The RIS sampling
for the Continuous-time model [7] picks a random node as the
source and invokes the Dijkstra’s algorithm to select nodes into
src(Rj ). When the edge (u, v) is first visited, the activation
time is sampled following its length distribution L(u,v) (t).
From the length distribution, we can compute the probability
p(u, v, T ) of an edge (u, v) having activation time at most T
Z T
p(u, v, T ) =
L(u,v) (t)dt
(29)
t=0
The importance sketching procedure for the Continuoustime model has the following components:
•
Probability of having a non-singular sample:
Y
γv = 1 −
(1 − p(u, v, T ))
(30)
u∈N in (v)
•
Source Sampling Rate:
Pr[src(Rj ) = v] = P
γv
(31)
γv
Sample a non-singular sample from a source.: Use a
bucket system on p(u, v, T ) similarly to IIS to select
the first in-neighbor u. The activation time of u follows
v∈V
•
(t)
. Subsequently,
the normalized density function (u,v)
γv
it continues by following RIS sampling [7].
VII. E XPERIMENTS
We demonstrate the advantages of our SKIS sketch through
a comprehensive set of experiments on the key influence
estimation and maximization problems. Due to space limit,
we report the results under the IC model and partial results
for the LT model. However, the implementations for all models
will be released on our website to produce complete results.
A. Experimental Settings
TABLE II: Datasets’ Statistics
Dataset
#Nodes
3
NetPHY
37 · 10
Epinions
75 · 103
DBLP
655 · 103
Orkut
3 · 106
Twitter [22] 41.7 · 106
Friendster 65.6 · 106
#Edges Avg. Degree
181 · 103
841 · 103
2 · 106
234 · 106
1.5 · 109
3.6 · 109
9.8
22.4
6.1
78.0
70.5
109.6
Datasets. We use 6 real-world datasets from [23], [22] with
size ranging from tens of thousands to as large as 65.6 million
nodes and 3.6 billion edges. Table II gives a summary.
Algorithms compared. On influence estimation, we compare our SKIS sketch with:
• RIS [5]: The well-known RIS sketch.
• SKIM [2]: Combined reachability sketch. We run SKIM
with default parameters in [2] (k = l = 64). SKIM is
modified to read graph from files instead of internally
computing the edge weights.
Following [3], we generate samples into SKIS and RIS until
the total size of all the samples reaches h · n log n where h is
a constant. Here, h is chosen in the set {5, 10}.
On influence maximization, we compare:
• PMC [24]: A Monte-Carlo simulation pruned method
with no guarantees. It only works on the IC model.
• IMM [7]: RIS-based algorithm with quality guarantees.
• DSSA [8]: The current fastest RIS-based algorithm with
approximation guarantee.
• DSSA+SKIS: A modified version of DSSA where SKIS
sketch is adopted to replace RIS.
We set = 0.5, δ = 1/n for the last three algorithms. For
PMC, we use the default parameter of 200 DAGs.
Metrics. We compare the algorithms in terms of the solution
quality, running time and memory usage. To compute the
solution quality of a seed set, we adopt the relative difference
|Î(S)−I(S)|
· 100%, where Î(S) is an
which is defined as max{I(S),
Î(S)}
estimate, and I(S) is the “ground-truth” influence of S.
Ground-truth Influence. Unlike previous studies [7], [2],
[24] using a constant number of cascade simulations, i.e.
10000, to measure the ground-truth influence with unknown
accuracy, we adopt the Monte-Carlo Stopping-Rule algorithm
[25] that guarantees an estimation error less than with probability at least 1 − δ where = 0.005, δ = 1/n. Specifically, let
W
Wj be the size of a random influence cascade and Zj = nj
with E[Zj ] = I(S)/n and 0 ≤ Zj ≤ 1. The Monte-Carlo
PT
method generates sample Zj until j=1 Zj ≥ 4(e−2) ln( 2δ ) 12
PT
Zj
and returns Î(S) = j=1
n as the ground-truth influence.
T
For Twitter and Friendster dataset, we set = 0.05, and δ =
1/n to compute ground-truth due to the huge computational
cost in these networks. For the other networks, we keep the
default setting of and δ as specified above.
Weight Settings. We consider two widely-used models:
• Weighted Cascade (WC) [6], [2], [7], [8]: The weight of
edge (u, v) is inversely proportional to the in-degree of
node v, din (v), i.e. w(u, v) = din1(v) .
• Trivalency (TRI) [2], [26], [27]: The weight w(u, v) is
selected randomly from the set {0.1, 0.01, 0.001}.
Environment. We implemented our algorithms in C++ and
obtained the implementations of others from the corresponding
authors. We conducted all experiments on a CentOS machine
with Intel Xeon E5-2650 v3 2.30GHz CPUs and 256GB RAM.
We compute the ground-truth for our experiments in a period
of 2 months on a cluster of 16 CentOS machines, each with
64 Intel Xeon CPUs X5650 2.67GHz and 256GB RAM.
TABLE III: Average relative differences (dnf: “did not finish”
within 24h). SKIS almost always returns the lowest errors.
WC Model
SKIS
RIS
TRI Model
SKIM
SKIS
RIS
SKIM
|S| Nets h(5) h(10) h(5) h(10) k(64) h(5) h(10) h(5) h(10) k(64)
l
PHY 6.2 3.7 14.0
Epin. 4.7 3.0 15.7
DBLP 3.8 4.1 13.7
Orkut 10.3 9.2 13.5
Twit. 10.9 10.5 21.4
Frien. 15.9 10.2 22.2
7.8 7.5 1.7 1.3 11.8
11.8 19.6 16.6 14.2 55.3
11.6 5.0 0.9 0.7 9.4
8.8 77.6 9.3 9.9 14.5
16.0 29.1 81.4 81.9 80.8
13.3 dnf 29.8 21.3 28.5
8.2 4.5
47.4 27.7
6.4 3.5
10.8 dnf
81.5 dnf
23.6 dnf
PHY
Epin.
2 DBLP
10
Orkut
Twit.
Frien.
0.9
1.0
0.9
0.9
1.1
0.9
0.6
0.7
0.6
0.6
1.2
0.7
1.0
1.0
1.9
1.1
1.3
0.9
0.7 2.1 0.3
1.0 7.6 0.2
1.4 5.0 0.8
0.7 56.5 0.1
1.1 60.2 4.3
0.7 dnf 1.9
0.2
1.5
0.7
0.2
3.1
1.9
1.1
4.4
5.5
4.2
6.4
0.6
0.9
1.8
5.3
0.9
5.5
2.0
1.8
2.8
5.5
dnf
dnf
dnf
PHY
Epin.
3 DBLP
10
Orkut
Twit.
Frien.
0.6
0.6
0.2
0.3
0.9
0.3
0.8
0.6
0.3
0.3
0.9
0.3
0.9
0.6
0.2
0.3
1.0
0.3
1.0 0.6 0.3
0.7 2.3 2.3
0.2 1.7 0.1
0.3 50.7 2.5
0.9 36.3 0.9
0.2 dnf 1.9
0.4
0.3
0.0
1.1
2.4
1.9
1.2
1.9
0.3
6.8
4.1
0.6
1.3
4.6
0.2
2.1
2.8
2.0
1.1
1.5
0.3
dnf
dnf
dnf
B. Influence Estimation
We show that SKIS sketch consumes much less time and
memory space while consistently obtaining better solution
quality, i.e. very small errors, than both RIS and SKIM.
1) Solution Quality: Table III and Figure 2 present the
relative estimation errors of all three sketches.
The solution quality of SKIS is consistently better than RIS
and SKIM across all the networks and edge models. As shown
in Table III, the errors of SKIS are 110% and 400% smaller
than those of RIS with k = 1 while being as good as or better
TABLE IV: Sketch construction time and index memory of
algorithms on different edge models. SKIS and RIS uses
roughly the same time and memory and less than that of SKIM.
Index Time
[second (or h for hour)]
SKIS
RIS
SKIM
Index Memory
[MB (or G for GB)]
SKIS
RIS
SKIM
M Nets h(5)h(10) h(5)h(10) k(64) h(5)h(10) h(5)h(10) k(64)
PHY
0
1
1
1
2 41
83
Epin.
1
1
1
1
10 63 126
DBLP 10
18
7
14
37 702 1G
WC
Orkut 92 157 69 148 0.6h 2G 5G
Twit. 0.6h 0.9h 0.4h 1.0h 5.2h 38G 76G
Frien. 0.8h 1.8h 0.8h 1.9h dnf 59G 117G
52 105 105
81 162 220
848 2G 2G
3G 5G 9G
42G 84G 44G
61G 117G dnf
PHY
0
1
1
2
Epin.
1
1
1
1
DBLP 11
34 18
36
TRI
Orkut 88 206 89 197
Twit. 0.6h 1.2h 0.5h 1.3h
Frien. 0.9h 2.3h 1.0h 2.4h
97 194
41
84
2G 5G
2G 4G
36G 69G
54G 108G
1
29
22
dnf
dnf
dnf
46
90
41
82
1G 2G
2G 4G
36G 69G
54G 108G
99
230
2G
dnf
dnf
dnf
than RIS for k = 100, 1000. On the other hand, SKIM shows
the largest estimation errors in most of the cases. Particularly,
SKIM’s error is more than 60 times higher than SKIS and
RIS on Twitter when |S| = 100. Similar results are observed
under TRI model. Exceptionally, on Twitter and Friendster,
the relative difference of RIS is slightly smaller than SKIS
with h = 5 but larger on h = 10. In TRI model, estimating
a random seed on large network as Twitter produces higher
errors since we have insufficient number of samples.
Figures 2b, c, and d draw the error distributions of sketches
for estimating the influences of random seeds. Here, we
generate 1000 uniformly random nodes and consider each
node to be a seed set. We observe that SKIS’s errors are
highly concentrated around 0% even when the influences are
small while errors of RIS and SKIM spread out widely. RIS
reveals extremely high errors for small influence estimation,
e.g. up to 80%. The error distribution of SKIM is the most
widely behaved, i.e. having high errors at every influence level.
Under TRI model (Figure 2a), SKIS also consistently provides
significantly smaller estimation errors than RIS and SKIM.
2) Performance: We report indexing time and memory of
different sketches in Table IV.
Indexing Time. SKIS and RIS use roughly the same
amount of time for build the sketches while SKIM is much
slower than SKIS and RIS and failed to process large networks
in both edge models. On larger networks, SKIS is slightly
faster than RIS. SKIM markedly spends up to 5 hours to build
sketch for Twitter on WC model while SKIS, or RIS spends
only 1 hour or less on this network.
Index Memory. In terms of memory, the same observations
are seen as with indexing time essentially because larger
sketches require more time to construct. In all the experiments,
SKIS consumes the same or less amount of memory with RIS.
SKIM generally uses more memory than SKIS and RIS.
In summary, SKIS consistently achieves better solution
quality than both RIS and SKIM on all the networks, edge
47.4
60
27.7
40
16.6
20
14.2
4.4
2.8
1.8
1.5
0.2
1.9
1.5
4.6
2.3 0.3
0
SKIM
50
0
-50
100
Relative difference (%)
55.3
100
Relative difference (%)
80
Relative difference (%)
Relative difference (%)
100
50
0
-50
SKIS h=5
SKIS h=10
-100
SKRIS SKRIS SKIS
SKIS
(h=5) (h=10) (h=5) (h=10)
(a) TRI model
1
50
0
-50
SKRIS h=5
SKRIS h=10
-100
10
100
Ground-truth influence
(b) SKIS-WC model
1
SKIM
-100
10
100
Ground-truth influence
1
(c) RIS-WC model
10
100
Ground-truth influence
(d) SKIM-WC model
0.1
0
SKIS (1K)
SKIS (10K)
SKIS (100K)
1
SKRIS (1K)
SKRIS (10K)
SKRIS (100K)
200 400 600 800
Number of seeds (k)
1000
(a) Epinions
1.5
0.05
0
1
200
5
5
Expected Influence (x10 )
0.2
0.1
Expected Influence (x10 )
0.3
Expected Influence (x105)
Expected Influence (x105)
Fig. 2: a) Relative difference on Epinions under TRI model and b), c), d) error distributions under WC model with |S| = 1.
SKIS has the lowest relative errors which highly concentrates around 0 while RIS’s and SKIM’s errors widely spread out.
1
0.5
0
200 400 600 800 1000
Number of seeds (k)
1
(b) NetPHY
200 400 600 800
Number of seeds (k)
100
0
1000
1
(c) DBLP
200 400 600 800
Number of seeds (k)
1000
(d) Twitter
Fig. 3: Efficiency of SKIS and RIS sketches in finding the maximum seed sets. SKIS sketch is up to 80% more efficient.
IMM
PMC
DSSA+SKIS
DSSA
Running time (s)
2
10
101
100
102
101
100
-1
10
-1
10
1
5000
10000 15000
Number of seeds (k)
20000
1
(a) Epinions
5000
10000 15000
Number of seeds (k)
20000
(b) DBLP
Fig. 4: Running time of algorithms under the IC model.
DSSA+SKIS
DSSA
IMM
DSSA+SKIS
101
Running time (s)
This subsection illustrates the advantage of IIS sketch in
finding the seed set with maximum influence. The results show
that IIS samples drastically speed up the computation time.
DSSA+SKIS is the first to handle billion-scale networks on
the challenging TRI edge model. We limit the running time
for algorithms to 6 hours and put “dnf” if they cannot finish.
1) Identifiability of the Maximum Seed Sets: We compare
the ability of the new IIS with the traditional RIS sampling in
terms of identifying the seed set with maximum influence.
We fix the number of samples generated to be in the set
{1000, 10000, 100000} and then apply the Greedy algorithm
to find solutions. We recompute the influence of returned
seed sets using Monte-Carlo method with precision parameters
= 0.005, δ = 1/n. The results is presented in Figure 3.
From Figure 3, we observe a recurrent consistency that IIS
samples return a better solution than RIS over all the networks,
k values and number of samples. Particularly, the solutions
provided by IIS achieve up to 80% better than those returned
by RIS. When more samples are used, the gap gets smaller.
2) Efficiency of SKIS on IM problem: Table V presents the
results of DSSA-SKIS, DSSA, IMM and PMC in terms of
running time, memory consumption and samples generated.
Running Time. From Table V, the combination
DSSA+SKIS outperforms the rest by significant margins on
all datasets and edge models. DSSA-SKIS is up to 10x faster
than the original DSSA. DSSA+SKIS is the first and only
algorithm that can run on the largest network on TRI model.
Running time (s)
C. Influence Maximization
IMM
PMC
DSSA+SKIS
DSSA
Running time (s)
models and seed set sizes while consuming the same or
less time/memory. The errors of SKIS is highly concentrated
around 0. In contrast, RIS is only good for estimating high
influence while incurring significant errors for small ranges.
100
-1
10
-2
10
1
5000
10000 15000
Number of seeds (k)
(a) Epinions
20000
DSSA
IMM
102
101
100
10-1
1
5000
10000 15000
Number of seeds (k)
20000
(b) DBLP
Fig. 5: Running time of algorithms under the LT model.
Figure 4 compares the running time of all IM algorithms
across a wide range of budget k = 1..20000 under IC and TRI
edge weight model. DSSA+SKIS always maintains significant
performance gaps to the other algorithms, e.g. 10x faster than
DSSA or 1000x faster than IMM and PMC.
Number of Samples and Memory Usage. On the same
line with the running time, the memory usage and number
of samples generated by DSSA+SKIS are much less than
those required by the other algorithms. The number of samples
generated by DSSA+SKIS is up to more 10x smaller than
DSSA on TRI model, 100x less than IMM. Since the memory
for storing the graph is counted into the total memory, the
memory saved by DSSA+SKIS is only several times smaller
than those of DSSA and IMM. PMC exceptionally requires
TABLE V: Performance of IM algorithms with k = 100 (dnf: “did not finish” within 6h, mem: “out of memory”).
Nets
Running Time [s (or h)]
Total Memory [M (or G)]
Expected Influence (%)
IMM PMC DSSA DSSA
+SKIS
IMM PMC DSSA DSSA
+SKIS
IMM PMC DSSA DSSA
+SKIS
#Samples [×103 ]
IMM DSSA DSSA
+SKIS
PHY
0.1
3.1
Epin.
0.2 10.5
DBLP 1.1 137.4
WC
Orkut 24.1 1.4h
Twit. 67.3 mem
Frien. dnf mem
0.0
0.0
0.1
2.6
5.5
78.3
0.0
0.0
0.1
0.9
6.3
43.6
31
86
39 130
162
60
4G
6G
30G mem
dnf mem
26
34
136
2G
17G
35G
9
17
113
2G
16G
36G
6.64
6.7
19.4 19.8
10.8 11.2
6.7
8.7
25.80 mem
dnf mem
5.33
17.9
9.3
5.7
24.1
0.35
5.34
16.6
8.5
5.1
21.0
0.35
103.3
39.8
93.0
174.4
54.0
mem
8.9
4.48
5.4
11.52
18.0
215.0
3.8
0.9
2.6
2.6
0.8
102.4
PHY
0.2
1.5
Epin. 13.9
6.9
DBLP 3.2 20.1
TRI
Orkut dnf 0.3h
Twit.
dnf mem
Frien. dnf mem
0.0
2.0
0.3
1.3h
5.2h
mem
0.0
0.6
0.2
0.2h
0.6h
3.1h
50
61
483
40
389
54
dnf 16G
dnf mem
dnf mem
30
72
191
28G
100G
mem
9
33
118
11G
28G
99G
1.77 1.73
5.7
5.9
0.32 0.31
dnf 67.3
dnf mem
dnf mem
1.4
5.47
0.28
67.9
24.2
mem
1.5
5.46
0.24
67.8
24.4
40.1
370.1
123.0
3171.0
dnf
dnf
dnf
35.8
8.9
348.2
1.4
3.4
mem
3.8
0.5
20.5
0.3
0.4
0.2
huge memory and is unable to run on two large networks.
Experiments on the Linear Threshold (LT) model. We
carry another set of experiments on the LT model with multiple
budget k. Since in LT, the total weights of incoming edge
to every node are bounded by 1, for each node, we first
normalized the weights of incoming edges and then multiply
them with a random number uniformly generated in [0, 1].
The results are illustrated in Figure 5. Similar observations
to the IC are seen in the LT model that DSSA+SKIS runs
faster than the others by orders of magnitude.
Overall, DSSA+SKIS reveals significant improvements
over the state-of-the-art algorithms on influence maximization.
As a result, DSSA+SKIS is the only algorithm that can handle
the largest networks under different models.
VIII. C ONCLUSION
We propose SKIS - a novel sketching tools to approximate
influence dynamics in the networks. We provide both comprehensive theoretical and empirical analysis to demonstrate the
superiority in size-quality trade-off of SKIS in comparisons
to the existing sketches. The application of SKIS to existing
algorithms on Influence Maximization leads to significant
performance boost and easily scale to billion-scale networks.
In future, we plan to extend SKIS to other settings including
evolving networks and time-based influence dynamics.
R EFERENCES
[1] B. Lucier, J. Oren, and Y. Singer, “Influence at scale: Distributed
computation of complex contagion in networks,” in KDD. ACM, 2015,
pp. 735–744.
[2] E. Cohen, D. Delling, T. Pajor, and R. F. Werneck, “Sketch-based
influence maximization and computation: Scaling up with guarantees,”
in CIKM. ACM, 2014, pp. 629–638.
[3] N. Ohsaka, T. Akiba, Y. Yoshida, and K.-i. Kawarabayashi, “Dynamic
influence analysis in evolving networks,” VLDB, vol. 9, no. 12, pp. 1077–
1088, 2016.
[4] D. Kempe, J. Kleinberg, and É. Tardos, “Maximizing the spread of
influence through a social network,” in KDD, 2003, pp. 137–146.
[5] C. Borgs, M. Brautbar, J. Chayes, and B. Lucier, “Maximizing social
influence in nearly optimal time,” in SODA. SIAM, 2014, pp. 946–957.
[6] Y. Tang, X. Xiao, and Y. Shi, “Influence maximization: Near-optimal
time complexity meets practical efficiency,” in SIGMOD. ACM, 2014,
pp. 75–86.
[7] Y. Tang, Y. Shi, and X. Xiao, “Influence maximization in near-linear
time: A martingale approach,” in SIGMOD, 2015, pp. 1539–1554.
[8] H. T. Nguyen, M. T. Thai, and T. N. Dinh, “Stop-and-stare: Optimal
sampling algorithms for viral marketing in billion-scale networks,” in
SIGMOD. New York, NY, USA: ACM, 2016, pp. 695–710.
[9] H. Tong, B. A. Prakash, T. Eliassi-Rad, M. Faloutsos, and C. Faloutsos,
“Gelling, and melting, large graphs by edge manipulation,” in CIKM.
ACM, 2012, pp. 245–254.
[10] S. R. A. Arora, S. Galhotra, “Debunking the myths of influence
maximization: Anin-depth benchmarking study,” in SIGMOD. ACM,
2017, pp. 75–86.
[11] E. Cohen and H. Kaplan, “Summarizing data using bottom-k sketches,”
in PODC. ACM, 2007, pp. 225–234.
[12] G. Cormode and S. Muthukrishnan, “An improved data stream summary:
the count-min sketch and its applications,” Journal of Algorithms,
vol. 55, no. 1, pp. 58–75, 2005.
[13] N. Du, L. Song, M. Gomez-Rodriguez, and H. Zha, “Scalable influence
estimation in continuous-time diffusion networks,” in NIPS, 2013, pp.
3147–3155.
[14] D. T. Nguyen, H. Zhang, S. Das, M. T. Thai, and T. N. Dinh, “Least
cost influence in multiplex social networks: Model representation and
analysis,” in ICDM. IEEE, 2013, pp. 567–576.
[15] T. N. Shen, Y.and Dinh, H. Zhang, and M. T. Thai, “Interest-matching
information propagation in multiple online social networks,” in CIKM.
ACM, 2012, pp. 1824–1828.
[16] H. T. Nguyen, P. Ghosh, M. L. Mayo, and T. N. Dinh, “Multiple infection
sources identification with provable guarantees,” in CIKM. ACM, 2016,
pp. 1663–1672.
[17] H. T. Nguyen, T. P. Nguyen, T. N. Vu, and T. N. Dinh, “Outward
influence and cascade size estimation in billion-scale networks,” in
SIGMETRICS. ACM, 2017, pp. 63–63.
[18] “Importance sketching of influence dynamics in billion-scale networks,”
https://www.dropbox.com/s/ssoq6ecngqky1v2/icdm17 sketch.pdf?dl=0
[19] H. T. Nguyen, M. T. Thai, and T. N. Dinh, “Cost-aware targeted viral
marketing in billion-scale networks,” in INFOCOM. IEEE, 2016, pp.
1–9.
[20] ——, “A billion-scale approximation algorithm for maximizing benefit
in viral marketing,” IEEE/ACM Transactions on Networking, vol. 25,
no. 4, pp. 2419–2429, 2017.
[21] G. Nemhauser and L. Wolsey, “Maximizing submodular set functions:
formulations and analysis of algorithms,” North-Holland Mathematics
Studies, vol. 59, pp. 279–301, 1981.
[22] H. Kwak, C. Lee, H. Park, and S. Moon, “What is twitter, a social
network or a news media?” in WWW. ACM, 2010, pp. 591–600.
[23] SNAP, http://snap.stanford.edu, 2017, stanford network analysis project.
[24] N. Ohsaka, T. Akiba, Y. Yoshida, and K.-i. Kawarabayashi, “Fast and
accurate influence maximization on large networks with pruned montecarlo simulations,” in AAAI, 2014.
[25] P. Dagum, R. Karp, M. Luby, and S. Ross, “An optimal algorithm for
monte carlo estimation,” SICOMP, pp. 1484–1496, 2000.
[26] W. Chen, C. Wang, and Y. Wang, “Scalable influence maximization for
prevalent viral marketing in large-scale social networks,” in KDD. New
York, NY, USA: ACM, 2010, pp. 1029–1038.
[27] K. Jung, W. Heo, and W. Chen, “Irie: Scalable and robust influence
maximization in social networks,” in ICDM, 2012, pp. 918–923.
[28] V. Vazirani, Approximation Algorithms. Springer, 2001.
where Rj (v) is a random IIS sketch with src(Rj (v)) = v.
Plugging this back into the computation of I(S) gives,
X
X
I(S) =
Pr[Rj (v) ∩ S 6= ∅]γv +
(1 − γv )
v∈V
=
v∈V
A PPENDIX
=
v∈V g∈ΩG
v∈V
ηg (S, v) Pr[g] +
g∈Ω∅
G (v)
g∈Ω̄∅
G (v)
Since our IIS sketching algorithm only generates samples
corresponding to sample graphs from the set Ω̄∅G (v), we define
Ω̄∅G (v) to be a graph sample space in which the sample graph
G]
Pr[ḡ ∈ Ω̄∅G (v)] = Pr[ḡ∈Ω
ḡ ∈ Ω̄∅G (v) has a probability
γv
P
of being realized (since
Pr[ḡ ∈ ΩG ] = γv is the
ḡ∈Ω̄∅
G (v)
normalizing factor to fulfill a probability distribution of a
sample space). Then, Eq. 32 is rewritten as follows,
X
X X
Pr[g ∈ ΩG ]
γv +
(1 − γv )
I(S) =
ηg (S, v)
γv
v∈S
v∈V g∈Ω̄∅ (v)
G
X X
X
=
ηḡ (S, v) Pr[ḡ ∈ Ω̄∅G (v)]γv +
(1 − γv )
v∈S
Now, from the node v in a sample graph ḡ ∈ Ω̄∅G (v), we
have a IIS sketch Rj (ḡ, v) starting from v and containing all
the nodes that can reach v in ḡ. Thus, ηḡ (S, v) = 1Rj (ḡ,v)∩S6=∅
where 1x is an indicator function returning 1 iff x 6= 0. Then,
X
ηḡ (S, v) Pr[ḡ ∈ Ω̄∅G (v)]
ḡ∈Ω̄∅
G (v)
=
X
ḡ∈Ω̄∅
G (v)
X
(1 − γv )
v∈S
X
(1 − γv )
(33)
v∈S
That completes the proof.
B. Proof of Lemma 4
Var[Xj (S)]
P
P
I(S) − v∈S (1 − γv )
I(S) − v∈S (1 − γv )
=
(1 −
)
Γ
Γ
P
2
X
(1 − γv )
I(S) I (S)
=
−
− v∈S 2
(Γ +
(1 − γv ) − 2I(S))
2
Γ
Γ
Γ
v∈S
Put this back into the variance of Zj (S) proves the lemma.
C. Proof of Lemma 5
P
v∈V g∈Ω̄∅ (v)
G
(32)
v∈V ḡ∈Ω̄∅ (v)
G
Pr[Rj (v) ∩ S 6= ∅] Pr[src(Rj ) = v]Γ +
From the basic properties of variance, we have,
P
Xj (S) · Γ + v∈S (1 − γv )
Var[Zj (S)] = Var[
]
n
2
Γ
= 2 Var[Xj (S)]
n
Since Xj (S) is a Bernoulli
random variable with its mean
P
value E[Xj (S)] = I(S)− v∈S (1−γv ) , the variance Var[Xj (S)]
X
ηg (S, v) Pr[g] . is computed as follows, Γ
In each g ∈ Ω∅G (v), the node v does not have any incoming
nodes,
thus, ηg (S, v) = 1 only if v ∈ P
S. Thus,
P
P
P we have that
ηg (S, v) Pr[g] =
Pr[g].
v∈V
g∈Ω∅
(v)
v∈S
g∈Ω∅
G
G (v)
Furthermore, the probability
of
a
sample
graph
which
has
no
P
incoming live-edge to v is g∈Ω∅ (v) Pr[g] = 1−γv . Combine
G
with the above equiation of I(S), we obtain,
X
X X
I(S) =
(1 − γv ) +
ηg (S, v) Pr[g ∈ ΩG ].
v∈S
v∈S
= Pr[Rj ∩ S 6= ∅] · Γ +
Given a stochastic graph G, recall that ΩG is the space of all
possible sample graphs g ∼ G and Pr[g] is the probability that
g is realized from G. In a sample graph g ∈ ΩG , ηg (S, v) = 1
if v is reachable from S in g. Consider the graph sample space
ΩG , based on a node v ∈ V \S, we can divide ΩG into two
partitions: 1) Ω∅G (v) contains those samples g in which v has
no incoming live-edges; and 2) Ω̄∅G (v) = ΩG \Ω∅G . We start
from the definition of influence spread as follows,
X X
I(S) =
ηg (S, v) Pr[g]
X X
X
v∈S
X
γv
Pr[Rj (v) ∩ S 6= ∅] Γ +
(1 − γv )
Γ
v∈V
A. Proof of Lemma 3
=
X
1Rj (ḡ,v)∩S6=∅ Pr[ḡ ∈ Ω̄∅G (v)] = Pr[Rj (v) ∩ S 6= ∅]
(1−γ )
v
v∈S
Since
Zj (S) takes values of either
or
P
n
Γ+ v∈S (1−γv )
I(S)
and
the
mean
value
E[Z
(S)]
=
,
i.e.
j
P
n
P
n
Γ+ v∈S (1−γv )
I(S)
v∈S (1−γv )
≤
≤
.
The
variance
of
Zj (S)
n
n
n
is computed as follows,
Var[Zj (S)]
I(S) P
Γ + P
I(S)
v∈S (1 − γv )
v∈S (1 − γv )
−
−
=
n
n
n
P n
P
I(S) Γ + v∈S (1 − γv )
v∈S (1 − γv )
≤
−
n
n
n
I(S) Γ
=
(34)
n n
D. Proof of Lemma 6
Lemma 2 in [7] states that:
Lemma 10. Let M1 , M2 , . . . be a martingale, such that
|M1 | ≤ a, |Mj − Mj−1 | ≤ a for any j ∈ [2, T ], and
T
X
Var[M1 ] +
Var[Mj |M1 , M2 , . . . , Mj−1 ] ≤ b,
(35)
j=2
where Var[.] denotes the variances of a random variable.
Then, for any λ > 0,
λ2
Pr[MT − E[MT ] ≥ λ] ≤ exp − 2
(36)
3 aλ + 2b
Note that uniform random variables are also a special type
of martingale and the above lemma holds for random variable
as well. Let p = I(S)
n . For RIS samples, since
• |M1 | ≤ 1,
• |Mj − Mj−1 | ≤ 1, ∀j ∈ [2, T ],
PT
• Var[M1 ]
+
=
j=2 Var[Mj |M1 , . . . , Mj−1 ]
Pi
Var[Y
(S)]
=
T
p(1
−
p)
≤
T
p,
j
j=1
applying Eq. 36 for λ = T p gives the following Chernoff’s
bounds,
T
hX
i
2
Pr
Xj (S) − T p ≥ T p ≤ exp −
2 T p , (37)
2 + 3
j=1
and,
T
hX
i
2
Pr
Xj (S) − T p ≤ −T p ≤ exp − T p .
2
j=1
(38)
However, for IIS samples in SKIS sketch, the corresponding
random variables Zj (S) replace Yj (S) and have the following
properties:
P
•
•
•
Γ+
(1−γ )
v
v∈S
≤ 1,
|M1 | ≤
n
|Mj − Mj−1 | ≤ 1, ∀j ∈ [2, T ],
The sum of variances:
T
X
Var[M1 ] +
Var[Mj |M1 , . . . , Mj−1 ]
j=2
=
i
X
Var[Zj (S)] = T p
j=1
Γ
n
(39)
Thus, applying the general bound in Eq. 36 gives,
T
i
hX
2
Xj (S) − T p ≥ T p ≤ exp − Γ 2 T p ,
Pr
2n + 3
j=1
(40)
and,
T
hX
i
2
Pr
Xj (S) − T p ≤ −T p ≤ exp − Γ T p .
2n
j=1
(41)
Note the factor Γn is added in the denominator of the terms in
the exp(.) function. Since 2 Γn dominates 32 , the concentration
bounds for Zj (S) for SKIS are tighter than those of Yj (S)
for RIS given in Eqs. 37 and 38.
E. Proof of Lemma 8
Since the function ÎR (S) contains two additive terms, it
is sufficient to show that each P
of them is monotone and
submodular. The second term
v∈S (1 − γv ) is a linear
function and thus, it is monotone and submodular. For the
Γ
is a constant and only
first additive term, we see that |R|·n
need to show that CR (S) is monotone and submodular. Given
the collection of IIS samples R in which Rj ∈ R is a list of
nodes, the function CR (S) is just the count of IIS samples
that intersect with the set S. In other words, it is equivalent
to a covering function in a set system where IIS samples
are elements and nodes are sets. A set covers an element
if the corresponding node is contained in the corresponding
IIS sample. It is well known that any covering function is
monotone and submodular [28] and thus, the CR (S) has the
same properties.
F. Improvements of SSA/DSSA using SKIS sketch
Recall that the original Stop-and-Stare strategy in [8] uses
two independent sets of RIS samples, called R and Rc .
The greedy algorithm is applied on the first set R to find
a candidate set Ŝk along with an estimate ÎR (Ŝk ) and the
second set Rc is used to reestimate the influence of Ŝk by
ÎRc (Ŝk ). Now, SSA and DSSA have different ways to check
the solution quality.
SSA. It assumes a set of fixed precision parameters 1 , 2 , 3
+2 +1 2 +3
(1 − 1/e) ≤ . The algorithm stops
such that 1(1+
1 )(1+2 )
when two conditions are met:
δ
−2
1) CR (Ŝk ) ≥ Λ1 where Λ1 = O(log tmax
3 ) and tmax is
a precomputed number depending on the size of the input
graph G.
2) ÎR (Ŝk ) ≤ (1 + 1 )ÎRc (Ŝk ).
⇒ Improvements using SKIS: Replacing RIS samples by
IIS samples to build R and Rc helps in both stopping
conditions:
δ
Γ
−2 ) using the tighter
• Reduce Λ1 to Λ1 = O( n log t
max 3
form of the Chernoff’s bounds in Lemma 6.
• Since IIS samples have better influence estimation accuracy, ÎR (Ŝk ) and ÎRc (Ŝk ) are closer to the true influence
I(Ŝk ). Thus, the second condition is met earlier than
using RIS samples.
DSSA. Instead of assuming precision parameters, DSSA
dynamically compute the error bounds 1 , 2 and 3 as follows:
•
1 =
•
2 =
•
3 =
ÎR (Ŝk )
− 1.
ÎR
qc (Ŝk )
n(1+)
.
2t−1
ÎRc (Ŝk )
q
n(1+)(1−1/e−)
(1+/3)2t−1 Î c (Ŝ ) .
k
R
Here, 1 measures the discrepancy of estimations using two
different sketches R and Rc while 2 and 3 are the error
bounds of estimating the influences of Ŝk and the optimal
solution Sk∗ using the number of samples contained in R and
Rc . The algorithm stops when two conditions are met:
δ
• CR (Ŝk ) ≥ Λ2 where Λ2 = O(log t
−2 ).
max
• (1 + 2 + 1 2 )(1 − 1/e − ) + (1 − 1/e)3 ≤ .
⇒ Improvement using SKIS: Similarly to SSA, applying
SKIS helps in both stopping conditions:
Γ
δ
• Reduce Λ2 to Λ2 = O( n log t
−2 ).
max
• Reduce the value of 1 , 2 and 3 due to better influence
estimations of ÎR (Ŝk ) and ÎRc (Ŝk ) by SKIS that leads
to earlier satisfaction of the second condition.
| 8cs.DS
|
REAL CLASS SIZES
arXiv:1803.01717v1 [math.GR] 1 Mar 2018
HUNG P. TONG-VIET
Abstract. In this paper, we study the structures of finite groups using some arithmetic conditions on the sizes of real conjugacy classes. We prove that a finite group
is solvable if the prime graph on the real class sizes of the group is disconnected.
Moreover, we show that if the sizes of all non-central real conjugacy classes of a
finite group G have the same 2-part and the Sylow 2-subgroup of G satisfies certain
condition, then G is solvable.
1. Introduction
Let G be a finite group. An element x ∈ G is said to be real if there exists an
element g ∈ G such that xg = x−1 . We denote by Re(G) the set of all real elements
of G. A conjugacy class xG containing x ∈ G is said to be real if x is a real element of
G or equivalently xG = (x−1 )G . The size of a real conjugacy class is called a real class
size. Several arithmetic properties of the real class sizes can be conveniently stated
using graph theoretic language. The prime graph on the real class sizes of a finite
group G, denoted by ∆∗ (G), is a simple graph with vertex set ρ∗ (G) the set of primes
dividing the size of some real conjugacy class of G and there is an edge between two
vertices p and q if and only if the product pq divides some real class size.
Now a prime p is not a vertex of ∆∗ (G), that is, p 6∈ ρ∗ (G) if and only if p divides
no real class size of G. In [6], the authors show that 2 is not a vertex of ∆∗ (G) if and
only if G has a normal Sylow 2-subgroup S (i.e., G is 2-closed) and Re(S) ⊆ Z(S).
For odd primes, a similar result is not that satisfactory. Combining results in [11],
[14] and [15], we can show that if an odd prime p is not a vertex of ∆∗ (G) and assume
′
further that when p = 3, SL3 (2) is not a composition factor of G, then O2 (G) has a
′
normal Sylow p-subgroup and Op (G) is solvable, in particular, G is p-solvable (see
Lemma 2.7). The proofs of the aforementioned results, especially, for odd primes,
are quite involved and depend heavily on the classification of finite simple groups.
This is in contrast to the similar result for all conjugacy classes, that is, if a prime p
does not divide the size of any conjugacy classes of G then G has a normal central
Sylow p-subgroup. The proof of this classical result is just an application of Jordan’s
theorem on the existence of derangements in finite permutation groups.
It is proved in [6] that ∆∗ (G) has at most two connected components. In our first
result, we will show that if ∆∗ (G) is disconnected, then G is solvable.
Date: March 6, 2018.
2010 Mathematics Subject Classification. Primary 20E45; Secondary 20D10.
Key words and phrases. Real conjugacy classes, real class sizes, prime graphs.
1
2
H. P. TONG-VIET
Theorem A. Let G be a finite group. If the prime graph on the real class sizes of G
is disconnected, then G is solvable.
We next study in more detail the real class sizes of finite groups with disconnected
prime graph on real class sizes.
Theorem B. Let G be a finite group. Suppose that ∆∗ (G) is disconnected. Then 2
divides some real class size and one of the following holds.
(1) G has a normal Sylow 2-subgroup.
′
′
(2) ∆∗ (O2 (G)) is disconnected and the real class sizes of O2 (G) are either odd
or powers of 2.
It follows from Theorem B that if the prime graph ∆∗ (G) of a finite group G
is disconnected, then 2 must be a vertex of ∆∗ (G). This confirms once again the
importance of the prime 2 in the study of real conjugacy classes of finite groups. In
′
the second conclusion of Theorem B, both connected components of ∆∗ (O2 (G)) are
complete and one of the components of this graph contains the prime 2 only. (See
Theorem 3.5). We should mention that it was proved in [7] that a finite group whose
all real class sizes are either odd or powers of 2 is solvable. However, in the proof
of (2), we will need the solvability from Theorem A. So, if one can prove part (2) of
Theorem B without using the solvability of the group, then we would have another
proof of Theorem A.
In [7], it is shown that if all non-central real conjugacy classes of a group have prime
sizes, then the group has a normal Sylow 2-subgroup or a normal 2-complement. The
next example shows that this is not the case if we only assume that all real class sizes
are prime powers.
′
Example. Let G = Alt4 : C4 be a solvable group of order 48. We have G = O2 (G),
G/Z(G) ∼
= Sym3 and the real class sizes of G are 1, 3 or 8. Clearly,
= Sym4 , G/O2 (G) ∼
G has no normal Sylow 2-subgroup nor normal 2-complement.
It follows from [1] that if the prime graph defined on all class sizes of a finite
group G is disconnected, then G/Z(G) is a Frobenius group with abelian kernel and
complement. By our example above, this does not hold for the prime graph on real
class sizes.
In our last result, we provide further evidence for a conjecture proposed in [16].
We will prove Conjecture C in [16] under some condition on the Sylow 2-subgroups.
Theorem C. Let G be a finite group. Suppose that the sizes of all non-central real
conjugacy classes of G have the same 2-part. Assume further that G has a Sylow
′
2-subgroup S with Re(S) ⊆ Z(S). Then G is solvable and O2 (G) has a normal
2-complement.
A conjecture due to G. Navarro states that a finite group G is solvable if G has
at most two real class sizes. Clearly, our Theorem C implies this conjecture with an
additional assumption that Re(S) ⊆ Z(S) for some Sylow 2-subgroup S of G. Finite
3
2-groups S with Re(S) ⊆ Z(S) have been studied by Chillag and Mann [2]. These are
exactly the finite 2-groups S for which if x, y ∈ S and x2 = y 2, then xZ(S) = yZ(S).
2. Real conjugacy classes
Our notation are more or less standard. If n is a positive integer, then π(n) is the
set of prime divisors of n. If π(n) ⊆ σ for some set of primes σ, then n is said to be
a σ-number. If n > 1 is an integer and p is a prime, then the p-part of n, denoted by
np , is the largest power of p dividing n.
Recall that Re(G) is the set of all real elements of G. We collect some properties
of real elements and real class sizes in the following lemma.
Lemma 2.1. Let G be a finite group and let N ✂ G.
If x ∈ Re(G), then every power of x is also real.
If x ∈ Re(G), then xt = x−1 for some 2-element t ∈ G.
If x ∈ Re(G) and |xG | is odd, then x2 = 1.
If x, y ∈ Re(G), xy = yx and (|xG |, |y G |) = 1, then xy ∈ Re(G). Furthermore, if (o(x), o(y)) = 1, then π(|xG |) ∪ π(|y G |) ⊆ π(|(xy)G |).
(5) If |G : N| is odd, then Re(G) = Re(N).
(6) Suppose that Nx is a real element in G/N. If |N| or the order of Nx in
G/N is odd, then Nx = Ny for some real element y ∈ G (of odd order if the
order of Nx is odd).
(1)
(2)
(3)
(4)
Proof. Let x ∈ G be a real element. Then xg = x−1 for some g ∈ G. If k is
any integer, then (xk )g = (xg )k = (x−1 )k = (xk )−1 so xk is real which proves (1).
Write o(g) = 2a m with (2, m) = 1 and let t = g m. Then t is a 2-element and
m
xt = xg = xg = x−1 as g 2 ∈ CG (x) and m is odd. This proves (2).
Parts (3)–(5) are in Lemma 6.3 of [6]. Finally, (6) is Lemma 2.2 in [11].
Fix 1 6= x ∈ Re(G), set C∗G (x) = {g ∈ G | xg ∈ {x, x−1 }}. Then C∗G (x) is a
2
subgroup of G containing CG (x). If g ∈ G such that xg = x−1 , then xg = x so
g 2 ∈ CG (x). Assume that x is not an involution. We see that g ∈ C∗G (x) \ CG (x)
and if h ∈ C∗G (x) \ CG (x), then xh = x−1 = xg so that hg −1 ∈ CG (x) or equivalently
h ∈ CG (x)g. Thus CG (x) has index 2 in C∗G (x) and hence |xG | is even. In particular,
this is the case if x is a nontrivial real element of odd order.
The next lemma is well-known. We will use this lemma freely without further
reference.
Lemma 2.2. Let G be a finite group and let N ✂ G. Then
(1) If x ∈ N, then |xN | divides |xG |.
(2) If Nx ∈ G/N, then |(Nx)G/N | divides |xG |.
The following lemma shows that a finite group G has no nontrivial real element of
odd order if and only if G has a normal Sylow 2-subgroup.
4
H. P. TONG-VIET
Lemma 2.3. ([6, Proposition 6.4]). The following are equivalent:
(1) Every nontrivial element in Re(G) has even order.
(2) Every element in Re(G) is a 2-element.
(3) G has a normal Sylow 2-subgroup.
The next lemma determines the number of connected components of the prime
graphs on real class sizes.
Lemma 2.4. ([6, Theorem 6.2]). For any finite group G, ∆∗ (G) has at most two
connected components.
If a finite group G is of even order, then it has a real element of order 2. If an odd
prime p dividing |G|, G may not have a real element of order p. However, if G has no
proper normal subgroup of odd index and G is p-solvable, Dolfi, Malle and Navarro
[5] show that G must contain a real element of order p.
′
Lemma 2.5. ([5, Corollary B]). Let G be a finite group with O2 (G) = G. Suppose
that p is an odd prime dividing |G|. If G is p-solvable, then G has a real element of
order p.
In the next two lemmas, we state the Itô-Michler theorem for real conjugacy classes.
Lemma 2.6. ([6, Theorem 6.1]). Let G be a finite group and let P be a Sylow 2subgroup of G. Then all real classes of G have odd size if and only if P ✂ G and
Re(P ) ⊆ Z(P ).
Lemma 2.7. Let G be a finite group and p be an odd prime. If p = 3, assume
in addition that G has no composition factor isomorphic to SL3 (2). If p does not
′
divide |xG | for all real elements of G, then G is p-solvable and Op (G) is solvable.
′
′
Furthermore, O2 (G) has a normal Sylow p-subgroup P and P ′ ≤ Z(O2 (G)).
Proof. The first claim is Theorem A in [11]. Now, Theorem B in that reference implies
that p does not divide χ(1) for all real-valued irreducible characters χ ∈ Irr(G). By
′
Theorem A in [15], we know that Op (G) is solvable. Finally, the last statement
follows from Theorem A in [14].
3. Proofs of Theorems A and B
Let G be a finite group. Suppose that ∆∗ (G) is disconnected. Then ∆∗ (G) has exactly two connected components by Lemma 2.4. The following will be used frequently
in our proofs.
Lemma 3.1. Let G be a finite group and suppose that ∆∗ (G) has two connected
components with vertex sets π1 and π2 , where 2 6∈ π2 . Then there exists an involution
i ∈ G such that |iG | > 1 is a π2 -number and CG (i) has a normal Sylow 2-subgroup.
Proof. Let p be a prime in π2 . Then p must divide |iG | for some nontrivial real
element i ∈ G. Clearly, every prime divisor of |iG | is adjacent to p ∈ π2 , this implies
5
that |iG | is a nontrivial π2 -number. Hence |iG | > 1 is odd and so i2 = 1 by Lemma
2.1(3) and thus i is an involution of G.
Assume that CG (i) has a nontrivial real element x of odd order. Then xi = ix and
|xG | is even so |xG | is a π1 -number and thus (|xG |, |iG |) = 1. Lemma 2.1(4) implies
that xi is a real element. Furthermore, since (o(x), o(i)) = 1, 2p divides |(ix)G | by
Lemma 2.1(4) again. This means that 2 ∈ π1 and p ∈ π2 are adjacent in ∆∗ (G),
which is impossible. Therefore, CG (i) has no nontrivial real element of odd order
and thus it has a normal Sylow 2-subgroup by Lemma 2.3.
Notice that if N ✂ G, then ∆∗ (N) is a subgraph of ∆∗ (G) by Lemma 2.2(1) and
the fact that Re(N) ⊆ Re(G). However, in general, it is not true that ∆∗ (G/N) is a
subgraph of ∆∗ (G). The involutions in G/N might produce extra vertices as well as
edges in ∆∗ (G/N). However, this is the case if |N| is odd.
Lemma 3.2. Let G be a finite group and let N ✂ G with |N| odd. Then ∆∗ (G/N)
is a subgraph of ∆∗ (G).
Proof. We first show that ρ∗ (G/N) ⊆ ρ∗ (G). Indeed, let p ∈ ρ∗ (G/N) and let
Nx ∈ G/N be a real element such that p divides |(Nx)G/N |. By Lemma 2.1(6),
there exists a real element y ∈ G such that Nx = Ny. Since |(Nx)G/N | = |(Ny)G/N |
divides |y G |, p divides |y G |, so p ∈ ρ∗ (G).
With a similar argument, we can show that if p 6= q ∈ ρ∗ (G/N) which are adjacent
in ∆∗ (G/N), then p, q are adjacent in ∆∗ (G) by using Lemma 2.1(6) again. Thus
∆∗ (G/N) is a subgraph of ∆∗ (G).
Let X be a subgroup or a quotient of a finite group G and suppose that ∆∗ (X)
is a subgraph of ∆∗ (G). Assume that ∆∗ (G) is disconnected having two connected
components with vertex sets π1 and π2 , respectively. To show that ∆∗ (X) is disconnected, it suffices to show that ρ∗ (X) ∩ πi 6= ∅ for i = 1, 2 or equivalently X has two
real elements ui , i = 1, 2 which both lift to real elements of G and π(|uX
i |) ∩ πi 6= ∅
for i = 1, 2.
For a finite group G and a prime p, G is said to be p-closed if it has a normal Sylow
p-subgroup and it is p-nilpotent if it has a normal p-complement.
Proposition 3.3. Let G be a finite group. Suppose that ∆∗ (G) has two connected
components with vertex sets π1 and π2 where 2 6∈ π2 . Then
(1) If G is not 2-closed and M ✂G with |G : M| odd, then ∆∗ (M) is disconnected.
′
(2) If N ✂G with |N| odd and assume further that G = O2 (G) is not 2-nilpotent,
then ∆∗ (G/N) is disconnected.
Proof. By Lemma 3.1, G has an involution i such that |iG | > 1 is a π2 -number and
CG (i) has a normal Sylow 2-subgroup S.
For (1), let M ✂ G with |G : M| being odd. Notice that ρ∗ (M) ⊆ ρ∗ (G) = π1 ∪ π2
and that if M is 2-closed, then G is also 2-closed. Thus we assume that M is not 2closed. By Lemma 2.6, 2 divides some real class size of M and hence 2 ∈ ρ∗ (M) ∩ π1 .
6
H. P. TONG-VIET
Now M contains every real element of G by Lemma 2.1(5), in particular, i ∈ M.
Moreover, ∆∗ (M) is a subgraph of ∆∗ (G). If M ≤ CG (i), then S ✂ M as |G : M| is
odd, so M is 2-closed, a contradiction. Thus, |iM | > 1 and hence ρ∗ (M) ∩ π2 6= ∅.
Therefore, ∆∗ (M) is disconnected as ρ∗ (M) ∩ πi is non-empty for each i = 1, 2.
For (2), suppose that N ✂ G with |N| odd. By Lemma 3.2, ∆∗ (G/N) is a subgraph
of ∆∗ (G). In particular, ρ∗ (G/N) ⊆ ρ∗ (G) = π1 ∪π2 . If G/N is 2-closed, then SN ✂G
′
is of odd index and thus G = SN since G = O2 (G). However, this would imply that
G is 2-nilpotent with a normal 2-complement N. Therefore, we assume that G/N is
not 2-closed.
Clearly, Ni ∈ G/N is an involution. If Ni is central in G/N, then G/N =
CG/N (Ni) = CG (i)N/N, where the latter equality follows from [12, Lemma 7.7].
Hence G/N has a normal Sylow 2-subgroup SN/N, a contradiction. Thus Ni is not
central in G/N and |(Ni)G/N | > 1, so ρ∗ (G/N) ∩ π2 6= ∅. Finally, as G/N is not
2-closed, 2 ∈ ρ∗ (G/N) ∩ π1 by applying Lemma 2.6 again. Therefore, ∆∗ (G/N) is
disconnected.
We are now ready to prove Theorem A which we restate here.
Theorem 3.4. Let G be a finite group. If ∆∗ (G) is disconnected, then G is solvable.
Proof. Let G be a counterexample with minimal order. Then G is non-solvable and
∆∗ (G) is disconnected. By Lemma 2.4, ∆∗ (G) has exactly two connected components
with vertex sets π1 and π2 , respectively. If G has a normal Sylow 2-subgroup, then it
is clearly solvable by Feit-Thompson theorem. Therefore, we can assume that G has
no normal Sylow 2-subgroup. Now it follows from Lemma 2.3 that G has a nontrivial
real element x of odd order. Then |xG | is divisible by 2 and hence 2 is always a vertex
of ∆∗ (G). We assume that 2 ∈ π1 . Hence all vertices in π2 are odd primes.
(1) By Lemma 3.1, G has an involution i such that |iG | > 1 is a π2 -number and
CG (i) has a normal Sylow 2-subgroup, say S. Clearly, S is also a Sylow 2-subgroup of
G as |iG | is odd. Notice that CG (i) is solvable. Now G has a nontrivial real element x
of odd order by Lemma 2.3. Clearly |xG | is even so |xG | must be a π1 -number. Thus
(|xG |, |iG |) = 1; therefore, G is not a nonabelian simple group by [8, Theorem 2].
′
′
(2) G = O2 (G). Since G is not 2-closed, ∆∗ (O2 (G)) is disconnected by Proposition
′
′
3.3(1). If O2 (G) < G, then O2 (G) is solvable by the minimality of |G|, hence G is
′
′
solvable since G/O2 (G) is solvable. This contradiction shows that G = O2 (G).
′
(3) O2′ (G) = 1. By (2) above, we have G = O2 (G) and since G is not solvable,
G is not 2-nilpotent so that by Proposition 3.3(2) ∆∗ (G) is disconnected, where
G = G/O2′ (G). If O2′ (G) is nontrivial, then |G| < |G| and thus by the minimality
of |G|, G is solvable and so is G. Hence O2′ (G) = 1 as required.
(4) If M is a maximal normal subgroup of G, then |G : M| = 2 and G = Mhii.
Let M be a maximal normal subgroup of G. Then G/M is a simple group.
7
(a) Assume first that G/M is abelian, then G/M ∼
= Cr for some prime r. Since
2′
O (G) = G by (2), we deduce that r = 2. Clearly, M is not solvable. We next claim
that i 6∈ M and hence we have that G = Mhii.
Suppose that by contradiction that i ∈ M. Then either |iM | > 1 or M = CG (i). If
the latter case holds, then M is solvable by (1) and thus G is solvable, a contradiction.
Assume that |iM | > 1. Notice that ∆∗ (M) is a subgraph of ∆∗ (G). We see that
ρ∗ (M) ∩ π2 6= ∅ as every prime divisor of |iM | is in π2 and i ∈ Re(M). Observe next
that M is not 2-closed so it has a nontrivial real element z of odd order by Lemma
2.3 and thus |z M | is even. In other words, 2 ∈ ρ∗ (M). Therefore ρ∗ (M) ∩ π1 6= ∅.
Hence we have shown that ∆∗ (M) is disconnected. Thus by induction, M is solvable,
which is a contradiction.
(b) Now, assume that G/M is a non-abelian simple group. Set G = G/M.
G
Assume first that i 6∈ M. Then i is an involution in G and |i | divides |iG |. Let
y be a nontrivial real element of G of odd order (such an element exists by Lemma
2.3). By Lemma 2.1(6), y lifts to a real element z ∈ G of odd order. Therefore |y G |
G
divides |z G |. Since (|z G |, |iG |) = 1, we deduce that (|yG |, |i |) = 1, contradicting [8,
Theorem 2].
Assume that i ∈ M. If G = MCG (i), then G/M ∼
= CG (i)/(M ∩ CG (i)) is a
non-abelian simple group, which is impossible as CG (i) is solvable by (1). Thus
H := MCG (i) < G and |G : H| = |G : H| divides |G : CG (i)| = |iG |. Let y ∈ G
be a real element of odd order and z ∈ G be a real element of odd order such that
y = z. Then |yG | = |z G | divides |z G | so (|G : CG (z)|, |G : H|) = 1 as (|z G |, |iG |) = 1.
Therefore, G = HCG (z), where H has odd index in G and H has a normal Sylow
2-subgroup S.
As every nontrivial real element of odd order of G lifts to a nontrivial real element
of odd order of G by Lemma 2.1(6), every prime divisor s of |G : H| (which lies in
π2 ) divides the size of no nontrivial real element of odd order of G, so G ∼
= PSL2 (q)
r
∼
with q = 2 − 1 a Mersenne prime and s | (q − 1)/2 or G = M23 with s = 5 by [11,
Theorem 4.1]. If the latter case holds, then |G : H| must be a power of 5. Inspecting
[3] shows that this is not the case. Thus G ∼
= PSL2 (q) with q = 2r − 1 a Mersenne
prime. It follows that r ≥ 3 and q ≥ 7 as G is non-solvable. Observe that S is a
Sylow 2-subgroup of G of order q + 1 = 2r ≥ 8 and so S ∼
= Dq+1 , which is a maximal
subgroup of G unless q = 7. Suppose first that q > 7. It implies that H = S so that
|G : H| = q(q − 1)/2. In particular, q divides |G : H| and thus q | (q − 1)/2 by the
claim above, which is impossible. Now assume that q = 7. Notice that the Sylow
2-subgroup of G ∼
= PSL2 (7) is self-normalizing but not maximal. Since S ✂ H, we
must have that H = S and we will get a contradiction as above.
(5) G = G′ hii and |G : G′ | = 2.
8
H. P. TONG-VIET
Since G is not non-abelian simple, G possesses a maximal normal subgroup W . It
follows from (4) that G = W hii and |G : W | = 2. It also follows from (4) that i does
not lie in any maximal normal subgroup of G so that G = hiG i.
Clearly G′ ≤ W as G/W ∼
= C2 is abelian. Moreover, G = hiG i ≤ G′ hii ≤ W hii =
G. It follows that W = G′ and the claim follows.
′
(6) O2 (G′ ) is a π1 -group.
Since |G : G′ | = 2 and G is non-solvable, G′ is non-solvable and thus ∆∗ (G′ ) is a
connected subgraph of ∆∗ (G) with 2 ∈ ρ∗ (G′ ) ⊆ π1 . Observe that π(G) = π(G′ ).
Let σ = π(G′ ) \ π1 . Now, if q ∈ σ, then q is odd and divides the size of no nontrivial
′
real conjugacy classes of G′ . Let K = O2 (G′ ).
If q ∈ σ and q > 3 or q = 3 and SL3 (2) is not a composition factor of G′ , then K
has a normal Sylow q-subgroup Q by Lemma 2.7. Clearly Q ✂ G and thus Q = 1 by
(3). Hence q does not divide |K|.
We now suppose that q = 3 and SL3 (2) is isomorphic to a composition factor of G′ .
As G′ /K is solvable, SL3 (2) is isomorphic to a composition factor of K. Let L be a
subnormal subgroup of K and U ✂ L such that L/U ∼
= SL3 (2). Using [3], we see that
L/U has a self-normalizing Sylow 2-subgroup T /U and a real element Uz ∈ L/U of
order 3 with |(Uz)L/U | = 7 · 8. There exists a real element y ∈ L of 3-power order
with Uz = Uy (see [16, Lemma 2.6]). Since |(Uz)L/U | divides |y L | and |y L | divides
|y G |, we see that 7 ∈ π1 .
By (1), CG (i) has a normal Sylow 2-subgroup S with S ∈ Syl2 (G). As L is
subnormal in G, S ∩ L is a Sylow 2-subgroup of L and thus (S ∩ L)U/U is a Sylow
2-subgroup of L/U. Observe that CG (i) contains a Sylow 7-subgroup, say Q, of G.
Hence (Q ∩ L)U/U is a Sylow 7-subgroup of L/U. Since Q normalizes S, we can see
that (Q∩L)U/U normalizes (S ∩L)U/U, which is impossible as the Sylow 2-subgroup
′
of L/U is self-normalizing. Thus we have shown that K = O2 (G′ ) is a π1 -group.
The final contradiction. By (5), we have G = G′ hii so CG (i) = CG′ (i)hii which
implies that |iG | = |G : CG (i)| = |G′ : CG′ (i)|. Since |iG | is a π2 -number and CG (i)
is solvable, CG′ (i) is also solvable and possesses a Hall π1 -subgroup T which is also
′
a Hall π1 -subgroup of G′ (as |G′ : CG′ (i)| is a π2 -number). As O2 (G′ ) ✂ G′ is a
′
′
π1 -subgroup by (6), we deduce that O2 (G′ ) ≤ T and thus O2 (G′ ) is solvable since
′
T ≤ CG′ (i) is solvable. Clearly, G′ /O2 (G′ ) is solvable by Feit-Thompson theorem,
which implies that G′ is solvable and hence G is solvable. This contradiction finally
proves the theorem.
′
In the next result, we show that if G = O2 (G) and ∆∗ (G) is disconnected, then
each connected component of ∆∗ (G) is complete and one of the components is just
{2}. In particular, every real class size of G is either odd or a 2-power.
′
Theorem 3.5. Let G be a finite group. Suppose that G = O2 (G) and ∆∗ (G) is
disconnected with vertex sets π1 and π2 where 2 6∈ π2 . Then π1 = {2} and π2 = π(|iG |)
for some non-central involution i ∈ G.
9
Proof. By Theorem 3.4, we know that G is solvable. Let i ∈ G be an involution as
in Lemma 3.1 and let S be a normal Sylow 2-subgroup of CG (i). Let σ be the set
of odd prime divisors p of |CG (i)| such that p does not divide |iG |. Since |iG | is a
π2 -number and |G| = |iG | · |CG (i)|, we see that
π1 \ {2} ⊆ π(G) \ (π2 ∪ {2}) ⊆ σ = π(G) \ ({2} ∪ π(|iG |)).
′
Assume that σ = ∅. Then π1 ⊆ {2} and π2 ⊆ π(|iG |). As G = O2 (G) and
∗
∆ (G) is disconnected, 2 divides some real class size of G, hence 2 ∈ π1 . Moreover
π(|iG |) ⊆ π2 ; therefore π1 = {2} and π(|iG |) = π2 as wanted.
Assume next that σ is not empty and let p ∈ σ. Then p is odd and p divides
|CG (i)| but does not divide |iG |. By Lemma 2.5, G has a real element x of order
p. Let P be a Sylow p-subgroup of CG (i). Since |G : CG (i)| = |iG | is not divisible
by p, P is also a Sylow p-subgroup of G. Replacing x by its G-conjugates, we can
assume x ∈ P ≤ CG (i) by Sylow theorem. We have that xi = ix, (o(x), o(i)) = 1
and (|xG |, |iG |) = 1 so that by applying Lemma 2.1(4), xi is real in G and
π(|xG |) ∪ π(|iG |) ⊆ π(|(xi)G |).
As |iG | > 1, we can find a prime r dividing |iG | and so r ∈ π2 . Now the previous
inclusion would imply that 2 ∈ π(|xG |) ⊆ π1 and r ∈ π2 are adjacent in ∆∗ (G), which
is impossible.
We now consider the situation when 2 is not a vertex of ∆∗ (G), where G is a finite
group. By Lemma 2.6, G has a normal Sylow 2-subgroup S with Re(S) ⊆ Z(S).
Lemma 3.6. Let G be a finite group. Suppose that G has a normal Sylow 2-subgroup
S with Re(S) ⊆ Z(S). Then ∆∗ (G) is connected.
Proof. If ∆∗ (G) has at most one vertex, then we are done. So, assume that ∆∗ (G)
has at least two vertices, i.e., |ρ∗ (G)| ≥ 2.
We argue by contradiction. Suppose that ∆∗ (G) is not connected. Then ∆∗ (G)
has two connected components with vertex sets π1 and π2 . Then we can find two
non-central real elements x, y ∈ Re(G) such that π(|xG |) ⊆ π1 and π(|y G|) ⊆ π2 so
(|xG |, |y G|) = 1.
Since S ✂ G and Re(S) ⊆ Z(S), Re(G) ⊆ Re(S) ⊆ Z(S) by using Lemma 2.1(5).
Thus all nontrivial real elements of G are involutions in Z(S). Let E be the set of
all involutions of Z(S) together with the identity. Then E is an elementary abelian
subgroup of Z(S). Indeed, E = Ω1 (Z(S)) and thus E ✂ G.
Clearly x, y ∈ E so that E is not central. In particular, CG (E) ✂ G is a proper
subgroup of G. Notice that E ≤ Z(S) and hence S ≤ CG (E). Therefore A :=
G/CG (E) is a nontrivial group of odd order and we can consider E as an F2 Amodule. By Maschke’s theorem, E is a completely reducible A-module. Observe
that |xA | = |A : CA (x)| = |G : CG (x)| = |xG | and |y A | = |y G |. Therefore, the two
A-orbits xA and y A have coprime sizes. By Theorem 1.1 in [4], the A-orbit (xy)A
has size |xA | · |y A |. Hence |(xy)G | = |xG | · |y G |, where xy ∈ E is an involution. This
10
H. P. TONG-VIET
implies that there is an edge between a prime in π1 and a prime in π2 , which is
impossible.
We are now ready to prove Theorem B.
Theorem 3.7. Let G be a finite group. Suppose that ∆∗ (G) is disconnected. Then
2 divides some real class size and one of the following holds.
(1) G has a normal Sylow 2-subgroup.
′
′
(2) ∆∗ (O2 (G)) is disconnected and the real class sizes of O2 (G) are either odd
or powers of 2.
Proof. Suppose that ∆∗ (G) is disconnected and let the vertex sets of the connected
components are π1 and π2 , respectively. Assume that 2 6∈ π2 . We see that ∆∗ (G) has
at least two vertices.
We first claim that 2 ∈ π1 . It suffices to show that 2 divides some real class size of
G. Suppose by contradiction that 2 divides no real class size. Then G has a normal
Sylow 2-subgroup S with Re(S) ⊆ Z(S) by Lemma 2.6. However, Lemma 3.6 implies
that ∆∗ (G) is connected, which is a contradiction.
Next, suppose that G has no normal Sylow 2-subgroup. We claim that part (2)
′
of the conclusion holds. Let K := O2 (G). By Proposition 3.3(1), ∆∗ (K) is dis′
connected. Since O2 (K) = K and ∆∗ (K) is disconnected, the result follows from
Theorem 3.5.
We suspect that if a finite group G has a normal Sylow 2-subgroup S, then ∆∗ (G)
is connected, that is, case (1) in Theorem 3.7 cannot occur. However, we are unable
to prove or disprove this yet. In view of Lemma 3.6, this is true if Re(S) ⊆ Z(S).
There are many examples of finite groups whose prime graphs on real class sizes
are disconnected. For the first example, let m > 1 be an odd integer; the dihedral
group D2n of order 2n, where n = m or n = 2m, has a disconnected prime graph on
real class sizes as its real class sizes are just 1, 2 and m. For another example, let
G be a Frobenius group with Frobenius kernel F and complement H, where both F
and H are abelian and |H| is even. In this case, all the nontrivial real class sizes of
G are |F | and |H| and since (|F |, |H|) = 1, ∆∗ (G) is disconnected.
4. Proof of Theorem C
Let G be a finite group. Observe that if x ∈ Z(G) is a real element of G, then
x2 = 1. We first begin with the following lemma.
Lemma 4.1. Let G be a finite group and let S ∈ Syl2 (G). Suppose that Re(S) ⊆ Z(S)
and |xG |2 = 2a ≥ 2 for all non-central real elements x ∈ G. If y is a nontrivial real
element of G whose order is a 2-power, then y is a central involution of G.
Proof. Let y be a nontrivial real element of G whose order is a 2-power. Then y t = y −1
for some 2-element t ∈ G by Lemma 2.1(2). As t normalizes hyi, U := hy, ti is a
11
2-subgroup of G. By Sylow theorem, U g ≤ S for some g ∈ G. If y g is a central
involution of G, then so is y. Thus we can assume that U ≤ S. Since y ∈ Re(U),
we have y ∈ Re(S) ⊆ Z(S) and so y 2 = 1. Hence y is an involution. Finally, since
y ∈ Z(S), |y G| is odd which forces y ∈ Z(G) as by assumption |xG | is even for all
non-central real elements x of G.
The following lemma is obvious.
Lemma 4.2. Let G be a finite group and let S ∈ Syl2 (G). Suppose that |xG |2 = 2a ≥
2 for all non-central real elements x ∈ G. Let K ✂ G be a normal subgroup of odd
index. Then |xK |2 = 2a for all non-central real elements x ∈ K.
Proof. Let K be a normal subgroup of G of odd index. By Lemma 2.1(5), we have
Re(G) ⊆ Re(K). Now let x ∈ K be a non-central real element of K and let C :=
CG (x). Let P ∈ Syl2 (C) and let S ∈ Syl2 (G) such that P ≤ S. Since |G : K| is odd,
we have P ≤ S ≤ K. In particular, S ∈ Syl2 (K). We see that CK (x) = K ∩ C and
P ≤ K ∩ C. Thus P ≤ CK (x) ≤ C and so P is also a Sylow 2-subgroup of CK (x).
Therefore |C|2 = |CK (x)|2 and hence |xG |2 = |G : C|2 = |S : P | = |K : CK (x)|2 =
|xK |2 .
Lemma 4.3. Let G be a finite group and let S ∈ Syl2 (G). Suppose that |xG |2 =
2a ≥ 2 for all non-central real elements x ∈ G. Assume further that G has a Sylow
2-subgroup S with Re(S) ⊆ Z(S). Then every nontrivial real element of G/O2′ (G)
of 2-power order lies in the center of G/O2′ (G).
Proof. Let N = O2′ (G)✂G and let Nx be a real element of G/N of order k := 2c ≥ 2.
Then Nx = Ny for some real element y ∈ G by Lemma 2.1(6). We see that y k ∈ N
and so (y k )m = 1 for some odd integer m ≥ 1. As (k, m) = 1, 1 = uk + vm for some
integers u, v. We have Nx = Ny = Ny uk+vm = (Ny uk )(Ny vm ) = Ny vm as y k ∈ N.
Clearly z := y vm is a nontrivial real element of G whose order divides k = 2c and
Nx = Ny = Nz. By Lemma 4.1, z is a central involution of G and thus Nx is also
a central involution of G/N.
In the next theorem, we show that if a finite group satisfies the hypothesis of
Theorem C, then it is solvable.
Theorem 4.4. Let G be a finite group. Suppose that |xG |2 = 2a for all non-central
real elements x ∈ G. Assume further that G has a Sylow 2-subgroup S with Re(S) ⊆
Z(S). Then G is solvable.
Proof. Let G be a minimal counterexample to the theorem and let S be a Sylow 2subgroup of G with Re(S) ⊆ Z(S). Then G is non-solvable and thus G has no normal
Sylow 2-subgroup. By Lemma 2.3, G has a nontrivial real element z of odd order.
Clearly, z is not central and thus |z G | is always even. Therefore, |z G |2 = 2a ≥ 2.
It follows from Lemma 4.3 that every nontrivial real element of G/O2′ (G) of 2power order lies in the center of G/O2′ (G). In particular, all involutions of G/O2′ (G)
are in the center of G/O2′ (G). Now we can apply results in [10]. Since O2′ (G/O2′ (G)) =
12
H. P. TONG-VIET
1, by the main theorem in [10], the last term of the derived series of G/O2′ (G), say
H/O2′ (G) is isomorphic to a direct product L1 × L2 × · · · × Ln , where each Li is
isomorphic to either SL2 (q) with q ≥ 5 odd or 2 · Alt7 , the perfect double cover of
Alt7 .
For each i, every real element of Li is also a real element of G/O2′ (G) as Li is
a subgroup of G/O2′ (G). Moreover, every nontrivial real element of Li of 2-power
order must lie in the center of G/O2′ (G) and hence must be in Z(Li ). Thus to obtain
a contradiction, we need to find a real element x ∈ Li of order 2m ≥ 4. Notice that
|Z(Li )| = 2 for all i ≥ 1.
Assume first that Li ∼
= 2 · Alt7 for some i ≥ 1. Using [3], Li has a real element x
of order 4. Assume next that Li ∼
= SL2 (q) for some q ≥ 5 odd. It is well known that
the Sylow 2-subgroup T of SL2 (q) is a generalized quaternion group of oder 2k+1 for
some k ≥ 2. (See, for example, Theorem 2.8.3 in [9]). Now T is generated by two
k−1
elements α and β such that o(β) = 2k , o(α) = 4, α2 = β 2
and β α = β −1 . Thus β
is a real element of SL2 (q) of order 2k ≥ 4. We can take x = β. The proof is now
complete.
We now prove the 2-nilpotence part of Theorem C.
Theorem 4.5. Let G be a finite group. Suppose that |xG |2 = 2a for all non-central
real elements x ∈ G. Assume further that G has a Sylow 2-subgroup S with Re(S) ⊆
′
Z(S). Then O2 (G) is 2-nilpotent.
′
Proof. By Lemma 4.2, we can assume that G = O2 (G). Let G = G/O2′ (G) and use
the ‘bar’ notation. Now G is solvable by Theorem 4.4. Let P = O2 (G). Since G
is solvable, it possesses a Hall 2′ -subgroup, say H. It follows from Lemma 4.3 that
every real element of G whose order is a power of 2 lies in the center of G. This
implies that H centralizes all real elements of order at most 4 of P and thus by [13,
Theorem B], H centralizes P . By Hall-Higman 1.2.3, H ≤ CG (P ) ≤ P which forces
H = 1. This means that G = P is a 2-group and so G is 2-nilpotent as required.
Finally, Theorem C follows by combining Theorems 4.4 and 4.5.
References
[1] Bertram, Edward A.; Herzog, Marcel; Mann, Avinoam. On a graph related to conjugacy
classes of groups. Bull. London Math. Soc. 22 (1990), no. 6, 569–575.
[2] Chillag, David; Mann, Avinoam. Nearly odd-order and nearly real finite groups. Comm. Algebra 26 (1998), no. 7, 2041–2064.
[3] Conway, J. H.; Curtis, R. T.; Norton, S. P.; Parker, R. A.; Wilson, R. A. Atlas of finite groups.
Maximal subgroups and ordinary characters for simple groups. With computational assistance
from J. G. Thackray. Oxford University Press, Eynsham, 1985.
[4] Dolfi, Silvio; Guralnick, Robert; Praeger, Cheryl E.; Spiga, Pablo. Coprime subdegrees for
primitive permutation groups and completely reducible linear groups. Israel J. Math. 195
(2013), no. 2, 745–772.
[5] Dolfi, Silvio; Malle, Gunter; Navarro, Gabriel. The finite groups with no real p-elements. Israel
J. Math. 192 (2012), no. 2, 831–840.
13
[6] Dolfi, Silvio; Navarro, Gabriel; Tiep, Pham Huu. Primes dividing the degrees of the real
characters. Math. Z. 259 (2008), no. 4, 755–774.
[7] Dolfi, Silvio; Pacifici, Emanuele; Sanus, Lucia. Finite groups with real conjugacy classes of
prime size. Israel J. Math. 175 (2010), 179–189.
[8] Fisman, Elsa; Arad, Zvi. A proof of Szep’s conjecture on nonsimplicity of certain finite groups.
J. Algebra 108 (1987), no. 2, 340–354.
[9] Gorenstein, Daniel. Finite groups. Second edition. Chelsea Publishing Co., New York, 1980.
[10] Griess, Robert L., Jr. Finite groups whose involutions lie in the center. Quart. J. Math. Oxford
Ser. (2) 29 (1978), no. 115, 241–247.
[11] Guralnick, Robert M.; Navarro, Gabriel; Tiep, Pham Huu. Real class sizes and real character
degrees. Math. Proc. Cambridge Philos. Soc. 150 (2011), no. 1, 47–71.
[12] Isaacs, I. Martin. Finite group theory. Graduate Studies in Mathematics, 92. American Mathematical Society, Providence, RI, 2008.
[13] Isaacs, I. M.; Navarro, Gabriel. Normal p-complements and fixed elements. Arch. Math. (Basel)
95 (2010), no. 3, 207–211.
[14] Isaacs, I. M.; Navarro, Gabriel. Groups whose real irreducible characters have degrees coprime
to p. J. Algebra 356 (2012), 195–206.
[15] Tiep, Pham Huu. Real ordinary characters and real Brauer characters. Trans. Amer. Math.
Soc. 367 (2015), no. 2, 1273–1312.
[16] Tong-Viet, H. P. Groups with some arithmetic conditions on real class sizes. Acta Math.
Hungar. 140 (2013), no. 1-2, 105–116.
Department of Mathematical Sciences, Binghamton University, Binghamton, NY
13902-6000, USA
E-mail address: [email protected]
| 4math.GR
|
Which NP-Hard SAT and CSP Problems Admit Exponentially
Improved Algorithms?
arXiv:1801.09488v1 [cs.DS] 29 Jan 2018
Victor Lagerkvist∗1 and Magnus Wahlström†2
1
2
Institut für Algebra, TU Dresden, Dresden, Germany
Department of Computer Science, Royal Holloway, University of London, Great Britain
Abstract
We study the complexity of SAT(Γ) problems for potentially infinite languages Γ closed under
variable negation, which we refer to as sign-symmetric languages Γ. Via an algebraic connection, this
reduces to the study of restricted partial polymorphisms we refer to as pSDI-operations (for partial,
self-dual and idempotent), under which the language Γ is invariant. First, we focus on the language
classes themselves. We classify the structure of the least restrictive pSDI-operations, corresponding to
the most powerful languages Γ, and find that these operations can be divided into levels, corresponding
to a rough notion of difficulty, where every level k has an easiest language class, containing the language
for (k − 1)-SAT, and a hardest language class, containing (among other things) constraints encoded as
roots of multivariate polynomials of degree (k − 1). Particular classes in each level correspond to the
natural partially defined versions of previously studied total algebraic invariants. In particular, the easiest
class on level k ≥ 3 corresponds to the partial k-ary near-unanimity (k-NU) operation, and a larger
class corresponds to the partial k-edge operation. The largest class at each level corresponds to a partial
operation uk we call k-universal. Furthermore, every sign-symmetric language Γ not preserved by uk
implements all k-clauses, hence SAT(Γ) is at least as hard as k-SAT; and if Γ is not preserved by uk for
any k, then SAT(Γ) is trivially SETH-hard (i.e., takes time O∗ (2n ) under SETH).
Second, we consider implications of this for the complexity of SAT(Γ). We find that particular
classes in the hierarchy correspond to previously known algorithmic strategies. In particular, languages
preseved by the partial 2-edge operation can be solved via Subset Sum-style meet in the middle, and
languages preserved by the partial 3-NU operation can be solved via fast matrix multiplication. These
results also hold for the correspondning non-Boolean CSP problems. We also find that symmetric 3-edge
languages reduce to finding a monochromatic triangle in an edge-coloured graph, which can be done
using algorithms for sparse matrix multiplication; and if the sunflower conjecture holds for sunflowers
with k petals, then the partial k-NU language has an improved algorithm via Schöning-style local search.
Complementing this, we show a lower bound, showing that for every level k there is a constant ck such
that for every partial operation p on level k, the problem SAT(Γ) with Γ = Inv(p) cannot be solved faster
than O∗ (cnk ) unless SETH fails. In particular, when Γ = Inv(2-edge), this gives us the first NP-hard SAT
problem which simultaneously has non-trivial upper and lower bounds on the running time, assuming
SETH. Finally, we note a possible conjecture: It is consistent with our present knowledge that SAT(Γ)
admits an improved algorithm if and only if Γ is preserved by uk for some constant k. However, to show
this in the positive poses some significant difficulty.
∗
†
[email protected]
[email protected]
1 Introduction
Significant attention has been paid to the exact time complexity of SAT and its various restrictions; in
particular CNF-SAT and k-SAT, but also other restrictions such as Not-All-Equal SAT, 1-in-k SAT, and
several more cases [15, 23, 25, 42, 49]. The usual focus is on an improved algorithm for some particular
variant, i.e., showing that the problem can be solved in time O∗ (cn ) for some c < 2, or, in some cases, that
such an improvement is not feasible, up to our current knowledge (i.e., it would require disproving the strong
exponential-time hypothesis, SETH; see below). Here, and in the sequel, the parameter n will in this context
always denote the number of variables in a given instance. But what is the general rule for when a SAT
problem admits such an improved algorithm? And can we say anything at all about lower bounds on such
improvements?
To refine the question, let us recall some terminology. A constraint language is a (possibly infinite) set Γ of
finitary relations R ⊆ D ar(R) over some domain D, where ar(R) denotes the arity of R. We will mainly focus
on the Boolean case, i.e., D = {0, 1}. Then SAT(Γ), occasionally called the parameterized satisfiability
problem, is the SAT problem where the constraints of the instance are applications of relations from Γ, i.e.,
the constraints are statements that R(x1 , . . . , xr ) must hold, for some R ∈ Γ and some variables x1 , . . . , xr
from the variable set (where we do allow repetitions of a variable). The multi-valued generalization of SAT,
the constraint satisfaction problem over Γ (CSP(Γ)) is defined in essentially the same way, except that Γ may
be non-Boolean. Full definitions of the problems under consideration follow in Section 2. Thus, for example,
3-SAT corresponds to SAT(Γ3SAT ) where Γ3SAT for each 3-clause in {(x ∨ y ∨ z), . . . , (¬x ∨ ¬y ∨ ¬z)}
contains the relation excluding only the tuple forbidden by that particular clause. Similarly, for k ≥ 3 let
ΓkSAT denote the constraint language of all k-clauses, i.e., SAT(ΓkSAT ) is equivalent to k-SAT.
Let us also tentatively define c(Γ) as the infimum over all constants c > 1 such that SAT(Γ) can be solved
in O(cn ) on n variables. Then the exponential time hypothesis (ETH), due to Impagliazzo and Paturi, states
that c(ΓkSAT ) > 1 for every k, and was shown to be equivalent to the statement that c(Γ3SAT ) > 1 [25]. It has
also been shown to be equivalent to the statement that c(Γ) > 1 for every Γ such that SAT(Γ) is NP-hard [31].
The strong exponential time hypothesis (SETH) is the statement that limk→∞ c(ΓkSAT ) = 2 [10, 25]. Then our
main research question can be rephrased as, for which constraint languages Γ is c(Γ) < 2, respectively, when
would c(Γ) < 2 contradict SETH? We say that SAT(Γ) allows an improved algorithm in the former case,
and that it is SETH-hard in the latter. Hence, our main interest is in exponential improvements rather than
subexponential improvements of the form O(2n−o(n) ) which have been proven to exist for CNF-SAT [17].
Before we discuss our approach for the general case, we consider a few examples. First of all, the
algorithms for k-SAT imply that c(Γ) < 2 for every finite language Γ. However, such bounds are also known
for some infinite languages. One example is Exact SAT, the language of 1-in-k-clauses of all arities, which
admits an improved algorithm [53]. As has been shown more recently, so does the problem where constraints
are encoded as the roots of bounded-degree multivariate polynomials over a finite field [42]. Thus, we need
a way to discuss properties of infinite arbitrary languages, and we need to consider the representation of
constraints from such a language. We address these issues in Section 1.2.
Lower bounds on c(Γ) for some Γ have been significantly harder to come by. Some SAT problems
have been shown to be SETH-hard, in particular Not-All-Equal SAT and problems related to SAT such as
Hitting Set [15]. It is also known that assuming ETH, the value of c(ΓkSAT ) increases infinitely often [25].
However, we do not even have conjectural evidence against any particular value of c(Γ) for any language
Γ such that SAT(Γ) is not SETH-hard, other than for trivial cases.1 We also are not aware of any previous
attempts to engage with the question of what makes a SAT problem SETH-hard or not in general.
In this paper, we study these questions using tools from universal algebra. It is known that the value
1By trivial cases, we mean problems where the natural search space is smaller than 2n but otherwise unrestricted. Consider a
language where every variable is involved in a disequality, e.g., the language of relations R′ (x1 , . . . , x2k ) ≡ (x1 6= xk+1 ) ∧ . . . ∧
(xk 6= x2k ) ∧ R(x1 , . . . , xk ) for arbitrary relations R. It is easy to see that under SETH, this problem has c(Γ) = 21/2 .
1
of c(Γ) is determined by algebraic invariants of Γ known as partial polymorphisms [31]. It is not difficult
to prove that if Γ has no interesting partial polymorphisms, then SAT(Γ) is trivially SETH-hard. We
study the converse to this question, to essentially ask, does the existence of even a single relevant partial
polymorphism p imply that SAT(Γ) has an improved algorithm? In particular, is it possible to design an
algorithm with an exponentially improved running time, whose correctness depends only on p? One of the
main strengths of using such an algebraic approach is that it makes the task of identifying languages Γ such
that c(Γ) < 2 considerably easier. In fact, as we discuss in Section 1.1, these languages can be succinctly
classified according to the expressive power of individual partial operations.
Our paper has two main contributions. First, we characterize the structure of the weakest non-trivial
invariants p. In this, we restrict ourselves to sign-symmetric languages (see below). This reveals a characterization of problem complexity, with close ties to several previously studied problems and algorithm
classes. Second, we use the framework to provide both upper and lower bounds on c(Γ) for the corresponding
languages Γ, under SETH. We show that algorithms from the literature can be extended to work for every language having a certain partial polymorphism p. In the negative direction, we are able to prove lower bounds
on c(Γ) for every language Γ characterised purely by its invariants. As a result, we produce the first language
Γ such that c(Γ) has both non-trivial upper and lower bounds under SETH. Finally, we make connections
between these SAT(Γ) problems and some problems in polynomial-time fine-grained complexity.
Our approach also implies some results for CSPs on a non-Boolean domain, but our main focus in the
present paper lies in studying the Boolean case.
1.1 Universal algebraic aspects of SAT problems
To make the discussion of our approach more precise, we need to review some notions from universal algebra.
This is simply intended as an introduction and overview to make the extended abstract self-contained; full
definitions follow later in the paper in Section 2. The universal algebraic approach to problem complexity
originates in research into the constraint satisfaction problem (CSP) [29]. Recall the definitions of a constraint
language Γ and the problem CSP(Γ) from the preceding section. Clearly, the complexity of CSP(Γ) varies
as a function of Γ: if Γ is simple enough, then CSP(Γ) is in P; and if Γ is rich enough, then CSP(Γ) is
NP-complete. The dichotomy conjecture, first posed by Feder and Vardi [19], states that these are the only
two cases and that no NP-intermediate CSP problems exist: for every fixed language Γ, CSP(Γ) is either in
P or is NP-complete. This conjecture has been the subject of intense research and the piece remaining to
complete the puzzle was recently resolved by two independent authors [7, 57].
The algebraic approach turned out to be central in this research programme. In short, this approach
boils down to the realization that properties of constraint languages can be expressed by properties of their
polymorphisms. Informally, a polymorphism of a constraint language Γ is an operation which yields a
method to combine satisfying assignments of instances of CSP(Γ). The algebraic reformulation of the
CSP dichotomy theorem then states that CSP(Γ) is tractable if there exists a non-trivial method to combine
solutions, and is NP-complete otherwise. More formally, we may define polymorphisms as follows. First, let
R ⊆ D n be a relation on D, and let p : Dr → D be an r-ary operation over D. We can then generalise p to an
operation (Dn )r → D n on tuples over D by p(x1 , . . . , xr )[i] = p(x1 [i], . . . , xr [i]) for every position i ∈ [n]
(where xj [i] denotes the ith element of the tuple xj ). Then p is a polymorphism of R if this generalised
operation preserves R, i.e., if p(x1 , . . . , xr ) ∈ R for any x1 , . . . , xr ∈ R. Note that if p is a projection, i.e.,
p(t1 , . . . , tr ) = ti for some i ∈ [r], then p preserves every possible relation. The notion of a polymorphism
easily extends to constraint languages, and we say that p is a polymorphism of the constraint language Γ if p
is a polymorphism of R for every relation R ∈ Γ, and let Pol(Γ) denote this set. It is then known that the
complexity of CSP(Γ), up to polynomial-time many-one reductions, is determined entirely by Pol(Γ) [28].
Theorem 1. Let Γ and ∆ be finite constraint languages over a finite domain D. If Pol(∆) ⊆ Pol(Γ), then
CSP(Γ) is polynomial-time many-one reducible to CSP(∆).
2
At this stage this result may seem slightly puzzling since we do not yet have a clear correspondence
between polymorphisms and their implications on constraint languages. However, there exists a dual concept
to polymorphisms on the relational side called implementations. Given a set of relations Γ over a domain D,
a k-ary relation R is definable by a primitive positive implementation over Γ (pp-definable) if there exists a
first-order formula making use of existential quantification and conjunctive constraints over Γ such that the
set of models of this formula is precisely R. Given a constraint language Γ we then let hΓi be the smallest
set of relations containing Γ and which is closed under taking pp-definitions. The polymorphisms of Γ then
characterize the power of pp-definitions over Γ in the following sense.
Theorem 2 ([5, 6, 21]). Let Γ and ∆ be two constraint languages. Then Γ ⊆ h∆i if and only if Pol(∆) ⊆
Pol(Γ).
This duality has two implications. First, note that an instance of CSP(Γ) can be viewed as a special case
of a pp-definition over Γ, hence the polymorphisms of Γ describe closure properties for the whole CSP(Γ)
problem, and can be used to design polynomial-time algorithms. This is in line with the intuition that a
polymorphism yields a method for combining satisfying assignments. Second, if R has a pp-definition in
Γ then there is a polynomial-time many-one reduction from CSP(Γ ∪ {R}) to CSP(Γ); essentially, the ppdefinition describes a classical “gadget reduction” between the problems obtained by replacing constraints
over R by the collection of constraints over Γ prescribed by the pp-definition. Therefore, dually to the
previous point, the absence of sufficiently interesting polymorphisms for Γ would imply a polynomial-time
reduction from an NP-hard problem CSP(Γ′ ), e.g., 3-SAT, to CSP(Γ).
In practice, for CSPs beyond the Boolean domain, the complexity landscape gets very complex and one
needs to apply a richer algebraic toolbox to make progress. However, it was realized early that not only
does the complexity of CSP(Γ) depend on Pol(Γ), but in fact only the identities satisfied by the operations
in Pol(Γ) [9]. In technical terms this means that the complexity of CSP(Γ) only depends on the variety
generated by Pol(Γ). We will not define these concepts formally since they are not needed to present the
main results; it is sufficient to know that the complexity of CSP(Γ) only depends on the identities satisfied
by the operations in Pol(Γ). For example, CSP(Γ) is solvable using k-consistency if Pol(Γ) contains a
majority operation, i.e., a ternary operation m satisfying the identities m(x, y, y) = y, m(y, x, y) = y,
m(y, y, x) = y [29]. Moreover, all operations resulting in tractable CSPs can be characterized using such
identities.
It is worth remarking that for the Boolean domain the situation is considerably simplified due to Post’s
classification of Boolean Pol(Γ) [46], and a large range of such problems have been proven to admit
dichotomies [14]. For example, Schaefers dichotomy theorem for SAT(Γ) [48] can be proven in an extremely
straightforward manner using this approach. However, for our purposes the above methods are too coarsegrained, since the precise running time O∗ (cn ) for a problem SAT(Γ) is not preserved by the introduction
of existentially quantified variables. Hence, we are in need of more fine-grained algebraic tools than usual,
which can be applied as follows.
A partial operation over D (of some arity r) is an operation p : X → D for some domain X ⊆ D r .
Similar to the total case we again extend it to a partial operation on tuples over D: for x1 , . . . , xr ∈ D n , we
let p(x1 , . . . , xr )[i] = p(x1 [i], . . . , xr [i]) if this is defined for every position i ∈ [n]; otherwise p(x1 , . . . , xr )
is undefined. Then p is a partial polymorphism of a relation R ⊆ D n if, for any x1 , . . . , xr ∈ R such that
p(x1 , . . . , xr ) is defined we have p(x1 , . . . , xr ) ∈ R. We will occasionally also say that R is invariant under
the partial operation p. A partial projection is a subfunction of a projection; such an operation preserves
every possible relation. A partial polymorphism of a constraint language Γ is a partial polymorphism of
every relation R ∈ Γ and we let pPol(Γ) denote the set of all partial polymorphisms of Γ. Similarly, given a
set of partial operations P we write Inv(P) to denote the set of relations invariant under P , and if P = {p}
is singleton we write Inv(p) instead of Inv({p}). Dually to this relaxed notion of a polymorphism, we have
a strengthened notion on the relational side: a quantifier-free primitive positive definition (qfpp-definition)
3
over Γ is a pp-definition without existential quantification. We let hΓi6∃ denote the smallest set of relations
containing Γ and which is closed under qfpp-definitions, and then obtain the following correspondence.
Theorem 3 ([21, 47]). Γ ⊆ h∆i6∃ if and only if pPol(∆) ⊆ pPol(Γ) for any constraint languages Γ and ∆.
With the help of this correspondence Jonsson et al. [31] proved that partial polymorphisms indeed can
be used for studying the fine-grained complexity of SAT and CSP.
Theorem 4. Let Γ and ∆ be two finite constraint languages. If pPol(Γ) ⊆ pPol(∆) then there is a
polynomial-time many-one reduction from CSP(∆) to CSP(Γ) which does not increase the number of
variables.
Unfortunately, this theorem is difficult to apply in practice since it requires a good understanding of the
structure of the closed sets pPol(Γ) for all possible choices of Γ. Despite advances made by several different
researchers [12, 13, 35, 51], no such classification is known even for Boolean Γ, and even less is known for Γ
such that SAT(Γ) is NP-hard. Hence, we propose a method inspired by the rich algebraic toolbox developed
for studying the classical complexity of CSP: does the SETH-hardness of SAT(Γ) and CSP(Γ) only depend
on the identities satisfied by the partial polymorphisms of Γ? On the one hand, it is easily verified that if the
only partial polymorphisms of Γ are the partial projections, then Γ can qfpp-define all k-clauses for every
k ≥ 1, and SAT(Γ) is SETH-hard. On the other hand, we would have to show that every non-trivial partial
polymorphism p allows the design of an algorithm that solves SAT(Γ) in O∗ (cn ) time for some c < 2.
One issue which speaks against the feasibility of this approach is that individual partial polymorphisms
are very weak restrictions. For one thing, it is known that for every finite set P of partial operations (that
does not imply any non-trivial total operation), the set Inv(P) of all relations that are invariant under P
contains a double-exponential number of relations as a function of the arity n [37, Lemma 35]. Note that for
k
a finite language such as k-SAT, there are in contrast only 2O(n ) distinct instances on n variables. Hence,
languages Inv(p) for a single partial operation p would be much richer than previously studied problems.
Very similarly, in a related study [36], it was shown that the existence of so-called polynomial kernels for
SAT(Γ) cannot be characterised by such a finite set P , whereas every finite problem, as well as Exact SAT
and problems defined via bounded-degree polynomials, have polynomial kernels [27].
Nevertheless, contrary to these earlier results, we will prove that the presence of certain individual partial
polymorphisms can be used to design improved algorithms for SAT problems. As a starting point we in
the first hand consider the partial analogues of well-studied polymorphisms resulting in tractable CSPs.
For example, a Maltsev operation is a ternary operation φ satisfying the two identities φ(x, x, y) = y and
φ(y, x, x) = y, and is well-known to result in tractable CSPs due to the algorithm by Bulatov and Dalmau [8].
We may then define the partial Maltsev operation over a domain D as the unique partial operation which
for all x, y ∈ D satisfies these two identities, but which is undefined otherwise. Similarly, it is possible to
define partial variants of k-ary near unanimity (k-NU) and k-ary edge (k-edge) operations. These classes
of operations are formally defined in Section 2.5 and at the moment we will simply regard them as wellbehaved operations resulting in tractable CSPs, but we remark that a 2-edge operation is equivalent to a
Maltsev operation and that a ternary NU-operation is nothing else than a majority operation. It may also be
interesting to observe that the partial operations defined in this manner are unique for every fixed domain,
even though there may exist a large number of total operations satisfying the identities.
1.2 Our results and structure of the paper
For a partial polymorphism p, let Inv(p)-SAT refer to the problem SAT(Γ) where Γ = Inv(p). Hence, in
this problem every involved relation is invariant under the given partial operation p. We will sometimes also
refer to the CSP-variants of these problems and denote these by Inv(p)-CSP (and tacitly assume that the
domain of the operation p is clear from the context, or is not relevant). We look at three related aspects of
the complexity of these problems. Let us first discuss our model more carefully.
4
Our questions and model. Since Γ is infinite we first need to fix a constraint representation. Let R ⊆
{0, 1}r be a relation. An explicit representation of R is a list of all tuples t ∈ R. For infinite languages the
explicit representation is not always the most natural one since a relation may contain exponentially many
tuples with respect to the arity. This is particuraly troublesome when proving lower bounds for Inv(p)-SAT
since we may not be able to construct relations of arbitrary arity in the required time bound. Hence, we also
consider an implicit representation. In this model of representation a contraint R(x1 , . . . , xr ) is represented
by an oracle consisting of a computable function which, given an assignment to variables X ⊆ {x1 , . . . , xr },
can determine if this assignment can be extended to an assignment to {x1 , . . . , xr } consistent with R.
Example 5. For each r ≥ 3 consider the relation Rr = {(x1 , . . . , xr ) ∈ {0, 1}r | x1 + . . . + xr is even}.
Even though |Rr | is exponential with respect to r it is not difficult to see that constraints over Rr can be
implicitly represented by computing the parity of the given assignment.
Given these definitions, we consider the following three notions of improved algorithms.
Definition 6. Let Γ be an infinite constraint language.
1. SAT(Γ) admits a non-uniform improved algorithm with running time O∗ (cn ), c < 2, if for every finite
Γ′ ⊂ Γ the problem SAT(Γ′ ) can be solved in O∗ (cn ) time.
2. SAT(Γ) admits an improved algorithm in explicit representation if SAT(Γ) admits an improved algorithm for the problem variant where every relation is provided in explicit representation.
3. SAT(Γ) admits an improved algorithm in the oracle model if SAT(Γ) admits an improved algorithm
when constraints are provided only as extension oracles.
Note that for a non-uniform improved algorithm, the representation does not matter. Also note that these
are gradually stronger requirements, and that in these terms, SETH states that CNF-SAT does not admit even
a non-uniform improved algorithm. On the other hand, allowing constraints of unbounded arity via oracle
P
access can be useful; for example, the n-ary constraint ( ni=1 xi = k) has a simple extension oracle, and if
included in the language, can be used to phrase optimisation problems as oracle-access SAT problems.
To restrict our scope, we focus on constraint languages that are closed under variable negation. Informally,
this means that whenever R ∈ Γ, in addition to constraints R(x1 , . . . , xr ) on only positive variables, we are
also allowed to impose constraints such as R(x1 , . . . , ¬xi , . . . , xr ) with some occurrences of variables xi
negated in the constraint. More formally, it means that for every R ∈ Γ, and for every subset S ⊆ [ar(R)]
of positions of R, the relation produced by negating every tuple t ∈ R in positions S is also contained in Γ.
In this case, we say that Γ is sign-symmetric. This is a natural restriction which holds for many well-studied
constraint language, e.g., the languges corresponding to k-SAT, 1-in-k-SAT and the roots of bounded-degree
polynomials are all sign-symmetric. Furthermore, it is known that the expressive power of a sign-symmetric
constraint language is characterised by a restricted kind of partial polymorphism which we refer to as pSDIoperations (for partial, self-dual and idempotent) [34, 38]. Thus, the restriction to sign-symmetric languages
corresponds directly to a restriction on the algebraic level. Most importantly, the Boolean partial operations
arising from system of identities of the form considered in Section 1.1 are guaranteed to be pSDI.
The fine-grained structure of NP-hard SAT problems. The first part of the paper, Section 3, is dedicated
to explaining the the structure of pSDI-operations. Due to the algebraic correspondence between partial
polymorphisms and qfpp-definability this also serves as a classification of the NP-hard SAT problems we
need to consider for constructing improved algorithms.
First, we study the structure of single pSDI-operations p that impose some non-trivial restrictions on the
expressive power of Γ. We particularly consider the weakest such operations, i.e., such that the language
5
Γ = Inv(p) is as rich as possible. In particular, we consider p such that every subfunction of p which is
pSDI is a partial projection. Let us refer to such an operation as being minimal. For example, the partial
variants of Maltsev, k-NU, and k-edge operations are all minimal. Equipped with this notion we then show
that minimal pSDI-operations are naturally organised into levels, with a structure as follows.
• There is a single minimal operation on level 2, which is the partial Maltsev, or, equivalently, the partial
2-edge operation. This is also equivalent to the 2-universal operation defined below.
• For every other minimal pSDI-operation p, there is a unique largest constant k such that p is implied
by the partial k-NU operation nuk . We refer to this as the level of k. Thus, the partial k-NU operation
is the strongest operation on level k ≥ 3.
• For every level k ≥ 2, there is also a unique weakest pSDI-operation uk which we refer to as the
k-universal operation, such that uk is implied by every operation on level k.
• The language ΓkSAT corresponding to k-SAT is preserved by the partial (k + 1)-NU operation, but not
by any operation on a previous level; and every sign-symmetric language Γ that is not preserved by
the k-universal operation can qfpp-define ΓkSAT .
• Finally, as an interesting case, roots of polynomials of degree at most d are preserved by the (d + 1)universal operation, but not by any other operation on a level up to d + 1.
Thus, the levels of minimal pSDI-operations correspond to a natural notion of difficulty. It also follows that
if a sign-symmetric language Γ is not preserved by the k-universal operation for any constant k, then SAT(Γ)
is trivially SETH-hard, whereas every other language Γ has some kind of restriction on its expressive power.
We also note that there is no known case of a problem known to be SETH-hard, which fits into a framework
of searching through the set {0, 1}n for a solution, and which is k-universal for any k. Hence, it is consistent
with our present knowledge that every k-universal problem SAT(Γ) admits an improved algorithm.
Last, we remark that although we in this paper are mainly interested in the time complexity of SAT, the
classification of minimal pSDI-operations in this section may be of independant interest for any Boolean
problem compatible with qfpp-definitions. In this vein, we also give a “vertical” result in the above hiearchy,
and show that every sign-symmetric constraint language Γ not preserved by the partial k-NU operation for
any k can qfpp-define either 1-in-k-clauses of all arities, or counting constraints modulo p of all arities for
some fixed prime p. This result is the main technical challenge in this section, and relies on an application
of Szemerédi’s theorem [54] to analyse the structure of symmetric relations R ∈
/ Inv(nuk ).
Upper and Lower Bounds on the SAT problem. Second, in Section 4 and Section 5, we consider the
strength of the problem Inv(p)-SAT for various pSDI-operations p, with an interest in bounding the value
c(Γ) for Γ = Inv(p) from above and below. The first question here is the matter of constraint representation.
As mentioned previously, the language Inv(p) contains a double-exponential number of relations of arity r
as a function of r; hence any fixed representation would in the worst case use 2O(r) bits just to encode the
relations. This becomes an issue when we allow constraints of unbounded arity. Recall that we consider
three alternatives for representation: explicit representation, extension oracles, and non-uniform algorithms
where the particular choice of representation does not matter. We then obtain the following results.
• When p is the partial 2-edge operation, we refer to Inv(p)-SAT as 2-edge-SAT. We show that 2-edgen
SAT can be solved in O∗ (2 2 ) time in the oracle setting using a meet-in-the-middle strategy combined
n
with the computation of a kind of canonical labels for partial assignments, similarly to the O∗ (2 2 )time algorithm for Subset Sum with n integers [24]. A similar improved algorithm is possible for the
generalisation to 2-edge-CSP, i.e., for fixed non-Boolean domains. Furthermore, if c(Inv(p)) < 21/2
6
1
in the extension oracle setting, then Subset Sum can be solved in O∗ (2( 2 −ε)n ) for some ε > 0, which
is a long-standing open problem.
• When p is the partial k-NU operation, we refer to Inv(p)-SAT as k-NU-SAT. For k = 3, this problem
is equivalent to 2-SAT, and hence in P, but the generalisation 3-NU-CSP to larger fixed domains is NPhard and admits an improved algorithm using fast matrix multiplication, similarly to the well-known
algorithm for the CSP problem over binary constraints.
• For k > 3, we show two conditional connections. First, if the (k, k − 1)-hyperclique problem
for hypergraphs with ground set of size n can be solved in time O(nk−ε ) for any ε > 0, then both
k-NU-SAT and k-NU-CSP admit improved algorithms in the oracle setting. Second, if the ErdősRado sunflower conjecture [18] holds for sunflowers with k sets, then k-NU-SAT admits an improved
algorithm via a local search strategy in the explicit representation, similar to Schöning’s algorithm for
k-SAT [52].
• We also investigate the case that p is the partial 3-edge operation e3 , and give a partial result. Assume
that every relation R in the input is either preserved by the partial 2-edge relation, or by nu3 , or R is
symmetric and preserved by e3 – i.e., whether t ∈ R depends only on the Hamming weight of t. Then
the SAT problem has an improved algorithm via a reduction to the problem of finding monochromatic
triangles in an edge-coloured graph, which in turn can be solved using fast algorithms for triangle
finding in sparse graphs. We do not know whether this strategy generalises to non-symmetric relations.
For further classes, we note that SAT(Γ) contains some highly challenging special cases. In particular an
algorithm for the k-universal languages for k > 2 would need to generalise the algorithm of Lokshtanov et
al. for bounded-degree polynomials [42], while only using the abstract properties guaranteed by uk .
Finally, we show lower bounds in the oracle extension model: for every minimal pSDI-operation p, we
get a concrete lower bound c(Inv(p)) ≥ ck > 1 assuming the randomized SETH, where k is the level of p.
(1−ε)n
) for any ε > 0 and any
That is, unless SETH is false, no algorithm can solve Inv(p)-SAT in time O∗ (ck
log k
p at level k. The bound ck converges to 2 at a rate of 2 − ck = Θ( k ).
A connection to polynomial-time problems. Finally, we make some connections between the Inv(p)SAT and Inv(p)-CSP problems and some problems in polynomial-time algorithms. We show that the
minimal pSDI-operations generalise not only to CSP problems on fixed domains, but to abstract conditions
on “CSP-like” problems on a domain of size n and with d = Θ(1) variables. We refer to this as the abstract
Inv(p)-problem. Any solution to such a problem that runs in time O(nd−ε ) for any ε > 0 implies an
improved algorithm for the corresponding Inv(p)-SAT and Inv(p)-CSP problems in the oracle setting for
every fixed domain. This lies behind the improved algorithms for 2-edge-CSP and 3-NU-CSP.
However, there is some indication that these problems may be tougher than the original problems, since
the reduction loses a significant amount of instance structure (e.g., the local search strategy for k-NU-SAT
cannot be lifted to the abstract problem). In fact, there are conjectures that would prevent improved algorithms
for most cases of the abstract problem considered in this article:
• The abstract k-NU problem is equivalent to (k, k − 1)-hyperclique, i.e., the problem of finding a
k-hyperclique in a (k − 1)-regular hypergraph. Thus, it has an improved algorithm for k = 3 but the
status is unknown for k > 3. Moreover, the general (l, k)-hyperclique problem for l > k has been
conjectured to require nl−o(1) time [40].
• The abstract 3-universal problem contains the problem of finding a zero-weight triangle in an edgeweighted graph with arbitrary edge weights. This does not admit an improved algorithm unless the
3-SUM conjecture fails (but SETH-hardness is not known) [56].
7
Considering the connections, we still consider it useful to ask which minimal pSDI-operations p suffice to
guarantee an improved algorithm for the abstract Inv(p)-problem. We leave this question for future work.
1.3 Technical notes and proof methods
Let us now give a few more details about the proofs of the above results.
The structural characterisation builds on a description of minimal non-trivial pSDI-operations (Lemma 27)
— they are precisely the operations produced by padding the partial k-NU operation by additional arguments.
The weakest and strongest operations on each level follow from this almost by definition. It also follows that
the operations on each level k are characterized by the presence or absence of each of roughly 2k possible
types of padding argument. Note that such a padding makes an operation weaker; e.g., in order to apply the
partial majority operation to a sequence of tuples t1 , . . . , tk ∈ R for some relation R, in a padded version
of arity r we require that R further contains a sequence of tuples tk+1 , . . . , tr determined by the padding
arguments from the tuples t1 , . . . , tk .
This also provides a way to think about the consequences of not being preserved by such an operation.
Assume e.g. that a relation R is not preserved by nuk . Then by definition there are t1 , . . . , tk ∈ R such that
nuk (t1 , . . . , tk ) = t is defined, and by sign-symmetry we may assume that t is the constant 0-tuple 0ar(R) .
Then the witness produces a partition of the arguments of R, in a way which can be used to implement a
relation R′ of arity k which accepts every tuple of weight 1 but none of weight 0. However, we have no
information at this point about the remaining tuples in R′ . Continuing this line of reasoning to derive a
consequence for an infinite sign-symmetric language Γ with nuk ∈
/ pPol(Γ) for every k, we first observe that
′′
we can define a symmetric relation R ∈
/ Inv(nuk ) as a conjunction of k! applications of R′ under argument
permutation, then (as announced) analyse the possibilities for families of such relations using Szemerédi’s
theorem. In particular, a broken arithmetic progression of i accepted weights in such a relation implies that
we can qfpp-define an i + 1-clause using R.
By contrast, if uk ∈
/ pPol(R), then the tuples t1 , . . . , t2k −1 ∈ R required by the arguments of uk imply
that such a relation R′ must have |R′ | = 2k − 1, i.e., it must be the relation corresponding to a k-clause.
Moving on to the algorithmic applications, most of the positive results are relatively straight-forward
applications of known ideas; the interesting aspect is that the applicability of these ideas follows from such
simple conditions as the minimal pSDI-operations. Here, we particularly wish to highlight the conjectural
connection to local search. Recall that Schöning’s algorithm [52] reduces k-SAT to several applications of
local search, i.e., given a starting point x ∈ {0, 1}n and a parameter t, find a satisfying assignment within
Hamming distance t of x. By sign-symmetry, for our problem this reduces to the case x = 0n (alternatively,
one could use monotone local search; cf. Fomin et al. [20]). Now, consider the set of all minimal tuples in
any relation R ∈ Inv(nuk ) with 0ar(R) ∈
/ R. It is easy to see that by the nuk -condition, this set does not
contain a sunflower of k sets, and by the sunflower conjecture, this implies that for every i there are at most
C i such minimal tuples in R of weight i for some C. A simple computation shows that a recursive algorithm
that finds an unsatisfied relation R, enumerates minimal tuples in it, and recursively proceeds from every
such tuple yields a total searching time of 2O(t) , which would be precisely sufficient to yield an improved
algorithm for k-NU-SAT. This algorithm uses the explicit representation in order to be able to enumerate
such minimal tuples. It is an interesting open question whether this can be achieved efficiently in the oracle
setting.
Finally, we move on to our lower bounds. These are of two kinds, a reduction from Subset Sum to
2-edge-SAT, and the generic lower bound under SETH against any problem Inv(p)-SAT. For the former,
recall that the partial 2-edge operation is equivalent to u2 , and thus contains all constraints which can be
phrased as linear equations, e.g., Subset Sum instances. But we are also required to provide an extension
oracle for each constraint, which is clearly infeasible if we plug in the Subset Sum equation as-is. However,
√
√
this is easily solved by splitting the binary expansion of the target number into O( n) blocks of O( n)
8
bits each. With some moderate guessing, each block reduces to one linear equation, and via the tabulation
√
algorithm for Subset Sum an extension oracle each such block can be produced with a query time of 2O( n) .
The generic bounds, in turn, work via a generic padding argument: we show that for every level k, and
any set X of n variables, there is a universal padding formula R(X, Y ) on |Y | = Θ(n) additional variables
such that R′ (X, Y ) ≡ R′ (X)∧ R(X, Y ) is k-NU for any relation R′ (X). Furthermore, random parity-check
variables suffice to produce this padding formula, allowing for an efficient extension oracle for the relation
R′ (X, Y ). Finally, by the regularity of the padding formula, we can reuse the same variables Y for all
constraints in an input instance of q-SAT, for any q, and only pay with |Y | = Θ(n) extra variables in total.
The fact that some such padding exists was previously known [37]. Recall that every operation p
considered has at least one tuple of values for which it is undefined. Then, if we add enough random
variables, for every attempt p(t1 , . . . , tr ) of finding a valid application of p on a relation R there will be a
padding variable j such that (t1 [j], . . . , tr [j]) takes the values of such a tuple, and p is undefined. The fact
that parity-check variables suffice in our case follows from the fact that p contains k arguments that form a
partial k-NU operation. It is easy to check that almost all parity-check variables form an undefined tuple
of values already over these arguments. This construction could be derandomized using a universal hash
family, possibly at the cost of a larger constant |Y |/|X|, but we do not pursue this.
1.4 Related work
Our work can be seen as an amalgamation of the following areas: fine-grained time complexity and lower
bounds under the SETH, and the algebraic approach for studying classical complexity of CSP.
Concerning the former, SETH has turned out to be a highly useful conjecture for exact algorithms since a
relative lower bound from SETH shows that any further improvements also implies a breakthrough speed-up
for SAT. Many different problems have been shown to admit lower bounds via the SETH, but in the current
context of SAT, in addition to the foundational works of Impagliazzo et al. [25, 26] it is worth mentioning
the lower bound for Not-all-equal SAT (NAE-SAT) by Cygan et al. [15] and the lower bound for Π2 3-SAT
by Calabro et al. [11]. However, to the best of our knowledge, all concrete lower bounds using SETH
for exponential-time algorithms falls into one of the following cases: either the lower bound matches the
running time of a trivial algorithm, as in the case of Hitting Set, NAE-SAT, and Π2 -3-SAT, showing that
no improvement is possible; or the lower bounds are with respect to a much more permissive complexity
parameter than n, such as treewidth [41]. The one other example we are aware of is from the study of
infinite-domain CSPs by Jonsson and Lagerkvist [30], who obtained upper bounds of the form O∗ (2f (n) ) for
non-linear functions f and a lower bound stating that the CSPs are not solvable in O(cn ) time for any constant
c. These bounds are therefore in a sense closer to non-subexponentiality results usually obtained from the
ETH. SETH and other conjectures have also seen significant applications over recent years in producing
conditional lower bounds for polynomial-time solvable problems, but these are only tangentially relevant
here.
With regards to the algebraic approach we wish to highlight a few related but different results. Partial
polymorphisms and the link to qfpp-definitions were first introduced to the CSP community by Schnoor &
Schnoor [50] even though these notions were well-known in the algebraic community much longer [21, 47].
However, the principal motivation by Schnoor & Schnoor was to obtain dichotomy theorems for CSPlike problems incompatible with existential quantification, and the explicit connection to fine-grained time
complexity of CSP was not realized until later by Jonsson et al. [31]. This work utilized a lattice-informed
approach which exploited the structure of the inclusion structure of closed sets of partial polymorphisms,
in order to identify an NP-complete SAT(Γ) problem such that c(Γ) ≤ c(∆) for every other NP-complete
SAT(∆). This problem was referred to as the easiest NP-complete SAT problem and was later generalized to
a broad class of finite-domain CSPs [32]. However, continued advancements in understanding this inclusion
structure revealed that even severely restricted classes of constraint languages had a very complicated
9
structure [12, 35]. In a similar vein of negative results it was also proven that (1) pPol(Γ) cannot be
generated by any finite set of partial operations whenever Γ is finite and SAT(Γ) is NP-hard, and (2) if
P is a finite set of partial operations such that Inv(P)-SAT is NP-hard, then any pp-definable relation
over Inv(P) can be transformed into a pp-definition using only a linear number of existentially quantified
variables [37]. In plain language, these results show that finite constraint languages result in complex partial
polymorphisms, and that simple partial polymorphisms result in complex constraint languages. A previous
attempt at grappling with this difficulty provided closure operators that generate pPol(Γ) for a finite Γ from
a finite basis [34], but this intrinsically uses that Γ is finite, and is not applicable in the current paper.
Our approach in this paper avoids the pitfalls of the lattice-informed approach since it is sufficient to
understand the behaviour of individual pSDI-operations. This is in line with how the research programme of
classifying the complexity of finite-domain CSPs evolved into a project of describing properties of operations
defined by system of identities (see the survey by Barto et al. for more details [3]).
Another related paper by the present authors investigates the existence of polynomial (or linear) kernels
for problems SAT(Γ), using ideas of extending the language Γ into a tractable CSP on a larger domain [36],
including extensions into 2-edge (i.e., Maltsev) and k-edge languages. However, there is no concrete
technical connection between that paper and this one, as having polynomial kernels turns out to be a much
more restricted property than admitting improved algorithms.
1.5 Concluding remarks and open questions
Our principal motivation in this paper is to study the SETH-hardness of the parameterized SAT(Γ) problem.
To simplify our study we restricted our focus to sign-symmetric constraint languages, which is a common
assumption for SAT problems studied in practice. Moreover, due to the connection between sign-symmetric
constraint languages and pSDI-operations, understanding the inclusion structure between sign-symmetric
constraint languages is tantamount to describing the expressive power of pSDI-operations. Even better,
pSDI-operations can in many cases be understood as the partial analouges of well-studied operations such
as Maltsev operations, NU-operations and edge-operations, making them easier to reason with.
The main open question is whether our results can be strengthened into a dichotomy for sign-symmetric
SAT problems. One direction is already clear: if Γ is not preserved by any k-universal operation then SAT(Γ)
is SETH-hard and does not admit an improved algorithm without breaking the SETH. The other direction
is harder and requires a substantially better understanding of languages invariant under a given k-universal
operation; such languages include, but are not limited to, relations expressible as roots of polynomial
equations of degree at most k + 1, where an improved algorithm is known [42]. It is not clear at this point
how much richer the set Inv(uk ) is, compared to this class of problems. Existing (conjectured) lower bounds
against polynomial-time problems captured by abstract Inv(p)-problems also indicate that the problem might
be more difficult for remaining cases. We also proved that the SAT problems under consideration admit lower
bounds under the SETH. To the best of our knowledge, this is the first result showcasing both a non-trivial
upper bound and a concrete lower bound under the SETH in terms of a natural parameter n. These bounds
were obtained in the extension oracle setting and it is currently unclear if matching bounds can also be
obtained if constraints are represented explicitly. The padding construction is still valid in this setting, but it
is a challenge to apply it without creating constraints with exponentially many tuples.
Last, our approach easily extends to finite-domain CSPs, as evidenced by the improved algorithms for
2-edge-CSP and 3-NU-CSP. The notion of a pSDI-operation is only relevant in the Boolean domain, but
a similar notion can likely be defined for arbitrary finite domains. For example, instead of self-duality,
essentially meaning that the partial operation is closed under negation, we would require that the operation is
closed under every unary operation over the domain. However, it is not clear if the inclusion structure of such
generalized pSDI-operations can be characterized in a similar hierarchy as the Boolean pSDI-operations.
10
2 Preliminaries
A k-ary relation over a domain D is a subset of D k . If t = (x1 , . . . , xn ) is a k-ary tuple we for every
1 ≤ i ≤ k let t[i] = xi , and if i1 , . . . , ik′ ∈ [k] = {1, . . . , k} we write Proji1 ,...,ik′ (t) = (t[i1 ], . . . , t[ik′ ])
for the projection of t on the coordinates i1 , . . . , ik′ . This notation easily extends to relations and we write
Proji1 ,...,ik′ (R) for the relation {Proji1 ,...,ik′ (t) | t ∈ R}.
A set of relations is called a constraint language, or simply a language, and will usually be denoted by
Γ and ∆. We will typically define relations either by their defining logical formulas or by their defining
equations. For example, the relation R1/3 = {(0, 0, 1), (0, 1, 0), (1, 0, 0)} may be defined by the expression
R1/3 ≡ x1 + x2 + x3 = 1. However, we will not always make a sharp distinction between relations and
their defining logical formulas and will sometimes treat e.g. a k-clause as a relation. We write ar(R) for the
arity of a relation R, and use the notation EqD to denote the equality relation {(x, x) | x ∈ D} over D.
A k-ary relation R is said to be totally symmetric, or just symmetric, if there exists a set S ⊆ [k] =
{1, . . . , k} such that (x1 , . . . , xk ) ∈ R if and only if x1 + . . . + xk ∈ S. For example, R1/3 is totally
symmetric as witnessed by the set S = {1}. Symmetric relations will prove to be useful since it is
sometimes considerably simpler to describe the symmetric relations invariant under a partial operation.
2.1 The parameterized SAT and CSP Problems
Let Γ be a Boolean constraint language. The parameterized satisfiability problem over Γ (SAT(Γ)) is the
computational decision problem defined as follows.
Instance: A set V of variables and a set C of constraint applications R(v1 , . . . , vk ) where R ∈ Γ, ar(R) = k,
and v1 , . . . , vk ∈ V .
Question: Is there a function f : V → {0, 1} such that (f (v1 ), . . . , f (vk )) ∈ R for each R(v1 , . . . , vk ) in C?
The constraint satisfaction problem over a constraint language Γ (CSP(Γ)) is defined analogously with
the only distinction that Γ is not necessarily Boolean. We write (d, k)-CSP for the CSP problem over a
domain with d elements where each constraint has arity at most k.
2.2 The extension oracle model
Recall from Section 1.2 that we consider two distinct representations of SAT and CSP instances. We now
define these in more detail. In the first representation each relation R occurring in a constraint R(x1 , . . . , xk )
is represented as a list of tuples. We call this representation the explicit representation. This is one of
the most frequently occurring representation methods in the algebraic approach to CSP, but it is fair to say
that it is not convenient in any practical application since a relation may contain exponentially many tuples
with respect to the number of arguments. We therefore consider a more implicit representation where each
constraint is represented by a procedure which can verify whether a partial assignment of its variables is
consistent with the constraint.
Definition 7. Let R be an n-ary relation over a set D. A computable function which given indices
′
i1 , . . . , in′ ∈ [n] and t ∈ Dn answers yes if and only if t ∈ Proji1 ,...,in′ (R) is called an extension oracle
representation of R.
Hence, given a constraint R(x1 , . . . , xn ) and a partial truth assignment f : X → D, X ⊆ {x1 , . . . , xn },
the extension oracle representation can be used to decide whether f can be completed into a satisfying
assignment of R(x1 , . . . , xn ).
Example 8. CNF-SAT can be succinctly represented in the extension oracle model. Consider e.g. a positive
clause (x1 ∨ . . . ∨ xn ) and a partial truth assignment f on {x1 , . . . , xn }. We can then answer yes if and
only if not every variable xi occurring in the clause is assigned the value 0.
11
2.3 Sign-symmetric constraint languages
An n-ary sign pattern is an tuple s where s[i] ∈ {+, −} for each 1 ≤ i ≤ n. If t is an n-ary Boolean tuple
and s an n-ary sign pattern then we let ts be the tuple where ts [i] = t[i] if s[i] = + and ts [i] = 1 − t[i] if
s[i] = −. Similarly, if if R is a Boolean relation and s an n-ary sign pattern we by Rs denote the relation
{ts | t ∈ R}. Last, for 1 ≤ i ≤ n and c ∈ {0, 1} we let Ri=c = {t | t ∈ R, t[i] = c} be the relation resulting
from freezing the ith argument of R to c.
Definition 9. A Boolean constraint language Γ is said to be sign-symmetric if (1) Rs ∈ Γ for every n-ary
R ∈ Γ and every n-ary sign pattern s and (2) Ri=c ∈ Γ for every c ∈ {0, 1} and every 1 ≤ i ≤ n.
2.4 Partial polymorphisms and quantifier-free primitive positive definitions
Let D be a finite set of values. A k-ary partial operation, or a partial function, f over D is a mapping
X → D where X ⊆ Dk . The set X is said to be the domain of f and we let domain(f ) = X denote
this set and ar(f ) = k denote the arity of f . If f and g are two n-ary partial operations over D such
that domain(g) ⊆ domain(f ) and g(x1 , . . . , xn ) = f (x1 , . . . , xn ) for every (x1 , . . . , xn ) ∈ domain(g)
then g is said to be a subfunction of g. For n ≥ 1 the i-ary projection, 1 ≤ i ≤ n, is the operation
πin (x1 , . . . , xi , . . . , xn ) = xi and a partial projection is any subfunction of a total projection.
If R is an n-ary relation over D and f a k-ary partial operation over D we say that f is a partial
polymorphism of R, that R is invariant under f , or that f preserves R, if f (t1 , . . . , tk ) ∈ t or f (t1 , . . . , tk ) is
undefined, for each sequence of tuples t1 , . . . , tk . We let pPol(R) be the set of all partial polymorphisms of
the relation R, and if Γ is a constraint language we let pPol(Γ) denote the set of partial operations preserving
each relation in Γ. The notion of a total polymorphism can be defined simply by requiring that f is total,
i.e., domain(f ) = D k , and we let Pol(Γ) be the set of all total polymorphsims of the constraint language Γ.
Similarly, if P is a set of partial operations we let Inv(P) be the set of all relations invariant under P . Each
set of partial operations P naturally induces a SAT problem SAT(Inv(P)) where each relation involved in
a constraint is preserved by every partial operation in P . Recall from Section 1.2 that we as a shorthand
denote this problem by Inv(P)-SAT. The two operators Inv(·) and pPol(·) are related by the following
Galois connection.
Theorem 10 ([21, 47]). Let Γ and ∆ be two constraint languages. Then Γ ⊆ Inv(pPol(∆)) if and only if
pPol(∆) ⊆ pPol(Γ).
The applicability of partial polymorphism in the context of fine-grained time complexity might not be
evident from these definitions. However, sets of the form Inv(P), called weak systems or weak co-clones, are
closed under certain restricted first-order formulas which are highly useful in this context. Say that a k-ary
relation R has a quantifier-free definition (qfpp-definition) over a constraint language Γ over a domain D if
R(x1 , . . . , xk ) ≡ R1 (x1 ) ∧ . . . ∧ Rm (xm ) where each Ri ∈ Γ ∪ {Eq D } and each xi is a tuple of variables
of length ar(Ri ). It is then known that Inv(P) for any set of partial operations P is closed under taking
qfpp-definitions. With this property the following theorem is then a straightforward consequence.
Theorem 11. [31] Let Γ and ∆ be two finite constraint languages. If pPol(Γ) ⊆ pPol(∆) then there exists
a polynomial-time many-one reduction from SAT(∆) to SAT(Γ) which maps an instance (V, C) of SAT(∆)
to an instance (V ′ , C ′ ) of SAT(Γ) where |V ′ | ≤ |V | and |C ′ | ≤ c|C|, where c depends only on Γ and ∆.
In particular this implies that if CSP(Γ) is solvable in O(cn ) time and pPol(Γ) ⊆ pPol(∆) then CSP(∆)
is solvable in O(cn ) time, too. We will now briefly describe the closure properties of pPol(Γ), which are
usually called strong partial clones. First, if f, g1 , . . . , gm ∈ pPol(Γ) where f is m-ary and each gi is n-ary,
then the composition f ◦ g1 , . . . , gm (x1 , . . . , xn ) = f (g1 (x1 , . . . , xn ), . . . , gm (x1 , . . . , xn )) is also included
12
in pPol(Γ). This operation will be defined on a tuple (x1 , . . . , xn ) ∈ D n if and only if each gi (x1 , . . . , xn )
is defined and the resulting application over f is defined. Second, pPol(Γ) contains every partial projection,
which is known to imply that pPol(Γ) is closed under taking subfunctions (i.e., if f ∈ pPol(Γ) then every
subfunction of f is included in pPol(Γ)). If P is a set of partial operations we write [P ]s = pPol(Inv(P))
for the smallest strong partial clone containing P .
2.5 Polymorphism patterns
In this section we describe a method for constructing partial polymorphisms that have a strong connection to
the sign-symmetric constraint languages defined in Section 2.3. As a shorthand we will sometimes denote
the k-ary constant tuple (d, . . . , d) by dk .
Definition 12. Let f be a Boolean partial operation. We say (1) that f is self-dual if x ∈ domain(f ) for
every x ∈ domain(f ) and f (x) = 1 − f (x), where x denotes the complement of the tuple x, and (2) that f
is idempotent if dk ∈ domain(f ) and f (dk ) = d for every d ∈ D.
In the sequel, we will call a Boolean partial operation which is both self-dual and idempotent a pSDIoperation, short for partial, self-dual, and idempotent operation. Let a polymorphism pattern of arity r be a
set of pairs (t, x) where t is an r-ary tuple of variables and where x occurs in t. We say that a r-ary partial
operation f over a set of values D satisfies an r-ary polymorphism pattern P if
domain(f ) = {(τ (x1 ), . . . , τ (xr )) | ((x1 , . . . , xr ), x) ∈ P, τ : {x1 , . . . , xr } → D}
and f (τ (x1 ), . . . , τ (xr )) = τ (x) for every ((x1 , . . . , xr ), x) ∈ P and every τ : {x1 , . . . , xr } → D.
A Boolean operation is pSDI if and only if it satisfies a polymorphism pattern. To see this, note that
if f is pSDI, then it is easy to create a polymorphism pattern P by letting each tuple t ∈ domain(f ) such
that f (t) = 0 correspond to an entry in P . Similarly, it is not difficult to show that any partial operation
satisfying a polymorphism pattern must be self-dual and idempotent. We then have the following link
between sign-symmetric constraint languages and partial operations satisfying polymorphism patterns.
Theorem 13. [38] Let f be a pSDI-operation. Then Inv(f) is sign-symmetric.
Hence, pSDI-operations provide a straightforward way to describe broad classes of sign-symmetric
constraint languages. It is also known that if Γ is sign-symmetric and SAT(Γ) is NP-hard, then every partial
polymorphism of Γ is a subfunction of a pSDI-operation preserving Γ [38][Theorem 3] (see Lagerkvist [34]
for a full proof). We will now define the pSDI-operations that will play a central role in our current pursuit.
Definition 14. Let k ≥ 2. A (k+1)-ary partial operation is a partial k-edge operation if it satisfies the pattern
consisting of ((x, x, y, y, y, . . . , y, y), y), ((x, y, x, y, y, . . . , y, y), y), and for each i ∈ {4, . . . , k + 1}, the
tuple ((y, . . . , y, x, y, . . . , y), y), where x appears in position i.
We will typically denote partial k-edge operations by ek , and, if the underlying set D is important, by
eD
.
k A partial 2-edge operation will sometimes be called a partial Maltsev operation.
Definition 15. Let k ≥ 3. A k-ary partial operation is a partial k-ary near-unanimity operation (partial kNU operation) if it satisfies the pattern which for each i ∈ {1, . . . , k} contains ((x, x, . . . , x, y, x, . . . , x), x),
where y occurs in position i.
We write nuD
k to denote this operation over the domain D, and nuk if the domain is clear from the
context, or not relevant. Ternary partial NU-operations will sometimes be called partial majority operations.
Note that the partial majority operation is total in the Boolean domain but is properly partial for every larger
domain. Last, we define the following class of self-dual partial operations. Say that the argument i of a k-ary
partial operation f is redundant if there exists j 6= i such that t[i] = t[j] for every t ∈ domain(f ).
13
Subset Sum
Linear Equations
1-in-k SAT
Sidon Sets
Degree-2 Polynomials
2-edge = 2-universal
2-SAT
(d, 2)-CSP
Graph k-Clique
3-NU
3-edge
3-universal
3-SAT
(d, 3)-CSP
(3, ℓ)-Hyperclique
4-NU
4-edge
4-universal
...
...
...
k-NU
k-edge
k-universal
(k − 1)-SAT
(d, k − 1)-CSP
(k − 1, ℓ)-Hyperclique
Degree-3 Polynomials
Degree-(k − 1) Polynomials
Figure 1: The inclusion structure between selected minimal pSDI-operations (solid outlines), and some
problems that reduce to the corresponding SAT or CSP problem (dotted outlines). Several classes on each
level k ≥ 3 have been omitted.
Definition 16. Let k ≥ 2. The k-universal operation uk is the Boolean (2k − 1)-ary pSDI-operation defined
on 2k +2 tuples such that (1) uk is not a partial projection and (2) uk does not have any redundant arguments.
While not immediate from the definition, the operation uk is in fact unique up to permutation of
arguments. To see this, simply take the k non-constant tuples t1 , . . . , tk ∈ domain(uk ) such that uk (t1 ) =
. . . = uk (tk ) = 0. Since uk is not a projection and is pSDI, it follows that there cannot exist i ∈ [2k − 1]
such that (t1 [i], . . . , tk [i]) = 0k . Hence, since uk does not have any redundant arguments, there for every
t ∈ {0, 1}k \ {0k } must exist a unique i ∈ [2k − 1] such that (t1 [i], . . . , tk [i]) = t.
Last, we remark that there is a connection between our notion of polymorphism patterns and the
operations studied in connection to the CSP dichotomy (see e.g. the survey by Barto et al. [3]). In technical
terms polymorphism patterns essentially matches strong Maltsev condititions where the right-hand side is
restricted to a single variable. Similar restrictions, called height-1 identities, have been considered earlier
and it is known that the complexity of a CSP(Γ) problem only depends on the height-1 identities satisfied by
the operations in Pol(Γ) [33].
3 Structure of Constraint Languages under Minimal Restrictions
We now properly begin the first part of the paper, investigating the structure of maximally expressive, yet
restricted sign-symmetric constraint languages. This investigation is performed via the study of the weakest
non-trivial pSDI-operations, including the operations defined in Section 2.5. As a preview of the structure,
and of some of the included problems, we refer to Figure 1. The problem and language inclusions illustrated
in this figure will be shown across the next two subsections.
More precisely, by “weakest” pSDI-operations, we mean partial operations that are minimal in the
following sense. Recall that for every pSDI-operation f and every subfunction f ′ of f , we have Inv(f) ⊆
Inv(f ′ ). This motivates the following definition.
14
Definition 17. Let f be a pSDI-operation. We say that f is trivial if it is a subfunction of a projection, and
a minimal non-trivial pSDI-operation if f is non-trivial but every proper subfunction f ′ of f which is a
pSDI-operation is trivial.
Our study in this section is focused on constraint languages Γ = Inv(f) where f is a single minimal
non-trivial pSDI-operation, since these are the most expressive sign-symmetric constraint languages that are
still restricted in expressive power. We begin by giving some examples for the particular classes of k-NU,
k-edge and k-universal partial operations defined in Section 2.5.
3.1 Properties of specific sign-symmetric constraint languages
In this section, we provide some illustrative examples of languages included in Inv(f) for particular pSDIoperations f . We first recall the following result from Lagerkvist & Wahlström.
Theorem 18. [37] Let F be a finite set of partial operations such that Inv(F)-SAT is NP-complete. Then any
n-ary Boolean relation has a pp-definition over Inv(F) using at most O(n) existentially quantified variables.
In effect, this implies that any constraint language Inv(F), where F is a finite set of pSDI-operations,
cn
is extremely expressive. One direct consequence is that Inv(F) contains at least 22 n-ary relations for
some constant 0 < c ≤ 1. This makes such constraint languages markedly different from finite constraint
languages, since for any finite constraint language Γ, the number of n-ary qfpp-definable relations over Γ is
bounded by O(2p(n) ) for a polynomial p depending on Γ. This also implies that there cannot exist a finite Γ
such that pPol(Γ) = [F]s . In fact, the relations of Inv(F) for such an F are dense enough that for any n-ary
relation R, a random padding of R by O(n) parity-check variables is enough to create a variable in Inv(F)
with high probability. This fact will be exploited in Section 5.3.
Despite this, we will see that the pSDI-operations defined in Section 2.5 do correspond roughly to natural
restrictions on the expressive power of a language Γ. We now illustrate the classes with a few examples.
In the process will occasionally refer to the language inclusions illustrated in Figure 1. Proofs of these
inclusions is given in Theorem 29 in Section 3.2. Let us now begin with a basic example.
Lemma 19. R ∈ Inv(nuk ) for every (k − 1)-ary relation R, k ≥ 3.
Proof. Let t1 , . . . , tk ∈ R be such that nuk (t1 , . . . , tk ) is defined, and for i ∈ [k − 1] let t(i) =
(t1 [i], . . . , tk [i]). For every i ∈ [k − 1], either t(i) is constant or there is a single index j where t(i) [j]
deviates from its other entries. By the pigeonhole principle, there is at least one index j ∈ [k] such that
t(i) [j] does not deviate from the majority for any i ∈ [k − 1]. Then we have nuk (t1 , . . . , tk ) = tj .
We also show a corresponding negative statement. By the inclusions shown in the next section, this will
imply that a k-clause is not preserved by any operation at “level k” of the hierarchy in Figure 1.
Lemma 20. Let R ⊂ {0, 1}k be a k-clause, i.e., |R| = 2k − 1, k ≥ 2. Then R is not preserved by the partial
k-universal operation.
Proof. By sign-symmetry, we assume that R = {0, 1}k \ {0k }. Let t1 , . . . , tk be the non-constant tuples in
domain(uk ) such that uk (ti ) = 0 for each i ∈ [k]. Then for each i ∈ [2k −1], the tuple t(i) = (t1 [i], . . . , tk [i])
defines a tuple of R; thus the application
k −1)
uk (t(1) , . . . , t(2
) = 0k
is defined and shows that R ∈
/ Inv(uk ).
Next, we consider a canonical example of a useful relation preserved by the partial 2-edge operation.
15
Lemma 21. Let R(x1 , . . . , xn ) ⊆ {0, 1}n be defined via a linear equation
n
X
αi xi = β
i=1
evaluated over a finite field F. Then R ∈ Inv(e2 ).
Proof. This is a special case of the notion of a Maltsev embedding of R previously investigated by the
authors [36]. It is known that a relation with a Maltsev embedding is closed under a family of partial
operations, of which e2 is the simplest.
A particular example of such relations is the Exact SAT problem. We show that its 1-in-k relations are
also not closed under nuk .
s
Lemma 22. Let R1/k = {(x1 , . . . , xk ) ∈ {0, 1}k | x1 + . . . + xk = 1}, and ΓXSAT = {R1/k
| k ≥ 1, s is
a k-ary sign-pattern}. Then ΓXSAT ⊆ Inv(e2 ) but is not preserved by nuk for any k.
Proof. The positive direction follows from Lemma 21, since R1/k can be phrased as a linear equation over
the integers mod p, for p ≥ k + 1. The negative direction is immediate: let R1/k = {t1 , . . . , tk }. Then
nuk (t1 , . . . , tk ) is defined and equals 0k .
Another example of a problem with the character of linear equations is Subset Sum. Even though an
instance of Subset Sum is defined by just a single linear equation rather than as a SAT(Γ) instance, we show
in Section 5.2 that the complexity of 2-edge-SAT and Subset Sum are closely connected. As for the class
Inv(ek ) for k ≥ 3, the inclusions illustrated in Figure 1 imply that this class contains both relations with
linear equation extensions and all (k − 1)-clauses.
Finally, we show two examples for the partial k-universal operation uk . The first is a previously studied
class of Lokshtanov et al. [42]. Note this problem does admit an improved algorithm.
Definition 23. Let Pd denote the set of Boolean relations such that each n-ary R ∈ Pd is the set of roots of
an n-variate polynomial equation where each polynomial has degree at most d.
Lemma 24. Let R ∈ Pd be an n-ary relation. Then R is preserved by ud+1 , but not by any other non-trivial
pSDI-operation of domain size at most 2d + 2.
Proof. For the first direction, let P (x1 , . . . , xn ) be the polynomial defining R, and let t1 , . . . , tr ∈ R be
such that ud+1 (t1 , . . . , tr ) = t′ is defined. Since the set of relations representable by bounded-degree
polynomials is sign-symmetric, we may assume for simplicity that t′ = 1n . The tuples (t1 , . . . , tr ) define
a new polynomial of degree at most d and with at most d + 1 variables, defined by identifying all pairs of
variables xi and xj that have the same pattern in (t1 , . . . , tr ), i.e., if ta [i] = ta [j] for every a ∈ [r]. We
also eliminate any variable xi such that tj [i] = 1 for every j ∈ [r] by replacing xi by the constant 1 in
P . Let P ′ be the resulting polynomial, and let R′ be the corresponding relation. If ar(R′ ) < d + 1, then
by Lemma 20 R′ is preserved by nud+1 and thus by ud+1 as well (see Theorem 29). Otherwise, for each
Q
I ⊂ [d + 1] let αI be the coefficient of the monomial i∈I xi in P ′ , and let χI ∈ {0, 1}d+1 be the tuple such
P
that χI [i] = 1 if and only if i ∈ I. Note that P ′ (χI ) = I ′ ⊆I αI ′ . We find that αI = 0 for every I. Indeed,
α∅ = 0 since 0d+1 ∈ R′ ; and α{i} = 0 for every i since P ′ (χ{i} ) = α{i} + α∅ = α − {i} = 0; and so on,
in order of increasing cardinality of I. Then P ′ is the constantly-zero polynomial, and 1d+1 ∈ R′ , hence
t′ = 1n ∈ R. We have thus shown that relations defined as roots of polynomials of degree d are preserved
by the (d + 1)-universal operation.
In the other direction, the same argument will show that for any pSDI-operation f with |domain(f )| ≤
2d+2 other than the (d+1)-universal operation, it is possible to define a polynomial on (|domain(f )|−2)/2
16
variables and of degree at most d such that the corresponding relation is not preserved by f . Indeed, let
n = (|domain(f ) − 2|)/2 and r = ar(f ), and let t1 , . . . , tr be tuples of arity n such that no tuple
(t1 [i], . . . , tr [i]) is constant and f (t1 , . . . , tr ) = 1n is defined. If n ≤ d, then we may simply consider the
Q
polynomial P (x1 , . . . , xn ) = i∈[n] xi , whose corresponding relation R is not preserved by f . Otherwise,
let I ⊂ [d + 1] be such that χI ∈
/ {t1 , . . . , tr }; this exists since f is not the (d + 1)-universal partial operation.
Let P ′ be the d + 1-variate polynomial with coefficients αJ = 0 if I 6⊆ J, and with αJ = (−1)|J|−|I|
otherwise, for all J ⊂ [d + 1]. Then P ′ (tI ) = 1, and it can be verified that P ′ (tJ ) = 0 for every J ⊂ [d + 1],
J 6= I, whereas P ′ (1d+1 ) = −(−1)d+1−|I| . Hence the relation corresponding to P ′ is not preserved by
f.
Finally, we give one example of a symmetric relation in Inv(u3 ) that has no obvious connection to roots
of polynomials. A Sidon set is a set S ⊆ {0, . . . , n} in which all sums i + j, i, j ∈ S are distinct.
Lemma 25. Let S ⊆ {0, . . . , n} be a Sidon set, and define a relation R(x1 , . . . , xn ) ⊆ {0, 1}n as
n
X
R(x1 , . . . , xn ) ≡ (
i=1
xi ∈ S).
Then R is preserved by u3 .
Proof. Assume that there exists t1 , . . . , t7 ∈ R such that u3 (t1 , . . . , t7 ) = t ∈
/ R. For i ∈ [n], let
xi = (t1 [i], . . . , t7 [i]) be the tuple of values taken by argument i of R in these tuples. Then the tuples xi take
up to 8 different values, partitioned as two constant tuples and three pairs of complementary tuples. Let Xj
for j = 1, 2, 3 be the set of arguments i ∈ [n] such that the tuple xi belongs to the j:th of these pairs, and let
nj be the difference in Hamming weight compared to t if flipping all values belonging to Xj . Let W be the
Hamming weight of t. Then S contains the values W + n1 , W + n2 , W + n1 + n3 and W + n2 + n + 3,
forming two pairs of weights with common difference n3 . Since n3 6= 0, we must have n1 = n2 . By
symmetry, we have n1 = n2 = n3 . But then S contains the values W + n1 , W + n1 + n2 = W + 2n1 , and
W + n1 + n2 + n3 = W + 3n1 , which is a contradiction. Thus nj = 0 for at least one j, hence W ∈ S and
t ∈ R, contradicting the original assumption.
3.2 Structure of minimal non-trivial pSDI-operations
Note that if f is a pSDI-operation, then |domain(f )| = 2k + 2 for some k, since f is defined on the two
constant tuples and since the tuples of the domain can be paired up as (t, t) where t is the complement
of t. Hence, we define the level of a minimal non-trivial pSDI operation f as (|domain(f )| − 2)/2. We
find no examples on level 0 or 1, and the only non-trivial example on level 2 is the 2-edge operation. At
each level k ≥ 3 the partial k-NU and k-universal operations are the unique strongest and weakest minimal
non-trivial pSDI-operation, respectively, whereas the k-edge operation is intermediate. This structure is also
illustrated in Figure 1. We also find that the k-universal operations uk are maximally weak in the sense that
any non-trivial pSDI-operation with a domain of size 2k + 2 can define uk .
We begin with the following lemma, which formalizes one of the main methods of constructing a
(k + 1)-ary partial operation from a k-ary partial operation. We refer to g as an argument padding of f .
Lemma 26. Let f be a k-ary partial operation and let g be a (k + 1)-ary partial operation such
that (1) Proj1,...,k (domain(g)) = domain(f ) and (2) f (x1 , . . . , xk ) = g(x1 , . . . , xk , xk+1 ) for every
(x1 , . . . , xk , xk+1 ) ∈ domain(g). Then g ∈ [f ]s .
Proof. Let f and g be as in the statement, and first construct the (k + 1)-ary partial operation
f ′ (x1 , . . . , xk , xk+1 ) = f (π1k+1 (x1 , . . . , xk , xk+1 ), . . . , πkk+1 (x1 , . . . , xk , xk+1 )).
17
Clearly, f ′ ∈ [f ]s , since it is a composition of f and the projections π1k+1 , . . . , πkk+1 , and it is not difficult to
see that Proj1,...,k (domain(f ′ )) = domain(f ) and that g can be obtained as a subfunction of f ′ . Since [f ]s
is closed under taking subfunctions it follows that g ∈ [f ]s .
The following will aid us in reasoning about minimal non-trivial pSDI-operations.
Lemma 27. Let f be a pSDI-operation with |domain(f )| = 2k + 2, k ≥ 3. Then f is a minimal non-trivial
operation if and only if f is an argument padding of nuk .
Proof. In the one direction, assume that f is a padding of nuk . It is not hard to verify that every subfunction
f ′ of f which is pSDI is a partial projection, and that f is non-trivial. Thus, f is minimal non-trivial. In
the other direction, assume that f is minimal and non-trivial, and let r = ar(f ). Let t1 , . . . , tk be the
non-constant tuples such that f (t1 ) = 0 is defined. For each i ∈ [k], let ji ∈ [r] be such that making f
undefined on ti and its complement ti leaves a subfunction of πjri . It follows that for all a ∈ [k], ta [ji ] 6= 0 if
and only if a = i. Then the arguments j1 , . . . , jk of f define the partial k-NU operation, and f is a padding
of it.
Our claims about the weakest and strongest operations follows from this.
Lemma 28. The following hold.
1. The unique non-trivial non-total pSDI-operation at level k < 3 is the partial 2-edge operation.
2. For any minimal non-trivial pSDI-operation f at level k ≥ 3, we have Inv(nuk ) ⊆ Inv(f) ⊆ Inv(uk ).
k −k−1
3. There are at most 22
distinct minimal non-trivial pSDI-operations at level k.
Proof. 1. It is easy to verify that no non-trivial operation is possible on level 1. Let f be a non-trivial
pSDI-operation on level 2, and let t1 , t2 ∈ domain(f ) be the non-constant tuples such that f (t1 ) = 0.
Consider the options for the pairs (t1 [i], t2 [i]) for i ∈ [ar(f )]. If two distinct positions i, i′ give identical
pairs, then t[i] = t[i′ ] for every t ∈ domain(f ) and i and i′ are redundant arguments in f , which we may
assume does not occur. If t1 [i] = t2 [i] = 0 for some i ∈ [ar(f )] then f is a partial projection. This leaves
three possible arguments, and unless all three exist, f will be a total operation. The remaining case is that
f = e2 .
2. By Lemma 27 f is a padding of nuk , which provides the first inclusion. For the second, we may
assume that f has no redundant arguments, since otherwise f is equivalent to an operation with fewer
arguments. But then by design, uk is a padding of f , and the second inclusion follows.
3. By Lemma 27, we can restrict our attention to paddings of nuk . Since f is a pSDI-operation, it is
defined by the values of the k non-constant tuples t in the domain with f (t) = 0. Let t1 , . . . , tk be those
tuples, and for i ∈ [ar(f )] let t(i) = (t1 [i], . . . , tk [i]). As above, we may assume that t(i) 6= t(j) for all
distinct i, j ∈ [ar(f )]. This leaves at most 2k possible arguments. Furthermore, t(i) cannot be all-zero unless
f is a partial projection, and k arguments are determined by nuk . This leaves 2k − k − 1 arguments, whose
presence or absence defines f .
The inclusion structure between the k-NU, k-edge and k-universal partial operations are now straightforward to prove with these results.
Theorem 29. Let k ≥ 3. Then the following inclusions hold.
1. Inv(e2 ) ⊂ Inv(ek ),
2. Inv(nuk ) ⊂ Inv(ek ) ⊂ Inv(uk ),
18
3. Inv(nuk ) ⊂ Inv(nuk+1 ),
4. Inv(ek ) ⊂ Inv(ek+1 ), and
5. Inv(uk ) ⊂ Inv(uk+1 ).
Proof. For the inclusions, the second item follows from Lemma 28, and every other inclusion follows from
Lemma 26. Indeed, it is readily verified that for every k ≥ 3, ek is an argument padding of ek−1 and nuk+1
is an argument padding of nuk . For the universal operations, let t1 , . . . , tk+1 be the non-constant tuples of
domain(uk+1 ) such that uk+1 (ti ) = 0, i ∈ [k + 1]. Then the tuples t(i) = (t1 [i], . . . , tk+1 [i]), i ∈ [2k+1 − 1]
spell out all (k + 1)-tuples except 0k+1 , without repetition. Consider the subset I ⊂ [ar(uk+1 )] consisting
of indices i such that tk+1 [i] = 0. Note that t(i) for i ∈ I enumerates all k-tuples except 0k , padded with a 0.
It follows that ProjI (uk+1 ) = domain(uk ) and that uk+1 is an argument padding of uk . By Lemma 26 the
inclusion follows.
To show that the inclusions are strict, consider the following: a k-clause is preserved by nu k+1 (Lemma 19)
but not by uk (Lemma 20); a 1-in-k constraint is preserved by e2 but not by nuk (Lemma 22); and the language
Pk−1 of roots of polynomials of degree at most k − 1 is preserved by uk but not by any other operation on
level k by Lemma 24.
Finally, we have an easy consequence in more general terms.
Corollary 30. Let f be a pSDI-operation with |domain(f )| = 2k + 2. Then Inv(f) ⊆ Inv(uk ).
Proof. Let f ′ be an arbitrary minimal pSDI-operation that is a subfunction of f . Then f ′ belongs to some
level k′ ≤ k, hence Inv(f) ⊆ Inv(uk′ ) ⊆ Inv(uk ) by Lemma 28 and Theorem 29.
3.3 Complementary consequences
We now consider some dual questions, i.e., what consequences can we (in general) draw from the information
that some sign-symmetric language Γ is not preserved by f , for some pSDI-operation f ? We begin with an
easy result, which forms the building block of later results.
Lemma 31. Let Γ be a sign-symmetric language which is not preserved by nuk , for some k ≥ 3. Then Γ can
qfpp-define a k-ary symmetric relation R such that R does not contain tuples of weight 0, but does contain
tuples of weight 2.
Proof. Let k ≥ 3 be an arbitrary constant, and let R ∈ Γ be a relation not preserved by nuk of some arity
n = ar(R). Let t1 , . . . , tk ∈ R be witnesses to this, i.e., t = nuk (t1 , . . . , tk ) is defined and t ∈
/ R. Define
t(i) = (t1 [i], . . . , tk [i]).
By sign-symmetry, we may assume that t = 0n . Furthermore, if there is an argument i ∈ [n] such that
t(i) = 0k , then we can find a smaller counterexample by fixing argument i of R to be constantly 0. Thus,
for every i ∈ [n], the tuple t(i) now contains precisely one non-zero value. Let us define a new relation
R′ (x1 , . . . , xk ) of arity k by identifying arguments according to this, i.e., for every position i ∈ [n] such
that t(i) is non-zero in position j ∈ [k], insert variable xj in position i in R. Additionally define R′′ as the
result of the conjunction of all k! applications of R′ with permuted argument order. Then R′′ is a symmetric
relation which contains all tuples of weight 1 but none of weight 0. Thus, Γ qfpp-defines a relation Rk = R′′
as described of every arity k ≥ 3.
By a similar strategy, we have an important result about languages not preserved by the k-universal
operation.
19
Lemma 32. Let Γ be a sign-symmetric language not preserved by uk for some k ≥ 2. Then Γ can qfpp-define
all k-clauses.
Proof. Let R ∈ Γ be a relation not preserved by uk , and let n = ar(R) and r = 2k − 1 be the arity of uk .
Let t1 , . . . , tr ∈ R be such that uk (t1 , . . . , tr ) = t is defined and t ∈
/ R. By sign-symmetry of Γ, we may
assume t = 0n . Create a new relation by identifying all variables xi and xj in R(x1 , . . . , xn ) for which
ta [i] = ta [j] for every a ∈ [r]. Also assume that there is no variable xi such that ta [i] = 0 for every a ∈ [r],
or else replace xi by the constant 0 in R (again by sign-symmetry). This defines a new relation R′ of arity
at most k. Since t ∈
/ {t1 , . . . , tr }, we find that R′ has arity precisely k and contains every possible k-tuple
except 0k , i.e., R′ qfpp-defines a k-clause. By sign-symmetry, Γ qfpp-defines all k-clauses.
3.3.1
Infinitary case
Finally, we consider consequences of a language not being preserved by any operation in a family of
operations.
Theorem 33. Let Γ be a sign-symmetric language that is not preserved by the partial k-NU operation, for
any k. Then one of the following holds.
1. Γ can qfpp-define all k-clauses for every k.
2. Γ can qfpp-define 1-in-k-clauses for every k.
3. There is a fixed prime p such that Γ can qfpp-define relations
k
X
i=1
xi ≡ a
(mod p)
for every 0 ≤ a < p, of every arity k.
Before we proceed with the proof, let us make a simple observation about qfpp-definitions among
symmetric relations.
Lemma 34. Let R be a symmetric n-ary relation, including tuples of weights S ⊆ {0, . . . , n}. Using R, we
can qfpp-define symmetric relations of the following descriptions.
1. Shift down: a relation of arity n − 1 accepting values S ′ = {x − 1 | x ∈ S, x > 0}.
2. Truncate: a relation of arity n − 1 accepting values S ′ = {x ∈ S | x < n}.
3. Grouping: for any integer p > 1, a relation of arity ⌊n/p⌋ accepting values S ′ = {x′ | x′ p ∈ S}.
Proof. These are implemented by, respectively, fixing an argument to 1 in R; fixing an argument to 0 in R;
and grouping arguments of R in groups of size p (after truncating ar(R) to an even multiple of p).
We can now show the result.
Proof of Theorem 33. Let k ≥ 3 be an arbitrary constant. By Szemerédi’s theorem [54] there is a constant
n = N (2k, 1/2(k +1)) such that every set S ⊆ [n] with |S| ≥ n/(2k +1) contains an arithmetic progression
a, a + p, . . . of at least 2k items. Let Rn be a relation produced by Lemma 31 of arity n, and let S be the
accepted weights for Rn . Say that an arithmetic progression a, a + p, . . . is complete in S if S contains all
values {x ∈ {0, . . . , n} | x ≡ a (mod b)}. We consider a few cases.
20
Case: S contains an incomplete arithmetic progression with at least k items. We show that in this case,
Γ can qfpp-define all k-clauses. Let a, a + p, . . . , a + (k − 1)p ∈ S be an arithmetic progression that in
one direction does not continue. If a ≥ p and a − p ∈
/ S, then by shifting, truncating and grouping we can
P
/ S, then we can
qfpp-define the k-ary relation ( i xi ≥ 1); in the other case, if a + kp ≤ n and a + kp ∈
P
similarly qfpp-define the k-ary relation ( i xi < k). In both cases, taking closure under sign-symmetry
shows that we can qfpp-define all k-clauses. This finishes this case.
Case: S is sparse. Assume that |S| < n/(2k + 2) and that S contains no incomplete arithmetic
progression of at least k items. By truncation, we can assume that n is an even multiple of k + 1. By
self-intersecting Rn by its shifted variant, if needed repeated up to k times, we can further ensure that S
contains no pairs i, i + 1, except possibly in a chain n − i, n − i + 1, . . . ending with n, while retaining that
n is a multiple of k. By only doing this as many times as needed, we can be sure that there is at least one
isolated weight w, 0 < w < n, such that tuples of weight w are accepted but not w − 1 or w + 1 (recall that
we start with a relation with 1 ∈ S).
Now partition {0, . . . , n} into windows (0, . . . , k), (k + 1, . . . , 2k + 1), . . . of length k + 1. By the
density of S (which did not increase during our modifications), at least half the windows contain no elements.
We may safely assume n ≥ 5k; thus there is an empty window that is not the first or the last. Let w be
an isolated weight. Then by sliding the window containing w towards the internal empty window, we must
eventually reach a window where there is an isolated weight which is either in position 1 or k − 1 of the
window. This lets us qfpp-define either a 1-in-k-clause or a (k − 1)-in-k-clause; and in the latter case we
get a 1-in-k-clause by negating all variables. Thus if S is sparse and contains no incomplete progressions of
length k, we can qfpp-define a 1-in-k-clause.
Case: S is dense. Finally, we assume that |S| ≥ n/(2k + 2) but does not contain any incomplete
progressions of length k. By Szemerédi’s theorem, S contains at least one complete progression {x | 0 ≤
x ≤ n, x ≡ a (mod p)} for some a and p, with at least 2k entries (i.e., (2k−1)p ≤ n). Let Rn−i be the (n−i)−(k−1)p
ary relation produced by shifting Rn i steps down and consider the relation Rn′ = Rn ∧ Rn−p ∧ . . . ∧ Rn
−i
of arity n − (k − 1)p, with applications of Rn and Rn padded with zeroes as necessary. Then Rn′ is the
union of complete progressions with difference p, since every weight w accepted by Rn′ corresponds to a
progression of length k in S. Furthermore, the same holds for any constant shift Rn′−i of Rn′ , i < p. Note
that Rn′ still has arity at least pk.
Let A ⊆ {0, . . . , p − 1} be the weights a such that Rn′ contains the complete progression with offset a.
Note that 0 ∈
/ A. By shifting and self-intersecting we can reduce to the case that |A| = 1, i.e., the remaining
P
relation is equivalent to ( xi ≡ a (mod p)) for some a and p.
If p ≥ k, then clearly Rn′ qfpp-defines a 1-in-k relation by further shifting and truncation. Thus, if the
difference p of the relations produced this way can grow without bound, then Γ qfpp-defines 1-in-k relations
of all arities k.
Otherwise, if none of the above cases applies infinitely often, then there is a fixed p such that this
P
process produces relations ( xi ≡ a (mod p)) of infinitely many arities k, which leads to the last case
in the theorem. Assume we are in this case. If p is not a prime, we fix a prime p′ that divides p, and let
a′ = a mod (p/p′ ). Shift the relation down by a′ and group the variables into blocks of size p/p′ . Then the
P
remaining relation is equivalent to ( i xi ≡ a′′ (mod p)′ ) for some a′′ . By shifting, and by starting from a
sufficiently large relation with period p′ , we can produce all relations as in the last case in the theorem.
Finally, we note that since k-clauses can qfpp-define the other two kinds of clauses, the same statement
holds with only 1-in-k clauses and the counting relations mod p.
Section summary. In summary of this section, towards the purpose of discussing sign-symmetric languages
Γ such that SAT(Γ) does, or does not, admit an improved algorithm under SETH, we conclude the following.
Recall that ΓkSAT denotes the language of all k-clauses. We find that ΓkSAT is preserved by every minimal
21
operation on level k′ > k (in particular, by nuk+1 ); not preserved by any operation on a level k′ ≤ k;
and that any sign-symmetric language Γ which is not preserved by the k-universal partial operation uk
can qfpp-define ΓkSAT. Assuming SETH, the minimal non-trivial pSDI-operations that preserve Γ therefore
appear to be reasonable proxies for the complexity of SAT(Γ).
Finally, for each level k, there is a language – namely the language of roots of polynomials of degree
less than k – which is preserved by uk but not by any other operation at level k′ ≤ k, and which does admit
an improved algorithm [42]. This shows that any “dichotomy” characterizing sign-symmetric languages
Γ for which SAT(Γ) admits an improved algorithm under SETH, cannot require a minimal non-trivial
pSDI-operation other than uk for some k.
It remains to show that these very mild restrictions, of requiring only the presence of a single non-trivial
pSDI-operation f preserving Γ, can be powerful enough to ensure that SAT(Γ) admits an improved algorithm.
This is our topic of study for the next section.
4 Upper bounds for sign-symmetric satisfiability problems
In this section, we consider the feasibility of designing an improved algorithm directly for Inv(f)-SAT and
Inv(f)-CSP for a minimal non-trivial pSDI-operation f , i.e., an improved algorithm that only uses the abstract
properties guaranteed by such an operation f .
We show this unconditionally for f = e2 and for f = nu3 , over arbitrary finite domains (where the latter
result is only interesting for the non-Boolean case, since the Boolean case is in P). The algorithms for these
cases use, respectively, a Subset Sum-style meet-in-the-middle algorithm and fast matrix multiplication over
exponentially large matrices. These algorithms all work in the extension oracle model.
We also show conditional or partial results. We show two conditional results for partial k-NU operations,
showing that k-NU-CSP admits an improved algorithm in the oracle model if the (k, k − 1)-hyperclique
problem admits an improved algorithm, and that k-NU-SAT admits an improved algorithm in the explicit
representation model if the Erdős-Rado sunflower conjecture [18] holds for sunflowers with k sets. The
first of these results is a direct generalisation of the matrix multiplication strategy; the second uses fast local
search in the style of Schöning [52]. Finally, we also consider the symmetric special case of 3-edge-SAT,
and show that this problem reduces to a problem of finding a unit-coloured triangle in an edge-coloured
graph. This, in turn, follows from fast algorithms for sparse triangle detection. Several of the algorithms
we reduce to have a running time that depends on the matrix multiplication exponent ω; the best currently
known value is ω < 2.373 [39, 55].
Before we begin, we need the following lemma, which shows that if a relation is preserved by a pSDIoperation, then it is possible to view the relation as a relation of smaller arity over a larger domain, which is
preserved by the corresponding partial operation over the larger domain.
Lemma 35. Let R be an n-ary relation over a set of values D, P a polymorphism pattern, and f a partial
operation preserving R and satisfying P . Let I1 . . . , Im be a partition of [n], and RI1 ,...,Im the m-ary
relation
RI1 ,...,Im = {(ProjI1 (t), . . . , ProjIm (t)) | t ∈ R}
over the set of values {ProjI1 (R) ∪ . . . ∪ Proj(Im )(R)}. Then every partial operation f ′ satisfying P over
{ProjI1 (R) ∪ . . . ∪ Proj(Im )(R)} preserves RI1 ,...,Im
Proof. Let k = ar(f ′ ) = ar(f ). Let t1 , . . . , tk ∈ R and let t′1 , . . . , t′k ∈ RI1 ,...,Im be the corresponding
tuples of RI1 ,...,Im . Assume that f ′ (t1 , . . . , tk ) is defined, i.e., (t1 [j], . . . , tk [j]) ∈ domain(f ′ ) for each
j ∈ [k]. Let i ∈ [n] and let Ij be the index set such that i ∈ Ij . Since f ′ (t1 [j], . . . , tk [j]) is defined it must
be an instantiation of a tuple p ∈ P . It follows that (t′1 [i], . . . , t′k [i]) must be an instantiation of p as well,
implying that f (t′1 [i], . . . , t′k [i]) is defined. Hence, f ′ preserves RI1 ,...,Im .
22
n
4.1 An O ∗ (|D| 2 ) algorithm for 2-edge-CSP
Given a binary relation R one can construct a bipartite graph where two vertices x and y have an edge between
them if and only if (x, y) ∈ R. Formally, the vertices V1 ∪ V2 of this graph will consist of the disjoint union
of Proj1 (R) and Proj2 (R), i.e., V1 = {(1, x) | x ∈ Proj1 (R)} and V2 = {(2, x) | x ∈ Proj2 (R)}.
However, whenever convenient, we will not make this distinction and instead assume that V1 = Proj1 (R)
and V2 = Proj2 (R). We say that a binary relation R is rectangular if its bipartite graph representation is a
disjoint union of bicliques.
Lemma 36. Let φD be the partial Maltsev operation over a domain D. Then every binary relation preserved
by φD is rectangular.
Proof. The proof is very similar to the total case, which is essentially folklore in universal algebra. First
note that R is rectangular if and only if a path of length 4 between nodes x, x′ , y, y ′ implies that there is an
edge between x and y ′ . Therefore, let (x, y), (x′ , y), (x′ , y ′ ) ∈ R. But then φD ((x, y), (x′ , y), (x′ , y ′ )) =
(φD (x, x′ , x′ ), φD (y, y, y ′ )) = (x, y ′ ), implying that (x, y ′ ) ∈ R since R is preserved by φD . Hence, R is
rectangular.
If R is an n-ary relation, I1 ∪ I2 a partition of [n], and s ∈ ProjI1 (R), t ∈ ProjI2 (R), we write
s I1×I2 t to denote the n-ary tuple in R satisfying ProjI1 (s I1 ×I2 t) = s and ProjI2 (s I1 ×I2 t) = t. Let
D = {d0 , d1 , . . . , dk−1 } be a finite set of values. We can then order D according to a total order <, by letting
d0 < d1 < . . . < dk−1 . This order easily extends to n-ary tuples s and t over D by letting s < t if and only
if there exists an i ∈ [n] such that Proj1,...,i (s) = Proj1,...,i (t) and s[i + 1] < t[i + 1]. Given a relation R
we say that the tuple t is lex-min if t ∈ R and there does not exist any t′ ∈ R such that t′ 6= t and t′ < t.
Lemma 37. Let R be an n-ary relation preserved by φD and let I1 ∪ I2 be a partition of [n]. Then there
exists a bipartite graph (V, E) where V is the disjoint union of ProjI1 (R) and ProjI2 (R) such that
1. (V, E) is a disjoint union of bicliques,
2. {s, t} ∈ E if and only if s I1×I2 t ∈ R,
3. for every s ∈ V occurring in a biclique C1 ∪ C2 a pair s0 ∈ C1 , t0 ∈ C2 such that s0 is lex-min in C1
and t0 lex-min in C2 can be computed in O(poly(n, |D|)) time in the extension oracle model.
Proof. Consider the binary relation RI1 ,I2 = {(ProjI1 (t), ProjI2 (t)) | t ∈ R} over the set of values
ProjI1 (R) ∪ ProjI2 (R). By Lemma 35 this relation is preserved by φ over the larger domain, and Lemma 36
then implies that RI1 ,I2 is rectangular. Take the bipartite graph representation (V1 ∪ V2 , E) of RI1 ,I2 (which
by the rectangularity property is a disjoint union of bicliques), and thus satisfies property (1). Property
number (2) then follows easily from the construction of the bipartite graph (V1 ∪ V2 , E) since two vertices s
and t are connected with an edge if and only if (s, t) ∈ RI1 ,I2 , which holds if and only if s I1 ×I2 t ∈ R.
For property (3) we need to show that we, given s ∈ V , can compute lex-min representatives of the
biclique C1 ∪ C2 containing s, in polynomial time with respect to n and |D|. Assume without loss of
generality that s ∈ V1 , and order I2 in ascending order as i1 , . . . , i|I2 | . Then determine the smallest value
d1 ∈ D such that s I1×{i1 } (d1 ) is included in the projection ProjI1 ∪{i1 } (R). This can be computed in
polynomial time using the extension oracle. Then continue, by for each i2 , . . . , ij determine the smallest
dj ∈ D such that s I1×{i1 } (d1 ) ∈ ProjI1 ∪{i1 ,...,ij } (R). Let t0 denote the resulting tuple, and observe that
t0 ∈ C2 and that {s, t0 } ∈ E. We then repeat this using the index set I1 in order to obtain a lex-min tuple
s0 such that {s0 , t0 } ∈ E, which again can be done in polynomial time in the extension oracle model.
n
Theorem 38. 2-edge-CSP is solvable in O∗ (|D| 2 ) time in both the extension oracle model and the explicit
representation.
23
Proof. Let (V, C) be an instance of 2-edge-CSP, where V = {x1 , . . . , xn } and C = {C1 , . . . , Cm }. Assume
without loss of generality that n is even, and let I = [ n2 ] and J = [n] \ I. Consider two sets P and Q
constructed as follows. Initially we let P and Q consist of all n2 -ary tuples over D. Then, for each p ∈ P ,
q ∈ Q we enumerate each constraint in the instance containing only variables indexed by I or J and check
whether p or q is contradicted by the constraint. If this is the case we remove p from P or q from Q. More
formally, if p ∈ P and Ri (xi1 , . . . , xik ) ∈ C, k = ar(Ri ), such that {i1 , . . . , ik } ⊆ I, we check whether
Proji1 ,...,ik (p) ∈ Proji1 ,...,ik (Ri ), and similarly for q ∈ Q. Each such step can be done in O(poly(k)) time
in the extension oracle model and in O(k + |Ri |) time if constraints are explicitly represented. By repeating
this for all elements in P and Q we will therefore obtain two sets of partial assignments that do not directly
contradict individual constraints in the input instance.
Next, for each p ∈ P and q ∈ Q create two m-ary tuples p′ and q ′ . By using Lemma 37 we for
each constraint Ci ∈ C will associate the ith element of p′ and q ′ with a representative of the biclique
corresponding to Ci , p, and q. Hence, let Ci = Ri (xi1 , . . . , xik ) ∈ C, k = ar(Ri ), be a constraint. We
distinguish between two cases. First, assume that {i1 , . . . , ik } ⊆ I or that {i1 , . . . , ik } ⊆ J. In this case
we for every t ∈ P ∪ Q let t′ [i] = 1. Second, assume that i1 , . . . , ik ∈ I ∪ J but that {i1 , . . . , ik } 6⊆ I
and {i1 , . . . , ik } 6⊆ J. In other words the constraint contains variables indexed by members of both I and
J. For every p ∈ P compute the lex-min representatives p0 and q0 of the biclique containing p, with respect
to the two index sets Pi = {j | ij ∈ I} and Qi = {j | ij ∈ J}. This can be done in polynomial time via
Lemma 37. Assign the ith value to the tuple p′ the value (p0 , q0 ), and then repeat this for every q ∈ Q.
Let P ′ = {p′ | p ∈ P } and Q′ = {q ′ | q ∈ Q} be the sets resulting from repeating this for every
constraint in the instance. We observe that the combination of p ∈ P and q ∈ Q satisfies a constraint
Ri (xi1 , . . . , xik ) ∈ C if and only if p′ [i] = q ′ [i], due to property (2) in Lemma 37. Hence, the instance is
n
satisfiable if and only if the two sets P ′ and Q′ intersect. Since P ′ and Q′ contain at most |D| 2 tuples, each
n
of length m, this test can easily be accomplished in O∗ (|D| 2 ) time using standard algorithms.
4.2 An O ∗ (|D|
ωn
3
) algorithm for 3-NU-CSP
The algorithm in Section 4.1 used the rectangularity property of binary relations in order to obtain an
ωn
improved algorithm for 2-edge-CSP. In this section we will devise an O∗ (|D| 3 ) time algorithm for 3NU-CSP by exploiting a structural property that is valid for all ternary relations preserved by nu3 . Here,
ω < 2.373 is the matrix multiplication exponent. We will need the following definition.
Definition 39. An n-ary relation R over D is k-decomposable if there for every t ∈
/ R exists an index set
I ⊆ [n], |I| ≤ k, such that ProjI (t) ∈
/ ProjI (R).
In the total case it is known that R is k-decomposable if R is preserved by a total k-ary NU-operation [29].
In general, this is not true for partial NU-operations, but we still obtain the following result.
Lemma 40. Let R be a k-ary relation preserved by nuk . Then R is (k − 1)-decomposable.
Proof. Let t be a k-ary tuple not included in R. Assume that ProjI (t) ∈ ProjI (R) for every index set I ⊆ [k],
|I| < k. But then there must exist t1 , . . . , tk ∈ R such that each ti differ from t in at most one position. This
furthermore implies that nuk (t1 , . . . , tk ) is defined, and therefore also that nuk (t1 , . . . , tk ) = t ∈
/ R. This
contradictions the assumption that nuk preserves R, and we therefore conclude that there must exist an index
set I ⊆ [k] of size at most k − 1, such that ProjI (t) ∈
/ ProjI (R).
ωn
Theorem 41. 3-NU-CSP is solvable in O∗ (|D| 3 ) time in both the extension oracle model and the explicit
representation, where ω < 2.373 is the matrix multiplication exponent.
24
Proof. Let (V, C) be an instance of 3-NU-CSP where V = {x1 , . . . , xn } and C = {C1 , . . . , Cm }. Partition
[n] into three sets I1 , I2 , I3 such that |Ii | = n3 (or, if this is not possible, as close as possible). Let F1 , F2 , F3
n
denote the set of all partial truth assignments corresponding to I1 , I2 , I3 , and observe that |Fi | ≤ |D| 3 . First,
for each partial truth assignment f ∈ Fi , remove it from the set Fi if there exists a constraint in the instance
which is not satisfied by f . This can be done in polynomial time with respect to the number of constraints in
the instance, using a extension oracle query for each constraint. Second, construct a 3-partite graph where
the node set is the disjoint union of F1 , F2 and F3 , and add an edge between two nodes in this graph if and
only if the combination of this partial truth assignment is not contradicted by any constraint in the instance.
Last, answer yes if and only if the resulting graph contains a triangle.
We begin by proving correctness of this algorithm and then analyse its complexity. We first claim that if
the combination of f1 ∈ F1 , f2 ∈ F2 , f3 ∈ F3 does not satisfy a constraint in the instance, then there exists
g1 , g2 ∈ F1 ∪ F2 ∪ F3 which do not satisfy the instance either. Hence, take a constraint R(xi1 , . . . , xik ) ∈ C,
k = ar(R), which is not satisfied by the combination of f1 , f2 , f3 . Let I1′ = {j | ij ∈ I1 }, I2′ = {j | ij ∈ I2 },
and I3′ = {j | ij ∈ I3 } and consider the relation RI1′ ,I2′ ,I3′ = {(ProjI1 (t), ProjI2 (t), ProjI3 (t)) | t ∈ R} over
the set of values ProjI1 (R) ∪ ProjI2 (R) ∪ ProjI3 (R). By Lemma 35 this relation is preserved by the nuk
operation over the larger domain, and it then follows from Lemma 40 that this relation is 2-decomposable.
But then it is easy to see that there must exist partial truth assignments y1 , y2 ∈ F1 ∪ F2 ∪ F3 such that y1
and y2 do not satisfy R(xi1 , . . . , xik ). Hence, if (V, C) is satisfiable, then there clearly exists a triangle in
the 3-partite graph, and if there exists a triangle, then by following the reasoning above, the instance must be
satisfiable.
For the complexity, we begin by enumerating the three sets of partial truth assignments, which takes
n
O(|D| 3 ) time. We then remove any partial truth assignment which is not consistent with the instance, which
increases this by a polynomial factor, depending only on the number of constraints and the extension queries
for each constraint. Similarly, when constructing the 3-partite graph we enumerate all binary combinations
of partial truth assignments from the three sets and check whether they are consistent. After this we check
n
nω
for the existence of a triangle in the resulting graph with O(|D| 3 ) nodes, which can be solved in O(|D| 3 )
time for ω < 2.373, using fast matrix multiplication.
4.3 Strategies for k-NU-SAT
It is easy to see that the strategy used in Theorem 41 extends to reducing k-NU-CSP problems to (k, k − 1)hyperclique, i.e., the problem of finding a k-vertex hyperclique in a (k − 1)-regular hypergraph. Thus we
get the following.
Lemma 42. Assume that (k, k − 1)-hyperclique on n vertices can be solved in time O∗ (nk−ε ) for some
ε > 0. Then k-NU-CSP admits an improved algorithm in the extension oracle model, i.e., an algorithm
′
running in time O∗ (|D|(1−ε )n ) on domain size D and on n variables, for some ε′ > 0.
However, it should be noted that this is a notoriously difficult problem, and there is some evidence against
such results [40]. Thus, we also investigate a less general algorithm that rests on a milder assumption.
4.3.1
k-NU-SAT via local search
We show that subject to a popular conjecture, k-NU-SAT admits an improved algorithm in the explicit
representation model via a local search strategy. To state this we need a few basic definitions. A sunflower
(with k sets) is a collection of k sets S1 , . . . , Sk with common intersection S = S1 ∩ . . . ∩ Sk , called the
core, such that for every pair i, j ∈ [k], i 6= j, we have Si ∩ Sj = S. Note that we may have S = ∅. The
sunflower conjecture [18], in the form we will need, states that for every k there is a constant Ck such that
for every n, every collection of at least Ckn sets of cardinality n contains a sunflower with k petals. This
25
conjecture was the subject of the Polymath 10 collaborative mathematics project, but remains a notorious
open problem. See Alon, Shpilka and Umans [1] for variations of the conjecture and connections to other
problems.
We first show a simple connection between the sunflower conjecture for sunflowers with k sets and
relations R ∈ Inv(nuk ). For convenience, for a set S ⊆ [n] we denote by χnS the tuple t ∈ {0, 1}n such that
for each i ∈ [n], t[i] = 1 is i ∈ S and t[i] = 0 otherwise.
Lemma 43. Let R ⊂ {0, 1}n be a relation with 0n ∈
/ R. Say that a tuple t = χnS is minimal in R if t ∈ R
′
n
/ R. For i ∈ [n], let Fi be the set of minimal tuples in R of Hamming
but for every S ⊂ S we have χS ′ ∈
weight i. If R is preserved by nuk , then Fi does not contain a sunflower of k sets.
Proof. Let Fi be as in the statement, and assume that R is preserved by nuk . Assume that there are distinct
sets S1 , . . . , Sk forming a sunflower with some core S, such that χSj ∈ Fi for every j ∈ [k]. But then
the operation nuk (χS1 , . . . χSk ) is defined, and produces the tuple χS . This contradicts that the tuples are
minimal in R.
We show that the sunflower conjecture is sufficient to allow an improved algorithm.
Lemma 44. Assume that the sunflower conjecture holds for sunflowers with k sets, with some constant Ck .
Let Γ be a sign-symmetric language preserved by nuk . Assume that for every n-ary relation R ∈ Γ and
every p ∈ [n], the minimal tuples in R of Hamming weight at most p can be enumerated in time O∗ (2O(p) ).
Then SAT(Γ) admits an improved algorithm.
Proof. We first show that the assumptions are sufficient to allow a solution for the local search problem for
SAT(Γ), in the following form. Let an instance (V, C) of SAT(Γ) with |V | = n, a tuple t ∈ {0, 1}n , and
an integer p ∈ [n] be provided. We can in O∗ (2O(p) ) time decide whether there is a tuple t′ ∈ {0, 1}n with
Hamming distance at most p from t that satisfies (V, C).
For this, we repeatedly perform the following procedure. Verify whether the present tuple t satisfies
(V, C), and if not, let R(X) be a constraint in C falsified by t, and let I ⊆ [n] be the set of indices
corresponding to the set of variables X. Let s be the sign pattern such that (ProjI (t))s = 0|X| . Note that
Rs ∈ Γ by assumption. We then enumerate the minimal tuples in Rs of Hamming weight at most p, and for
every such tuple t′ , of weight i, let t′′ be the tuple t with bits flipped according to t′ , and recursively solve
the local search problem from tuple t′′ with new parameter p − i. Correctness is clear, since the search is
exhaustive (because we loop through all minimal tuples). We argue that this solves the local search problem
itself in O∗ (2O(p) ) time. For the running time, assume for simplicity that producing the tuples takes O∗ (cp )
time and, for the same constant c, there are at most ci minimal tuples of weight i (by Lemma 43). Up to
polynomial factors, the running time is then bounded by a recurrence
T (p) = cp +
p
X
i=1
ci T (p − i),
which is bounded as T (p) ≤ (2c)p .
From here on, well-known methods can be used to complete the above into an improved algorithm; cf.
Schöning’s algorithm for k-SAT [52] and its derandomization [16], or even restrict the above to monotone
local search instead of arbitrary local search and apply the method of Fomin et al. [20].
In particular, this is allows for an algorithm in the explicit representation model.
Theorem 45. Assume that the sunflower conjecture holds for sunflowers with k sets. Then k-NU-SAT admits
an improved algorithm in the explicit representation model.
26
We leave it as an open question whether access to an extension oracle (also known as an interval oracle)
suffices to solve the local search problem in single-exponential time. The problem, of course, is that the
bounds above only apply to the minimal tuples, and while it is easy to find a single minimal tuple using an
extension oracle, it is less obvious how to test for the existence of a minimal tuple within a given interval.
Meeks [43] showed how a similar result is possible, but her method would require an oracle for finding
minimal satisfying tuples of weight exactly i, which is also not clear how to do.
4.3.2
k-NU-SAT and bounded block sensitivity
Finally, we briefly investigate connections between the nuk partial operation and a notion from Boolean
function analysis known as block sensitivity, introduced by Nisan [44]. See also the book by O’Donnell [45].
We first introduce some temporary notation. For any relation R ⊆ {0, 1}n , let fR : {0, 1}n → {0, 1}
be a function defined as fR (t) = [t ∈ R], i.e., fR (t) = 1 if t ∈ R and fR (t) = 0 otherwise. For a
tuple t ∈ {0, 1}n and a set S ⊆ [n], let tS denote the tuple t with the bits of S flipped. A function
f : {0, 1}n → {0, 1} has block sensitivity bounded by b if for every t ∈ {0, 1}n there are at most b disjoint
sets S1 , . . . , Sb ⊆ [n] such that f (tSi ) 6= f (t) for every i ∈ [b]. We show that nuk can be seen as a one-sided
version of block sensitivity.
Lemma 46. Let R ⊆ {0, 1}n be a relation. Then fR has block sensitivity less than k if and only if both R
and its complement R := {0, 1}n \ R are preserved by nuk .
Proof. In the first direction, assume that f has block sensitivity at least k. Let t ∈ {0, 1}n be a tuple and let
[n] = X0 ∪ . . . ∪ Xk be a partition of [n] into blocks such that for each 1 ≤ i ≤ k, we have f (tXi ) 6= f (t).
Then if f (t) = 1, then the tuples tXi form a witness that R is not preserved by nuk , and if f (t) = 0 they
form a witness against R being preserved by nuk . In the other direction, let t1 , . . . , tk ∈ R be such that
nuk (t1 , . . . , tk ) = t is defined and t ∈
/ R. For i ∈ [k], let Xi be the positions j where t[j] 6= ti [j]. Then
X1 ∪ . . . ∪ Xk forms a subpartition of [n], showing that f has block sensitivity at least k. The case that R is
not preserved by nuk , instead of R, is completely dual.
It is known that a block sensitivity of at most b implies a certificate complexity of at most b2 , i.e., for any
/ R [44]. This
relation R ∈ Inv(nuk ) and any tuple t ∈ R, there are at most b2 bits in t that certify that t ∈
suggests a branching or local search algorithm for SAT(Γ) where Γ contains such relations. However, more
strongly, it implies that R has a decision tree of bounded depth [44], and thus, since k is a constant, that R
only depends on constantly many arguments. Thus, block sensitivity is a significantly stronger restriction
than what nuk imposes.
However, one related question remains. Assume that R is an n-ary relation preserved by nuk , and
which does depend on all its arguments. Is there a non-trivial upper bound on |R|, e.g., does it hold that
|R| ≤ (2 − εk )n for some εk depending on k? A positive answer to this question would imply a trivial
improved algorithm for k-NU-SAT via enumeration of satisfying assignments, constraint by constraint.
4.4 Symmetric 3-edge-SAT
We finish this section with a result showing that a number of special cases of 3-edge-CSP admits an improved
algorithm via sparse triangle finding. The class in particular contains 3-edge-SAT for symmetric relations
R ∈ Inv(e3 ). We begin by characterising the symmetric relations in Inv(e3 ).
Lemma 47. Let R ⊆ {0, 1}n be a symmetric relation preserved by e3 , Let S ⊆ {0, . . . , n} be the weights
accepted by R. Then either S is a complete arithmetic progression (possibly a trivial one, of length 1), or
S = {a, a + b} or S = {n − a, n − a − b} for some a < b.
27
Proof. Let us first make a simpler claim: If a, a + b ∈ S is a pair that does not extend to a complete
progression in S, then either a − b < 0 or a + 2b > n.
To see this, let a, a + b ∈ S, and assume a + 2b ∈
/ S, a + 2b ≤ n. First assume a ≥ b. We subpartition
[n] into one set T0 of size a − b ≥ 0 and three sets Ti of size b, i = 1, 2, 3. This is possible since
a − b + 3b = a + 2b ≤ n. Let t = χT0 ∪...∪T3 and for i = 1, 2, 3 let ti = tTi . Finally, let t4 = tT1 ∪T2 .
Then e3 (t1 , . . . , t4 ) is defined and produces t. Thus we conclude a < b, i.e., a − b < 0. By the symmetric
argument, if a, a + b ∈ S with a − b ≥ 0 and a − b ∈
/ S, then a + 2b > n. This finishes the claim.
Next, assume that |S| > 2 and that S contains some pair a, a + b such that the progression does not
continue. Let b > 0 be the smallest value such that such a pair exists, and again by symmetry assume that
a+ 2b ≤ n; thus a− b < 0. Let c ∈ S \{a, a+ b}. First assume c > a+ 2b. Then we may, similarly to above,
pack sets with |T0 | = a, |T1 | = |T2 | = b, and |T3 | = c − a − 2b, and we have a witness showing a + 2b ∈ S.
But in the remaining cases, c must be involved in a complete progression with either a or a + b, by the choice
of a and b. It is easy to check that this implies the existence of a value c′ ∈ S with a < c′ < a + b, and
that iterating the claim eventually produces an arithmetic progression of step size dividing b, covering a and
a + b, contradicting the assumption that a + 2b ∈
/ S. Thus |S| = 2, i.e., S = {a, a + b}.
In particular, this lemma shows that every symmetric relation in Inv(e2 ) is a simple arithmetic progression.
It also shows that R has a simple-to-compute 2-edge embedding, i.e., R̂ ⊇ R, R̂ ∩ {0, 1}ar(R) = R, and R̂
is preserved by a total 2-edge operation [36], produced by extending S into a complete progression.
We now describe the algorithm. Let R be a relation with arguments X. For a partition X = X1 ∪ X2
and an assignment f to X1 , we refer to the 2-edge label of f as the pair (f0 , g0 ) produced by first extending
f to a lex-min assignment g0 such that (f, g0 ) ∈ R, then extending g0 to a lex-min assignment f0 such that
(f0 , g0 ) ∈ R. Note that this is the same procedure used in the algorithm for 2-edge-CSP.
We extend this to 3-partite graphs as follows. Let the variable set be partitioned as [n] = X ∪ Y ∪ Z,
and define a graph G = (V, E) with partition V = VX ∪ VY ∪ VZ , where the nodes of each part represent
partial assignments as in Section 4.2. For each edge, verify that the corresponding partial assignment is
consistent with each relation in the input instance. We proceed to give labels to edges of G for each relation
R as follows. We assume that for each relation, the “type” of R is known to us (2-edge, 3-NU, or symmetric
3-edge). If R ∈ Inv(nu3 ), all edges get the same label. Otherwise, let R̂ ⊇ R be the 2-edge-embedding of
R (with R̂ = R if R is already 2-edge). Let pq be an edge in G, corresponding to partial assignments p, q.
If one of these assignments, say p, is an assignment to X, then we set the label of pq to the 2-edge label of p
in the partition X ∪ (Y ∪ Z). Otherwise, p ∪ q is an assignment to Y ∪ Z, and we set the label of pq to the
2-edge label of this assignment in X ∪ (Y ∪ Z). We show that this label scheme captures our language.
Lemma 48. Let R be a relation with arguments U , for some U ⊆ [n], and let G = (V, E) and X ∪ Y ∪ Z
be as above. If either R ∈ Inv(e2 ), or R ∈ Inv(nu3 ), or R is Boolean, symmetric and R ∈ Inv(e3 ), then a
triple (f, g, h) with f ∈ VX , g ∈ Vy , h ∈ Vz satisfies R if and only if f gh is a triangle in G where the edges
f g, f h, gh all have the same label.
Proof. Refer to a triangle f gh with all edge labels identical as a single-label triangle. We will also
slightly abuse notation by treating R as a 3-ary relation taking values from VX × VY × VZ . First assume
that R ∈ Inv(e2 ), and recall that R is rectangular. Let f gh be a single-label triangle with shared label
L = (f0 , g0 h0 ); we show that (f, g, h) ∈ R. Since L is the label of the edge gh, it must be that
(f0 , g, h), (f0 , g0 , h0 ) ∈ R, and by the edges f g and f h it must be that (f, g0 , h0 ) ∈ R as well. By the partial
2-edge operation, this implies (f, g, h) ∈ R. Thus every single-label triangle corresponds to a satisfying
assignment.
In the other direction, let (f, g, h) ∈ R. Since R is rectangular, there is a unique lex-min pair (f0 , g0 h0 )
in the biclique containing (f, gh), and both extensions (f, g0 h0 ) and (f0 , gh) are compatible with R. Thus
all three edges get the same label and the algorithm works for R ∈ Inv(e2 ).
28
The case R ∈ Inv(nu2 ) is trivial. Since such a relation is 2-decomposable, the entire verification of
R happens in the stage where edges are filtered, and in the remaining graph, every triangle represents a
satisfying assignment and every triangle is single-label.
Finally, assume R ∈ Inv(e3 ) and is symmetric. If R ∈ Inv(e2 ), then we argue as above. Otherwise, by
Lemma 47, either S = {a, a + b} or S = {n − a, n − a − b} for a < b, and R̂ verifies that each assignment
(f, g, h) has the correct weight when computed mod b. First assume that f gh is a single-label triangle in
G. First assume S = {a, a + b}. By the edge-filtering step, we know that for each of the edges f g, gh, f h
the corresponding partial assignment has weight at most a + b. Thus the total weight of (f, g, h) is at most
(a + b)(3/2) ≤ a + b + (a + b)/2 < a + 2b. Dually, assume S = {n − a − b, n − a}. No edge in f gh has
more than a + b zeroes, thus the total assignment has weight greater than n − a − 2b. In both case, since the
edge-labels work to verify the value mod b, we conclude (f, g, h) ∈ R.
On the other hand, assume (f, g, h) ∈ R. Since the edge labels verify the more permissive relation R̂,
the triangle f gh is a single-label triangle.
The remaining problem can now be solved via algorithms for triangle-finding in sparse graphs.
Theorem 49. Assume a CSP or SAT problem with the following characteristic: for every relation R, either
R ∈ Inv(e2 ) and R is labelled with type e2 , or R ∈ Inv(nu3 ) and R is labelled with type nu3 , or the
language is Boolean, R is a symmetric relation in Inv(e3 ) and R is labelled with type e3 . This problem can
ω+3
be solved in time O∗ (|D| 6 ) in the extension oracle model, where ω < 2.373 is the matrix multiplication
exponent.
Proof. By the description above, we create a 3-partite graph G on 3|D|n/3 vertices (where |D| = 2 in the
Boolean case), and for every edge in G we give it a vector of labels, one label per relation in the input instance.
We refer to this vector as the colour of the edge. Note that a symmetric relation R can be “inspected” using
its extension oracle to find out the set S of accepted weights. By Lemma 48, the instance has a satisfying
assignment if and only if G has a triangle where all edges have the same colour.
This we solve as follows. For every colour c used by an edge in G, we generate the graph Gc consisting of
all edges of colour c. Let mc be the number of edges of Gc , and let N ≤ 3|D|n/3 be the number of vertices
in G. We check if Gc contains a triangle. If Gc is dense enough, then we use the usual triangle-finding
algorithm for this, with running time O∗ (N ω ), otherwise we use an algorithm for triangle finding in sparse
2ω/(ω+1)
graphs. Alon, Yuster and Zwick [2] show such an algorithm with running time O(mc
), where
ω < 2.373 is the matrix multiplication exponent. Hence, the crossover point at which we use the dense
P
algorithm is mc ≥ N (ω+1)/2 =: N α . Summing over all colours, we have c mc ≤ N 2 . Since the algorithm
for sparse graphs has a super-linear running time, the worst case is when we are at the crossover density and
use the sparse algorithm N 2−α times for a cost of O(N ω ) each time. This works out to a total running time
O(N (ω+3)/2 ) for triangle-finding, i.e., the CSP is solved in time O∗ (|D|(ω+3)n/6 ) = O∗ (|D|0.896n ) using
ω = 2.373.
We do not know whether this strategy can be extended to arbitrary relations R ∈ Inv(e3 ), even for a
non-uniform algorithm.
Section summary. We have proven that it is indeed feasible to construct improved algorithms for Inv(p)SAT and Inv(p)-CSP for individual pSDI-operations p. A crucial step for constructing algorithms of this
form is first to identify non-trivial properties of relations invariant under p, which for the partial 2-edge
operation turned out be rectangularity, and for the partial 3-NU operation 2-decomposability. However,
it might not always be the case that every invariant relation satisfies such a clear-cut property, and for
3-edge-SAT we had to settle for an improved algorithm for symmetric relations.
29
For k-NU-CSP and k-NU-SAT we also gave conditional improvements in terms of (k, k−1)-hyperclique
and the sunflower conjecture. At the present, it is too early to say whether these algorithms constitute the
only source of improvement or if more direct arguments are applicable.
5 Lower Bounds
In this section we turn to the problem of proving lower bounds for sign-symmetric SAT problems.
5.1 Lower bounds based on k-SAT
As an easy warm-up, we first consider languages Γ such that SAT(Γ) is at least as hard as k-SAT for some
k. For each k ≥ 3 let ck ≥ 0 denote the infimum of the set {c | k-SAT is solvable in O(2cn ) time}. Under
the ETH, ck > 0 for each k ≥ 3, and for each k ≥ 3 there exists k′ > k such that ck′ > ck [25]. The best
known upper bounds yield ck ≤ 1 − Θ(1/k), but no methods for lower-bounding the values ck are known.
Recall that Lemma 32 gives a condition under which a language Γ can qfpp-define all k-clauses. We
observe the immediate consequence of this.
Lemma 50. Let Γ be a sign-symmetric constraint language not preserved by the k-universal partial operation.
Then SAT(Γ) cannot be solved in time O∗ (2cn ) for any c < ck , even in the non-uniform model.
Proof. By Lemma 32, Γ can qfpp-define all k-clauses. More concretely, there is a finite set Γ′ ⊆ Γ of
relations such that every k-clause has a fixed, finite-sized gadget implementation over Γ′ . Thus, given a
k-SAT instance on n variables, we can produce an equivalent instance of SAT(Γ′ ) in linear time, with the
same variable set.
As a consequence, ck is also a lower bound on the running time for Inv(f)-SAT for every minimal pSDIoperation at level k + 1 and higher. However, this above lemma applies to any sign-symmetric constraint
language, and not just to the special case when Γ = Inv(f). We can also observe a similar consequence for
SETH-hardness.
Corollary 51. Let Γ be a sign-symmetric constraint language not preserved by the k-universal partial
operation for any k. Then assuming SETH, SAT(Γ) does not admit an improved algorithm, even in the
non-uniform model.
Proof. By SETH, there is for every ε > 0 a constant k such that k-SAT cannot be solved in O∗ ((2 − ε)n )
time. By Lemma 50, there is a reduction from k-SAT to SAT(Γ) for this k. Thus, SAT(Γ) does not admit
an improved non-uniform algorithm.
5.2 2-edge-SAT and Subset Sum
Next, we sharpen the connection between Subset Sum and 2-edge-SAT. Recall that an instance of Subset
Sum consists of a set S = {x1 , . . . , xn } of n numbers and a target integer t, with the question of whether
P
there is a set X ′ ⊆ S such that X ′ = t. This can also be phrased as asking for z1 , . . . , zn ∈ {0, 1} such
that
n
X
zi xi = t.
i=1
Also recall from Lemma 21 that such a relation is contained in Inv(e2 ). However, this does not by itself
imply a problem reduction, since an instance or 2-edge-SAT assumes the existence of an extension oracle for
every constraint. We show that such a reduction can be implemented by splitting the above equation apart
into several equations, based on the bit-expansion of t.
30
Theorem 52. If 2-edge-SAT is solvable in O(2cn ) time for c > 0 in the extension oracle model, then Subset
Sum is solvable in O(2(c+ε)n ) time for every ε > 0.
Proof. Let x1 , . . . , xn , t ∈ N be the input to a Subset Sum instance. We will reduce this instance in
subexponential time to a disjunction over 2-edge-SAT instances on n variables each.
We proceed as follows. Harnik and Naor [22] give a randomized procedure for this that reduces a Subset
Sum instance to bit length at most 2n + log ℓ, where ℓ is the bit length of the input. If ℓ ≥ 2n , then we solve
the instance by brute force in time polynomial in the input length, otherwise we are left with an instance of
bit length ℓ′ ≤ 3n.
√
Next, set a parameter k = n, and split the binary expansion of the input integers into k blocks of equal
√
√
length, giving n blocks of length O( n). For each block guess the contribution of the solution to the
target value. Note that the maximum overflow that can carry over to the next block is n, which means that for
a single block there are O(n2 ) options for the contribution within the block. We get at most O(n2k ) = 2o(n)
√
P
guesses in total, after which we have replaced the original equation i zi xi = t by the conjunction of n
√
linear equations, each with a target integer of O( n)
√ bits. This allows us to implement an extension oracle
for every such constraint with a running time of 2O( n) , using the well-known tabulation approach.
This encodes an instance of 2-edge-SAT in the extension oracle model with n variables. Using an
algorithm for this problem, and multiplying its running time by the time required for answering an oracle
query, yields the claimed running time for Subset Sum.
Given that the running time for 2-edge-SAT in the extension oracle model given in this paper matches
the best known running time for Subset Sum, and given that improving the latter is a long-open problem, it
seems at the very least that an improvment to 2-edge-SAT would require significant new ideas.
5.3 Padding formulas
We now give a combinatorial interlude, showing how relations R ⊆ {0, 1}n can be padded with additional
variables such that the new relation lies in Inv(f), for any non-total partial operation f . This will be
leveraged in the next section to finally provide concrete lower bounds on the running time of Inv(f)-SAT for
pSDI-operations f .
For a partial operation p, say of arity k, and a sequence of tuples t1 , . . . , tk , we say that p(t1 , . . . , tk )
is a projective application if p(t1 , . . . , tk ) is either undefined or p(t1 , . . . , tk ) ∈ {t1 , . . . , tk }. Similarly, if
p(t1 , . . . , tk ) is defined and p(t1 , . . . , tk ) ∈
/ {t1 , . . . , tk } we call p(t1 , . . . , tk ) a non-projective application.
Definition 53. Let R ⊆ {0, 1}n be a relation and P a set of Boolean partial operations. A padding of R
with respect to P is an (n + m)-ary relation PR such that (1) Proj1,...,n (PR ) = R, (2) |PR | = |R|, and (3)
PR ∈ Inv(P). A universal padding formula for n ≥ 1 with respect to P is an (n + m)-ary relation U P P
which (1) is a padding of the relation {0, 1}n and (2) p(t1 , . . . , tar(p) ) is a projective application for every
partial operation p ∈ P and every sequence of tuples t1 , . . . , tar(p) ∈ U P P .
Note that if R is a relation and p a k-ary partial operation such that p(t1 , . . . , tk ) is a projective application
for every sequence t1 , . . . , tk ∈ R, then R ∈ Inv(P). In particular this implies that U P P ∈ Inv(P) for every
universal padding formula U P P of P . Also, critically, if U P P is an (n + m)-ary universal padding formula
for a set of partial operations P , and R is an n-ary relation, then the relation R′ (x1 , . . . , xn , y1 , . . . , ym ) ≡
R(x1 , . . . , xn ) ∧ U P P (x1 , . . . , xn , y1 , . . . , ym ) is a padding formula for R. Hence, a universal padding
formula can be viewed as a blueprint which can be applied to obtain a concrete padding formula for any
relation. It is known that if P contains no total operation, then a universal padding formula can be constructed
using a universal hash family [37].
31
Lemma 54. Let P be a finite set of partial operations such that the only total functions in [P ]s are projections.
For every n ≥ 1 there exists an (n + m)-ary universal padding formula U P P such that m ≤ c · n, for a
constant c depending on P .
Proof. See Lagerkvist & Wahlström [37, Lemma 35].
A quick note is in place on the role of universal padding formulas in obtaining lower bounds for Inv(P)SAT, when P is a finite set of partial operations. Note that in a standard “gadget” reduction from CNF-SAT
to some problem SAT(Γ), one would introduce some number of local variables for every clause of the input,
to create an equivalent output formula that only uses constraints from Γ. The existence of padding formulas
does allow us to do this for Inv(P)-SAT, but for lower bounds under SETH this is not useful since we have
no control over the number of additional variables created this way. However, the universality property of
universal padding formulas allow us to reuse the padding variables between different constraints, to produce
an output which only has n + m = O(n) variables in total. The details are given in the next section, but first
we investigate concrete values of the constant c for partial k-edge and k-NU operations.
L
Lemma 55. Let X = {x1 , . . . , xn } be a set of variables, and let y = i∈S xi be the parity sum for a set
S ⊆ [n] chosen uniformly at random. For any tuple t ∈ {0, 1}n , let t′ be t padded by y. Let p be a partial
operation as specified below, let r = ar(p), and let (t1 , . . . , tr ) be a sequence of tuples in {0, 1}n such that
p(t1 , . . . , tr ) is a non-projective application. Then the following hold.
1. If p is the partial 2-edge operation, with r = 3, then the probability that p(t′1 , t′2 , t′3 ) is defined is 3/4.
2. If p is the partial 3-edge operation, with r = 4, then the probability that p(t′1 , . . . , t′4 ) is defined is 1/2.
3. If p is the partial k-NU operation, k ≥ 4, then the probability that p(t′1 , . . . , t′r ) is defined is (2k+2)/2k .
For every weaker operation, e.g., for the partial k-edge or k-universal operations, the probability is
at most this high.
4. If p is the partial k-universal operation, k ≥ 3, then the probability that p(t′1 , . . . , t′r ) is defined is
(k + 1)/2k .
L
Proof. Throughout the proof, we write y(t) = i∈S t[i]. Let us consider each case in turn.
1. We have ar(p) = 3. Let I respectively J be the set of indices i ∈ [n] such that t1 [i] = t2 [i] 6= t3 [i],
respectively, t1 [i] 6= t2 [i] = t3 [i]. Note that both I and J are non-empty since p(t1 , t2 , t3 ) is a non-projective
application. Then p(y(t1 ), y(t2 ), y(t3 )) is undefined if and only if the parity of S ∩ I and S ∩ J are both
odd. Since I and J are disjoint, the probability of this is exactly 1/4.
2. For the partial 3-edge operation, recall from Theorem 28 that p can be constructed by adding a
fictitious argument to the partial 3-NU operation. Hence, the arguments i ∈ [n] such that (t1 [i], . . . , t4 [i]) is
non-constant partition into three sets I1 , I2 , I3 ⊆ [n], and since p(t1 , . . . , t4 ) is a non-projective application,
all three sets must be nonempty. It can be verified that p(y(t1 ), . . . , y(t4 )) is defined if and only if S ∩ Ii is
odd for at most one i ∈ [3]. This happens with exactly 1/2 probability.
3. For the partial k-NU operation, we have ar(p) = k; let p(t1 , . . . , tk ) = t. There are k non-empty
pairwise disjoint sets I1 , . . . , Ik such that ti [j] 6= t if and only if j ∈ Ii , for each i ∈ [k], j ∈ [n]. The tuple
(y(t1 ), . . . , y(tk )) has one value, say b, in every row i ∈ [k] where S ∩ Ii is odd, and another value, 1 − b,
in every row i where S ∩ Ii is odd. Thus p(y(t1 ), . . . , y(tk )) is defined if either S ∩ Ii is odd for at most
one index or S ∩ Ii is even for at most one index; these are 2k + 2 possibilities. For all other 2k − (2k + 2)
possibilities, the operation is undefined. Note that all these possibilities happen with equal probability, since
the sets Ii are non-empty and pairwise disjoint.
4. We have ar(p) = 2k − 1 = r, with the non-constant parts of domain(p) partitioned into k pairs. Let
Ii , i ∈ [k] be the sets of indices j ∈ [n] such that (t1 [j], . . . , tr [j]) belongs to the ith of these pairs, in some
32
enumeration. We claim that p(y(t1 ), . . . , y(tr )) is defined if and only if S ∩ Ii is odd for at most one i ∈ [k].
On the one hand, if this holds, then (y(t1 ), . . . , y(tr )) is contained in pair number i or is constant, and it is
clear that the operation is defined. Otherwise, let S ∩ Ii and S ∩ Ij both be odd, i 6= j. Let t = p(t1 , . . . , tr );
let a ∈ [r] be the argument such that ta [i] 6= t[i] if and only if i ∈ Ii ; let b ∈ [r] be the argument such that
tb [i] 6= t[i] if and only if i ∈ Ij ; and let c ∈ [r] be the argument such that tc [i] 6= t[i] if and only if i ∈ Ii ∪ Ij .
Then the three positions y(ta ), y(tb ), y(tc ) have a pattern that is not compatible with any domain element of
p. It follows that the probability that p(t′1 , . . . , t′r ) is defined is exactly (k + 1)/2k .
Lemma 56. Let p be a partial operation. There are (|domain(p)|)n sequences (t1 , . . . , tar(p) ) of tuples in
{0, 1}n such that p(t1 , . . . , tar(p) ) is defined.
Proof. For every argument i ∈ [n], we choose which element from domain(p) the tuple (t1 [i], . . . , tar(p) [i])
will correspond to. Every such choice results in a distinct sequence of tuples.
Lemma 57. Let R(x1 , . . . , xn , y1 , . . . , ym ) be a padding formula for {0, 1}n , where each yi is a a parity bit
over {x1 , . . . , xn } chosen uniformly at random. Then the following hold.
1. For the partial 2-edge operation, R(x1 , . . . , xn , y1 , . . . , ym ) is a universal padding formula with
probability at least 1 − ε if m ≥ 6.23n + log(1/ε).
2. For the partial 3-edge operation, R(x1 , . . . , xn , y1 , . . . , ym ) is a universal padding formula with
probability at least 1 − ε if m ≥ 3n + log(1/ε).
3. For the partial k-NU operation, k ≥ 4, and for any operation weaker than it, R(x1 , . . . , xn , y1 , . . . , ym )
is a universal padding formula with exponentially small failure probability if m = Ω( logk k n).
Proof. 1. By Lemma 56, there are 6n triples such that p is defined. For each such triple such that the
application of p is non-projective, the probability that it remains defined after the addition of a single random
parity bit is 3/4. Thus after adding t parity bits, the expected number of non-projective triples is at most
6n (3/4)t = 2n log 6−t log(4/3) .
With t = (n log 6)/(log 4/3)+d, this number equals 1/2d , which means that with probability at least 1−1/2d ,
no defined triples remain. The constant factor works out to (log 6)/(log(4/3)) = (1 + log 3)/(2 − log 3) <
6.23.
2. There are 8n tuples (t1 , . . . , t4 ) such that p is defined, and for each of them which is non-projective
the probability of remaining defined after the addition of a single parity bit is 1/2. Thus adding 3n + d parity
bits leaves in expectation at most
8n (1/2)3n+d = 2−d
non-projective tuples, and the probability that no non-projective tuples remain is at least 1 − 1/2d .
3. In the general case, there are (2k + 2)n = 2(1+log(k+1))n defined tuples, and the probability of a
non-projective tuple remaining defined after the addition of a random parity bit is O(k/2k ). Note that
(ck/2k )t = 2(log c+log k−k)t . Thus the expected number of non-projective tuples after t parity bits is at most
′
2(1+log(k+1))n−(k−log k−c )t ,
and it suffices to let t = Ω( logk k n).
We remark that with a padding strategy other than simple parity bits, a significantly lower scaling ratio
may be possible for the partial k-universal operation. However, the advantage of paddding with parity bits is
that the padding can be efficiently inverted, allowing for efficient extension oracles for the padded relation.
33
5.4 Lower bounds in the extension oracle model
In this section we use the bounds obtained in Section 5.3 to obtain lower bounds for Inv(P)-SAT in the
extension oracle model.
Lemma 58. Let U P P be an (n + m)-ary universal padding formula via the construction in Lemma 57.
Let R = {0, 1}k \ {t} for a k-ary tuple t ∈ {0, 1}k . Then there is a polynomial-time extension oracle for
R(x1 , . . . , xk ) ∧ U P P (x1 , . . . , xn , y1 , . . . , ym ).
Proof. Let α : X → {0, 1}, X ⊆ {x1 , . . . , xk , y1 , . . . , ym }, be a partial truth assignment. We need to show
that we can decide if α is consistent with R(x1 , . . . , xk ) ∧ U P P (x1 , . . . , xn , y1 , . . . , ym ) in polynomial time.
First, we check whether α is consistent with the constraint R(x1 , . . . , xk ), which is easy to do due to the
L
representation of R. Second, recall that there for each yi exists an index set Si such that yi = s∈Si xs .
Hence, the partial assignment α together with R(x1 , . . . , xk ) ∧ U P P (x1 , . . . , xn , y1 , . . . , ym ) induces a
system of linear equations over GF(2) where the unknown variables are those unassigned by α. We may thus
solve this system and check whether it has any solution f where f [i] 6= t[i] for some i ∈ [k].
Theorem 59. Let P be a set of partial operations, and set m ≥ cn + log n such that a random paritypadded formula U P P (x1 , . . . , xn , y1 , . . . , ym ) is a universal padding formula with high probability. Then
Inv(P)-SAT cannot be solved in time O∗ (2(1/(c+1)−ε)n ) for any ε > 0, assuming the randomized version of
the SETH is true. In particular, we have the following lower bounds for specific problems:
1. 2-edge-SAT cannot be solved in O(2(c−ε)n time for any ε > 0, where c ≈ 1/7.28.
2. 3-edge-SAT cannot be solved in O(2(c−ε)n time for any ε > 0, where c = 1/3.
3. For k ≥ 4, k-NU-SAT cannot be solved in O(2(c−ε)n ) time for any ε > 0, where c = 1 − Θ( logk k ),
and the same bound holds for the harder problems k-edge-SAT and k-universal SAT.
Proof. Let F be a CNF-SAT instance on variable set X, |X| = n, and compute a random padding formula
U P P (x1 , . . . , xn , y1 , . . . , ym ), with m as stated. We assume that the construction is successful, i.e., that
the resulting relation is a universal padding formula with respect to P . For every clause in the input,
defined on a tuple of variables (xi1 , . . . , xir ), let R(xi1 , . . . , xir ) be the corresponding relation, and let
R′ (xi1 , . . . , xir ) ∧ U P P (x1 , . . . , xn , y1 , . . . , ym ) be the relation as in Lemma 58 (up to the ordering of
variables). Note that we do not need to explicitly enumerate the tuples in this relation, since we may simply
provide the extension oracle proven to exist in Lemma 58. Then the output is a conjunction of Inv(P)-SAT
relations, with a polynomial-time extension oracle for each one, and the resulting instance is equivalent to F.
Since the output instance has n + m = (c + 1) · n variables, an algorithm solving Inv(P)-SAT faster than
the time stated would imply an improved algorithm for CNF-SAT. The bounds for specific problems follow
from the bounds for universal padding formulas computed in Lemma 57.
Finally, we note that the convergence of the lower bounds for k-NU-SAT towards 2n , assuming SETH,
is at a slower rate than the upper bounds for the best known algorithms for k-SAT, which scale as ck ≤
1 − Θ(1/k) [25]. There are also significant differences in problem model (finite language versus infinite
language, and concrete constraints versus extension oracles). It would be interesting to improve these results,
to either improve the convergence rate or provide bounds in some explicit representation model, assuming
SETH.
34
Section summary. We have proven lower bounds under SETH. The bounds obtained in Theorem 59 are
only valid in the extension oracle model, and it does not appear entirely straightforward to extend them to
the explicit representation. However, for 2-edge-SAT we also gave a lower bound subject to the Subset Sum
n
problem, which as remarked is strong evidence that the O∗ (2 2 ) algorithm from Theorem 38 is the best we
could reasonably hope for.
6 Discussions and Conclusions
We have investigated the structure of constraint languages under fine-grained reductions, with a focus on
sign-symmetric Boolean languages, and applied the results to an analysis of the time complexity of NP-hard
SAT problems, in a general setting.
The structural analysis uses an algebraic connection to analyse constraint languages via their partial
polymorphisms. Thereby the structural conclusions are relevant for any problem that takes as input a
constraint formula over some fixed constraint language, under just a few assumptions: (1) that the constraints
in the formula are “crisp” rather than soft, and are required to all be satisfied (as opposed to problems
such as MAX-SAT, where a feasible solution may falsify some constraints); (2) that there are no structural
restrictions of the formula itself (e.g., no bounds on the number of occurrences per variable); and (3) that
the constraint language is sign-symmetric, i.e., allows the free application of negated variables and the use
of constants in constraints. Thus it naturally applies to SAT(Γ) problems, but would also be relevant for
the analysis of problems such as #SAT and optimisation problems, or even parameterized problems such as
Local Search SAT(Γ) – is there a solution within distance k of a given non-satisfying assignment t?
Structural results. The expressive power of sign-symmetric languages is characterised by the restricted
partial polymorphisms in this paper referred to as pSDI-operations. We characterise the structure of all
minimal non-trivial pSDI-operations, and find that they are organised into a hierarchy, whose levels correspond to the problem complexity, with close connections to being able to express the k-SAT languages.
Moreover, we described the weakest and strongest operations on each level. We find that particular families
of pSDI-operations correspond to partially defined versions of well-known algebraic conditions from the
study of CSPs; in particular, the strongest operation at each level k corresponds to the k-NU condition.
Finally, we also give a result in the “vertical” direction of the hierarchy, giving a simple characterisation of
languages not preserved by the partial k-NU operation for any k. By the above discussion, this result should
be of interest also for other inquiries.
Complexity of SAT(Γ) problems. We apply our results to an analysis of the fine-grained time complexity
of SAT(Γ) for sign-symmetric languages, under SETH. We consider previously studied languages with
improved algorithms – i.e., such that SAT(Γ) can be solved in time O∗ (cn ) for some c < 2 – and find that
they correspond well to particular classes of the hierarchy. Conversely, every known language Γ such that
SAT(Γ) is SETH-hard – i.e., admits no improved algorithm assuming SETH – lives entirely outside of the
hierarchy. We also show the feasibility of giving improved algorithms whose correctness relies only and
directly on the above-mentioned pSDI-operations, by showing that known algorithmic strategies such as fast
matrix multiplication and (conjecturally) fast local search can be extended to work for such classes.
Finally, we give complementary lower bounds – for every invariant f as above, there is a constant cf
such that Inv(f)-SAT cannot be solved in O∗ (cn ) time for any c < cf , assuming SETH. These results are
arguably the first of their kind; every previously known concrete lower bound under SETH has either been
for showing that a problem admits no non-trivial algorithm, or has been applied to problems analysed under
more permissive parameters such as treewidth. In particular, 2-edge-SAT is the first SAT problem which
simultaneously has non-trivial upper and lower bounds on the running time under SETH.
35
6.1 The abstract problem and polynomial-time connections
Finally, let us make a short detour to consider what we may call the abstract problem. We have noted that
for every Boolean pSDI-operation f , there is a set of equational conditions that characterise f , similarly to
definitions of varieties in universal algebra, and for every larger domain D, these conditions will uniquely
determine a partial operation over the domain D. Furthermore, these conditions are preserved under taking
powers of the domain, which we have exploited for particular cases of Inv(f)-SAT and Inv(f)-CSP to reduce
input instances to instances of polynomial-time solvable problems on exponentially many variables.
These polynomial-time problem will in general be search problems, like CSPs, and will be preserved by
the same type of operation f , but have a fixed number of variables d and with an unbounded domain size n.
Let us refer to this as the abstract Inv(f)-problem. The question can be raised, for which pSDI-operations f
does such a problem allow improved polynomial-time algorithms?
We refrain from phrasing the question formally, because the polynomial-time complexity may be strongly
affected by details such as constraint representation, but we note that the class of problems defined this way,
unlike the original problems SAT(Γ), contain several problems conjectured not to have such an improvement.
First, we note that every constraint of arity less than d is preserved by the k-NU-type partial operation
with k ≥ d. This in particular includes the k-hyperclique problem for (k − 1)-uniform hypergraphs, which
has been conjectured not to be solvable in time O(nk−ε ) for any ε > 0 and k > 3 [40]. Thus the abstract
d-NU problem does not admit an improved algorithm for d > 3 under this conjecture.
Second, it can be verified that the problem of finding a zero-weight triangle, under arbitrary large edge
weights, if viewed as a single constraint of arity d = 3, is preserved by the corresponding 3-universal partial
operation. It is known that subject to the 3SUM conjecture, this problem cannot be solved in O(n3−ε ) for
any ε > 0 [56].
If we restrict ourselves to the minimal non-trivial pSDI-operations f defined for the Boolean domain
in this paper, this leaves only a small number of concrete problems open under the above conjectures. By
the inclusions we have established, any operation f at a level k > 3 yields an abstract problem as hard as
the k-NU operation. Furthermore, the abstract 3-NU problem does admit an improved algorithm via fast
matrix multiplication. It can be easily checked that up to argument permutation, there are only eight distinct
pSDI-operations f at level 3 of the hierarchy; and by the above discussion, the easiest and the hardest are
(conjecturally) resolved. We consider it an interesting question to investigate the complexity of the problem
for these remaining cases.
6.2 Regarding a dichotomy for sign-symmetric SAT problems
Ignoring for the moment the lower bounds discussed in the previous section, the results throughout our paper
suggest a simple potential dichotomy between NP-complete SAT problems solvable in O(2cn ) time for c < 1
and SAT problems not solvable in O(2cn ) time for any c < 1 unless SETH fails. We can formulate this
conjecture as follows. To simplify the conjecture we restrict ourselves to the non-uniform model.
Conjecture 60. Let Γ be a possibly infinite sign-symmetric Boolean constraint language such that SAT(Γ)
is NP-complete. Then SAT(Γ) admits a non-uniform algorithm with running time in O(2cn ) time for c < 1
if and only if Γ is preserved by a non-trivial pSDI-operation.
Note that by Corollary 51, the negative direction of this conjecture is already known, up to SETH. It
thus remains to consider whether k-universal SAT admits a non-uniform improved algorithm for every k.
Furthermore, as discussed in the Introduction, the class of constraints definable as the roots of bounded-degree
multivariate polynomials represents an example which by Lemma 24 is directly associated with k-universal
SAT, and which has an improved algorithm by Lokshtanov et al. [42]. Thus, the above conjecture at least
represent a kind of Occam’s razor-type extrapolation of least mathematical surprise.
36
However, at the moment this conjecture seems difficult to settle. An extreme negative result, such as the
conclusion that the full problem Inv(f)-SAT admits an improved algorithm only when the abstract Inv(f)problem does, would by Theorem 45 need to refute the sunflower conjecture. A full positive resolution
would need to generalise the result of Lokshtanov et al. [42] to apply based only on a weak abstract condition,
whereas their present algorithm strongly uses properties specific to polynomials. Intermediate outcomes are
of course possible, but would raise further questions of which pSDI-operations f are powerful enough to
guarantee the existence of an improved algorithm.
6.3 Future work
The investigations in this paper leave several concrete open questions, and significant avenues for future
work, regarding all parts of the paper. Let us highlight a few.
Structural aspects. Assuming that the class of partial k-edge operations turn out to be relevant for
the analysis of future problems, it would be valuable to have a set of canonical consequences to a language
not being preserved by any partial k-edge operation, similarly to Theorem 33. To this aim, it may also be
enlightening to fully describe the symmetric relations contained in various classes in the hierarchy.
Another concrete question is regarding the structure of Inv(nuk ) for k > 3. Assume that R ∈ Inv(nuk )
is an n-ary Boolean relation, which depends on every argument. Is there a non-trivial upper bound on |R|?
Extension to CSPs. Many questions remain regarding an extension of the project to CSPs on nonBoolean domains. While the minimal non-trivial pSDI-operations defined in this paper do have higherdomain analogues, via polymorphism patterns, and while these analogues do in some cases have useful
consequences for the complexity of the corresponding CSP, it is not clear that they are in general the only
kind of condition that is relevant for the fine-grained complexity of CSPs. In particular, in the Boolean
domain there is a known correspondence between pSDI-operations and sign-symmetric languages. No such
correspondence has been shown for CSPs in general.
In a different vein, for higher-domain CSPs there are also classes of NP-hard problems whose time
complexity is far better than O∗ (|D|n ), e.g., k-Colouring corresponds to a CSP of domain size |D| = k
and can be solved in O∗ (2n ) time for every k [4]. Arguably, we do not have a good understanding of when
this occurs in general, and we cannot claim that an O(cn ) time algorithm for c < |D| is necessarily an
improvement. A reasonable starting point to mitigate some of these technical difficulties is to initially only
consider consider constraint languages whose total polymorphisms are the projections.
Problems. Let us mention a few concrete algorithmic questions. First of all, by Lemma 25, symmetric
relations defined by Sidon sets are preserved by the 3-universal operation, but they do not seem to be captured
by currently known algorithms for problems in this class. Does the language consisting of all such relations
admin an improved algorithm?
Another problem is to find a generalisation of the algorithm for constraints defined via bounded-degree
polynomials [42], without explicitly using properties specific to polynomials. A different generalisation of
this class was considered by the present authors (see the arXiv version of [36]), in the form of relations
with bounded-degree Maltsev embeddings. Since this properly generalises bounded-degree polynomials, it
is natural to ask whether this class admits an improved algorithm.
More broadly, as remarked earlier, the classification of the expressiveness of sign-symmetric constraint
languages may be of interest for questions other than just satisfiability. The algorithm for 2-edge-SAT, for
instance, can be used to solve the corresponding counting problem, showing that pSDI-operations may be
powerful enough also in other settings. Concrete questions to consider here include improved algorithms for
the counting problem #SAT(Γ) and the parameterized problem Local search SAT(Γ).
Lower bounds. Can the padding scheme be improved to give better asymptotics with respect to the level
k? Recall that the lower bound behaves as a bound of 2 − Θ((log k)/k), whereas all known algorithmic
strategies yield running times of the form (2 − Θ(1/k))n .
37
It would also be very interesting to have a SETH-based lower bound in the explicit representation model.
As discussed earlier the padding construction is valid also in this representation, but is difficult to implement
in practice since the resulting relations may contain exponentially many tuples with respect to the number of
variables.
References
[1] N. Alon, A. Shpilka, and C. Umans. On sunflowers and matrix multiplication. Computational Complexity, 22(2):219–243, 2013.
[2] N. Alon, R. Yuster, and U. Zwick. Finding and counting given length cycles. Algorithmica, 17(3):209–
223, 1997.
[3] L. Barto, A. Krokhin, and R. Willard. Polymorphisms, and How to Use Them. In A. Krokhin and
S. Zivny, editors, The Constraint Satisfaction Problem: Complexity and Approximability, volume 7
of Dagstuhl Follow-Ups, pages 1–44. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, Dagstuhl,
Germany, 2017.
[4] A. Björklund, T. Husfeldt, and M. Koivisto. Set partitioning via inclusion-exclusion. SIAM Journal on
Computing, 39(2):546–563, 2009.
[5] V. G. Bodnarchuk, L. A. Kaluzhnin, V. N. Kotov, and B. A. Romov. Galois theory for Post algebras. I.
Cybernetics, 5:243–252, 1969.
[6] V. G. Bodnarchuk, L. A. Kaluzhnin, V. N. Kotov, and B. A. Romov. Galois theory for Post algebras. II.
Cybernetics, 5:531–539, 1969.
[7] A. Bulatov. A dichotomy theorem for nonuniform CSPs. In Proceedings of the 58th Annual Symposium
on Foundations of Computer Science (FOCS-2017). IEEE Computer Society, 2017.
[8] A. Bulatov and V. Dalmau. A simple algorithm for Mal’tsev constraints. SIAM Journal On Computing,
36(1):16–27, 2006.
[9] A. Bulatov, P. Jeavons, and A. Krokhin. Classifying the complexity of constraints using finite algebras.
SIAM Journal on Computing, 34(3):720–742, Mar. 2005.
[10] C. Calabro, R. Impagliazzo, and R. Paturi. The complexity of satisfiability of small depth circuits.
In Parameterized and Exact Computation, 4th International Workshop (IWPEC 2009), pages 75–85,
2009.
[11] C. Calabro, R. Impagliazzo, and R. Paturi. On the exact complexity of evaluating quantified k-CNF.
Algorithmica, 65(4):817–827, Apr 2013.
[12] M. Couceiro, L. Haddad, V. Lagerkvist, and B. Roy. On the interval of Boolean strong partial clones
containing only projections as total operations. In Proceedings of the 47th International Symposium
on Multiple-Valued Logic (ISMVL-2017), pages 88–93. IEEE Computer Society, 2017.
[13] M. Couceiro, L. Haddad, K. Schölzel, and T. Waldhauser. Relation graphs and partial clones on a
2-element set. In Proceedings of the 44th International Symposium on Multiple-Valued Logic (ISMVL2014), pages 161–166. IEEE Computer Society, 2014.
38
[14] N. Creignou and H. Vollmer. Boolean constraint satisfaction problems: When does Post’s lattice help?
In N. Creignou, P. G. Kolaitis, and H. Vollmer, editors, Complexity of Constraints, volume 5250 of
Lecture Notes in Computer Science, pages 3–37. Springer Berlin Heidelberg, 2008.
[15] M. Cygan, H. Dell, D. Lokshtanov, D. Marx, J. Nederlof, Y. Okamoto, R. Paturi, S. Saurabh, and
M. Wahlström. On problems as hard as CNF-SAT. ACM Transactions on Algorithms, 12(3):41:1–
41:24, 2016.
[16] E. Dantsin, A. Goerdt, E. A. Hirsch, R. Kannan, J. M. Kleinberg, C. H. Papadimitriou, P. Raghavan, and
U. Schöning. A deterministic (2 − 2/(k + 1))n algorithm for k-SAT based on local search. Theoretical
Computer Science, 289(1):69–83, 2002.
[17] E. Dantsin and A. Wolpert. Derandomization of Schuler’s algorithm for SAT. In Proceedings of Theory
and Applications of Satisfiability Testing (SAT-2004), pages 80–88, 2005.
[18] P. Erdős and R. Rado. Intersection theorems for systems of sets. Journal of the London Mathematical
Society, s1-35(1):85–90, 1960.
[19] T. Feder and M. Vardi. The computational structure of monotone monadic SNP and constraint satisfaction: A study through datalog and group theory. SIAM Journal on Computing, 28(1):57–104,
1998.
[20] F. V. Fomin, S. Gaspers, D. Lokshtanov, and S. Saurabh. Exact algorithms via monotone local search.
In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing (STOC 2016),
pages 764–775, 2016.
[21] D. Geiger. Closed systems of functions and predicates. Pacific Journal of Mathematics, 27(1):95–100,
1968.
[22] D. Harnik and M. Naor. On the compressibility of NP instances and cryptographic applications. SIAM
Journal on Computing, 39(5):1667–1713, 2010.
[23] T. Hertli. 3-SAT faster and simpler - unique-SAT bounds for PPSZ hold in general. SIAM Journal on
Computing, 43(2):718–729, 2014.
[24] E. Horowitz and S. Sahni. Computing partitions with applications to the knapsack problem. Journal
of the ACM, 21(2):277–292, Apr. 1974.
[25] R. Impagliazzo and R. Paturi. On the complexity of k-SAT. Journal of Computer and System Sciences,
62(2):367 – 375, 2001.
[26] R. Impagliazzo, R. Paturi, and F. Zane. Which problems have strongly exponential complexity? Journal
of Computer and System Sciences, 63:512–530, 2001.
[27] B. M. P. Jansen and A. Pieterse. Optimal sparsification for some binary CSPs using low-degree
polynomials. In Proceedings of the 41st International Symposium on Mathematical Foundations of
Computer Science (MFCS-2016), volume 58, pages 71:1–71:14, 2016.
[28] P. Jeavons. On the algebraic structure of combinatorial problems. Theoretical Computer Science,
200:185–204, 1998.
[29] P. Jeavons, D. Cohen, and M. Gyssens. Closure properties of constraints. Journal of the ACM,
44(4):527–548, July 1997.
39
[30] P. Jonsson and V. Lagerkvist. An initial study of time complexity in infinite-domain constraint satisfaction. Artificial Intelligence, 245:115–133, 2017.
[31] P. Jonsson, V. Lagerkvist, G. Nordh, and B. Zanuttini. Strong partial clones and the time complexity
of SAT problems. Journal of Computer and System Sciences, 84:52 – 78, 2017.
[32] P. Jonsson, V. Lagerkvist, and B. Roy. Time complexity of constraint satisfaction via universal algebra.
In Proceedings of the 42nd International Symposium on Mathematical Foundations of Computer
Science (MFCS-2017), pages 17:1–17:15, 2017.
[33] M. P. L. Barto, J. Oprsal. The wonderland of reflections. Israel Journal of Mathematics. To appear.
[34] V. Lagerkvist. Strong Partial Clones and the Complexity of Constraint Satisfaction Problems: Limitations and Applications. PhD thesis, Linköping University, The Institute of Technology, 2016.
[35] V. Lagerkvist and B. Roy. A Preliminary Investigation of Satisfiability Problems Not Harder than
1-in-3-SAT. In Proceedings of the 41st International Symposium on Mathematical Foundations of
Computer Science (MFCS-2016), pages 64:1–64:14, 2016.
[36] V. Lagerkvist and M. Wahlström. Kernelization of constraint satisfaction problems: A study through
universal algebra. In Principles and Practice of Constraint Programming - 23rd International Conference (CP 2017), pages 157–171, 2017.
[37] V. Lagerkvist and M. Wahlström. The power of primitive positive definitions with polynomially many
variables. Journal of Logic and Computation, 27(5):1465–1488, 2017.
[38] V. Lagerkvist, M. Wahlström, and B. Zanuttini. Bounded bases of strong partial clones. In Proceedings
of the 45th International Symposium on Multiple-Valued Logic (ISMVL-2015), pages 189–194, 2015.
[39] F. Le Gall. Powers of tensors and fast matrix multiplication. In Proceedings of the International
Symposium on Symbolic and Algebraic Computation (ISSAC-2014), pages 296–303, 2014.
[40] A. Lincoln, V. Vassilevska Williams, and R. Williams. Tight hardness for shortest cycles and paths
in sparse graphs. In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA-2018), pages 1236–1252, 01 2018.
[41] D. Lokshtanov, D. Marx, and S. Saurabh. Known algorithms on graphs of bounded treewidth are
probably optimal. In Proceedings of the Twenty-second Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA-2011), pages 777–789, 2011.
[42] D. Lokshtanov, R. Paturi, S. Tamaki, R. R. Williams, and H. Yu. Beating brute force for systems of
polynomial equations over finite fields. In P. N. Klein, editor, Proceedings of the Twenty-Eighth Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA 2017), pages 2190–2202, 2017.
[43] K. Meeks. Randomised enumeration of small witnesses using a decision oracle. In 11th International
Symposium on Parameterized and Exact Computation (IPEC 2016), pages 22:1–22:12, 2016.
[44] N. Nisan. CREW PRAMs and decision trees. SIAM Journal On Computing, 20(6):999–1007, 1991.
[45] R. O’Donnell. Analysis of Boolean Functions. Cambridge University Press, 2014.
[46] E. Post. The two-valued iterative systems of mathematical logic. Annals of Mathematical Studies,
5:1–122, 1941.
40
[47] B. Romov. The algebras of partial functions and their invariants. Cybernetics, 17(2):157–167, 1981.
[48] T. Schaefer. The complexity of satisfiability problems. In Proceedings of the 10th Annual ACM
Symposium on Theory Of Computing (STOC-1978), pages 216–226. ACM Press, 1978.
[49] D. Scheder and J. P. Steinberger. PPSZ for general k-SAT - making Hertli’s analysis simpler and 3-SAT
faster. In Proceedings of the 32nd Computational Complexity Conference (CCC-2017), pages 9:1–9:15,
2017.
[50] H. Schnoor and I. Schnoor. Partial polymorphisms and constraint satisfaction problems. In N. Creignou,
P. G. Kolaitis, and H. Vollmer, editors, Complexity of Constraints, volume 5250 of Lecture Notes in
Computer Science, pages 229–254. Springer Berlin Heidelberg, 2008.
[51] K. Schölzel. Dichotomy on intervals of strong partial Boolean clones. Algebra Universalis, 73(34):347–368, 2015.
[52] U. Schöning. A probabilistic algorithm for k-SAT and constraint satisfaction problems. In Proceedings
of the 40th Annual Symposium on Foundations of Computer Science (FOCS-1999), pages 410–414,
1999.
[53] R. Schroeppel and A. Shamir. A T = O(2n/2 ), S = O(2n/4 ) algorithm for certain NP-complete
problems. SIAM Journal On Computing, 10(3):456–464, 1981.
[54] E. Szemerédi. On sets of integers containing no k elements in arithmetic progression. Acta Arithmetica,
27:199–245, 1975.
[55] V. V. Williams. Multiplying matrices faster than Coppersmith-Winograd. In Proceedings of the 44th
Symposium on Theory of Computing Conference (STOC 2012), pages 887–898, 2012.
[56] V. V. Williams and R. Williams. Finding, minimizing, and counting weighted subgraphs. SIAM Journal
On Computing, 42(3):831–854, 2013.
[57] D. Zhuk. The proof of CSP dichotomy conjecture. In Proceedings of the 58th Annual Symposium on
Foundations of Computer Science (FOCS-2017). IEEE Computer Society, 2017.
41
| 8cs.DS
|
Compressing and Indexing Stock Market Data
Gianni Decaroli1 and Giovanni Manzini1,2
arXiv:1606.05724v2 [cs.DS] 13 Sep 2016
1
Computer Science Institute, University of Eastern Piedmont, Italy
2
IIT-CNR, Pisa, Italy
Abstract
We show how to build a compressed index for the offline Order Preserving Pattern
Matching problem. Our solution is based on the new approach of decomposing the
sequence to be indexed into an order component, containing ordering information, and
a δ component, containing information on the absolute values. The Order Preserving
Matching problem is then transformed into an exact matching problem on the order
component followed by a verification phase on the δ component. Experiments show
that this approach is viable and it is the first one offering simultaneously small space
usage and fast retrieval.
1
Introduction
The problem of Order Preserving Pattern Matching consists in finding, inside a numerical
sequence T , all subsequences whose elements are in a given relative order. For example, if
the pattern is P = (1, 2, 3, 4, 5) we need to find all increasing subsequences of length five;
so if T = (10, 20, 25, 30, 31, 50, 47, 49) we have a first match starting with the value 10, a
second match starting with the value 20, and no others.
This problem is a natural generalization of the classic exact matching problem where
we search for subsequences whose values are exactly those of the pattern. Order preserving
matching is useful to search for trends in time series like stock market data, biomedical
sensor data, meteorological data, etc..1 In the last few years this problem has received
much attention. Not surprisingly, most of the results are generalization of algorithms and
techniques used for exact matching. In [1, 6, 16, 18] the authors propose solutions inspired
by the classical linear time Knuth-Morris-Pratt and Boyer-Moore algorithms [15, Chap. 2].
In [7] the authors consider the offline problem in which T can be preprocessed and propose
an index that generalizes the classical Suffix Tree data structure [15, Chap. 5]. Finally
in [2, 3, 4, 5, 9, 11] the authors consider approaches based on the concept of filtration and
seminumerical matching [15, Chap. 4].
In this paper we extend to Order Preserving Matching another well known idea of
exact matching: simultaneously compressing and indexing a sequence of values [20]. We
show how to compactly represent a sequence T so that given a pattern P we can efficiently
report all subsequences of T whose elements are in the same relative order as the elements
of P . Our contribution is based on the new idea of decomposing the sequence T into two
components: the order component and the δ component. Informally, the order component
1
The title of the paper is an attempt to find an easy-to-remember name for this problem: we do not
claim our results are particularly suitable to stock market data.
1
stores the information about the relative order of the elements of T inside a window of a
preassigned size, while the δ component contains the information required for reconstructing T given the order component. The order component is stored into a compressed suffix
array while the δ component is stored using an ad-hoc compression technique.
To search for a pattern we compute its ordering information and then we search it in
the compressed suffix array of the order component. Since the information in the order
component is only partial, this search gives us a list of potential candidates which are
later verified using the δ component. In other words, the search in the compressed suffix
array is a sort of filtering phase that uses the index to quickly select a set of candidates,
discarding all other subsequences in T that certainly do not match. This approach can
be seen as a generalization of the offline strategy in [4] as we will comment in the next
section.
The overall efficiency of our approach depends on some parameters of the algorithm
whose influence will be experimentally analyzed in Section 6. The bottom line is that
our index takes roughly the same space as the compressor gzip and can report the order
preserving occurrences of a pattern order of magnitude faster than a scan based algorithm.
2
Problem formulation and previous results
Let T [1, n] denote a sequence of n = |T | numerical values. We write T [i] to denote the i-th
element and T [j, k] to denote the subsequence T [j]T [j + 1] · · · T [k]. Given two sequences
P , Q, we say that they are order isomorphic, and write P ≈ Q, if |P | = |Q| and the
relative order of P ’s and Q’s elements is the same, that is
P [i] ≤ P [j] ⇐⇒ Q[i] ≤ Q[j]
for 1 ≤ i, j ≤ |P |.
(1)
Hence (1, 3, 4, 2) is order isomorphic to (100, 200, 999, 101) but not to (1, 3, 4, 5). Given
a reference sequence T [1, n] and a pattern P [1, p] the order preserving pattern matching
problem consists in finding all subsequences T [i + 1, i + p] such that T [i + 1, i + p] ≈ P [1, p].
In this paper we consider the offline version of the problem in which the sequence T is given
in advance and we are allowed to preprocess it in order to speedup subsequent searches.
For the offline problem, the authors of [7] describe a generalization of the Suffix Tree
data structure that reports the occurrences of a pattern P in O(|P | + occ) time, where
occ is the number of (order-preserving) occurrences. Like the classical Suffix Tree, the
one described in [7] has the drawback of using O(n log n) bits of space. The only other
offline solution we are aware of is the one in [4] where the authors build an index which is
used as a filter to get a list of potential matches. The index used in [4] is a FM-index [10]
built on the binary sequence t1 · · · tn−1 such that ti = 1 iff T [i + 1] > T [i]. The pattern to
be searched is also transformed with the same rule and the corresponding binary string
is searched in the index. This provides a list of potential matches that are later verified
using the actual sequence T .
In this paper, following a well established research line [20], we consider the problem of
designing a compressed index for the Order Preserving Matching problem. We describe a
data structure that uses roughly 30% of the space used by the sequence T , and supports the
efficient retrieval of order isomorphic subsequences. Our starting point is a generalization
of the index in [4]. However, instead of extracting a binary sequence from T we use a
sliding window of size q and we extract an order component defined over the alphabet
{0, 1, . . . , q − 1}. In addition, we compute a δ component containing the information not
2
stored in the order component. We show that these components can be compactly stored
and at the same time used as an index for Order Preserving Matching.
Our solution makes use of compressed suffix array implemented using a FM-index [10].
The reader should be familiar with the basic concepts of this data structure, namely the
fact that it works in terms of ranges of Suffix Array rows (even if the Suffix Array itself is
not available) and its two basic operations: backward search and LF mapping [20].
3
Basic results
In the following we assume that all sequences contain distinct values. If this is not the
case, we force this property assuming that if i < j and entries T [i] and T [j] contain the
same value then T [i] is considered smaller than T [j]. This is the same assumption used
in [16]. In some applications equal values are important and it is natural to require that
equal values in the pattern correspond to equal values in T . Since our algorithms are
slightly more complex for this case we postpone their discussion to Section 5.1.
Given a positive parameter q and a sequence T [1, n] we define a new sequence To [1, n]
with elements over the set {0, 1, . . . , q − 1}. Let iq = max(1, i − q + 1); for i = 1, . . . , n we
define
(
0 if T [i] = min T [iq , i]
To [i] =
(2)
k if T [i − k] = max{T [j] | iq ≤ j < i ∧ T [j] < T [i]}.
In other words, we set To [i] = k > 0 when T [i − k] is the immediate predecessor of T [i] in
the subsequence T [iq , i], that is, T [i − k] is the largest element smaller than T [i] in T [iq , i].
If T [i] is the smallest element in the sequence, then it has no predecessor and To [i] = 0.
For example, if q = 4 for T = (3, 8, 3, 5, −2, 9, 6, 6) it is To = (0, 1, 2, 1, 0, 2, 3, 1) (recall
we are assuming T [1] < T [3] and T [7] < T [8]). We call To the order component for T since
it encodes ordering information for T ’s elements within a size-q window. Formally the
sequence To depends also on q, but since we use the same q for all sequences for simplicity
we omit it from the notation. Representations similar to To , but without a sliding window,
have been used in [7, 16].
Obviously, if P and Q have the same length and P ≈ Q then Po = Qo . However,
we are interested in finding the order preserving occurrences of P within a (much) longer
reference sequence T , so we will make use of the following more general result.
Lemma 1. Let P [1, p] be any pattern and T a reference sequence. Let i be such that
P [1, p] ≈ T [i + 1, i + p]. It is To [i + j] = Po [j] for every j such that
2≤j≤p
(j − To [i + j]) ≥ 1.
and
Proof. Let j be such that the above condition is satisfied and let w = min(q, i + j) and
v = min(q, j). Note that w (resp. v) is the size of the subsequence which is considered for
determining To [i + j] (resp. Po [i + j]). Clearly w ≥ v.
If To [i + j] = 0, then T [i + j] is the smallest element in the subsequence T [i + j −
w + 1, i + j]. A fortiori T [i + j] is the smallest element in T [i + j − v + 1, i + j]. The
hypothesis P [1, p] ≈ T [i + 1, i + p] implies that likewise P [j] must be the smallest element
in P [j − v + 1, j] so it must be Po [j] = 0.
Assume now To [i + j] > 0 and let ℓj = j − To [i + j]. By construction T [i + ℓj ] is the
immediate predecessor of T [i + j] in the subsequence T [i + j − w + 1, i + j]. The condition
3
j − To [i + j] ≥ 1 implies 1 ≤ ℓj < j. We want to show that Po [j] = To [i + j], that is, P [ℓj ]
is the immediate predecessor of P [j] in P [j − v + 1, j].
Since P [1, p] ≈ T [i + 1, i + p] and T [i + ℓj ] < T [i + j] we have P [ℓj ] < P [j]. To complete
the proof we need to show that there is no k, j − v < k < j, such that P [ℓj ] < P [k] < P [j].
But if this were the case we would have T [i + ℓj ] < T [i + k] < T [i + j] contrary to the
assumption that T [i+ℓj ] is the immediate predecessor of T [i+j] in T [i+j −w+1, i+j].
Since it is always To [i+j] ≤ q−1, the above lemma tells us that if P [1, p] ≈ T [i+1, i+p]
we must have To [i + j] = Po [j] for every j, q ≤ j ≤ p. For j = 2, 3, . . . , q − 1 the lemma
establishes that To [i + j] must be either equal to Po [j] or larger than j − 1. Summing up
we have the following corollary.
Corollary 2. Given a reference T [1, n] and a pattern P [1, p] with p > q, if P [1, p] ≈
T [i + 1, i + p] then we must have
To [i + 2, i + p] = x2 x3 · · · xq−1 Po [q] Po [q + 1] · · · Po [p]
(3)
where for j = 2, . . . , q − 1 either xj = Po [j] or xj ≥ j.
In view of the above corollary, our strategy to solve the order preserving matching
problem is to build a compressed index for the sequence To . Then, given a pattern P [1, p]
with p > q, we compute Po [1, p] and then the set of positions i1 , i2 , . . . , im satisfying (3).
Clearly, we can have P [1, p] ≈ T [i + 1, i + p] only if i ∈ {i1 , i2 , . . . , im }. However, since
Corollary 2 states only a necessary condition, some of these positions could be false positives and a verification step is necessary. For the verification step we need the actual
values of the sequence T . Since we are interested in indexing and compressing T , instead
of the original representation we save space using a representation that takes advantage
of the values in To that are stored in the index.
Given T [1, n] and the corresponding ordering component To [1, n], we define a new
sequence Tδ [1, n] as follows. Let Tδ [1] = T [1]. For i = 2, . . . , n let iq = max(1, i − q + 1)
and:
(
min T [iq , i − 1] − T [i] if To [i] = 0
(4)
Tδ [i] =
T [i] − T [i − To [i]]
if To [i] > 0.
Observe that for i ≥ 2, Tδ [i] ≥ 0. Indeed, if To [i] = 0 then by (2) T [i] ≤ min T [iq , i − 1].
If To [i] > 0 since T [i − To [i]] is the immediate predecessor of T [i] in T [iq , i] it is Tδ [i] ≥ 0.
The sequence Tδ is called the δ component of T . While To provides information on the
ordering of T ’s elements, Tδ contains information on the absolute values of T ’s elements.
Together these two sequences contain the same information as T , as shown by the following
lemma.
Lemma 3. Given To [1, n] and Tδ [1, n] it is possible to retrieve T [1, n] in linear time.
Proof. It is T [1] = Tδ [1]. For i = 2, . . . , n let iq = max(1, i − q + 1). From (4) we get
(
min T [iq , i − 1] − Tδ [i] if To [i] = 0,
T [i] =
Tδ [i] + T [i − To [i]]
if To [i] > 0.
Summing up, our approach to compress and index an integer sequence T and support
order preserving pattern matching is the following:
4
1. Select a window size q and build the ordering component To and the δ component Tδ .
2. Build a compressed full-text index for To and compress Tδ taking advantage of To .
The compressed index for To and the compressed representation of Tδ constitute our
index. To search the order preserving occurrences of a pattern of length at least q using
our index:
1. Given P [1, p] compute the corresponding ordering component Po .
2. Use the full-text index for To to find the set of candidate positions i1 , i2 , . . . , im
satisfying Corollary 2.
3. For each candidate position ik use To and Tδ to retrieve T [ik + 1, ik + p] and verify
whether P [1, p] ≈ T [ik + 1, ik + p].
The above description is quite general and can be realized in many different ways. In the
following sections we describe our particular implementation and experimentally measure
its effectiveness.
4
Representation of the components To and Tδ
We represent To [1, n] using a Compressed Suffix Array (csa) consisting of a Huffman shaped
Wavelet Tree [14] built on the BWT of the sequence To (in our experiments we used
the csa wt class from the sdsl-lite library [12]). This approach guarantees a reasonable
compression of To and supports very efficiently the search inside To . More precisely, given
a pattern p the above csa computes in O(p) time the range of rows [b, e] of the Suffix Array
of To which are prefixed by p (see [10, 20] for details). To find the actual position in To of
each occurrence of p, the csa stores the set of Suffix Array entries containing text positions
which are multiple of a previously chosen block size B. Then, for each row r ∈ [b, e] we
walk backward in the text using the LF-map until we reach a marked Suffix Array entry
from which we can derive the position in To of the occurrence that prefixes row r. The
above scheme uses O(n + (n/B) log n) bits of space and can find the position of all (exact)
occurrence of p in To in O(|p| + B occ) time, where occ = e − b + 1 is the number of
occurrences. Clearly, the parameter B offers a trade-off between space usage and running
time.
Having chosen a representation for To we now consider the problem of compactly
storing the information in Tδ [1, n]. We do not need to search inside Tδ however, during
the verification phase, we do need to extract (decompress) the values in random portions
of Tδ . For this reason we split Tδ in blocks of size B (ie the same size used for the blocks
in the csa of To ) and we compress each block independently. The k-th block consists of
the subsequence Tδ [kB + 1, kB + B], except for the last block which has size n mod B.
Additionally, we use a header storing the starting position of each block. Hence, given a
block number we can decompress it in O(B) time.
To compactly represent a block of Tδ we take advantage of the fact that the corresponding values in To are available during compression and decompression. Recalling the
definition of Tδ [i] in (4), we partition the values in T into three classes:
1. those such that T [i] = min T [iq , i] are called minimal;
2. those such that T [i] = max T [iq , i] are called maximal;
5
Name
ibm
prices
tmax
ecg
rwalk
rand
# Values
2,167,147
31,559,990
15,476,000
20,140,000
50,000,000
50,000,000
Description
ibm stock tick data from Sep. 2011 to Feb. 2016
daily, hourly, and 5min US stock prices
max daily temperature from 424 US stations
22 hours and 23 minutes of ECG data
random walk with integer steps in the range [−20, 20]
random integers in the range [−20, 20]
Figure 1: Files used in our experiments. All values are 32-bit integers so the size in bytes
of the files is four times the number of values. All test files and the source code of our
algorithms are available at https://people.unipmn.it/manzini/stockmarket/.
3. all other values are called intermediate.
The class of a value can be easily determined by both compressor and decompressor:
minimal values are those such that Tδ [i] = 0; maximal values are those such that Tδ [i] 6= 0
and T [i − Tδ [i]] = max T [iq , i − 1]; all other values are intermediate. For each block we
define
m = max{Tδ [i] | i is minimal},
M = max{Tδ [i] | i is maximal};
and we store these two values at the beginning of the block. When we encounter a minimal
(resp. maximal) value T [i] we know that the corresponding value Tδ [i] will be in the range
[0, m + 1) (resp. [0, M + 1)). The interval is semi-open since the right extreme is excluded.
When we encounter a intermediate value T [i] we know that Tδ [i] will be in the range
[0, v − T [i − To [i]]) where v is the smallest element in T [iq , i − 1] largest than T [i − To [i]]
(note that if v = T [i − To [i]] + 1 it is Tδ [i] = 0 and there is nothing to encode).
Summing up, compressing a block of the sequence Tδ amounts to compressing a sequence of non-negative integers ℓ1 , ℓ2 , . . . , ℓB with the additional information that for
i = 1, . . . , B, ℓi < wi where the values w1 , . . . , wB are known during both compression and
decompression.2 Let bin(k) denote the binary representation of the integer k. We have
experimented with three different integer encoders.
lsk: Log-skewed coding described in [19] (but probably known for a long time). Encodes
an integer ℓ ∈ [0, w) using at most |bin(w − 1)| bits. If w is not a power of two the
smallest values in the range [0, w) are encoded using less than |bin(w − 1)| bits. The
encoding takes at most log(w) + O(1) bits.
dlt: Delta coding [8]. Encodes an integer ℓ ∈ [0, w) using 2|bin(ℓ + 1)| − 1 bits consisting
of the unary representation of |bin(ℓ + 1)| followed by the bits of bin(ℓ + 1) except
for the most significant bit. This encoding takes 2 log(ℓ + 1) + O(1) bits, is efficient
for small ℓ but does not take advantage of w.
lsd: Log-skewed-delta coding: a new approach combining lsk and dlt. Encodes ℓ ∈ [0, w)
by first using lsk to encode the integer |bin(ℓ + 1)| ∈ [0, |bin(w)|] followed by the
bits of bin(ℓ + 1) except for the most significant bit. The encoding takes at most
log(ℓ + 1) + log log w + O(1) bits.
6
We have tested the above three coders for different values of the window size q and of
the block size B on the collection of test files described in Fig. 1. The results for B = 32
are shown in Table 1 together with the space occupancy of the csa for the sequence To .
We can summarize this first set of experiments as follows:
1. Among the three encoders lsk is the one that compresses better, closely followed
by lsd, while dlt is clearly worse. We have also tested a mixed approach in which
intermediate values are encoded with lsk while minimal and maximal values are
encoded with lsd, but this strategy did not improve over the simpler lsk encoder.
For this reason in rest of the paper delta values are always encoded with lsk.
2. As the window size q increases the cost of storing Tδ decreases while the csa size
increases. This was to be expect since a larger q means that more information is
contained in To .
Summing up, for a given block size B and window size q our “stock market index”
(smi from now on, again with no claim that it is particularly suitable for stock market
data) consists of the csa for To and a compressed representation of Tδ obtained using the
lsk encoder. Table 2 shows the overall space of our index for different values of B and q
compared to the space used by gzip and by the state of the art compressor xz. We can
see that smi’s space usage is essentially at par with gzip’s: it can be smaller or larger
depending on the block size B. As expected xz compression is clearly superior to both.
These data show that smi uses an amount of space similar to modern compressed full text
indices for exact pattern matching [13].
5
Searching a Stock Market Index
Given a pattern P [1, p], to find its order-preserving occurrences in T we compute its order
component Po and use Corollary 2. For example, if q = 4 and Po = (0, 0, 1, 3, 2, 0),
Corollary 2 tells us that in correspondence to Po [4, p] = (3, 2, 0) To must be (3, 2, 0)
(an exact match), while in correspondence of Po [3] = 1, To must be either 1 or 3, and in
correspondence of Po [2] = 0, To must be either 0, 2, or 3. In other words if T [i+1, i+p] ≈ P
then To [i + 2, i + p] must be one of the following 6 sequences:
(0, 1, 3, 2, 0)
(0, 3, 3, 2, 0)
(2, 1, 3, 2, 0)
(2, 3, 3, 2, 0)
(3, 1, 3, 2, 0)
(3, 3, 3, 2, 0).
In general, if the window size is q the number of sequences satisfying Corollary 2 is (q −1)!.
Clearly, not all the (q − 1)! subsequences will necessarily occur inside To . Using To ’s csa
we compute the range of Suffix Array rows which are prefixed by each one of the (q − 1)!
subsequences mentioned above. Recall that the basic operation of a csa is the backward
search in which, given the range of rows prefixed by a substring α and a character c, we
find in O(1) time the range of rows prefixed by cα. This suggests to compute the desired
set of row ranges with a two steps procedure: first (Phase 1) with p−q +1 backward search
steps we compute the range of rows prefixed by P [q, p]; then (Phase 2) with additional
backward search steps we compute the range of rows prefixed by x2 x3 · · · xq−1 P [q] · · · P [p]
for each (q − 2)-tuple x2 , . . . , xq−1 satisfying the conditions of Corollary 2. Phase 2 can
2
While Tδ blocks are compressed independently, To values are defined globally. Hence for the first q − 1
values of a block To may point to a value in the previous block. In this case we store in Tδ also the position
of the predecessor inside the block.
7
B = 32
q=4
B = 32
q=8
B = 32
q = 12
csa
lsk
dlt
lsd
csa
lsk
dlt
lsd
csa
lsk
dlt
lsd
ibm
0.09
0.18
0.19
0.18
0.11
0.17
0.18
0.17
0.12
0.17
0.18
0.17
prices
0.10
0.25
0.30
0.26
0.13
0.24
0.29
0.25
0.14
0.23
0.28
0.25
tmax
0.10
0.22
0.25
0.23
0.13
0.20
0.23
0.21
0.15
0.19
0.23
0.21
ecg
0.09
0.18
0.20
0.19
0.11
0.17
0.20
0.18
0.13
0.17
0.19
0.18
rwalk
0.10
0.24
0.28
0.26
0.13
0.23
0.27
0.25
0.15
0.22
0.27
0.24
rand
0.11
0.24
0.28
0.26
0.14
0.22
0.25
0.23
0.16
0.21
0.25
0.22
Table 1: Space usage of To ’s csa and of Tδ compressed with three different encoders: logskewed (lsk), delta coding (dlt), and log-skewed delta (lsd). The reported values are the
ratio between the size of compressed file over the size of the test file (both expressed in
bytes).
smi q = 4
B = 32 smi q = 8
smi q = 12
smi q = 4
B = 64 smi q = 8
smi q = 12
smi q = 4
B = 96 smi q = 8
smi q = 12
gzip --best
xz --best
ibm
0.26
0.27
0.28
0.20
0.21
0.21
0.18
0.19
0.19
0.22
0.13
prices
0.35
0.37
0.38
0.29
0.30
0.31
0.27
0.28
0.28
0.37
0.24
tmax
0.32
0.33
0.34
0.26
0.26
0.27
0.24
0.24
0.25
0.24
0.17
ecg
0.27
0.29
0.30
0.21
0.22
0.23
0.19
0.20
0.21
0.19
0.12
rwalk
0.34
0.36
0.37
0.28
0.29
0.30
0.25
0.27
0.28
0.36
0.23
rand
0.35
0.36
0.37
0.28
0.28
0.29
0.26
0.26
0.26
0.24
0.18
Table 2: Space usage of the Stock Market Index (smi) for different values of B and q. Each
value is the ratio between the size of the index over the size of the test file, both expressed
in bytes. For completeness we show also the space usage for gzip and xz.
8
require up to q! backward search steps, but the number of steps is also upper bounded by
q times the number of row ranges obtained at the end of the phase, which is usually much
smaller.
At the end of Phase 2 we are left with a set of rows each one representing a position
in T where an order preserving match can occur. To verify if there is actually a match
we have to decompress the corresponding subsequence of T and compare it with P . This
is done (Phase 3) taking again advantage of the properties of To ’s csa. Given a row index
r representing a position in To prefixed by a string x2 x3 · · · xq−1 P [q] · · · P [p] we use the
LF-map to move backward in To until we reach a marked position, that is, a position in
To (and hence in T ) which is a multiple of the block size B (say position ℓB) and marks
the beginning of block ℓ. Each time we apply the LF-map we also obtain a symbol yi of
To hence when we reach the beginning of the block we also have the sequence
y1 y2 · · · yk x2 x3 · · · xq−1 P [q] · · · P [p]
of To values from the beginning of the block till the position corresponding to P [p]. Using
this information and the compressed representation of Tδ (whose blocks are compressed
independently) we are able to retrieve the corresponding T values
T [ℓB + 1]T [ℓB + 2] · · · T [ℓB + k]T [ℓB + k + 1] · · · T [ℓB + k + p − 1]
and determine if there is an actual order preserving match between the last p values of the
above sequence and P [1, p]. If there is match the algorithm outputs the starting position
in T of the matching sequence (the value ℓB + k in the above example).
Phase 3 is usually the most expensive step since the algorithm has to consider one
candidate at a time and for each candidate we need to reach the beginning of the block
containing it. We can therefore expect that its running time is linearly affected by the
block size B. Note that in our implementation Phase 2 and 3 are interleaved: as soon
as we have determined a range of rows prefixed by one of the patterns in Corollary 2 we
execute Phase 3 for all rows in the range before considering any other row range.
5.1
Strict index: Taking care of equal values
In this section we discuss the modifications which must be made to our index and algorithms in the general case when we cannot assume all values in our sequences are distinct.
If equal values can occur, according to definition (1), to have P [1, p] ≈ T [i + 1, i + p] whenever P [j] = P [k] we must have T [i+j] = T [i+k]. In Section 3 we side-stepped this problem
assuming that later occurrences of a value are greater that previous occurrences: Under
that assumption, if P [j] = P [k] with j < k it is acceptable also that T [i + j] < T [i + k].
Since, if P [1, p] ≈ T [i + 1, i + p] according to the stricter definition they are order
isomorphic also under the assumption used in Section 3, we can build an index for T
as described above and only change the verification of the single candidates (Phase 3).
However, there is a non-trivial improvements that can be made to the algorithm to reduce
the number of candidates. The improvement is a consequence of the following lemma.
Lemma 4. In the presence of equal values, if P [1, p] ≈ T [i+1, i+p] and P [j −Po [j]] = P [j]
with Po [j] > 0, then it is To [i + j] = Po [j].
Proof. If Po [j] > 0 and P [j − Po [j]] = P [j] then P [j − Po [j]] must be rightmost value to
the left of P [j] which is equal to P [j]. The hypothesis P [1, p] ≈ T [i + 1, i + p] implies that
the same should be true for T [i + j − Po [j]] with respect to T [i + j]. Hence we must have
To [i + j] = Po [j] as claimed.
9
ssmi q = 4
B = 32 ssmi q = 8
ssmi q = 12
ssmi q = 4
B = 64 ssmi q = 8
ssmi q = 12
ssmi q = 4
B = 96 ssmi q = 8
ssmi q = 12
gzip --best
xz --best
ibm
prices
tmax
ecg
rwalk
rand
0.27
0.28
0.29
0.20
0.21
0.22
0.18
0.19
0.20
0.22
0.13
0.35
0.37
0.39
0.29
0.31
0.32
0.27
0.29
0.30
0.37
0.24
0.32
0.34
0.35
0.26
0.27
0.28
0.24
0.25
0.26
0.24
0.17
0.28
0.30
0.31
0.21
0.23
0.24
0.19
0.21
0.22
0.19
0.12
0.34
0.36
0.38
0.27
0.29
0.30
0.25
0.27
0.28
0.36
0.23
0.35
0.36
0.37
0.28
0.29
0.30
0.26
0.26
0.27
0.24
0.18
Table 3: Space usage of the Strict Stock Market Index (ssmi) for different values of B and
q. Each value is the ratio between the size of the index over the size of the test file, both
expressed in bytes. For completeness we show also the space usage for gzip and xz.
Because of the above Lemma, in the presence of equal values Corollary 2 should be
modified so that for j = 2, . . . , q − 1 we should consider the case xj > j only if Po [j] = 0
or P [j − Po [j]] 6= P [j]. Indeed, by Lemma 4 whenever Po [j] 6= 0 and P [j − Po [j]] = P [j]
we must have To [i + j] = Po [j]. Hence, in the presence of equal values in the pattern we
need to do consider a smaller number of candidates (see the discussion of Phase 2 of the
search algorithm at the beginning of Section 5).
If we are only interested to search in strict mode, ie considering equal values, we can
take advantage of the presence of equal values in T to completely avoid the encoding of
some of the δ components. This will yield a smaller compressed file and faster search. To
achieve this result we slightly change the definition of the order component given in (2) as
follows. When computing To [i], assume the immediate predecessor of T [i] in T [iq , i − 1] is
some value y that appears at least twice in T [iq , i−1]. Say y = T [k1 ] = T [k2 ] = · · · = T [kh ]
with iq ≤ kh < · · · k2 < k1 ≤ i−1. If this is the case, if T [i] = y we set as usual To [i] = i−k1 ;
whereas if T [i] > y we set To [i] = i − k2 (ie we skip an occurrence of y).
Using this scheme, whenever To [i] = k > 0 and there is another occurrence of T [i−k] in
T [iq , i−1], both the coder and the decoder can tell whether T [i] = T [i−k] or T [i] > T [i−k]
depending on the position of the other occurrence of T [i − k]. In particular, if both the
coder and decoder know that T [i] = T [i − k] there is no need to compute and store the
corresponding value Tδ [i]. The space saving for Tδ can be significant when T contain long
stretches of equal values. If the modified procedure is used to compute both To and Po
the search algorithm does not need to be modified.
A completely different approach to search in strict mode is to add an additional bit to
each nonzero entry of To to distinguish whether T [i − To [i]] is equal to T [i] or not. More
precisely, we define the order component as follows:
if T [i] < min T [iq , i − 1]
0
To [i] = 2k
if T [i − k] = max{T [j] | iq ≤ j < i ∧ T [j] ≤ T [i]}, T [i − k] 6= T [i].
2k − 1 if T [i − k] = max{T [j] | iq ≤ j < i ∧ T [j] ≤ T [i]}, T [i − k] = T [i].
With this approach To contains values in the range [0, 2q − 1] so we expect its index to take
10
more space. However, we also expect Tδ to takes less space since values corresponding to
odd entries in To don’t have to be encoded. We have implemented this variant, from now
on called the strict index, and its space usage is shown in Table 3. Comparing Table 3
with Table 2 we see that for the same values B and q the strict index takes only slightly
more space, the difference being about 1% of the original (uncompressed) file size.
6
Experimental results
All tests have been performed on a desktop PC with eight Intel-I7 3.40GHz CPUs running
Linux-Debian 8.3. All tests used a single CPU. Note that Phase 3 of our algorithm can be
easily parallelized using multiple threads (and to some extent also Phase 2), but we leave
this development to future research. All tests involved 1000 patterns of length 15, 20,
and 50 extracted from the same file where the patterns are later searched, so every search
reports at least one occurrence. The patterns where extracted selecting 1000 random
position in the in file. Note that patterns occurring more often are more likely to be
selected so this setting is the least favorable for our algorithm: like all algorithms based
on an index, it is much faster when the pattern does not occur, or occurs relatively few
times.
6.1
Filtering effectiveness
Since Phases 1 and 2 of our algorithms produce a set of candidates that must be verified
in Phase 3, in our first experiment we measure how effective are the first two Phases
in producing only a small number of candidates which are later discarded (that is, how
effective are Phases 1 and 2 in producing a small number of false positives). The results
of this experiment for smi are reported in Table 4. We see that for patterns of length 15
the average number of occurrences is surprisingly high, with a peak of 73000+ average
occurrences for the file ecg (recall the random selection of patterns favors those which
occur more often). Despite this, we see that in most cases for Phase 2 the ratio is smaller
than 1.50, that is, the number of candidates at the end of Phase 2 is less than one and a
half times the number or actual occurrences. Note that the poor performance of Phase 1
for q = 12 was to be expected since when |P | = 15 Phase 1 will only search in To ’s csa the
last 4 symbols of Po which cannot provide an effective filter.
Not surprisingly for patterns of length 20 and 50 both phases are much more effective.
In particular, the number of false positives at the end of Phase 2 is usually smaller than
10% the number of occurrences (ratio below 1.10). Note that in terms of filtering power
there is not a clear winner among the different window sizes. The reason is that a larger
window implies that the information stored in To is “more accurate” but at the same time
Phase 2 will start earlier and be less effective since it will generate (q − 1)! row ranges.
For the search in strict mode the number of false positives as a function of the window
size q follows a similar trend. The data are shown in Table 7 in the Appendix.
6.2
Running times
Our search algorithms have no direct competitors since they are the first ones to combine compression and indexing. We expect them to be faster than scan based (online)
algorithms [2, 6, 16, 18, 5], at least for sufficiently long files, and slower than the offline
algorithms [4, 7] designed without considering the problem of compressing the reference
11
|P | = 15
ibm
prices
tmax
1
2
1
2
1
2
1155.43
2.22
1.23
7.00
1.21
42.73
1.28
293.83
1.97
1.21
8.99
1.19
170.71
1.31
1469.18
1.17
1.01
1.58
1.01
5.20
1.02
73054.84
4.36
1.00
1.33
15.87
13.07
1.06
6.45
6.24
2.46
139.62
5.79
1.05
3.52
1.09
7.75 13745.64 2620.21
1.06
13.32
1.79
|P | = 20
ave # occs
q = 4 Phase 1
Phase 2
ibm
273.43
1.69
1.11
prices
55.45
1.17
1.05
tmax
1168.09
1.05
1.00
ecg
24299.10
1.21
1.05
Phase 1
Phase 2
4.95
1.14
3.42
1.08
1.41
1.01
q = 12 Phase 1
Phase 2
14.38
1.20
9.66
1.11
ibm
1.00
2.14
1.08
2.47
1.00
4.86
1.00
prices
1.00
1.00
1.00
1.00
1.00
1.00
1.00
ave # occs
q = 4 Phase
Phase
q = 8 Phase
Phase
q = 12 Phase
Phase
q=8
|P | = 50
ave # occs
q = 4 Phase
Phase
q = 8 Phase
Phase
q = 12 Phase
Phase
1
2
1
2
1
2
ecg
rwalk
rand
rwalk
1.01
1.40
1.19
rand
1.00
1.02
1.01
2.31
1.04
1.92
1.02
1.00
1.00
1.56
1.01
4.57
1.06
89.67
1.07
1.01
1.00
tmax
380.55
1.02
1.00
1.07
1.00
1.12
1.01
ecg
5.45
2.05
1.78
1.77
1.14
2.53
1.05
rwalk
1.00
1.00
1.00
1.00
1.00
1.00
1.00
rand
1.00
1.00
1.00
1.00
1.00
1.00
1.00
Table 4: False positives as a function of the window size q. The first row shows the average
number of actual occurrences for the patterns in the test set. The other rows show the
ratios between candidates and actual occurrences at the end of Phase 1 and 2.
sequence. We plan a comprensive experimental analysis in a future study; in this section
we are mainly interested in optimizing our algorithms and understanding the influence of
the window size q and of the block size B on the running times.
Table 5 reports the running times of our stock market index (smi) for different values
of q and B. As a reference we report also the running times of a simple scan based
search algorithm in which we check each text position with the verification algorithm
outlined in [5, Sec. 3] (this is the same verification algorithm used by Phase 3 of our
algorithm). We observe that for the largest files (prices, rwalk, rand) smi is at least two
order of magnitudes faster than scan, the difference being more evident when the average
number of occurrences is small. On the other hand, for the file ecg with |P | = 15 there
are 73000+ average occurrences per pattern, and our algorithm is slower than scan. The
explanation is that just to extract 73000 length-15 subsequences smi decompresses more
than one million input values which is roughly 1/20 of the entire file. Since subsequences
are extracted in random order and there are additional overheads due to compression, it is
12
|P | = 15
ave # occs
ibm
1155.43
prices
293.83
tmax
1469.18
ecg
73054.84
rwalk
4.36
rand
1.00
q=4
B = 32 q = 8
q = 12
q=4
B = 64 q = 8
q = 12
scan
3.64
5.09
7.34
6.49
9.19
12.88
17.28
0.94
1.38
2.32
1.71
2.54
4.02
206.44
3.01
4.50
5.77
6.01
8.99
11.55
101.90
220.34
318.10
399.17
407.17
604.65
768.24
129.71
0.15
0.15
1.09
0.26
0.25
1.55
300.89
0.04
0.02
0.08
0.07
0.03
0.11
352.94
|P | = 20
ave # occs
q=4
B = 32 q = 8
q = 12
q=4
B = 64 q = 8
q = 12
scan
ibm
273.43
0.78
1.14
1.64
1.39
2.04
2.93
17.10
prices
55.45
0.14
0.22
0.29
0.26
0.39
0.51
208.34
tmax
1168.09
2.44
4.13
4.73
4.69
7.24
9.79
101.65
ecg
24299.10
77.52
114.90
143.78
139.45
210.68
268.84
127.16
rwalk
1.01
0.01
0.02
0.03
0.02
0.03
0.04
303.01
rand
1.00
0.01
0.02
0.03
0.02
0.03
0.04
367.82
|P | = 50
ave # occs
q=4
B = 32 q = 8
q = 12
q=4
B = 64 q = 8
q = 12
scan
ibm
1.00
0.01
0.02
0.03
0.01
0.02
0.03
16.76
prices
1.00
0.02
0.03
0.04
0.02
0.03
0.04
202.29
tmax
380.55
0.91
1.44
1.92
1.52
2.40
3.20
100.38
ecg
5.45
0.04
0.05
0.07
0.07
0.07
0.09
120.99
rwalk
1.00
0.02
0.03
0.04
0.02
0.04
0.05
296.46
rand
1.00
0.02
0.03
0.04
0.03
0.04
0.05
380.01
Table 5: smi vs scan: Average running times, in milliseconds, for searching 1000 random
patterns of length 15, 20, and 50. Running times do not include the time to load the
compressed index (for smi) or the uncompressed text (for scan).
13
|P | = 15
ave # occs
ibm
92.63
prices
59.70
tmax
1392.66
ecg
2142.49
rwalk
2.38
rand
1.00
q=4
B = 32 q = 8
q = 12
q=4
B = 64 q = 8
q = 12
scan
0.23
0.25
0.29
0.43
0.49
0.57
16.47
0.17
0.21
0.44
0.31
0.41
0.64
252.64
3.04
3.80
3.77
6.39
8.01
8.65
122.92
7.73
9.39
11.98
14.26
18.89
23.65
135.78
0.06
0.09
0.72
0.11
0.13
0.88
406.92
0.02
0.03
0.05
0.03
0.03
0.06
419.95
|P | = 20
ave # occs
q=4
B = 32 q = 8
q = 12
q=4
B = 64 q = 8
q = 12
scan
ibm
24.22
0.06
0.07
0.08
0.11
0.13
0.17
16.59
prices
1.49
0.02
0.03
0.04
0.02
0.04
0.05
246.37
tmax
988.39
2.15
2.73
2.70
4.26
5.54
5.82
121.52
ecg
106.78
0.47
0.56
0.81
0.85
1.07
1.46
132.68
rwalk
1.00
0.01
0.03
0.05
0.02
0.03
0.06
403.77
rand
1.00
0.01
0.03
0.04
0.02
0.04
0.05
410.79
|P | = 50
ave # occs
q=4
B = 32 q = 8
q = 12
q=4
B = 64 q = 8
q = 12
scan
ibm
1.00
0.02
0.02
0.03
0.01
0.02
0.03
15.92
prices
1.00
0.02
0.03
0.05
0.02
0.04
0.06
241.06
tmax
371.99
0.87
1.01
0.97
1.55
1.90
1.87
114.22
ecg
1.00
0.02
0.03
0.04
0.02
0.03
0.04
131.76
rwalk
1.00
0.02
0.04
0.06
0.02
0.04
0.06
392.10
rand
1.00
0.02
0.04
0.06
0.03
0.05
0.07
345.17
Table 6: ssmi vs scan: Average running times, in milliseconds, for searching 1000 random
patterns of length 15, 20, and 50. Running times do not include the time to load the
compressed index (for ssmi) or the uncompressed text (for scan).
14
not surprising that a simple scan is faster. This suggests that to efficiently deal with this
almost pathological number of occurrences it is necessary to use ad hoc techniques as it
was done for other indices [17, 21]. The results for the two smallest files in our collection
(ibm 2 million values, tmax 15 million values) show that our approach clearly outperforms
scan also when the number of occurrences is relatively large and the size of the input is
relatively small.
The results in Table 5 show that doubling the block size B from 32 to 64 usually
doubles smi’s running time which is and indirect confirmation that Phase 3 (the only
phase influenced by B) is the most time consuming. We also see that the smallest running
times are obtained consistently with q = 4, the only exception being rwalk and rand for
|P | = 15. The likely explanation here is that for a smaller q all operations on the Wavelet
Tree underlying the csa are faster and this makes up for the better filtration obtained
in some instances by Phase 1 and 2 for larger values of q. Since, according to Table 1,
q = 4 also gives the highest compression, among the tested values q = 4 appears to be the
optimal choice. We have tested smi also for q = 3 (smi requires q > 2). We do not report
the results here, but it turns out that q = 3 yields essentially the same compression as
q = 4 and very similar search speed (there is not a clear winner among the two).
Table 6 shows the search performance of our strict stock market index (ssmi). Since
the order preserving matching is strict, the average number of occurrences of each pattern
is much smaller (up to 200 times for ecg for |P | = 15) and is at most 2142. Hence, we do
not have the pathological cases we had previously and ssmi was always faster than scan.
For B = 32 and q = 4 the minimum speedup was 17 (for ecg and |P | = 15). For ssmi
we see that q = 4 always gives the smallest space usage and the fastest speed, and that
doubling the block size doubles the running times, showing again that also for ssmi the
cost of Phase 3 dominates the overall cost.
7
Concluding Remarks
In this paper we have proposed a compressed index for the order preserving pattern matching problem, considering both the weak and the strict order preserving matching models.
Our approach is based on the new idea of splitting the original sequence into two complementary components: the order component and the δ component. The problem of finding
the order preserving occurrences of a pattern is transformed into an exact search problem
on the order component followed by a verification phase using the δ component. Experiments show that our index has a space usage similar to gzip and can find order preserving
occurrences much faster than a sequential scan.
Our approach is quite general and improvements could be obtained by changing some
implementation choices. For example, we index the order component using a Wavelet-Tree
based FM-index; to improve the performances for inputs with many (order preserving)
repetitions we can use a different compressed full-text index. The compression of the δ
component can also be radically changed without altering the overall scheme.
References
[1] Djamal Belazzougui, Adeline Pierrot, Mathieu Raffinot, and Stéphane Vialette. Single
and multiple consecutive permutation motif search. In ISAAC, volume 8283 of Lecture
Notes in Computer Science, pages 66–77. Springer, 2013.
15
[2] Domenico Cantone, Simone Faro, and M. Oguzhan Külekci. An efficient skip-search
approach to the order-preserving pattern matching problem. In Stringology, pages
22–35. Department of Theoretical Computer Science, Czech Technical University in
Prague, 2015.
[3] Tamanna Chhabra, Emanuele Giaquinta, and Jorma Tarhio. Filtration algorithms
for approximate order-preserving matching. In SPIRE, volume 9309 of Lecture Notes
in Computer Science, pages 177–187. Springer, 2015.
[4] Tamanna Chhabra, M. Oguzhan Külekci, and Jorma Tarhio. Alternative algorithms
for order-preserving matching. In Stringology, pages 36–46. Department of Theoretical
Computer Science, Czech Technical University in Prague, 2015.
[5] Tamanna Chhabra and Jorma Tarhio. A filtration method for order-preserving matching. Inf. Process. Lett., 116(2):71–74, 2016.
[6] Sukhyeun Cho, Joong Chae Na, Kunsoo Park, and Jeong Seop Sim. A fast algorithm
for order-preserving pattern matching. Inf. Process. Lett., 115(2):397–402, 2015.
[7] Maxime Crochemore, Costas S. Iliopoulos, Tomasz Kociumaka, Marcin Kubica,
Alessio Langiu, Solon P. Pissis, Jakub Radoszewski, Wojciech Rytter, and Tomasz
Walen. Order-preserving incomplete suffix trees and order-preserving indexes. In
SPIRE, volume 8214 of Lecture Notes in Computer Science, pages 84–95. Springer,
2013.
[8] Peter
Elias.
Universal
codeword
sets
and
representations
of
the
integers.
IEEE Trans. Information Theory,
21(2):194–203,
1975.
URL:
http://dx.doi.org/10.1109/TIT.1975.1055349,
doi:10.1109/TIT.1975.1055349.
[9] Simone Faro and M. Oguzhan Külekci.
Efficient algorithms for
the order preserving pattern matching problem.
In Algorithmic Aspects in Information and Management, volume 9778, 2016.
URL:
ttp://www.springerlink.com/index/10.1007/978-3-642-02158-9,
doi:10.1007/978-3-642-02158-9.
[10] Paolo
Ferragina
and
Giovanni
Manzini.
An
experimental study of a compressed index.
Inf. Sci.,
135(1-2):13–28,
2001.
URL:
http://dx.doi.org/10.1016/S0020-0255(01)00098-6,
doi:10.1016/S0020-0255(01)00098-6.
[11] Pawel Gawrychowski and Przemyslaw Uznanski. Order-preserving pattern matching
with k mismatches. In CPM, volume 8486 of Lecture Notes in Computer Science,
pages 130–139. Springer, 2014.
[12] Simon Gog, Timo Beller, Alistair Moffat, and Matthias Petri. From theory to practice:
Plug and play with succinct data structures. In SEA, volume 8504 of Lecture Notes
in Computer Science, pages 326–337. Springer, 2014.
[13] Simon Gog and Matthias Petri.
Optimized succinct data structures for
massive data.
Softw., Pract. Exper., 44(11):1287–1314, 2014.
URL:
http://dx.doi.org/10.1002/spe.2198, doi:10.1002/spe.2198.
16
[14] Roberto Grossi, Ankur Gupta, and Jeffrey Scott Vitter. High-order entropycompressed text indexes. In SODA, pages 841–850. ACM/SIAM, 2003.
[15] D. Gusfield. Algorithms on Strings, Trees, and Sequences: Computer Science and
Computational Biology. Cambridge University Press, 1997.
[16] Jinil Kim, Peter Eades, Rudolf Fleischer, Seok-Hee Hong, Costas S. Iliopoulos, Kunsoo Park, Simon J. Puglisi, and Takeshi Tokuyama. Order-preserving matching.
Theor. Comput. Sci., 525:68–79, 2014.
[17] Sebastian Kreft and Gonzalo Navarro. On compressing and indexing repetitive sequences. Theor. Comput. Sci., 483:115–133, 2013.
[18] Marcin Kubica, Tomasz Kulczynski, Jakub Radoszewski, Wojciech Rytter, and
Tomasz Walen. A linear time algorithm for consecutive permutation pattern matching. Inf. Process. Lett., 113(12):430–433, 2013.
[19] Giovanni Manzini and Marcella Rastero.
A simple
compressor.
Softw., Pract. Exper., 34(14):1397–1411,
http://dx.doi.org/10.1002/spe.619, doi:10.1002/spe.619.
and fast DNA
2004.
URL:
[20] Gonzalo Navarro and Veli Mäkinen. Compressed full-text indexes. ACM Comput. Surv., 39(1), 2007. URL: http://doi.acm.org/10.1145/1216370.1216372,
doi:10.1145/1216370.1216372.
[21] Gonzalo Navarro, Simon J. Puglisi, and Jouni Sirén. Document retrieval on repetitive
collections. In ESA, volume 8737 of Lecture Notes in Computer Science, pages 725–
736. Springer, 2014.
17
Appendix: Additional tables
|P | = 15
ave # occs
q = 4 Phase
Phase
q = 8 Phase
Phase
q = 12 Phase
Phase
|P | = 20
ave # occs
q = 4 Phase
Phase
q = 8 Phase
Phase
q = 12 Phase
Phase
1
2
1
2
1
2
ibm
92.63
2.10
1.01
9.63
1.02
120.22
1.03
prices
59.70
2.43
1.04
6.58
1.03
304.25
1.09
tmax
1392.66
1.15
1.00
1.50
1.00
2.19
1.00
ecg
2142.49
2.28
1.07
8.99
1.04
60.70
1.06
rwalk
2.38
10.20
4.04
142.49
2.63
19105.57
7.68
1
2
1
2
1
2
ibm
24.22
2.42
1.01
7.87
1.02
31.38
1.03
prices
1.49
1.13
1.02
42.19
1.01
140.24
1.01
tmax
988.39
1.05
1.00
1.44
1.00
1.59
1.00
ecg
106.78
3.31
1.24
21.75
1.09
93.16
1.15
rwalk
1.00
1.14
1.07
1.26
1.00
39.29
1.01
rand
1.00
1.00
1.00
1.00
1.00
1.00
1.00
ibm
prices
tmax
ecg
rwalk
rand
1.00
1.00
1.00
1.00
1.00
2.18
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
371.99
1.02
1.00
1.07
1.00
1.12
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
1.00
|P | = 50
ave # occs
q = 4 Phase
Phase
q = 8 Phase
Phase
q = 12 Phase
Phase
1
2
1
2
1
2
rand
1.00
3.58
2.02
1.43
1.00
436.38
1.02
Table 7: False positives as a function of the window size q for the Strict Stock Market Index
(ssmi). The first row shows the average number of actual occurrences for the patterns in
the test set. The other rows show the ratios between candidates and actual occurrences
at the end of Phase 1 and 2.
18
| 8cs.DS
|
1
Knowledge Projection for Effective Design of
Thinner and Faster Deep Neural Networks
arXiv:1710.09505v1 [cs.CV] 26 Oct 2017
Zhi Zhang, Guanghan Ning, and Zhihai He
Abstract—While deeper and wider neural networks are actively pushing the performance limits of various computer vision
and machine learning tasks, they often require large sets of
labeled data for effective training and suffer from extremely
high computational complexity. In this paper, we will develop
a new framework for training deep neural networks on datasets
with limited labeled samples using cross-network knowledge
projection which is able to improve the network performance
while reducing the overall computational complexity significantly.
Specifically, a large pre-trained teacher network is used to
observe samples from the training data. A projection matrix
is learned to project this teacher-level knowledge and its visual
representations from an intermediate layer of the teacher network to an intermediate layer of a thinner and faster student
network to guide and regulate its training process. Both the
intermediate layers from the teacher network and the injection
layers from the student network are adaptively selected during
training by evaluating a joint loss function in an iterative manner.
This knowledge projection framework allows us to use crucial
knowledge learned by large networks to guide the training of
thinner student networks, avoiding over-fitting, achieving better
network performance, and significantly reducing the complexity.
Extensive experimental results on benchmark datasets have
demonstrated that our proposed knowledge projection approach
outperforms existing methods, improving accuracy by up to 4%
while reducing network complexity by 4 to 10 times, which is
very attractive for practical applications of deep neural networks.
Index Terms—Deep neural networks, knowledge projection,
transfer learning, network distillation.
I. I NTRODUCTION
R
ECENTLY, large neural networks have demonstrated
extraordinary performance on various computer vision
and machine learning tasks. Visual competitions on large
datasets such as ImageNet [1] and MS COCO [2] suggest that
wide and deep convolutional neural networks tend to achieve
better performance, if properly trained on sufficient labeled
data with well-tuned hyper-parameters, at the cost of extremely
high computational complexity. Over-parameterization in large
networks seems to be beneficial for the performance improvement [3], [4], however, the requirements for large sets of
labeled data for training and high computational complexity
pose significant challenges for us to develop and deploy deep
neural networks in practice.
First, low power devices such as mobile phones, cloud based
services with high throughput demand, and real-time systems,
have limited computational resources, which requires that the
Z. Zhang, G. Ning and Z. He are with the Department of Electrical and
Computer Engineering, University of Missouri, Columbia, MO, 65203 USA.
network inference or testing should have low computational
complexity. Besides the complexity issue, a large network
often consumes massive storage and memory bandwidth.
Therefore, smaller and faster networks are often highly desired
in real-world applications. Recently, great efforts have been
made to address the network speed issue. A variety of model
compression approaches [5], [6], [7], [8], [9] were proposed
to obtain faster networks that mimic the behavior of large
networks. Second, in practical applications, we often have
access to very limited labeled samples. It is very expensive
to obtain human labeled ground-truth samples for training.
In some applications domains, it is simply not feasible to
accumulate enough training examples for deep networks [10],
[11], [12], [13] .
Interestingly, these two problems are actually coupled together. The network capacity is often positively correlated to
its task complexity. For instance, we would expect a small
network classifier of two classes (e.g. , dog and cat) to achieve
a similar level of accuracy as a significantly larger network
for tens of thousand classes of objects. Existing solutions to
obtaining a fast network on new tasks is often based on a twostep approach: train the network on a large dataset, then apply
model compression or distillation to the network after finetuning or transfer learning on the new dataset [9]. Each step
is performed separately and they are not jointly optimized.
Therefore, how to jointly address the problems of network
compression, speed up, and domain adaptation becomes a very
important and intriguing research problem.
A successful line of work [14], [9], [15], [16], [17] suggest
that cumbersome large neural networks, despite their redundancy, have very robust interpretation of training data. By
switching learning targets from labels to interpreted features
in small networks, we have observed not only speed-ups but
also performance improvements. Inspired by this phenomenon,
we are interested to explore if this interpretation power is still
valid across different (at least similar) domains, and to what
extent of performance a newly trained student network can
achieve with the help of a large model pre-trained on different
datasets.
In this paper, we propose a Knowledge Projection Network
(KPN) with a two-stage joint optimization method for training
small networks under the guidance of a pre-trained large
teacher network, as illustrated in Figure 1. In KPN, a knowledge projection matrix is learned to extract distinctive representations from the teacher network, and used to regularize the
training process of the student network. We carefully design
the teacher-student architecture and joint loss function so that
the smaller student network can benefit from extra guidance
2
Learned Projection
Large
Teacher
Network
Testing
[OFF]
Desired
Network
Training
[ON]
Training
Small domain dataset
Fig. 1. System overview. We apply learned projection during training to
guide a standard thinner and faster network for inference on a smaller domain
dataset.
while learning towards specific tasks. Our major observation
is that, by learning necessary representations from a teacher
network which is fully trained on a large dataset, a student
network can disentangle the explanatory factors of variations
in the new data and achieve more precise representation of
the new data from a smaller number of examples. Thus,
same level performance can be achieved using a smaller network. Extensive experimental results on benchmark datasets
have demonstrated that our proposed knowledge projection
approach outperforms existing methods, improving accuracy
by up to 4% while reducing network complexity by 4 to 10
times, which is very attractive for practical applications of
deep neural networks.
Our contributions in this paper are summarized as follows:
(1) we propose a new architecture to transfer the knowledge
from a large teacher network pre-trained on a large dataset into
a thinner and faster student network to guide and facilitate
its training on a smaller dataset. Our approach addresses
the issues of network adaptation and model compression at
the same time. (2) We have developed a method to learn a
projection matrix which is able to project the visual features
from the teacher network into the student network to guide
its training process and improve its overall performance. (3)
We have developed an iterative method to select the optimal
path for knowledge projection between the teacher and student
networks. (4) We have implemented the proposed method in
MXNet and conducted extensive experiments on benchmark
datasets to demonstrate that our method is able to significantly
reduce the network computational complexity by 4-10 times
while largely maintaining or even improving the network
performance by a significant margin.
The rest of this paper is organized as follows. Related work
is reviewed in Section II. We present the proposed Knowledge
Projection Network in Section III. Experimental results are
presented in Section IV. Finally, Section V concludes this
paper.
II. R ELATED W ORK
Large neural networks have demonstrated extraordinary
performance on various computer vision and machine learning
tasks. During the past a few years, researchers have been
investigating how to deploy these deep neural networks in
practice. There are two major problems that need to be
carefully addressed: the high computational complexity of the
deep neural network and the large number labeled samples
required to train the network [18], [19]. Our work is closely
related to domain adaptation and model compression, which
are reviewed in this section.
To address the problem of inadequate labeled samples for
training, methods for network domain adaptation [20], [12],
[21] have been developed, which enable learning on new
domains with few labeled samples or even unlabeled data.
Transfer learning methods have been proposed over the past
several years, and we focus on supervised learning where
a small amount of labeled data is available. It has been
widely recognized that the difference in the distributions of
different domains should be carefully measured and reduced
[21]. Learning shallow representation models to reduce domain discrepancy is a promising approach, however, without
deeply embedding the adaptation in the feature space, the
transferability of shallow features will be limited by the taskspecific variability. Recent transfer learning method coupled
with deep networks can learn more transferable representations
by embedding domain adaptations in the architecture of deep
learning [22] and outperforms traditional methods by a large
margin. Tzeng et al. [13] optimizes domain invariance by
correcting the marginal distributions during domain adaptation. The performance has been improved, but only within
a single layer. Within the context of deep feed-forward neural networks, fine-tune is an effective and overwhelmingly
popular method [23], [24]. Feature transferability of deep
neural networks has been comprehensively studied in [25].
It should be noted that this method does not apply directly
to many real problems due to insufficient labeled samples in
the target domain. There are also some shallow architectures
[26], [27] in the context of learning domain-invariant features.
Limited by representation capacity of shallow architectures,
the performance of shallow networks is often inferior to that
of deep networks [21].
With the dramatically increased demand of computational
resources by deep neural networks, there have been considerable efforts to design smaller and thinner networks from
larger pre-trained network in the literature. A typical approach
is to prune unnecessary parameters in trained networks while
retaining similar outputs. Instead of removing close-to-zero
weights in the network, LeCunn et al. proposed Optimal Brain
Damage (OBD) [5] which uses the second order derivatives
to find trade-off between performance and model complexity.
Hassibi et al. followed this work and proposed Optimal Brain
Surgeon (OBS) [6] which outperforms the original OBD
method, but was more computationally intensive. Han et al.
[28] developed a method to prune state-of-art CNN models
without loss of accuracy. Based on this work, the method of
deep compression [7] achieved better network compression
3
ratio using ensembles of parameter pruning, trained quantization and Huffman coding, achieved 3 to 4 times layer-wise
speed up and reduced the model size of VGG-16 [29] by
49 times. This line of work focuses on pruning unnecessary
connections and weights in trained models and optimizing for
better computation and storage efficiency.
Various factorization methods have also been proposed to
speed up the computation-intensive matrix operations which
are the major computation in the convolution layers. For example, methods have been developed to use matrix approximation
to reduce the redundancy of weights. Jenderberg et al. [8] and
Denton et al. [30] use SVD-based low rank approximation. For
example, Gong et al. [31] use a clustering-based product quantization to reduce the size of matrices by building an indexing.
Zhang et al. [32] successfully compressed very deep VGG-16
[29] to achieve 4 times speed up with 0.3% loss of accuracy
based on Generalized Singular Value Decomposition and special treatment on non-linear layers. This line of approaches can
be configured as data independent processes, but fine-tuned
with training data to improve the performance significantly.
In contrast to off-line optimization, Ciresan et al. [33] trained
a sparse network with random connections, providing good
performance with better computational efficiency than densely
connected networks.
Rather than pruning or modifying parameters from existing
networks, there has been another line of work in which a
smaller network is trained from scratch to mimic the behavior
of a much larger network. Starting from the work of Bucila
et al. [14] and Knowledge Distillation (KD) by Hinton et al.
[9], the design of smaller yet efficient networks has gained
a lot of research interest. Smaller networks can be shallower
(but much wider) than the original network, performing as
well as deep models, as shown by Ba and Caruna in [34]. The
key idea of knowledge distillation is to utilize the internal
discriminative feature that is implicitly encoded in a way
not only beneficial to original training objectives on source
training dataset, but also has a side-effect of eliminating
incorrect mappings in networks. It has been demonstrated in
[9] that small networks can be trained to generalize in the
same way as large networks with proper guidance. FitNets [15]
achieved better compression rate than knowledge distillation
by designing a deeper but much thinner network using trained
models. The proposed hint-based training is one step further
beyond knowledge distillation which uses a finer network
structure. Nevertheless, training deep networks has proven to
be challenging [35]. Significant efforts have been devoted
to alleviate this problem. Recently, adding supervision to
intermediate layers of deep networks is explored to assist the
training process [36], [37]. These methods assume that source
and target domains are consistent. It is still unclear whether
the guided training is effective when the source and target
domains are significantly different.
In this paper, we consider a unique setting of the problem.
We use a large network pre-trained on a large dataset (e.g.
, the ImageNet) to guide the training of a thinner and faster
network on a new smaller dataset with limited labeled samples,
involving adaptation over different data domains and model
compression at the same time.
III. K NOWLEDGE P ROJECTION N ETWORK
In this section, we present the proposed Knowledge Projection Network (KPN). We start with the KPN architecture
and then explain the knowledge projection layer design. A
multi-path multi-stage training scheme coupled with iterative
pruning for projection route selection is developed afterwards.
A. Overview
An example pipeline of KPN is illustrated in Figure 2.
Starting from a large teacher network pre-trained on a large
dataset, a student network is designed to predict desired
outputs for the target problem with guidance from the teacher
network. The student network uses similar buiding blocks as
the teacher network, such as Residue [38], Inception [39] or
stacks of plain layers [29], sub-sampling and BatchNorm [40]
layers. The similarity in baseline structure ensures smooth
transferability. Note that the convolution layers consume most
of the computational resources. Their complexity can be
modeled by the following equation
C=
N
−1
X
Ci · Hi · Wi · Ci+1 · KiH · KiW ,
(1)
i=1
where the computational cost is multiplicatively related to
the number of input Ci and output channels Ci+1 , the spatial
size of input feature map Hi · Wi where Hi and Wi are the
height and width of the feature map at the i-th layer, and kernel
size KiH · KiW . The student network is designed to be thinner
(in terms of filter channels) but deeper to effectively reduce
network capacity while preserves enough representation power
[34], [15]. We depict the convolutional blocks in Figure 3
that are used to build the thin student networks. In contrast
to standard convolutional layers, a squeeze-then-expand [41],
[42] structure is effective in reducing the channel-wise redundancy by inserting spatially narrow (1 × 1) convolutional
layers between 3 × 3 standard convolutional layers. We denote
this structure as bottleneck Type A and extend it to a more
compact squeeze-expand-squeeze shape, namely bottleneck
Type B. With (1), we can calculate the proportional layerwise computation cost for the standard convolutional layer,
bottleneck Type A and B, respectively. For simplicity, feature
map dimensions are denoted in capital letters, and we use
identical size for kernel height and width, denoted as K,
without loss of generality:
Cstandard = C · H · W · C 0 · K 2 ,
(2)
CT ypeA = C · H · W · X + X · H · W · C 0 · K 2 ,
(3)
CT ypeB = C ·H ·W ·X +X 2 ·H ·W ·K 2 +X ·H ·W ·C 0 . (4)
Combining (2), (3) and (4), we define the reductions in
computation for Type A and B as
X
X
X
CT ypeA
= 0
+
≈ ,
Cstandard
C · K2
C
C
(5)
4
1
1
32
100
256
192
96
Teacher network
64
Image
1
64
Knowledge
Projection Layer
1
1
LKP
3
2
16
Forward path
Backward path
Cross-entropy
loss
Student network
32
128
10
64
64
16
192
128
128-channel
output
96
64
Knowledge
Projection with
64×96×1×1 kernel
1
2
Conditional paths
Gradients
Fig. 2. KPN architecture. Solid arrows showing the forward data-flow, dotted arrows showing the paths for gradients.
C×H×W
C×H×W
C×H×W
3×3 Conv
1x1 Conv
1x1 Conv
C’ × H × W
X×H×W
X×H×W
3x3 Conv
3x3 Conv
C’ × H × W
X×H×W
1x1 Conv
C’ × H × W
Fig. 3. Left: Standard 3x3 Convolutional layer. Middle: Bottleneck type A.
computational cost while preserve the dimension of feature
map and receptive field, and the layer-wise reduction is controlled by X. For example, by cutting the bottleneck channels
by half, i.e. , X = C2 , we have the approximate reduction
rate 21 for Type A, 41 ∼ 18 for Type B. In practice, the
output channel C 0 is equal to or larger than input channel
C: C 0 ∈ [C, 2C]. We replace standard convolutional layers
by bottleneck structures A and B in the teacher network according to computational budget and constitute corresponding
student network. Layer-wise width multipliers α = X
C are
the major contributor to model reduction. We use smaller α
in deep layers where the feature is sparse and computational
expensive layers where the gain is significant. The flexibility
of bottleneck structures and elastic value range of α ensured
we have enough degrees of freedom controlling the student
network capacity. In our KPN, the student network is trained
by optimizing the following joint loss function:
Right: Bottleneck type B. H and W are feature spatial height and width,
C, X, C 0 are input, reduced and output channels for this building block,
Ws∗ = arg min λ · LKP (Ws , Wk ) + Lp (Ws ) + R,
(7)
Ws
respectively. For simplicity, batch-norm and activation layers are omitted in
this figure.
CT ypeB
X
X2
X
X2
= 0
+
+
≈
, (6)
Cstandard
C · K2
C · C0
C · K2
C · C0
Bottleneck structures A and B can effectively reduce the
where LKP and Lp are loss from the knowledge projection
layer and problem specific loss, respectively. For example, for
the problem-specific loss, we can choose the cross-entropy loss
in many object recognition tasks. λ is the weight parameter
decaying during training, Wk is the trained teacher network, R
is a L−2 regularization term, and Ws∗ is the trained parameters
in the student network. Unlike traditional supervised training,
the knowledge projection loss LKP plays an important role in
5
guiding the training direction of KPN, which will be discussed
in more detail in the following section.
network by generating a strong and explicit gradient applied
to backward path to the injection layer in the following form
∂LKP
,
∂Ws,i
∆Ws,i = −λ ·
B. Knowledge Projection Layer Design
In this work, the pre-trained teacher network and the student
network analyze the input image simultaneously. To use the
teacher network to guide the student network, we propose to
map the feature FT of size N learned at one specific layer of
the teacher network into a feature vector FS of size M and
inject it into the student network to guide its training process.
For the mapping, we choose linear projection
FS = P · FT ,
(8)
where P is an N × M matrix. In deep convolutional neural
networks, this linear projection matrix P can be learned by
constructing a convolution layer between the teacher and
student network. Specifically, we use a convolutional layer
to bridge teacher’s knowledge layer and student’s injection
layer. A knowledge layer is defined as the output of a
teacher’s hidden convolutional layer responsible for guiding
the student’s learning process by regularizing the output of
t
student’s injection convolutional layer. Let Oht , Ow
and Oct be
the spatial height, spatial width, and number of channels of the
knowledge layer output in the teacher network, respectively.
s
and Ocs be the corresponding sizes of student’s
Let Ohs , Ow
injection layer output, respectively. Note that there are a
number of additional layers in the student network to further
analyze the feature information acquired in the inject layer and
contribute to the final network output. We define the following
loss function:
LKP (Ws ,Wk ) = h[µ(x; Wk )]·|r[µ(x; Wk ); WKP ] − v[x; Ws ]| ,
(9)
(
1, if x ≥ 0,
h(x) =
(10)
η, otherwise
where µ and v represent the deep nested functions (stacks of
convolutional operations) up to the knowledge and injection
layer with network parameters Wk and Ws , respectively. r[·]
is the knowledge projection function applied on µ[·] with
parameter WKP which is another convolution layer in this
work. µ, v and r must be comparable in terms of spatial
dimensionality.
The knowledge projection layer is designed as a convolutional operation with a 1 × 1 kernel in the spatial domain.
As a result, WKP is a Oct × Ocs × 1 × 1 tensor. As a
comparison, a fully connected adaptation layer will require
t
s
Oht × Ow
× Oct × Ohs × Ow
× Ocs parameters which is not
feasible in practice especially when the spatial size of output
is relatively large in the early layers. Using the convolutional
adaptation layer is not only beneficial for lower computational complexity, but also provides a more natural way to
filter distinctive channel-wise features from the knowledge
layers while preserve spatial consistency. The output of the
knowledge projection layer will guide the training of student
(11)
where Ws,i is the weight matrix of injection layer in student
network. Note that in (9), h[µ(x; Wk )] is applied to LKP with
respect to the hidden output of knowledge projection layer as a
relaxation term. For negative responses from µ(x; Wk ), LKP
is effectively reduced by the slope factor η, which is set to 0.25
by cross-validation. Overall, LKP acts as a relaxed L1 loss.
Compared to L2 loss, LKP is more robust to outliers, but still
has access to finer level representations in r[µ(x; Wk ); WKP ].
Student subnet
Teacher subnet
Projection layer
convolution
convolution
3
convolution
1
activation
activation
convolution
Subsampling /2
4
activation
convolution
2
Subsampling /2
convolution
5
1 2
Knowledge output
3 4 5
Projected guidance
1
3 1
4 2
5
Candidate routes
1
5 2
3 2
4
Discarded routes
Fig. 4. Candidate Routes of Knowledge Projection. Candidate routes are paths
from teacher’s knowledge layer to student’s injection layer. Only one route
will survive after iterative pruning.
C. Multi-Path Multi-Stage Training
In the student network, layers after the injection layer are
responsible for adapting the projected feature to the final network output. This adaptation must be memorized throughout
the training process. Those network layers before the injection
layer aim to learn distinctive low-level features. Therefore,
in our KPN framework, the student network and knowledge
projection layer are randomized and trained in two stages:
initialization stage and end to end joint training stage.
In the initialization stage, Path 2 in Figure 2 is disconnected, i.e. the knowledge projection layer together with
the lower part of student network is trained to adapt the
intermediate output of teacher’s knowledge layer to the final
target by minimizing Lp , which is the loss for target task, e.g.,
softmax or linear regression loss. The upper part of student
6
network is trained sorely by minimizing LKP . In this stage,
we use the projection matrix as an implicit connection between
upper and lower parts in the student network. The upper
student network layers are always optimized towards features
interpreted by the projection matrix, and have no direct access
to targets. This strategy prevents the student network from
over-fitting quickly during the early training stage which is
very hard to correct afterwards.
After the initialization stage, we then disconnect Path 1
and reconnect Path 2 , the training now involves jointly
minimizing the objective function described in (7). Using the
results from stage 1 as the initialization, the joint optimization
process aims to establish smooth transitions inside the student
network from the input to the final output. The loss LKP
injected into the student network continues to regularize the
training process. In this way, the student network is trained
based on a multi-loss function which has been used in the
literature to regulate deep networks [43].
D. Iterative Pruning for Projection Route Selection
One important question in knowledge projection between
the teacher and student networks is to determine which layers
from the teacher network should be chosen as the knowledge
layer and which layers from the students should be chosen
for the injection layer. In this work, we propose to explore
an iterative pruning and optimization scheme to select the
projection route.
Assume that the teacher network Nt and the student network
Ns have Lt and Ls layers, respectively. Candidate projection
routes are depicted in Figure 4. We use only convolution layers
as candidates for the knowledge and injection layers. To satisfy
the constraints on spatial size and receptive field, candidate
knowledge projection routes are computed and denoted as
Ri,j ∈ G, where i is the index of knowledge layer in the
teacher network, j is the index of injection layer in the student
network, and G is the set of all candidate routes. We follow
the procedure for computing the center of receptive field in
[44] for calculating the size of receptive field in layer L:
SL =
L p−1
X
Y
(
Sq )(Fp − 1),
(12)
p=1 q=1
where Sq and Fp are the layer-wise stride and kernel size,
assuming they are identical along x and y directions for
simplicity. Routes with constrained receptive filed are kept
after calculation with a small tolerance β = 0.2:
(1 − β) · Si ≤ Sj ≤ (1 + β) · Si .
(13)
For example, in Figure 4, we have
{R1,3 , R1,4 , R2,5 } ⊂ G
(14)
and the rest routes in this figure are not valid due to mismatched spatial shapes. The idea of iterative pruning for the
projection route selection is to traverse all possible routes
with same training hyper-parameters, and determine the best
route for knowledge-injection pair on-the-fly. Specifically, we
randomly initialize |G| KPNs according to each Ri,j .
Each KPN stores a student network Ws , knowledge projection parameter WKP and routing Ri,j , teacher network Wt
is shared across all KPNs to save computation and memory
overhead. The target is to find the KPN setting with minimum
joint loss
0
0
{Ws0 , WKP
, Ri,j
}=
arg min
(λ · LKP + Lp ).
(15)
{Ws ,WKP ,Ri,j }
We assume that the pre-trained teacher network Wt is responsible for guiding the training of a specifically designed student
network Ws which satisfies the computational complexity
requirement. According to (13), we can generate a list L
of candidate KPNs. Each KPN is a copy of the designed
student network Ws with different projection routing Ri,j
and corresponding parameters WKP . Within a period of k
epochs, the KPNs are optimized separately using Stochastic
Gradient Descend to minimize the joint loss described in (15).
Note that even though the optimization target is a joint loss,
as depicted in Fig. 2, the upper and bottom layers of the
student network are receiving different learning targets from
the teacher network and dataset distribution, respectively. At
the end of k epochs, the joint loss of each KPN computed
on the validation dataset is used to determine which KPN to
prune. The same procedure is applied on the remaining KPNs
in the list L iteratively. This iterative pruning procedure is
summarized in Algorithm 1:
Only one KPN will survive after the iterative pruning
process. We continue the multi-stage training with or without
adjusting the batch-size depending on the released memory
size after sweeping out bad KPNs. The stopping criteria can
either be plateau of validation accuracy or a pre-defined end
epoch.
IV. E XPERIMENTAL R ESULTS
In this section, we provide comprehensive evaluations of our
proposed method using three groups of benchmark datasets.
Each group consists of two datasets, the large dataset Dt
used to train the teacher network and the smaller dataset Ds
used to train the student network. The motivation is that, in
practical applications, we often need to learn a network to
recognize or classify a relatively small number of different
objects and the available training dataset is often small. We
also wish the trained network to be fast and efficient. The
large dataset is often available from existing research efforts,
for example, the ImageNet. Both the large and the small
datasets have the same image dimensions so that pre-trained
models are compatible with each other in terms of shape.
We use the existing teacher network model already trained
by other researchers on the public dataset Dt . We compare
various algorithms on the benchmark dataset Ds where stateof-the-art results have been reported. Performance reports on
small datasets are rare, thus we choose existing large famous
benchmark datasets in following experiments, and aggressively
reduce the size of training set to simulate the shortage of
labeled data in real world scenarios.
7
Algorithm 1: Iterative pruning algorithm for projection
route selection.
Input : List L of KPNs, as in form
{Ws,n , WKP,n , Rin ,jn }, where n = 1, ..., |G|,
and teacher network Wt
∗
∗
Output: Ws∗ , WKP
and Ri,j
1
Configure all KPNs as initialization stage.
2
while |L| > 1 do
3
for k epochs do
to be 0.6, and gradually decays to 0. The pruning frequency
is 10000 and we also randomly revoke the initialization stage
during joint training stage, to repetitively adjusting network
guidance strength.
For fine-tuning, we test with a wide variety of experimental
settings. Starting from pre-trained networks, we adjust the last
layer to fit to the new dataset, and randomly initialize the last
layer. The reshaped network is trained with standard backpropagation with respect to labels on the new dataset, and
unfreeze one more layer from the bottom one at a time. The
best result from all configurations was recorded. To make sure
all networks are trained using the optimal hyper-parameter set,
we extensively try a wide range of learning rates, and repeat
experiments on the best parameter set for at least 5 times.
The average performance of the best 3 runs out of 5 will be
reported. Data augmentation is limited to random horizontal
flip if not otherwise specified.
for Batch x in Data do
4
5
Forward teacher: yt ← µ(x; Wk );
6
for {Ws , WKP , Ri,j } ∈ L do
Forward-backward w.r.t. Ws , WKP ;
7
end
8
end
9
10
end
11
0
0
, Ri,j
} ← arg min(λ · LKP + Lp );
{Ws0 , WKP
12
0
0
, Ri,j
} in L;
Remove {Ws0 , WKP
13
end
14
∗
∗
return {Ws∗ , WKP
, Ri,j
} in L;
A. Network Training
We have implemented our KPN framework using the
MXNet [45], a deep learning framework designed for both
efficiency and flexibility. The dynamically generated computational graph in MXNet allows us to modify network
structures during run time. The KPNs are trained on NVidia
Titan X 12GB with CUDNN v5.1 enabled. Batch-sizes vary
from 16 to 128 depending on the KPN group size. For all
experiments, we train using the Stochastic Gradient Descend
(SGD) with momentum 0.9 and weight decay 0.0001 except
the knowledge projection layers. The weight decay for all
knowledge projection layers is 0.001 in the initialization stage
and 0 for the joint training stage. 40% of iterations are used
for the initialization stage, and the rest goes to be joint training
stage. The weight controller parameter λ for joint loss is set
B. Results on the CIFAR-10 Dataset
We first evaluate the performance of our method on the
CIFAR-10 dataset guided by a teacher network pre-trained on
CIFAR-100 dataset. The CIFAR-10 and CIFAR-100 datasets
[50] have 60000 32×32 color images with 10 and 100 classes,
respectively. They were both split into 50K-10K sets for
training and testing. To validate our approach, we trained a
38-layer Resnet on the CIFAR-100 as reported in [38], and
use it to guide a 50-layer but significantly slimmer Resnet on
the CIFAR-10. We augment the data using random horizontal
flip and color jittering. Table I summarizes the results, with
comparisons against the state-of-the-art results which cover a
variety of optimization techniques including Layer-sequential
unit-variance initialization [49], pooling-less [48], generalized
pooling [47] and maxout activation [46]. We choose different
sizes ST of the training set and list the accuracy. For network
complexity, we compute its number of model parameters
NP ara and the number of multiplication and additions NM A
needed for the network inference. It should be noted that for
methods in the literature we do not have their accuracy results
on down-sized training sets.
We do not apply specific optimization techniques used
in the state-of-the-art methods due to some structures not
reproducible in certain conditions. To compare, we trained
a standard 38-layer Residue Network, a 50-layer slimmer
version of ResNet (each convolutional layer is half the capacity of the vanilla ResNet) and a fine-tuned model of 38layer ResNet (from CIFAR-100) on CIFAR-10 with different
amount of training samples. With all 50000 training data, our
proposed method outperforms direct training and best finetuning results and still match the state-of-the-art performance.
We believe the performance gain specified in [47], [49] can
be also applied to our method, i.e. , ensemble of multiple
techniques could achieve better performance. The proposed
KPN method has improved the accuracy by up to 1.2% while
significantly reducing the network size by about 11 times,
from 3.1M network parameters to 273K parameters. It also
demonstrated strong robustness against aggressive reduction
of labeled training samples.
8
TABLE I
CIFAR-10 ACCURACY AND NETWORK CAPACITY COMPARISONS WITH STATE - OF - THE - ART METHODS . R ESULTS USING RANDOMLY SAMPLED SUBSETS
FROM TRAINING DATA ARE ALSO REPORTED .
N UMBER OF NETWORK PARAMETERS ARE CALCULATED BASED ON REPORTS IN RELATED WORK .
Accuracy with Different ST
NP ara
NM A
-
9M
379M
-
-
0.86M
53M
-
-
-
2.5M
107M
93.95
-
-
-
3.5M
362M
ALL-CNN-C [48]
92.7
-
-
-
1.0M
257M
Good Init [49]
94.16
-
-
-
2.5M
166M
ResNet-50 slim
87.53
71.92
55.86
48.17
0.27M
31M
ResNet-38
90.86
75.28
61.74
51.62
3.1M
113M
ResNet-38 fine-tune
91.15
89.61
86.26
83.45
3.1M
113M
Our method
92.37
90.35
88.73
87.61
0.27M
31M
Methods
50000
5000
1000
500
Maxout [46]
90.18
-
-
FitNets-11 [15]
91.06
-
FitNets [15]
91.61
GP CNN [47]
(1)
(2)
(3)
(4)
(5)
(6)
Fig. 5. (1)(2): CIFAR-100/10 sample images; (3): Imagenet 2012; (4) Pascal
VOC 2007; (5) MNIST; (6) Omniglot;
C. Results on the Pascal VOC 07 Dataset
We evaluate the proposed method on the PASCAL Visual
Object Classes Challenge(VOC) dataset [54] with a VGG16 model [29] pre-trained on the ILSVRC 2012 dataset
[1]. The pre-training usually takes several weeks, thus we
downloaded and converted the teacher network from the Caffe
model available online. We compare our method with stateof-the-art results obtained on this dataset in the literature,
including the VGG16+SVM method [29], the segment hy-
potheses based multi-label HCP-VGG method [52], and the
FisherNet-VGG16 method [53] which encodes CNN feature
with fisher vector. These papers have reported results on the
original whole dataset with 5011 images. To test the learning
capability of the network on smaller datasets with reduced
samples, we also implement the fine-tuning method. We try
different combination of network update scheme and learning
parameters and use the best result for performance comparison
with our method. We conducted our experiments on the entire
training set with 5011 images and test set with 4952 images.
In addition, we randomly sample 50 and 10 images from each
class, generating two small datasets with 1000 and 200 training
images, respectively. The results are summarized in Table II.
We list the test accuracy of the network for each configuration.
We compute the corresponding complexity of the network,
including the number of model parameters NP ara and the
number of multiplication and additions NM A . It should be
noted that for methods in the literature we do not have their
accuracy results on down-sized training sets. It can be seen
that our proposed method outperforms standard training and
fine-tuning by a large margin while reducing the model size
by 2 times and improving the inference speed by 4.6 times.
D. Results on the Ommniglot Dataset
We are interested in how the proposed KPN method works
on very small datasets, for example, the Ommniglot handwritten recognition dataset. The MNIST [55] is a famous
handwritten digits dataset, consists of 60000 training images
and 10000 test images, 28x28x1 in size, organized into 10
9
TABLE II
PASCAL VOC 2007 TEST OBJECT CLASSIFICATION PERFORMANCES COMPARISON . R ESULTS USING RANDOMLY SAMPLED SUBSETS FROM TRAINING
DATA ARE ALSO REPORTED .
N UMBER OF CONVOLUTION LAYER PARAMETERS ARE LISTED FOR FAIR COMPARISON BASED ON REPORTS IN RELATED
WORK .
Accuracy at Different ST
NP ara
NM A
-
6.5M
2483M
-
-
14.7M
15470M
89.3
-
-
21.8M
15470M
HCP-VGG [52]
90.9
-
-
14.7M
15470M
FisherNet-VGG16 [53]
91.7
-
-
14.7M
15470M
VGG16 standard BP
83.5
65.2
<30
14.7M
15470M
Fine-tune VGG16 last layer (softmax)
89.6
87.4
85.7
14.7M
15470M
Fine-tune VGG16 2+ learnable layers
90.2
86.3
82.8
14.7M
15470M
Our method
91.2
88.4
86.5
8M
3361M
Methods
5011
1000
200
Chatfield et al. [51]
82.4
-
VGG16+SVM [29]
89.3
VGG19+SVM [29]
classes. The Omniglot [56] is a similar but much smaller
dataset, containing 1623 different handwritten characters from
50 alphabets. Each of the 1623 characters was drawn online
via Amazon’s Mechanical Turk by 20 different people. All
images are binarized and resized to 28×28×1 with no further
data augmentation. We use all 70000 images from MNIST for
training a 5-layer Maxout convolutional model as the teacher
network Nt as proposed in [46]. We report experimental
results of various algorithms across a wide range of number
of training examples, from 19280 to merely 1000, shown
in Table III. Note that we use class dependent shuffling to
randomly select training subsets, which is critical to avoid
unbalanced class distribution in Omniglot due to the limited
number of samples for each class. We can see that the proposed
KPN is able to reduce the error rate by 1.1-1.3%. Table III
also provides some interesting insights of how models are
transferred to different tasks. First, the fine-tuning methods
are all affected by the number of learnable parameters and
training samples. Smaller training set will result in significant
over-fitting, thus breaking the fragile co-adaptation between
layers. If the training set is large enough, the number of
learnable parameters are positively related to the performance.
This phenomenon is also discussed in [25], where transferring
knowledge from the pre-trained model to an exactly same
network is extensively tested.
E. Algorithm Parameter Analysis
In this section, we study how the performance of the our
method is impacted by the selection of major parameters.
TABLE III
T EST ERROR RATE COMPARISONS BETWEEN EXPERIMENTAL SETTINGS
AND BASELINE METHODS .
Error Rates at Different ST
Methods
19280
5000
1000
Deep CNN [56]
13.5%
-
-
Deep Siamese CNN [56]
8.0 %
-
-
Large CNN standard BP
9.3%
12.9%
19.4%
Small CNN standard BP
12.1%
18.5%
23.8%
Fine-tuned from MNIST
6.8%
7.4%
9.2%
Our method
5.9%
6.6%
7.9%
(1) Trade-off between Performance and Efficiency. To
evaluate how the size of network affects the performance, we
measure the test accuracy, number of parameters, and network
speed up ratio of various student networks on the CIFAR10 dataset. Figure 6 shows the results. Student networks are
designed based on a multi-layer Resnet denoted as N - or N -, where N is the number of layers, - and - - indicate it’s
a slim or slimmer version of Resnet. The detailed network
configurations are listed in Table IV. As expected, deeper and
slimmer networks are more difficult to train with limited train-
10
TABLE IV
N ETWORK CONFIGURATIONS FOR EXTENSIVE BENCHMARKS ON O MNIGLOT DATASET. N - DENOTES SLIM NETWORK WITH N LAYERS , SIMILARLY, N
LAYER SLIMMER NETWORK IS DENOTED AS
N - -. N OTE THAT 1 × 1 ADAPTIVE CONVOLUTIONS FOR RESIDUE MODULES ARE NOT INCLUDED IN THIS
TABLE .
# Layers
50
50-
50- -
44-
44- -
38-
38- -
32-
32- -
26-
26- -
Conv3 × 3 /s1
16
16
16
16
16
16
16
16
16
16
16
ResConv3 × 3 /s2
32 ×16
32 ×16
16 ×16
32 ×14
16 ×14
32 ×12
16 ×12
32 ×10
16 ×10
32 ×8
16 ×8
ResConv3 × 3 /s1
64 ×16
32 ×16
32 ×16
32 ×14
32 ×14
32 ×12
32 ×12
32 ×10
32 ×10
32 ×8
32 ×8
ResConv3 × 3 /s2
128 ×16
64 ×16
48 ×16
64 ×14
48 ×14
64 ×12
48 ×12
64 ×10
48 ×10
64 ×8
48 ×8
Conv3 × 3 /s1
256
128
96
128
96
128
96
128
96
128
96
S U R S R V H G
V W D Q G D U G E S
R F F X U D Q F H V R Y H U W H V W V
$ F F X U D F \
103
0 H D Q H U U R U
3 D U D P V N
102
6 S H H G X S U D W L R
Fig. 7. Iterative pruning analysis. Top: occurrences of projection route t-s
W
1 H W Z R U N F R Q I L J X U D W L R Q V
over 32 standalone tests. Bottom: mean classification error of projection route
t-s by disable iterative pruning. t-s: network with knowledge layer t from
teacher to injection layer s from student.
Fig. 6. Network capacity and performance analysis. Top: test accuracies with
proposed KPN and normal training with standard back-propagation; Middle:
number of parameters (×103 ), note that the y-axis is in logarithmic scale;
Bottom: actual inference speed up ratio with respect to Resnet-50. Network
notations: t is teacher network, N - denotes slim network with N layers,
similarly, N layer slimmer network is denoted as N - -.
ing data. However, with proposed method enabled, the depth
is beneficial, and networks are less suffered from performance
drop. Impressively, we could obtain a model which is 34 times
faster using less than 2% parameters, with about 3% accuracy
loss, compared to the teacher network.
(2) Analysis of Iterative Pruning for Automatic Route
Selection. The knowledge projection route is critical for the
network training and test performance. Intuitively, the projection route should not be too shallow or too deep. Shallow
layers may contain only low-level texture features, while deep
layers close to output may be too task specific. To study how
the iterative pruning works during training, we record the
pruning results and compare them with respect to manually
defined projection routes, shown in Figure 7. We can see that
the statistics of survival projection routes is highly correlated
to the training accuracy, which is evaluated by manually
defining projection route from t to s and disabling iterative
pruning during training. The result also indicates that choosing
the middle layers for projection is potentially better. Reducing
11
the size of training data also affects the pruning results. This
might relate to the difficulty of fitting knowledge projection
layer to the target domain when very limited data is presented.
As a result, projection layers tend to appear more on very deep
layers close to the output, so that the penalty from adaptation
loss will not dominate. The bottom line is, even though the
iterative pruning method is a random optimization process, it
is reliably producing satisfactory results.
[2] T. Lin, M. Maire, S. J. Belongie, L. D. Bourdev, R. B. Girshick,
J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft
COCO: common objects in context,” CoRR, vol. abs/1405.0312, 2014.
[Online]. Available: http://arxiv.org/abs/1405.0312 1
[3] M. Denil, B. Shakibi, L. Dinh, N. de Freitas et al., “Predicting parameters in deep learning,” in Advances in Neural Information Processing
Systems, 2013, pp. 2148–2156. 1
F. Discussion and Future Work
Our KPN is designed in a highly modular manner. The
training of projection layers is removed during actual network
testing, and the network capacity is highly configurable for
performance/speed trade-off. This KPN method can be easily
extended to other problems such as object detection, object
segmentation, and pose estimation by replacing softmax loss
layer used in the classification problems. Since the deployed
network is a pure standard network, another research direction
is to apply KPN as a building block in traditional model
compression techniques to reshape the network in a new
perspective. Although we have focused on the advantage of
KPN with thinner networks on smaller datasets, there are
potential benefits to apply KPN on large network and relatively
large datasets, for example, performance oriented situations
where speed is not an issue.
[4] G. E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and R. R.
Salakhutdinov, “Improving neural networks by preventing co-adaptation
of feature detectors,” arXiv preprint arXiv:1207.0580, 2012. 1
[5] Y. LeCun, J. S. Denker, and S. A. Solla, “Optimal brain damage,” in
Advances in Neural Information Processing Systems 2, D. S. Touretzky,
Ed.
Morgan-Kaufmann, 1990, pp. 598–605. [Online]. Available:
http://papers.nips.cc/paper/250-optimal-brain-damage.pdf 1, 2
[6] B. Hassibi, D. G. Stork et al., “Second order derivatives for network
pruning: Optimal brain surgeon,” Advances in neural information processing systems, pp. 164–164, 1993. 1, 2
[7] S. Han, H. Mao, and W. J. Dally, “Deep compression: Compressing
deep neural networks with pruning, trained quantization and huffman
V. C ONCLUSION
We have developed a novel knowledge projection framework for deep neural networks the address the issues of domain adaptation and model compression in training simultaneously. We exploit the distinctive general features produced by
the teacher network trained on large dataset, and use a learned
matrix to project them into domain relevant representations
to be used by the student network. A smaller and faster
student network is trained to minimize joint loss designed for
domain adaptation and knowledge distillation simultaneously.
Extensive experimental results have demonstrated that our
unified training framework provides an effective way to obtain
fast high-performance neural networks on small datasets with
limited labeled samples.
coding,” arXiv preprint arXiv:1510.00149, 2015. 1, 2
[8] M. Jaderberg, A. Vedaldi, and A. Zisserman, “Speeding up convolutional neural networks with low rank expansions,” arXiv preprint
arXiv:1405.3866, 2014. 1, 3
[9] G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural
network,” arXiv preprint arXiv:1503.02531, 2015. 1, 3
[10] S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang, “Domain adaptation via
transfer component analysis,” IEEE Transactions on Neural Networks,
vol. 22, no. 2, pp. 199–210, 2011. 1
[11] K. Zhang, B. Schölkopf, K. Muandet, and Z. Wang, “Domain adaptation
under target and conditional shift.” in ICML (3), 2013, pp. 819–827. 1
R EFERENCES
[12] X. Wang and J. Schneider, “Flexible transfer learning under support and
model shift,” in Advances in Neural Information Processing Systems,
[1] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
2014, pp. 1898–1906. 1, 2
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and
[13] E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko, “Simultaneous
L. Fei-Fei, “ImageNet Large Scale Visual Recognition Challenge,”
deep transfer across domains and tasks,” in Proceedings of the IEEE
International Journal of Computer Vision (IJCV), vol. 115, no. 3, pp.
International Conference on Computer Vision, 2015, pp. 4068–4076. 1,
211–252, 2015. 1, 8
2
12
[14] C. Bucilu, R. Caruana, and A. Niculescu-Mizil, “Model compression,”
processing systems, 2014, pp. 3320–3328. 2, 9
in Proceedings of the 12th ACM SIGKDD international conference on
[26] H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, and M. Marchand,
ACM, 2006, pp. 535–541. 1,
“Domain-adversarial neural networks,” arXiv preprint arXiv:1412.4446,
Knowledge discovery and data mining.
3
2014. 2
[15] A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y. Ben-
[27] M. Ghifary, W. B. Kleijn, and M. Zhang, “Domain adaptive neural net-
gio, “Fitnets: Hints for thin deep nets,” arXiv preprint arXiv:1412.6550,
works for object recognition,” in Pacific Rim International Conference
2014. 1, 3, 8
on Artificial Intelligence.
Springer, 2014, pp. 898–904. 2
[16] Y. Hou, Z. Li, P. Wang, and W. Li, “Skeleton optical spectra based action
[28] S. Han, J. Pool, J. Tran, and W. Dally, “Learning both weights and con-
recognition using convolutional neural networks,” IEEE Transactions on
nections for efficient neural network,” in Advances in Neural Information
Circuits and Systems for Video Technology, 2016. 1
Processing Systems, 2015, pp. 1135–1143. 2
[17] C. Xiong, L. Liu, X. Zhao, S. Yan, and T.-K. Kim, “Convolutional fusion
[29] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
network for face verification in the wild,” IEEE Transactions on Circuits
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
and Systems for Video Technology, vol. 26, no. 3, pp. 517–528, 2016. 1
3, 8, 9
[18] K. Kim, S. Lee, J.-Y. Kim, M. Kim, and H.-J. Yoo, “A configurable
[30] E. L. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R. Fergus,
heterogeneous multicore architecture with cellular neural network for
“Exploiting linear structure within convolutional networks for efficient
real-time object recognition,” IEEE Transactions on Circuits and Sys-
evaluation,” in Advances in Neural Information Processing Systems,
tems for Video Technology, vol. 19, no. 11, pp. 1612–1622, 2009. 2
2014, pp. 1269–1277. 3
[19] N. Sudha, A. Mohan, and P. K. Meher, “A self-configurable systolic
[31] Y. Gong, L. Liu, M. Yang, and L. Bourdev, “Compressing deep
architecture for face recognition system based on principal component
convolutional networks using vector quantization,” arXiv preprint
neural network,” IEEE transactions on circuits and systems for video
arXiv:1412.6115, 2014. 3
technology, vol. 21, no. 8, pp. 1071–1084, 2011. 2
[32] X. Zhang, J. Zou, K. He, and J. Sun, “Accelerating very deep convolu-
[20] S. J. Pan and Q. Yang, “A survey on transfer learning,” IEEE Trans-
tional networks for classification and detection,” IEEE transactions on
actions on knowledge and data engineering, vol. 22, no. 10, pp. 1345–
pattern analysis and machine intelligence, vol. 38, no. 10, pp. 1943–
1359, 2010. 2
1955, 2016. 3
[21] M. Long, Y. Cao, J. Wang, and M. I. Jordan, “Learning transferable
[33] D. C. Cireşan, U. Meier, J. Masci, L. M. Gambardella, and J. Schmidhu-
features with deep adaptation networks.” in ICML, 2015, pp. 97–105. 2
ber, “High-performance neural networks for visual object classification,”
[22] Y. Ganin and V. Lempitsky, “Unsupervised domain adaptation by
backpropagation,” arXiv preprint arXiv:1409.7495, 2014. 2
[23] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in European conference on computer vision. Springer,
2014, pp. 818–833. 2
[24] M. Oquab, L. Bottou, I. Laptev, and J. Sivic, “Learning and transferring
mid-level image representations using convolutional neural networks,”
in Proceedings of the IEEE conference on computer vision and pattern
recognition, 2014, pp. 1717–1724. 2
arXiv preprint arXiv:1102.0183, 2011. 3
[34] J. Ba and R. Caruana, “Do deep nets really need to be deep?” in
Advances in neural information processing systems, 2014, pp. 2654–
2662. 3
[35] D. Erhan, P.-A. Manzagol, Y. Bengio, S. Bengio, and P. Vincent, “The
difficulty of training deep architectures and the effect of unsupervised
pre-training.” in AISTATS, vol. 5, 2009, pp. 153–160. 3
[36] C.-Y. Lee, S. Xie, P. W. Gallagher, Z. Zhang, and Z. Tu, “Deeplysupervised nets.” in AISTATS, vol. 2, no. 3, 2015, p. 5. 3
[25] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson, “How transferable are
[37] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
features in deep neural networks?” in Advances in neural information
V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,”
13
in Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2015, pp. 1–9. 3
[49] D. Mishkin and J. Matas, “All you need is a good init,” arXiv preprint
arXiv:1511.06422, 2015. 7, 8
[38] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
[50] A. Krizhevsky and G. Hinton, “Learning multiple layers of features
recognition,” in Proceedings of the IEEE Conference on Computer Vision
from tiny images,” Master’s thesis, Department of Computer Science,
and Pattern Recognition, 2016, pp. 770–778. 3, 7
University of Toronto, 2009. 7
[39] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking
[51] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, “Return of
the inception architecture for computer vision,” in Proceedings of the
the devil in the details: Delving deep into convolutional nets,” arXiv
IEEE Conference on Computer Vision and Pattern Recognition, 2016,
preprint arXiv:1405.3531, 2014. 9
pp. 2818–2826. 3
[52] Y. Wei, W. Xia, M. Lin, J. Huang, B. Ni, J. Dong, Y. Zhao, and S. Yan,
[40] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
“Hcp: A flexible cnn framework for multi-label image classification,”
network training by reducing internal covariate shift,” arXiv preprint
IEEE transactions on pattern analysis and machine intelligence, vol. 38,
arXiv:1502.03167, 2015. 3
no. 9, pp. 1901–1907, 2016. 8, 9
[41] F. N. Iandola, S. Han, M. W. Moskewicz, K. Ashraf, W. J. Dally,
and K. Keutzer, “Squeezenet: Alexnet-level accuracy with 50x fewer
parameters and¡ 0.5 mb model size,” arXiv preprint arXiv:1602.07360,
2016. 3
[42] J. Redmon and A. Farhadi, “Yolo9000: Better, faster, stronger,” arXiv
preprint arXiv:1612.08242, 2016. 3
[43] C. Xu, C. Lu, X. Liang, J. Gao, W. Zheng, T. Wang, and S. Yan, “Multiloss regularized deep neural network,” IEEE Transactions on Circuits
and Systems for Video Technology, vol. 26, no. 12, pp. 2273–2283, 2016.
6
[44] K. Lenc and A. Vedaldi, “R-cnn minus r,” arXiv preprint
arXiv:1506.06981, 2015. 6
[45] T. Chen, M. Li, Y. Li, M. Lin, N. Wang, M. Wang, T. Xiao, B. Xu,
C. Zhang, and Z. Zhang, “Mxnet: A flexible and efficient machine
learning library for heterogeneous distributed systems,” arXiv preprint
arXiv:1512.01274, 2015. 7
[46] I. J. Goodfellow, D. Warde-Farley, M. Mirza, A. C. Courville, and
Y. Bengio, “Maxout networks.” ICML (3), vol. 28, pp. 1319–1327, 2013.
7, 8, 9
[47] C.-Y. Lee, P. W. Gallagher, and Z. Tu, “Generalizing pooling functions in
convolutional neural networks: Mixed, gated, and tree,” in International
conference on artificial intelligence and statistics, 2016. 7, 8
[48] J. T. Springenberg, A. Dosovitskiy, T. Brox, and M. Riedmiller,
“Striving for simplicity: The all convolutional net,” arXiv preprint
arXiv:1412.6806, 2014. 7, 8
[53] P. Tang, X. Wang, B. Shi, X. Bai, W. Liu, and Z. Tu, “Deep fishernet
for object classification,” arXiv preprint arXiv:1608.00182, 2016. 8, 9
[54] M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, “The pascal visual object classes (voc) challenge,” International
journal of computer vision, vol. 88, no. 2, pp. 303–338, 2010. 8
[55] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning
applied to document recognition,” Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, 1998. 8
[56] B. M. Lake, R. Salakhutdinov, and J. B. Tenenbaum, “Human-level
concept learning through probabilistic program induction,” Science, vol.
350, no. 6266, pp. 1332–1338, 2015. 9
| 7cs.IT
|
INTRODUCTION TO THE SP
THEORY OF INTELLIGENCE
arXiv:1802.09924v1 [cs.AI] 24 Feb 2018
Aiming to simplify and integrate observations and
concepts across artificial intelligence, mainstream
computing, mathematics, and human learning,
perception, and cognition.
J Gerard Wolff∗
February 28, 2018
Abstract
This article provides a brief introduction to the SP Theory of Intelligence and its realisation in the SP Computer Model. The overall goal
of the SP programme of research, in accordance with long-established
principles in science, has been the simplification and integration of
observations and concepts across artificial intelligence, mainstream
computing, mathematics, and human learning, perception, and cognition. In broad terms, the SP system is a brain-like system that
takes in “New” information through its senses and stores some or all
of it as “Old” information. A central idea in the system is the powerful concept of SP-multiple-alignment, borrowed and adapted from
bioinformatics. This the key to the system’s versatility in aspects of
intelligence, in the representation of diverse kinds of knowledge, and in
the seamless integration of diverse aspects of intelligence and diverse
kinds of knowledge, in any combination. There are many potential
benefits and applications of the SP system. It is envisaged that the
∗
Dr Gerry Wolff, BA (Cantab), PhD (Wales), CEng, MBCS, MIEEE; CognitionResearch.org, Menai Bridge, UK; [email protected]; +44 (0) 1248 712962; +44 (0)
7746 290775; Skype: gerry.wolff; Web: www.cognitionresearch.org.
1
system will be developed as the SP Machine, which will initially be a
software virtual machine, hosted on a high-performance computer, a
vehicle for further research and a step towards the development of an
industrial-strength SP Machine.
1
Introduction
The SP theory of intelligence and its realisation in the SP computer model is
a system that has been under development since about 1987, with a break
between early 2006 and late 2012.
Potential benefits and applications of the SP system include versatility
in aspects of intelligence, versatility in the representation of diverse forms
of knowledge, seamless integration of aspects of intelligence and kinds of
knowledge in any combination, helping to solve nine problems with big data,
providing a model for aspects of neuroscience, solving several problems with
deep learning, and many more.
A key idea in the SP system is the powerful concept of SP-multiplealignment, borrowed and adapted from the concept of ‘multiple sequence
alignment’ in bioinformatics. This may prove to be as significant for an
understanding of human intelligence as is DNA in biological sciences: SPmultiple-alignment may prove to be the “double helix” of intelligence.
Although there are still residual problems to be solved [13, Section 3.3],
it is envisaged that the SP computer model will be the basis for an SP machine as a software virtual machine, hosted on a high-performance computer.
This would be a vehicle for further research and a basis for an industrialstrength SP machine with many potential benefits and application, as shown
schematically in Figure 1. A programme of development for the SP machine
is described in [6].
2
SP theory and SP computer model
High parallel
In the cloud
SP MACHINE
Open source
Good user interface
Natural language processing
Representation of knowledge
Several kinds of reasoning
Planning & problem solving
Unsupervised learning
Information compression
Pattern recognition
Information retrieval
MANY APPLICATIONS
Figure 1: Schematic representation of the development and application of
the SP machine. Reproduced from Figure 2 in [13], with permission.
2
Background
The overall goal of the SP programme of research, in accordance with longestablished principles in science, has been the simplification and integration
of observations and concepts across artificial intelligence, mainstream computing, mathematics, and human learning, perception, and cognition.
From the beginning, a unifying theme in the SP research has been that
all kinds of processing would be done by compression of information via a
search for patterns that match each other and via the merging or ‘unification’
of patterns that are the same. The reason for the emphasis on the matching
and unification of patterns is that this seems to provide a better handle
on possible mechanisms in natural or artificial systems than do the more
mathematically-oriented approaches to information compression.
The main motivation for the focus on information compression is research
by Fred Attneave [1], Horace Barlow [2, 3], and others, showing the importance of information compression in the workings of brains and nervous
systems. Solomonoff’s seminal work on the development of algorithmic probability theory [8, 9] is also important.
Since people often ask, the name “SP” stands for Simplicity and Power,
two ideas which, together, mean the same as information compression. This
is because information compression may be seen to be a process of maximising ‘simplicity’ in a body of information, by reducing redundancy in that
3
information, whilst at the same time retaining as much as possible of its
non-redundant expressive ‘power’.
3
The SP system
The SP system is described in outline here and in Appendix I of [20], in
more detail in [13], and in even more detail in [11]. Distinctive features and
advantages of the SP system are described in [20].
In broad terms, the SP system is a brain-like system that takes in New
information through its senses and stores some or all of it as Old information,
as shown schematically in Figure 2.
Old
(compressed)
New
(uncompressed)
Figure 2: Schematic representation of the SP system from an ‘input’ perspective. Reproduced, with permission, from Figure 1 in [13].
3.1
SP-multiple-alignment
A central idea in the SP system, is the concept of SP-multiple-alignment,
borrowed and adapted from the concept of ‘multiple sequence alignment’
in bioinformatics. Probably the best way to explain the idea is by way of
examples, shown in Figures 3 and 4.
Figure 3 shows an example of multiple sequence alignment in bioinformatics. Here, there are five DNA sequences which have been arranged one above
the other, and then, by judicious ‘stretching’ of one or more of the sequences
in a computer, symbols that match each other across two or more sequences
have been brought into line. A ‘good’ multiple sequence alignment, like the
one shown, is one with a relatively large number of matching symbols from
4
row to row. The process of discovering a good multiple sequence alignment
is normally too complex to be done by exhaustive search, so heuristic methods are needed, building multiple sequence alignments in stages and, at each
stage, selecting the best partial structures for further processing.
G
|
G
|
A |
|
G
|
G
G
|
G
|
G
|
G
|
G
A
|
| G
|
A C T
|
A A
|
C A
G
C
|
|
G C C C
| | | |
G C C C
|
|
|
C
A
|
A
|
A
|
A
|
A
G
|
G
|
G
|
G
|
G
G
|
G
|
G
|
G
|
G
G
|
G
|
G
|
G
|
G
A
|
A
|
|
|
A
|
A
G
|
G
|
G
|
G
|
G
G
|
G
|
G
|
G
|
G
A
T
|
|
A
|
|
|
| G C T
|
|
A
| A
|
C
G
G
G G A
|
|
| | |
G G C G
G G A
|
|
| | |
G G A | G A
|
|
| | |
G
G
G G A
|
|
| | |
G
G
G G A
Figure 3: A ‘good’ multiple alignment amongst five DNA sequences. Reproduced with permission from Figure 3.1 in [11].
Figure 4 shows an example of an SP-multiple-alignment, superficially similar to the one in Figure 3, except that sequences are called SP-patterns and,
more importantly, one of the SP-patterns is New information and is normally
shown in row 0, while the remaining SP-patterns are Old information, and
these are shown in the remaining rows.
0
1
2
3
4
5
6
f
|
|
|
|
|
|
|
N 4 f
|
NP 2 N
|
S 0 NP
o
|
|
|
|
|
|
|
o
r
|
|
|
|
|
|
|
r
t
|
|
|
|
|
|
|
t
u
|
|
|
|
|
|
|
u
n
|
|
|
|
|
|
|
n
e
|
|
|
|
|
|
|
e #N
|
#N #NP
|
#NP
f a v o u r
| | | | | |
Vr 6 f a v o u r #Vr
|
|
V 7 Vr
#Vr
|
VP 3 V
|
|
|
|
|
VP
7
8
9
s
|
|
|
s #V
|
#V NP
|
|
|
|
|
|
|
|
|
NP 1 D
|
D 8
t
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
t
h
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
h
e
|
|
|
|
|
|
|
|
|
|
|
|
|
|
N 5
|
|
| #D N
| |
e #D
b
|
|
|
|
|
|
|
|
|
|
|
|
|
b
r
|
|
|
|
|
|
|
|
|
|
|
|
|
r
a
|
|
|
|
|
|
|
|
|
|
|
|
|
a
v
|
|
|
|
|
|
|
|
|
|
|
|
|
v
e
0
|
|
1
|
|
2
|
|
#NP #VP
3
|
|
|
|
|
|
4
|
|
|
|
|
|
5
|
|
|
|
| #VP #S 6
|
|
e #N |
7
|
|
#N #NP
8
9
Figure 4: The best SP-multiple-alignment produced by the SP computer
model with a New SP-pattern, ‘f o r t u n e f a v o u r s t h e b r
a v e’, representing a sentence to be parsed and a repository of user-supplied
Old SP-patterns representing grammatical categories, including morphemes
and words. Reproduced with permission from Figure 2 in [19].
As can be seen from this example, the building of an SP-multiple-alignment
may achieve the effect of parsing a sentence (‘f o r t u n e f a v o u r
s t h e b r a v e’ in this example) into its grammatical parts and subparts. But as we shall see later, the SP system has strengths in several
5
different aspects of intelligence, and in the representation of several different
kinds of knowledge—and most of this versatility flows from the building of
SP-multiple-alignments.
To create an SP-multiple-alignment like the one shown in Figure 4, the
SP system starts with a relatively large repository of Old SP-patterns, each
one representing a syntactic structure in English, which may be a morpheme,
a word, or a higher-level structure. The Old SP-patterns would ideally be
learned by the system, but pending full development of the learning processes,
the Old SP-patterns may be supplied to the system by the user.
With a repository of Old SP-patterns in place, the SP system is supplied
with the New SP-pattern (‘f o r t u n e f a v o u r s t h e b r a v
e’) and the system tries to build one or more SP-multiple-alignments, each
of which allows the New SP-pattern to be encoded economically in terms the
Old SP-patterns in the SP-multiple-alignment.
The details of how the encoding is done need not detain us here, but it
is relevant to note that the SP-multiple-alignment construct, in conjunction
with unsupervised learning in the SP system (outlined below), appears to
provide a means of achieving relatively high levels of information compression
with many kinds of data.
As with the building of ‘good’ multiple sequence alignments in bioinformatics, the creation of one or more ‘good’ SP-multiple-alignments is normally too complex to be done by any exhaustive process. As with multiple
sequence alignments in bioinformatics, heuristic search is needed, building
SP-multiple-alignments in stages and, at each stage, selecting the best partial structures for further processing. With this approach, it is not normally
possible to guarantee that the best possible SP-multiple-alignment has been
found, but it is normally possible to create SP-multiple-alignments that are
‘good enough’.
3.2
Learning in the SP system
In the SP system, learning is special. Instead of it being a by-product of
the building of SP-multiple-alignments it is a process of creating grammars,
where each grammar is a collection of Old SP-patterns (many of which would
normally be derived from partial matches between SP-patterns within SPmultiple-alignments), and each grammar is scored in terms of its effectiveness
via SP-multiple-alignment in the economical encoding of a target set of New
SP-patterns. As with the building of SP-multiple-alignments, the process is
too complex for exhaustive search so heuristic methods are needed.
In the SP system, learning is normally ‘unsupervised’, deriving structures
from incoming sensory information without the need for any kind of ‘teacher’,
6
or ‘reinforcement’, or anything equivalent. But in case this seems unduly
narrow, it appears that unsupervised learning is the most general kind of
learning and that, within the framework of unsupervised learning in the SP
system, there is potential to model other kinds of learning such as ‘supervised’
learning, ‘reinforcement’ learning, and more.
3.3
SP-neural
A potentially useful feature of the SP system is that it is possible to see how
abstract constructs and processes in the system may be realised in terms of
neurons and their interconnections. This is the basis for SP-neural, a ‘neural’
version of the SP system, described in an early form in [11, Chapter 11], and
in an updated and more detailed form in [19], and illustrated in Figure 5.
In this connection, it is relevant to mention that the SP system, in both
its abstract and neural forms, is quite different from deep learning in artificial neural networks [7] and has substantial advantages compared with such
systems, as described in [20, Section V] and in [23]. Some examples of those
advantages are outlined in Section 8, below.
3.4
Generalising the SP system for two-dimensional
SP-patterns, both static and moving
This brief description of the SP system and how it works may have given the
impression that it is intended to work entirely with sequences of SP-symbols,
like multiple sequence alignments in bioinformatics. But it is envisaged that,
in future development of the system, two-dimensional SP-patterns will be introduced, with potential to represent and process such things as photographs
and diagrams, and structures in three dimensions as described in [14, Section 6.1 and 6,2], and procedures that work in parallel as described in [15,
Sections V-G, V-H, and V-I, and Appendix C].
It is envisaged that, at some stage, the SP system will be generalised to
work with sequences of two-dimensional ‘frames’ from moving visual media.
7
NP 1 D #D N #N #NP
D 9 o n e #D
N 5 b r a v e #N
D 1 t h e #D
N 9 t a b l e #N
… t .... h .... e .... b .... r .... a .... v .... e …
|
|
|
|
|
|
|
|
t
h
e
b
r
a
v
e
receptor array
sensory data
Figure 5: A schematic representation of a partial SP-multiple-alignment
in SP-neural, as discussed in [19, Section 4]. Each broken-line rectangle
with rounded corners represents a pattern assembly—corresponding to an
SP-pattern in the main SP theory; each character or group of characters enclosed in a solid-line ellipse represents a neural symbol corresponding to an
SP-symbol in the main SP theory; the lines between pattern assemblies represent nerve fibres with arrows showing the direction in which impulses travel;
neural symbols are mainly symbols from linguistics such as ‘NP’ meaning
‘noun phrase, ‘D’ meaning a ‘determiner’, ‘#D’ meaning the end of a determiner, ‘#NP’ meaning the end of a noun phrase, and so on. Reproduced with
permission from Figure 3 in [19].
8
4
Versatility in aspects of intelligence
Strengths and potential of the SP system are summarised in this section and
in those that follow. Further information may be found in [13, Sections 5 to
12], [11, Chapters 5 to 9], [20], and in other sources referenced in the sections
that follow. This section outlines the SP system’s versatility in aspects of
intelligence.
The system has, first, strengths and potential in the ‘unsupervised’ learning of new knowledge. As we saw in Section 3.2, this is an aspect of intelligence in the SP system that is different from others because it is not a
by-product of the building of multiple alignments but is, instead, achieved
via the creation of grammars, drawing on information within SP-multiplealignments.
Secondly, other aspects of intelligence exhibited by the SP system are
modelled via the building of SP-multiple-alignments. These other aspects of
intelligence include: the analysis and production of natural language; pattern
recognition that is robust in the face of errors in data; pattern recognition at
multiple levels of abstraction; computer vision [14]; best-match and semantic
kinds of information retrieval; several kinds of reasoning (next paragraph);
planning; and problem solving.
Thirdly, kinds of reasoning that may be modelled in the SP system include: one-step ‘deductive’ reasoning; chains of reasoning; abductive reasoning; reasoning with probabilistic networks and trees; reasoning with ‘rules’;
nonmonotonic reasoning and reasoning with default values; Bayesian reasoning with ‘explaining away’; causal reasoning; reasoning that is not supported
by evidence; the inheritance of attributes in class hierarchies; and inheritance
of contexts in part-whole hierarchies. Where it is appropriate, probabilities
for inferences may be calculated in a straightforward manner ([11, Section
3.7], [13, Section 4.4]).
There is also potential in the system for spatial reasoning [15, Section
IV-F.1], and for what-if reasoning [15, Section IV-F.2].
It seems unlikely that the features of intelligence mentioned above are the
full extent of the SP system’s potential to imitate what people can do. The
close connection that is known to exist between information compression and
concepts of prediction and probability [8, 9, 5], the central role of information
compression in the SP-multiple-alignment framework, and the versatility of
the SP-multiple-alignment framework in aspects of intelligence suggests that
there there are more insights to come.
9
5
Versatility in the representation of knowledge
Although SP-patterns are not very expressive in themselves, they come to
life in the SP-multiple-alignment framework. Within that framework, they
may serve in the representation of several different kinds of knowledge, including: the syntax of natural languages; class-inclusion hierarchies (with or
without cross classification); part-whole hierarchies; discrimination networks
and trees; if-then rules; entity-relationship structures [12, Sections 3 and 4];
relational tuples (ibid., Section 3), and concepts in mathematics, logic, and
computing, such as ‘function’, ‘variable’, ‘value’, ‘set’, and ‘type definition’
([11, Chapter 10], [17, Section 6.6.1], [22, Section 2]).
As previously noted, the addition of two-dimensional SP patterns to the
SP computer model is likely to expand the representational repertoire of the
SP system to structures in two-dimensions and three-dimensions, and the
representation of procedural knowledge with parallel processing.
As with the SP system’s generality in aspects of intelligence, it seems
likely that the SP system is not constrained to represent only the forms of
knowledge that have been mentioned. The generality of information compression as a means of representing knowledge in a succinct manner, the central
role of information compression in the SP-multiple-alignment framework, and
the versatility of that framework in the representation of knowledge, suggest
that the SP system may prove to be a means of representing all the kinds of
knowledge that people may work with.
6
Seamless integration of diverse aspects of
intelligence, and diverse kinds of knowledge, in any combination
An important third feature of the SP system, alongside its versatility in
aspects of intelligence and its versatility in the representation of diverse kinds
of knowledge, is that there is clear potential for the SP system to provide
seamless integration of diverse aspects of intelligence and diverse kinds of
knowledge, in any combination. This is because diverse aspects of intelligence
and diverse kinds of knowledge all flow from a single coherent and relatively
simple source: the SP-multiple-alignment framework.
It appears that seamless integration of diverse aspects of intelligence and
diverse kinds of knowledge, in any combination, is essential in any artificial
system that aspires to the fluidity, versatility and adaptability of the human
10
mind.
Figure 6 shows schematically how the SP system, with SP-multiplealignment centre stage, exhibits versatility and integration.
e
heory of intellig
t
P
en
S
Unsupervised learning
e
c
Th
Several kinds of reasoning:
deductive; abductive; chains of reasoning;
probabilistic networks and trees; with ‘rules’;
nonmonotonic; Bayesian;
Pattern recognition:
causal; inheritance of
robust against
attributes; with
errors in data;
potential for
at multiple levels
SPmore.
of abstraction.
multipleInformation
alignment
Computer vision
retrieval:
scene analysis
best-match
and more.
and ‘semantic’.
Seamless integration of diverse kinds of
intelligence and knowledge, in any
combination.
Analysis and production
of natural language
Figure 6: A schematic representation of versatility and integration in the SP
system, with SP-multiple-alignment centre stage.
7
Potential benefits and applications of the
SP system
Apart from its strengths and potential in modelling aspects of the human
mind, it appears that, in more humdrum terms, the SP system has several
potential benefits and applications. These include:
• Big data. Somewhat unexpectedly, it has been discovered that the SP
system has considerable potential to help solve nine significant problems associated with big data [16]. These are: overcoming the problem
11
of variety in big data; the unsupervised learning of structures and relationships in big data; interpretation of big data via pattern recognition,
natural language processing and more; the analysis of streaming data;
compression of big data; model-based coding for efficient transmission
of big data; potential gains in computational and energy efficiency in
the analysis of big data; managing errors and uncertainties in data;
and visualisation of structure in big data and providing an audit trail
in the processing of big data.
• Autonomous robots. The SP system opens up a radically new approach
to the development of intelligence in autonomous robots [15];
• An intelligent database system. The SP system has potential in the
development of an intelligent database system with several advantages
compared with traditional database systems [12]. In this connection,
the SP system has potential to add several kinds of reasoning and other
aspects of intelligence to the ‘database’ represented by the World Wide
Web, especially if the SP machine were to be supercharged by replacing
the search mechanisms in the foundations of the SP machine with the
high-parallel search mechanisms of any of the leading search engines.
• Medical diagnosis. The SP system may serve as a vehicle for medical
knowledge and to assist practitioners in medical diagnosis, with potential for the automatic or semi-automatic learning of new knowledge
[10];
• Computer vision and natural vision. The SP system opens up a new
approach to the development of computer vision and its integration
with other aspects of intelligence. It also throws light on several aspects
of natural vision [14];
• Neuroscience. As outlined in Section 3.3, abstract concepts in the
SP theory map quite well into concepts expressed in terms of neurons
and their interconnections in a version of the theory called SP-neural
([19], [11, Chapter 11]). This has potential to illuminate aspects of
neuroscience and to suggest new avenues for investigation.
• Commonsense reasoning. In addition to the previously-described
strengths of the SP system in several kinds of reasoning, the SP system
has strengths in the surprisingly challenging area of “commonsense reasoning”, as described by Ernest Davis and Gary Marcus [4]. How the
SP system may meet the several challenges in this area is described in
[18].
12
• Other areas of application. The SP system has potential in several
other areas of application including [17]: the simplification and integration of computing systems; applications of natural language processing; best-match and semantic forms of information retrieval; software
engineering [22]; the representation of knowledge, reasoning, and the
semantic web; information compression; bioinformatics; the detection
of computer viruses; and data fusion.
• Mathematics. The concept of information compression via the matching and unification of patterns provides an entirely novel interpretation
of mathematics [21]. This interpretation is quite unlike anything described in existing writings about the philosophy of mathematics or its
application in science. There are potential benefits in science from this
new interpretation of mathematics.
8
Other strengths of the SP system
Apart the strengths of the SP system described in the previous four sections,
there are others described mainly in [20]. A selection of what appear to be
the more important ones are outlined here:
• The amounts of data and processing required for learning. There is now
widespread recognition of the unreasonably large amounts of data that
are currently required by deep learning systems to learn anything useful, and the correspondingly large amount of processing that is needed.1
Although there is still work to be done in the development of unsupervised learning in the SP system, it is clear from the overall approach
and from what has been achieved already that the learning of meaningful knowledge by the SP system is likely to require substantially less
data and processing than with deep learning systems [20, Sections V-D
and V-E].
• Transparency in the representation of knowledge, and in processing.
It is now widely recognised that a major problem with deep learning
systems is that the way in which learned knowledge is represented in
deep learning systems is far from being transparent and comprehensible
by people, and that the way in which deep learning systems arrive at
their conclusions is difficult or impossible for people to understand.
1
See, for example, “Greedy, brittle, opaque, and shallow: the downsides to deep
learning”, Wired, 2018-02-03, bit.ly/2nM1ccg.
13
These deficiencies are of concern for reasons of safety, legal liability,
and more.2
By contrast, and as with big data (Section 7), knowledge in the SP
system is represented in a manner that is familiar to people, using
such devices as class-inclusion hierarchies, part-whole hierarchies, and
others. And there is an audit trail for all processing in the SP system,
so that it is explicit and comprehensible by people.
• Catastrophic forgetting and continuous learning. A major problem with
deep learning—‘catastrophic forgetting’—is that new leaning wipes out
old learning. A related problem is that, to be practical, a learning system, like a person, should be able to learn continuously from its environment without old knowledge being disturbed by new knowledge.3,4
A robust solution to this problem is intrinsic to the design of the SP
system: new learning does not disturb old learning.
• Learning with generalisation, and learning from ‘dirty data’. A general
problem with any system for unsupervised learning is how to generalise beyond the finite body of information that has been seen since the
‘birth’ of the learning system, and how to correct over-generalisations
and under-generalisations without the provision of a ‘teacher’ or anything equivalent. A related problem is how ‘correct’ knowledge may be
learned from ‘dirty data’, meaning data that contains ‘errors’.
With systems for deep learning, several different solutions to the generalisation problem have been proposed, but none of them have any
good theoretical underpinning [20, Section V-H]. It appears that, with
deep learning systems, no solutions have been offered to the problem
of learning from ‘dirty data’.
By contrast, the SP system offers a relatively simple solution to both
problems that derives directly from the unifying principe—compression
of information—that lies at the heart of the SP system. How compression of information may generalise correctly from its raw data, and how
it may achieve learning from ‘dirty data’ is described in [11, Section
9.5.3] and [13, Section 5.3]. There is also relevant discussion in [20,
2
See, for example, “Inside DARPA’s push to make artificial intelligence explain
itself”, CET US News, 2017-08-10, bit.ly/2FQMoAr.
3
Even if, for example, a person changes their name, the system should be able to
retain both names and the date when the name was changed.
4
It appears that this problem is a matter of concern to military planners as
described, for example, in “DARPA seeking AI that learns all the time”, IEEE Spectrum,
2017-11-21, bit.ly/2BdERfZ.
14
Sections V-H and XI-C]. Experiments with computational models of
learning suggest that the analysis is sound.
9
Conclusion
It seems that the overarching goal of this research—the simplification and
integration of observations and concepts across artificial intelligence, mainstream computing, mathematics, and human learning, perception, and
cognition—has, to a large extent, been achieved.
The SP system provides a favourable combination of simplicity and power:
the concept of SP-multiple-alignment, together with some relatively simple
procedures for unsupervised learning, have proved to be remarkably versatile
across diverse aspects of intelligence, in the representation of diverse kinds of
knowledge, and in the seamless integration of diverse aspects of intelligence,
and diverse kinds of knowledge, in any combination.
That last feature—seamless integration of diverse aspects of intelligence
and diverse kinds of knowledge—appears to be essential in any artificial system that aspires to the fluidity, versatility, and adaptability of the human
mind.
The SP system, compared with AI-related alternatives considered in [20],
appears to provide a relatively firm foundation for the development of general
human-like intelligence.
References
[1] F. Attneave. Some informational aspects of visual perception. Psychological Review, 61:183–193, 1954.
[2] H. B. Barlow. Sensory mechanisms, the reduction of redundancy, and
intelligence. In HMSO, editor, The Mechanisation of Thought Processes,
pages 535–559. Her Majesty’s Stationery Office, London, 1959.
[3] H. B. Barlow. Trigger features, adaptation and economy of impulses.
In K. N. Leibovic, editor, Information Processes in the Nervous System,
pages 209–230. Springer, New York, 1969.
[4] E. Davis and G. Marcus. Commonsense reasoning and commonsense
knowledge in artificial intelligence. Communications of the ACM,
58(9):92–103, 2015.
15
[5] M. Li and P. Vitányi. An Introduction to Kolmogorov Complexity and
Its Applications. Springer, New York, 3rd edition, 2014.
[6] V. Palade and J. G. Wolff. Development of a new machine for artificial
intelligence. Technical report, CognitionResearch.org, 2017. Submitted
for publication. bit.ly/2tWb88M, arXiv:1707.0061 [cs.AI].
[7] J. Schmidhuber. Deep learning in neural networks: an overview. Neural
Networks, 61:85–117, 2015.
[8] R. J. Solomonoff. A formal theory of inductive inference. Parts I and II.
Information and Control, 7:1–22 and 224–254, 1964.
[9] R. J. Solomonoff. The discovery of algorithmic probability. Journal of
Computer and System Sciences, 55(1):73–88, 1997.
[10] J. G. Wolff. Medical diagnosis as pattern recognition in a framework of information compression by multiple alignment, unification and
search. Decision Support Systems, 42:608–625, 2006. bit.ly/1F366o7,
arXiv:1409.8053 [cs.AI].
[11] J. G. Wolff. Unifying Computing and Cognition: the SP Theory and Its
Applications. CognitionResearch.org, Menai Bridge, 2006. ISBNs: 09550726-0-3 (ebook edition), 0-9550726-1-1 (print edition). Distributors,
including Amazon.com, are detailed on bit.ly/WmB1rs.
[12] J. G. Wolff. Towards an intelligent database system founded on the SP
theory of computing and cognition. Data & Knowledge Engineering,
60:596–624, 2007. bit.ly/1CUldR6, arXiv:cs/0311031 [cs.DB].
[13] J. G. Wolff. The SP theory of intelligence: an overview. Information,
4(3):283–341, 2013. bit.ly/1NOMJ6l, arXiv:1306.3888 [cs.AI].
[14] J. G. Wolff. Application of the SP theory of intelligence to the understanding of natural vision and the development of computer vision. SpringerPlus, 3(1):552–570, 2014. bit.ly/2oIpZB6, arXiv:1303.2071
[cs.CV].
[15] J. G. Wolff. Autonomous robots and the SP theory of intelligence. IEEE
Access, 2:1629–1651, 2014. bit.ly/18DxU5K, arXiv:1409.8027 [cs.AI].
[16] J. G. Wolff. Big data and the SP theory of intelligence. IEEE Access,
2:301–315, 2014. bit.ly/2qfSR3G, arXiv:1306.3890 [cs.DB]. This paper,
with minor revisions, is reproduced in Fei Hu (Ed.), Big Data: Storage,
16
Sharing, and Security (3S), Taylor & Francis LLC, CRC Press, 2016,
pp. 143–170.
[17] J. G. Wolff. The SP theory of intelligence: benefits and applications.
Information, 5(1):1–27, 2014. bit.ly/1FRYwew, arXiv:1307.0845 [cs.AI].
[18] J. G. Wolff. Commonsense reasoning, commonsense knowledge, and the
SP theory of intelligence. Technical report, CognitionResearch.org, 2016.
Submitted for publication. bit.ly/2eBoE9E, arXiv:1609.07772 [cs.AI].
[19] J. G. Wolff. Information compression, multiple alignment, and the representation and processing of knowledge in the brain. Frontiers in Psychology, 7:1584, 2016. bit.ly/2esmYyt, arXiv:1604.05535 [cs.AI].
[20] J. G. Wolff. The SP theory of intelligence: its distinctive features
and advantages. IEEE Access, 4:216–246, 2016. bit.ly/2qgq5QF,
arXiv:1508.04087 [cs.AI].
[21] J. G. Wolff. On the “mysterious” effectiveness of mathematics in science. Technical report, CognitionResearch.org, 2017. Submitted for
publication. bit.ly/2otrHD0, viXra:1706.0004, hal-01534622.
[22] J. G. Wolff. Software engineering and the SP theory of intelligence. Technical report, CognitionResearch.org, 2017. bit.ly/2w99Wzq,
arXiv:1708.06665 [cs.SE].
[23] J. G. Wolff. Solutions to problems with deep learning. Technical report,
CognitionResearch.org, 2018. bit.ly/2AJzu4j, arXiv:1801.05457 [cs.LG].
17
| 2cs.AI
|
arXiv:1610.05016v1 [cs.NE] 17 Oct 2016
Weekly maintenance scheduling using exact and
genetic methods
Andrew W. Palmer, Robin Vujanic, Andrew J. Hill, Steven J. Scheding
October 18, 2016
Abstract
The weekly maintenance schedule specifies when maintenance activities should be performed on the equipment, taking into account the
availability of workers and maintenance bays, and other operational constraints. The current approach to generating this schedule is labour intensive and requires coordination between the maintenance schedulers and
operations staff to minimise its impact on the operation of the mine. This
paper presents methods for automatically generating this schedule from
the list of maintenance tasks to be performed, the availability roster of the
maintenance staff, and time windows in which each piece of equipment
is available for maintenance. Both Mixed-Integer Linear Programming
(MILP) and genetic algorithms are evaluated, with the genetic algorithm
shown to significantly outperform the MILP. Two fitness functions for the
genetic algorithm are also examined, with a linear fitness function outperforming an inverse fitness function by up to 5% for the same calculation
time. The genetic algorithm approach is computationally fast, allowing
the schedule to be rapidly recalculated in response to unexpected delays
and breakdowns.
1
Introduction
Maintenance activities are a significant cost in the mining sector, making up
between 30% and 60% of the total operating cost of a mine (Singleton & Krellis,
1998; Lewis & Steinberg, 2001; Dhillon, 2008). In addition to direct costs, poorly
planned maintenance can have a large impact on mine productivity—equipment
spending longer in maintenance than needed effectively reduces the total tonnes
of ore that a mine can produce. As a result, generating maintenance schedules
that minimise the impact on the operation of the mine is of great importance to
mine operators. Existing work on planning maintenance in the mining industry
has looked at modelling the reliability of the equipment (Summit & Halomoan,
2015), or focused on long timescales, allocating equipment to tasks with the
objective to minimise the expected maintenance costs over the lifetime of the
mine (Topal & Ramazan, 2010). This paper examines the problem of generating
the weekly maintenance schedule for the multitude of activities to be performed
1
on each piece of equipment, with the objective of minimising the impact of
the maintenance on the mine operations. The proposed approach is aware of
constraints such as the number of maintenance bays and the availability roster
of the maintainers, and is demonstrated on real and simulated datasets that
reflect practical problem sizes.
The current approach to producing weekly maintenance schedules is largely
manual, and requires the dedicated schedulers to coordinate with the operations
planners to develop a schedule that satisfies operational constraints. While the
intention is that the schedule should be generated once for the week, they are frequently revised to incorporate unexpected breakdown of equipment and delays
(Tomlingson, 2007). The aim of this work is to automatically and quickly generate the schedule given the list of maintenance tasks that are to be performed,
the availability roster of the maintenance staff, and time windows for each piece
of equipment in which the equipment can be taken down for maintenance with
minimal impact on the mine operations.
To the best of the authors’ knowledge, the weekly maintenance scheduling problem has not been studied in the literature. Some related maintenance
scheduling problems are examined by Gopalakrishnan et al. (2001); Deris et al.
(1999); Ben Ali et al. (2011); Jiu et al. (2013); Jin et al. (2009); Aissani et al.
(2009); Pandey et al. (2011); Najid et al. (2011). Many of these authors used
meta-heuristics such as tabu-search (Gopalakrishnan et al. , 2001) and Genetic
Algorithms (GAs) (Deris et al. , 1999; Ben Ali et al. , 2011; Jiu et al. , 2013)
to generate a schedule. These meta-heuristic approaches were shown to produce near-optimal solutions in reasonable calculation times. An option-based
cost model was used by Jin et al. (2009), while a novel reinforcement learning
approach was utilised by Aissani et al. (2009). Methods for jointly optimising
maintenance and production were developed by Pandey et al. (2011); Najid
et al. (2011). Najid et al. (2011) formulated this as a Mixed-Integer Linear
Program (MILP) and solved it using commercial optimisation software. However, the authors pointed out that only small problem instances were solvable,
and stated that they intend to investigate meta-heuristics for solving larger
instances.
The major difference between the above problems and producing a weekly
mine maintenance schedule is in the size of the problem. Typical weekly maintenance schedules can contain over 100 pieces of equipment, with up to 50 individual activities per piece of equipment. Durations for each activity can range
from half an hour to several days, and approximately 25 different types of workers, such as fitters, electricians, and boilermakers, are required. The availability
of these workers varies both day to day and between the day and night shifts.
One of the characteristics of mining in general is that it is very dynamic, with
unexpected events such as equipment failures occurring frequently. Thus, one of
the requirements of the scheduling system is that it should be computationally
fast to allow rapid replanning in response to these unexpected events.
Two approaches to this problem are proposed in this paper—a MILP formulation that is solved using Gurobi (Gur, n.d.), and a GA. While the MILP
approach generates optimal solutions, it will be shown to be computationally
2
infeasible for realistic problem sizes, motivating the use of the GA. The specific
contributions of this paper are:
• A MILP formulation of the problem
• A GA approach that utilises a greedy heuristic to ensure that feasible
schedules are generated from the chromosomes
• A comprehensive evaluation of the MILP method and the GA using two
different fitness functions
The remainder of this paper is structured as follows: Section 2 presents a
MILP formulation of the problem, and Section 3 develops the GA. A comparison
of the methods is presented in Section 4, and concluding remarks are provided
in Section 5.
2
MILP Formulation
The weekly maintenance problem is to schedule a given set of tasks (equipment)
i ∈ I with associated subtasks (work orders) j ∈ Ji . Each task has a ready time
that specifies the earliest a task can be started, and a deadline that specifies
when it should be completed by. Some tasks require a maintenance bay, and
subtasks have a specified duration and worker requirement. Subtasks can have
a precedence requirement where they require other subtasks of the task to be
performed before it can be performed. An example task with 7 subtasks is
shown in Figure 1. The worker requirements of the subtasks in the example are
denoted by the colours and specified numbers in the subtasks, and the subtasks
have been scheduled to respect the precedence of the subtasks. The schedule is
split into discrete, equally-sized time periods. The objective of the problem is to
schedule the subtasks to minimise the sum of the makespan and lateness of the
tasks. Makespan is simply the length of a task, and lateness is incurred if the
task is completed after the deadline. The rest of this section formally presents
the MILP model.
2.1
Indices and Sets
t∈T
time periods, where T = {0, . . . , tmax }
i∈I
tasks
j∈J
subtasks
p∈P
worker types
The set of tasks that require a maintenance bay is M ⊆ I, the set of subtasks
that compose task i is Ji ⊆ J, and the set of subtasks that must be performed
before j is Kj ⊆ J. Subtasks can only belong to one task, so Ji ∩ Jk = ∅ ∀i ∈
I, k ∈ I\{i}.
3
Figure 1: An example task with its subtasks scheduled as early as possible
taking into account the precedence constraints. Examples of the precedence
constraints seen in this example are the vehicle must be washed before further
tasks are performed, and the tyres must be removed and replaced before and
after the brake service. The colour of the subtask specifies the type of personnel
required (5 in the example, which could include electrician, fitter, boilermaker,
etc.), and the number of people required is specified by the number in each
subtask.
2.2
Parameters
dj
rj,p,s,t
ap,t
m
bi
ci
fi
gi
2.3
the duration of subtask j ∈ J in time periods
the number of workers of type p ∈ P required for subtask j ∈ J in timestep t ∈ T if j started in timestep
s ∈ {max(0, t − dj , . . . , t)}
the number of workers of type p ∈ P available in time
period t ∈ T
the number of maintenance bays
ready time for task i ∈ I
deadline for task i ∈ I
the weighting of the makespan of task i ∈ I
the weighting of the lateness of task i ∈ I
Optimisation Variables
xstart
∈ {0, 1}
j,t
xi,t ∈ {0, 1}
xstart
∈ {0, 1}
i,t
xfinish
∈ {0, 1}
i,t
yimakespan ∈ R
yilateness ∈ R
1 iff subtask j ∈ J starts in time period t ∈ T
1 iff task i ∈ I is being performed in time period t ∈ T
1 iff task i ∈ I starts in time period t ∈ T
1 iff task i ∈ I finishes in time period t ∈ T
makespan of task i ∈ I
lateness of task i ∈ I
4
2.4
Objective function
The objective function is the sum of the weighted makespan and lateness of
each task:
X
J=
fi yimakespan + gi yilateness
(1)
i∈I
2.5
2.5.1
Constraints
Subtasks
The first set of constraints specify when a subtask can start. Firstly, (2) enforces
that each subtask has to start exactly once, and (3) specifies that it has to be
completed before the end of the schedule. Constraint (4) encodes the precedence
constraints as specified in Kj .
X
xstart
=1
∀j ∈ J
(2)
j,t
t
X
(txstart
j,t ) + dj ≤ tmax + 1
∀j ∈ J
(3)
t
X
X
(txstart
(txstart
j,t ) ≥
k,t ) + dk
t
2.5.2
∀j ∈ J, k ∈ Kj
(4)
t
Tasks
The next set of constraints deal with the tasks. Constraint (5), similar to (2),
specifies that a task can only start once. The period that the task starts in is
defined in (6) as the earliest starting period of its subtasks. The makespan of
the task is calculated in (7) by specifying that the task must finish no earlier
than any of its subtasks. Constraint (8) enforces that the task can only finish
once, and (9) specifies that the period in which the task is finished in is no earlier
than its start period plus its makespan. The lateness of the task is calculated
in (10). Constraint (11) calculates whether a task is being performed in a time
period or not based on whether it was being performed in the previous time
period, and the time period in which the task starts and finishes. For (11) to
work when t = 0, a dummy variable is added for each task in the time period
t = −1 and assigned the value of 0 in (12).
X
xstart
=1
∀i ∈ I
(5)
i,t
t
X
X
(txstart
(txstart
i,t ) ≤
j,t )
t
∀i ∈ I, j ∈ Ji
X
X
makespan
(txstart
≥
(txstart
i,t ) + yi
j,t ) + dj
t
(6)
t
t
5
∀i ∈ I, j ∈ Ji
(7)
X
xfinish
=1
i,t
∀i ∈ I
(8)
t
X
X
(txfinish
)≥
i,t
t
makespan
(txstart
i,t ) + yi
∀i ∈ I
(9)
t
yilateness ≥
X
(txfinish
) − ci
i,t
∀i ∈ I
(10)
t
xi,t = xi,t−1 + xstart
− xfinish
i,t
i,t
∀i ∈ I
xi,−1 = 0
2.5.3
∀i ∈ I, t ∈ T
(11)
(12)
Resources
There are two resource constraints, namely a limited number of maintenance
bays, and a limited number of workers. Constraint (13) enforces a limit on the
number of maintenance bays being used in each time period. Note that only
a subset of the tasks require a maintenance bay. Constraint (14) encodes the
limit on the number of each type of worker in each time period.
X
xi,t ≤ m
∀t ∈ T
(13)
i∈M
X
X
xstart
j,s rj,p,s,t ≤ ap,t
∀p, t
(14)
j∈J s∈{max(0,t−dj ),...,t}
2.6
Optimisation Problem
The maintenance scheduling problem is therefore expressed as the optimisation
model P :
(
minimise (1)
P =
subject to (2) − (14)
Two solution methods to this optimisation problem are proposed—a branch
and bound approach using commercial software package Gurobi Gur (n.d.), and
a GA that is developed in the next section. As will be shown later in the results,
this optimisation problem is hard for Gurobi, motivating the development of the
genetic algorithm.
6
3
Genetic Algorithm
Genetic algorithms are a class of meta-heuristics that aim to emulate the process of natural selection. They use a population of chromosomes to represent
possible solutions to a problem, with each chromosome consisting of a set of
genes that describe the solution. To create a new generation of chromosomes,
chromosomes from the current generation are randomly combined in proportion
to their fitness. In this way, traits from the strongest chromosomes are most
likely to be carried forward to the next generation (Mitchell, 1998). Genes are
also randomly mutated to avoid being trapped in local optima. This process is
outlined in Figure 2. A common addition to GAs is elitism, in which the best
chromosome(s) from the current generation are propagated to the next generation without modification. Elitism guarantees that solution quality will not
decrease (Baluja & Caruana, 1995).
The set of chromosomes in a given population will be denoted by N . The
rest of this section first discusses possible chromosome representations and how
a maintenance schedule is generated from a chromosome, before presenting the
methods used in each step of the GA.
3.1
Chromosome representation and schedule generation
Several different chromosome representations were considered. An obvious representation to choose is a direct representation, where the genes consist of binary
variables corresponding to the optimisation variables introduced in Section 2.
However, most of the values that this chromosome representation can take result in infeasible schedules due to the constraints on the availability of workers
and maintenance bays. Thus, it was desirable to develop a representation that
would reliably produce feasible schedules.
The chromosome representation proposed for this application is an ordered
list of tasks. The tasks are then scheduled using a greedy heuristic in the
order specified by the chromosome to determine the values of the optimisation
variables outlined in Section 2.3. The greedy heuristic places the subtasks of
each task as early as possible while satisfying the constraints in Section 2.5.
Provided there is always at least 1 maintenance bay and the minimum number
of workers required for any subtasks available in each time period, then this
representation is guaranteed to always produce a feasible solution. Figure 3
shows the schedules resulting from using the greedy heuristic on the two possible
chromosomes for a scenario with two tasks.
The greedy heuristic can fail to produce a feasible schedule in certain cases.
For example, if the type of worker required for a specific subtask is only available
for a small number of time periods at the beginning of the schedule, and the
task requires a maintenance bay, then if the task is too late in the chromosome
the greedy heuristic can fail to find a valid spot for the subtasks to be placed.
7
Initial population
Calculate fitness
Select parents
Mate parents
using crossover
Mutate children
No
Final
generation?
Yes
Best solution
Figure 2: Flowchart outlining the steps in the genetic algorithm.
8
(a) Tasks scheduled in order 1,2
(b) Tasks scheduled in order 2,1
Figure 3: This figure shows two tasks being scheduled by the greedy heuristic. Task 1 becomes available for maintenance in time period 0, and task 2 in
time period 2 (denoted by the striped area). The tasks are otherwise identical,
consisting of two subtasks that each require 1 worker to complete. The first subtask of each task has a duration of 1 time period, and the second subtask has
a duration of 2 time periods. For this example, there is only 1 worker available
and the tasks do not require maintenance bays. In (a), the tasks are greedily
scheduled in the order 1,2, while in (b) the tasks are greedily scheduled in the
order 2,1.
9
3.2
Initial population
The characteristics of the initial population of chromosomes can have a large
impact on the performance of the algorithm (Diaz-Gomez & Hougen, 2007).
There is a trade-off between having an initial population that contains good
initial solutions, which improves the probability of finding a good final solution
(Burke et al. , 2004), and having a diverse initial population, which helps avoid
premature convergence (Leung et al. , 1997). To strike a balance between diversity and quality, two of the chromosomes in the initial population are selected
using heuristics, while the remaining are randomly generated.
The first of the heuristics creates a chromosome by sorting the tasks by
their ready times. The reasoning behind this heuristic is that it should avoid
situations like Figure 3b where the task with the later ready time is performed in
between the subtasks of the other task, leading to a large makespan for the first
task. The second heuristic is primarily aimed at producing a feasible schedule
in scenarios where many of the chromosomes are unable to be converted into a
feasible schedule by the greedy heuristic. Tasks that have subtasks that can only
be performed in a limited range of times are placed earlier in the chromosome
so that they are likely to be able to be scheduled by the greedy heuristic.
3.3
Calculating the fitness of a chromosome
Two fitness functions were investigated. The first fitness function calculates the
fitness of the n-th chromosome, φ1n , as:
φ1n = max{Jν ∀ν ∈ N } − Jn
(15)
where Jn is calculated using the objective function from Eq. (1). Note that the
max term considers only the chromosomes that result in a feasible schedule.
The second function calculates the fitness, φ2n , as:
φ2n =
Jn −
P
1
fi min-makespani
(16)
i∈I
In this formula, the minimum possible makespan cost is subtracted from the
value of the objective function. A perfect schedule in which all tasks take the
minimum possible time and are completed before their deadlines would therefore
result in a denominator value of 0, and a corresponding infinite fitness value.
Note that if a perfect schedule is found, this can simply be returned as the best
schedule without running the GA to completion.
3.4
Parent selection
A pair of parents are randomly selected for each chromosome in the next generation in proportion to their fitness using the roulette wheel selection method
(Davis, 1991). Chromosomes that do not result in feasible schedules do not have
a fitness and are not considered as valid parents.
10
3.5
Crossover
Two-point crossover is used to generate a new chromosome from the two parents,
as illustrated in Figure 4. First, one of the parents is selected at random to be
the dominant parent and two crossover points in the chromosome are randomly
selected. If the number of tasks between the crossover points is larger than the
number of tasks outside the crossover points, then the dominant parent’s tasks
between the crossover points are copied to the child chromosome. Otherwise,
the dominant parent’s tasks outside of the crossover points are copied to the
child chromosome. The tasks that were not copied across are then greedily
placed into the child chromosome in the order in which they appear in the
non-dominant parent.
3.6
Mutation
The final step in each round of the GA is to mutate the children. A probabilistic
check against the mutation rate is performed for each gene in the chromosome.
If the check succeeds, then the gene is swapped with another randomly selected
gene in the chromosome. The result of this is shown in Figure 4d.
4
Computational Study
Two weekly maintenance schedules from a sample mine were used to generate
the input datasets used in this paper. Each weekly schedule had just under 100
pieces of equipment, and each piece of equipment had between 1 and 50 subtasks,
yielding a total of approximately 800 subtasks. 25 different types of people
were required for performing the work, each with differing availability levels.
Some types of people were only available during the day-shift, while others were
available during both the day-shift and night-shift, but in varying numbers.
Approximately 1/3rd of the pieces of equipment required a maintenance bay,
while the remaining were serviced in the field. The mine site under consideration
had 5 bays in the maintenance shed available for equipment being serviced.
This section first evaluates the performance of the commercial solver and
GA approaches as the number of tasks is varied, followed by a comprehensive
comparison of the methods on randomly generated datasets with the task and
worker tightnesses varied. The commercial solver used was Gurobi (Gur, n.d.),
and, unless otherwise specified, a time limit of 600s was used for solving the
model. The methods tested were the GA with the fitness function defined
in (15) (GA1), the GA with the fitness function defined in (16) (GA2), the
Gurobi (MILP), and Gurobi seeded with an initial solution using a heuristic
(MILP+H). The heuristic used in the MILP+H method simply sorted the tasks
by their ready time.
The GA used a population size of 100 chromosomes, 60 generations, a mutation rate of 0.1% per gene, and elitism of 1 chromosome. These values were
experimentally found to give good performance. These parameters were not
11
Parent 1
1
2
3
4
5
6
7
8
9
10
Parent 2
10
9
8
7
6
5
4
3
2
1
Child
(a) Before crossover
Parent 1
1
2
3
4
5
6
7
8
9
10
Parent 2
10
9
8
7
6
5
4
3
2
1
2
3
4
5
6
7
Child
(b) Dominant parent’s genes copied across
Parent 1
1
2
3
4
5
6
7
8
9
10
Parent 2
10
9
8
7
6
5
4
3
2
1
Child
10
2
3
4
5
6
7
9
8
1
(c) Non-dominant parent’s genes copied across
Child
10
2
3
4
5
6
7
9
8
1
Mutated
10
2
7
4
5
6
3
9
8
1
(d) Mutation
Figure 4: Demonstration of the crossover and mutation process on chromosomes
consisting of 10 genes. Parent 1 is selected as the dominant parent, and crossover
points between the 1st and 2nd genes, and between the 7th and 8th genes are
chosen. In (b), the genes between the crossover points in parent 1 are copied to
the child. In (c), the genes not currently in the child chromosome are inserted
in the order they appear in parent 2. Finally, in (d), a random mutation swaps
the 3rd and 7th genes.
12
varied for the different fitness functions as the intention was to show the difference due only to the choice of fitness function. The purpose of this paper was
also not to determine the optimal parameters for the GA, as these should be
tuned for the specific scenario under consideration. Unless otherwise noted, all
calculation times are from an i7-4810MQ with 16GB of RAM, and Gurobi was
set to use a maximum of 4 threads. The GA was programmed by the authors
in Python and is single threaded.
4.1
Varying the number of tasks
The performance of the approaches as the size of the problem was varied was
first examined. The number of tasks was varied to correspond to approximately
1, 2, 3, 4, 5, 6, and 7 days worth of tasks in the original schedule. Figure 5 shows
the optimality gap for each method, where the optimality gap was calculated
based on the best lower bound calculated by Gurobi using the following formula:
best solution cost
−1
(17)
best bound
Gurobi was only able to find a solution in the first case, in which it found
the optimal solution. When provided with an initial solution, it found the
optimal solution in the first two cases, but was unable to improve upon the
heuristically generated initial solution in the remaining cases. Figure 6 shows
the calculation times of each method. As can be seen, Gurobi very quickly
hit the time limit of 600s, while the maximum calculation time for the GA
approaches was approximately 230s.
To highlight the infeasibility of the commercial solvers for realistic problem
sizes, Gurobi was run for 6 hours using 8 threads on the full 7 days worth of
tasks. In this case, the models have over 300,000 binary variables to solve for.
Even running for 6 hours, Gurobi was unable to find a feasible solution for the
MILP model, or improve from the initial solution for the MILP+H model. In
addition, Gurobi was unable to improve the lower bound from the lower bound
produced by the initial root relaxation of the model.
optimality gap =
4.2
Random scenarios
A set of random scenarios was generated from the input schedules to comprehensively test the performance of each method. Two aspects of the problem were
varied—the tightness of the deadline, and the tightness of the worker availability
constraints. While the tightness of the deadline impacts the objective value of
the optimal schedule, it does not impact on the difficulty of the problem. The
tightness of the worker availability constraints, on the other hand, impacts both
the objective value of the optimal schedule as well as the problem difficulty. 100
random scenarios were created by randomly sampling the ready time of each
task. The deadline of each task was generated from the ready time, minimum
possible makespan of the task, and the specified deadline tightness, π, using the
following formula:
13
Optimality gap versus number of tasks
GA1
GA2
MILP
MILP+H
40
35
Optimality gap (%)
30
25
20
15
10
5
0
10
20
30
40
50
60
70
Number of tasks
80
90
100
Figure 5: Optimality gap versus number of tasks. Note that a feasible solution
was only found in the first case for Gurobi solving the MILP.
Calculation time versus number of tasks
600
Calculation time (s)
500
400
GA1
GA2
MILP
MILP+H
300
200
100
0
10
20
30
40
50
60
70
Number of tasks
80
90
Figure 6: Calculation time versus number of tasks.
14
100
ci = bi + π × min-makespani
(18)
Values of π ∈ {1, 1.5, 2} were tested. Three levels of worker tightness were
also considered. The tightest level used the actual number of workers available
for the supplied schedules—in this case, there were between 1 and 6 workers of
most types available, depending on whether it was the day shift of night shift,
while 1 or 2 of the types of workers had 10 workers available. The next tightest
level assumed that there were 10 workers of every type available. In this way,
most of the worker availability constraints become somewhat irrelevant, reducing the difficulty of the problem. Finally, the loosest level of worker tightness
assumed that 15 workers of every type available.
Each of the 100 random scenarios was tested using every combination of
deadline tightness and worker tightness, yielding 900 different scenarios in total.
A summary of the results are displayed in Tables 1 and 2. Note that the results
for the MILP method are omitted as Gurobi was only able to find a feasible
solution in 6 of the 900 instances. The heuristic used to seed the MILP+H
method produced feasible solutions in all but 3 of the instances, and in 14 cases
Gurobi found the optimal solution. It should be noted, though, that the GA1
and GA2 methods also found the optimal solution in those 14 cases, suggesting
that they were particularly easy instances.
Table 1 shows the average optimality gap of each method calculated using
(17) and the best lower bound found by Gurobi. It is clear that Gurobi is
significantly outperformed by the GA approaches, while, in general, the GA1
approach outperformed the GA2 approach. Similar to the scenarios in Section
4.1, Gurobi struggled to improve upon both the heuristically generated initial
solution and the initial lower bound. It is therefore difficult to comment on how
close to the true optimal solution the GA approaches were.
Table 2 shows the relative performance of each method using the GA1
method as the reference point. As can be seen, the GA1 method clearly outperforms the GA2 approach when both the deadlines and worker constraints
are tight, with an almost 5% difference in their costs. As these constraints are
loosened, the performance difference between the methods decreases. This is a
characteristic of the different fitness functions used—when the cost of the schedules are high, the fitness function in (15) provides much better discrimination
between chromosomes than the fitness function in (16). This is highlighted in
Figure 7a, where the GA1 approach is shown to converge significantly quicker
than the GA2 approach. On the other hand, when the schedule costs are very
low, the GA2 method can converge significantly quicker than the GA1 method,
as shown in Figure 7b. While GA2 converges faster however, the cost of the
best chromosome progresses similarly with the number of generations for both
methods.
15
Table 1: Average optimality gap to best known bound for random scenarios
Deadline tightness
Tight
→
Loose
Worker tightness
Method
Tight
MILP+H
GA1
GA2
100.1%
48.6%
55.6%
100.4%
43.1%
46.6%
63.0%
22.1%
24.0%
↓
MILP+H
GA1
GA2
52.2%
23.5%
25.1%
42.7%
15.3%
15.1%
20.7%
3.3%
3.6%
Loose
MILP+H
GA1
GA2
15.3%
5.3%
5.9%
16.8%
4.9%
4.9%
7.5%
0.3%
0.6%
Table 2: Average performance deficit relative to GA1 for random scenarios.
Bold entries are the cases where the GA1 method was outperformed by the
respective method.
Deadline tightness
Tight
→
Loose
Worker tightness
Method
Tight
MILP+H
GA1
GA2
34.4%
0.0%
4.9%
39.2%
0.0%
2.5%
32.6%
0.0%
1.5%
↓
MILP+H
GA1
GA2
21.6%
0.0%
1.3%
22.3%
0.0%
-0.1%
16.4%
0.0%
0.4%
Loose
MILP+H
GA1
GA2
9.2%
0.0%
0.5%
11.0%
0.0%
0.0%
7.0%
0.0%
0.2%
16
Cost versus Generation
22000
GA1 min
GA1 avg
GA1 max
GA2 min
GA2 avg
GA2 max
20000
Cost
18000
16000
14000
12000
10000
8000
0
10
20
30
Generation
40
50
60
(a) High cost
Cost versus Generation
9000
GA1 min
GA1 avg
GA1 max
GA2 min
GA2 avg
GA2 max
8000
Cost
7000
6000
5000
4000
3000
0
10
20
30
Generation
40
50
60
(b) Low cost
Figure 7: Performance of the GA1 and GA2 methods as the number of generations is increased. The minimum, average, and maximum costs in each generation are reported. In (a), the solution costs are significantly higher than the
lowest possible schedule cost of approximately 2900, while in (b) the solution
costs are very close to it.
17
5
Conclusion
This paper developed and compared several methods for automatically generating a maintenance schedule for a typical fleet of mining equipment. Current
approaches for scheduling maintenance are manual and time intensive. A mixedinteger linear programming model of the problem was formulated. Commercial
optimisation software was unable to sufficiently solve the model, even when
supplied with a heuristically generated starting schedule, necessitating the development of alternative strategies. To this end, a genetic algorithm approach
was developed, which significantly outperformed the commercial optimisation
software. Two fitness functions were also compared, with a linear fitness function shown to in general outperform an inverse fitness function.
There are several avenues for future work. Further improvements to the
performance of the genetic algorithm without increasing the computation time
can be achieved by switching from Python to a compiled language such as C++.
Genetic algorithms are also embarrassingly parallelisable, so further reductions
in the computation time can be achieved in this way. On the algorithmic side,
running multiple independent genetic algorithms in parallel may improve the
robustness of the approach by helping to avoid local minima. Dynamically
switching the fitness function used based on the current chromosome costs could
also yield potential improvements in the convergence rate of the algorithm.
Finally, methods for incorporating the previous schedule as a starting point
when replanning could be investigated.
Acknowledgements
This work was supported by the Rio Tinto Centre for Mine Automation and
the Australian Centre for Field Robotics, University of Sydney, Australia.
References
Gurobi Optimization.
Aissani, N., Beldjilali, B., & Trentesaux, D. 2009. Dynamic scheduling of maintenance tasks in the petroleum industry: A reinforcement approach. Engineering Applications of Artificial Intelligence, 22(7), 1089–1103.
Baluja, Shumeet, & Caruana, Rich. 1995. Removing the genetics from the
standard genetic algorithm. Icml, 1–11.
Ben Ali, M., Sassi, M., Gossa, M., & Harrath, Y. 2011. Simultaneous scheduling
of production and maintenance tasks in the job shop. International Journal
of Production Research, 49(13), 3891–3918.
Burke, Edmund K., Gustafson, Steven, & Kendall, Graham. 2004. Diversity in
Genetic Programming: An Analysis of Measures and Correlation with Fitness.
IEEE Transactions on Evolutionary Computation, 8(1), 47–62.
18
Davis, Lawrence. 1991. Handbook of genetic algorithms. Van Nostrand Reinhold.
Deris, Safaai, Omatu, Sigeru, Ohta, Hiroshi, Shaharudin Kutar, Lt.Cdr, & Abd
Samat, Pathiah. 1999. Ship maintenance scheduling by genetic algorithm
and constraint-based reasoning. European Journal of Operational Research,
112(3), 489–502.
Dhillon, Balbir S. 2008. Mining equipment reliability, maintainability, and
safety. Springer Science & Business Media.
Diaz-Gomez, Pa, & Hougen, Df. 2007. Initial Population for Genetic Algorithms:
A Metric Approach. Proceedings of the 2007 International Conference on
Genetic and Evolutionary Methods, 43–49.
Gopalakrishnan, M, Mohan, S, & He, Z. 2001. A tabu search heuristic for preventive maintenance scheduling. Computers & Industrial Engineering, 40(12), 149–160.
Jin, Xiaoning, Li, Lin, & Ni, Jun. 2009. Option model for joint production and
preventive maintenance system. International Journal of Production Economics, 119(2), 347–353.
Jiu, Song, Zhou, Zhili, & Liu, Jiyin. 2013. The equipment maintenance scheduling problem in a coal production system. International Journal of Production
Research, 51(17), 5309–5336.
Leung, Yee, Gao, Yong, & Xu, Zong Ben. 1997. Degree of population diversity A perspective on premature convergence in genetic algorithms and its Markov
chain analysis. IEEE Transactions on Neural Networks, 8(5), 1165–1176.
Lewis, Michael W., & Steinberg, Luiz. 2001. Maintenance of mobile mine equipment in the information age. Journal of Quality in Maintenance Engineering,
7(4), 264–274.
Mitchell, Melanie. 1998. An Introduction to Genetic Algorithms. MIT press.
Najid, Najib M., Alaoui-Selsouli, Marouane, & Mohafid, Abdelmoula. 2011. An
integrated production and maintenance planning model with time windows
and shortage cost. International Journal of Production Research, 49(8), 2265–
2283.
Pandey, Divya, Kulkarni, Makarand S, & Vrat, Prem. 2011. A methodology for
joint optimization for maintenance planning, process quality and production
scheduling. Computers & Industrial Engineering, 61(4), 1098–1106.
Singleton, T, & Krellis, O. 1998. Mine maintenance - the cost of operation. Coal
Operators Conference, 81–90.
19
Summit, Raymond, & Halomoan, David. 2015. Reliability modelling for maintenance scheduling of mobile mining equipment. Pages 526–540 of: Proceedings
of the 11th Biennial Engineering Mathematics and Applications Conference,
vol. 55.
Tomlingson, Paul D. 2007. Mine Maintenance Management Reader. Society for
Mining, Metallurgy, and Exploration.
Topal, Erkan, & Ramazan, Salih. 2010. A new MIP model for mine equipment
scheduling by minimizing maintenance cost. European Journal of Operational
Research, 207(2), 1065–1071.
20
| 9cs.NE
|
Simple groups, product actions, and generalised quadrangles
arXiv:1702.07308v1 [math.GR] 23 Feb 2017
John Bamberg, Tomasz Popiel, Cheryl E. Praeger
Abstract. The classification of flag-transitive generalised quadrangles is a long-standing open problem
at the interface of finite geometry and permutation group theory. Given that all known flag-transitive
generalised quadrangles are also point-primitive (up to point–line duality), it is likewise natural to seek
a classification of the point-primitive examples. Working towards this aim, we are led to investigate
generalised quadrangles that admit a collineation group G preserving a Cartesian product decomposition of the set of points. It is shown that, under a generic assumption on G, the number of factors of
such a Cartesian product can be at most four. This result is then used to treat various types of primitive
and quasiprimitive point actions. In particular, it is shown that G cannot have holomorph compound
O’Nan–Scott type. Our arguments also pose purely group-theoretic questions about conjugacy classes
in non-Abelian finite simple groups, and about fixities of primitive permutation groups.
1. Introduction
Generalised quadrangles are point–line incidence geometries introduced by Tits [25] in an attempt
to find geometric models for simple groups of Lie type. The classical generalised quadrangles arise
in this way [22, Section 3]. Each admits one of the simple classical groups T = PSp(4, q) ∼
= Ω5 (q),
PSU(4, q) ∼
(q)
or
PSU(5,
q)
acting
transitively
on
flags
(incident
point–line
pairs).
Moreover,
the
= PΩ−
6
point and line stabilisers are certain maximal subgroups of T , so T acts primitively on both points and
lines. The classification of flag-transitive generalised quadrangles is a long-standing open problem. In
addition to the classical families, only two other flag-transitive examples are known (up to point–line
duality), each admitting an affine group acting point-primitively but line-imprimitively. Hence, all of
the known flag-transitive generalised quadrangles are also point-primitive (up to duality), and so it is
natural to seek a classification of the point-primitive examples. Indeed, this is arguably a more difficult
problem, because one begins with essentially no information about the action of the collineation group
on lines, nor any notion of what ‘incidence’ means, whereas in a flag-transitive point–line geometry,
points and lines correspond to cosets of certain subgroups of the collineation group, and incidence is
determined by non-empty intersection of these cosets.
Here we prove the following theorem. The abbreviations HS (holomorph simple), HC (holomorph
compound), SD (simple diagonal), CD (compound diagonal), PA (product action), AS (almost simple)
and TW (twisted wreath) refer to the possible types of non-affine primitive permutation group actions,
in the sense of the O’Nan–Scott Theorem as stated in [19, Section 6]. In the second column of
Table 1, soc(G) denotes the socle of the group G, namely the subgroup generated by its minimal
normal subgroups. By fixΩ (h) we mean the number of elements fixed by a permutation h of the set
Ω, and Q− (5, 2) is the unique generalised quadrangle of order (2, 4).
Theorem 1.1. If Q is a thick finite generalised quadrangle with a non-affine collineation group G
that acts primitively on the point set P of Q, then the action of G on P does not have O’Nan–Scott
type HC, and the conditions in Table 1 hold for the remaining O’Nan–Scott types.
2010 Mathematics Subject Classification. primary 51E12; secondary 20B15, 05B25.
Key words and phrases. generalised quadrangle, primitive permutation group, finite simple group, centraliser, fixity.
The first author acknowledges the support of the Australian Research Council (ARC) Future Fellowship
FT120100036. The second author acknowledges the support of the ARC Discovery Grant DP140100416. The research
reported in the paper forms part of the ARC Discovery Grant DP140100416 of the third author. We thank Elisa Covato
and Tim Burness for making available to us the results quoted in Section 6, and Luke Morgan for helpful discussions.
1
2
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
Type soc(G)
Necessary conditions
HS
T is of Lie type Aε5 , Aε6 , B3 , C2 , C3 , Dε4 , Dε5 , Dε6 , Eε6 , E7 or F4
T is a sporadic simple group or T ∼
= Altn with n 6 18; or
T is an exceptional Lie type group; or
T has Lie type A1 , Aεn with 2 6 n 6 8; Bn or Cn with 2 6 n 6 4; or Dεn with 4 6 n 6 8
r = 2 and T ∼
= Altn with n 6 9; or
r = 2 and T is a sporadic simple group with T ∼
6 Suz, Co2 , Fi22 , Fi23 , B or M; or
=
r = 2 and T has Lie type A1 , Aε2 , Aε3 , B2 , 2 B2 , 2 F4 , G2 or 2 G2 ; or
r = 3 and T ∼
= J1 or T is of Lie type A1 or 2 B2
T ×T
SD
Tk
CD
(T k )r
PA
Tr
r = 2 and fixΩ (h) < |Ω|3/5 for all h ∈ H \ {1}; or
3 6 r 6 4, T is a group of Lie type, and fixΩ (h) < |Ω|1−r/5 for all h ∈ H \ {1}; or
3 6 r 6 4, H = T ∼
= Altp with point stabiliser p. p−1
2 for a prime p ≡ 3 (mod 4)
AS
T
TW
Tr
fixP (g) < |P|4/5 for all g ∈ G \ {1}; or
Q = Q− (5, 2) with T ∼
= PSU4 (2)
fixP (g) < |P|4/5 for all g ∈ G \ {1}
Table 1. Conditions for Theorem 1.1. Here T is a non-Abelian finite simple group, k > 2
and r > 2. If G acts primitively of type CD (respectively PA) on P, then G 6 H ≀ Symr for
some primitive group H 6 Sym(Ω) of type SD (respectively AS) with socle T k (respectively T ).
Note also that, in the notation for finite simple groups of Lie type used in Table 1 (and throughout
+
−
−
2
+
−
2
2
the paper), ε = ± and A+
n = An , An = An , Dn = Dn , Dn = Dn , E6 = E6 , E6 = E6 .
Before we proceed, a remark is in order about the assumption in Theorem 1.1 that G not be
an affine group. If G is affine, then the generalised quadrangle Q necessarily arises from a so-called
pseudo-hyperoval in a projective space PG(3n − 1, q) with q even [12]. In joint work with Glasby [6],
we were able to classify the generalised quadrangles admitting an affine group that acts primitively
on points and transitively on lines: they are precisely the two flag-transitive, point-primitive, lineimprimitive generalised quadrangles mentioned above. However, without the extra assumption of
transitivity on lines, the problem is equivalent to the classification of the pseudo-hyperovals that have
an irreducible stabiliser. As explained in [6, Remark 1.3], this latter problem would appear to be
extremely difficult, and possibly intractable. It also has a rather different flavour to the cases treated
in the present paper, and so we do not consider it further here.
Let us now establish some definitions and notation, before discussing further. By a point–line
incidence geometry we mean a triple Γ = (P, L, I ), where P and L are sets whose elements are called
points and lines, respectively, and I ⊆ P × L is a symmetric binary relation called incidence. We write
Γ = (P, L) instead of (P, L, I ) when we do not need to refer to the incidence relation explicitly. Two
points (respectively lines) of Γ are said to be collinear (respectively concurrent) if they are incident
with a common line (respectively point). A collineation of Γ is a permutation of P ∪ L that preserves
P and L setwise and preserves the incidence relation. By a collineation group of Γ we mean a subgroup
of the group of all collineations of Γ, which is called the full collineation group.
A generalised quadrangle is a point–line incidence geometry Q = (P, L) that satisfies the following
two axioms: (i) two distinct points are incident with at most one common line, and (ii) given a point
P and a line ℓ not incident with P , there exists a unique point incident with ℓ that is collinear with P .
The second axiom implies that every pair in P ∪ L is contained in an ordinary quadrangle, and that
Q contains no triangles. All generalised quadrangles considered in this paper are assumed to be finite,
in the sense that P and L are finite sets. If every point is incident with at least three lines, and every
line is incident with at least three points, then Q is said to be thick. In this case, there exist constants
s > 2 and t > 2 such that every point is incident with exactly t + 1 lines and every line is incident with
exactly s + 1 points [27, Corollary 1.5.3]. The pair (s, t) is called the order of Q. Observe also that
there is a natural concept of point–line duality for generalised quadrangles: if (P, L) is a generalised
quadrangle, then so is (L, P); and if (P, L) has order (s, t), then (L, P) has order (t, s).
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
3
Let us now discuss Theorem 1.1 further. The primitive permutation groups on a finite set ∆ are
classified into eight types according to the O’Nan–Scott Theorem as presented in [19, Section 6]. In
2012, Bamberg et al. [3] showed that if a thick finite generalised quadrangle admits a collineation
group G that acts primitively on both points and lines, then G must be an almost simple (AS type)
group. That is, T 6 G 6 Aut(T ) for some non-Abelian finite simple group T . Given that there
exist point-primitive generalised quadrangles that are line-transitive but line-imprimitive, our initial
aim was to extend the result of [3] by relaxing the line-primitivity assumption to line-transitivity. In
addition to handling the affine (HA type) case with Glasby [6], we were also able to show that no such
examples arise if the point action has type HS or HC [8].
Theorem 1.1 significantly strengthens and expands upon the results of [3, 8]. The idea behind its
proof begins with the following observations. A primitive group G 6 Sym(∆) of O’Nan–Scott type HC,
CD, PA or TW preserves a Cartesian product decomposition ∆ = Ωr , for some set Ω and some r > 2.
Therefore, in studying point-primitive generalised quadrangles, we are led in particular to consider
generalised quadrangles with collineation groups that preserve a Cartesian product decomposition of
the point set. The following theorem shows that the number of factors of such a decomposition becomes
severely restricted under a fairly generic assumption on the group. Here a semiregular permutation
group action is one in which only the identity element fixes a point, and if H1 , . . . , Hr are
Qrpermutation
groups on sets Ω1 , . . . , Ωr , respectively, then the product action of the direct product i=1 Hi on the
Q
Cartesian product ri=1 Ωi is the action (ω1 , . . . , ωr )(h1 ,...,hr ) = (ω1h1 , . . . , ωrhr ). We also recall that a
permutation group is said to act regularly if it acts transitively and semiregularly.
Theorem 1.2. Let Ω1 , . . . , Ωr be finite sets with 2 6 |Ω1 | 6 · · · 6 |Ωr |, where r > 1, and let
Hi 6 Sym(Ωi ) for each i ∈ {1, . . . , r}. Assume
further that H1 is non-trivial and that its action on
Q
Ω1 is not semiregular. Suppose that N = ri=1 Hi is a collineation group
Q of a thick finite generalised
quadrangle Q = (P, L) of order not equal to (2, 4), such that P = ri=1 Ωi and N has the product
action on P. Then r 6 4, and every non-identity element of H1 fixes less than |Ω1 |1−r/5 points of Ω1 .
The proof of Theorem 1.2 relies on the existence of a non-identity element h1 of H1 that fixes at
least one point
(h1 , 1, . . . , 1) ∈ N of Q that
Q
Q of Ω1 . If r > 2, one can then construct a collineation
fixes at least ri=2 |Ωi | points of the Cartesian product P = ri=1 Ωi . Theorem 1.2 is then deduced
from the following result, which bounds the number of points fixed by a non-identity collineation of
an arbitrary thick finite generalised quadrangle. The proofs of both theorems are given in Section 2.
Theorem 1.3. Let θ be a non-identity collineation of a thick finite generalised quadrangle Q =
(P, L). Then either θ fixes less than |P|4/5 points of Q, or Q is the unique generalised quadrangle
Q− (5, 2) of order (2, 4) and θ fixes exactly 15 of the 27 points of Q.
Remark 1.4. Theorem 1.3 improves a particular case of a recent result of Babai on automorphism
groups of strongly regular graphs [1, Theorem 1.7]. If Q has order (s, t) then its collinearity graph,
namely the graph with vertex set P and two vertices adjacent if and only if they are collinear in Q,
is a strongly regular graph with parameters v = |P| = (s + 1)(st + 1), k = s(t + 1), λ = s − 1 and
µ = t + 1. Roughly speaking, we have v ≈ s2 t and k ≈ st, so the condition k 6 n3/4 in assertion (b)
of [1, Theorem 1.7] becomes t 6 s2 , which is just Higman’s inequality for generalised quadrangles (see
Lemma 2.1(ii)). Babai’s result, which applies far more generally to strongly regular graphs that are
non-trivial, non-graphic and non-geometric, therefore implies that a non-identity collineation θ of Q
can fix at most O(|P|7/8 ) points. Theorem 1.3 sharpens the 7/8 exponent in this bound to 4/5 in the
case of collinearity graphs of generalised quadrangles. (Note also that assertion (a) of [1, Theorem 1.7]
sharpens the 7/8 exponent to 5/6 when, roughly,
t >√
s: the condition k > n2/3 roughly translates to
√
t > s, and the corresponding bound is O( kn), with kn ≈ s3/2 t > (s2 t)5/6 ≈ |P|5/6 when t > s.)
To aid our discussion, let us now state the following immediate corollary of Theorem 1.2.
Corollary 1.5. Let Ω be a finite set with |Ω| > 2, and suppose that H 6 Sym(Ω) is non-trivial
and not semiregular. Suppose that N = H r , r > 1, is a collineation group of a thick finite generalised
quadrangle Q = (P, L) of order not equal to (2, 4), such that P = Ωr and N has the product action on
P. Then r 6 4, and every non-identity element of H fixes less than |Ω|1−r/5 points of Ω.
4
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
T
r=1
r=2
r=3
Altn
sporadic
exceptional Lie type
PSLn+1 (q)
PSUn+1 (q)
PSp2n (q) or Ω2n+1 (q)
PΩ±
2n (q)
5 6 n 6 18
any
any
16n68
26n68
26n64
46n68
56n69
56n66
any except Suz, Co2 , Fi22 , Fi23 , B, M J1
2
2
F4 (q), G2 (q), 2 G2 (q), 2 B2 (q)
B2 (q)
16n63
n = 1, q 6= 7
26n63
—
n=2
—
—
—
Table 2. Possibilities for a non-Abelian finite simple group T with the property that
|CT (x)| < |T |1−r/5 for all x ∈ T \ {1}, for r equal to one of 1, 2 or 3.
We apply Corollary 1.5 to groups N that arise as subgroups of certain types of primitive groups.
This in turn motivates certain questions about non-Abelian finite simple groups. As illustration,
consider the case where Ω = T for some non-Abelian finite simple group T , with H = T × T acting
′
on Ω via ω (x,x ) = x−1 ωx′ . This situation arises when N is the socle (the subgroup generated by the
minimal normal subgroups) of a primitive group of type HS (r = 1) or HC (r > 2). If x′ = x then
the element (x, x′ ) = (x, x) ∈ H fixes precisely |CT (x)| points of Ω, where CT (x) is the centraliser of
x in T . Corollary 1.5 therefore implies that r 6 4, and that |CT (x)| < |T |1−r/5 for all x ∈ T \ {1}.
We therefore ask which non-Abelian finite simple groups T satisfy this condition. If r = 4 then we
require that |CT (x)| < |T |1/5 for all x ∈ T \ {1}, which is false for every non-Abelian finite simple
group T . Indeed, it is well known that every non-Abelian finite simple group T contains an involution
x with |CT (x)| > |T |1/3 (in fact, every involution in T has this property [21, Proposition 2.4]). For
r ∈ {1, 2, 3}, we verify the following result in Section 3. Although this result follows from routine
calculations, we include it here in case it proves to be a convenient reference.
Proposition 1.6. Let r ∈ {1, 2, 3} and let T be a non-Abelian finite simple group. Then either
|CT (x)| > |T |1−r/5 for some x ∈ T \ {1}, or T is one of the groups listed in Table 2.
Our new results about generalised quadrangles with point-primitive collineation groups are proved
in Sections 4–6. Corollary 1.5 is applied not only to actions of type HS or HC as illustrated above, but
also to types SD, CD and PA. In particular, the proof of Theorem 1.1 is free from the Classification
of Finite Simple Groups (CFSG) to the extent that, for G of type HC, CD or PA with socle T r × T r ,
T (k−1)r or T r respectively, the proof that r 6 4 depends only on Corollary 1.5. The CFSG is, however,
needed to prove Proposition 1.6 and some of the results in Section 5. For type PA, the group H in
Corollary 1.5 is an almost simple primitive group, so we are led to consider lower bounds on the fixity
of such a group, namely the maximum number of fixed points of a non-identity element. In Section 6,
we discuss how refinements of a recent result of Liebeck and Shalev [21, Theorem 4] on this problem,
currently being carried out by Elisa Covato at the University of Bristol as part of her PhD research
[11], can be adapted to further improve the bound r 6 4 in this case. In particular, for r ∈ {3, 4}
we are able to show that T cannot be a sporadic simple group, and to rule out the case T ∼
= Altn
except in one specific action when n is a prime congruent to 3 modulo 4 (see Table 1). The proof of
Theorem 1.1 is presented in Section 7.
Section 8 concludes the paper with a discussion and some open problems. In light of the growing
body of work towards a classification of point-primitive generalised quadrangles, and the possible
avenues outlined in Remark 5.11, Remark 6.4 and Section 8 for attacking the cases left open by
Theorem 1.1, we feel that the following conjecture can be made with a reasonable amount of confidence.
Conjecture 1.7. If a thick finite generalised quadrangle Q admits a collineation group G that
acts primitively on the point set of Q, then G is either affine or almost simple.
2. Bounding the number of points fixed by a collineation
The facts summarised in the following lemma are well known. (The existence of an order is proved
in [27, Corollary 1.5.3], and proofs of assertions (i)–(iii) may be found in [22, Section 1.2].)
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
5
Lemma 2.1. Let Q be a thick finite generalised quadrangle. Then Q has an order (s, t), and the
following properties hold:
(i) Q has (s + 1)(st + 1) points and (t + 1)(st + 1) lines,
(ii) s1/2 6 t 6 s2 6 t4 (Higman’s inequality),
(iii) s + t divides st(st + 1).
A point–line incidence geometry S = (P, L, I ) is called a grid if there exist positive integers
s1 and s2 such that P can be written as {Pij | 0 6 i 6 s1 , 0 6 j 6 s2 }, L can be written as
{ℓk | 0 6 k 6 s1 } ∪ {ℓ′k | 0 6 k 6 s2 }, and we have Pij I ℓk if and only if i = k, and Pij I ℓ′k if and
only if j = k. Each point of S is then incident with exactly two lines, and |P| = (s1 + 1)(s2 + 1).
Let us say that such a grid has parameters s1 and s2 . Note that a grid with parameters s1 = s2 is
a generalised quadrangle of order (s1 , 1). A dual grid is defined analogously, by swapping the roles
of points and lines. That is, there exist positive integers t1 and t2 such that L can be written as
{ℓij | 0 6 i 6 t1 , 0 6 j 6 t2 }, P can be written as {Pi | 0 6 i 6 t1 } ∪ {Pj′ | 0 6 j 6 t2 }, Pk I ℓij if
and only if i = k, and Pk′ I ℓij if and only if j = k. In this case, each line is incident with exactly two
points, and |P| = (t1 + 1) + (t2 + 1). Let us say that such a dual grid has parameters t1 and t2 .
If θ is a collineation of a generalised quadrangle Q = (P, L), then it makes sense to consider the
point–line incidence geometry Qθ = (Pθ , Lθ ) with Pθ = {P ∈ P | P θ = P }, Lθ = {ℓ ∈ L | ℓθ = ℓ}, and
incidence inherited from Q. Here we call Qθ the substructure of Q fixed by θ. It may happen that Qθ
is a grid or a dual grid, or a generalised quadrangle. More specifically, we have the following result,
based on the description of the possible structures of Qθ given by Payne and Thas [22, 2.4.1].
Lemma 2.2. Let Q = (P, L) be a thick finite generalised quadrangle of order (s, t). Let θ be a
non-identity collineation of Q, and let Qθ = (Pθ , Lθ ) be the substructure of Q fixed by θ. Then at least
one of the following conditions holds.
(i) Pθ is empty.
(ii) Lθ is empty and Pθ is a set of pairwise non-collinear points. In particular, |Pθ | 6 st + 1.
(iii) All points of Qθ are incident with a common line, and |Pθ | 6 s + 1.
(iv) All points of Qθ are collinear with a common point, and |Pθ | 6 s(t + 1) + 1.
(v) Qθ is a grid. In this case, either |Pθ | = (s + 1)2 and s 6 t, or |Pθ | < s2 .
(vi) Qθ is a dual grid, and |Pθ | 6 2(t + 1).
(vii) Qθ is a thick generalised quadrangle, and |Pθ | 6 (s + 1)(t + 1).
In particular, either |Pθ | 6 (s + 1)(t + 1); or s > t + 3, Qθ is a grid and |Pθ | < s2 .
Proof. The possible structures (i)–(vii) of Qθ are given by [22, 2.4.1]. We verify the claimed upper
bounds for |Pθ |. The bounds in cases (iii) and (iv) are immediate, because every line of Q is incident
with exactly s + 1 points, and every point of Q is incident with exactly t + 1 lines. For case (ii), note
[22, Section 2.7] that the maximum size of a set of pairwise non-collinear points in Q is st + 1. For
(v), if Qθ is a dual grid with parameters t1 and t2 , then t1 6 t and t2 6 t, and hence |Pθ | 6 2(t + 1).
Now suppose that Qθ is a grid with parameters s1 and s2 , noting that s1 6 s and s2 6 s, and
assuming (without loss of generality) that s1 > s2 . If s1 = s2 = s then |Pθ | = (s + 1)2 , and Qθ
is a generalised quadrangle of order (s, 1), so [22, 2.2.2(i)] implies that s 6 t. The case s2 = s − 1
cannot occur, because if θ fixes s points incident with a line then it must also fix the final point; and
if s2 6 s − 2 then |Pθ | 6 (s + 1)(s − 1) < s2 . Finally, suppose that Q is a thick finite generalised
quadrangle, and let (s′ , t′ ) denote its order. Then |Pθ | = (s′ + 1)(s′ t′ + 1) by Lemma 2.1(i). If t′ = t
then s′ < s because θ 6= 1, so [22, 2.2.1] implies that s′ t = s′ t′ 6 s, and hence |Pθ | 6 (s/t + 1)(s + 1) 6
(t2 /t + 1)(s + 1) = (s + 1)(t + 1), where for the second inequality we use Lemma 2.1(ii). If t′ < t then
the dual statement of [22, 2.2.1] yields s′ t′ 6 t, so |Pθ | = (s′ + 1)(s′ t′ + 1) 6 (s + 1)(t + 1).
The final assertion is deduced by comparing the upper bounds on |Pθ | established in each case.
We observe that |Pθ | 6 (s + 1)(t + 1) except possibly in the second case of (v), where our bound is
|Pθ | < s2 . However, if s 6 t + 2 then in this case we have |Pθ | < s2 < (s + 1)(t + 1).
Remark 2.3. We mention a paper of Frohardt and Magaard [17, Section 1.3], in which results
analogous to Lemma 2.2 are obtained for generalised d-gons with d ∈ {6, 8} (that is, generalised
hexagons and octagons). The known examples of such geometries admit point- and line-primitive
6
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
actions of almost simple groups with socle 3 D4 (q) or G2 (q) (for d = 6) and 2 F4 (q) (for d = 8). Frohardt
and Magaard use the aforementioned results to determine upper bounds for fixities of primitive actions
of groups G with generalised Fitting subgroup 3 D4 (q), G2 (q) or 2 F4 (q) (and they also treat the other
exceptional Lie type groups of Lie rank 1 or 2). By comparison, we instead apply Lemma 2.2 to
determine which groups might act primitively on the points of a generalised d-gon (with d = 4 in our
case). (We remark that we have also investigated point-primitive generalised hexagons and octagons,
although via different methods than in the present paper [7, 18].)
We now use Lemma 2.2 to prove Theorem 1.3, from which we deduce Theorem 1.2.
Proof of Theorem 1.3 Let (s, t) be the order of Q, and let Qθ = (Pθ , Lθ ) be the substructure of Q
fixed by θ. We must show that either |Pθ | < |P|4/5 , or (s, t) = (2, 4) and |Pθ | = 15 for Q = Q− (5, 2).
First suppose that s 6= 2. By Lemma 2.2, we have either |Pθ | < s2 or |Pθ | 6 (s + 1)(t + 1). If
|Pθ | < s2 then |Pθ | < |P|4/5 since |P| = (s+1)(st+1) > s2 t > s5/2 by Lemma 2.1. If |Pθ | 6 (s+1)(t+1)
then it suffices to show that the function f (s, t) = ((s + 1)(st + 1))4/5 − (s + 1)(t + 1) is positive for all
s > 3, for all s1/2 6 t 6 s2 . This is readily checked when s ∈ {3, 4}, so assume that s > 5. Thinking
of s and t as real variables, we have
(s + 1)(4s − h(s, t))
∂f
(s, t) =
,
∂t
h(s, t)
where h(s, t) = 5((s + 1)(st + 1))1/5 .
Since s and t are positive, this derivative is positive if and only if 4s − h(s, t) > 0. Since s > 5 and 2 6
11
1/5 s4/5 . Hence, 4s − h(s, t) > s4/5 (4s1/5 − 5( 33 )1/5 ).
t 6 s2 , we have h(s, t) 6 5( 65 s)1/5 ( 10
st)1/5 6 5( 33
25 )
25
4125
The right-hand side of this inequality is positive if s > ( 45 )5 ( 33
25 ) = 1024 ≈ 4.028, and so certainly
∂f
1/2 6 t 6 s2 . Since f (s, t) > f (s, s1/2 ) and f (s, s1/2 ) > 0 for s > 5, it
∂t (s, t) > 0 when s > 5 and s
follows that f (s, t) > 0 for all s > 5, for all s1/2 6 t 6 s2 .
Now suppose that s = 2. Then t ∈ {2, 4} by Lemma 2.1. There exist unique generalised quadrangles of orders (2, 2) and (2, 4), namely the symplectic space W(3, 2) and the elliptic quadric Q− (5, 2),
respectively [22, 5.2.3 and 5.3.2]. The full collineation groups of these generalised quadrangles are
PΓSp4 (2) and PΓU4 (2), respectively. One may use the package FinInG [4] in the computer algebra
system GAP [14] to check that every non-identity collineation of W(3, 2) fixes at most 7 points. Since
W(3, 2) has a total of 15 points and 154/5 ≈ 8.73 > 7, the claimed inequality |Pθ | < |P|4/5 holds for every non-identity collineation θ in this case. On the other hand, there exist 36 non-identity collineations
of Q− (5, 2) that fix 15 points, but the total number of points of Q− (5, 2) is 27 and 274/5 ≈ 13.97 < 15.
We also remark that the substructure fixed by such a collineation is, in fact, a generalised quadrangle
of order (2, 2). Every other non-identity collineation of Q− (5, 2) fixes at most 9 points.
Proof of Theorem 1.2. Since the action of H1 on Ω1 is not semiregular, there exists h1 ∈ H1 \ {1}
fixing at least one point of Ω1 . Let f1 be the number of points of Ω1 fixed by h1 . Let θ = (h1 , 1, . . . , 1) ∈
N , and let f be the number of points of Q fixed by θ. If Q
r = 1 then Theorem 1.3Qimplies that
f1 = f < |P|4/5 = |Ω1 |4/5 . If r > 2 then Theorem 1.3 gives f1 ri=2 |Ωi | = f < |P|4/5 = ( ri=1 |Ωi |)4/5 ,
Q
Q
so f1 ( ri=2 |Ωi |)1/5 < |Ω1 |4/5 . Since ( ri=2 |Ωi |)1/5 > |Ω1 |(r−1)/5 , it follows that f1 < |Ω1 |1−r/5 . In
particular, 1 − r/5 > 0 because f1 > 1, and so r 6 4.
We also use Lemma 2.2 to sharpen the 4/5 exponent bound in Theorem 1.3 in some special cases.
The proofs are just modifications of the proof of Theorem 1.3, but since the details are somewhat
tedious to check, we include them in Appendix A to save the reader having to reproduce them. We
also remark that the exponent 94/125 = 0.752 in case (i) of Proposition 2.5 could be changed to 3/4+ǫ
for any ǫ > 0 at the expense of increasing the upper bound on s in case (ii), but that this would not
have been useful for our arguments in Section 5.
Proposition 2.4. Let Q = (P, L) be a finite generalised quadrangle of order (s, t), let θ be any
non-identity collineation of Q, and let Qθ = (Pθ , Lθ ) be the substructure of Q fixed by θ. Then either
(i) |Pθ | < |P|7/9 ,
(ii) s ∈ {2, 3}, t = s2 and Qθ is a generalised quadrangle of order (s, s), or
(iii) s > t + 3, Qθ is a grid and |Pθ | < s2 .
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
7
Proposition 2.5. Let Q = (P, L) be a finite generalised quadrangle of order (s, t), let θ be any
non-identity collineation of Q, and let Qθ = (Pθ , Lθ ) be the substructure of Q fixed by θ. Then either
(i) |Pθ | < |P|94/125 ,
(ii) s < 2.9701 × 1015 , or
(iii) s > t + 3, Qθ is a grid and |Pθ | < s2 .
Proposition 2.6. Let Q = (P, L) be a finite generalised quadrangle of order (s, t), let θ be any
non-identity collineation of Q, and let Pθ denote the set of points fixed by θ. Suppose that t = s + 2.
Then |Pθ | < |P|7/9 if s > 3, and |Pθ | < |P|13/18 if s > 5.
3. Centraliser orders in non-Abelian finite simple groups
Here we verify a series of lemmas about centraliser orders in non-Abelian finite simple groups,
from which Proposition 1.6 is deduced. Specifically, we need to know which non-Abelian finite simple
groups T contain non-identity elements x with ‘large’ centralisers, in the sense that |CT (x)| > |T |1−r/5
for r equal to one of 1, 2 or 3. This question is readily and exactly answered for alternating groups
and sporadic simple groups in the following two lemmas. Note that we treat the Tits group 2 F4 (2)′
in Lemma 3.2 along with the sporadic groups.
Lemma 3.1. Let T ∼
= Altn with n > 5. Then
4/5
(i) |CT (x)| < |T |
for all x ∈ T \ {1} if and only if n 6 18,
(ii) |CT (x)| < |T |3/5 for all x ∈ T \ {1} if and only if n 6 9,
(iii) |CT (x)| < |T |2/5 for all x ∈ T \ {1} if and only if n 6 6.
Proof. If n > 19 and x ∈ T is a 3-cycle, then |CT (x)| = 23 (n − 3)! > ( 12 n!)4/5 = |T |4/5 . The remaining
assertions are readily verified using GAP [14].
Lemma 3.2. Let T be either
(i) |CT (x)| < |T |4/5 for all
(ii) |CT (x)| < |T |3/5 for all
(iii) |CT (x)| < |T |2/5 for all
a sporadic finite simple group or the Tits group 2 F4 (2)′ . Then
x ∈ T \ {1},
x ∈ T \ {1} if and only if T ∼
6 Suz, Co2 , Fi22 , Fi23 , B or M,
=
x ∈ T \ {1} if and only if T ∼
= J1 .
Proof. This is readily verified upon checking maximal centraliser orders in the ATLAS [10].
Next we consider the exceptional Lie type groups, namely those of type E8 , E7 , Eε6 (where ε = ±),
F4 , 2 F4 , G2 , 2 G2 , 3 D4 or 2 B2 . Note that we make no attempt to check the converse of assertion (i)
(although this could be done using standard references including those cited here).
Lemma 3.3. Let T be a finite simple group of exceptional Lie type.
(i) If T has type E8 , E7 , Eε6 , F4 or 3 D4 , then there exists x ∈ T \ {1} with |CT (x)| > |T |3/5 .
(ii) |CT (x)| < |T |2/5 for all x ∈ T \ {1} if and only if T has type 2 B2 .
Proof. (i) For T ∼
= E8 (q), F4 (q) or 3 D4 (q), take x ∈ T to be a unipotent element of type A1 in the
sense of [20, Tables 22.2.1 and 22.2.4] and [23], respectively. Then |CT (x)| = q 57 |E7 (q)|, q 15 |C3 (q)|
or q 12 (q 6 − 1), respectively, and it is readily checked that |CT (x)| > |T |3/5 in each case. Now suppose
that T ∼
= E7 (q) or Eε6 (q), and write G := Inndiag(T ). Take x ∈ T to be a unipotent element of type
A1 in the sense of [20, Tables 22.2.2 and 22.2.3], respectively. Then xT = xG by [9, Corollary 17.10],
so |CT (x)| = |CG (x)|/|G : T | = q 33 |D6 (q)|/ gcd(2, q − 1) or q 21 |Aε5 (q)|/ gcd(3, q − ε), respectively, and
again one can check that |CT (x)| > |T |3/5 in each case.
(ii) If T ∼
= 2 B2 (q) then |CT (x)| 6 q 2 < (q 2 (q 2 + 1)(q − 1))2/5 = |T |2/5 for all x ∈ T \ {1} [24]. It
remains to check that |CT (x)| > |T |2/5 for some x ∈ T \ {1} when T has type 2 F4 , G2 or 2 G2 . In
these respective cases, take x to be a unipotent element of type (Ã1 )2 , A1 or (Ã1 )3 in the sense of [20,
Tables 22.2.5–22.2.7], so that |CT (x)| = q 10 |2 B2 (q)|, q 5 |A1 (q)| or q 3 .
Finally, we consider the finite simple classical groups. Again, we do not check the converses of
assertions (i) or (ii), remarking only that one could do so using the monograph [9] of Burness and
Giuidici, where the conjugacy classes of elements of prime order in these groups are classified.
8
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
Lemma 3.4. Let T be a finite simple classical group.
(i) If T has type Aεn , Dn or 2 Dn with n > 9, or type Bn or Cn with n > 5, then there exists
x ∈ T \ {1} with |CT (x)| > |T |4/5 .
(ii) If T has type Aεn with n > 4, type Bn or Cn with n > 3, or type Dn or 2 Dn with n > 4, then
there exists x ∈ T \ {1} with |CT (x)| > |T |3/5 .
(iii) |CT (x)| < |T |2/5 for all x ∈ T \ {1} if and only if T ∼
= PSL2 (q) with q 6= 7.
Proof. Throughout the proof, we write q = pf with p a prime and f > 1. First suppose that T has
type A1 . That is, T ∼
= PSL2 (q), with q > 4. The smallest non-trivial conjugacy class of T has size
1 2
q(q − 1), 2 (q − 1) or 12 q(q − 1) according as whether p = 2, q ≡ 1 (mod 4) or q ≡ 3 (mod 4). Hence,
every non-trivial conjugacy class of T has size greater than |T |3/5 if and only if q 6= 7. Equivalently,
|CT (x)| < |T |2/5 for all x ∈ T \ {1} if and only if q 6= 7.
Now suppose that T has type Aεn with n > 2. That is, T ∼
= PSLεn+1 (q) (where L+ := L, L− := U).
ε
Let x ∈ G := PGLn+1 (q) be an element of order p with one Jordan block of size 2 and n − 1 Jordan
blocks of size 1. That is, a1 = n − 1, a2 = 1 and a3 = · · · = ap = 0 in the notation of [9, Section 3.2.3].
Then x ∈ T , and xT = xG by [9, Propositions 3.2.7 and 3.3.10], so by [9, Tables B.3 and B.4] we have
|CT (x)| = |CG (x)|/|G : T | = d1 |CG (x)| = 1d q 2n−1 |GLεn−1 (q)|, where d := gcd(n + 1, q − ε). Therefore,
(1)
n−1
n+1
1 n(n+1)/2 Y i
1 n(n+1)/2 Y i
i
(q − ε ) and |T | = q
(q − εi ).
|CT (x)| = q
d
d
i=1
i=2
For n ∈ {2, 3} we must show that |CT (x)| >
If n = 2 then d 6 3, so |CT (x)| > 13 q 3 (q − ε)
while |T | 6 q 3 (q 2 − ε2 )(q 3 − ε3 ). This implies that |CT (x)| > |T |2/5 for all q > 7, and one may check
directly that this inequality also holds for q < 7. If n = 3 then d 6 4, so |CT (x)| > 41 q 6 (q − ε)(q 2 − ε2 )
while |T | 6 q 6 (q 2 − ε2 )(q 3 − ε3 )(q 4 − ε4 ). This implies that |CT (x)| > |T |2/5 for all q > 3, and a direct
calculation shows that this inequality also holds for q = 2. Now suppose that 4 6 n 6 8. We must
show that |CT (x)| > |T |3/5 . Since q > 2, (1) gives
2
1 3 n n2 +2n
1 qn
q
,
and
|T
|
6
|CT (x)| >
d 2n−1
d 2
|T |2/5 .
2
and so it suffices to show that q 2n −6n > d2 22n−5 33n . Indeed, since d 6 n + 1, it suffices to show that
2
q 2n −6n > (n + 1)2 22n−5 33n . This inequality holds for all q > 2 if n ∈ {7, 8}, for all q > 3 if n = 6,
for all q > 4 if n = 5, and for all q > 11 if n = 4. In the remaining cases, where (n, q) = (6, 2), (5, 2),
(5, 3) or (4, q) with q < 11, one may check directly that |CT (x)| > |T |3/5 . It remains to show that
|CT (x)| > |T |4/5 for all q > 2 when n > 9. If q > 3 then (1) gives
1 4 n n2 +2n
1 2 n−1 n2
and |T | 6
q
q
,
|CT (x)| >
d 3
d 3
2
so it suffices to show that q n −8n > d · 23n+5 3n−5 . Indeed, since d 6 n + 1, we can just show that
2
q n −8n > (n + 1)23n+5 3n−5 . This inequality holds for all q > 3 if n > 11; if n = 10, it holds for all
q > 5, and if n = 9, it holds for all q > 29. For n = 10 with 2 6 q < 5 and n = 9 with 2 6 q < 29, one
may check directly that |CT (x)| > |T |4/5 . Finally, we must check that |CT (x)| > |T |4/5 when q = 2
255 i
i
and n > 9. Since n > 9, and since 256
q 6 q i − ε 6 257
256 q for i > 8, (1) gives
7
|CT (x)| >
1 255 n−8 n2 −28 Y i
(q − 1) and
q
d 256
i=1
7
|T | 6
1 257 n−8 n2 +2n−27 Y i
(q − 1).
q
d 256
i=2
Noting also that d 6 3, we see that it suffices to show that
2n
2 −8n−32
2555n−40
7
Y
i=2
(q i − 1) > 3 · 256n−8 2574n−32 .
This inequality holds for all n > 9, and so the proof of the Aεn case is complete.
Next, suppose that T has type Cn , where n > 2. That is, T ∼
= PSp2n (q). Write G := PGSp2n (q),
noting that |G : T | = gcd(2, q − 1). If p > 2, take x ∈ G of order p with one Jordan block of size 2 and
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
9
2(n − 1) Jordan blocks of size 1. That is, a1 = 2(n − 1), a2 = 1 and a3 = · · · = ap = 0 in the notation of
[9, Section 3.4.3]. Then x ∈ T , and by [9, Proposition 3.4.12], xG splits into two T -conjugacy classes
and hence |CT (x)| = 2|CG (x)|/|G : T | = 12 q 2n−1 |Sp2(n−1) (q)|. If p = 2 then T = G and we take x
to be an involution of type b1 as in [9, Table 3.4.1], so that |CT (x)| = q 2n−1 |Sp2(n−1) (q)|. Hence, for
every p, we have
n−1
1 n2 Y 2i
(q − 1) and
|CT (x)| = q
d
(2)
i=1
n
1 2 Y 2i
|T | = q n
(q − 1),
d
i=1
1 4 2
2 q (q
− 1) and |T | 6 q 4 (q 2 − 1)(q 4 − 1),
where d = gcd(2, q − 1) 6 2. If n = 2 then |CT (x)| >
and it follows that |CT (x)| > |T |2/5 for all q > 2. Similarly, for n ∈ {3, 4} one may check that
|CT (x)| > |T |3/5 for all q > 2. Now suppose that n > 5. Since q 2i > 4, we have q 2i − 1 > 34 q 2i for all
2
2
i > 1, and so |CT (x)| > 21 ( 43 )n−1 q 2n −n , while |T | < q 2n +n . Hence, to show that |CT (x)| > |T |4/5 , it
2
suffices to show that ( 34 )5n−5 q 2n −9n > 2. This inequality holds for all q > 2 when n > 6, and for all
q > 4 when n = 5; for (n, q) = (5, 2) and (5, 3), one may check directly that |CT (x)| > |T |4/5 .
Now suppose that T has type Bn , where n > 2. That is, T ∼
= Ω2n+1 (q) with q odd. For q ≡ 1 or 3
(mod 4), let x ∈ G := PGO2n+1 (q) be an involution of type tn or t′n , respectively, in the sense of [9,
Sections 3.5.2.1 and 3.5.2.2]. Then x ∈ T and xT = xG , so |CT (x)| = |CG (x)|/|G : T | = 21 |CG (x)| =
|SO±
2n (q)| by [9, Table B.8]. Now,
n
|SO±
2n (q)| = q
(3)
2 −n
(q n ∓ 1)
n−1
Y
i=1
(q 2i − 1) >
n−1
1 n2 Y 2i
(q − 1),
q
2
i=1
and the right-hand side above is the value of |CT (x)| that we obtained in the Cn case. Since
|Ω2n+1 (q)| = |PSp2n (q)|, we therefore reach the same conclusions as for type Cn .
Now suppose that T has type Dεn , namely T ∼
= PΩε2n (q) with n > 4. Let G := Inndiag(PΩε2n (q)),
as defined on [9, p. 56]. Assume first that p > 2, noting that |G : T | divides 4. Take x ∈ G of order p
with one Jordan block of size 2(n − 2) and two Jordan blocks of size 2. That is, a1 = 2(n − 2), a2 = 2
and a3 = · · · = ap = 0 in the notation of [9, Section 3.5.3]. Then x ∈ T , and [9, Propositions 3.5.14(i)
and (ii,b)] imply that xT = xG . Therefore, [9, Table B.12] gives |CT (x)| = |CG (x)|/|G : T | >
1 4n−7
1
1
|Oε2(n−2)
(q)||Sp2 (q)|, where the value of ε1 = ± depends on n and q as described
4 |CG (x)| = 8 q
1
there. Multiplying the inequality in (3) by 2 to get a lower bound for |Oε2(n−2)
(q)|, it follows that
n−3
(4)
Y
1 2
|CT (x)| > q n −2 (q 2 − 1)
(q 2i − 1),
8
i=1
where d =
(5)
gcd(4, q n
− ε). Since
q 2i
n−1
Y
1
while |T | = q n(n−1) (q n − ε)
(q 2i − 1),
d
i=1
qn
34
> 9 for all i > 1, and in particular
> = 81, we have
82 2
1 8 n−2 2n2 −5n+6
q
and |T | < q 2n −n .
|CT (x)| >
8 9
81
2
For 4 6 n 6 8 we need |CT (x)| > |T |3/5 , so by (5) it suffices to show that ( 98 )5n−10 q 4n −22n+30 >
82 3
) , which holds unless (n, q) = (4, 3) or (4, 5). For (n, q) = (4, 5), (4) implies that |CT (x)| > |T |3/5 ;
85 ( 81
for (n, q) = (4, 3), a GAP [14] calculation shows that there exist elements x ∈ T \ {1} for which this
inequality holds. For n > 9 we claim that |CT (x)| > |T |4/5 , and we now have q n > 39 = 19683, so we
82
8 5n−10 2n2 −21n+30
19684 4
can replace the 81
in (5) by 19684
q
> 85 ( 19683
) .
19683 to see that it suffices to show that ( 9 )
If n > 10 then this inequality holds for all q > 3, and if n = 9 then it holds for q > 127. For n = 9 and
Qn−3 2i
2
(q − 1), which
q < 127, using the equality in (3) we obtain |CT (x)| > 14 q n −n (q n−2 − 1)(q 2 − 1) i=1
4/5
implies that |CT (x)| > |T |
except when q = 3 and ε = +. However, in this case we have |G : T | = 2
(compare [9, Figure 2.5.1 and Lemma 2.2.9], noting that the discriminant of a hyperbolic quadratic
form on F2n
q with (n, q) = (9, 3) is ⊠, in the notation used there, because n(q − 1)/4 = 9 is odd), so
1
the 4 in the above estimate for |CT (x)| may be replaced by 21 , and we again obtain |CT (x)| > |T |4/5 .
Finally, suppose that T ∼
= PΩε2n (q) with q even, noting that T = G in this case. Take x ∈ G
to be an involution of type a2 as in [9, Table 3.5.1]. Then |CT (x)| = q 4n−7 |Ωε2(n−2) (q)||Sp2 (q)| and
10
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
gcd(4, q n − ε) = 1, so instead of (4) we have
n−3
(6)
Y
1 2
|CT (x)| > q n −2 (q 2 − 1)
(q 2i − 1)
4
i=1
and |T | = q
n(n−1)
n
(q − ε)
n−1
Y
i=1
(q 2i − 1).
1
4
(In the bound for |CT (x)| we drop a factor of because |G : T | = 1 for q even, but pick up a factor
of 21 because Ωε2(n−2) (q) has index 2 in SOε2(n−2) (q).) For 4 6 n 6 8 we need |CT (x)| > |T |3/5 . Since
2
3
q i > 2 for all i > 1, and in particular q n > 16, it suffices to show that ( 34 )5n−10 q 4n −22n+30 > 45 ( 17
16 ) .
This inequality holds unless (n, q) = (4, 2) or (4, 4), in which cases a direct calculation shows that
|CT (x)| > |T |3/5 . For n > 9 we must show that |CT (x)| > |T |4/5 . We now have q n > 512, and so
2
4
it suffices to show that ( 43 )5n−10 q 2n −21n+30 > 45 ( 513
512 ) . This inequality holds unless (n, q) = (10, 2)
or n = 9 and q 6 28 . One may use (6) to check that |CT (x)| > |T |4/5 in each of these cases except
(n, q) = (9, 2), in which case the desired inequality may be verified by a direct calculation.
4. Quasiprimitive point actions of type SD or CD
We now apply Corollary 1.5 to permutation groups N that arise as subgroups of certain types
of primitive groups. In some cases, we are also able to treat quasiprimitive groups, namely those in
which every non-trivial normal subgroup is transitive. In this section, we consider the case where the
group N in Corollary 1.5 has a ‘diagonal’ action. Specifically, we work under the following hypothesis.
Hypothesis 4.1. Let T be a non-Abelian finite simple group, let k > 2, and write H = T k . Let
Ω = {(y1 , . . . , yk−1 , 1) | y1 , . . . , yk−1 ∈ T } 6 H, and let H act on Ω by
(7)
−1
(y1 , . . . , yk−1 , 1)(x1 ,...,xk ) = (x−1
k y1 x1 , . . . , xk yk−1 xk−1 , 1).
Suppose that N = H r is a collineation group of a thick finite generalised quadrangle Q = (P, L, I ) of
order (s, t), such that P = Ωr and N has the product action on P.
This situation arises when N is the socle of a primitive permutation group G 6 Sym(Ω) of type
HS, HC, SD or CD. For type HS (respectively HC) we have k = 2 and r = 1 (respectively r > 2),
G has two minimal normal subgroups, each isomorphic to T r , and the socle of G is T r × T r , which
is isomorphic to N . For type SD (respectively CD) we have k > 2 and r = 1 (respectively r > 2),
and G has a unique minimal normal subgroup, isomorphic to T kr ∼
= N . Note that the notation k
and r is consistent with that of Table 1. Of course, G must (usually) satisfy certain other conditions
[19, Section 6] in order to actually be primitive, but these conditions are not needed for the proof of
Proposition 4.2. It suffices that there is a subgroup of the form N . In particular, we are also able
to treat quasiprimitive groups [19, Section 12], because the (action of the) socle of G is the same as
in the respective primitive types. (Note that a quasiprimitive group of type HS or HC is necessarily
primitive, but a quasiprimitive group of type SD or CD need not be primitive.)
Proposition 4.2 shows, in particular, that the parameter r in Hypothesis 4.1 can be at most 3.
As illustrated after Corollary 1.5, the proof relies on the information about centraliser orders in nonAbelian finite simple groups given in Proposition 1.6. We also observe that when r = 3, there always
exists a solution (s, t) = (|Ω| − 1, |Ω| + 1) of the equation |Ω|3 = |Ω|r = |P| = (s + 1)(st + 1), and this
solution satisfies properties (ii) and (iii) of Lemma 2.1. Hence, although we are unable to rule out the
case r = 3 completely, we verify that this ‘obvious’ situation cannot occur.
Proposition 4.2. If Hypothesis 4.1 holds then r 6 3 and |CT (x)| < |T |1−r/5 for all x ∈ T \ {1},
and in particular T must appear in Table 2. Moreover, if r = 3 then (s, t) 6= (|Ω| − 1, |Ω| + 1).
Proof. Note first that |P| = |Ω|r = |T |(k−1)r . In particular, the excluded case (s, t) = (2, 4) in
Corollary 1.5 does not arise, because |P| > |T | > |Alt5 | = 60 > (2 + 1)(2 · 4 + 1). If we take x := x1 =
· · · = xk 6= 1 in (7), then (y1 , . . . , yk−1 , 1) ∈ Ω is fixed if and only if y1 , . . . , yk−1 ∈ CT (x). That is,
(x, . . . , x) ∈ H fixes precisely |CT (x)|k−1 elements of Ω (and, in particular, the action of H on Ω is not
semiregular). Corollary 1.5 therefore implies that r 6 4 and |CT (x)|k−1 < |Ω|1−r/5 = |T |(k−1)(1−r/5) ,
namely |CT (x)| < |T |1−r/5 , for all x ∈ T \ {1}. If r = 4 then we have a contradiction because
every non-Abelian finite simple group T contains a non-identity element x with |CT (x)| > |T |1/5 . For
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
11
example, it is well known that every non-Abelian finite simple group T contains an involution x with
|CT (x)| > |T |1/3 (in fact, every involution has this property [21, Proposition 2.4]). Therefore, r 6 3.
In particular, Proposition 1.6 tells us that T must be one of the groups appearing in Table 2. To
prove the final assertion, suppose towards a contradiction that r = 3 and (s, t) = (|Ω| − 1, |Ω| + 1).
Take any x ∈ T with |CT (x)| > |T |1/3 . Then ((x, . . . , x), (1, . . . , 1), (1, . . . , 1)) ∈ H r = H 3 = N fixes
|CT (x)|k−1 |T |2(k−1) > |T |7(k−1)/3 = |P|7/9 points of Q, contradicting Proposition 2.6.
The following immediate consequence of Proposition 4.2 (and the preceding observations) implies
the SD and CD cases of Theorem 1.1.
Proposition 4.3. Let Q = (P, L) be a thick finite generalised quadrangle admitting a collineation
group G that acts quasiprimitively of type SD or CD on P. Then the conditions in Table 1 hold.
5. Primitive point actions of type HS or HC
We now consider the case where k = 2 in Hypothesis 4.1 in more detail. As explained above, this
case arises when N is the socle of a primitive permutation group G 6 Sym(Ω) of type HS (r = 1) or
HC (r > 2). When k = 2 it is natural to simplify the notation of Hypothesis 4.1 by identifying the set
Ω with T r , so we first re-cast the hypothesis in this way and also establish some further notation.
(8)
Hypothesis 5.1. Let T be a non-Abelian finite simple group and let N = T r × T r act on T r by
y (u1 ,u2 ) = u−1
2 yu1 .
Let M = {(u, 1) | u ∈ T r } 6 N , so that M may be identified with T r acting regularly on itself by
right multiplication. Suppose that N is a collineation group of a thick finite generalised quadrangle
Q = (P, L, I ) of order (s, t) with P = T r . Let P1 ⊂ P denote the set of points collinear with but
not equal to the identity element 1 ∈ T r = P, and let L1 ⊂ L denote the set of lines incident with 1.
Given a line ℓ ∈ L, let ℓ̄ ⊂ P denote the set of points incident with ℓ.
The following lemma may essentially be deduced from [26, Lemma 10] upon observing that the
assumption gcd(s, t) > 1 imposed there is not necessary (as far as we can tell, and at least not in
our more restrictive setting). We include a proof to make it clear that we do not need to make this
assumption. Our notation differs from that of [26, p. 654] as follows: the point-regular group G is our
M∼
= T r , and the point O is our point 1, so that ∆ is our P1 \ {1}.
Lemma 5.2. Suppose that Hypothesis 5.1 holds. Let x ∈ P1 \ {1}, and let ℓx be the unique line in
r
L1 incident with x. Then, for every i ∈ {1, . . . , |x| − 1}, the conjugacy class (xi )T is contained in P1 .
Moreover, the collineation (x, 1) ∈ M fixes ℓx .
Proof. Let us first establish some notation. Given u ∈ T r = P, write
fixP (u) = {P ∈ P | P (u,1) = P },
collP (u) = {P ∈ P | P (u,1) is collinear with but not equal to P },
fixL (u) = {ℓ ∈ L | ℓ(u,1) = ℓ},
concL (u) = {ℓ ∈ L | ℓ(u,1) is concurrent with but not equal to ℓ}.
Since the subgroup M = {(u, 1) | u ∈ T r } of N acts regularly on P, fixP (u) is empty. Moreover,
−1
−1
P ∈ collP (u) if and only if P (P ,1) = 1 and (P u)(P ,1) = P uP −1 are collinear, which is if and only
r
if P uP −1 ∈ uT ∩ P1 . Since for g, h ∈ T r we have gug−1 = huh−1 if and only if g−1 h ∈ CT r (u), it
follows that
r
| collP (u)| = |uT ∩ P1 ||CT r (u)|,
as in the proof of [26, Lemma 3]. Then (again, as in that proof) [22, 1.9.2] implies that
(9)
r
| collP (u)| = (s + 1)| fixL (u)| + | concL (u)| = |uT ∩ P1 ||CT r (u)|
(for every u ∈ T r ).
Now, since x ∈ P1 , we have u−1 xu = x(u,u) ∈ P1 for every collineation of the form (u, u) ∈ N ,
because such collineations (are precisely those that) fix the point 1. That is, every T r -conjugate of x
r
r
is in P1 . In other words, xT ∩ P1 = xT , and so (9) implies that
(10)
r
| collP (x)| = (s + 1)| fixL (x)| + | concL (x)| = |xT ||CT r (x)| = |T r | = |P| = (s + 1)(st + 1).
12
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
In particular, we have collP (x) = P; that is, every point of Q is mapped to a collinear point under the
collineation (x, 1) ∈ M . We now claim that concL (x) is empty. If not, then some line ℓ is concurrent
with its image under the collineation (x, 1). Let P denote the unique point incident with both ℓ and
ℓ(x,1) . Then P x−1 is incident with ℓ, being the image of P under the collineation (x, 1)−1 = (x−1 , 1),
and P x−1 6= P because x 6= 1 and M acts regularly on P. Since Q is thick, there exists a third
(x,1)
point P3 incident with ℓ, distinct from P and P x−1 . Since collP (x) = P, the points P3
= P3 x and
P3 are collinear. Moreover, P3 x is collinear with P , because both of these points are incident with
ℓ(x,1) . Therefore, P3 x is collinear with two distinct points that are incident with ℓ, namely P3 and
P , and so P3 x is itself incident with ℓ because Q contains no triangles. This, however, means that
P3 x is incident with both ℓ and ℓ(x,1) , which forces P3 x = P and hence P3 = P x−1 , a contradiction.
Therefore, | concL (x)| = 0 as claimed, and so (10) implies that
(11)
r
| fixL (x)| = st + 1.
Next, we show that (xi )T ⊆ P1 for all i ∈ {1, . . . , |x| − 1}. For each such i, we certainly have
fixL (x) ⊆ fixL (xi ), because if the collineation (x, 1) fixes a line then so too does (x, 1)i = (xi , 1). In
particular, | fixL (xi )| > | fixL (x)| = st + 1, by (11). On the other hand, no two lines fixed by (xi , 1) can
be concurrent, because if they were, then the unique point incident with both lines would be fixed by
(xi , 1), a contradiction since M acts regularly on P. Hence, the total number of points that are incident
with some line in fixL (xi ) is (s + 1)| fixL (xi )|. As this number cannot exceed |P| = (s + 1)(st + 1), we
must also have | fixL (xi )| 6 st + 1. Therefore, | fixL (xi )| = st + 1. Now (9) implies, on the one hand,
that
| collP (xi )| = (s + 1)| fixL (xi )| + | concL (xi )| = |P| + | concL (xi )|.
Since | collP (xi )| 6 |P|, this implies that | concL (xi )| = 0, and then in turn that |P| = | collP (xi )|.
r
Appealing again to (9), we now deduce that |(xi )T ∩ P1 ||CT r (xi )| = |P| = |T r |, which implies that
r
(xi )T ⊆ P1 as required. The first assertion is therefore proved.
Finally, we must show that the collineation (x, 1) fixes the unique line ℓx ∈ L1 incident with x. If
|x| = 2, then (x, 1) fixes ℓx because it fixes setwise the subset {1, x} of points incident with ℓx . That
is, it maps 1 to 1(x,1) = 1x = x and x to x(x,1) = x2 = 1. Now suppose that |x| > 2. Then the point
2
1(x,1) = x2 6= 1 is collinear with x because x is collinear with 1. On the other hand, (x2 )Tr ⊆ P1 by the
first assertion, so in particular x2 is collinear with 1. Therefore, x2 is collinear with two distinct points
incident with ℓx (namely 1 and x), and so is itself incident with ℓx because Q contains no triangles.
Hence, (x, 1) fixes ℓx because it maps two points incident with ℓx , namely 1 and x, to another two
points incident with ℓx , namely x and x2 .
Hypothesis 5.1 imposes the following restrictions on the points and lines incident with the identity
element of T r = P, and on the order (s, t) of Q.
Lemma 5.3. The following assertions hold under Hypothesis 5.1.
(i) P1 is a union of T r -conjugacy classes.
(ii) Every line ℓ ∈ L1 has the property that ℓ̄ is a subgroup of T r . Specifically,
ℓ̄ = {u ∈ T r | (u, 1) ∈ M fixes ℓ}.
(iii) Every line ℓ ∈ L1 is incident with an involution.
(iv) If some line in L1 is incident with representatives of every T r -conjugacy class of involutions
in P1 , then N acts transitively on the flags of Q and r > 2.
(v) T r has at least three conjugacy classes of involutions.
(vi) If T r has exactly three conjugacy classes of involutions, then either P1 contains exactly two
of these classes, or N acts transitively on the flags of Q and r > 2.
(vii) gcd(s, t) = 1 and t > s + 1.
Proof. (i) This follows immediately from Lemma 5.2.
(ii) If u ∈ ℓ̄ then the collineation (u, 1) ∈ M fixes ℓ by Lemma 5.2. Conversely, if (u, 1) fixes ℓ
then, because 1 ∈ P is incident with ℓ, so too is 1(u,1) = u; that is, u ∈ ℓ̄.
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
13
(iii) If ℓ ∈ L1 is not incident with any involution, then ℓ̄, which is a subgroup of T r by (ii), must
have odd order. That is, s + 1 = |ℓ̄| must be odd. However, (s + 1)(st + 1) = |T |r is even by the
Feit–Thompson Theorem [13], so s must be odd and hence s + 1 must be even, a contradiction.
(iv) If ℓ ∈ L1 is incident with representatives of every conjugacy class of involutions in P1 , then ℓ
can be mapped to any other line in L1 by some element of the stabiliser N1 = {(u, u) | u ∈ T r } in N
of the point 1 ∈ P = T r . Since N acts transitively on P, this means that N acts transitively on the
flags of Q. If r = 1, this contradicts the main result of our earlier paper [8, Theorem 1.1], so r > 2.
(v) Suppose towards a contradiction that T r contains at most two conjugacy classes of involutions.
Then r = 1, because if r > 2 then any involution x ∈ T gives rise to the three pairwise non-conjugate
involutions (x, 1, . . . , 1), (1, x, 1, . . . , 1) and (x, x, 1, . . . , 1) in T r . Hence, by (iii), T must have exactly
two conjugacy classes of involutions, say xT and y T , and both must be contained in P1 . Without loss
of generality, x and y commute, because at least one of them centralises a Sylow 2-subgroup of T .
Therefore, xy is an involution, and so must be collinear with 1 ∈ P. Since 1 is collinear with x, the
images of 1 and x under the collineation (y, 1) ∈ M are collinear. That is, 1(y,1) = y is collinear with
x(y,1) = xy. Similarly, 1 and y are collinear, and hence so too are 1(x,1) = x and y (x,1) = yx = xy.
Since the involution xy is also collinear with 1 and Q contains no triangles, the points 1, x, y and xy
must be incident with a common line. In particular, x and y are incident with a common line in L1 .
Since r = 1, this contradicts (iv).
(vi) Let x, y and z denote representatives of the three T r -conjugacy classes of involutions. If P1
contains exactly one of these classes, then N acts flag-transitively by (iv), and it follows from [8,
Theorem 1.1] that r > 2. Now suppose that P1 contains all three of xT , y T and z T . Without loss of
generality, x centralises a Sylow 2-subgroup of T r and both y and z commute with x, so xy = yx and
xz = zx are involutions. Arguing as in the proof of (iii), we deduce that 1, x, y and xy are incident
with a common line ℓ ∈ L1 . Replacing y by z in this argument, we see that z is also incident with ℓ,
so (iv) again implies that N acts flag-transitively (and it follows as above that r > 2).
(vii) If gcd(s, t) > 1 then [26, Lemma 6(i)] implies that every T r -conjugacy class intersects P1 .
However, assertion (i) then implies that P1 = P, which is impossible. Therefore, gcd(s, t) = 1. In
particular, to show that t > s + 1 it suffices to show that t > s. The proof of this assertion is adapted
from that of [8, Corollary 2.3]. Choose two distinct lines ℓ1 , ℓ2 ∈ L1 , so that ℓ̄1 and ℓ̄2 are subgroups
of T r by (ii). For brevity, we now abuse notation slightly and identify ℓ1 and ℓ2 with ℓ̄1 and ℓ̄2 ,
respectively, dropping the ‘bar’ notation. Since ℓ1 is a subgroup of T r and right multiplication by any
element of T r is a collineation of Q (identified with an element of M ), we have in particular that every
right coset ℓ1 g2 of ℓ1 with g2 ∈ ℓ2 corresponds precisely to the set of points incident with some line
of Q. Similarly, left multiplications (identified with elements of {1} × T r 6 N ) are collineations, so
every left coset g1 ℓ2 of ℓ2 with g1 ∈ ℓ1 is a line of Q. Therefore, L′ = {g1 ℓ2 | g1 ∈ ℓ1 } ∪ {ℓ1 g2 | g2 ∈ ℓ2 }
is a subset of L. Consider also the subset P ′ = ℓ1 ℓ2 of P = T r , and let I ′ be the restriction of I
to (P ′ × L′ ) ∪ (L′ × P ′ ). If we can show that Q′ = (P ′ , L′ , I ′ ) is a generalised quadrangle of Q of
order (s, 1), then [22, 2.2.2(i)] will imply that t > s. Let us first check that Q′ satisfies the generalised
quadrangle axiom. Let ℓ ∈ L′ and take P ∈ P ′ not incident with ℓ. Then, since Q satisfies the
generalised quadrangle axiom, there is a unique point P0 ∈ P incident with ℓ and collinear with P .
Since ℓ ⊂ P ′ , we have P0 ∈ P ′ , and so Q′ also satisfies the generalised quadrangle axiom. It remains
to check that Q′ has order (s, 1). Every line in L′ is incident with s + 1 points in P ′ , being a coset of
either ℓ1 or ℓ2 , so it remains to show that every point in P ′ is incident with exactly two lines in L′ .
Given P = g1 g2 ∈ P ′ , where g1 ∈ ℓ1 , g2 ∈ ℓ2 , each line ℓ ∈ L′ incident with P is either of the form
h1 ℓ2 for some h1 ∈ ℓ1 or ℓ1 h2 for some h2 ∈ ℓ2 , and since P ∈ ℓ, we must have h1 = g1 or h2 = g2 ,
respectively. Therefore, P is incident with exactly two lines in L′ , namely g1 ℓ2 and ℓ1 g2 .
Proposition 4.2 restricts the possibilities for the simple group T in Hypothesis 5.1 to those listed
in Table 2. The following result shows that, furthermore, T must be a Lie type group.
Proposition 5.4. If Hypothesis 5.1 holds then T is a Lie type group.
Proof. We have |P| = |T |r , and r ∈ {1, 2, 3} by Proposition 4.2. For each of the alternating and
sporadic simple groups T in Table 2, we check computationally for solutions of |T |r = (s + 1)(st + 1)
satisfying s > 2, t > 2 and properties (ii) and (iii) of Lemma 2.1 (see Remark 5.5). If r = 3 then the
14
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
r
T
s
t
s(t + 1)
1
1
2
2
2
Alt7
Alt8
Alt6
M11
J1
11
19
19
89
419
19
53
341
7831
175141
220
1026
6498
697048
73384498
Table 3. Alternating and sporadic groups in the proof of Proposition 5.4.
only such solutions have the form (s, t) = (|T | − 1, |T | + 1) = (|Ω| − 1, |Ω| + 1), and this contradicts
the final assertion of Proposition 4.2. If r ∈ {1, 2} then the possibilities for T and (s, t) are as in
Table 3. By Lemma 5.3(i), P1 is a union of T r -conjugacy classes, and so we must be able to partition
|P1 | = s(t + 1) into a subset of the sizes of these classes (respecting multiplicities). When r = 1 and
T ∼
= Alt7 or Alt8 , this is impossible: the non-trivial conjugacy class sizes not exceeding s(t + 1) are 70,
105 and 210 in the first case, and 105, 112 and 210 in the second (with each occurring exactly once).
Similarly, if r = 2 and T ∼
= J1 , one may check computationally that there is no partition of s(t + 1),
where (s, t) = (419, 175141), into non-trivial T 2 -conjugacy class sizes. Hence, it remains to consider
the cases where r = 2 and T ∼
= Alt6 or M11 . Here we first determine computationally the possible
partitions of s(t + 1) into non-trivial T 2 -conjugacy class sizes to obtain a list of possible partitions P1
into T 2 -conjugacy classes. Now, because the point graph of Q is a strongly regular graph in which
adjacent vertices have λ := s − 1 common neighbours and non-adjacent vertices have µ := t + 1
common neighbours, P1 must be a partial difference set of T 2 with these parameters. That is, each
non-identity element y ∈ T 2 must have exactly λ representations of the form y = zi zj−1 for zi , zj ∈ P1
if y ∈ P1 , and exactly µ such representations if y 6∈ P1 . A computation verifies that this condition is
violated for each of the partitions of P1 determined in the previous step.
Remark 5.5. In the proof of Proposition 5.4, and at several other points in Sections 5 and 6,
we need to check computationally whether certain positive integers X can be equal to the number
of points of a thick finite generalised quadrangle. That is, we check for integral solutions (s, t) of
the equation (s + 1)(st + 1) = X subject to the constraints s > 2, t > 2, s1/2 6 t 6 s2 6 t4 and
s + t | st(st + 1) imposed by Lemma 2.1. In Section 5, X has the form |T |m for some non-Abelian
finite simple group T and some m 6 3, and in Section 6 we instead have X = Y m with m 6 4 and Y
the index of a maximal subgroup of an almost simple group. The above inequalities imply that s must
lie between X 1/4 − 1 and X 5/2 , so it suffices to consider every integer s in this range and determine
whether t = ((X − 1)/s − 1)/(s + 1) is an integer and, if so, whether s + t | st(st + 1). We remark
that we found it useful to also observe that s must divide X − 1, because it turned out that X − 1
had only a very small number of divisors in many of the cases that we had to consider.
We now show that r cannot equal 3, and deduce some further restrictions on T when r ∈ {1, 2}.
Proposition 5.6. If Hypothesis 5.1 holds then r 6 2 and T is a Lie type group with the property
that |CT (x)| < |T |1−2r/9 for all x ∈ T \ {1}.
Proof. By Propositions 4.2 and 5.4, T is a Lie type group and r 6 3. We now show that |CT (x)| <
|T |1−2r/9 for all x ∈ T \ {1} and deduce from this that r 6= 3. Suppose, towards a contradiction, that
there exists x ∈ T \ {1} with |CT (x)| > |T |1−2r/9 . Define w = (x, 1, . . . , 1) ∈ T r and let Qθ = (Pθ , Lθ )
be the substructure of Q fixed by θ = (w, w) ∈ N1 . Then Pθ = CT (x) × T r−1 , and hence
(12)
|Pθ | > |T |(1−2r/9)+(r−1) = |T |7r/9 = |P|7/9 .
Proposition 2.4 then says that either s > t + 3, or (s, t) ∈ {(2, 4), (3, 9)}. The first of these conditions
contradicts Lemma 5.3(vii); the second implies that |T |r = (s + 1)(st + 1) ∈ {27, 112}, which is
impossible because |T | > 60. Hence, every x ∈ T \ {1} must satisfy |CT (x)| < |T |1−2r/9 . For r = 3
this says that |CT (x)| < |T |1/3 for all x ∈ T \ {1}, a contradiction because we can always find some x
with |CT (x)| > |T |1/3 (as noted in the proof of Proposition 4.2). Therefore, r 6 2.
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
15
Proposition 5.6 allows us to further reduce the list of candidates for the simple group T in Hypothesis 5.1 in the remaining cases r ∈ {1, 2}. Let us first consider r = 2.
Corollary 5.7. If Hypothesis 5.1 holds with r = 2 then T is of Lie type A1 , Aε2 , 2 B2 or 2 G2 . In
particular, T has a unique conjugacy class of involutions.
Proof. The result is verified by straightforward calculations involving the bound on centraliser orders
imposed by Proposition 5.6, but we include the details in Appendix A.
We can now prove the HC case of Theorem 1.1.
Theorem 5.8. If Q is a thick finite generalised quadrangle with a collineation group G that acts
primitively on the point set P of Q, then the action of G on P does not have O’Nan–Scott type HC.
Proof. As explained above, the socle of G is a group N = T r ×T r as in Hypothesis 5.1, for some r > 2.
However, Corollary 5.7 tells us that r = 2 and that T has a unique conjugacy class of involutions.
In particular, T r = T 2 has exactly three conjugacy classes of involutions, with representatives (x, 1),
(1, y) and (x, y), where x and y are involutions in T . Now, [8, Theorem 1.1] says that G cannot act
transitively on the flags of Q, so in particular N cannot act transitively on the flags of Q. Lemma 5.3(vi)
therefore implies that P1 = T 2 must contain exactly two T 2 -conjugacy classes. Hence, without loss
2
of generality, P1 contains the class (x, 1)T . Since G acts primitively on P, it induces a subgroup of
Aut(T 2 ) = Aut(T ) ≀ Sym2 that swaps the two simple direct factors of T 2 . Therefore, P1 also contains
2
2
the class (1, y)T , and so does not contain the class (x, y)T . In particular, no line ℓ ∈ L1 can be
incident with both a conjugate of (x, 1) and a conjugate of (1, y), because by Lemma 5.3(ii), ℓ would
then also be incident with the product of these elements, a conjugate of (x, y). Hence, L1 is partitioned
into two sets of lines: those incident with conjugates of (x, 1), and those incident with conjugates of
(1, y). Since G1 swaps these sets, G acts flag-transitively, in contradiction with [8, Theorem 1.1].
For r = 1 we are left with the following list of candidates for T .
Corollary 5.9. Suppose that Hypothesis 5.1 holds with r = 1. Then T is of Lie type A1 , Aεn
with 2 6 n 6 6, B2 , C2 , C3 , Dεn with 4 6 n 6 6, or exceptional Lie type other than E8 .
Proof. By Propositions 4.2 and 5.4, T is one of the Lie type groups in the first column of Table 2.
By arguing as in the proof of Proposition 5.6 but applying Proposition 2.5 instead of Proposition 2.4
in the first paragraph, we conclude that one of the following conditions must also hold:
(i) every non-identity element x ∈ T satisfies |CT (x)| < |T |94/125 , or
(ii) s 6 2.9701 × 1015 .
By choosing appropriate elements x ∈ T as in the proofs in Section 3, we are able to use this to deduce
that T does not have type Aε7 , Aε8 , B4 , C4 , Dε7 , Dε8 or E8 . We rule out E8 here as an example, and include
details of the remaining cases in Appendix A. If T ∼
= E8 (q) then (s + 1)4 > |P| = |T | > |E8 (2)| ≈
3.378 × 1074 and hence s > |E8 (2)|1/4 − 1 ≈ 4.287 × 1018 , contradicting (ii), so (i) must hold. However,
as noted in the proof of Lemma 3.3, there exists x ∈ T ∼
= E8 (q) with |CT (x)| = q 57 |E7 (q)| ∼ q 190 ,
while |T |94/125 ∼ (q 248 )94/125 < q 187 . Indeed, one may check that |CT (x)| > |T |94/125 for all q > 2.
Finally, we use Lemma 5.3 to reduce the list of candidates for T in Corollary 5.9 to those given in
the first row of Table 1, thereby proving the HS case of Theorem 1.1.
Proposition 5.10. Let Q = (P, L) be a thick finite generalised quadrangle admitting a collineation
group G that acts primitively of type HS on P, with socle T × T for some non-Abelian finite simple
group T . Then T has Lie type Aε5 , Aε6 , B3 , C2 , C3 , Dε4 , Dε5 , Dε6 , Eε6 , E7 or F4 .
Proof. We are assuming that Hypothesis 5.1 holds with r = 1, so T must be one of the groups listed
in Corollary 5.9. It remains to show that, further, T cannot have Lie type A1 , Aε2 , Aε3 , Aε4 , 2 B2 , 2 G2 ,
2 F , G or 3 D . This follows from Lemma 5.3(v), because in each of these cases T has at most two
4
2
4
conjugacy classes of involutions. (This may be verified using, for example, [16, Table 4.5.1] for odd
characteristic and [20] for even characteristic.)
16
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
Remark 5.11. Proposition 5.10 begs the obvious question of whether we can rule out the last
remaining candidates for T listed there. We are confident that we will eventually be able to do this, but
it seems that it will require even more new ideas and a detailed case-by-case analysis. Of course, some
of the remaining groups can be ruled out in certain cases using Lemma 5.3(v); in particular, if T has
Lie type C2 , C3 , F4 or Eε6 in characteristic p, then we must have p = 2, because in odd characteristic
these groups have only two conjugacy classes of involutions. When T has exactly three conjugacy
classes of involutions, we can begin by applying Lemma 5.3(v) (because we know from [8] that N
cannot act transitively on the flags of Q), and then the arguments in the proof of Lemma 5.3(iv–
vi) can be extended to deduce some restrictions on which involutions can appear in P1 . However,
even with this extra information, we have thus far been unable to completely rule out any of the
remaining candidates for T . These kinds of arguments become more difficult when T has more than
three conjugacy classes of involutions, and in any case, it seems that it will be necessary to treat each
group individually, and to use the structure of its involution centralisers in some detail. Although not
an ideal state of affairs, we therefore leave the remaining cases for a future project.
6. Primitive point actions of type PA
We now apply Corollary 1.5 to the case where N is the socle of a primitive permutation group G
of O’Nan–Scott type PA. The notation of Hypothesis 6.1 coincides with that of Table 1.
Hypothesis 6.1. Let Q = (P, L) be a thick finite generalised quadrangle of order (s, t) admitting
a collineation group G that acts primitively of type PA on P, writing T r 6 G 6 H ≀ Symr for some
almost simple primitive group H 6 Sym(Ω) with socle T , where r > 2.
Proposition 6.2. If Hypothesis 6.1 holds then 2 6 r 6 4 and every non-identity element of H
fixes less than |Ω|1−r/5 points of Ω.
Proof. The socle of G is N = T r and the action of H on Ω is not semiregular, so the result follows
immediately from Corollary 1.5.
To say more than this, we would like to have generic lower bounds for the fixity f (H), namely
the maximum number of fixed points of a non-identity element, of an almost simple primitive group
H 6 Sym(Ω). This problem was investigated in a recent paper of Liebeck and Shalev [21], who proved
that f (H) > |Ω|1/6 except in a short list of exceptions. This lower bound is not quite large enough
to force further restrictions on r in Proposition 6.2, because to rule out r = 4 (as we did for types
HC and CD) we would need f (H) to be at least |Ω|1/5 . However, Liebeck and Shalev remark (after
[21, Theorem 4]) that their |Ω|1/6 bound could potentially be improved generically to around |Ω|1/3 ,
which would be sufficient for this purpose. Work in this direction is currently being undertaken by
Elisa Covato at the University of Bristol as part of her PhD research [11], with the aim of classifying
the almost simple primitive permutation groups H 6 Sym(Ω) containing an involution that fixes at
least |Ω|4/9 points. As of this writing, the alternating and sporadic cases have been completed, and
so we are able to apply these results to sharpen Proposition 6.2 as follows.
Proposition
a sporadic simple
r ∈ {3, 4}, H = T
6.3. Suppose that Hypothesis 6.1 holds with r > 2 and T an alternating group or
group, and let S 6 H denote the point stabiliser in the action of H on Ω. Then
∼
= Altp with p a prime congruent to 3 modulo 4, and S ∩ T = p. p−1
2 .
Proof. Since r > 2, Proposition 6.2 tells us that r ∈ {3, 4}, and that the fixity f (H) of H must be at
most |Ω|1−r/5 . If f (H) 6 |Ω|1−3/5 = |Ω|2/5 then Covato’s results [11] imply that either (i) T ∼
= Altp
p−1
with p ≡ 3 (mod 4) a prime and S ∩ T = p. 2 , or (ii) H and S are in Table 4.
In case (i) we can at least deal with the situation where H = Symp . Indeed, by the argument in
[21, Section 6], there is an involution u ∈ S = p.(p − 1) fixing 2(p−3)/2 ( p−3
2 )! elements of Ω, which is
2/5
2/5
greater than |Ω|
= (2(p − 2)!)
provided that p > 7. If p = 7 then we observe that u still fixes
more than |Ω|1/3 elements. This rules out r = 4, because then 1/3 > 1 − r/5 = 1/5. For r = 3
we apply Proposition 2.6. We have |Ω| = 2 · 5! = 120 and hence |P| = |Ω|3 = 1203 , and the only
solution of 1203 = (s + 1)(st + 1) satisfying s > 2, t > 2 and properties (ii) and (iii) of Lemma 2.1 is
(s, t) = (119, 121), so Proposition 2.6 implies that every non-identity collineation of Q fixes at most
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
H
Alt9
J1
J1
He
He.2
Th
S
2
3 : SL2 (3)
23 .7.3
7:6
72 : 2.PSL2 (7)
72 : 2.PSL2 (7).2
25 .PSL5 (2)
H
S
J3 .2
O’N.2
M23
Th
B
19.9
31.15
23.11
31.15
47.23
17
Table 4. Actions with small fixity in Proposition 6.3.
|P|7/9 points. However, the collineation (u, 1, 1) ∈ G fixes more than |Ω|1/3 |Ω|2 = |P|7/9 points, a
contradiction. Therefore, if we are in case (i) then we must have H = Altp , as per the assertion.
Now suppose that H and S are in Table 4. First consider the six cases on the left-hand side of
the table. In each of these cases, f (H) is at least |Ω|1/5 , so r = 4 is ruled out. For r = 3, we apply
Proposition 2.6 as above. Since f (H) > |Ω|1/5 , we have in particular f (H) > |Ω|1/6 . Choose u ∈ H
fixing at least |Ω|1/6 elements of Ω, and consider the collineation (u, 1, 1) ∈ G, which fixes at least
|Ω|1/6+2 = |Ω|13/6 = |P|13/18 points of Q. Since the only solutions of |Ω|3 = (s + 1)(st + 1) satisfying
s > 2, t > 2 and properties (ii) and (iii) of Lemma 2.1 are those with t = s+2, Proposition 2.6 provides
a contradiction. Now consider the five cases on the right-hand side of Table 4. The actions of J3 .2,
O’N.2 and Th all have fixity greater than |Ω|1/6 [21, Lemma 5.3], so these are ruled out for both r = 3
and r = 4 by the same arguments as above. Now consider the action of M23 . Here |Ω| = 40320, and for
r = 4 there are no solutions of |Ω|r = (s + 1)(st + 1) satisfying s > 2, t > 2 and properties (ii) and (iii)
of Lemma 2.1. If r = 3, the only solution is (s, t) = (40319, 40321). By [21, Lemma 5.3], we have
f (H) = 5, realised by an element u of order 11, and so we can construct a collineation θ = (u, 1, 1) ∈ G
of Q fixing 5|Ω|2 = 8128512000 points. However, s = t − 2 < t + 3, so the final assertion of Lemma 2.2
implies that |Pθ | 6 (s + 1)(s + 3) = 1988752683 < 8128512000, and we have a contradiction. Finally,
consider the given action of B, for which |Ω| = 3843461129719173164826624000000. For r = 4
there is no admissible solution of |Ω|r = (s + 1)(st + 1). For r = 3 the only admissible solution
is (s, t) = (|Ω| − 1, |Ω| + 1), and so the final assertion of Lemma 2.2 implies that any non-identity
collineation of Q fixes at most (|Ω| + 1)(|Ω| + 3) points. However, [21, Lemma 5.3] tells us that
f (H) = 22, so we can construct a collineation with 22|Ω|2 points to yield a contradiction.
Remark 6.4. Further improvements to Proposition 6.2 will be made in a future project. In the
first instance, we hope to use Covato’s results [11] on fixities of Lie type groups (once available), to
complete our treatment of the cases r = 3 and r = 4. We also note that it is straightforward to handle
the case r = 2 with T a sporadic simple group, and likewise the almost simple case with sporadic
socle, computationally along the lines of [3, Section 6] (but assuming only point-primitivity and not
line-primitivity). However, we omit these computations from the present paper for brevity.
7. Proof of Theorem 1.1
Let us now summarise the proof of Theorem 1.1. If the primitive action of G on P has O’Nan–
Scott type AS or TW, then the conditions stated in Table 1 follow immediately from Theorem 1.3.
Types HS, HC and PA are treated in Proposition 5.10, Theorem 5.8 and Proposition 6.3, respectively.
Types SD and CD are treated together in Proposition 4.3.
8. Discussion and open problems
We feel that the results presented in this paper represent a substantial amount of progress towards
the classification of point-primitive generalised quadrangles, but there is evidently still a good deal
of work to do. We conclude the paper with a brief discussion, and outline some open problems that
could be investigated independently and then potentially applied to our classification program.
As discussed in Remark 5.11, we are confident that we will eventually be able to finish the HS case,
and it is at least somewhat clear how this might be done. The SD and CD cases would appear to be
18
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
more difficult, however. The arguments used in Section 5 do not work in these cases, because the proof
of Lemma 5.2 (and therefore Lemma 5.3) relies in a crucial way on having k = 2 in Hypothesis 4.1,
so that conjugation by an element of the underlying point-regular group M is a collineation. We have
thus far been unable to find a way to work around this difficulty in any sort of generality. On the other
hand, a primitive (respectively quasiprimitive) group of type SD must induce a primitive (respectively
transitive) permutation group on the set of simple direct factors of its socle T k , and it seems that it
should be possible to use this extra structure to say more about the SD and CD types, at least in
the primitive case (especially since we have already reduced the list of candidates for T to those in
Table 1). Although we have made some preliminary investigations along these lines, we do not yet
know how to finish the SD and CD cases, and so we leave this task for a future project.
8.1. Point-regular collineation groups, and a number-theoretic problem. There is, of
course, a potential — but apparently extremely challenging — way to deal with all of the types HS, SD
and CD, and also with type TW, in one fell swoop. In each of these cases, the full collineation group
must have a point-regular subgroup of the form T m , for some m, with T a non-Abelian finite simple
group. Hence, it would certainly be sufficient to show that such a group cannot act regularly on the
points of a generalised quadrangle. However, this would appear to be a very difficult problem in light of
the (limited) existing literature. Yoshiara [26] managed to show that a generalised quadrangle of order
(s, t) with s = t2 cannot admit a point-regular group, while Ghinelli [15] considered the case where
s is even and t = s, showing that such a group must have trivial centre and cannot be a Frobenius
group. Beyond this, not much else seems to be known in the way of restrictions on groups that can act
regularly on the points of a generalised quadrangle (though certainly many of the known generalised
quadrangles admit point-regular groups [5], and the Abelian case is understood [12]). Although
Yoshiara [26] has an extensive suite of lemmas that one might attempt to use to investigate (in
particular) the possibility that a group of the form T n acts point-regularly on a generalised quadrangle
Q, the bulk of these lemmas assume that the order (s, t) of Q satisfies gcd(s, t) 6= 1. Although this
condition holds under Yoshiara’s intended assumption that s = t2 , it seems to be difficult to guarantee
in general. Indeed, according to Lemma 5.3(ii) (and perhaps not surprisingly), it must fail in our HS
case. On the other hand, one might seek a contradiction by examining the arithmetic nature of the
equation |T |m = (s + 1)(st + 1) subject to the constraints s > 2, t > 2, s1/2 6 t 6 s2 6 t4 and
s + t | st(st + 1) imposed by Lemma 2.1, and asking when it can be guaranteed that a solution must
satisfy gcd(s, t) 6= 1. More generally, one might simply ask whether this equation can have any such
solutions at all. This motivates the following problem.
Problem 8.1. Determine for which non-Abelian finite simple groups T , and which positive integers
m, there exist integral solutions (s, t) of the equation
(13)
|T |m = (s + 1)(st + 1)
with
s > 2, t > 2, s1/2 6 t 6 s2 6 t4 and s + t | st(st + 1).
Failing this, determine when such a solution must satisfy gcd(s, t) 6= 1.
As noted before Proposition 4.2, there is always an ‘obvious’ solution of (13) when m is divisible
by 3, namely (s, t) = (|T |m/3 − 1, |T |m/3 + 1), and gcd(s, t) = 1 in this case because |T | is even. It
would be useful even to know whether this is the unique solution in this particular situation. We do
know that (13) has solutions for certain T when m = 1 or 2, as demonstrated by Table 3, but we
do not recall encountering any solutions apart from the aforementioned ‘obvious’ ones when m > 3.
Moreover, it is straightforward to run numerical computations that suggest that certain combinations
of families of T and values of m will never yield a solution of (13). For example, if T ∼
= PSL2 (q) and
m = 1 then there is no solution if q < 106 , but we do not see how to go about proving that there is
no solution for any q.
One might also ask about gearing Problem 8.1 towards the PA and AS cases, by seeking solutions of
(13) with |T | replaced by |H : S| for H an almost simple group with socle T and S a maximal subgroup
of H (compare Hypothesis 6.1, which reduces to the AS case if r is taken to be 1). However, solutions
of (13) seem to be rather more common in this setting, and so other methods are needed to rule out
cases where solutions arise. For example, if we take H = T = McL (the McLaughlin sporadic simple
group) then there are five (classes of) maximal subgroups S of H for which |H : S|2 = (s + 1)(st + 1)
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
19
has an ‘admissible’ solution: four maximal subgroups of order 40320, which yield (s, t) = (296, 5644),
and the maximal subgroup PSU4 (3), for which (s, t) = (24, 126).
8.2. Fixities of primitive groups of type TW. We conclude with a brief discussion of the
TW case. Let N = T1 × · · · × Tr , where T1 ∼
= ··· ∼
= Tr ∼
= T for some non-Abelian finite simple
group T . A primitive permutation group G 6 Sym(Ω) of type TW is a semidirect product N ⋊ P
with socle N acting regularly by right mutlitplication, and P 6 Symr acting by conjugation in such a
way that T1 , . . . , Tr are permuted transitively. Certain other rather complicated conditions must also
hold [2], and in particular T must be a section of P . If we intend to apply Theorem 1.1 to classify
the generalised quadrangles with a point-primitive collineation group of TW type, then we will need
‘good’ lower bounds for fixities of primitive TW-type groups. Liebeck and Shalev [21, Section 4] show
that, for every T and r, the fixity of G is at least |T |r/3 . Although this is very far away from the
4/5 exponent bound imposed by Theorem 1.1, we would be interested to know under what conditions
it could be improved to something ‘close’ to 4/5, so that we could at least rule out some of the
subgeometries listed in Lemma 2.2 and then perhaps use the underlying point-regular group to say
more. In [21, Section 4], Liebeck and Shalev consider an involution x ∈ P (which must exist because
T is a section of P and |T | is even) and observe that x induces a permutation of {T1 , . . . , Tr } that fixes
at least |T |ca+b elements of Ω ≡ T r , where the induced permutation has cycle structure (1a , 2b ) and
every involution g ∈ Aut(T ) satisfies |CT (g)| > |T |c . By [21, Proposition 2.4], we can take c = 1/3
independently of T , and so because a/3+b > (r−2)/3+2/3 = r/3, it follows that x fixes at least |T |r/3
elements. Now, c can certainly be taken larger than 1/3 in at least some non-Abelian finite simple
groups T (though presumably never as large as 4/5), and if we happen to have c > 1/2 then ca + b is
maximised when b = 1 (else it is maximised when a = 0). Hence, roughly speaking, if c happens to
be somewhat large (for a given T ) and we happen to be able to guarantee that x can be chosen with
b quite small, then we might have a useful bound on the fixity of G to work with. Bounds on c can
certainly be determined on a case-by-case basis from standard results about involution centralisers,
but in light of the rather involved necessary and sufficient conditions for P to be a maximal subgroup
of G, it is not clear to us what can be said about the cycle structure of permutations of {T1 , . . . , Tr }
induced by involutions in P . We therefore pose the following (somewhat vaguely worded) problem.
Problem 8.2. Under what conditions can a primitive permutation group of type TW and degree
d be guaranteed to have large fixity, where by “large” we mean, say, d3/4 or more?
Appendix A. Additional proofs
Proof of Proposition 2.4. Suppose that we are not in case (iii). Then, by the final assertion of
Lemma 2.2, we have |Pθ | 6 (s + 1)(t + 1), and we argue as in the proof of Theorem 1.3. We must
show that we are either in case (ii), or that f (s, t) = ((s + 1)(st + 1))7/9 − (s + 1)(t + 1) is positive.
2/9 . If s > 13 then
We have ∂f
∂t (s, t) = (s + 1)(7s − h(s, t))/h(s, t), where h(s, t) = 9((s + 1)(st + 1))
14 2/9 27
2/9 s8/9 , and so 7s − h(s, t) >
s) ( 26 st)2/9 6 9( 189
(using also 2 6 t 6 s2 ) we have h(s, t) 6 9( 13
169 )
189 2/9
8/9
1/9
2
s (7s − 9( 169 ) ). The right-hand side is positive if and only if s > ( 79 )9 ( 189
169 ) ≈ 12.01 > 12, and
so it follows that f (s, t) > 0 for all s > 13, for all s1/2 6 t 6 s2 . If 4 6 s 6 12 then a direct calculation
shows that f (s, t) > 0 for all s1/2 6 t 6 s2 , so it remains to consider s ∈ {2, 3}. If s = 2 then, by
the final paragraph of the proof of Theorem 1.3, either t = 2 and every non-identity collineation of Q
fixes at most 7 < |P|7/9 = 157/9 ≈ 8.22 points, or t = 4 and we are in case (ii). Finally, if s = 3 then
31/2 6 t 6 32 , and a direct calculation shows that f (3, t) > 0 for 31/2 6 t 6 7. Moreover, Q cannot
have order (s, t) = (3, 8) by Lemma 2.1(iii). If (s, t) = (3, 9) then Q is the elliptic quadric Q− (5, 3)
[22, 5.3.2], and a FinInG [4] calculation shows that (up to conjugacy) there is a unique non-identity
collineation θ fixing 40 > |P|7/9 = 1127/9 ≈ 39.25 points. Moreover, Qθ is a generalised quadrangle of
order (3, 3), and every other non-identity collineation of Q fixes at most 16 < 1127/9 points.
Proof of Proposition 2.5. Suppose that we are not in case (ii) or (iii). Then |Pθ | 6 (s + 1)(t + 1) by
the final assertion of Lemma 2.2. We show that f (s, t) = ((s+1)(st+1))94/125 −(s+1)(t+1) is positive.
31/125 . Let a =
We have ∂f
∂t (s, t) = (s + 1)(94s − h(s, t))/h(s, t), where h(s, t) = 125((s + 1)(st + 1))
31/125 ( 2a+1 st)31/125 6
2.9701×1015 . Then s > a, so (using also 2 6 t 6 s2 ) we have h(s, t) 6 125( a+1
a s)
2a
20
JOHN BAMBERG, TOMASZ POPIEL, CHERYL E. PRAEGER
)31/125 s124/125 , and hence 94s − h(s, t) > s124/125 (94s1/125 − 125( (2a+1)(a+1)
)31/125 ). The
9( (2a+1)(a+1)
2a2
2a2
125 ( (2a+1)(a+1) )31 ≈ 2.97009 × 1015 , and it follows
right-hand side is positive because s > a > ( 125
94 )
2a2
that f (s, t) > 0 for all s > a, for all s1/2 6 t 6 s2 .
Proof of Proposition 2.6. Since s = t − 2 < t + 3, the final assertion of Lemma 2.2 implies that
|Pθ | 6 (s + 1)(t + 1) = (s + 1)(s + 3). The result follows upon comparing this with |P| = (s + 1)3 .
Proof of Corollary 5.7. By Propositions 4.2 and 5.4, T is one of the Lie type groups in the second
column of Table 2. However, by Proposition 5.6, we must also have |CT (x)| < |T |5/9 for all x ∈ T \{1}.
We use this to show that T cannot have type Aε3 , B2 = C2 , 2 F4 or G2 .
If T ∼
= G2 (q) then q > 3 (because G2 (2) is not simple), |T | = q 6 (q 6 − 1)(q 2 − 1), and we can
choose x ∈ T with |CT (x)| = q|A1 (q)| = q 6 (q 2 − 1)/ gcd(2, q − 1) as in the proof of Lemma 3.3(ii).
If q is even or q > 19 then |CT (x)| > |T |5/9 , and if q ∈ {3, 5, 7, 9, 11, 13, 17, 19} then there is no
solution of |T |2 = (s + 1)(st + 1) satisfying s > 2, t > 2 and properties (ii) and (iii) of Lemma 2.1.
If T ∼
= 2 F4 (q) then q = 22n+1 with n > 1 (because 2 F4 (2) is not simple and 2 F4 (2)′ was treated
in Proposition 5.4), |T | = q 12 (q 6 + 1)(q 4 − 1)(q 3 + 1)(q − 1) and, as in the proof of Lemma 3.3(ii),
we can choose x ∈ T with |CT (x)| = q 10 |2 B2 (q)| = q 12 (q 2 + 1)(q − 1). This yields |CT (x)| > |T |5/9
for all q. If T ∼
= PSp4 (q) ∼
= Ω5 (q) then q > 3 (because PSp2 (2) ∼
= Sym6 is not simple), |T | =
4
4
2
q (q − 1)(q − 1)/ gcd(2, q − 1), and taking x ∈ T with |CT (x)| = q 4 (q 2 − 1)/ gcd(2, q − 1) as in
the proof of Lemma 3.4 yields |CT (x)| > |T |5/9 for all q > 3. Finally, if T ∼
= PSLε4 (q) (where
+
−
6
4
2
3
L := L and L := U), then |T | = q (q − 1)(q − 1)(q − ε)/ gcd(4, q − ε) and we can choose
x ∈ T with |CT (x)| = q 5 |GLε2 (q)|/ gcd(4, q − ε) = q 6 (q 2 − 1)(q − ε)/ gcd(4, q − ε) as in the proof of
Lemma 3.4. This yields |CT (x)| > |T |5/9 unless ǫ = + and q = 2, and in this case there is no solution
of |T |2 = (s + 1)(st + 1) satisfying s > 2, t > 2 and properties (ii) and (iii) of Lemma 2.1.
Proof of Corollary 5.9 (continued). Now suppose that T ∼
= PSLε (q) with n ∈ {7, 8}, and
n+1
choose x ∈ T as in the proof of Lemma 3.4, so that (1) holds. If n = 8 then |CT (x)| > |T |7/9 for all
q > 2, so Proposition 5.6 gives a contradiction. If n = 7 then |CT (x)| > |T |94/125 for all q > 2, so
condition (ii) must hold. That is, s 6 2.9701 × 1015 , so
(14)
|T | = |P| < (s + 1)4 6 (2.9701 × 1015 + 1)4 < 7.78188 × 1061 .
This implies that q 6 9, in which case there is no solution of |T | = (s + 1)(st + 1) satisfying s > 2,
t > 2 and properties (ii) and (iii) of Lemma 2.1.
If T ∼
= PSp8 (q) then, as per (2), |T | = q 16 (q 2 − 1)(q 4 − 1)(q 6 − 1)(q 8 − 1)/ gcd(2, q − 1) and we can
choose x ∈ T with |CT (x)| = q 16 (q 2 − 1)(q 4 − 1)(q 6 − 1)/ gcd(2, q − 1). This yields |CT (x)| > |T |94/125
for all q > 2, so again (14) must hold. This implies that q 6 53, in which case there is no solution of
|T | = (s + 1)(st + 1) satisfying s > 2, t > 2 and properties (ii) and (iii) of Lemma 2.1. Similarly, if
T ∼
= Ω9 (q) then we can take x ∈ T as in the proof of Lemma 3.4, so that with |CT (x)| = |SO±
2n (q)| =
12
q (q 4 ± 1)(q 2 − 1)(q 4 − 1)(q 6 − 1) as in (3). This yields |CT (x)| > |T |94/125 for all q > 2, so (14) must
hold, and we immediately have a contradiction because |Ω9 (q)| = |PSp8 (q)|.
Finally, suppose that T ∼
= PΩ±
2n (q) with n ∈ {7, 8}, and choose x ∈ T as in the proof of Lemma 3.4.
If n = 8 then by using (4) (for q odd) and (6) (for q even), one may check that |CT (x)| > |T |94/125
for all q > 2. Hence, (14) must hold, and this implies that q ∈ {2, 3}, in which case there is no
solution of |T | = (s + 1)(st + 1) satisfying s > 2, t > 2 and properties (ii) and (iii) of Lemma 2.1.
Now suppose that n = 7. Then (14) holds if and only if q 6 4, and in this case there is no solution of
|T | = (s + 1)(st + 1) satisfying s > 2, t > 2 and properties (ii) and (iii) of Lemma 2.1 for q ∈ {2, 4}.
Therefore, we must have q > 5. However, in this case |CT (x)| > |T |94/125 , so we have a contradiction.
(To check this, note that |CT (x)| > c · q 42 (q 5 − 1)(q 2 − 1)2 (q 4 − 1)(q 6 − 1)(q 8 − 1) where c = 12 or 41
according as q is even or odd.)
References
[1] L. Babai, “On the automorphism groups of strongly regular graphs II”, J. Algebra 421 (2015) 560–578.
[2] R. W. Baddeley, “Primitive permutation groups with a regular non-abelian normal subgroup”, Proc. London Math.
Soc. 67 (1993) 547–595.
SIMPLE GROUPS, PRODUCT ACTIONS, AND GENERALISED QUADRANGLES
21
[3] J. Bamberg, M. Giudici, J. Morris, G. F. Royle, and P. Spiga, “Generalised quadrangles with a group of automorphisms acting primitively on points and lines”, J. Combin. Theory Ser. A 119 (2012) 1479–1499.
[4] J. Bamberg, A. Betten, P. Cara, J. De Beule, M. Lavrauw, and M. Neunhöffer, FinInG — Finite Incidence Geometry,
Version 1.0, 2014.
[5] J. Bamberg and M. Giudici, “Point regular groups of automorphisms of generalised quadrangles”, J. Combin.
Theory, Ser. A 118 (2011) 1114–1128.
[6] J. Bamberg, S. P. Glasby, T. Popiel, and C. E. Praeger, “Generalised quadrangles and transitive pseudo-hyperovals”,
J. Combin. Des. 24 (2016) 151–164.
[7] J. Bamberg, S. P. Glasby, T. Popiel, C. E. Praeger and C. Schneider, “Point-primitive generalised hexagons and
octagons”, J. Combin. Theory, Ser. A 147 (2017), 186–204.
[8] J. Bamberg, T. Popiel, and C. E. Praeger, “Point-primitive, line-transitive generalised quadrangles of holomorph
type”, to appear in J. Group Theory, doi:10.1515/jgth-2016-0042.
[9] T. C. Burness and M. Giudici, Classical groups, derangements, and primes, Australian Mathematical Society Lecture
Series 25, Cambridge University Press, Cambridge, 2016.
[10] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker and R. A. Wilson, ATLAS of finite groups, Oxford University
Press, Oxford, 1985.
[11] E. Covato, The involution fixity of simple groups, PhD Thesis, University of Bristol, in preparation.
[12] S. De Winter and K. Thas, “Generalized quadrangles with an abelian Singer group”, Des. Codes Cryptogr. 39 (2006)
81–87.
[13] W. Feit and J. G. Thompson, “Solvability of groups of odd order”, Pacific J. Math. 13 (1963) 775–1029.
[14] The GAP Group, GAP — Groups, Algorithms, and Programming, Version 4.7.8, 2015.
[15] D. Ghinelli, “Regular groups on generalized quadrangles and nonabelian difference sets with multiplier −1”, Geom.
Dedicata 41 (1992) 165–174.
[16] D. Gorenstein, R. Lyons and R. Solomon, The classification of the finite simple groups, number 3, Mathematical
Surveys and Monographs 40, American Mathematical Society, Providence, 1998.
[17] D. Frohardt and K. Magaard, “Fixed point ratios in exceptional groups of rank at most two”, Comm. Algebra 30
(2002) 571–602.
[18] L. Morgan and T. Popiel, “Generalised polygons admitting a point-primitive almost simple group of Suzuki or Ree
type”, Electronic J. Combin. 23 (2016) P1.34.
[19] C. E. Praeger, C.-H. Li and A. C. Niemeyer, “Finite transitive permutation groups and finite vertex-transitive
graphs”, in: G. Hahn and G. Sabidussi (Eds.), Graph symmetry, Kluwer, 1997, pp. 277–318.
[20] M. W. Liebeck and G. M. Seitz, Unipotent and nilpotent classes in simple algebraic groups and Lie algebras, Mathematical Surveys and Monographs 180, American Mathematical Society, Providence, RI, 2012.
[21] M. W. Liebeck and A. Shalev, “On fixed points of elements in primitive permutation groups”, J. Algebra 421 (2015)
438–459.
[22] S. E. Payne and J. A. Thas, Finite generalized quadrangles, Pitman, London, 1984.
[23] N. Spaltenstein, “Caractères unipotents de 3 D4 (Fq )”, Comment. Math. Helvetici 57 (1982) 676–691.
[24] M. Suzuki, “On a class of doubly transitive groups”, Annals Math. 75 (1962) 105–145.
[25] J. Tits, “Sur la trialité et certains groupes qui s’en déduisent”, Inst. Hautes Études Sci. Publ. Math. 2 (1959) 13–60.
[26] S. Yoshiara, “A generalized quadrangle with an automorphism group acting regularly on the points”, European J.
Combin. 28 (2007) 653–664.
[27] H. Van Maldeghem, Generalized polygons, Birkháuser, Basel, 1998.
Centre for the Mathematics of Symmetry and Computation, School of Mathematics and Statistics,
The University of Western Australia, 35 Stirling Highway, Crawley, W.A. 6009, Australia.
Email: {john.bamberg, tomasz.popiel† , cheryl.praeger‡ }@uwa.edu.au
†
Current address: School of Mathematical Sciences, Queen Mary University of London, Mile End
Road, London E1 4NS, United Kingdom.
‡
Also affiliated with King Abdulaziz University, Jeddah, Saudi Arabia.
| 4math.GR
|
Memory Based Online Learning of Deep Representations from Video Streams
arXiv:1711.07368v1 [cs.CV] 17 Nov 2017
Federico Pernici, Federico Bartoli, Matteo Bruni and Alberto Del Bimbo
MICC – Media Integration and Communcation Center
University Of Florence
Abstract
Memory
1000110101
0010101000
We present a novel online unsupervised method for face
identity learning from video streams. The method exploits
deep face descriptors together with a memory based learning mechanism that takes advantage of the temporal coherence of visual data. Specifically, we introduce a discriminative feature matching solution based on Reverse Nearest
Neighbour and a feature forgetting strategy that detect redundant features and discard them appropriately while time
progresses. It is shown that the proposed learning procedure is asymptotically stable and can be effectively used in
relevant applications like multiple face identification and
tracking from unconstrained video streams. Experimental
results show that the proposed method achieves comparable results in the task of multiple face tracking and better
performance in face identification with offline approaches
exploiting future information. Code will be publicly available.
1010010101
0010101001
1010101001
0101010100
1010101001
1010011010
...
1000101001
1001101010
0010101001
1010011010
Figure 1. Memory based appearance representation. Left: Each
element in the memory consists of a descriptor with an associated
identity (indicated by box color) and an associated scalar value reflecting the degree of redundancy (indicated by grey box area) with
respect to the current representation. Right: The shaded regions
represent the original appearance representation (i.e. VGGface).
The descriptors outside those regions are learned from the video
and extend the original appearance representation.
tically similar information. This provides a variety of conditions in which an object can be framed, and therefore a
comprehensive representation of its appearance can be obtained. Accordingly, tracking a subject in the video could, at
least in principle, support a sort of unsupervised incremental learning of its appearance. This would avoid or reduce
the cost of annotation as time itself would provide a form
of self-supervision. However, this solution is not exempt
of problems [5]. On the one hand, parameter re-learning
of Deep Networks, to adequately incorporate the new information without catastrophic interference, is still an open
challenge [6, 7], especially when re-learning should be done
in real time while tracking, without the availability of labels
and with data coming from a stream which is often nonstationary. On the other hand, classic object tracking [8] has
substantially divergent goals from continuous incremental
learning. While in tracking the object appearance is learned
only for detecting the object in the next frame (the past information is gradually forgotten), continuous incremental
learning would require that all the past visual information
of the object observed so far is collected in a comprehensive
and cumulative representation. This requires that tracking
does not drift in the presence of occlusions or appearance
1. Introduction
Visual data is massive and is growing faster than our
ability to store and index it, nurtured by the diffusion and
widespread use of social platforms. Their fundamental
role in advancing object representation, object recognition
and scene classification research have been undoubtedly assessed by the achievements of Deep Learning [1]. However, the cost of supervision remains the most critical factor
for the applicability of such learning methods as linear improvements in performance require an exponential number
of labelled examples [2]. Efforts to collect large quantities
of annotated images, such as ImageNet [3] and Microsoft
coco [4], while having an important role in advancing object recognition, don’t have the necessary scalability and are
hard to be extended, replicated or improved. They may also
impose a ceiling on the performance of systems trained in
this manner. Semi or unsupervised Deep Learning from image data still remains hard to achieve.
An attracting alternative would be to learn the object appearance from video streams with no supervision, both exploiting the large quantity of video available in the Internet and the fact that adjacent video frames contain seman1
changes and that incremental learning should be asymptotically stable in order to converge to an univocal representation.
In this paper, we present a novel online unsupervised
method for face identity learning from unconstrained video
streams. The method exploits CNN based face detectors
and descriptors together with a novel incremental memory
based learning mechanism that collects descriptors and distills them based on their redundancy with respect to the current representation. This allows building a sufficiently compact and complete appearance representation of the individual identities as time advances (Fig. 1).
While we obtained comparable results with offline approaches exploiting future information in the task of multiple face tracking, our model is able to achieve better performance in face identification from unconstrained video. In
addition to this, it is shown that the proposed learning procedure is asymptotically stable and the experimental evaluation confirms the theoretical result. In the following, in
Section 2, we cite a few works that have been of inspiration
for our work. In Section 3 we highlight our contributions,
in Section 4 we expounded the approach in detail and finally, in Section 5, experimental results are given.
2. Related Work
Memory Based Learning: Inclusion of a memory
mechanism in learning [9] is a key feature of our approach.
On domains that have temporal coherence like Reinforcement Learning (RL) memory is used to store the past experience with some priority and to sample mini-batches to perform incremental learning [10] [11]. This makes it possible
to break the temporal correlations by mixing more and less
recent experiences. More recently, Neural Turing Machine
architectures have been proposed in [12, 13] and [14] that
implement an augmented memory to quickly encode and retrieve new information. These architectures have the ability
to rapidly bind never-before-seen information after a single
presentation via an external memory module. However, in
these cases, training data are still provided supervisedly and
the methods don’t scale with massive video streams.
Open Set: In addition to the incremental learning procedure, the system needs to have the capability to discriminate between already known and unknown classes (open
set) [15]. The open set classification is a problem of balancing known space (specialization) and unknown open space
(generalization) according to the class rejection option. Formalization for open space risk is considered as the relative
measure of open space compared to the overall measure
space [15, 16, 17, 18]. The underlying assumption in these
approaches is that data is I . I . D . which allows sampling the
overall space uniformly. However, in a continuously data
stream context, as in this paper, data is no longer independent and identically distributed, therefore balancing the
known space vs the unknown space is more difficult since
space with absence of data may be misinterpreted for open
space. Storing data in a memory module can limit these effects [19, 20].
Open World: The other fundamental problem is incorporating the identified novel classes into the learning system
(open world) [21]. This requirement favors non-parametric
methods, since they can quickly learn never seen before information by simply storing examples. The Nearest Class
Mean (NCM) classifier proposed in [22], has been shown
to work well and be more robust than standard parametric classifiers in an incremental learning setting [22] [23]
[24]. NCM’s main shortcomings are: it is not a discriminative classifiers and nonlinear data representation and/or non
I . I . D . data streams limit the effectiveness of using the mean.
We adopt from NCM the idea of prototype-based classification. However, the prototypes we use are not the average
features vectors but we keep a representative non redundant
discriminative subset.
Multiple Object Tracking: All the methods we described so far make use of ground truth labels and typically address the categorization problem in which data is
manually cropped around the object boundary. An alternative approach that in principle accomplishes the classincremental learning criteria expounded above (i.e. open
set and open world) but with the addition of unknown labels and with data coming from the output of a detector
(i.e. no manual cropped data) is Multiple Object Tracking (MOT) [25, 26]. Recent Multiple Object Tracking algorithms typically adopt appearance and motion cues into
an affinity model to estimate and link detections to form
tracklets which are afterwards combined into final trajectories [27, 28, 29, 30, 31, 32]. Most existing MOT methods are applied to pedestrian tracking and either use simple
color histogram features [28, 33, 34, 35, 36] or hand-crafted
features [37, 38, 39, 40] as the appearance representation
of objects and have simple appearance update mechanisms.
Few exceptions can operate online and use deep features
[41, 42, 43, 44] but they still assume continuous motion and
do not update the appearance. MOT methods are not suited
to abrupt changes across different shots or scenes since the
assumptions of continuous motion no longer hold. Abrupt
changes across different shots are typically handled offline
by exploiting tracklets into predetermined non-overlapping
shots as in clustering face descriptors [45] [46] [47] [48].
Long Term Object Tracking: Finally, another relevant
research subject to our learning setting is long-term object
tracking [49]. The aim of long-term object tracking is to
track a specific object over time and re-detect it when the
target leaves and re-enters the scene. Only a few works
on tracking have reported drift-free results on on very long
video sequences ([50, 51, 52, 53, 54] among the few), and
only few of them have provided convincing evidence on
6
Face
Detection
neighbor described in the next section. Unmatched descriptors of the faces in the incoming frame are stored in the
memory module with a new Id. They ideally represent hypothesys of new identities that have not been observed yet
and will eventually appear in the following frames. In order
to learn a cumulative and comprehensive identity representation of each observed subject, two distincts problems are
addressed. They are concerned with matching in consecutive frames and control of the memory module. These are
separately addressed in the following subsections respectively.
1
Controller
Memory
New Ids
Generation
Descriptor
Extraction
ko
Discriminative
Matching
ok
4.1. Reverse Nearest Neighbour Matching
Figure 2. Block diagram presenting the major work flow and functional components in the proposed method. The gray shaded region highlights the components discussed in this paper. The memory module and the matching strategy run on the GPU.
the possibility of incremental appearance learning strategies that are asymptotically stable [50][52]. However, all of
these works only address tracking and perform incremental
learning to detect the target in the next frame.
3. Contributions
1. We firstly combine in a principled manner Multiple
Object Tracking in an online Open World learning system
in which the learning strategy is shown to be asymptotically
stable.
2. The proposed method performs very well with respect
to offline clustering methods which exploits future information.
3. Different from several existing approaches, our proposed method operates online and and hence have a wider
range of applications particularly face recognition with
auto-enrollment of unrecognized subjects.
4. The proposed approach
In our system, deep face descriptors are computed on
face regions detected by a face detector and stored in a
memory module as:
N (t)
M(t) = {(xi , Idi , ei , ai )}i=1
While tracking in consecutive frames, it is likely that the
face of the same individual will have little differences from
one frame to the following. In this case, highly similar descriptors will be stored in the memory and quickly a new
face descriptor of the same individual will have comparable
distances to the nearest and the second nearest descriptor
already in the memory. In this case, a discriminative classifier like the Nearest Neighbor (NN) based on the distanceratio criterion [55] does not work properly and matching
cannot be assessed. We solved this problem by performing
descriptor matching according to Reverse Nearest Neighbour (ReNN) [56]:
n
o
||x −1NN (x )||
M? = (xi , Idi , ei , ai ) ∈ M(t) | ||xii −2NNIIt (xii )|| < ρ̄,
(2)
t
where ρ̄ is the distance ratio threshold for accepting a
match, xi is a deep face descriptor in the memory module
and 1NNIt (xi ) and 2NNIt (xi ) are respectively its nearest
and second nearest neighbor deep face descriptor in the incoming frame It .
Fig. 3 shows the effects of this change of perspective:
here two new observations are detected (two distinct faces,
respectively marked as o1 and o2 ). They both have distance ratio close to 1 to the nearest xi s in the memory (the
dots inside the grey region S). Therefore both their matchings are undecidable. Differently from NN, ReNN is able
to correctly determine the nearest descriptor for each new
(1)
where xi is the deep descriptor, Idi is the object identity (an
incremental number), ei is the eligibility factor (discussed
in the following), ai tracks the age of items stored in memory and N (t) is the number of descriptors at time t in the
memory module.
The block diagram of the proposed system is shown in
Fig. 2. As video frames are observed, new faces are detected and their descriptors are matched with those already
in the memory. Each newly observed descriptor will be assigned with the object identity of its closest neighbour according to a discriminative strategy based on reverse nearest
𝐨1
1
𝑆
𝐱𝑖
𝐨2
Figure 3. Reverse Nearest Neighbor for a repeated temporal visual
structure (S) with the distance ratio criterion. All elements xi
match with o1 , for clarity only one of them is highlighted to show
the distances (thick black lines).
descriptor in the incoming frame. In fact, with ReNN, the
roles of xi and oi are exchanged and the distance ratio is
computed between each xi and the oi as shown in figure
for one of the xi s (the yellow dot is associated to the newly
observed red dot). Due to the fact that with ReNN a large
number of descriptors (those accumulated in the memory
module) is matched against a relatively small set of descriptors (those observed in the current image), calculation of
the ratio between distances could be computationally expensive if sorting is applied to the entire set. However, minimum distances can be efficiently obtained by performing
twice a brute force search, with parallel implementation on
GPU [57]. This technique not only leverages the very efficient CUDA matrix multiplication kernel for computing the
squared distance matrix but it also exploits the GPU parallelism since each query is independent. GPU limited bandwidth is not an issue being the memory incrementally populated.
The other important advantage of using ReNN is that
all the descriptors xi of the shown repeated structure S of
Fig. 3 match with the descriptor o1 resulting in a one to
many correspondence: {o1 } ↔ {xi }. This capability provides a simple and sound method in the selection of those
redundant descriptors that need to be condensed into a more
compact representation. The feature o1 will be used, as described in the next section, to influence the other matched
(redundant) features xi regarding the fact that they belong
to the same repeated structure. Therefore not only ReNN restores the discriminative matching capability under the distance ratio criterion but it also creates the foundation for
the development of memory control strategies to correctly
forget the redundant feature information.
4.2. Memory Control
Descriptors that have been matched according to ReNN
ideally represent different appearances of a same subject
face. However, collecting these descriptors indefinitely
could quickly determine memory overload. To detect redundant descriptors and discard them appropriately, we defined a dimensionless quantity ei referred to as eligibility.
This is set to ei = 1 as a descriptor is entered in the memory module and hence decreased at each match with a newly
observed descriptor, proportionally to the distance ratio:
ei (t + 1) = ηi ei (t).
(3)
When doing this, we also re-set the age: ai = 0. Eligibility
allows to take into account both discriminative spatial redundancy at a rate proportional to the success of matching
in consecutive frames. In fact, as the eligibility ei of a face
descriptor xi in the memory drops below a given threshold
ē (that happens after a number of matches), that descriptor with its associated identity, age and relative eligibility is
6
1
𝐨1
𝐨2
Low
𝐱𝑖
Figure 4. The shape of the density (here in 2D) down-weighting
the eligibility associated to each matched descriptor in the memory. Features xi in proximity of the observed descriptor o1 have
their eligibility decreased to encourage their redundancy. The
asymmetric shape of the density encourages more diversity in the
open space far from the identity o2 rather than close.
removed from the memory module:
if (ei < ē) then M(t + 1) = M(t) \ {(xi , Idi , ei , ai )}. (4)
The value ηi is computed according to:
ηi =
1 d1i
ρ̄ d2i
α
,
(5)
where d1i and d2i are respectively the distances between xi
and its first and second nearest neighbour oi , the value ρ̄ is
the distance-ratio threshold of Eq. 2 here used to normalize ηi in the unit interval. The value of α emphasizes the
effect of the distance-ratio. With every memory update we
also increment the age ai of all non-matched elements by
1. Eq. 5 defines a density that weights more the eligibility
around the matched features and less the eligibility far apart
from their second nearest neighbor. This definition is similar to discriminative distance metric learning in which the
features belonging to two different classes are expected to
be separated as much as possible in the feature space. The
density defined by Eq. 5 can be visualized in Fig. 4 for some
values of the distance ratio below the matching threshold ρ̄.
Each 2D circle in the figure visually represents the density
weighting the eligibility of the matching descriptors. The
geometric shape of the density is a generalization to multiple dimensions of the Apollonious circle1 . In particular, the
asymmetric shape of the density induced by the distance ratio encourages learning feature diversity in the open space.
Therefore not only the matching is discriminative and indicated for rejecting hypotheses (Open Set) but also well
suited for learning in an Open World.
1 Apollonius of Perga (c. 262 BC - c. 190 BC) showed that a circle may
also be defined as the set of points in a plane having a constant ratio of
distances to two fixed foci.
1
𝐨1
𝐱𝑖
𝐨2
Figure 5. Matching with multiple identities. The identity o1
matches with two identities (yellow and green). The ambiguity
is resolved by assigning o1 with the Id having the largest number
of matched descriptors (i.e. the yellow identity).
4.3. Temporal Coherence in Image Space
The model previously described exploits video temporal coherence in the deep descriptor space, further spatiotemporal coherence is exploited in the image space introducing the following constraints:
1. Id novelty: Potential novel identities in the current
frame are included in the memory only if at least one known
identity is recognized in the current frame. This allows introducing novel identity information which is known to be
reasonably different from the recognized ones.
2. Id temporal coherence: An identity is assigned and
included in the memory only if has been assigned in two
consecutive frames. After the assignment (i.e. memory
inclusion) it must match at least once in the following 3
frames, otherwise it is discarded.
3. Id uniqueness: Duplicated Ids in the current frame
are not considered.
4. Id ambiguity: A subject may match with multiple
identities. This ambiguity is resolved by assigning all the
matched descriptors with the Id having the largest number
of matched descriptors as shown in Fig. 5.
Bounding box overlap, typically used in multiple object tracking, is not exploited since not effective in unconstrained video with abrupt motions. Video temporal coherence in the image space is explicitly enforced by the 2nd
constraint.
4.4. Memory Overflow Control
Our method, operating online, does not require any prior
information about how many identity classes will occur, and
can run for an unlimited amount of time. However, since the
memory requirement is dominated by the size of the stored
exemplars, if the number of identities increases indefinitely
the exemplar removal based on eq. 4 may not be sufficient
in handling redundancy and the system may overflow its
bounded memory. In this condition the system is forced
to remove some stored exemplars by the memory limita-
tions. To overcome this issue we follow a strategy similar
to [14, 58] that involves the use of a policy based on removing from the memory the Least Recently Used Access
(LRUA) exemplars. This is achieved by finding memory
items with maximum age ai in the memory, and write to
one of those. Therefore the system preserves recently encoded information according to the Eligibility strategy, or
writes to the last used location according to the LRUA strategy. The latter can function as an update of the memory
with newer, possibly more relevant information by avoiding
the deletion of rare but useful descriptors. A benefit of the
LRUA strategy is that of handling those features collected
in the memory that will never obtain matches. This effect is
largely due to scene occluders or with descriptors extracted
from bounding boxes computed from false positives of the
face detector. In the long run such features may waste critical space in the memory buffer.
4.5. Asymptotic stability
Under the assumption that descriptors are sufficiently
distinctive (as in the case of deep face descriptors), the
incremental learning procedure described above stabilizes
asymptotically around the probability density function of
the descriptors of each individual subject face. This can be
proved by studying the discrete dynamic system of Eq. 3
relating e(t + 1) to e(t) by the map T : X 7→ X as
e(t + 1) = T (e(t)). A fixed point of T corresponds to
an equilibrium state of the discrete dynamical system. In
particular if T is a contraction there is a unique equilibrium
state and the system approaches this state as time goes to
infinity starting from any initial state. In this case the fixed
point is globally asymptotically stable. More formally:
Theorem (Contraction Mapping) 1 Let (X, d) be a complete metric space and T : X 7→ X be the map of Eq. 3
such that d(T (e), T (e0 )) ≤ c · d(e, e0 ) for some 0 < c ≤ 1
and all e and e0 ∈ X. Then T has a unique fixed point in
X. Moreover, for any e(0) ∈ X the sequence e(n) defined
as e(n + 1) = T (e(n)), converges to the fixed point of T .
The key element that guarantees such theoretical asymptotic
stability is that the ReNN distance ratio is always below 1.
In fact, it is easy to demonstrate that the updating rule of
Eq. 3 is a contraction and converges to its unique fixed point
0 according to the Contraction Mapping theorem (Banach
fixed-point theorem).
The asymptotic stability of the method is illustrated in
Fig. 6 with a simple one-dimensional case. Two patterns
of synthetic descriptors, respectively modeling the case of a
distinctive identity (red curve) and a non distinctive identity
(black curve) are generated by two distinct 1D Gaussian distributions. The learning method was ran for 1000 iterations
for three different configurations of the two distributions.
The configurations reflect the limit case in which the dis-
Figure 7. MOTA for each video sequence in the BBT dataset.
Figure 6. Asymptotic stability of
. incremental learning of a face
identity in a sample sequence
tinctiveness assumption of the deep descriptors no longer
holds. Mismatches might therefore corrupt the distinctive
identity. The blue points represent the eligibility of the distinctive identity. The histogram in yellow represents the distribution of the distinctive identity as incrementally learned
by the system. The three figures represent distinct cases in
which the non distinctive identity is progressively overlapping the distinctive one. The ReNN matching mechanism
and the memory control mechanism still keep the learned
distinctive identity close to its ground truth pdf.
5. Quantitative Experiments
We focus on tracking/identifying multiple faces according to their unknown identities in unconstrained videos consisting of many shots typically taken from different cameras. We used the Music-dataset in [48] which includes 8
music videos downloaded from YouTube with annotations
of 3,845 face tracks and 117,598 face detections. We also
add the first 6 episodes from Season 1 of the Big Bang Theory TV Sitcom (referred as BBT01-06) [36]. Each video
is about more than 20 minutes long with 5-13 people and is
taken mostly indoors. The main difficulty lies in identifying
faces of the same subject from a long video footage.
The two algorithm parameters in Eq. 5 are set empirically to: ρ̄ = 1.6 and α = 0.01. Deep face descriptor are
extracted according to [62]. We firstly show the capability
of the proposed method to perform online learning without
drifting using the long sequences of the BBT dataset. This
consists on monitoring the performance degradation of the
system as time advances. A decrease in performance may
eventually hinder learning being the system in a condition
from which is not possible to recover. In order to build a pic-
ture of the performance over time we evaluate the method
with the metric set commonly used in multiple object tracking [63]. In particular we report the MOTA: The Multiple
Object Tracking Accuracy that takes into account false positives, wrongly P
rejected identities and identity switches as:
(FNt +FPt +IDSt )
where GTt , FNt , FPt and
MOTA = 1 − t P
t GTt
IDSt are respectively the number of ground truth objects,
the number of false negatives, the number of false positives
and the number of identity switches at time t. The identity switches are defined as the total number of times that a
tracked trajectory changes its matched GT identity. Fig. 7
shows the MOTA curves as time progresses for each video
sequence of the BBT dataset for about 30000 frames. Each
individual frame is used to test the model before it is used
for training by the incremental learning procedure [64]. As
can be seen from the figure the curves reveal the stability
of the learning mechanism confirming the theoretical result
of Sec. 4.5. The initial fluctuations typically vary from sequence to sequence and reflect the approximate invariance
of the original representation. That is, the few descriptors
entering in the memory at the beginning of each sequence
do not provide substantial improvement with respect to the
original representation. However, as time advances, the reduction of fluctuations reveal that the proposed method is
able to learn by collecting all the non-redundant descriptors
it can from the video stream until no more improvement is
possible.
We further compare the proposed algorithm with other
state-of-the-art MOT trackers, including modified versions
of TLD [65], ADMM [60], IHTLS [61]. We specifically
compare with two multi-face tracking methods using the
TLD method implemented as described in [48]. The first
method, called mTLD, runs multiple TLD trackers for all
targets, and each TLD tracker is initialized with the ground
truth bounding box in the first frame. The second method,
referred as mTLD2, is used to generate shot-level trajectories within each shot initializing TLD trackers with untracked detections, and link the detections in the following
frames according to their overlap scores with TLD outputs.
The methods indicated as Pre-trained, SymTriplet,
Table 1. Quantitative comparison with other state-of-the-art multi-object tracking methods on the Music video dataset
Method
mTLD [59]
ADMM [60]
IHTLS [61]
Pre-Trained [48]
mTLD2 [59]
Siamese [48]
Triplet [48]
SymTriplet [48]
MuFTiR-dpm
MuFTiR-tiny
Method
mTLD
ADMM
IHTLS
Pre-Trained
mTLD2
Siamese
Triplet
SymTriplet
MuFTiR-dpm
MuFTiR-tiny
A PINK
Mode
IDS ↓
Offline
31
Offline
179
Offline
173
Offline
100
Offline
173
Offline
124
Offline
140
Offline
78
Online
121
Online
191
MOTA ↑
-2.2
72.4
74.9
54.0
77.4
79.0
78.9
80.0
21.8
55.1
G IRLS A LOUD
Mode
IDS ↓ MOTA ↑
Offline
9
-1.1
Offline
487
46.6
Offline
396
51.8
Offline
138
42.7
Offline
322
46.7
Offline
112
51.6
Offline
80
51.7
Offline
64
51.6
Online
51
-2.7
Online
339
49.3
Method
mTLD
ADMM
IHTLS
Pre-Trained
mTLD2
Siamese
Triplet
SymTriplet
MuFTiR-dpm
MuFTiR-tiny
B RUNO M ARS
Mode
IDS ↓ MOTA ↑ MOTP ↑
Method
Offline
35
-8.7
65.3
mTLD
Offline
428
50.6
85.7
ADMM
Offline
375
52.7
85.8
IHTLS
Offline
151
48.3
88.0
Pre-Trained
Offline
278
52.6
87.9
mTLD2
Offline
126
56.7
87.8
Siamese
Offline
126
56.6
87.8
Triplet
Offline
105
56.8
87.8
SymTriplet
Online
78
4.5
61
MuFTiR-dpm
Online
420
48.8
65.5
MuFTiR-tiny
H ELLO B UBBLE
MOTP ↑
Method
Mode
IDS ↓ MOTA ↑ MOTP ↑ Method
71.0
mTLD
Offline
7
-3.5
66.5 mTLD
87.1
ADMM
Offline
115
47.6
69.9 ADMM
87.2
IHTLS
Offline
109
52.0
69.9 IHTLS
87.7
Pre-Trained
Offline
71
36.6
68.5 Pre-Trained
88.2
mTLD2
Offline
139
52.6
70.5 mTLD2
87.8
Siamese
Offline
105
56.3
70.6 Siamese
87.8
Triplet
Offline
82
56.2
70.5 Triplet
87.8
SymTriplet
Offline
69
56.5
70.5 SymTriplet
61
MuFTiR-dpm Online
170
4.0
59.0 MuFTiR-dpm
66.1
MuFTiR-tiny Online
88
51.4
69.9 MuFTiR-tiny
TARA
W ESTLIFE
Mode
IDS ↓ MOTA ↑ MOTP ↑ Method
Mode
IDS ↓ MOTA ↑
Offline
130
1.4
67.9 mTLD
Offline
20
-34.7
Offline
251
29.4
63.8 ADMM
Offline
223
62.4
Offline
218
35.3
63.8 IHTLS
Offline
113
60.9
Offline
143
57.3
72.4 Pre-Trained
Offline
85
57.0
Offline
251
56.0
72.6 mTLD2
Offline
177
58.1
Offline
106
58.4
72.5 Siamese
Offline
74
64.1
Offline
94
59.0
72.5 Triplet
Offline
89
64.5
Offline
75
59.2
72.4 SymTriplet
Offline
57
68.6
Online
124
15
68 MuFTiR-dpm Online
47
-0.2
Online
270
39.5
76.4 MuFTiR-tiny Online
76
58.9
MOTP ↑
71.2
76.1
76.1
75.5
76.3
76.3
76.3
76.3
61
65.4
Method
mTLD
ADMM
IHTLS
Pre-Trained
mTLD2
Siamese
Triplet
SymTriplet
MuFTiR-dpm
MuFTiR-tiny
Triplet and Siamese refers to the four alternatives methods proposed in [48]. In these methods including ADMM,
mTLD, mTLD2 and IHTLS, shot changes are detected and
the video is divided into non-overlapping shots. Within
each shot, a face detector is applied and adjacent detections are linked into tracklets. The recovered collection of
tracklets are used as face pairs (Siamese) or face triplets
(Triplet and SymTriplet) to fine-tune a CNN initial face feature descriptor based on the AlexNet architecture trained
on the CASIA-WebFace (Pre-trained). Then, appearance of
each detection is represented with the fine-tuned feature descriptors and tracklets within each shot are linked into shotlevel tracklets according to a global multiple object tracking [34, 66]. Finally tracklets across shots are subsequently
merged across multiple shots into final trajectories according to the Hierarchical Agglomerative Clustering.
We reported two alternative versions using the (Deformable Part Model) DPM [67] and the Tiny [68] face detectors. They are indicated as MuFTiR-DPM and MuFTiRTINY respectively. For such comparisons we also include
the multiple target metric MOTP: The Multiple Object
Tracking Precision. MOTP is the average dissimilarity between all true positives and their corresponding ground truth
targets. MOTP is a measure of localization precision. Given
the quite different nature between offline and online this
comparison is to be considered a proof-of-concept. However, given the good performance of the offline methods we
DARLING
Mode
IDS ↓ MOTA ↑
Offline
24
-22.0
Offline
412
53.0
Offline
381
62.7
Offline
115
42.7
Offline
278
59.8
Offline
214
69.5
Offline
187
69.2
Offline
169
70.5
Online
64
2.2
Online
449
62.1
P USSYCAT D OLLS
Mode
IDS ↓ MOTA ↑
Offline
24
3.1
Offline
287
63.2
Offline
248
70.3
Offline
128
65.1
Offline
296
68.3
Offline
107
70.3
Offline
99
69.9
Offline
82
70.2
Online
55
-13.5
Online
83
30.7
MOTP ↑
69.9
88.4
88.4
88.5
89.3
88.9
88.9
88.9
63.7
66.0
MOTP ↑
71.3
63.5
63.5
64.9
64.9
64.9
64.9
64.9
61.1
62.7
MOTP ↑
56.9
87.5
87.5
88.2
88.1
88.0
88.0
88.1
61.5
66.1
compare to, it is certainly non-trivial for our online method
to do any better. Table 1 shows that our online tracking
algorithm does reasonably well with respect to offline algorithms, although there are some exceptions. In H EL LO B UBBLE, B RUNO M ARS , DARLING , TARA and W ESTL IFE our best performing method has the MOTA score similar to the ADMM and IHTLS methods with little less identity switches. Despite the on par performance, our method
achieves the results without exploiting future information.
Performance are still good in A PINK, the identity switches
are still comparable despite a decrease in MOTA. Excluding
Siamese, Triplet and SymTriplet that use a refined descriptor specifically tailored to the clustered identities extracted
with the multiple passes over the sequence, our method is
on par with the other offline methods. Our main observation
is that with modern CNN based face detector and descriptor, the state-of-the-art offline trackers do not have expected
advantages over the simpler online ones. Advantages further thin when processing long video sequences that do not
fit into memory.
Results are confirmed in the BBT dataset as shown in Table 2. As in the previous comparison on the Music dataset,
except for the Siamese, Triplet and SymTriplet the overall performance are very good. In the Episode four we
achieved better results. Considering that CNN descriptor
fine-tuning takes around 1 hour per sequence on a modern
GPU, our method perform favorably in those applications
Table 2. Quantitative comparison with other state-of-the-art multi-object tracking methods on the BBT dataset.
BBT _ S 01 E 02
BBT _ S 01 E 01
Method
mTLD [59]
ADMM [60]
IHTLS [61]
Pre-Trained [48]
mTLD2 [59]
Siamese [48]
Triplet [48]
SymTriplet [48]
MuFTiR-tiny
Mode
Offline
Offline
Offline
Offline
Offline
Offline
Offline
Offline
Online
IDS ↓
1
323
312
171
223
144
164
156
24
MOTA ↑
-16.3
42.5
45.7
41.9
58.4
69.0
69.3
72.2
59.9
MOTP ↑
74.8
64.0
64.0
73.3
73.8
73.7
73.6
73.7
70.3
MOTA ↑
-15.9
9.7
13.3
0.1
11.6
23.0
18.0
19.5
53.2
MOTP ↑
76.8
65.8
65.8
66.3
66.3
66.4
66.4
66.4
69.6
Method
mTLD
ADMM
IHTLS
Pre-Trained
mTLD2
Siamese
Triplet
SymTriplet
MuFTiR-tiny
BBT _ S 01 E 04
Method
mTLD
ADMM
IHTLS
Pre-Trained
mTLD2
Siamese
Triplet
SymTriplet
MuFTiR-tiny
Mode
Offline
Offline
Offline
Offline
Offline
Offline
Offline
Offline
Online
IDS ↓
0
298
295
46
103
85
103
77
84
IDS ↓ MOTA ↑ MOTP ↑
1
-7.6
82.8
395
41.3
71.3
394
42.4
71.4
130
27.4
74.5
174
43.6
75.9
116
60.4
75.8
143
60.2
75.7
102
61.6
75.7
57
45.1
68.8
BBT _ S 01 E 05
Mode
IDS ↓ MOTA ↑ MOTP ↑
Offline
1
-15.5
76.9
Offline
380
37.4
68.2
Offline
360
33.8
68.2
Offline
98
32.3
75.0
Offline
169
46.4
74.9
Offline
128
60.7
75.0
Offline
118
60.5
74.9
Offline
90
60.9
74.9
Online
36
44.5
69.3
Mode
Offline
Offline
Offline
Offline
Offline
Offline
Offline
Offline
Online
Method
mTLD
ADMM
IHTLS
Pre-Trained
mTLD2
Siamese
Triplet
SymTriplet
MuFTiR-tiny
operating on real time streaming data. Currently our approach runs at 10 frame per second on 800x600 video frame
resolution on a Nvidia GeForce GTX TITAN X (Maxwell).
MOTA2 , while largely used to evaluate performance
in multiple object tracking, it is not fully appropriate to
evaluate the performance of identification in a open world
scenario. In fact, it does not explicitly handle target reidentification. Different identities assigned to the same individual in two distinct scenes are not accounted as an identity switch. This effect has particular impact with videos
obtained from multiple cameras or with many shots. In order to take into account this case, for each sequence we
alsoPreport the weighted cluster purity, defined as: W =
1
c mc pc , where c is the identity cluster, mc the number
M
of assigned identities, pc the ratio between the most frequently occurred identity and mc , and M denotes the total number of identity detections in the video. Table 3 and
2 Provided
by www.motchallenge.org
Table 3. Clustering results on Music Dataset. Weighted purity of
each video is measured on ideal number of clusters.
Videos
HOG
AlexNet
Pre-trained
VGG-Face
Siamese
Triplet
SymTriplet
MuFTiR-tiny
Apink
0.20
0.22
0.29
0.24
0.48
0.60
0.72
0.51
B. Mars
0.36
0.36
0.50
0.44
0.88
0.83
0.90
0.96
M USIC DATASET
Darling Girls A. Hello B.
0.19
0.29
0.35
0.18
0.30
0.31
0.24
0.33
0.34
0.20
0.31
0.29
0.46
0.67
0.54
0.49
0.67
0.60
0.70
0.69
0.64
0.73
0.89
0.59
P. Dolls
0.28
0.31
0.31
0.46
0.77
0.77
0.78
0.97
T-ara
0.22
0.25
0.31
0.23
0.69
0.68
0.69
0.72
Westlife
0.27
0.37
0.37
0.27
0.54
0.52
0.56
0.98
Table 4. Clustering results on Big Bang Theory Dataset. Weighted
purity of each video is measured on ideal number of clusters.
Episodes
HOG
AlexNet
Pre-trained
VGG-Face
Siamese
Triplet
SymTriplet
MuFTiR-tiny
BBT01
0.37
0.47
0.62
0.91
0.94
0.94
0.94
0.98
B IG BANG T HEORY
BBT02 BBT03 BBT04
0.32
0.38
0.35
0.32
0.45
0.35
0.72
0.73
0.57
0.85
0.83
0.54
0.95
0.87
0.74
0.95
0.92
0.74
0.95
0.92
0.78
0.98
0.98
0.85
BBT05
0.29
0.29
0.52
0.65
0.70
0.68
0.85
0.98
BBT06
0.26
0.26
0.52
0.46
0.70
0.70
0.75
0.94
BBT _ S 01 E 03
Method
mTLD
ADMM
IHTLS
Pre-Trained
mTLD2
Siamese
Triplet
SymTriplet
MuFTiR-tiny
Method
mTLD
ADMM
IHTLS
Pre-Trained
mTLD2
Siamese
Triplet
SymTriplet
MuFTiR-tiny
IDS ↓ MOTA ↑
5
-2.1
370
30.8
376
33.5
110
17.8
142
38.0
109
52.6
121
50.7
126
51.9
14
43.6
BBT _ S 01 E 06
Mode
IDS ↓ MOTA ↑
Offline
0
-3.9
Offline
527
47.5
Offline
515
43.2
Offline
191
27.8
Offline
192
37.7
Offline
156
46.2
Offline
185
45.4
Offline
196
47.6
Online
222
42.9
Mode
Offline
Offline
Offline
Offline
Offline
Offline
Offline
Offline
Online
MOTP ↑
69.4
68.1
68.0
67.5
67.9
67.9
67.8
67.8
68.4
MOTP ↑
89.3
97.6
97.7
98.2
97.8
97.9
98.0
98.0
69.2
Figure 8. Four frames from the B RUNO M ARS video sequence with
the superimposed estimated identities are shown.
4 show the quantitative results of the comparison with the
Music and the BBT datasets. HOG, AlexNet and VGGface
indicate the method [48] using alternative descriptors. HOG
uses a conventional hand-crafted feature with 4356 dimensions, AlexNet uses a generic feature representation with
4096 dimensions. Our proposed approach achieves the best
performance in six out of eight videos in the Music dataset
and it achieves state of the art in all the BBT video sequences.
Finally, Fig. 8 shows four frames of the of the B RUNO
M ARS sequence with the learned identities superimposed.
Faces appear sensibly diverse (see f.e. individual number
1), nonetheless it can be observed that the learning mechanism is capable to extend the original representation to preserve identities under large pose variations including face
profiles not included in the original representation.
6. Conclusion
In this paper we exploited deep CNN based face detection and descriptors coupled with a novel memory based
learning mechanism that learns face identities from video
sequences unsupervisedly, exploiting the temporal coherence of video frames. Particularly, all the past observed
information is learned in a comprehensive representation.
We demonstrate the effectiveness of the proposed method
with respect multiple face tracking on the Music and BBT
datasets. The proposed method is simple, theoretically
sound, asymptotically stable and follows the cumulative
and convergent nature of human learning. It can be applied in principle to any other context for which a detectordescriptor combination is available (i.e. car, person, boat,
traffic sign).
Acknowledgment
This research is based upon work supported in part by
the Office of the Director of National Intelligence (ODNI),
Intelligence Advanced Research Projects Activity (IARPA),
via IARPA contract number 2014-14071600011. The views
and conclusions contained herein are those of the authors
and should not be interpreted as necessarily representing the official policies or endorsements, either expressed
or implied, of ODNI, IARPA, or the U.S. Government.
The U.S. Government is authorized to reproduce and distribute reprints for Governmental purpose notwithstanding
any copyright annotation thereon.
References
[1] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton.
Imagenet classification with deep convolutional neural networks. In NIPS, volume 1, page 4, 2012. 1
[2] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in
deep learning era. In The IEEE International Conference on
Computer Vision (ICCV), Oct 2017. 1
[3] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li,
and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Computer Vision and Pattern Recognition,
2009. CVPR 2009. IEEE Conference on, pages 248–255.
IEEE, 2009. 1
[4] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays,
Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence
Zitnick. Microsoft coco: Common objects in context. In
European Conference on Computer Vision, pages 740–755.
Springer, 2014. 1
[5] Greenberg Andy. Watch a 10-year-old’s face unlock his
mom’s iphone x. https://www.wired.com/story/
10-year-old-face-id-unlocks-mothersiphone-x/, November 2017.
[Online; posted 14November-2017]. 1
[6] Zhizhong Li and Derek Hoiem. Learning without forgetting.
In European Conference on Computer Vision, pages 614–
629. Springer, 2016. 1
[7] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins,
Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks.
arXiv preprint arXiv:1606.04671, 2016. 1
[8] Matej Kristan, Ales Leonardis, Jiri Matas, Michael Felsberg,
Roman Pflugfelder, Luka Cehovin Zajc, Tomas Vojir, Gustav
Hager, Alan Lukezic, Abdelrahman Eldesokey, and Gustavo
Fernandez. The visual object tracking vot2017 challenge results. In The IEEE International Conference on Computer
Vision (ICCV), Oct 2017. 1
[9] Dharshan Kumaran, Demis Hassabis, and James L McClelland. What learning systems do intelligent agents need?
complementary learning systems theory updated. Trends in
cognitive sciences, 20(7):512–534, 2016. 2
[10] 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. 2
[11] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. In International Conference on Learning Representations, Puerto Rico, 2016. 2
[12] Alex Graves, Greg Wayne, and Ivo Danihelka. Neural turing
machines. arXiv preprint arXiv:1410.5401, 2014. 2
[13] Alex Graves, Greg Wayne, Malcolm Reynolds, Tim
Harley, Ivo Danihelka, Agnieszka Grabska-Barwińska, 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. 2
[14] Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan
Wierstra, and Timothy Lillicrap.
Meta-learning with
memory-augmented neural networks. In International conference on machine learning, pages 1842–1850, 2016. 2, 5
[15] Walter J Scheirer, Anderson de Rezende Rocha, Archana
Sapkota, and Terrance E Boult. Toward open set recognition. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 35(7):1757–1772, 2013. 2
[16] Walter J. Scheirer, Lalit P. Jain, and Terrance E. Boult. Probability models for open set recognition. IEEE Transactions
on Pattern Analysis and Machine Intelligence (T-PAMI), 36,
November 2014. 2
[17] Abhijit Bendale and Terrance E. Boult. Towards open set
deep networks. In The IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), June 2016. 2
[18] Ethan M Rudd, Lalit P Jain, Walter J Scheirer, and Terrance E Boult. The extreme value machine. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.
2
[19] Antoine Cornuéjols. Machine Learning: The Necessity of
Order (is order in order ?). In E. Lehtinen & T. O’Shea (Eds.)
F. Ritter, J. Nerb, editor, In order to learn: How the sequences of topics affect learning. Oxford University Press,
2006. 2
[20] Antoine Cornuéjols. On-line learning: where are we so
far? In Ubiquitous knowledge discovery, pages 129–147.
Springer, 2010. 2
[21] Abhijit Bendale and Terrance Boult. Towards open world
recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1893–1902,
2015. 2
[34] Chang Huang, Bo Wu, and Ramakant Nevatia. Robust object
tracking by hierarchical association of detection responses.
In European Conference on Computer Vision, pages 788–
801. Springer, 2008. 2, 7
[22] Thomas Mensink, Jakob Verbeek, Florent Perronnin, and
Gabriela Csurka. Distance-based image classification: Generalizing to new classes at near-zero cost. IEEE transactions
on pattern analysis and machine intelligence, 35(11):2624–
2637, 2013. 2
[35] Yuan Li, Chang Huang, and Ram Nevatia. Learning to associate: Hybridboosted multi-target tracker for crowded scene.
In Computer Vision and Pattern Recognition, 2009. CVPR
2009. IEEE Conference on, pages 2953–2960. IEEE, 2009.
2
[23] Thomas Mensink, Jakob Verbeek, Florent Perronnin, and
Gabriela Csurka. Metric learning for large scale image classification: Generalizing to new classes at near-zero cost.
Computer Vision–ECCV 2012, pages 488–501, 2012. 2
[36] Baoyuan Wu, Siwei Lyu, Bao-Gang Hu, and Qiang Ji. Simultaneous clustering and tracklet linking for multi-face
tracking in videos. In Proceedings of the IEEE International
Conference on Computer Vision, pages 2856–2863, 2013. 2,
6
[24] Marko Ristin, Matthieu Guillaumin, Juergen Gall, and Luc
Van Gool. Incremental learning of ncm forests for largescale image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
3654–3661, 2014. 2
[25] Laura Leal-Taixé, Anton Milan, Konrad Schindler, Daniel
Cremers, Ian Reid, and Stefan Roth. Tracking the trackers:
An analysis of the state of the art in multiple object tracking.
arXiv preprint arXiv:1704.02781, 2017. 2
[37] Markus Roth, Martin Bäuml, Ram Nevatia, and Rainer
Stiefelhagen. Robust multi-pose face tracking by multi-stage
tracklet association. In Pattern Recognition (ICPR), 2012
21st International Conference on, pages 1012–1016. IEEE,
2012. 2
[26] Wenhan Luo, Xiaowei Zhao, and Tae-Kyun Kim. Multiple
object tracking: A review. CoRR, abs/1409.7618, 2014. 2
[38] Bing Wang, Gang Wang, Kap Luk Chan, and Li Wang.
Tracklet association by online target-specific metric learning and coherent dynamics estimation. IEEE transactions on
pattern analysis and machine intelligence, 39(3):589–602,
2017. 2
[27] William Brendel, Mohamed Amer, and Sinisa Todorovic.
Multiobject tracking as maximum weight independent set.
In Computer Vision and Pattern Recognition (CVPR), 2011
IEEE Conference on, pages 1273–1280. IEEE, 2011. 2
[39] Cheng-Hao Kuo and Ram Nevatia. How does person identity
recognition help multi-person tracking? In Computer Vision
and Pattern Recognition (CVPR), 2011 IEEE Conference on,
pages 1217–1224. IEEE, 2011. 2
[28] Li Zhang, Yuan Li, and Ramakant Nevatia. Global data
association for multi-object tracking using network flows.
In Computer Vision and Pattern Recognition, 2008. CVPR
2008. IEEE Conference on, pages 1–8. IEEE, 2008. 2
[40] Ramazan Gokberk Cinbis, Jakob Verbeek, and Cordelia
Schmid. Unsupervised metric learning for face identification in tv video. In Computer Vision (ICCV), 2011 IEEE
International Conference on, pages 1559–1566. IEEE, 2011.
2
[29] Anton Andriyenko, Konrad Schindler, and Stefan Roth.
Discrete-continuous optimization for multi-target tracking.
In Computer Vision and Pattern Recognition (CVPR), 2012
IEEE Conference on, pages 1926–1933. IEEE, 2012. 2
[30] Yuan Li, Haizhou Ai, Chang Huang, and Shihong Lao. Robust head tracking with particles based on multiple cues fusion. In European Conference on Computer Vision, pages
29–39. Springer, 2006. 2
[31] Yuan Li, Haizhou Ai, Takayoshi Yamashita, Shihong Lao,
and Masato Kawade. Tracking in low frame rate video: A
cascade particle filter with discriminative observers of different life spans. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 30(10):1728–1740, 2008. 2
[32] Seung-Hwan Bae and Kuk-Jin Yoon. Robust online multiobject tracking based on tracklet confidence and online discriminative appearance learning. In Proceedings of the
IEEE conference on computer vision and pattern recognition, pages 1218–1225, 2014. 2
[33] Horesh Ben Shitrit, Jerome Berclaz, Francois Fleuret, and
Pascal Fua. Tracking multiple people under global appearance constraints. In Computer Vision (ICCV), 2011 IEEE
International Conference on, pages 137–144. IEEE, 2011. 2
[41] Fengwei Yu, Wenbo Li, Quanquan Li, Yu Liu, Xiaohua Shi,
and Junjie Yan. Poi: Multiple object tracking with high performance detection and appearance feature. In European
Conference on Computer Vision, pages 36–42. Springer,
2016. 2
[42] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and
Ben Upcroft. Simple online and realtime tracking. In 2016
IEEE International Conference on Image Processing (ICIP),
pages 3464–3468, 2016. 2
[43] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese.
Tracking the untrackable: Learning to track multiple cues
with long-term dependencies. In The IEEE International
Conference on Computer Vision (ICCV), Oct 2017. 2
[44] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple
online and realtime tracking with a deep association metric.
CoRR, abs/1703.07402, 2017. 2
[45] Baoyuan Wu, Yifan Zhang, Bao-Gang Hu, and Qiang Ji.
Constrained clustering and its application to face clustering
in videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3507–3514,
2013. 2
[46] Makarand Tapaswi, Omkar M Parkhi, Esa Rahtu, Eric Sommerlade, Rainer Stiefelhagen, and Andrew Zisserman. Total
cluster: A person agnostic clustering method for broadcast
videos. In Proceedings of the 2014 Indian Conference on
Computer Vision Graphics and Image Processing, page 7.
ACM, 2014. 2
[47] Shijie Xiao, Mingkui Tan, and Dong Xu. Weighted blocksparse low rank representation for face clustering in videos.
In European Conference on Computer Vision, pages 123–
138. Springer, 2014. 2
[48] Shun Zhang, Yihong Gong, Jia-Bin Huang, Jongwoo Lim,
Jinjun Wang, Narendra Ahuja, and Ming-Hsuan Yang.
Tracking persons-of-interest via adaptive discriminative features. In European Conference on Computer Vision, pages
415–433. Springer, 2016. 2, 6, 7, 8
[49] Long term detection and tracking workshop. In Conjunction
With The IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) Workshops (LTDT2014), June 2014. 2
[50] Z. Kalal, J. Matas, and K. Mikolajczyk. P-n learning: Bootstrapping binary classifiers by structural constraints. In
CVPR, june 2010. 2, 3
[51] Thang Ba Dinh, Nam Vo, and G. Medioni. Context tracker:
Exploring supporters and distracters in unconstrained environments. In CVPR, june 2011. 2
[52] Federico Pernici and Alberto Del Bimbo. Object tracking by
oversampling local features. IEEE transactions on pattern
analysis and machine intelligence, 36(12):2538–2551, 2014.
2, 3
[53] Yang Hua, Karteek Alahari, and Cordelia Schmid. Occlusion
and motion reasoning for long-term tracking. In Computer
Vision–ECCV 2014, pages 172–187. Springer, 2014. 2
[54] Zhibin Hong, Zhe Chen, Chaohui Wang, Xue Mei, Danil
Prokhorov, and Dacheng Tao. Multi-store tracker (muster):
A cognitive psychology inspired approach to object tracking.
June 2015. 2
[55] David G. Lowe. Distinctive image features from scaleinvariant keypoints. IJCV, 60:91–110, 2004. 3
[56] Flip Korn and S. Muthukrishnan. Influence sets based on
reverse nearest neighbor queries. In Proceedings of the 2000
ACM SIGMOD International Conference on Management of
Data, SIGMOD ’00, pages 201–212, New York, NY, USA,
2000. ACM. 3
[57] Shengren Li and Nina Amenta. Brute-force k-nearest neighbors search on the gpu. In International Conference on Similarity Search and Applications, pages 259–270. Springer,
2015. 4
[58] Lukasz Kaiser, Ofir Nachum, Aurko Roy, and Samy Bengio.
Learning to remember rare events. International Conference
on Learning Representations (ICLR), 2017. 5
[59] Z Kalal, J Matas, and K Mikolajczyk. Tracking-learningdetection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2011. 7, 8
[60] Caglayan Dicle, Octavia I Camps, and Mario Sznaier. The
way they move: Tracking multiple targets with similar appearance. In Proceedings of the IEEE International Conference on Computer Vision, pages 2304–2311, 2013. 6, 7, 8
[61] Mustafa Ayazoglu, Mario Sznaier, and Octavia I Camps. Fast
algorithms for structured robust principal component analysis. In Computer Vision and Pattern Recognition (CVPR),
2012 IEEE Conference on, pages 1704–1711. IEEE, 2012.
6, 7, 8
[62] O. M. Parkhi, A. Vedaldi, and A. Zisserman. Deep face
recognition. In British Machine Vision Conference, 2015.
6
[63] Laura Leal-Taixé, Anton Milan, Ian Reid, Stefan Roth,
and Konrad Schindler.
Motchallenge 2015: Towards
a benchmark for multi-target tracking. arXiv preprint
arXiv:1504.01942, 2015. 6
[64] João Gama, Indrė Žliobaitė, Albert Bifet, Mykola Pechenizkiy, and Abdelhamid Bouchachia. A survey on concept drift adaptation. ACM Comput. Surv., 46(4):44:1–44:37,
March 2014. 6
[65] Zdenek Kalal, Krystian Mikolajczyk, and Jiri Matas.
Tracking-learning-detection. IEEE transactions on pattern
analysis and machine intelligence, 34(7):1409–1422, 2012.
6
[66] Junliang Xing, Haizhou Ai, and Shihong Lao. Multiobject tracking through occlusions by local tracklets filtering and global tracklets association with detection responses.
In Computer Vision and Pattern Recognition, 2009. CVPR
2009. IEEE Conference on, pages 1200–1207. IEEE, 2009.
7
[67] Markus Mathias, Rodrigo Benenson, Marco Pedersoli, and
Luc Van Gool. Face detection without bells and whistles. In
European Conference on Computer Vision, pages 720–735.
Springer, 2014. 7
[68] Peiyun Hu and Deva Ramanan. Finding tiny faces. In The
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 7
| 1cs.CV
|
Embedding AC Power Flow in the Complex Plane
Part I: Modelling and Mathematical Foundation
arXiv:1604.03425v2 [cs.SY] 18 Jul 2016
Sina S. Baghsorkhi, Member, IEEE, and Sergey P. Suetin
Abstract—Part I of this paper embeds the AC power flow
problem with voltage control and exponential load model in the
complex plane. Modeling the action of network controllers that
regulate the magnitude of voltage phasors is a challenging task
in the complex plane as it has to preserve the framework of
holomorphicity for obtention of these complex variables with
fixed magnitude. The paper presents two distinct approaches to
modelling the voltage control of generator nodes. Exponential (or
voltage-dependent) load models are crucial for accurate power
flow studies under stressed conditions. This new framework for
power flow studies exploits the theory of analytic continuation,
especially the monodromy theorem for resolving issues that have
plagued conventional numerical methods for decades. Here the
focus is on the indispensable role of Padé approximants for
analytic continuation of complex functions, expressed as power
series, beyond the boundary of convergence of the series. The
zero-pole distribution of these rational approximants serves as
a proximity index to voltage collapse. Finally the mathematical
underpinnings of this framework, namely the Stahl’s theory and
the rate of convergence of Padé approximants are explained.
Index Terms—AC power flow, voltage control, exponential load
model, power flow feasibility, saddle-node bifurcation, algebraic
curves, analytic continuation, monodromy, continued fractions,
Padé approximants, reduced Gröbner basis, Stahl’s compact set.
I. I NTRODUCTION
Power flow, the most fundamental concept in power system
engineering, is at the heart of studies ranging from daily
operation to long-term planning of electricity networks. The
AC power flow problem is a system of nonlinear algebraic
equations that mathematically models the steady-state relations
between the phasor representation of parameters and unknown
states in an AC circuit. The parameters typically consist of
the power generated and consumed by source and sink nodes
and the electrical properties, i.e. the impedance, of lines
that connect these nodes. The unknown states are primarily
voltage phasors but could also include continuous or discrete
variables associated with network controllers, e.g. FACTS
devices and tap-changing or phase-shifting transformers. The
accurate and reliable determination of these states is imperative
for control and thus for efficient and stable operation of the
network. In certain studies it is equally vital to determine
for which parameter values the power flow problem becomes
infeasible as this condition is intimately linked to saddlenode bifurcation and the voltage collapse phenomenon where
the system loses structural stability [3]–[5]. The concept of
structural stability is only applicable to dynamical systems [6],
[7]. However as demonstrated previously in the context of the
differential-algebraic equations that model a power system,
the analysis of the static model, i.e. the algebraic equations,
is sufficient to determine where exactly in the parameter
space the system loses structural stability. The dynamical
model of load and generators are only needed to capture
oscillatory instability phenomena, such as Hopf bifurcation,
that can arise from the interaction of generators and their
controllers with the network [8]. Thus the distance in the
parameter space to power flow infeasibility might be1 regarded
as the margin of voltage stability [5]. Voltage collapse and
bifurcation is certainly one of the most theoretical areas in
electrical engineering. As conventional power systems undergo
a fundamental transformation by large-scale highly-variable
wind and solar generation distributed across the network, this
field may experience a resurgence [9]. Given the inherent limitations of traditional methods, deeper understanding of these
complicated phenomena requires new theoretical approaches
rooted in complex analysis and algebraic geometry.
The basis of power flow is Kirchhoff’s current law which
states that for every node i in N , the set of all nodes, Ii , the
net current flowing out of that node, is related to its voltage
Vi and those of its adjacent nodes Vk in the following way:
Ii =
X
Iik =
k∈N (i)
X
X Vi − Vk
Vk Yik
=
Zik
k∈N (i)
(1)
k∈N [i]
N (i) and N [i] are the open and closed neighborhoods of
node i. Iik is the current flow through the line connecting
node i and k and Zik = Rik + jXik is the impedance of that
line which is used to construct the diagonal and off-diagonal
elements of the admittance matrix as,
X 1
1
Yii =
, Yik = −
(2)
Zik
Zik
k∈N (i)
Since complex power is Si = Pi + jQi = Vi Ii∗ , the power
flow problem in its complex form can be expressed as,
X
Vi∗ Vk Yik
∀i ∈ N − {r}
(3)
Si∗ =
k∈N [i]
Here r is the voltage reference node with |Vr | = constant
and arg(Vr ) = 0. It also serves as the slack node meaning
that Sr is a free parameter that accounts for the mismatch of
complex power and its losses throughout the network.
The numerical methods, developed historically to solve
this problem, take the polynomial system of (3) out of its
complex form by reformulating it in rectangular or polar
forms. These techniques, all based on Newton’s method or its
variants, iteratively linearize the equations and approximate
the solution, starting from an initial guess.
There are two inherent shortcomings in such methods that
can arise near the feasibility boundary of (3) characterized by
the saddle-node bifurcation manifold in its parameter spaces.
1 As demonstrated in Part II of this paper, the power flow feasibility and
saddle-node bifurcation boundaries are not necessarily equivalent. In that case,
the power flow would be still feasible, albeit with a non-physical solution,
beyond the closest saddle-node bifurcation.
Physically, proximity to the feasibility boundary corresponds
to a network operating close to its loadability limit such
as periods of peak electricity demand. The first issue is
the increased likelihood of non-convergence even though the
operating point is still feasible. The second issue near the
feasibility boundary where different algebraic branches coalesce is convergence to solutions that lie on other algebraic
branches. Although dependent on the dynamical model of the
physical system, these solutions in power systems typically
signify unstable [5] or low voltage [10] operating points.
Most of these operating points cannot be physically realized
and are thus false solutions. The region of initial guesses
in Newton’s method that converge to a particular solution
has a fractal boundary [11]. The multiple fractal domains of
convergence are pressed together near the bifurcation manifold
which explains erratic behavior of such methods in finding the
desirable, i.e. stable, solution even with seemingly reasonable
initial guesses.
Recently a semidefinite relaxation of rectangular power flow
has been reformulated as a special case of optimal power flow
where the objective function of the semidefinite programming
(SDP) is minimizing active power loss [12]. This addresses
the convergence failure of iterative methods but has its own
serious drawbacks. First, the relaxation may not be tight and
yield a high rank matrix where it is impossible to recover
any solution to the original power flow, let alone the desirable
one. Second, if the solution of the relaxed problem is high
rank, nothing can be concluded on the feasibility of the power
flow in the same vein as non-convergence of iterative methods
cannot rule out the existence of solutions. Third, the suggested
heuristic, i.e. active power loss minimization, does not always
find the stable solution branch. The first two problems can be
remedied, at least in theory, by obtaining higher-order and thus
tighter relaxations of power flow equations. The computation
cost however, explodes with the order of relaxation and the
number of variables. Reference [13] discusses the theoretical
underpinning of this approach in the context of the generalized
moment problem and highlights its connection to real algebraic geometry which we see as an obstacle to distinguishing
the desirable solution branch for algebraic systems.
Among the above issues, the challenge of finding the
solution on the desirable branch, more than anything else,
underlines the significance of embedding the power flow
problem in the complex plane where the extraordinary potentials of analytic continuation theory for multi-valued complex
functions can be tapped. This is pioneered by the idea of
holomorphic embedding load flow (HELM) which builds on
the fact that under no load/no generation (Si = 0 ∀i ∈ N ),
the network has a trivial non-zero solution for voltage phasors.
This corresponds to all currents Iik being zero and reference
voltage Vr propagated across the network and this trivial
solution characterizes the stable branch2. Analytic continuation
of this solution (or more accurately speaking, the germs
developed at z = 0) is guaranteed by monodromy theorem
to yield the desirable solution all the way to the saddle-node
bifurcation in the parameter space of (3) where there is a
non-trivial monodromy (Appendix) and the physical solution
ceases to exist.
2 These
concepts are explained more concretely in Part II of this paper.
Although the idea of HELM has aroused significant interest
in the power system community, it still needs much further development before it can prove its superiority over conventional
methods. Here we demonstrate how the magnitude of power
flow complex variables, considered as functions of a single
complex variable z, can be held fixed while preserving the
framework of holomorphicity. This is an important step in the
embedding of power flow as it models the action of network
controllers in the complex plane. Under stressed conditions,
i.e. near the feasibility boundary, voltage magnitudes tend to
deviate far below their nominal values. Hence the modelling
of voltage-dependent or more generally exponential load is
another crucial aspect that is developed here. We also show
the indispensable role of Padé approximants in the cases of
voltage control and exponential load models. Throughout the
paper we refer to this method as PA to highlight the central role
of rational approximation of functions of a complex variable
for recovering the power flow solution. With this abbreviation
we also wish to emphasize the critical direction of research and
potential challenges for further development of this method.
The Part I of this paper is organized as follows. In Section
II we review the main ideas of HELM as presented in the
original paper [1], i.e. for the PQ buses, introduce the concept
of rational approximation of analytic functions in relation to
power series and continued fractions and explore the zeropole structure of Padé approximants for a 3-bus example. In
Section III we introduce the mathematical static model of the
most prevalent controller in the network, the automatic voltage
regulator (AVR) of the generator. We demonstrate through
modification of the previous 3-bus network, this time with
a generator (PV bus), how the approximation of functions of
a single complex variable is essential for analytic continuation
of the voltage phasors with fixed magnitude. We start from the
analysis of the parameterized polynomial system of equations
and obtain their corresponding algebraic curves via reduced
Gröbner basis method. We obtain the critical points of these
curves, interpret the zero-pole structure of Padé approximants
and its transformation as the solution reaches the feasibility (or
bifurcation) boundary and explain the significance of the zeropole distribution of the Padé approximants in terms of voltage
stability margin at a given operating point. In Section IV we
introduce an alternative approach for modelling the voltage
magnitude constraints. In Section V we address the shortcomings of a previous formulation in the literature that attempts
to incorporate the PV buses into the general framework of
HELM. In section VI we introduce the exponential load model
and its special case the ZIP load. In the Appendix we explain
the mathematical underpinning of this paper including the
concept of germ and the monodromy of multi-valued algebraic
functions in relation to the Stahl’s theory. We also explain
the rate of convergence of Padé approximants in the context
of the Stahl’s maximal domain and its Green’s function.
As illustrated in Section III, the method of embedding the
equations and the resulting structure of the analytic arcs
(branch cuts) has implications for the rate of convergence
of Padé approximants and may hinder the effective analytic
continuation of the developed germs. The numerical values of
power flow variables and parameters presented in this paper
are all in per unit.
II. E MBEDDING
S YSTEM OF E QUATIONS
C OMPLEX P LANE
THE
IN THE
Consider the following parametrization of complex power
flow equations in terms of z ∈ C with Vi∗ replaced with
independent variables Wi ,
X
Wi Vk Yik
∀i ∈ N − {r}
(4a)
zSi∗ =
k∈N [i]
zSi =
X
Vi Wk Yik∗
k∈N [i]
∀i ∈ N − {r}
(4b)
From a geometric point of view, the 2n equations of (4) define generically an affine algebraic curve in (z, V1 , V2 , ..., Wn ).
It follows from the Kirchhoff’s current law and the existence
of the voltage reference node (with Vr appearing in (4) as a
parameter)
that the polynomials
P on the right side of (4a)-(4b)
P
(i.e. k∈N [i] Wi Vk Yik and k∈N [i] Vi Wk Yik∗ ∀i ∈ N − {r})
are algebraically independent. To establish this algebraic independence in relation to the reference node requires rigorous
analysis, a task which lies outside the scope of this paper.
Taking the algebraic independence of these polynomials for
granted, degenerate cases where the equations of (4) define not
an algebraic curve but a higher-dimension algebraic variety can
only arise when the power flow problem is ill-defined as in the
case of networks with disconnected graphs. This is in line with
the physical intuition that in the absence of a reference voltage,
voltages are floating and a given Vi can assume any value in C.
The equations of (4) generate an ideal and give a starting basis
for finding the corresponding reduced Gröbner basis [14]. For
any lexicographic order, such as ... > Vi > z, this gives a
last basis element which is a bivariate polynomial fi (Vi , z)
for well-defined problems. fi (Vi , z) = 0 can be solved for an
algebraic (multi-valued) function Vi = Vi (z) which has holomorphic branches where ∂fi (Vi , z)/∂Vi 6= 0. By permuting
the order, we arrive at 2n algebraic functions Vi = Vi (z),
Wi = Wi (z), i = 1, ..., n, giving an algebraic parametrization
of the curve defined by (4). A z-critical point of this curve will
be where any of the components Vi (z) or Wi (z) has a branch
point, i.e., fi (Vi , z) = 0 and ∂fi (Vi , z)/∂Vi = 0 (similarly for
the Wi ) [15]. The branch point closest to the point z0 at which
the Taylor series expansion of any single-valued branch is
developed, determines the radius of convergence of the series.
Branch points play a critical role in the analytic continuation of
these functions and the PA method (Appendix). This analysis
also extends to the case of voltage control where we introduce
new functions V i (z), Qi (z) and Si (z).
Since Vi (z) is analytic in z, (Vi (z ∗ ))∗ is also analytic in
z and identical to the conjugate of Vi (z) on the real axis.
Hence the solution process involves analytic continuation of
the functions of a single complex variable from z = 0 to z = 1
in the following system,
X
zSi∗
Vk (z)Yik
=
∗
∗
(Vi (z ))
k∈N [i]
∀i ∈ N − {r} (5)
P∞ [i]
P∞ [i]
By defining Vi (z) = n=0 cn z n , 1/Vi (z) = n=0 dn z n
P
∞
∗ [i] n
and 1/(Vi (z ∗ ))∗ =
n=0 dn z , and requiring that
∗ ∗
(Vi (z )) 6= 0 this system is adequately described by the
following set of power series relations,
zSi∗
∞
X
∞
X
X
cn [k] z n ) ∀i ∈ N − {r} (6a)
d∗n [i] z n = (Yik
n=0
k∈N [i]
∞
∞
X
X
n
n
(
c[i]
d[i]
n z )(
n z )
n=0
n=0
n=0
∀i ∈ N − {r} (6b)
=1
The procedure to obtain the coefficients of (6) starts by
setting
z = 0 in (6a). This gives the linear system of
P
[k]
= 0 which always yields the trivial solution
k∈N [i] Yik c0
[i]
[i]
[i]
(i.e. c0 = Vr ∀i ∈ N − {r}). Next d0 = 1/c0 by setting
z = 0 in (6b). The higher order coefficients are progressively
obtained by solving the linear system of (7a) (∀i ∈ N − {r})
which itself is obtained by differentiating (6a) with respect to
z and evaluating at z = 0 and the convolution formula of (7b).
Si∗ d∗n−1 [i] =
X
Yik cn [k]
k∈N [i]
d[i]
n =
−
[i]
[i]
m=0 cn−m dm
[i]
c0
Pn−1
∀i ∈ N − {r}
(7a)
∀i ∈ N − {r}
(7b)
The radius of convergence is R = limn→∞ |cn |/|cn+1 |,
if the limit exists. This marks the distance from the origin
to the closest branch point. Notice that when an analytic
function does not have a closed-form expression as in this
case, its representation as a power series expansion can be
approximated by a partial sum of a finite order. Since this
approximation for Vi (z) does not converge for |z| > R
the analytic continuation of these complex functions toward
z = 1 requires an alternative representation of these analytic
functions. One such representation, with superior convergence
properties, is a continued C-fraction which is approximated by
truncation. The relation between these two representations is
crucial for understanding Padé approximants and is described
below [17],
For a given power series V (z) = c0 + c1 z + c2 z 2 + ...,
assume the existence of the reciprocal relation between the
original series, as modified below, and a new series indexed
by superscript (1) ,
c2 z
c3 z 2
(1)
(1)
1+
+
+ ... = (1 + c1 z + c2 z 2 + ...)−1 (8)
c1
c1
Now the original power series can be expressed as,
(0)
c1 z
(9)
c0 + c1 z + c2 z 2 + ... = c0 +
(1)
(1)
1 + c1 z + c2 z 2 + ...
Next assume the existence of another reciprocal relation
between the modified series from the denominator of the
fraction in (9) and a new series indexed by superscript (2) ,
(1)
1+
c2 z
(1)
c1
(1)
+
c3 z 2
(1)
c1
(2)
(2)
+ ... = (1 + c1 z + c2 z 2 + ...)−1
(10)
This allows the expansion of the denominator of (9) in terms
of another fraction,
(0)
c1 z
c0 + c1 z + c2 z 2 + ... = c0 +
(1)
c1 z
1+
(2)
(2)
1 + c1 z + c2 z 2 + ...
(11)
By successively forming the reciprocal series we obtain a
C-fraction, written in a compact form as,
(2)
Bus 2
S2
−0.20−j0.10
(0)
1
(1)
(c0 (c1
(2)
(13)
(0)
Bus 3 (Reference)
V3 =1.00 0
(0) (2)
+ c1 ) + c1 )z + c1 c1 z 2
(1)
(2)
1 + (c1 + c1 )z
(1)
(2)
where c1 = c1 , c1 = −c2 /c1 and c1 = (c22 −c1 c3 )/(c1 c2 ).
The diagonal Padé approximant of degree M of V (z),
hereafter appearing frequently in the text, is the (2M +1)th
convergent of its C-fraction representation in (12),
A2M (z)
PA[M/M ]V (z) =
(14)
B2M (z)
In general a given analytic function can be approximated
by PA[L/M ](z) where L and M are not necessarily equal,
L+M
X
00
j1.
(0)
c0 + (c0 c1 + c1 )z
A2 (z)
=
,
(1)
B2 (z)
1+c z
c0 +
A3 (z)
=
B3 (z)
1
L
Fig. 1: 3-bus network with no voltage magnitude constraint
1.4
1.2
1
0.8
0.6
0.4
0.2
0
a0 + a1 z + ... + aL z
+ O(z L+M+1 ) (15)
b0 + b1 z 1 + ... + bM z M
−0.2
Setting b0 = 1, the denominator coefficients b1 , ..., bM are
obtained by cross-multiplying (4), equating the coefficients of
z L+1 ,z L+2 ,...,z L+M to zero and solving the resulting linear
system. Next the numerator coefficients a0 , a1 , ..., aL are obtained similarly by considering the coefficients of z 0 ,z 1 ,...,z L.
Now consider the network of Figure 1 where the per-unit
values of power flow parameters are shown and the unknown
states are voltage phasors V1 and V2 . The Taylor series for
V1 (z) and V2 (z) are obtained based on (7) which are then
used to compute the Padé coefficients. The concentration of
zeros (o) and poles (∗) of the diagonal Padé approximant,
shown in Figure 2, defines the closest common branch point
of V1 (z) and V2 (z) at zb ≈ 1.2 which is also given by Fabry’s
theorem [19] as zb = limn→∞ cn /cn+1 . Note that the branch
point is the common limit point of the sequences of zeros
and poles (See the discussion of the Stahl’s compact set in
the Appendix). A closer inspection of the zeros and poles
reveals the exact location of the branch point at 1.21510.
This means that if the loading in the network is increased
by this factor the new operating point is infeasible. When
the loading is increased by 1.21509, PA[100/100] recovers
the solution of the network with active and reactive power
mismatches smaller than 10−5 . Here, as it is often the case
for the class of problems in (5), i.e. for networks with only
PQ buses and a voltage reference, analytic continuation of
the germs by rational approximation is unnecessary as all the
power series already converge at z = 1 (highlighted as a red
dot) and thus are sufficient for computing V1 and V2 . However,
Padé approximants are much more efficient for evaluating
these functions as they converge to a given function at a much
−0.8
n=0
cn z n =
S1
−0.10−j0.00
0+
1.0
By truncating the C-fraction in (12) we obtain its convergents which are rational fractions in z. For example the first
4 convergents of (12) are given as,
A0 (z)
A1 (s)
(0)
= c0 ,
= c0 + c1 z,
B0 (z)
B1 (z)
(1)
Bus 1
0.20 + j0.30
j1.
00
(1)
c1 z c1 z c1 z
1 + 1 + 1 +...
(12)
0.8
0+
(0)
c0 + c1 z + c2 z 2 + ... = c0 +
higher rate than the original power series does [19] and can
also discover the analytic structure of a given multi-valued
function [20].
−0.4
−0.6
−1
−1.2
−1.4
−1.4 −1.2 −1 −0.8 −0.6 −0.4 −0.2
0
0.2 0.4 0.6 0.8
1
1.2 1.4 1.6 1.8
2
2.2 2.4 2.6 2.8
Fig. 2: Zero-pole distribution of PA[100/100] corresponding
to the network of Fig. 1
III. E MBEDDING THE VOLTAGE M AGNITUDE
C ONSTRAINTS IN THE C OMPLEX P LANE3
For a generator node i ∈ G ⊂ N , the real (active) power
output, Pi = Re(Si ), is fixed whereas the imaginary (reactive)
power, Qi = Im(Si ), is a free parameter which is adjusted so
as to fix the magnitude of the voltage phasor Vi at a given
setpoint value Mi . Extending the general framework of the
holomorphic embedding to this case is particularly challenging
as here the magnitude of a holomorphic function, Vi (z), is to
be held fixed. Such a function has to be constant by the open
mapping theorem [18] as the image of Vi (z) in the complex
plane is a subset of a circle and thus Vi can no longer be an
open map. To resolve this contradiction we define an analytic
P∞ [i] n
function V i (z) =
independently of Vi (z) for
n=0 cn z
i ∈ G. Note that V i (z) 6= (Vi (z ∗ ))∗ and this distinction is
the essential concept behind embedding voltage constraints
and allows Vi (z) to adopt the value of Vr at z = 0 with
its magnitude approaching Mi as z increases. However the
challenge remains as how these two independent functions
3 The modelling approaches defined here and in the following section are
the improved versions and evolution of our earliest formulation in [21].
3
Bus 2
Bus 1
|V1 |=1.00
0.20 + j0.30
G
S2
−0.50−j0.30
0
1.0
+j
+j
1.0
0
Re(S1)=1.00
0.8
0
0
1.0
should be related to Mi so that at z = 1, V i (z) = (Vi (z ∗ ))∗
for i ∈ G. An approach that enforces Vi (z)V i (z) = Mi2 ,
leaves the possibility that at z = 1 Vi (z)V i (z) = Mi2 while
V i (z) 6= (Vi (z ∗ ))∗ . In other words the relaxation of V ∗ into
W may not yield a tight solution for the original algebraic
equations. To remedy this problem we relax V ∗ into W for a
given generator node i in the following relations,
X
Vk∗ Yik∗ + Vi Wi Yii∗
(16a)
Pi + jQi = Vi
k∈N (i)
∗
(Pi + jQi ) = Wi
X
Vk Yik + Vi Wi Yii
(16b)
Bus 3 (Reference)
V3 =1.00 0
k∈N (i)
Vi Wi = Mi2
(16c)
Note that here we have made no assumption that Qi is realvalued. Since P is real-valued, one can check that V ∗ = W if
and only if Qi is real-valued, i.e. when (Pi +jQi )∗ = Pi −jQi
and by equating the expressions of Q from (16a) and (16b)
we reach the following identity relating Vi , Wi and Mi ,
X
X
Vk Yik + Vi Yii∗ ) = 2Pi − Mi2 Yii − Vi Vk∗ Yik∗ (17a)
Wi (
k∈N (i)
k∈N (i)
Now if we incorporate (16) into the existing embedding
framework of (4) and demonstrate that Q is algebraic in z
then (17a) guarantees that the Taylor series expansion of Q
has real coefficients. Consider the following parametrization of
the power flow equations augmented by the generator nodes,
X
Wi Vk Yik
∀i ∈ N − {r} − G (18a)
zSi∗ =
k∈N [i]
zSi =
X
Vi Wk Yik∗
k∈N [i]
z(Pi − jQi ) =
z(Pi + jQi ) =
Wi (
(18b)
(18c)
X
Wi Vk Yik
∀i ∈ G
X
Vi Wk Yik∗
∀i ∈ G (18d)
k∈N [i]
k∈N [i]
X
∀i ∈ N − {r} − G
X
Vi Wk Yik∗
Vk Yik + Vi Yii∗ ) = 2Pi − Mi2 Yii −
Fig. 3: 3-bus network with voltage control at bus 1
and other variables associated with generator nodes such as
Q and V are unambiguously defined in relation to this germ.
As analytic functions their sole purpose is to facilitate the
construction of Si∗ (V, Mi , Pi ). Also notice that W and V
are eliminated in both (19a) and (19b) and this rules out
the possibility of ghost solutions where Wi (z) 6= (Vi (z ∗ ))∗ .
With this clarification we propose the following embedding
for networks with generator nodes4 ,
X
zSi∗
Vk (z)Yik
=
(Vi (z ∗ ))∗
k∈N [i]
∀i ∈ N − {r} (20a)
z(Si (z ∗ ))∗ X
Vk (z)Yik
=
(Vi (z ∗ ))∗
k∈N [i]
X
Si (z)
= V i (z)Yii∗ +
(Vk (z ∗ ))∗ Yik∗
Vi (z)
k∈N (i)
X
Vk (z)Yik + Vi (z)Yii∗ ) =
V i (z)(
∀i ∈ G
(20b)
∀i ∈ G
(20c)
∀i ∈ G
(20d)
k∈N (i)
X
Vi (z)(Vk (z ∗ ))∗ Yik∗
2Pi − Mi2 Yii −
k∈N (i)
k∈N (i)
k∈N (i)
∀i ∈ G
(18e)
Notice that in contrast to (4) here for each generator
node there are three algebraically independent equations
and three variables. This parameterization of power flow
equations defines generically an affine algebraic curve in
(z, V1 , V2 , ..., Wn , Qi , ...). Here again the reduced Gröbner
basis for a lexicographic order, such as ... > Qi > z, gives a
bivariate polynomial fi (Qi , z) for well-defined problems and
thus it follows that Qi is a multi-valued algebraic function of
z and can be represented as a power series expansion. The
general principle for defining and analytically continuing a
given solution germ can be summed up by reducing the system
of (18) in (19).
X
zSi∗
Vk (z)Yik
∀i ∈ N − {r} (19a)
=
(Vi (z ∗ ))∗
k∈N [i]
zSi∗ (V, Mi , Pi ) X
Vk (z)Yik
=
(Vi (z ∗ ))∗
k∈N [i]
∀i ∈ G (19b)
Note that a given set of voltage phasors V that satisfies (19)
uniquely characterizes a germ to the expanded system of (18)
Here Qi (z), by construction, has real coefficients and
Si (z) = Pi + jQi (z) ∀i ∈ G. The embedding of (20) is
based on the parametrization introduced in (18) and at z = 1
sufficiently determines the AC power flow relations in a network with load and generation. Notice the different embedding
of (20b) and (20c). The former corresponds to (19b) and in
combination with (20a) is used to develop the germ of the
stable branch and ensures that Vi (z) has the trivial solution Vr
at z = 0. The latter constructs Si (z) = Pi + jQi (z) based on
the relation that is enforced between Vi (z) and V i (z) in (20d).
Thanks to this unique construction of Qi (z) and subsequently
Si (z) one can easily inspect that the combination of (20b)
and (20c) enforces V i (z) = (Vi (z ∗ ))∗ at z = 1 and as such
we succeed in implementing the reduced system of (19) while
enforcing the voltage magnitude and active power constraints
of the generators. The resulting algebraic system is adequately
described by the following set of power series relations,
4 There are a number of minor variations to this specific formulation that
might be more advantageous from a computational point of view. We chose
this formulation as it was simpler to mathematically justify the key condition
of V i (z) = (Vi (z ∗ ))∗ at z = 1.
∞
∞
X
X
X
zSi∗ d∗n [i] z n = (Yik cn [k] z n ) ∀i ∈ N−{r}−G
n=0
n=0
k∈N [i]
∞
∞
∞
X
X [i]
X X
z(
g ∗ n z n )( d∗n [i] z n ) = (Yik cn [k] z n )∀i ∈ G
n=0
n=0
k∈N [i] n=0
∞
∞
X
X
n
n
(
c[i]
d[i]
∀i ∈ N − {r}
n z )(
n z )= 1
n=0
n=0
∞
∞
∞
X
X
X
X
n
(
(Yik cn [k] z n ) + Yii∗ cn [i] z n ) =
c[i]
n z )(
n=0
n=0
n=0
k∈N (i)
∞
∞
X
X
X
Mi2 Yii − 2Pi + (Yik∗ c∗n [k] z n )( cn [i] z n )∀i ∈ G
n=0
n=0
k∈N (i)
∞
∞
X
X
n
(
gn[i] z n )(
d[i]
n z ) =
n=0
n=0
∞
∞
X
X
X
∗
n
(Y
c∗n [k] z n )
∀i ∈ G
c[i]
z
+
Yii∗
ik
n
n=0
k∈N (i) n=0
(21a)
(21b)
(21c)
(21d)
(21e)
[i]
The key coefficients cn (∀i ∈ N − {r}) are progressively
obtained by differentiating (21a) and (21b) with respect to z,
evaluating at z = 0 and solving the resulting linear system
[i]
[i]
[i]
which itself requires the prior knowledge of dm , cm and gm
for m = 1, ..., n − 1. These coefficients are already obtained
at previous stages. Once this linear system of size |N | − 1
[i]
[i]
is solved for cn (∀i ∈ N − {r}), cn (∀i ∈ G) are obtained
[i]
from (21d). Then we can compute g n (∀i ∈ G) from (21e) and
repeat the previous steps to obtain the next set of coefficients.
[i]
[i]
[i]
[i]
Notice that by construction g0 = Pi +jq0 and gn = jqn for
n > 1 where q [i] , the coefficients of Qi (z), are real-valued. It
is straightforward to enforce the reactive limit of a generator
node i which changes (Si (z ∗ ))∗ to Si∗ in (20b) and removes
the corresponding equations in (20c)-(20d). Accordingly the
corresponding power series on the left side of (21b) is replaced
by (21a) and the ones in (21d) and (21e) are eliminated.
Now consider the modified network of Figure 3 where bus
1 has a generator that regulates its voltage magnitude at 1.00
and generates P1 = 1.00. Consider the embedding of (24).
∗
∗
∗
z(P1 + jQ1 ) = V1 (W1 Y11
+ W2 Y12
+ V3∗ Y13
)
z(P1 − jQ1 ) = W1 (V1 Y11 + V2 Y12 + V3 Y13 )
(24a)
(24b)
∗
∗
∗
P1 + jQ1 = V1 (V 1 Y11
+ W2 Y12
+ V3∗ Y13
)
(24c)
∗
V 1 (V2 Y12 + V3 Y13 + V1 Y11
)=
2
∗
∗
2P1 − M1 Y11 − V1 (W2 Y12 + V3∗ Y13
)
∗
∗
∗ ∗
zS2 = V2 (W1 Y12 + W2 Y22 + V3 Y23 )
zS2∗ = W2 (V1 Y12 + V2 Y22 + V3 Y23 )
(24d)
(24e)
(24f)
Here z is the embedding parameter and the unknown
states are (V1 , W1 , V2 , W2 , V 1 , Q1 ). All other quantities are
parameters of the power flow problem. Notice that (24a)-(24d)
represent the embedding of the power flow relations of bus 1
(PV) with (24c)-(24d) modeling the action of the generator
AVR in the complex plane enforcing |V1 | = |V 1 | = M1 at
z = 1. Equations (24e)-(24f) correspond to the power flow
relations of bus 2 (PQ).
From a geometric point of view, the algebraically
independent equations of (24) define an affine algebraic variety of dimension 1, i.e. an algebraic curve in
(z, V1 , W1 , V2 , W2 , V 1 , Q1 ). There is a polynomial ideal I ⊂
C[z, V1 , W1 , V2 , W2 , V 1 , Q1 ] corresponding to this algebraic
variety and the equations of (24) are only one basis, among
many different bases, that generate I [14]. These equations
can be the starting basis for finding the corresponding reduced
Gröbner basis of I with the unique property that for a given
lexicographic (lex) order, such as ... > V1 > z, the last
element of this special basis is a bivariate polynomial f (V1 , z).
The elements of the reduced Gröbner basis have a triangular
structure in terms of the appearance of the unknown states and
z and, as such, the algorithms for obtaining this special basis
are the nonlinear generalization of the Gaussian elimination in
linear algebra.
To highlight the significance of this analysis for voltage
collapse studies let us compute f (V2 , z) first for the set of
power flow parameters shown in Figure 3 and next for when
the power flow reaches its feasibility boundary at P1 = 2.6785.
The coefficient matrices
these two bivariate polynomials
P6 ofP
6
which have the form i=0 k=0 aik V2i z k are shown in (22)
and (23) respectively. Note that the rows of the matrices
correspond to (1, V2 , V22 , ..., V26 ) and the columns correspond
to (1, z, z 2, ..., z 6 ). So the element in the ith row and the kth
column is the coefficient of V i−1 z k−1 . We can recover all
solutions of the power flow by simply setting z = 1 and
solving for the roots of the resulting univariate polynomial
in V2 . This will include all valid solutions of V2 as well
as solutions where V2 6= W2∗ . By evaluating the next basis
element at z = 1 and a given solution of V2 we can obtain
the numerical value of the next unknown in the lex order, for
example V1 if the lex order was chosen as ... > V1 > V2 > z.
This value of V1 corresponds to that particular solution of
V2 . However this is not the focus of our paper. The critical
points of the projection of these algebraic curves onto C, the
extended complex plane, are where ∂fi (Vi , z)/∂Vi = 0. A
subset of these so-called z-critical points [15] for the two
cases of P1 = 1.0000 and P1 = 2.6785 are shown in Table I.
These points, as explained later, are the branch points of power
flow variables, considered as analytic functions of a single
complex variable, i.e. z and play an important role in the
analytic continuation of the germ from a trivial solution to the
actual power flow solution and contain vital information on the
voltage stability margin of the operating point. However there
is absolutely no need to compute the reduced Gröbner basis in
order to locate these branch points in C. In fact computation
of these basis is exponential space complete [16] and requires
time that is at least exponential in the number of solutions of
the polynomial system. Hence there is little prospect in the
near future for applying the concept of reduced Gröbner basis
to industrial power flow studies.
Instead we develop the germs of the unknown states
(V1 , V2 , V 1 , Q1 ), according to (21). Figure 4 shows the zeropole distribution of the diagonal Padé approximant for V1 (z)
forming the Stahl’s compact sets corresponding to the network
of Figure 3 for P1 = 1.0000 (Figure 4a) and P1 = 2.6785
(Figure 4b). Notice that the analytic arcs (branch cuts) of voltage phasors are highlighted by the distribution of zeros (o) and
0
0
0.00106−0.00018i
0.00065−0.00842i
−0.03691+0.00834i
0.03275+0.16840i
0.27347−0.00447i
0
0.00251−0.00171i
0
0.06219−0.04894i
0
0.07360−0.19166i
0
0
0.10368+0.00834i −0.04226−0.04332i
0.01831−0.02968i −0.05020−0.56373i −0.46670−0.47703i
0.38542−0.41374i −0.22562+0.08679i
−0.03951−0.10721i −1.53037+0.00275i −0.93086+1.70712i −0.68418−1.05204i
−0.19620+0.26933i 0.72469+3.04445i 3.55535−1.19261i −0.91634+0.33716i
0.01889+0.58108i
0.33003+0.12123i
0.61071+0.63059i 1.58947−2.87746i −2.37011−1.93190i
−0.48693−1.03924i −1.37499+0.56371i 0.15033+0.76564i
0.11278+0.83042i
0.00650−0.14824i
0.08156−0.12479i
0
0
0
−0.27099−0.16367i −0.22579+0.38863i
0.29999+0.04936i
0
0
0
0.00012+0.00009i
0.00091−0.00078i
−0.00452−0.00286i
−0.01357+0.02012i
0.02779+0.02690i
0.00058+0.00030i
0.00681−0.00255i
0.01500+0.00427i
0.09960−0.01779i
−0.00130−0.00754i
−0.01718+0.11334i
0
0
0.31730−0.43239i
0.45355−0.23719i
0.67023+0.07099i
0.03391−0.03513i
0.25444+0.19442i
0
0.01547−0.01417i
0
0
0
TABLE I
2.5742
−0.8496
−0.3672+0.6263i
−0.3672−0.6263i
−0.4644+1.2672i
−0.4644−1.2672i
0.04075−0.04565i
0.01180−0.06678i
0
0
0
0
0.18759+0.15906i −1.14552−0.28566i −0.26170+0.94492i
P1 = 1.0000
0.02509−0.00623i
0
0.03594−0.05669i −0.02112+0.10596i 0.02158−0.06639i −0.00479+0.00533i
0.14498−0.10685i 0.07458+0.50024i −0.00671−0.34180i 0.03068+0.01416i
0.19685+0.00828i 0.39108+0.66032i −0.20047+0.44572i
0.35001+0.33909i −1.13221+0.70117i 0.54878+0.53354i
−0.24357−0.30435i −0.37291+0.24887i
−0.01073−0.04347i −0.06144+0.01628i 0.02506+0.03494i
Branch
Points
zB
z1
z2
z3
z4
z5
z6
z7
(22)
0
0
0
0.00019+0.00524i
P1 = 2.6785
Bifurcation
1.0000
−0.2266
−0.6170+1.1219i
−0.6170−1.1219i
0.7611+0.9593i
0.7611−0.9593i
0.9638+0.2129i
0.9638−0.2129i
poles (∗) of the truncated C-fraction, i.e. Padé approximant.
The points of infinite density of the zeros and poles are exactly
(within 5 decimal digits for PA[1000/1000]) the branch points
(cf. Table I) of V1 (z) which, in this case, are common with
V 1 (z), V2 (z) and Q1 (z). These branch points are a subset
the z-critical points of the algebraic curve that we computed
earlier as the last element of reduced Gröbner basis of the
polynomial system of (24). An analytic arc emanates from
each branch point and culminates in a different branch point or
in a Chebotarev’s point of the Stahl’s compact set. The region
of convergence is a disk bounded by the closest branch points,
a pair of complex conjugate points at zb = −0.3672±j0.6263.
In contrast to the previous case where the region of convergence of the series contained z = 1, here, the concept
of analytic continuation by Padé approximants is elegantly
illustrated. Since limn→∞ |cn+1 |/|cn | ≈ 1.4, the coefficients
tend to explode rapidly. Without Padé approximants based on
these otherwise useless coefficients, it is impossible to recover
the power flow solution.
Figure 4b shows the transformation of the Stahl’s compact
set as the power flow solution reaches the feasibility boundary
at P1 = 2.6785. The branch point on the positive real axis
has now moved to z = 1. Since past the branch point, there
is a non-trivial monodromy, examining the PA solutions, as
the degree of the diagonal Padé approximants is increased,
reveals whether the power flow problem has a stable solution
or not. This procedure is shown concretely in sections V and
VI. The location of this branch point can also serve as a
proximity index to the feasibility boundary where the saddle-
0.06940−0.00611i
0.02803−0.02035i
0
0
(23)
node bifurcation, i.e. loss of structural stability, occurs. It
is worth mentioning that the power flow is still feasible at
P1 = 2.6785 and the PA method recovers the solution whereas
Newton-Raphson method fails to converge for P1 > 2.6750.
IV. A N A LTERNATIVE A PPROACH TO E MBEDDING THE
VOLTAGE M AGNITUDE C ONSTRAINTS IN THE C OMPLEX
P LANE
In Section III we essentially constructed Si∗ (V, Mi , Pi )
(∀i ∈ G) to be incorporated into (5). This involved, at each
stage, a matrix-vector multiplication to obtain the coefficients
of voltage phasors V which were subsequently used to obtain
the coefficients of V i and later the real-valued coefficients
of Qi (∀i ∈ G). The matrix dimension for developing the
power series expansion of voltage phasors was the same as the
number of load and generator nodes, i.e. |N |−1. An alternative
approach is to develop the power series expansion of V , V and
Q simultaneously. For Q to have real-valued coefficients, the
auxiliary variables V are extended to load nodes, in contrast
to the previous approach which V was exclusively defined for
generator nodes. The embedded equations take the following
form,
X
zSi∗
Vk (z)Yik
=
∗
∗
(Vi (z ))
k∈N [i]
∀i ∈ N − {r} (25a)
X
z(Si (z ∗ ))∗
Vk (z)Yik
=
∗
∗
(Vi (z ))
∀i ∈ G (25b)
X
zSi
=
V k (z)Yik∗
∗
∗
(V i (z ))
k∈N [i]
∀i ∈ N − {r} (25c)
X
zSi (z)
V k (z)Yik∗
=
∗
∗
(V i (z ))
k∈N [i]
∀i ∈ G (25d)
k∈N [i]
Vi (z)V i (z) = Vr2 + z(Mi2 − Vr2 )
∗
∗
0
∀i ∈ G (25e)
where Si (z) = Pi + jQ(z) and (Si (z )) = Pi − jQ(z)
in (25b) and (25d) with Qi (z) coefficients being real-valued
a fact that follows from the symmetry of embedded equations (25a) through (25d) in combination with (25e) which
2
1.5
4
1
2
0.5
1
0
B
−0.5
3
−1
5
−1.5
−2
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
(a) The Stahl’s compact set when the operating point is far from the power flow feasibility boundary (P2 = 1.0000).
2
1.5
2
1
4
0.5
6
1
0
B
7
−0.5
−1
5
3
−1.5
−2
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
(b) Transformation of the Stahl’s compact set when the operating point is precisely on the feasibility boundary (P2 = 2.6785)
Fig. 4: Zero-pole distributions of PA[1000/1000] depicting the Stahl’s compact set, i.e. the analytic structure and the common
branch points of V1 (z), V 1 (z), V2 (z) and Q2 (z) (corresponding to the network of Fig. 3).
0.5
0
−0.5
−1
−0.5
0
0.5
1
1.5
2
Fig. 5: Zero-pole distribution of PA[1000/1000] depicting the Chebotarev’s point reaching z = 1 at bifurcation point
(corresponding to Fig. 3 and the approach defined by (27a))
enforces a complex-conjugate relationship5 between the coefficients of V i (z) and those of Vi (z). Note that at z = 0
Vi (z) = V i (z) = Vr and this characterizes the unstressed
(zero-current) state of the network. The resulting algebraic
system is adequately described by the following set of power
series relations,
∞
∞
X
X
X
zSi∗ d∗n [i] z n = (Yik cn [k] z n ) ∀i ∈ N−{r}−G (26a)
n=0
z(
∞
X
n=0
n=0
∞
X
k∈N [i]
n=0
∞
X
n=0
k∈N [i]
n=0
∞
X
X
∗ [i]
dn z n = (Yik∗
zSi
z(
n=0
k∈N [i]
∞
∞
X
X X
n
∗ [i] n
(Y
cn [k] z n )∀i ∈ G (26b)
z
)(
d
z
)
=
g ∗ [i]
ik
n
n
cn [k] z n ) ∀i ∈ N−{r}−G (26c)
n=0
∞
∞
X
X X
∗ [i]
gn[i] z n )( dn z n ) = (Yik∗ cn [k] z n ) ∀i ∈ G (26d)
n=0
k∈N [i]
n=0
∞
∞
X
X
( cn [i] z n )( cn [i] z n )=Vr2 + z(Mi2 − Vr2 ) ∀i ∈ G (26e)
n=0
∞
X
(
n=0
∞
X
(
n=0
n=0
n
c[i]
n z )(
∞
X
n=0
∞
X
n
c[i]
n z )(
n
d[i]
n z )= 1
[i]
dn z n ) = 1
n=0
[i]
[i]
∀i ∈ N − {r} (26f)
∀i ∈ N − {r} (26g)
[i]
The coefficients cn and cn (∀i ∈ N−{r}) and gn (∀i ∈ G)
are progressively obtained by differentiating (26a) through
(26e) with respect to z, evaluating at z = 0 and solving
the resulting linear system which itself requires the prior
[i]
[i] [i]
[i]
[i]
knowledge of cm , dm , cm , dm and gm for m = 1, ..., n−1, already obtained at previous stages. This involves matrix-vector
multiplication with the size of matrix being 2|N | + |G| − 2
(Contrast with |N |−1 in Section III). Notice that here, similar
[i]
[i]
to the previous approach defined in (21), g0 = Pi + jq0 and
[i]
[i]
gn = jqn for n > 1 where q [i] , the coefficients of Qi (z), are
real-valued.
The distinction between this approach and the one proposed
earlier in Section III is primarily related to the required degree
of Padé approximants to achieve a certain level of solution
accuracy. We have noticed in all cases examined so far that
with the exception of a few this alternative approach requires
lower degrees of Padé approximants. However, the advantage
of the first method is the smaller dimension of the matrix used
in developing the voltage phasor power series. This dimension
is expected to be critical for the computational performance of
the embedding method on extremely large networks. The Padé
degree requirement of each of these two approaches will be
examined in the context of medium and large-scale networks
in Part III of this paper.
V. A P ROBLEM
WITH A PREVIOUSLY PROPOSED APPROACH
Previously, a different approach to modeling PV (generator)
nodes was proposed that eliminates reactive power by manipulating the power flow equations of the generator node in the
5 This condition can be exploited to enhance the computational performance
of this approach and we will explain this in a future publication on the
computational aspects of different embedding approaches.
following way [22],
X
X
Yik∗ Vk∗
Yik Vk = 2Pi Vi − Vi2
Mi2
k∈N [i]
(27a)
k∈N [i]
This formulation can be incorporated into the general
framework of the holomorphic embedding for developing the
germ solution. However there are two major issues with this
approach. First (27a) does not adequately constrain the voltage
magnitude and one can check that the analytically continued
power series of the network of Figure 3, obtained based on the
specific embedded form of (27a) as presented in equation (20)
of reference [22], satisfies neither the voltage magnitude nor
the active power constraints. So there is a need to consider
V i (z). However, since reactive power is not considered, it
is difficult to see how a relation similar to (20d) can be
enforced in this case to guarantee V i (z) = (Vi (z ∗ ))∗ at z = 1.
Even under the assumption that a stronger relation is enforced
between V i (z), Vi (z) and Mi , for example V i (z)Vi (z) = Mi2
or V i (z)Vi (z) = (1 − z)Vr2 + zMi2 in the case of the
network of Figure 3, there arises an even more fundamental
problem. In Figure 4b we showed that the branch point
on the real-axis reaches z = 1 as the system experiences
saddle-node bifurcation. In the approach based on (27a) it is
the Chebotarev’s point that reaches z = 1 (see Figure 5).
Chebotarev’s point is a 0-density point for the equilibrium
measure that is concentrated on the Stahl’s compact set. In
the generic case, from each Chebotarev’s point exactly three
analytic arcs emanate and each arc culminates at either another
Chebotarev’s point or a branch point of f . In contrast, from
each branch point of f only a single analytic arc emanates and
culminates at either a Chebotarev’s point or another branch
point. Thus the local structure of the Stahl’s compact set
at a branch point is different from its local structure at a
Chebotarev’s point.
The rate of convergence of Padé approximants at a given
point z declines as that point gets closer to the Stahl’s compact
set and this can be explained in terms of the value of the
Green’s function gS (z, 0) (see the Appendix). However this is
not the unique reason for the rate of convergence to decline.
The structure of the Stahl’s compact set can also cause the
rate of convergence to decline. Notice the peculiar form of
the analytic arcs in Figure 5 forming a pair of pincers that
encompass the segment of the real axis from z ≈ 0.1 to z = 1.
In particular notice the small gap at the opening of the pincers.
As this gap shrinks, the value of the Green’s function for the
space contained within the pincers gets smaller. No matter
how large the space inside the pincers is, the small gap at
the opening of the pincers causes the rate of convergence of
Padé approximants to suffer drastically and this makes the
effective analytic continuation of the germ impossible when
the Chebotarev’s point is close to z = 1. This has implications
for power flow studies near the feasibility boundary where
embedding the power flow in the complex plane is most
needed as other methods tend to fail. By contrasting these
two approaches we intend to highlight the significance of the
correct embedding approach both for efficiently solving the
power flow and also for obtaining a reliable proximity index
to power flow infeasibility and voltage collapse based on the
zero-pole distribution of the rational approximants.
VI. E XPONENTIAL AND ZIP LOAD MODELS
In real systems loads are voltage dependent and their
representation as constant parameters may render the power
flow analysis grossly inaccurate. In this section we describe
how the exponential load model and its special case ZIP load
model can be incorporated into the framework of embedding
the power flow in the complex plane. Exponential load model
is described as,
P = |V |a P0
(28a)
Q = |V |b Q0
(28b)
where a and b are rational constants and each can be expressed
as a fraction of two relatively prime integers (m, n). Consider
the following relations assuming a = b (we later relax this
constraint),
m
V I ∗ = |V | n S0
∗
V I = |V |
m
n
S0∗
(29b)
2
V mn (I ∗ )mn = |V |m S0mn
2
|V |2mn I mn = |V |m V mn (S0∗ )mn
(30a)
(30b)
m−n
From (29) we have II ∗ = S0 S0∗ (V V ∗ ) n which transforms (30) into a purely phasor relation between the current,
voltage and apparent power of an exponential load given as,
V m (S0∗ )2n
(V ∗ )2n−m
(31)
Now suppose a = m/n and b = r/s in (28a) then we can
consider I as the sum of Ip and Iq given by,
V m P02n
(V ∗ )2n−m
r
V (−jQ0 )2s
=
(V ∗ )2s−r
Ip2n =
Iq2s
(32a)
(32b)
One can check that the case of m = 2, n = 1 in (31) corresponds to constant-impedance load and the case of m = 1,
n = 1 models constant-current load. By putting m = 0 in (31)
we obtain the familiar constant-power load model. Hence the
ZIP load model can be expressed as a relation between each
current component satisfying (31) for its corresponding (m, n)
and the net current being the sum of each component current
similar to I = Ip + Iq in (32).
Now that we have established the algebraic relations between the phasor quantities of various components of voltagedependent load model we can consider each component current as an algebraic function in z. The general principle is
similar to the case of voltage magnitude constraint where
we develop the germ of voltage phasors by constructing
Si∗ (V, Mi , Pi ) for a given generator node i. Here we construct
Ii (Vi , Si , m, n) which is again unambiguously defined in
relation to the germ of the stable solution. The corresponding
equation that should be added to (19) is as follows,
z
X
k∈L[i]
Iik (Vi , Sik , mk , nk ) =
X
k∈N [i]
∞
∞
∞
X
X
X
(
xk z k )2 = (S0∗ )2 (
ck z k )(
d∗k z k )
k=0
Vk (z)Yik
∀i ∈ E (33)
k=0
(34)
k=0
P∞
k
The power series I(z) =
is developed by
k=0 xk z
successive application of the concept of convolution. The
coefficient xk is obtained based on cj , dj and xj for j =
1, ..., k − 1, all given from previous stages, as follows,
(29a)
which are equivalent to,
I 2n =
where L[i] is the set of load components of node i each
characterized by its S0 and (m, n) and E is the set of nodes
with exponential load components.
Suppose a given node has a constant-current load where
m = n = 1 in (31) then the corresponding power series
relation is given as,
(S0∗ )2
xk =
k
X
j=0
ck−j d∗j −
2x0
k−1
X
j=1
xk−j xj
(35)
Note that the order of power series convolution in the right
side of (35) increases as m and n assume larger values in (31).
Thus the corresponding formula for obtaining xk can be much
more complicated. The increasing complexity is related to the
partitioning of integers. For example to compute x100 , the
101st coefficient of I, for n = 7, we need to have obtained,
beforehand, all partitions of integer 100 with number of parts
less than or equal to 2n = 14. One such partition has only a
single part which is 100 itself and its corresponding term is
2nx2n−1
x100 . All other partitions have corresponding terms
0
that are solely expressed in terms of x0 , x1 , ..., x99 , all obtained in previous stages. This task is, however, independent of
the problem itself and is purely in the realm of number theory
and combinatorics. In fact there are generating algorithms for
these partitions [23]. All that is needed is to somehow compute
and store these partitions for relevant combinations of (m, n).
We expect that I coefficients are obtained quite efficiently via
parallel computing when m and n are small integers but we
leave these aspects, including the need for higher numerical
precision, for future publications. Here we focus on the ZIP
load model, especially its constant-current component.
Figure 6a shows the Stahl’s compact set corresponding
to the network of Figure 3 where bus 2 has a constantcurrent load, i.e. P2 = −0.50|V2 | and Q2 = −0.30|V2 |, and
P1 = 1.00. The Padé approximated solution of V2 (z) and
I2 (z) are 0.7795 + j0.3032 and −0.5747 + j0.0983 and hence
|I2 | = |S2 | as expected. For this operating point the voltage
dependency of the load enhances the stability margin as seen
by the location of the branch point on the positive real axis
at zb = 9.5700). This is because as the system is stressed
by increased loading and as the voltage at bus 2 declines,
the realized load which is proportional to voltage magnitude
decreases further and further from the load at nominal voltage.
This drop in the actual load is clearly beneficial for the voltage
stability of the system. The situation is the opposite when
the system is stressed by increasing active power generation.
Under this condition where the voltage magnitude of the load
is suppressed by increasing the injection of active power a
voltage-dependent load would further stress the system by
accentuating the effect of increased active power injection.
2
1.5
1
0.5
0
−0.5
−1
−1.5
−2
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
9
2
9.5
10
10.5
11
(a) Zero-pole concentration forms the Stahl’s compact set highlighting the common branch points of V1 (z), V 1 (z), S1 (z) and V2 (z) (P1 = 1.00).
2
1.5
1
0.5
0
−0.5
−1
−1.5
−2
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
(b) Transformation of the Stahl’s compact set on the feasibility boundary (P1 = 2.6785)
Fig. 6: Zero-pole distribution of PA[1000/1000] depicting the analytic structure of V1 (z) (corresponding to Fig. 3)
.
This becomes evident by contrasting Figure 6b where bus 2
has a constant-current load and Figure 4b where bus 2 has
a constant-power load. Notice that under constant-power load
the network can absorb P1 = 2.6785 whereas this level of
generation is no longer feasible under constant-current load
as clearly shown by the analytic arc on real axis covering
z = 1. This contrast highlights the significance of appropriate
load models for voltage stability studies.
A PPENDIX
O N THE S TAHL’ S T HEORY AND THE RATE OF
C ONVERGENCE OF PAD É A PPROXIMANTS
Assume we are given a germ at point z0 of a multi-valued
analytic function f . Typically this means that we are given the
following power series,
S(z) ≃
∞
X
k=0
ck (z − z0 )k
(36)
This is the Taylor series expansion of the analytic function
developed at z = z0 and its radius of convergence might be
calculated via Cauchy–Hadamard formulae as,
1
= lim |ck |1/k .
R k→∞
(37)
Pn
Let Sn (z)= k=0 ck (z − z0 )k be a partial sum of the power
series and S∞ (z) the limit of Sn (z) as n → ∞. Then f can
be evaluated in the disk DR := {z ∈ C : |z − z0 | < R} as,
f (z) = S∞ (z) for |z − z0 | < R
(38)
Now suppose that f is a multi-valued analytic function on
the Riemann sphere C punctured at a finite set Σ = Σf of
the remarkable points of f at least one of which is a branch
point of f . Let Σ = {b1 , . . . , bp } and suppose z0 ∈
/ Σ. The
case when multi-valued function f is an algebraic function
is of great interest to us. Function f is an algebraic of
degree m if there exists an irreducible complex polynomial
P (z, w) in two complex variables z and w and of degree m
in w such that P (z, f (z)) ≡ 0 for z ∈
/ Σ. One can invoke
Cauchy’s argument principle to show that an algebraic function
(assuming m > 2) is also a multi-valued analytic function.
Recall that the essential idea of holomorphic embedding is
to analytically continue the germ of voltage phasors from
z = z0 = 0 where the power flow has a trivial stable solution
to z = 1. So hereafter we assume z0 = 0. By definition,
analyticity of f in the domain G := C \ Σ implies that a germ
of f given at the point z = 0 might be continued analytically
from z = 0 to each point z = a, a ∈
/ Σ, along every path γ
such that γ avoids points of the set Σ, i.e. γ ⊂ C \ Σ. Since
function f is multi-valued in G, two different paths, say γ1
and γ2 , such that γ1 , γ2 ⊂ G and both of γ1 and γ2 connect
the original point z = 0 with the end point z = a, might
lead to two different germs f1 and f2 of f at just the same
point z = a. In others words, one can obtain f1 (z) 6= f2 (z)
in some neighborhood of a. This is always the case when two
paths γ1 and γ2 form a closed curve γ = γ1 ∪ γ2 such that
it encircles exactly one branch point bj ∈ Σ of f . This is
an example of a nontrivial monodromy of the closed path γ.
Now suppose that we fix some simply connected subdomain
D of G such that {0, a} ⊂ D. Recall that simple connectivity
means that C \ D, the complement of D, is a connected set.
From this it follows immediately that for each path γ in D
its monodromy is trivial. Then from the classical monodromy
theorem it follows that for every two paths γ1 and γ2 which
both lead from z = 0 to z = a the corresponding germs f1
and f2 are identical, i.e. f1 (z) ≡ f2 (z) in some neighborhood
of the point z = a.
From now on we shall restrict our attention to the case of
algebraic functions only.
Given a germ at z = 0 of a multi-valued analytic function f
with a finite set Σ = {b1 , . . . , bp } of remarkable points,
one can evaluate the function f (z) for |z| < R, where
R = min{|bj |, j = 1, . . . , p}, via the equality f (z) = S∞ (z).
Since the disk DR = {|z| < R} is a simply connected domain,
from the monodromy theorem it follows that for each closed
path γ from the disk DR its monodromy is trivial.
There exists a powerful method to evaluate the analytic
function via its germ, given at the point z = 0, beyond the
boundary of the disk of convergence of the corresponding
power series. In this method which is also classical and is
based on the notion of continued fractions (to be more precise
continued C-fractions), we start from the given power series
and use the classical Viskovatov algorithm to obtain (under
some additional assumptions of nondegeneracy of a given
germ as it is generically always the case) the formal expansion,
(0)
S(z) ≃ c0 + c1 z + c2 z 2 + · · · ≃ c0 +
c1 z
(1)
1+
c1 z
≃ C(z)
(2)
1+
c1 z
1 + ...
(39)
This classical approach of evaluating an analytic function
from its germ was known, in some partial forms, since the
time of Jacobi and Gauss as continued fraction expansion
(or J-fraction expansion named after Jacobi). But at that time
it was used to expand only special functions, in particular
the hypergeometric functions. They had recognized that Jfraction expansions give the single valued continuation of a
germ of a multivalued analytic function from the origin into
an unknown domain which is much larger than the initial disk
of convergence.
Let us now consider a germ of an algebraic function f ,
developed at z = 0, and let Cn (z) be the n-th truncate of the
C-fraction (39), i.e.
(0)
c1 z
(40)
Cn (z) = c0 +
(1)
c1 z
1+
(2)
c1 z
1+
(n−2)
c1
z
..
.
(n−1)
1 + c1
z
Let C∞ (z) be the limit of the truncated Cn (z) as n → ∞.
Then a number of fundamental problems arise around the
equality (cf. (38))
f (z) = C∞ (z)
(41)
The main problems are as follows. In what domain D of the
complex variable z and in what sense the equality (41) holds
true? Since all Cn (z) functions are rational in z and thus
single-valued on C, the limit function C∞ (z) is also singlevalued. This is in contrast to that fact that the initial function
f is multi-valued. It is well-known that in general for each
n there is a finite number of the so-called spurious zero-pole
pairs of Cn (z) that do not correspond to any singularity of the
given function f (and neither correspond to a pole or a zero
of f ; see [24] and also [17, Chapter 2, § 2.2] and [26]). Such
pairs are usually referred to as ‘Froissart doublets’. As n tends
to infinity, pole and zero in such a pair come close to each
other. On one hand, they cancel each other as n → ∞, but
on the other, for a fixed n, they are distinct from each other
and as n → ∞ they are dense everywhere on the Riemann
sphere C. For a Riemann surface of genus 1 they make some
kind of ’winding of the torus’. By that reason there can not
be a pointwise convergence of Cn (z) to f (z) for z ∈ D, i.e.
the pointwise equality f (z) = C∞ (z) in D should not be
expected at all.
Recall once again that for an algebraic function f the
number of Froissart doublets is finite, i.e. is independent of
n, and depends on f only. For example, let f be given
√ by the
algebraic equation (1−z 2 )w2 −1 ≡ 0, i.e. f (z) = 1/ 1 − z 2 ,
and let us fix the germ at z = 0 by the equality f (0) = 1.
Then all zeros and poles of Cn (z) belong to the complement
R \ [−1, 1] of the closed segment [−1, 1] and its inverses has
the limit distribution on the segment [−1, 1] that coincide with
dx
1
, x ∈ [−1, 1]. There
Chebyshev measure given by √
π 1 − x2
are no Froissart doublets in that case and thus the equality
f (z) = C∞ (z) holds true pointwise for z ∈ D := C \ [−1, 1].
In contrast, the equality f (z) = S∞ (z) holds true only for the
unit disk, i.e. for |z| < 1.
The problem of equality in (41) f (z) = C∞ (z) for an
arbitrary multi-valued function with a finite set of branch
points Σ = {b1 , . . . , bp } was completely solved by H. Stahl
in 1985–1986 (see [29]– [33], [35] and also [20], [26]).
Given a germ6 f of a multi-valued analytic function f
with a finite number of branch points, the seminal Stahl’s
theorem7 gives the complete answer to the problem of limit
zero-pole distribution of the classical Padé approximants and
the equality of C∞ to f . The keystone of the Stahl’s theorem
is the existence of a unique so-called ‘maximal domain’ of
holomorphy of a given multi-valued function f , i.e. a domain
D = D(f ) ∋ 0 such that the given germ f can be continued as
holomorphic (i.e. analytic and single-valued) function from a
neighborhood of z = 0 to D (i.e. the function f is continued
analytically along each path that belongs to D). ‘Maximal’
means that ∂D is of ‘minimal capacity’ (with respect to the
point z = 0) among all compact sets ∂G such that G is a
domain, G ∋ 0 and f ∈ H (G). Such ‘maximal’ domain D
is unique up to a compact set of zero capacity. Compact set
S = S(f ) := ∂D is now called the ‘Stahl’s compact set’ or the
‘Stahl’s S-compact set’ and D is called the ‘Stahl’s domain’.
The crucial properties of S for the Stahl’s theorem to be true
are the following: the complement D = C \ S is a domain,
S consists of a finite number of analytic arcs, and finally S
possesses some special property of ‘symmetry’8.
From the Stahl’s theorem it follows that the limit points of
the zero and pole distributions of rational functions Cn (z)
as n → ∞ exists and coincides with a unique so-called
probability equilibrium measure for the Stahl’s compact set
S. From the numerical point of view this means, first, that
zeros and poles of Cn (z) are attracted as n → ∞ to the Stahl’s
compact set S. Second, they accumulate to each9 branch point
bk ∈ S ∩Σ of f with a density similar to that of the Chebyshev
dx
for S = [−1, 1] at the end points ±1.
measure π1 √1−x
2
On the Stahl’s compact set S there are also a finite number
of the so-called Chebotarev’s points that do not correspond
to any branch point of the initial function f and at these
points the equilibrium measure
of S has a zero density similar
√
to that of the measure 1 − x2 dx at the end points ±1.
These Chebotarev’s points are the transcendental parameters
6 That
is the convergent power series at the point z = z0 .
the reason of its very general character and the various subjects of
complex analysis involved into the proof of the theorem, it is sometimes
considered as ‘Stahl’s Theory’.
8 Compact sets of such type are usually referred to as ‘S-compact sets’ or
‘S-curves’, see [28], [27].
9 To be more precise, to each ‘active’ branch point. All the so-called
‘inactive’ branch points are hidden on the other ‘nonphysical’ sheets of the
Riemann surface of the given function; see [37].
7 By
of the problem, i.e. they can not be recovered from the branch
points of a given function over elementary functions but only
over transcendental functions. For example in the case of the
function f (z) = (b1 − z)1/3 (b2 − z)1/3 (b3 − z)−2/3 with the
three branch points b1 , b2 , b3 this means that the Chebotarev’s
point v is uniquely determined from the condition that the both
periods of the Abelian integral are purely imaginary:
Z
z
r
v − η dη
B3 (η) η
(42)
Q3
where B3 (η) := j=1 (bj − η). Chebotarev’s points, jointly
with the branch points of f , determine the Stahl’s twosheeted hyperelliptic Riemann surface associated with f . Thus
numerically in a neighborhood of a Chebotarev’s point, zeros
and poles of Cn (z) are very sparse compared to the number
n. It might be concluded that zeros and poles of Cn (z) as
n becomes large enough, eventually will recover numerically
the complete structure of the Stahl’s compact set S and the
Stahl’s domain D.
Finally in the Stahl’s theorem it is proved that the equality
f (z) = C∞ (z) holds true in the Stahl’s domain D = D(f )
not pointwise but ‘in capacity’. The convergence in capacity
inside the Stahl’s domain D (recall that f ∈ H (D)) means
that for every compact set K ⊂ D and for every small ε > 0
the following holds,
cap{z ∈ K : |f (z)−[n/n]f (z)| > ε > 0} → 0, n → ∞, z ∈ D
(43)
where cap(·), is the logarithmic capacity [36]. The only reason
for this specific mode of convergence is the existence of a finite
number of Froissart doublets. In fact the truncated Cn (z) of the
C-fraction in (38) gives a very good numerical approximation
of f (z) in all the points z of the Stahl’s domain up to a finite
number of ‘wandering’ Froissart doublets.
The rate of the convergence in (43) is completely characterized by the equality
cap
|(f − [n/n]f )(z)|1/n −→ e−2gS (z,0) , n → ∞, z ∈ D
(44)
where gS (z, 0) is the Green’s function of the domain D with a
logarithmic singularity at the point z = 0. gS (z, 0) is defined
in relation to a given branch point bj in the following way,
gS (z, 0) := Re
Z
z
bj
r
v − η dη
B3 (η) η
(45)
Thus the rate of convergence at a given point z ∈ D depends
on the value of the Green’s function gS (z, 0) for the domain
D ∋ 0 at that point. The closer the point z gets to the boundary
S := ∂D of D, the smaller the rate of convergence becomes.
In the disk DR the convergence of Cn (z) to f (z) is much
faster than the convergence of partial sums Sn (z). Finally
we should mention that in generic cases for an algebraic
function f given by polynomial equation P (z, f (z)) ≡ 0,
the corresponding Stahl’s compact set is stable under small
perturbations of complex coefficients of the polynomial P .
R EFERENCES
[1] A. Trias, “The Holomorphic Embedding Load Flow method,” Proceedings
of Power and Energy Society General Meeting, 22-26 July 2012.
[2] A.I. Markushevich, Theory of Functions of a Complex Variable, Translated by R.A. Silverman, 2nd Edition, American Mathematical Society,
2005.
[3] V.A. Venikov, V.A. Stroev, V.I. Idelchick, and V.I. Tarasov, “Estimation
of electrical power system steady-state stability in load flow calculations,”
IEEE Transactions on Power Apparatus and Systems, vol.94, no.3,
pp.1034-1041, May 1975.
[4] P.W. Sauer and M.A. Pai, “Power system steady-state stability and the
load-flow Jacobian,” IEEE Transactions on Power Systems, vol.5, no.4,
pp.1374-1383, Nov. 1990.
[5] I.A. Dobson, et al.,“Chapter 2: Basic Theoretical Concepts,” in Voltage
Stability Assessment: Concepts, Practices and Tools, IEEE-PES, 2002.
[6] A.A. Andronov, A.A. Vitt and S.E. Khaikin, Theory of Oscillators,
(Translation from Russian), Pergamon Press, 1966.
[7] V.I. Arnold, Geometrical Methods in the Theory of Ordinary Differential
Equations, (Translation from Russian), Springer-Verlag, 1983.
[8] I. Dobson, “The irrelevance of electric power system dynamics for the
loading margin to voltage collapse and its sensitivities,” Bulk power
system voltage phenomena III, voltage stability, security & control,
ECC/NSF workshop, Davos, Switzerland, August 1994
[9] S.S. Baghsorkhi, “Computing Saddle-Node and Limit-Induced Bifurcation Manifolds for Subtransmission and Transmission Wind Generation,”
Proceedings of the IEEE Power and Energy Society General Meeting,
Denver, CO, July 2015.
[10] P.W. Sauer, B.C. Lesieutre and M.A. Pai, “Maximum Loadability and
Voltage Stability in Power Systems,” International Journal of Electrical
Power and Energy Systems, vol. 15, pp.145-154 1993.
[11] J.S. Thorp and S.A. Naqavi, “Load-flow fractals draw clues to erratic
behavior,” IEEE Computer Applications in Power, pp. 59-62, Jan. 1997.
[12] D.K. Molzahn, “Application of Semidefinite Optimization Techniques to
Problems in Electric Power Systems,” Ph.D. Dissertation, University of
Wisconsin-Madison, Department of Electrical Engineering, August 2013.
[13] J.B. Lasserre, Moments, Positive Polynomials and Their Applications,
Imperial College Press, 2010.
[14] D. Cox, J. Little and D. O’Shea, Ideals, Varieties, and Algorithms, 3rd
Edition, Springer, 2006.
[15] J.-D. Boissonnat and M. Teillaud, Effective Computational Geometry for
Curves and Surfaces (Mathematics and Visualization), Springer, 2006.
[16] K. Kuhnle and E.W. Mayr, “Exponential space computation of Gröbner
bases,” Proceedings of the 1996 International Symposium on Symbolic
and Algebraic Computation (ISAAC96), pp 63-71. ACM Press, New York
(1996)
[17] G.A. Baker and P. Graves-Morris, Padé Approximants. Cambridge
University Press, 1996.
[18] E.M. Stein and R. Shakarchi, Complex Analysis. Princeton University
Press, 2003.
[19] S.P. Suetin, “Padé approximants and the effective analytic continuation
of a power series,” Russian Math. Surveys, 57:1 (2002), 43-141.
[20] A.I. Aptekarev, V.I. Buslaev, A. Martinez-Finkelshtein, S.P. Suetin,
“Padé approximants, continued fractions, and orthogonal polynomials,”
Russian Math. Surveys, 66:6 (2011), 1049-1131.
[21] S.S. Baghsorkhi, S.P. Suetin, “Embedding AC Power Flow with Voltage
Control in the Complex Plane: The Case of Analytic Continuation via
Padé Approximants”, 2015 , 9 pp., arXiv: 1504.03249
[22] M.K. Subramanian, F. Yang, and D. Tylavsky “PV bus modeling in
a holomorphically embedded power-flow formulation,” North American
Power Symposium (NAPS), pp. 1-6. 2013.
[23] G.E. Andrews and K. Eriksson, Integer Partitions. Cambridge University
Press, 2004.
[24] M. Froissart, “Approximation de Pade: application a la physique des
particules elementaires”, Recherche Cooperative sur Programme (RCP),
9, eds. Carmona, J., Froissart, M., Robinson, D.W., Ruelle, D., Centre
National de la Recherche Scientifique (CNRS), Strasbourg, 1969 pages
3, 1-13.
[25] A.A. Gonchar, “Rational Approximations of Analytic Functions”,
Sovrem. Probl. Mat., 1, Steklov Math. Inst., RAS, Moscow, 2003, 83106; Proc. Steklov Inst. Math., 272:, suppl. 2 (2011), S44-S57 pages.
[26] N.R. Ikonomov, R.K. Kovacheva, S.P. Suetin, “On the limit zero distribution of type I Hermite-Padé polynomials”, 2015, 67, arXiv: 1506.08031
pages 3.
[27] Kuijlaars, Arno B.J.; Silva, Guilherme L.F., “S-curves in polynomial
external fields”, J. Approx. Theory, 191 (2015), 1-37 pages 4.
[28] E.A. Rakhmanov, “Orthogonal polynomials and S-curves” (Recent advances in orthogonal polynomials, special functions and their applica-
tions), Contemp. Math., 578, Amer. Math. Soc., Providence, RI pages 4,
2012, 195-239.
[29] H. Stahl, “Extremal domains associated with an analytic function. I”,
Complex Variables Theory Appl., 4 (1985), 311-324 pages 3.
[30] H. Stahl, “Extremal domains associated with an analytic function. II”,
Complex Variables Theory Appl., 4 (1985), 325-338 pages.
[31] H. Stahl, “Structure of extremal domains associated with an analytic
function, Complex Variables Theory Appl., 4 (1985), 339-354 pages.
[32] H. Stahl, “Orthogonal polynomials with complex valued weight function. I”, Constr. approx., 2 (1986), 225-240 pages.
[33] H. Stahl, “Orthogonal polynomials with complex valued weight function. II”, Constr. approx., 2 (1986), 241-251 pages 3.
[34] H. Stahl, “Diagonal Padé approximants to hyperelliptic functions”, Ann.
Fac. Sci. Toulouse Math. (6), 1996, Special Issue, 121-193 pages.
[35] H. Stahl, “The convergence of Padé approximants to functions with
branch points”, J. Approx. Theory, 91:2 (1997), 139-204 pages 3.
[36] E. Saff and V. Totik, Logarithmic Potentials with External Fields.
Grundlehren der Matematischen Wissenschaften, 316, Springer, 1997.
[37] H. Stahl, “Sets of Minimal Capacity and Extremal Domains”,
arXiv:1205.3811, 2012, 112 pages.
| 3cs.SY
|
Delft University of Technology
Software Engineering Research Group
Technical Report Series
Applying and Combining Three Different
Aspect Mining Techniques
M. Ceccato, M. Marin, K. Mens, L. Moonen,
P. Tonella, and T. Tourwé
Report TUD-SERG-2006-002
SERG
TUD-SERG-2006-002
Published, produced and distributed by:
Software Engineering Research Group
Department of Software Technology
Faculty of Electrical Engineering, Mathematics and Computer Science
Delft University of Technology
Mekelweg 4
2628 CD Delft
The Netherlands
ISSN 1872-5392
Software Engineering Research Group Technical Reports:
http://www.se.ewi.tudelft.nl/techreports/
For more information about the Software Engineering Research Group:
http://www.se.ewi.tudelft.nl/
c copyright 2006, Software Engineering Research Group, Department of Software Technology, Faculty
of Electrical Engineering, Mathematics and Computer Science, Delft University of Technology. All rights
reserved. No part of this series may be reproduced in any form or by any means without prior written
permission of the publisher.
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
Applying and Combining Three Different
Aspect Mining Techniques
M. Ceccato1 , M. Marin2 , K. Mens3 , L. Moonen2,4 , P. Tonella1 , and T. Tourwé4
1
ITC-irst, Trento, Italy
Delft University, The Netherlands
3
Université catholique de Louvain, Belgium
4
CWI, The Netherlands
[email protected], [email protected], [email protected],
[email protected], [email protected], [email protected]
2
Abstract. Understanding a software system at source-code level requires understanding the different concerns that it addresses, which in
turn requires a way to identify these concerns in the source code. Whereas
some concerns are explicitly represented by program entities (like classes,
methods and variables) and thus are easy to identify, crosscutting concerns are not captured by a single program entity but are scattered over
many program entities and are tangled with the other concerns. Because
of their crosscutting nature, such crosscutting concerns are difficult to
identify, and reduce the understandability of the system as a whole.
In this paper, we report on a combined experiment in which we try to
identify crosscutting concerns in the JHotDraw framework automatically.
We first apply three independently developed aspect mining techniques
to JHotDraw and evaluate and compare their results. Based on this analysis, we present three interesting combinations of these three techniques,
and show how these combinations provide a more complete coverage of
the detected concerns as compared to the original techniques individually. Our results are a first step towards improving the understandability
of a system that contains crosscutting concerns, and can be used as a
basis for refactoring the identified crosscutting concerns into aspects.
1
Introduction
The increasing popularity of aspect-oriented software development (AOSD) is
largely due to the fact that it recognises that some concerns cannot be captured
adequately using the abstraction mechanisms provided by traditional programming languages. Several examples of such crosscutting concerns have been identified, ranging from simple ones such as logging, to more complex ones such as
transaction management [1] and exception handling [2, 3].
An important problem with such crosscutting concerns is that they affect the
understandability of the software system, and as a result reduce its evolvability
and maintainability. First of all, crosscutting concerns are difficult to understand,
because their implementation can be scattered over many different packages,
TUD-SERG-2006-002
1
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
II
classes and methods. Second, in the presence of crosscutting concerns, ordinary
concerns become harder to understand as well, because they get tangled with
the crosscutting ones: particular classes and methods do not only deal with the
primary concern they address, but also may need to take into account some
secondary, crosscutting concerns.
Several authors have presented automated code mining techniques, generally
referred to as aspect mining techniques, that are able to identify crosscutting
concerns in the source code [4]. The goal of these techniques is to provide an
overview of the source-code entities that play a role in a particular crosscutting concern. This not only improves the understandability of the concern in
particular and of the software in general, but also provides a first step in the
migration towards applying aspect-oriented software development techniques.
However, since the research field is still in its infancy, very few experiments have
been conducted on real-world case studies, comparisons of different techniques
are lacking, and no agreed-upon benchmark is available that allows to evaluate
the existing techniques.
This paper reports on an experiment involving three independently developed aspect mining techniques: fan-in analysis [5, 6], identifier analysis [7, 8] and
dynamic analysis [9]. In the experiment, each of these techniques is applied to
the same case study: the JHotDraw graphical editor framework. The goal of the
experiment is not to identify the “best” aspect mining technique, but rather to
mutually compare the individual techniques and assess their major strengths
and weaknesses. Additionally, by identifying where the techniques overlap and
where they are complementary, the experiment allows us to propose interesting
combinations and to apply these combinations on the same benchmark to verify
whether they actually perform better.
The JHotDraw framework which we selected as benchmark case was originally developed to illustrate good use of object-oriented design patterns [10] in
Java programs. This implies that the case study has been well-designed and that
care has been taken to cleanly separate concerns and make it as understandable
as possible. Nevertheless, JHotDraw exposes some of the modularisation limitations present even in well-designed systems, and contains some quite interesting
crosscutting concerns.
The contributions of this paper can be summarised as follows:
– We provide an overview of the major strengths and weaknesses of three aspect mining techniques. This information is valuable for developers using
these techniques, as it can help them choosing a technique that suits their
needs. Other aspect mining researchers can take this information into account to compare their techniques to ours, or to fine-tune our techniques;
– We discuss how the individual techniques can be combined in order to perform better, and validate whether this is indeed the case by applying the
combined techniques on the same benchmark application and comparing
the results;
– We present a list of all crosscutting concerns that the three techniques identified in the JHotDraw framework. Such information is valuable for other
2
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
III
interface A {
public void m();
}
class B implements A {
public void m() {};
}
class C1 extends B {
public void m() {};
}
class C2 extends B {
public void m() { super.m();};
}
class D {
void f1(A a) { a.m(); }
void f2(B b) { b.m(); }
void f3(C1 c) { c.m(); }
}
Fig. 1. Various (polymorphic) method calls.
aspect mining researchers who want to validate their techniques, and might
lead to JHotDraw becoming a de-facto benchmark for aspect mining techniques;
The paper is structured as follows. Section 2 introduces the necessary background concepts required to understand the three aspect mining techniques explained in Section 3. Section 4 presents the results of applying each technique on
the common benchmark, while Section 5 uses these results for discussing the benefits and drawbacks of each technique with respect to the others. Based on this
discussion, Section 6 presents useful combinations of the techniques, and reports
on the experience of applying such combinations on the benchmark application.
Section 7 presents our conclusions. For an overview of related work concerning
aspect mining, we refer to the papers discussing the individual techniques [5–9]
and to an initial survey on aspect mining [4].
2
2.1
Background concepts
Fan-in
The fan-in metric, as defined by Henderson-Sellers, counts the number of locations from which control is passed into a module [11]. In the context of objectorientation, the module-type to which this metric is applied is the method. We
define the fan-in of a method M as the number of distinct method bodies that
can invoke M . Because of polymorphism, one call site can affect the fan-in of
several methods: a call to method M contributes to the fan-in of M , but also to
all methods refined by M , as well as to all methods that are refining M [6].
TUD-SERG-2006-002
3
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
IV
Method
A.m
B.m
C1.m
C2.m
Potential
D.f1, D.f2,
D.f1, D.f2,
D.f1, D.f2,
D.f1, D.f2
callers
Fan-in
D.f3
3
D.f3, C2.m
4
D.f3
3
2
Fig. 2. Fan-in values for program in Figure 1.
As an example, Figure 2 shows the calculated fan-in for the methods named
m in the program of Figure 1. Note that D.f 3 is reported among the potential
callers of B.m, even though this situation cannot actually occur at run-time.
However, the resulting effect of having higher fan-in values reported for methods
in super-classes is arguably positive for the purpose of the present analysis, as it
emphasizes the concern implemented by the super-class method, which generally
is addressed by its overriding methods as well.
2.2
Concept analysis
Formal concept analysis (FCA) [12] is a branch of lattice theory that can be used
to identify meaningful groupings of elements that have common properties.5
Programming lang. object-oriented functional logic static typing dynamic typing
√
√
Java
√
√
Smalltalk
√
√
C++
√
√
Scheme
√
√
Prolog
Table 1. Programming languages and their supported programming paradigms.
FCA takes as input a so-called context, which consists of a (potentially large,
but finite) set of elements E, a set of properties P on those elements, and a
Boolean incidence relation T between E and P . An example of such a context is
given in √
Table 1, which relates different programming languages and properties.
A mark in a table cell means that the element (programming language) in the
corresponding row has the property of the corresponding column.
Starting from such a context, FCA determines maximal groups of elements
and properties, called concepts, such that each element of the group shares the
properties, every property of the group holds for all of its elements, no other
5
4
We use the terms element and property instead of object and attribute used in
traditional FCA literature, because these latter terms have a very specific meaning
in OO software development.
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
V
element outside the group has those same properties, nor does any property
outside the group hold for all elements in the group.
√ Intuitively, a concept corresponds to a maximal ‘rectangle’ containing only
marks in the table, modulo
any permutation of the table’s rows and columns.
Formally, the starting context is a triple (E, P, T ), where T ⊆ E × P is a
binary relation between the set of all elements E and the set of all considered
element properties P . A concept c is defined as a pair of sets (X, Y ) such that:
X = {e ∈ E | ∀p ∈ Y : (e, p) ∈ T }
(1)
Y = {p ∈ P | ∀e ∈ X : (e, p) ∈ T }
(2)
where X is said to be the extent of the concept (Ext[c]) and Y is said to be
its intent (Int[c]). It should be noticed that the definition above is not “constructive”, being mutually recursive between X and Y . However, given a pair
(X, Y ), it allows deciding whether it is a concept or not. FCA algorithms provide
constructive methods to determine all pairs (X, Y ) satisfying the constraints (1)
and (2).
{}
{Java, Smalltalk, C++, Scheme, Prolog}
{OO}
{dynamic typing}
{Java, C++, Smalltalk}
{Scheme, Prolog, Smalltalk}
{static typing, OO}
{dyn. typing, OO}
{dyn. typing, funct.}
{dyn. typing, logic}
{Java, C++}
{Smalltalk}
{Scheme}
{Prolog}
{OO, funct., logic, static typing, dyn. typing}
{}
Fig. 3. The concept lattice for Table 1.
The containment relationship between concept extents (or, equivalently, intents) defines a partial order over the set of all concepts, which can be shown to
be a lattice [12]. Figure 3 shows the concept lattice corresponding to Table 1.
The lattice’s bottom concept contains those elements that have all properties.
Since there is no such programming language in our example, that concept contains no elements (its extent is empty). Similarly, the top concept contains those
properties that hold for all elements. Again, there is no such property (the concept’s intent is empty). Other concepts represent related groups of programming
languages, such as the concept ({Java, C++}, {static typing, OO }), which
groups all statically-typed object-oriented languages, a sub-concept of all OO
languages. Intuitively, the sub-concept relationship can thus be interpreted as a
specialization of more general notions. Elements (resp. properties) in boldface
TUD-SERG-2006-002
5
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
VI
are those that are most concept-specific, being attached to the largest lower
bound (resp. least upper bound) concept. When using the so-called sparse labeling of the concept lattice, only these boldface labels are retained, without loss
of information.
More precisely, when using sparse labeling, a node c is marked with an element
e ∈ Ext[c] only if it is associated with the most specific (i.e., lowest) concept c
having e in the extent; a node c is marked with a property p ∈ Int[c] only if it is
associated with the most general (i.e., highest) concept c having p in its intent.
The (unique) node of a lattice L marked with a given element e is thus:
γ(e) = inf{c ∈ L | e ∈ Ext[c]}
(3)
where inf gives the infimum (largest lower bound) of a set of concepts. Similarly,
the unique lattice node marked with a given property p is:
µ(p) = sup{c ∈ L | p ∈ Int[c]}
(4)
where sup gives the supremum (least upper bound) of a set of concepts. The set
of elements in the extent of a lattice node c can then be computed as the set
of all elements at or below c, while the set of properties in its intent are those
marking c or any node above c.
The labeling introduced by the functions µ and γ give the most specific
concept for a given element (resp. property). Thus, with sparse labeling, the
elements and properties that label a given concept are those that characterize it
most specifically. Sometimes it is convenient to get the labels of a given concept
through the following functions:
α(c) = {p ∈ P | µ(p) = c}
(5)
β(c) = {e ∈ E | γ(e) = c}
(6)
α(c) gives the set of properties labeling a concept c, while β(c) gives the concept’s
elements, according to the sparse labeling.
2.3
Terminology
We conclude this background section by introducing some terminology that will
be used throughout the remainder of this paper.
A concern is a collection of related source-code entities, such as classes, methods, statements or expressions, that implement a particular functionality or
feature of the application. A crosscutting concern is a concern whose entities
are not captured into a single localised abstraction, but are scattered over
many different locations and tangled with other concerns.
A (concern) seed is a single source-code entity, such as a method, or a collection of such entities, that strongly connotes a crosscutting concern. It offers
a starting point for further exploration and understanding the whole extent
of that concern’s implementation.
6
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
VII
A candidate seed is identified by an automated aspect mining technique as a
potential concern seed but is not yet confirmed to be an actual concern seed
or rather a false positive.
Seed expansion is the manual or automated process of completing the set
of source-code entities constituting a seed into the entire set of source-code
entities of which the crosscutting concern corresponding to that seed consists.
3
The three aspect mining techniques
In this section, we give a brief overview of three techniques, developed independently by different research groups, that support the automated discovery of
crosscutting concerns in the source code of a software system that is written in
a non aspect-oriented way.
3.1
Fan-in analysis
Crosscutting functionality can occur at different levels of modularity. Classes,
for instance, can assimilate new concerns by implementing multiple interfaces or
by implementing new methods specific to super-imposed roles. At the method
level, crosscutting in many cases resides in calls to methods that address a different concern than the core logic of the caller. Typical examples include logging,
tracing, pre- and post-condition checks, and exception handling. It is exactly
this type of crosscutting that fan-in analysis tries to capture.
When we study the mechanics of AOSD, we see that it employs the so-called
advice construct to eliminate crosscutting at method level. This construct is used
to acquire control of program execution and to add crosscutting functionality to
methods without an explicit invocation from those methods. Rather, the crosscutting functionality is isolated in a separate module, called aspect, and woven
with the method implicitly based on the advice specification.
Fan-in analysis reverses this line of reasoning and looks for crosscutting
functionality that is explicitly invoked from many different methods scattered
throughout the code. The hypothesis is that the amount of calls to a method
implementing this crosscutting functionality (fan-in) is a good measure for the
importance and scattering of the discovered concern.
To perform the fan-in analysis, a fan-in metric was implemented as a plug-in
for the Eclipse platform6 , and integrated it into an iterative process that consists
of three steps:
1. Automatic computation of the fan-in metric for all methods in the investigated system.
2. Filtering of the results from the previous step by
– eliminating all methods with fan-in values below a chosen threshold (in
the experiment, a threshold of 10 was used);
6
http://swerl.tudelft.nl/view/AMR/FINT
TUD-SERG-2006-002
7
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
VIII
– eliminating the accessor methods (methods whose signature matches a
get* /set* pattern and whose implementation only returns or sets a reference);
– eliminating utility methods, like toString() and collection manipulation methods, from the remaining subset.
3. (Partially automated) analysis of the methods in the resulting, filtered set by
exploring the callers, call sites, naming convention used, the implementation
and the comments in the source code.
Besides code exploration, the tool supports automatic recognition of a number of relations between the callers of a method, such as common roles,
consistent call positions, etc.
The result of the fan-in analysis is a set of candidate seeds, represented as methods with high fan-in.
3.2
Identifier analysis
In the absence of designated language constructs for aspects, naming conventions are the primary means for programmers to associate related but distant
program entities. This is especially the case for object-oriented programming,
where polymorphism allows methods belonging to different classes to have the
same signature, where it is good practice to use intention-revealing names [13],
and where design and other programming patterns provide a common vocabulary known by many programmers.
Identifier analysis relies on this assumption and identifies candidate seeds
by grouping program entities with similar names. More specifically, it applies
FCA with as elements all classes and methods in the analyzed program (except
those that generate too much noise in the results, like test classes and accessor
methods), and as properties the identifiers associated with those classes and
methods.
The identifiers associated with a method or class are computed by splitting
up its name based on where capitals appear in it. For example, a method named
createUndoActivity yields three identifiers create, undo and activity. In addition, we apply the Porter stemming algorithm [14] to make sure that identifiers
with the same root form (like undo and undoable) are mapped to one single representative identifier or ‘stem’. It is these stems that are used as properties for
the concept analysis.
The FCA algorithm then groups entities with the same identifiers. When such
a group contains a certain minimum number of elements (in the experiment, a
threshold of 4 was used) and the entities contained in it cut across multiple class
hierarchies, the group is considered a candidate seed. The only remaining but
most difficult task is that of deciding manually whether a candidate seed is a
real seed or a false positive. To help the developer in this last task, the DelfSTof
source-code mining tool presents the concepts in such a way that they can be
browsed easily by a software engineer and so that he or she can readily access
the code of the classes and methods belonging to a discovered seed.
8
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
IX
3.3
Dynamic analysis
Formal concept analysis has been used to locate ‘features’ in procedural programs [15]. In that work, the goal was to identify the computational units (procedures) that specifically implement a feature (i.e., requirement) of interest. Execution traces obtained by running the program under given scenarios provided
the input data (dynamic analysis).
In a similar way, dynamic analysis can be used to locate aspects in program
code [9] according to the following procedure. Execution traces are obtained
by running an instrumented version of the program under analysis, for a set of
scenarios (use-cases). The relationship between execution traces and executed
computational units (methods) is subjected to concept analysis. The execution
traces associated with the use-cases are the elements of the concept analysis
context, while the executed methods are the properties. In the resulting concept
lattice (with sparse labeling), the use-case specific concepts are those labeled by
at least one trace for some use-case (i.e. α contains at least one element), while
the concepts with zero or more properties as labels (those with an empty α) are
regarded as generic concepts. Thus, use-case specific concepts are a subset of the
generic ones.
Both use-case specific concepts and generic concepts carry information potentially useful for aspect mining, since they group specific methods that are
always executed under the same scenarios. When the methods that label one
such concept (using the sparse labeling) crosscut the principal decomposition, a
candidate aspect is determined.
Formally, let C be the set of all the concepts and let Cs be the set of use-case
specific concepts (|α(c)| > 0). A concept c is considered a candidate seed iff:
Scattering: ∃p, p0 ∈ β(c) | pref (p) 6= pref (p0 )
Tangling: ∃p ∈ β(c), ∃c0 ∈ Ω, ∃p0 ∈ β(c0 ) | c 6= c0 ∧ pref (p) = pref (p0 )
where Ω = Cs for the use-case specific seeds, while Ω = C for the generic seeds.
The first condition (scattering) requires that more than one class contributes to
the functionality associated with the given concept (pref(p) is the fully scoped
name of the class containing the method p). The second condition (tangling)
requires that the same class addresses more than one concern.
In summary, a concept is a candidate seed if: (1) scattering: more than one
class contributes to the functionality associated with the given concept; (2) tangling: the class itself addresses more than one concern.
The first condition alone is typically not sufficient to identify crosscutting
concerns, since it is possible that a given functionality is allocated to several
modularized units without being tangled with other functionalities. In fact, it
might be decomposed into sub-functionalities, each assigned to a distinct module.
It is only when the modules specifically involved in a functionality contribute
to other functionalities as well (i.e. the second condition) that crosscutting is
detected, hinting for a candidate seed.
TUD-SERG-2006-002
9
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
X
Concern type
# Seed’s description
Consistent behavior
4 Methods implementing the consistent behavior
shared by different callers, such as checking and refreshing figures/views that have been affected by the
execution of a command.
4 Method implementing a contract that needs to be enforced, such as checking the reference to the editor’s
active view before executing a command.
Contract enforcement
Undo
1 Methods checking whether a command is undoable/redoable and the undo method in the superclass, which is invoked from the overriding methods
in subclasses.
Persistence and resurrection 1 Methods implementing functionality common to persistent elements, such as read/write operations for
primitive types wrappers (e.g., Double, Integer, etc.)
which are referenced by the scattered implementations of persistence/resurrection.
Command design pattern
Observer design pattern
Composite design pattern
Decorator design pattern
Adapter design pattern
1 The execute method in the command classes and command constructors.
1 The observers’ manipulation methods and notify
methods in classes acting as subject.
2 The composite’s methods for manipulating child components, such as adding a new child.
1 Methods in the decorator that pass the calls on to the
decorated components.
1 Methods that manipulate the reference from the
adapter (Handle) to the adaptee (Figure).
Table 2. Summary of the results of the fan-in analysis experiment.
4
Results of the aspect mining
In this section, we present the results of applying each technique to version
5.4b1 of JHotDraw, a Java program with approximately 18,000 non-commented
lines of code and around 2800 methods. We mutually compare the results of
the techniques, and discuss the limitations of each technique as well as their
complementarity.
4.1
The fan-in analysis experiment
As described in Subsection 3.1, fan-in analysis first performs a number of successive steps to filter the methods in the analyzed system. The threshold-based
filtering, which selects methods with high fan-in values, kept around 7% of the
total number of methods. The filters for accessors and utility methods eliminated
around half of the remaining methods. In the remaining subset, more than half
of the methods (52%) were categorized as seeds, based on manual analysis.
10
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XI
Table 2 gives an overview of the types of crosscutting concerns that were
identified and the seeds that led to their identification. Several of these concern
types, such as consistent behavior or contract enforcement [16], have more than
one instance in JHotDraw; that is, multiple unrelated (crosscutting) concerns
exist that conform to the same general description. For example, one instance of
contract enforcement checks a priori conditions to a command’s execution, while
another instance verifies common requirements for activating drawing tools. The
number of different instances that were detected is indicated in the # column.
We distinguish three different ways in which the fan-in metric can be associated with the crosscutting structure of a concern implementation (also indicated
in Table 2):
1. The crosscutting functionality is implemented through a method and the
crosscutting behavior resides in the explicit calls to this method. Examples
in this category include consistent behavior and contract enforcement.
2. The implementation of the crosscutting concern is scattered throughout the
system, but makes use of a common functionality. The crosscutting resides
in the call sites, and can be detected by looking at the similarities between
the calling contexts and/or the callers. Examples of concerns in this category
are persistence and undo [6].
3. The methods reported by the fan-in analysis are part of the roles superimposed to classes that participate in the implementation of a design pattern.
Many of these roles have specific methods associated to them: the subject
role in an Observer design pattern is responsible to notify and manage the
observer objects, while the composite role defines specific methods for manipulating child components. In general, establishing a relation between these
seed-methods and the complete concern to which they appertain might require a better familiarity of the human analyzer with the code being explored,
than for the previous two categories. However, many of these patterns are
well-known and have a clear defined structure, which eases their recognition [17].
For more details regarding fan-in analysis and a complete discussion of the JHotDraw results, we refer to [6].
4.2
The identifier analysis experiment
Applying the identifier analysis technique of Subsection 3.2 on JHotDraw yielded
230 concepts and took about 31 seconds when using a threshold of 4 for the
minimum number of elements in a concept. With a threshold of 10, the number
of concepts produced was significantly less: only 100 concepts remained after
filtering, for a similar execution time.7 In both cases, 2193 elements and 507
properties were considered. It is a good sign that the number of properties is
7
Whereas the threshold of 4 was chosen arbitrarily, the threshold of 10 was determined
experimentally: below that threshold the amount of concepts that were regarded as
noise was significantly higher than above the threshold.
TUD-SERG-2006-002
11
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XII
Crosscutting concern Concept(s)
#elements Some elements
Observer
change(d)
check
listener
release
command executed
execut(abl)e
undo(able)
redo(able)
visit
file
storable
load
register
67
14
65
12
4
51
53
14
12
15
5
8
7
figureChanged(e)
checkDamage()
createDesktopListener()
...
commandExecuted(...)
commandExecutable(...)
createUndoActivity()
redo()
visit(FigureVisitor)
registerFileFilters(c)
readStorable()
loadRegisteredImages
loadRegisteredImages
draw
move
112
36
draw(g)
moveBy(x,y)
moveSelection(dx,dy)
5
iterator(), listIterator(), . . .
Command execution
Undo
Visitor
Persistence
Drawing figures
Moving figures
Iterating over collections iterator
Table 3. Selection of results of the identifier analysis experiment.
significantly smaller than the total number of elements considered, as it implies
that there is quite some overlap in the identifiers of the different source-code
entities, which was one of the premisses of the identifier analysis technique.
The manual part of the experiment, i.e. deciding which concepts were real
seeds, was much more time-consuming. Overall, this took about three days for
the experiment with threshold 4, where 230 seed candidates needed to be investigated. For each of the discovered concepts, the code of the entities in its
extent had to be inspected to decide whether (most of) these entities addressed a
similar concern. Other than allowing to browse the source code of the elements
in the extent of a concept, the DelfSTof code mining tool provided no direct
support for this.
Table 3 presents some of the seeds discovered by manually analyzing the
classes and methods belonging to the extent of the concepts produced by the
FCA algorithm. The first column names the concern, the second column shows
the identifiers shared by the elements belonging to the concept(s) corresponding
to that concern. The third column shows the size of the extent for each concept.
Finally, for illustration purposes, the fourth column shows some program entities
appearing in the extent of the discovered concepts.
Out of 230 candidate seeds, 41 seeds were retained, when using a threshold of
4 for the minimum number of elements in a concept. These discovered concerns
were classified in three different categories:
1. Some of these concerns looked like aspects in the more traditional sense (e.g.,
observer, undo and persistence).
12
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XIII
2. Many other concerns seemed to represent a crosscutting functionality that
was part of the business logic (e.g., drawing figures, moving figures). The
distinction between these two first categories was somewhat subjective, however.
3. Three Java-specific concerns were discovered (e.g., iterating over collections)
that are difficult to factor out into an aspect because they rely on or extend
specific Java code libraries.
4.3
The dynamic analysis experiment
The dynamic analysis technique of Subsection 3.3 is supported by the Dynamo
aspect mining tool8 . The first step required by Dynamo is the definition of a
set of use-cases. To accomplish this task, the documentation associated with
the main functionalities of JHotDraw was used to define a use-case for each
functionality described in the documentation. Amongst others, a use-case was
created to draw a rectangle, one to draw a line using the scribble tool, one
to create a connector between two existing figures, one to attach a URL to
a graphical element, and so on. In total, 27 use-cases were obtained. When
executed they exercised 1262 methods belonging to JHotDraw classes, so that
the initial context for the concept analysis algorithm contained 27 elements and
1262 properties. The resulting concept lattice contained 1514 nodes.
Crosscutting concern
Concepts Methods
Undo
Bring to front
Send to back
Connect text
2
1
1
1
36
3
3
18
Persistence
Manage handles
Manage figure change event
Move figure
Command executability
Connect figures
Figure observer
Add text
Add URL to figure
Manage figures outside drawing
Get attribute
Set attribute
Manage view rectangle
Visitor
1
4
3
1
1
1
4
1
1
1
1
1
1
1
30
60
8
7
25
55
11
26
10
2
2
2
2
6
Table 4. Summary of the results of the dynamic analysis experiment.
8
Available from http://star.itc.it/dynamo/ under GPL.
TUD-SERG-2006-002
13
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XIV
Among the concepts in the lattice, 11 satisfied the crosscutting conditions
(scattering and tangling), described in Section 3, for the use-case specific concepts, while 56 (including the 11 above) satisfied the conditions for the generic
concepts. Next, both the use-case specific and generic concepts were revisited
manually, to determine which ones could be regarded as plausible seeds and
which ones should be considered false positives. The criterion followed in this
assessment was the following: a concept satisfying the crosscutting conditions is
considered a seed if
– it can be associated to a single, well-identified functionality (this usually
accounts for the possibility to give it a short description that labels it), and
– some of the classes involved in such a functionality have a different primary
responsibility (indicating crosscutting with respect to the principal decomposition).
Of course, due to the nature of crosscutting concerns and the related design
decisions, some level of subjectivity still remains (as is the case for the other
techniques).
In the end, the list of candidate seeds shown in Table 4 was obtained. The
four topmost concerns are use-case specific. As apparent from the second column
of the table, and as was the case for the identifier analysis experiment, some
crosscutting concerns were detected by multiple concepts. In total, among the 56
generic concepts satisfying the crosscutting conditions, 24 concepts were judged
to be associated with 18 crosscutting concerns.
The methods associated with each candidate seed (counted in the last column
of Table 4) are indicative of the “aspectizable” functionality. Although they may
be not the complete list (dynamic analysis is partial) and may contain false
positives, they represent a good starting point for a refactoring intervention
aimed at migrating the application to AOSD.
5
Comparing the results
In this section we discuss some selected concerns that were identified by the
different techniques. We selected some concerns that were detected by all three
techniques, as well as a representative set of concerns that were detected by some
techniques but not by others. This allows us to clearly pinpoint the strengths
and weaknesses of each individual technique.
5.1
Selected concerns
Table 5 summarises the concerns we selected. The first column names the concern. The other columns show by what technique(s) the concern was discovered:
if a technique discovered the concern, we put a + sign in the corresponding
column, otherwise a - sign is in the table.
14
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XV
Concern
Observer
Undo
Persistence
Consistent behavior /
Contract enforcement
Command execution
Bring to front / Send to back
Manage handles
Move Figures
Fan-In
Analysis
+
+
+
+
+
+ (discarded)
Identifier Dynamic
Analysis Analysis
+
+
+
+
+
+
+
+
+
+
+
+
+
Table 5. A selection of detected concerns in JHotDraw.
Observer The Observer design pattern is an example of a concern reported by
all techniques. Other examples include Command execution, Undo functionality
and Persistence, whose implementation in JHotDraw is described in [6]. Their
identification should come as no surprise, because they correspond to well-known
aspects, frequently mentioned in AOSD literature, or to functionalities for which
an AOSD implementation looks quite natural.
Concerns identified by all three techniques are probably the best starting
point for migrating a given application to AOSD, because developers can be
quite confident that the concern is very likely to be an aspect. However, the
fact that only four of such concerns were discovered, stresses the need for an
approach that combines the strengths of different techniques.
Contract enforcement / consistent behavior The contract enforcement
and consistent behavior concerns [16] generally describe common functionality
required from, or imposed on, the participants in a given context, such as a
specific pre-condition check on certain methods in a class hierarchy. An example from the JHotDraw case is the Command hierarchy for which the execute
methods contain code to ensure the pre-condition that an ‘active view’ reference
exists (is not null).
We classify these concerns as a combination of contract enforcement and
consistent behavior since these types often have very similar implementations,
and choosing a particular type depends mainly on the context and on (personal)
interpretation.
Fan-in analysis is particularly suited to address this kind of scattered, crosscutting functionalities, which involve a large number of calls to the same method,
while the other two techniques potentially miss it. In fact, contract enforcement
and consistent behavior are usually associated with method calls that occur in
every execution scenario, so that they cannot be discriminated by any specific
use-case. On the other hand, identifier analysis will miss those cases where the
methods that enforce a given contract or ensure consistent behavior do not share
a common naming scheme.
TUD-SERG-2006-002
15
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XVI
Command execution This concern deals with the executability and the actual execution of objects whose class belongs to the Command hierarchy. Identifier analysis identified a concept which contains exactly the execute methods
in the Command hierarchy. Dynamic analysis identified the classes containing
isExecutable methods. Indeed, the execute methods all have the same name and
manual inspection showed they exhibit similar behavior: they nearly all make
a super call to an execute method, invoke a checkDamage method and (though
not always) invoke a setUndoAcivity and getUndoActivity method. A similar
argument can be made for isExecutable.
Hence, whereas identifier and dynamic analysis may not detect the more
generic Contract enforcement / Consistent behavior aspect directly, they can
identify some locations (pointcuts) where potentially such an aspect could be
introduced.
Bring to front / Send to back The functionality associated with this concern
consists of the possibility to bring figures to the front or send them to the back
of an image. When exercised, it executes specific methods that have a low fan-in,
hence they were not detected by fan-in analysis. Identifier analysis also missed
them, because there were not enough methods with a sufficiently similar name
to surpass the threshold. Hence, dynamic analysis is the only technique that
identified this concern. This example is a good representative of crosscutting
concerns that are reported only by dynamic analysis: whenever the methods
involved in a functionality are not characterized by a unifying naming scheme
(or there are not enough of them), neither do they have high fan-in, the other
two techniques are likely to fail.
Manage handles A crosscutting functionality is responsible for managing the
handles associated with the graphical elements. Such handles support interactive
operations, such as resizing of an element, conducted by clicking on the handle
and dragging the mouse. This seed is interesting because it is detected by dynamic analysis and by identifier analysis, but in different ways. Identifier analysis
detects this concern based on the presence of the word ‘handle’ in identifiers.
Consequently, it misses methods such as north(), south(), east(), west(),
which are clearly related to this concern, but do not share the lexicon with the
others. On the other hand, dynamic analysis reports both the latter methods
and (some of) those containing the word ‘handle’. However, since not all possible handle interactions have been exercised, the output of dynamic analysis is
partial and does not include all the methods reported by identifier analysis.
The manage handles concern was missed by the fan-in analysis because the
calls are too specific: they are similar but different calls instead of one single
called method with a high fan-in.
Moving figures The three techniques discard concerns on different bases: some
of the concerns are filtered automatically while others are excluded manually.
16
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XVII
The move figures concern, seeded by the moveBy method in the Figure classes,
is one example where different, subjective decisions can be made depending on
whether the concept is classified either as a candidate aspect or as part of the
principal decomposition. The moveBy methods allow to move a figure with a
given offset. The team which used fan-in analysis argued that the original design
seems to consider this functionality as part of a Figure’s core logic. The other
two teams considered it as part of a crosscutting functionality and included it
in the list of reported seeds.
This example highlights the difficulty of deciding objectively on what is and
what is not an aspect and corroborates our choice to conduct a qualitative,
instead of a quantitative, comparison.
5.2
Limitations
As a consequence of applying each technique to the same case, some of the
limitations of the respective techniques have become obvious. For example, we
obtained a better idea of potential ‘false negatives’, i.e. concerns that were not
identified by a particular technique but that were identified by another. Below, we summarise some of the discovered limitations. In the next section we
then describe how to partly overcome these limitations by combining different
techniques.
Fan-in analysis mainly addresses crosscutting concerns that are largely scattered and that have a significant impact on the modularity of the system. The
downside of this characteristic is that concerns with a small code footprint and
thus with low fan-in values associated, will be missed. For example, the identification of Observer design pattern instances is dependent on the number of
classes implementing the observer role. These classes contain calls to specific
methods in the subject class for registering as listeners to the subject’s changes.
The number of observer classes will determine to a large extent the number of
calls to the registration method in the subject role. A collateral effect is the
anticipated unsuitability of the technique for analysing small case studies.
Identifier analysis tends to produce a lot of detailed results. However, these
results typically contain too much noise (false positives), so a more effective filtering of the discovered concepts, as well as of the elements inside those concepts,
is needed. In addition, the discovered concepts are often incomplete, in the sense
that they do not completely “cover” an aspect or crosscutting concern. Often,
more than one concept is needed to describe a single concern, as was the case
for the Observer aspect. The individual concepts themselves may also need to
be completed with additional elements that are not contained in those concepts.
This was the case for the Undo aspect: in addition to the methods with ‘undo’
or ‘undoable’ in their name, some of the methods calling these undo methods
need to be considered as part of the core aspect as well.
TUD-SERG-2006-002
17
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XVIII
Dynamic analysis is partial (i.e., not all methods involved in an aspect are
retrieved), being based on specific executions, and it can determine only aspects
that can be discriminated by different execution scenarios (e.g., aspects that are
exercised in every program execution cannot be detected). Additionally, it does
not deal with code that cannot be executed (e.g., code that is part of a larger
framework, but that is not used in a specific application).
5.3
Complementarity
The three proposed techniques address symptoms of crosscutting functionality,
such as scattering and tangling, in quite different ways. As shown in Table 6,
fan-in analysis and dynamic analysis show largely complementary result sets:
among the 30 concerns identified by either dynamic or fan-in analysis, only 4
are identified by both techniques. This is an expected result. Fan-in analysis focuses on identifying those methods that are called at multiple places. However,
when a method is called many times, it is likely to occur in most (if not all) execution traces. Hence, no specific use-case can be defined to isolate the associated
functionality, and dynamic analysis will fail to identify it as a seed.
Identifier analysis is the least discriminating of the three techniques and has
a large overlap with the other two techniques. When a concern can be identified
through fan-in analysis and/or dynamic analysis, identifier analysis can often
isolate it too, since a common lexicon is often used in the names of the involved
methods.
In the next section, we will use these observations to propose a new aspect
mining technique that is a clever combination of the three individual techniques.
Technique
Concerns
Dynamic analysis
18
Fan-in analysis S
16
Dynamic analysis T Fan-in analysis
30
Dynamic analysis
Fan-in analysis
4
Table 6. Concerns identified by either dynamic or fan-in analysis.
6
Toward interesting combinations
Based on the discussion in the previous section, this section presents three combined aspect mining techniques and reports on the results of applying these
combined techniques on the JHotDraw application. Based on the analysis indicators of recalled methods and seed quality we compare whether these combined
techniques provide a more complete coverage of the detected concerns than each
of the original techniques individually.
18
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XIX
6.1
Motivation
As has been explained in the previous sections, the fan-in analysis and dynamic
analysis techniques are largely complementary, and address different symptoms
of crosscutting. An obvious and interesting combination of these techniques thus
consists of simply applying each technique individually and taking the union of
the results. Additionally, the seeds in the intersection of the results (if any) are
likely to represent the best aspect candidates, because both techniques identify
them. This was illustrated in our experiment, in which both techniques identified
the Observer, Undo, Persistence and Command execution candidates.
As for other combinations of the techniques, two interesting observations were
considered. First, the manual intervention required by identifier analysis is very
time-consuming and is not justified by the fact that it produces more interesting
results. This makes the technique less suited than the others for large(r) cases.
Second, both fan-in analysis and dynamic analysis identify only candidate seeds
that serve as a starting point for seed expansion. Dynamic analysis in particular
suffers from this problem as it is based on a (necessarily partial) list of execution
scenarios. Similarly, fan-in analysis is only focused on invocations of high fanin methods, which represent just a portion of the whole concern. Interestingly,
while performing fan-in analysis and dynamic analysis, we observed that the
classes and methods in the seed expansion often exhibited similar identifiers.
Consequently, we believe better results can be obtained if we use identifier
analysis as a seed expansion technique for the seeds identified by either fan-in
analysis or dynamic analysis, or by the seeds identified by both these techniques.
In this way, the search space for identifier analysis is reduced significantly, and
more automation is provided for the manual seed expansion needed by both
fan-in analysis and dynamic analysis. A final manual refinement step is anyway
necessary, since the expanded seeds may contain false positives and negatives.
In the remainder of this section, we will present three different techniques: a
combination of fan-in analysis with identifier analysis, of dynamic analysis with
identifier analysis, and of the union of fan-in analysis and dynamic analysis with
identifier analysis.
6.2
Definition of the combined techniques
The combined techniques work as follows:
1. Identify interesting candidate seeds by applying fan-in analysis, dynamic
analysis or both to the application;
– For candidate seeds identified by dynamic analysis, (manually) filter out
those methods that do not pertain to the concern;
2. For each method in the candidate seed, find its enclosing class, and compute
the identifiers occurring in the method and the class name, according to the
algorithm used by identifier analysis;
3. Apply identifier analysis to the application, and search for a concept, among
the concepts it reports, that is “nearest”. The nearest concept is the concept
TUD-SERG-2006-002
19
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XX
that contains most of the identifiers generated in the previous step. If more
than one nearest concept exists, take the union of all their elements.
4. Add the methods contained in the nearest concept(s) to the candidate seed.
5. Revise the expanded list of candidate seeds manually to remove false positives and add missing seeds (false negatives).
In what follows, we experimentally validate these techniques on the JHotDraw case.
6.3
Analysis indicators
Before applying the combined techniques, we define two measures to validate
the results. The goal is to measure how identified seeds change in terms of precision and recall. Unfortunately, this requires information about all crosscutting
concerns present in the application, and this is not available. Therefore, we have
chosen alternative metrics, which we call recalled methods and seed quality.
Recalled methods is the number of methods reported in a seed that actually
belong to the crosscutting concern.
Seed quality is the percentage of a seed’s recalled methods with respect to the
total number of methods in the seed. This indicator estimates how difficult
it is to spot a concern in the methods provided by the seed.
With respect to the definitions above, it is important to remark that for fanin, two interpretations of seeds are possible: the first takes only the callees with
high fan-in into account; the second interpretation includes, besides the callees
with high fan-in, also all callers to these methods. These differences stem from
the fact that the fan-in technique is actually based on the call-relation and the
interpretations use either one or both sides of the relation in seed representations.
During exploration these differences aren’t that important because we can easily
navigate from caller to callee and vise versa. However, when we start assessments
based on counting elements, these interpretations do have considerable impact.
In the first case, the number of recalled methods will be low (since callsites are not considered in the seeds), and the seed quality will always be 100%
since the high fan-in callees belong to the concern by definition. The second
interpretation will result in higher values of recall and yields a more complete
picture of the concern. However lower values for seed quality are possible since
not all calls may be caused by a crosscutting concern.
Section 6.4 describes the results of applying combined techniques on the
JHotDraw appication, and evaluates the above indicators before and after the
experiment. We include results for both interpretations of fan-in seeds discussed
above.
6.4
Experimental results
Table 7 shows the values of the indicators before and after the completion experiment (based on the first interpretation of seeds for fan-in). Although the
20
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XXI
Concerns
Technique
Undo
Command execution
Recalled
Seed
Recalled
Seed
Methods? Quality? Methods? Quality?
Dynamic analysis
23
64%
20
80%
Fan-in
analysis
3
100%
3
100%
S
Dyn
Fan-in
24
63%
22
81%
Dyn + Identifier
183
55%
132
80%
Fan-inS+ Identifier
94
100%
132
80%
(Dyn
Fan-in) + Identifier
183
55%
132
80%
Concerns
Technique
Persistence
Observer
Recalled
Seed
Recalled
Seed
Methods? Quality? Methods? Quality?
Dynamic analysis
29
97%
3
100%
Fan-in
analysis
6
100%
10
100%
S
Dyn
Fan-in
32
97%
13
100%
Dyn + Identifier
104
100%
121
14%
Fan-inS+ Identifier
104
100%
146
15%
(Dyn
Fan-in) + Identifier
104
100%
146
15%
Table 7. Recalled methods and seed quality before and after completion (? based on
the first interpretation of seeds for fan-in)
completion technique can be applied to all concerns identified by either fan-in
analysis or dynamic analysis, we performed the experiment only on the concerns
identified by all three techniques. The sole reason is that we need to assess how
the completion technique influences the recalled methods and seed quality indicators as compared to their initial values, which can only be done for the Undo,
Command execution, Persistence and Observer concerns.
When looking at the common results, it is important to note that fan-in
seeds point to distinct crosscutting concerns sorts that can occur as parts of
more complex structures like implementations of the Observer pattern [18, 19].
In the experiments, these are grouped to obtain the same level of granularity
obtained by the other techniques.
A deeper look into the results of the completion with identifier analysis reveals interesting information: For the Undo concern, the results of both fan-in
analysis and dynamic analysis improve a lot in terms of recalled methods (from
23 and 3 up to 183 and 94). There is a negative impact on the seed quality for
(completed) dynamic analysis (from 64% down to 55%), but the seed quality for
fan-in plus identifier analysis remains at 100%. For the Command execution and
Persistence concerns, the number of recalled methods increases significantly for
the completion technique (from 20 and 3 up to 132 and from 29 and 6 up to
104), while the seed quality remains at the same level.
For the Observer concern, the results are less encouraging than for the other
concerns. Even though the number of recalled methods increases for the completion technique, the quality of the seeds drops to an unacceptable level (from
TUD-SERG-2006-002
21
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XXII
Seed
Recalled Seed
Methods Quality
Undo (callee #1)
Undo (callee #2)
Undo (callee #3)
24
25
24
92%
88%
83%
Undo (combined)
73
88%
Observer (combined) 83
100%
Table 8. Recalled methods and seed quality for fan-in analysis based on the second
interpretation of seeds for fan-in
100% down to 14% and 15%). Clearly, the completion does not provide a good expansion of the original seeds. Closer inspection reveals that no clearly distinctive
naming convention has been used to implement the Observer concern. The Undo,
Command execution and Persistence concerns employ distinctive identifiers such
as undo/undoable, execute/command and store/storable, which are used extensively only within the concern implementation. Consequently, the completion
provided by identifier analysis gives good seed expansions. However, the identifiers used for the Observer concern are the more general figure/update/...
that are used extensively in throughout the application, and not only in the
concern implementation. Therefore, identifier analysis is not able to provide a
good expansion for the seeds found by the other techniques.
An overview of results based on the second interpretation of seeds for fan-in,
i.e. taking also the call-sites into account, is shown in Table 8. For the Undo
concern, we show both the individual values for each of the three high fan-in
callees reported as seeds earlier and the recall and seed quality of the combination
of these three. The seed quality is lower than 100% in these cases since some of
the calls found were not considered to be part of the actual crosscutting concern.
For the Observer concern we only show the value for the combined high fan-in
callees since it would go too far to go over all individual values here. The seed
quality is 100% in these cases since there are no calls from outside this concern
to the reported callees.
For a detailed discussion of these measurements and an assessment of various
quality metrics, we refer to [20] and the fan-in website9 .
7
Summary and future work
The purpose of the paper was to compare three different aspect mining techniques, discuss their respective strengths and weaknesses by applying them to
a common benchmark application, and develop combined techniques based on
this discussion.
9
22
http://swerl.tudelft.nl/amr/
TUD-SERG-2006-002
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XXIII
We observed that all three techniques were able to identify seeds for wellknown crosscutting concerns, but that interesting differences arose for other
concerns. These differences are largely due to the different ways in which the
techniques work. Fan-in analysis is good at identifying seeds that are largely
scattered throughout the system and that involve a lot of invocations of the
same method, but it cannot be used to analyse smaller applications. Identifier
analysis is able to identify seeds when the associated methods have low fanin, but only if these methods share a common lexicon. The main drawback of
this technique is the large number of reported seeds that had to be inspected
manually. Finally, dynamic analysis is able to find seeds in the absence of high
fan-in values and common identifiers, but the technique is only partial because
it relies on execution traces.
We also observed that the three techniques are quite complementary: fan-in
analysis and dynamic analysis require a manual effort to expand the seeds into
full concerns, whereas identifier analysis covers a large part of a concern, but
requires extensive filtering of the reported seeds. Hence, to improve automation
of both fan-in analysis and dynamic analysis, and to reduce the search space for
identifier analysis, we proposed a combined technique in which seeds from either
fan-in analysis or dynamic analysis are expanded automatically by applying
identifier analysis. To verify the performance of this combined technique, we
applied it to JHotDraw and interpreted the results in terms of two indicators:
recalled methods and seed quality. The measures show that for three out of the
four concerns we considered, the combined technique outperforms the individual
techniques. In only one case, the combined technique performed worse.
Future work mainly consists of extending our comparison with other aspect
mining techniques, and potentially proposing new interesting combinations with
such techniques. This will not only allow us to come up with better (combined)
aspect mining techniques, but will also allow us to evaluate the three considered
techniques even better, as new concerns will be identified that we were not aware
of. Additionally, we could come up with extra quality indicators that complement
the recalled methods and seed quality indicators, and empirically establish their
validity by considering other benchmark applications as well.
References
1. Fabry, J.: Modularizing Advanced Transaction Management - Tackling Tangled
Aspect Code. PhD thesis, Vrije Universiteit Brussel (2005)
2. Lippert, M., Lopes, C.V.: A study on exception detection and handling using
aspect-oriented programming. In: Proceedings of the International Conference on
Software Engineering (ICSE), ACM Press (2000) 418–427
3. Bruntink, M., Deursen, A., Tourwé, T.: Discovering faults in idiom-based exception handling. In: Proceedings of the 28th International Conference on Software
Engineering (ICSE) (to appear), ACM Press (2006)
4. Kellens, A., Mens, K.: A survey of aspect mining tools and techniques. Technical
report, INGI 2005-07, Université catholique de Louvain, Belgium (2005)
TUD-SERG-2006-002
23
SERG
Ceccato et al – Applying and Combining Three Different Aspect Mining Techniques
XXIV
5. Deursen, A., Marin, M., Moonen, L.: Aspect mining and refactoring. In: Proceedings of the First International Workshop on REFactoring: Achievements, Challenges, Effects (REFACE03). (2003)
6. Marin, M., Deursen, A., Moonen, L.: Identifying aspects using fan-in analysis.
In: Proc. of the 11th IEEE Working Conference on Reverse Engineering (WCRE
2004), IEEE Computer Society (2004)
7. Mens, K., Tourwé, T.: Delving source-code with formal concept analysis. Elsevier
Journal on Computer Languages, Systems & Structures 31(3–4) (2005) 183–198
Special Issue: Smalltalk.
8. Tourwé, T., Mens, K.: Mining aspectual views using formal concept analysis. In:
Proc. of the Fourth IEEE International Workshop on Source Code Analysis and
Manipulation (SCAM 2004), IEEE Computer Society (2004)
9. Tonella, P., Ceccato, M.: Aspect mining through the formal concept analysis of
execution traces. In: Proceedings of the 11th IEEE Working Conference on Reverse
Engineering (WCRE 2004), IEEE Computer Society (2004)
10. Gamma, E., Helm, R., Johnson, R., Vlissides, J.: Design Patterns: Elements of
Reusable Object-Oriented Software. Addison-Wesley (1994)
11. Henderson-Sellers, B.: Object-oriented metrics: measures of complexity. PrenticeHall (1996)
12. Ganter, B., Wille, R.: Formal Concept Analysis: Mathematical Foundations.
Springer-Verlag (1999)
13. Beck, K.: Smalltalk: best practice patterns. Prentice-Hall (1997)
14. Porter, M.: An algorithm for suffix stripping. Program 14(3) (1980) 130–137
15. Eisenbarth, T., Koschke, R., Simon, D.: Locating features in source code. IEEE
Transactions on Software Engineering 29(3) (2003) 195–209
16. The AspectJ Team: The AspectJ Programming Guide. Palo Alto Research Center.
(2003) Version 1.2.
17. Hannemann, J., Kiczales, G.: Design pattern implementation in Java and AspectJ.
In: Proceedings of the 17th Annual ACM conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA), ACM Press (2002) 161–
173
18. Marin, M., Moonen, L., Deursen, A.: An approach to aspect refactoring based on
crosscutting concern types. In: Proceedings of the First International Workshop
on the Modeling and Analysis of Concerns in Software, International Conference
on Software Engineering, St. Louis, USA (2005)
19. Marin, M., L.Moonen, Deursen, A.: A classification of crosscutting concerns.
In: Proceedings International Conference on Software Maintenance (ICSM 2005),
IEEE Computer Society (2005)
20. Marin, M.: Reasoning about assessing and improving the seed quality of a generative aspect mining technique. In: Proc. of the Second Workshop on Linking Aspect
Technology and Evolution at AOSD 2006. (2006)
24
TUD-SERG-2006-002
TUD-SERG-2006-002
ISSN 1872-5392
SERG
| 2cs.AI
|
The classification of fused links
Timur R. Nasybullov∗†
arXiv:1511.00136v3 [math.GT] 30 Dec 2015
April 2, 2018
Abstract
We construct the complete invariant for fused links. It is proved that the
set of equivalence classes of n-component fused links is in one-to-one correspondence with the set of elements of the abelization U V Pn /U V Pn0 up to
conjugation by the elements from the symmetric group Sn < U V Bn .
Keywords: Fused links, unrestricted virtual braid group, knot invariant.
1
Introduction
Knot invariants are functions of knots that do not change under isotopies. The study
of knot invariants is at the core of knot theory. Indeed, the isotopy class of a knot
is, tautologically, a knot invariant. During last years different authors constructed
vast number of knot invariants: the (self) linking number, the unknotting number,
the knot group, the knot quandle, the Jones polynomial, the Conway polynomial
and so on (see, for example, [19, 7, 12]). The disadvantage of a large number of
easy countable invariants is that they are not complete, i. e. they do not distinguish
some knots or links.
At the same time there are few examples of complete knot invariants, but usually
it is difficult to understand if the value of the invariant on the knot coincides with
the value of this invariant on the another knot. The complement of a knot itself (as
a topological space) is known to be a complete invariant of the knot by the theorem
of C. Gordon and J. Luecke [10] in the sense that it distinguishes the given knot from
all other knots up to ambient isotopy and mirror image. Some invariants associated
with the knot complement include the knot group which is just the fundamental
group of the complement. The knot quandle is also a complete invariant in this
sense [12] but it is difficult to determine if two quandles are isomorphic.
∗
†
The author is supported by Russian Science Foundation (project 14-21-00065)
e-mail: [email protected], [email protected]
1
Thus an important problem in knot theory is to construct a complete knot invariant which can be easily found and used.
Recently some generalizations of classical knots and links were defined and studied: singular links [20, 5], virtual links [11, 18], welded links [8] and fused links
[13, 3, 17]. The problem of constructing invariants is also important for all of this
knot theories.
One of the ways of studying classical links is to study the braid group. Singular
braids [5, 2], virtual braids [18, 21], welded braids [8] and unrestricted virtual braids
[16, 13] were defined similar to the classical braid group adding the extra generators
and relations. Theorem of A. A. Markov [6, §2.2] reduces the problem of classification of links to some algebraic problems of the theory of braid groups. There
are generalizations of Markov theorem for virtual links, welded links and fused links
[14].
In the paper we study fused links, which were defined by L. Kauffman and
S. Lambropoulou in [17], and their invariants. Fused links are represented as generic
immersions of circles in the plane (fused link diagrams) where double points can be
classical (with the usual information on overpasses and underpasses) or virtual (see
Fig. 1).
Figure 1: Crossings in the double welded knot diagram
Fused link diagrams are equivalent under ambient isotopy and some types of local
moves (generalized Reidemeister moves): classical Reidemeister moves (see Fig. 2),
virtual Reidemeister moves (see Fig. 3), mixed Reidemeister moves (see Fig. 4) and
Forbidden moves (see. Fig. 5).
Figure 2: Classical Reidemmeister moves
Figure 3: Virtual Reidemeister moves
2
Figure 4: Mixed Reidemeister moves
Figure 5: Forbidden moves
In the theory of fused links every knot is equivalent to the unknot [15]. However
not every link is equivalent to the trivial link. For example, trivial 2-component link,
Hopf link and Hopf link with one virtual crossing and with one classical crossing all
are different (see Fig. 6).
Figure 6: Different 2-component fused links
The full classification of fused links is not (completely) trivial. In particular, A. Fish
and E. Keyman proved that the fused link with classical crossings only is completely
determined by the linking numbers of each pair of components [9]. It means that the
set of linking numbers for each pair of components of fused links is a full invariant
for fused links which have only classical crossings.
In the present paper we find all non-equivalent classes of fused links and construct
an easy computable full invariant for fused links. We use the following proposition,
which is implicitly formulated in [3].
Proposition. There exists a map %∗ : U V B∞ → U V P∞ , such that the closures
of the braids β and %∗ (β) are equivalent as fused links.
Denote by T the set of coset representatives of U V Pn /U V Pn0 and for the element
α ∈ U V Pn denote by α ∈ T the unique coset representative αU V Pn0 = αU V Pn0 .
Then the main result of the paper can be formulated in the following form.
Theorem. Let α and β be unrestricted virtual braids. Then their closures α
b and
∗
∗
b
β are equivalent as fused links if and only if % (α) and % (β) are conjugated by the
element from Sn < U V Bn .
Sn
Thus the map α
b → %∗ (α) is a complete invariant for fused links and in order
to understand that two fused links α
b an βb are equivalent or not we just need to
3
Sn
Sn
compare two finite sets %∗ (α) and %∗ (β) , or equivalently, to understand that the
Sn
element %∗ (α) belongs to the finite set %∗ (β) or not.
At the almost same time B. Audoux, P. Bellingeri, J.-B. Meilhan and E. Wagner
found the full classification of fused links independently from the author [1, Theorem
2.5]. Using the different from the author’s methods they proved that every fused
link is completely determined by the set of virtual linking numbers for each pair of
components. The result is the same (however in different formulation) and these
two approaches seem to us complementary and both interesting.
The author is grateful to Valeriy Bardakov for multiple helpful advices during
the work on the paper. Also the author would like to thank the University of
Bologna (Italy) where a part of this work was completed and especially Prof. Michele
Mulazzani for his help and support.
2
Definitions and results
In this section we fix notation and recall basic definitions and known results about
different generalizations of braid groups. The classical braid group Bn on n strands
(n > 1) is the group generated by the elements σ1 , . . . , σn−1 (see Fig. 7)
Figure 7: Geometric braids representing σi (on the left) and σi−1 (on the right)
with the following defining relations.
i = 1, . . . , n − 2;
|i − j| ≥ 2.
σi σi+1 σi = σi+1 σi σi+1
σi σj = σj σi
(B1 )
(B2 )
There exists a homomorphism ι : Bn → Sn from the braid group Bn onto the
symmetric group Sn on n letters. This homomorphism maps the generator σi to the
transposition τi = (i, i + 1) for i = 1, 2, . . . , n − 1. The kernel of this homomorphism
is called pure braid group on n strands and denoted by Pn .
The virtual braid group V Bn is a group obtained from Bn adding new generators
ρ1 , . . . , ρn−1 (see Fig. 8)
and additional relations
i = 1, . . . , n − 2;
|i − j| ≥ 2;
i = 1, . . . , n − 1;
ρi ρi+1 ρi = ρi+1 ρi ρi+1
ρi ρj = ρj ρi
ρ2i = 1
4
(P1 )
(P2 )
(P3 )
|i − j| ≥ 2;
i = 1, . . . , n − 2.
σi ρj = ρj σi
ρi ρi+1 σi = σi+1 ρi ρi+1
(M1 )
(M2 )
It is easy to verify that the elements ρ1 , . . . , ρn−1 generate the symmetric group Sn .
Also it is known that the elements σ1 , . . . , σn−1 generate the braid group Bn . In the
paper [11] it is proved that the relations
i = 1, . . . , n − 2;
i = 1, . . . , n − 2;
ρi σi+1 σi = σi+1 σi ρi+1
ρi+1 σi σi+1 = σi σi+1 ρi
(F1 )
(F2 )
do not hold in the group V Bn . According to [8] the welded braid group W Bn on n
strands is a quotient of the group V Bn by the forbidden relation (F1 ), i. e. it is a
group with the generators σ1 , . . . , σn−1 , ρ1 , . . . , ρn−1 and relations (B1 )–(F1 ). If we
add to the group W Bn the second forbidden relation (F2 ) the we get the unrestricted
virtual braid group U V Bn .
The elements σ1 , . . . , σn−2 , ρ1 , . . . , ρn−2 generate the subgroup U V Bn−1 in the
group U V Bn . Then we have the following chain of inclusions.
[
U V B2 < U V B3 < U V B4 < · · · < U V B∞ =
U V Bn
n≥2
The homomorphism ι can be extended to the homomorphism U V Bn → Sn by the
rule ι : σi , ρi 7→ τi = (i i + 1). The kernel of this homomorphism is called pure
unrestricted virtual braid group and is denoted by U V Pn . The group U V P∞ is
defined analogically to the group U V B∞ .
[
U V P2 < U V P3 < U V P4 < · · · < U V P∞ =
U V Pn
n≥2
The symmetric group Sn acts on the set {1, . . . , n}. By the symbol π(k) we
denote an image of the integer k ∈ {1, . . . , n} under the permutation π. If for the
braid α we have ι(α)(k) = k, then we say that α fixes k or k is fixed by α.
We say that the braid α does not involve the strand j if α belongs to the group
hσ1 , . . . , σj−2 , σj+1 , . . . , σn−1 , ρ1 , . . . , ρj−2 , ρj+1 , . . . , ρn−1 i. It is obvious that the braid
α ∈ hρ1 , . . . , ρn−1 i which fixes the strands n − 1, n does not involve an n-strand.
We say that the braid β ∈ U V Bn+1 is obtained from the braid α ∈ U V Bn
by right stabilization of positive (negative, virtual) type if β = ασn (β = ασn−1 ,
Figure 8: Geometric braid representing ρi
5
β = αρn respectively). In this case we say that the braid α is obtained from the
braid β by the opposite to the right stabilization of positive (negative, virtual) type
transformation. S. Kamada proved an analogue of Markov theorem for welded links
in [14, Theorem 2].
Theorem 1. Closures of two welded braids α and β are equivalent as welded links
if and only if they are related by the finite sequence of the following transformations.
1. A conjugation in the welded braid group,
2. A right stabilization of positive, negative or virtual type,
3. An opposite to a right stabilization of positive, negative or virtual type transformation.
This theorem also holds for fused links since every relation of welded braid group
is fulfilled in the unrestricted virtual braid group.
Let us define some element of U V Bn . For i = 1, . . . , n − 1 :
λi,i+1 = ρi σi−1 ,
λi+1,i = ρi λi,i+1 ρi = σi−1 ρi .
For 1 ≤ i < j − 1 ≤ n − 1:
λi,j = ρj−1 ρj−2 . . . ρi+1 λi,i+1 ρi+1 . . . ρj−2 ρj−1 ,
λj,i = ρj−1 ρj−2 . . . ρi+1 λi+1,i ρi+1 . . . ρj−2 ρj−1 .
The elements λi,j and λj,i for i < j belong to the pure unrestricted virtual braid
group U V Pn and have the following geometric interpretation (see Fig. 9).
Figure 9: Geometric braids representing λi,j (on the left) and λj,i (on the right)
The following lemma is proved in [4, Lemma 1] for the corresponding elements in
V Bn and therefore is also true in the quotient U V Bn .
Lemma 1. The following conjugating rules are fulfilled in U V Bn :
6
1. for 1 ≤ i < j ≤ n and k < i − 1 or i < k < j − 1 or k > j:
ρk λi,j ρk = λi,j ,
ρk λj,i ρk = λj,i ;
2. for 1 ≤ i < j ≤ n:
ρi−1 λi,j ρi−1 = λi−1,j ,
ρi−1 λj,i ρi−1 = λj,i−1 ;
3. for 1 ≤ i < j − 1 ≤ n:
ρi λi,i+1 ρi = λi+1,i ,
ρi λi,j ρi = λi+1,j ,
ρi λi+1,i ρi = λi,i+1 ,
ρi λj,i ρi = λj,i+1 ;
4. for 1 ≤ i + 1 < j ≤ n:
ρj−1 λi,j ρj−1 = λi,j−1 ,
ρj−1 λj,i ρj−1 = λj−1,i ;
5. for 1 ≤ i < j ≤ n:
ρj λi,j ρj = λi,j+1 ,
ρj λj,i ρj = λj+1,i .
The following result on the structure of the pure unrestricted virtual braid group
U V Pn is presented in [3, Theorem 2.7].
Theorem 2. The group U V Pn has a presentation with generators λk,l for 1 ≤ k 6=
l ≤ n, and defining relations: λi,j commutes with λk,l if and only if {i, j} =
6 {k, l}.
Let Uj be the subgroup of U V Pn generated by the elements {λi,j , λj,i | 1 ≤ i < j}.
Then by Theorem 2 we have
Uj = hλ1,j , λj,1 i × hλ2,j , λj,2 i × · · · × hλj−1,j , λj,j−1 i = F2 × · · · × F2 ,
(1)
U V Pn = U2 × U3 × · · · × Un .
(2)
The structure of the unrestricted virtual braid group U V Bn follows from Lemma
1 and the theorem 2 and is also given in [3, Theorem 2.4].
Theorem 3. The group U V Bn is isomorphic to the semidirect product U V Pn hSn ,
where the symmetric group Sn acts by permutations on the indices of generators λi,j
of U V Pn .
Denote by Bi,j = ρj−1 ρj−2 . . . ρi+1 ρi if i < j and Bi,j = 1 in other cases. Using
simple calculations in the group Sn it is easy to see that for i < j the image ι(Bi,j )
is a cycle (i i + 1 . . . j).
7
3
Construction of %∗
In this section we construct the map %∗ : U V B∞ → U V P∞ , which is implicitly
constructed in [3]. At first we prove the following lemma.
Lemma 2. Let α ∈ U V Bn and s ∈ {1, . . . , n} be a maximal number, such that
ι(α)(s) 6= s. Then α can be uniquely expressed in the form
α = γxs Bks ,s xs+1 xs+2 . . . xn ,
where xi ∈ Ui and the braid γ ∈ U V Bn does not involve the strands s, s + 1, . . . , n.
Proof. By the theorem 3 for certain elements β ∈ U V Pn and π ∈ Sn we have
α = βπ.
(3)
Let ks = π(s) = ι(βπ)(s) = ι(α)(s) 6= s. Since ι(α) is a bijection of {1, . . . , n} and
ι(α)(s) = ks , then we have ι(α)(ks ) 6= ks . Since s is a maximal number with the
condition ι(α)(s) 6= s then ks < s and therefore Bks ,s 6= 1.
Since the permutations π and Bk−1
act identically on the integers s + 1, . . . , n,
s ,s
−1
then the permutation δ = πBks ,s also acts identically on the integers s + 1, . . . , n.
Moreover the image of the integer s under the permutation δ is equal to s
Bk−1
s ,s
π
s −−−→ ks −−−−→ s,
therefore δ = πBk−1
acts identically on the integers s, . . . , n.
s ,s
By the equality (2) for certain elements yj ∈ Uj we have β = y2 y3 . . . yn , and
therefore we can rewrite the equality (3).
α = βπ = y2 y3 . . . yn δBks ,s = δ(y2 y3 . . . yn )δ Bks ,s
(4)
Since (y2 . . . yn ) is a pure braid, then (y2 y3 . . . yn )δ is a pure braid and therefore
by the equality (2) we have (y2 y3 . . . yn )δ = z2 . . . zn for certain elements zj ∈ Uj .
The braids z2 , . . . , zs−1 do not involve the strands s, s + 1, . . . , n, therefore the braid
γ = δz2 . . . zs−1 does not involve the strands s, s + 1, . . . , n. Then the equality (4)
can be rewritten in the following form.
α = δ(y2 y3 . . . yn )δ Bks ,s = δz2 . . . zs−1 zs zs+1 . . . zn Bks ,s
B
B
ks ,s
= γzs zs+1 . . . zn Bks ,s = γzs Bks ,s (zs+1 . . . zn )Bks ,s = γzs Bks ,s zs+1
. . . zn ks ,s
Since zj ∈ Uj = hλ1,j , λj,1 i×hλ2,j , λj,2 i×· · ·×hλj−1,j , λj,j−1 i, then by the theorem
B
3 for every j = s + 1, . . . , n we have zj ks ,s ∈ Uj . Therefore, if we denote xs = zs ,
Bks ,s
B
xs+1 = zs+1
, . . . , xn = zn ks ,s then we have
α = γxs Bks ,s xs+1 . . . xn .
8
and we proved that the braid α can be written in the form from the formulation of
the lemma.
To prove that such a representation is unique we consider another representation
of α in the form from the formulation of the lemma
α = γxs Bks ,s xs+1 . . . xn = ηys Bts ,s ys+1 . . . yn
(5)
and prove that γ = η, ks = ts and xj = yj for every j = s, . . . , n.
From the equality (5) we have
η −1 γ = ys Bts ,s ys+1 . . . yn (xs Bks ,s xs+1 . . . xn )−1
(6)
Without loss of generality consider that ts ≤ ks and look at the images of the
braids from the right and left sides of this equality under the homomorphism ι.
The permutation ι(ys Bts ,s ys+1 . . . yn (xs Bks ,s xs+1 . . . xn )−1 ) maps s to s if ts = ks and
maps s to ts if ts < ks
ι(ys Bt
ι((xs Bk
,s ys+1 ...yn )
−1 )
,s xs+1 ...xn )
s
s −−−−−−
−−−−−−→ ts −−−−−−s−−−−−−−−−→ ts .
At the same time since η and γ does not involve the strands s, . . . , n, then s is
fixed by the permutation ι(η −1 γ) and therefore ts = ks . Then from the equality (6)
we have ι(ys Bts ,s ys+1 . . . yn (xs Bks ,s xs+1 . . . xn )−1 ) = 1 and ι(η −1 γ) = 1. Therefore
η −1 γ ∈ U V Pn and since it does not involve the strands s, s + 1, . . . n, we have
η −1 γ ∈ hU2 , . . . , Us−1 i.
On the other side the braid ys Bts ,s ys+1 . . . yn (xs Bks ,s xs+1 . . . xn )−1 can be rewritten in the following form.
−1
−1 −1
ys Bts ,s ys+1 . . . yn (xs Bks ,s xs+1 . . . xn )−1 = ys Bts ,s ys+1 . . . yn x−1
n . . . xs+1 Bts ,s xs
−1
−1 −1 Bts ,s −1
= ys Bts ,s Bt−1
(ys+1 . . . yn x−1
xs
n . . . xs+1 xs )
s ,s
−1
−1 −1 Bts ,s −1
= ys (ys+1 . . . yn x−1
xs
n . . . xs+1 xs )
B −1
−1
Bts ,s
By the theorem 3 for every j = s + 1, . . . , n we have yj ts ,s ∈ Uj , (x−1
∈ Uj ,
j )
therefore
ys Bts ,s ys+1 . . . yn (xs Bks ,s xs+1 . . . xn )−1 ∈ hUs , . . . , Un i.
From the equality (2) we have hU2 , . . . , Us−1 i ∩ hUs , . . . , Un i = 1, therefore η = γ
and xs Bks ,s xs+1 . . . xn = ys Bts ,s ys+1 . . . yn . The lemma is proved.
The following lemma has a bit different formulation in the paper [3] and it
completely proved there. However here we formulate the lemma in the other words
and repeat the proof from [3] since this lemma is extremely important in our paper.
9
Lemma 3. There exists a map % : U V B∞ → U V B∞ with the following conditions.
1. For any α the closures of α and %(α) are equivalent as double welded links.
2. If α ∈ U V P∞ , then %(α) = α.
3. If α ∈ U V Bn \ U V Pn , then %(α) ∈ U V Bn−1 .
Proof. We explicitly construct the image of the braid α ∈ U V Bn under the
map %. If α ∈ U V Pn , then %(α) = α and there is nothing to construct. So let
α ∈ U V Bn \ U V Pn , then the algorithm of finding %(α) has the following steps.
Step 1. Find a maximal number s ∈ {1, . . . , n} with the condition ι(α)(s) = ks 6= s.
n−s
Step 2. Conjugate the braid α by the element B1,n
.
s−n
n−s
α1 = B1,n
δB1,n
The fused link α
b1 is equivalent to α
b, and the permutation induced by the braid
α1 maps n to kn = n − s + ks < n
s−n
B1,n
n−s
B1,n
α
n −−−−→ s −−−→ ks −−−−→ n − s + ks = kn .
Step 3. By Lemma 2 express the braid α1 in the form
α1 = γw1 . . . wn−1 ρn−1 Bkn ,n−1 ,
where the braid γ does not involve the strand n and wi ∈ hλi,n , λn,i i for
i = 1, . . . , n (see Fig. 10).
Figure 10: The braid α1
By Lemma 2 we have α1 = γxn Bkn ,n , where the braid γ does not involve
an n-strand and xn ∈ Un . By the equality (1) we have xn = w1 . . . wn−1 for
wi ∈ hλi,n , λn,i i. Sine Bkn ,n 6= 1, then Bkn ,n = ρn−1 Bkn ,n−1 .
10
Step 4. Change the braid α1 by the braid
α2 = γw1 . . . wn−2 ρn−1 Bkn ,n−1 .
From the figure 10 it is obvious that all the crossings of wn−1 belong to the
same component of α
b1 , therefore, we can virtualize all the classical crossings
of wn−1 ρn−1 using Kanenobu’s technique [15, Proof of the theorem 1], i. e. to
change σn−1 to ρn−1 in the representation of wn−1 and to obtain a braid α2
such that α
b2 and α
b1 are equivalent as fused links.
−1
−1
Since wn−1 ∈ hλn−1,n , λn,n−1 i and λn−1,n = ρn−1 σn−1
, λn,n−1 = σn−1
ρn−1 , then
after virtualization of classical crossings the braid wn−1 ρn−1 involved an odd
number of ρn−1 and we obtain ρn−1 instead of wn−1 ρn−1 . Therefore the closure
of the braid α2 = γw1 . . . wn−2 ρn−1 Bkn ,n−1 is equivalent to the closure of the
braid α1 as fused link.
Step 5. Construct %(α).
Note that by Lemma 1 for all λj,n and λn,j with j < n − 1 we have:
λj,n ρn−1 = ρn−1 λj,n−1 ,
λn,j ρn−1 = ρn−1 λn−1,j .
ρ
0
Bkn ,n−1 , where wi0 = wi n−1 is a braid from
Therefore α2 = γρn−1 w10 . . . wn−2
hλi,n−1 , λn−1,i i and hence does not involve an n-strand (see Fig. 11).
Figure 11: The braid α2
In the braid α2 there is only one (virtual) crossing on the n-strand, so, using
0
Markov moves we obtain a new braid α3 = γw10 . . . wn−2
Bkn ,n−1 whose closure
is again equivalent to α
b and has n − 1 strands. We have constructed the braid
0
Bkn ,n−1 .
%(α) = α3 = γw10 . . . wn−2
The lemma is proved.
11
Remark 1. It is obvious that the equality %(α) = %(α1 ) = %(α2 ) holds in Lemma
2.
Let α ∈ U V Bn , then the sequence %(α), %2 (α), . . . is stabilized on some step. Let
%∗ be such a map which maps the braid α to the braid %k (α), where k is a minimal
integer such that %k (α) = %k+1 (α). Every step of finding %(α) is clearly defined,
therefore %∗ is a correct function.
Using simple induction on the number n it is easy to prove the following fact.
Remark 2. Let α ∈ U V Bn be the braid, such that %∗ (α) ∈ U V Pm . If α = βγ for
β ∈ U V Bn , γ ∈ U V Pn0 , then %∗ (α) = %∗ (β)δ, where δ ∈ U V Pm0 .
From Lemma 3 We have the following corollary which is formulated and proved
in the paper [3].
Proposition 1. Let α ∈ U V Bn be an unrestricted virtual braid, such that its
closure α
b has m components. Then there exists a pure unrestricted virtual braid
b
β ∈ U V Pm , such that α
b = β.
Proof. β = %∗ (α).
4
Proof of the main results
Lemma 4. Let α ∈ U V Pn and u, v ∈ hλn−1,n , λn,n−1 i. Then the closure of the
braids α and α[u, v] are equivalent as fused links.
Proof. Let γ be the following braid from U V Bn+2
γ = αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 Bn−1,n+2 .
Here the braid αu−1 ρn−1 does not involve the strands n + 1, n + 2 and the braid
uρn ρn−1 belongs to hλn,n+1 , λn+1,n i. Let us find the braid %(γ).
Step 1. Since ι(γ)(n + 2) = n − 1, then the maximal number which is not fixed by γ
is equal to n + 2.
Step 2. γ1 = γ.
Step 3. γ = αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 Bn−1,n+2 , where the braid αu−1 ρn−1 uρn ρn−1 Bn−1,n+1
does not involve the strand n + 2.
Step 4. γ2 = αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 ρn+1 Bn−1,n+1 .
Step 5. %(γ) = αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 Bn−1,n+1 .
12
Let us find the braid %2 (γ).
Step 1. Since ι(%(γ))(n + 1) = n, then the maximal number which is not fixed by %(γ)
is equal to n + 1.
Step 2. %(γ)1 = %(γ).
Step 3. The braid %(γ) can be rewritten
%(γ) = αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 Bn−1,n+1
= αu−1 ρn−1 uρn ρn−1 ρn−1 ρn
= αu−1 ρn−1 uρn ρn−1 ρn−1 ρn = αu−1 uρn ρn ,
where the braid αu−1 does not involve the strand n + 1 and uρn belongs to
hλn−1,n+1 , λn+1,n−1 i.
Step 4. %(γ)2 = αu−1 uρn ρn .
Step 5. %2 (γ) = α.
Since %2 (γ) = α is a pure braid, then %∗ (γ) = α. Let w = v ρn−1 Bn−1,n+1 Bn−1,n+2 , then
by Lemma 1 the braid w belongs to hλn+2,n+1 , λn+1,n+2 i. Denote by δ = wγw−1 and
find the braid %∗ (δ). At first find the braid %(δ).
Step 1. Since w is a pure braid, then the maximal number, which is not fixed by δ is
equal to the maximal number which is not fixed by γ and is equal to n + 2.
Step 2. δ1 = δ.
Step 3. The braid δ can be rewritten in details
δ = wγw−1 = wαu−1 ρn−1 uρn ρn−1 Bn−1,n+1 Bn−1,n+2 w−1
−1
= αu−1 ρn−1 uρn ρn−1 wBn−1,n+1 (w−1 )Bn−1,n+2 Bn−1,n+2
−1
= αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 wBn−1,n+1 (w−1 )Bn−1,n+2 Bn−1,n+2
−1
= αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 (w−1 )Bn−1,n+2 wBn−1,n+1 Bn−1,n+2 ,
−1
where the braid αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 (w−1 )Bn−1,n+2 does not involve the
strand n + 2 and the braid wBn−1,n+1 belongs to hλn+2,n , λn,n+2 i.
−1
Step 4. δ2 = αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 (w−1 )Bn−1,n+2 wBn−1,n+1 ρn+1 Bn−1,n+1 .
−1
Step 5. %(γ) = αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 (w−1 )Bn−1,n+2 wBn−1,n+1 ρn+1 Bn−1,n+1 .
Let us fin the braid %2 (δ)
13
Step 1. Since ι(%(δ))(n + 1) = n, then n + 1 is maximal number which is not fixed by
%(δ).
Step 2. %(δ)1 = %(δ).
Step 3. The braid %(δ) can be rewritten in details
−1
%(δ) = αu−1 ρn−1 uρn ρn−1 Bn−1,n+1 (w−1 )Bn−1,n+2 wBn−1,n+1 ρn+1 Bn−1,n+1
−1
−1
−1
−1
−1
−1
= αu−1 ρn−1 uρn ρn−1 (w−1 )Bn−1,n+2 Bn−1,n+1 wBn−1,n+1 ρn+1 Bn−1,n+1 Bn−1,n+1 Bn−1,n+1
= αu−1 ρn−1 uρn ρn−1 (w−1 )Bn−1,n+2 Bn−1,n+1 wBn−1,n+1 ρn+1 Bn−1,n+1 ρn−1 ρn
−1
−1
−1
= αu−1 uρn (w−1 )Bn−1,n+2 Bn−1,n+1 ρn−1 wBn−1,n+1 ρn+1 Bn−1,n+1 ρn−1 ρn
−1
−1
−1
= αu−1 (w−1 )Bn−1,n+2 Bn−1,n+1 ρn−1 uρn wBn−1,n+1 ρn+1 Bn−1,n+1 ρn−1 ρn ,
−1
−1
where the braid αu−1 (w−1 )Bn−1,n+2 Bn−1,n+1 ρn−1 does not involve the strand n+1
−1
and the braid uρn wwBn−1,n+1 ρn+1 Bn−1,n+1 ρn−1 belongs to hλn−1,n+1 , λn+1,n−1 i.
−1
−1
−1
Step 4. %(δ)2 = αu−1 (w−1 )Bn−1,n+2 Bn−1,n+1 ρn−1 uρn wBn−1,n+1 ρn+1 Bn−1,n+1 ρn−1 ρn .
Step 5. The braid %2 (δ) follows
−1
−1
−1
%2 (δ) = αu−1 (w−1 )Bn−1,n+2 Bn−1,n+1 ρn−1 uwBn−1,n+1 ρn+1 Bn−1,n+1 ρn−1 ρn .
−1
−1
Since w = v ρn−1 Bn−1,n+1 Bn−1,n+2 , then (w−1 )Bn−1,n+2 Bn−1,n+1 ρn−1 = v −1 and
−1
wBn−1,n+1 ρn+1 Bn−1,n+1 ρn−1 ρn = v
−1
ρn−1 Bn−1,n+1 Bn−1,n+2 Bn−1,n+1 ρn+1 Bn−1,n+1
ρn−1 ρn
= v ρn−1 Bn−1,n+1 Bn−1,n+2 Bn−1,n+1 ρn+1 ρn−1 ρn ρn−1 ρn
= v ρn−1 Bn−1,n+1 Bn−1,n+2 Bn−1,n+1 ρn+1 ρn ρn−1
= v ρn−1 Bn−1,n+2 Bn,n+2 Bn−1,n+1 ρn+1 ρn ρn−1
= v ρn−1 Bn−1,n+2 ρn+1 ρn−1 ρn+1 ρn ρn−1
= v ρn−1 Bn−1,n+2 ρn−1 ρn ρn−1 = v ρn−1 ρn+1 ρn−1 = v ρn+1
=v
(since v ∈ hλn−1,n , λn,n−1 i)
Therefore %2 (δ) = αu−1 v −1 uv = α[u, v]. Since %2 (δ) is a pure braid, then
%∗ (δ) = %2 (δ) = α[u, v].
−1
Since the closures of the braids γ and δ = γ w define equivalent fused links, then
the closures of the braids %∗ (γ) = α and %∗ (δ) = α[u, v] define equivalent fused links.
The lemms is proved.
Denote by T the set of coset representatives of U V Pn /U V Pn0 and for the element
α ∈ U V Pn denote by α ∈ T the unique coset representative αU V Pn0 = αU V Pn0 . The
following statement almost immediately follows from Lemma 4.
14
Corollary 1. Let α, β ∈ U V Pn be unrestricted virtual braids, such that α and β
are conjugated by the element from Sn . Then the fused links α
b and βb are equivalent.
Proof. Since α and β are conjugated by the element from Sn then the closures
of the braids α and β are equivalent fused links. Therefore it is enough to prove
that the closure of the braid α is equivalent to the closure of the braid α and the
closure of the braid β is equivalent to the closure of the braid β.
Since αU V Pn0 = αU V Pn0 , then α = α[u1 , v1 ][u2 , v2 ] . . . [uk , vk ] for certain elements ui , vi ∈ U V Pn , i = 1, . . . , k. By Theorem 2 we can consider that ui , vi ∈
hλri ,si , λsi ,rI i. Denote by α0 = α, α1 = α0 [u1 , v1 ], α2 = α1 [u2 , v2 ], . . . , αk =
αk−1 [uk , vk ] = α. If we prove that the closure of αi and the closure of αi−1 define the same fused link for every i = 1, . . . , k, then we prove that the closures of
the braids α and α are equivalent as fused links. Therefore we can consider that
α = α[u, v] for some u, v ∈ hλr,s , λs,r i.
The closure of the braid α is equivalent to the closure of the braid αµ for any
µ ∈ Sn < U V Bn . If µ is a permutation which maps s to n − 1 and maps r to n, then
the closure of the braid α is equivalent to the closure of the braid αµ = αµ [uµ , v µ ]. By
Theorem 3 the braids uµ , v µ belong to hλn−1,n , λn,n−1 i, thus by Lemma 4 the closure
of the braid α is equivalent to the closure of the braid αµ and is also equivalent
to the closure of α. The closure of β is equivalent to the closure of β by the same
reasons. The corollary is proved.
Proposition 2. Let α, β ∈ U V B∞ be unrestricted virtual braids, such that α
b
∗
∗
b
and β are equivalent fused links with m components. Then % (α) and % (β) are
conjugated by the element from Sn .
Proof. Since the links α
b and βb are equivalent, then by Theorem 1 the braids α
and β are related by the finite sequence of Markov’s transformations. It means that
there is a finite sequence of braids
α = α0 , α1 , . . . , αk = β
such that αj+1 is obtained from αj by conjugation in U V B∞ , by right stabilization
or by inverse to right stabilization transformation.
If we prove that %∗ (αj ) and %∗ (αj+1 ) are conjugated by the element from Sn for
every j = 0, . . . , k − 1, than we prove that %∗ (α) and %∗ (β) are conjugated by the
element from Sn , therefore we can consider that β is obtained from α using only one
Markov’s transformation.
Case 1. The braid β is obtained from the braid α by right stabilization of positive,
negative or virtual type. We consider only the case of right stabilization of positive
type (β = ασn ), the cases of right stabilization of negative and virtual type are
similar.
Let us count the image of β under the map %:
15
Step 1. Since β = ασn , then ι(β)(n + 1) = ι(ασn )(n + 1) = n 6= n + 1
α
σ
n
→ n,
n + 1 −−−→ n + 1 −−−
therefore n + 1 is a maximal number which is not fixed by β.
n+1−(n+1)
Step 2. β1 = B1,n+1
n+1−(n+1)
βB1,n+1
= β.
Step 3. We can rewrite the braid β1 in the following form
β1 = β = ασn = ασn ρn ρn = αλ−1
n,n+1 ρn ,
where the braid α does not involve the strand n+1 and λ−1
n,n+1 ∈ hλn,n+1 , λn+1,n i.
Step 4. β2 = αρn .
Step 5. %(β) = α.
Since %(β) = α, then %∗ (β) = %∗ (α) (and they the braids %∗ (β) and %∗ (α) are
certainly conjugated).
Case 2. The braid β is obtained from the braid α by an opposite to a right
stabilization of positive, negative or virtual type transformation.
In this case the braid α is obtained from the braid β by right stabilization of
positive (negative, virtual) type and by the case 1 we have %∗ (α) = %∗ (β).
Case 3. The braid β is obtained from the braid α by conjugation in U V Bn .
Since the braid α and β are conjugated, then for some braid θ we have β = αθ .
Since σi = ρi λi,i+1 , then θ = y1 y2 . . . yr , where
±1
±1
yi ∈ {ρ1 , . . . , ρn−1 , λ±1
1,2 , λ2,3 , . . . , λn−1,n }.
If we denote δ1 = α, δ2 = δ1y1 , δ3 = δ2y2 , . . . , δr+1 = δryr = β and prove that %∗ (δj )
is conjugated with %∗ (δj+1 ) by the element from Sn for every j = 1, . . . , r + 1, then
we prove that %∗ (α) is conjugated with %∗ (β) by the element from Sn . Therefore we
can consider that β is obtained from α conjugating by ρi or λ±1
i,i+1 for some i.
We use the induction by the parameter n − m, i. e. by the difference between
the number of strands in the braid α and the number components in the link α
b.
∗
∗
If n−m = 0, then α and β are pure braids and % (α) = α, % (β) = β. Let β = αµ
for some braid µ ∈ U V Bn , then by Theorem 3 for certain elements β ∈ U V Pn and
π ∈ Sn the braid µ can be presented as a product µ = βπ. Therefore
%∗ (β) = β = αµ = µ−1 αµ = π −1 β −1 αβπ
= π −1 αβ −1 [β −1 , α]βπ = π −1 α[β −1 , α]β π = π −1 %∗ (α)[β −1 , αβ ]π
16
Since α, β ∈ U V Pn , then [β −1 , αβ ] ∈ U V Pn0 , therefore %∗ (α)[β −1 , αβ ] = %∗ (α) and
π
%∗ (β) = %∗ (α) .
If n − m > 0, then α, β ∈ U V Bn \ U V Pn , %∗ (α) 6= α, %∗ (β) 6= β and we have the
following cases
Case 3.1. The braid β is obtained from α conjugating by ρi . Let us find %(α).
Step 1. Let s be a maximal number which is not fixed by α and ι(α)(s) = ks .
s−n
n−s
Step 2. Since α1 = B1,n
αB1,n
, then ι(α1 )(n) = ks + n − s = kn 6= n.
s−n
B1,n
n−s
B1,n
α
n −−−−→ s −−−→ ks −−−−→ ks + n − s = kn .
Step 3. Express the braid α1 in the form
n−s
s−n
= γw1 . . . wn−1 ρn−1 Bkn ,n−1 ,
αB1,n
α1 = B1,n
(7)
where the braid γ does not involve the strand n and wi ∈ hλi,n , λn,i i.
Step 4. α2 = γw1 . . . wn−2 ρn−1 Bkn ,n−1 .
Step 5. The braid %(α) has the following form
ρ
ρ
n−1
%(α) = γw1 n−1 . . . wn−2
Bkn ,n−1 ,
(8)
ρ
where wj n−1 ∈ Un−1 does not involve the strand n.
Case 3.1.1. i ≤ ks − 2. Let us count the braid %(β) = %(ρi αρi ).
Step 1. Since i ≤ ks − 2, then ρi fixes s and ks , therefore s is a maximal number which
is not fixed by β and ι(β)(s) = ks .
ρi
ρi
α
s −−−→ s −−−→ ks −−−→ ks
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
, then the permutation ι(β) maps n to kn 6= n.
s−n
B1,n
n−s
B1,n
β
n −−−−→ s −−−→ ks −−−−→ ks + n − s = kn
Step 3. For i ≤ ks − 2 we have
ρi B1,n = ρi ρn−1 . . . ρ1 = ρi ρn−1 . . . ρi+2 ρi+1 ρi ρi−1 . . . ρ1
= ρn−1 . . . ρi+2 ρi ρi+1 ρi ρi−1 . . . ρ1
= ρn−1 . . . ρi+2 ρi+1 ρi ρi+1 ρi−1 . . . ρ1
= ρn−1 . . . ρi+2 ρi+1 ρi ρi−1 . . . ρ1 ρi+1 = B1,n ρi+1 ,
17
n−s
n−s
therefore ρi B1,n
= B1,n
ρi+n−s and hence we have
s−n
n−s
s−n
n−s
s−n
n−s
β1 = B1,n
βB1,n
= B1,n
ρi αρi B1,n
= ρi+n−s B1,n
αB1,n
ρi+n−s
= ρi+n−s γw1 . . . wn−1 ρn−1 Bkn ,n−1 ρi+n−s
ρi+n−s ρi+n−s ρi+n−s
= γ ρi+n−s (w1 . . . wn−2 )ρi+n−s wn−1
ρn−1 Bkn ,n−1
Since i ≤ ks − 2, then i + n − s ≤ ks − 2 + n − s = kn − 2 < n − 2, therefore
ρ
ρi+n−s
ρi+n−s
Bkni+n−s
= wn−1 , ρn−1
= ρn−1 and the braid γ ρi+n−s does
,n−1 = Bkn ,n−1 , wn−1
not involve the strand n. Since ρi+n−s fixes n, then (w1 . . . wn−2 )ρi+n−s belongs
to hλ1,n , λn,1 i × · · · × hλn−2,n , λn,n−2 i. Then we have
β1 = γ ρi+n−s (w1 . . . wn−2 )ρi+n−s wn−1 ρn−1 Bkn ,n−1 .
Step 4. β2 = γ ρi+n−s (w1 . . . wn−2 )ρi+n−s ρn−1 Bks ,n−1
Step 5. Since i + n − s < n − 2, then ρn−1 and ρi+n−s commute, therefore the braid
%(β) has the following form
%(β) = γ ρi+n−s (w1 . . . wn−2 )ρi+n−s ρn−1 Bkn ,n−1
= γ ρi+n−s (w1 . . . wn−2 )ρn−1 ρi+n−s Bkn ,n−1
ρ
= γ ρi+n−s ((w1 . . . wn−2 )ρn−1 )ρi+n−s Bkni+n−s
,n−1
= (γ(w1 . . . wn−2 )ρn−1 Bkn ,n−1 )ρi+n−s = %(α)ρi+n−s
Therefore %(β) = %(α)ρi+n−s and by the induction hypothesis the braids %∗ (α) and
%∗ (β) are conjugated by the element from Sn .
Case 3.1.2. i = ks − 1. Let us find the braid %(β).
Step 1. Since ks ≤ s − 1, then i = ks − 1 ≤ s − 2 and ρi fixes the strands s. Then the
image of s under the permutation β = ρks −1 αρks −1 is equal to ks − 1.
ρk
−1
ρk
α
−1
s
s
s −−−
−→ s −−−→ ks −−−
−→ ks − 1.
Therefore s is a maximal number which is not foxed by β.
n−s
s−n
and ι(β1 )(n) = kn − 1.
Step 2. We have β1 = B1,n
βB1,n
s−n
B1,n
n−s
B1,n
β
n −−−−→ s −−−→ ks − 1 −−−−→ n − s + ks − 1 = kn − 1
Step 3. Using the same arguments as in the case 3.1.1 we have
n−s
n−s
n−s
ρks −1 B1,n
= B1,n
ρn−s+ks −1 = B1,n
ρkn −1 .
18
Therefore the braid β1 has the following form
n−s
s−n
n−s
s−n
n−s
s−n
ρkn −1
αB1,n
= ρkn −1 B1,n
ρks −1 αρks −1 B1,n
= B1,n
βB1,n
β1 = B1,n
= ρkn −1 γw1 . . . wn−1 ρn−1 Bkn ,n−1 ρkn −1
ρkn −1 ρkn −1 ρkn −1
ρn−1 Bkn ,n−1
= γ ρkn −1 (w1 . . . wn−2 )ρkn −1 wn−1
ρ
ρ
−1
kn −1
= ρkn −1 Bkn −1,n , then we have
Bknkn,n−1
Since ρn−1
ρ
kn −1
ρkn −1 Bn,kn −1
β1 = γ ρkn −1 (w1 . . . wn−2 )ρkn −1 wn−1
ρkn −1
ρkn −1 (w1 . . . wn−2 )wn−1 Bn,kn −1
=γ
= γ ρkn −1 ρkn −1 (w1 . . . wn−2 )wn−1 ρn−1 Bn−1,kn −1
Since kn ≤ n − 1, then kn − 1 ≤ n − 2, therefore the braid γ ρkn −1 ρkn −1 does
not involve the strand n.
Step 4. β2 = γ ρkn −1 ρkn −1 (w1 . . . wn−2 )ρn−1 Bn−1,kn −1
Step 5. The braid %(β) has the following form.
%(β) = γ ρkn −1 ρkn −1 (w1 . . . wn−2 )ρn−1 Bn−1,kn −1
= γ ρkn −1 (w1 . . . wn−2 )ρn−1 ρkn −1 ρkn −1 Bn−1,kn −1
ρkn −1
= γ ρkn −1 (w1 . . . wn−2 )ρn−1 ρkn −1 Bn−1,k
= %(α)ρkn −1
n
Therefore %(β) = %(α)ρkn −1 and by the induction hypothesis the braids %∗ (α) and
%∗ (β) are conjugated by the element from Sn .
Case 3.1.3. i = ks .
Case 3.1.3.1. ks ≤ s − 2. Let us find the image of the braid β = ρks αρks under
the map %.
Step 1. Since ks ≤ s−2, then and ρks fixes s and the image of s under the permutation
ι(β) is equal to ks + 1.
ρk
α
ρk
s −−−s→ s −−−→ ks −−−s→ ks + 1
Therefore s is a maximal number which is not fixed by β.
s−n
n−s
βB1,n
and ι(β1 )(n) = kn + 1.
Step 2. We have β1 = B1,n
s−n
B1,n
n−s
B1,n
β
n −−−−→ s −−−→ ks + 1 −−−−→ kn + 1,
19
n−s
n−s
Step 3. Analogically to the case 3.1.1, since ks ≤ s − 2, then ρks B1,n
= B1,n
ρkn .
Therefore the braid β1 has the following form.
s−n
n−s
s−n
n−s
s−n
n−s
β1 = B1,n
βB1,n
= B1,n
ρks αρks B1,n
= ρkn B1,n
αB1,n
ρk n
= ρkn γw1 . . . wn−1 ρn−1 Bkn ,n−1 ρkn
ρ
ρkn ρkn
ρn−1 Bknkn,n−1
= γ ρkn (w1 . . . wn−2 )ρkn wn−1
ρ
ρ
kn
Bknkn,n−1 ρkn Bkn +1,n , then we have
Since ρn−1
ρ
kn
ρkn Bkn +1,n
β1 = γ ρkn (w1 . . . wn−2 )ρkn wn−1
ρkn
= γ ρkn (w1 . . . wn−2 )wn−1 Bkn +1,n .
Since ks ≤ s − 2, then kn ≤ n − 2 and the braid γ ρkn ρkn does not involve the
strand n.
Step 4. β2 = γ ρkn ρkn (w1 . . . wn−2 )ρn−1 Bkn +1,n−1
Step 5. The braid %(β) has the form
%(β) = γ ρkn ρkn (w1 . . . wn−2 )ρn−1 Bkn +1,n−1
= γ ρkn (w1 . . . wn−2 )ρn−1 ρkn ρkn Bkn +1,n−1
ρ
= γ ρkn (w1 . . . wn−2 )ρn−1 ρkn Bknkn,n−1 = %(α)ρkn
Therefore %(β) = %(α)ρkn and by the induction hypothesis the braids %∗ (α) and
%∗ (β) are conjugated by the element from Sn .
Case 3.1.3.2. ks = s − 1.
Case 3.1.3.2.1. ι(α)(s − 1) = s. From the formulas (7) and (8) we have
α1 = γw1 . . . wn−1 ρn−1 Bkn ,n−1 ,
%(α) = γ(w1 . . . wn−2 )ρn−1 Bkn ,n−1 .
Let us rewrite this equalities in more details. Since ι(α)(s) = s − 1, then
ι(α1 )(n) = n − 1 and Bkn ,n = ρn−1 .
s−n
Bs,n
n−s
Bs,n
α
n −−−−→ s −−−→ s − 1 −−−−→ n − 1
Since ι(α)(s − 1) = s then ι(α1 ) maps (n − 1) to n.
s−n
Bs,n
α
n−s
Bs,n
n − 1 −−−−→ s − 1 −−−→ s −−−−→ n
Therefore, since γ = α1 (Bkn ,n w1 . . . wn−1 )−1 , then ι(γ)(n − 1) = n − 1
−1
(Bkn ,n w1 ...wn−1 )
n − 1 −−−→ n −−−−−−−−−−−−−→ n − 1
α1
20
and by Lemma 2 we have
γ = ηv1 . . . vn−2 ,
where η does not involve the strand n−1 and vj ∈ hλj,n−1 , λn−1,j i for j = 1, . . . , n−2.
Therefore
α1 = ηv1 . . . vn−2 w1 . . . wn−1 ρn−1 ,
%(α) = ηv1 . . . vn−2 (w1 . . . wn−2 )ρn−1 .
Let us find the braid %(β).
Step 1. Since ks = s−1 and ι(α)(s−1) = s, then the image of s under the permutation
ι(β) = ι(ρks αρks ) = ι(ρs−1 αρs−1 ) is equal to s − 1.
ρs−1
ρs−1 a
α
s −−−−→ s − 1 −−−→ s −−−−→ s − 1,
then s is a maximal number which is not fixed by β.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = n − 1.
s−n
B1,n
n−s
B1,n
β
n −−−−→ s −−−→ s − 1 −−−−→ n − 1,
n−s
n−s
Step 3. Since ρs−1 B1,n
= B1,n
ρn−1 , then we have
s−n
n−s
s−n
n−s
s−n
n−s
β1 = B1,n
βB1,n
= B1,n
ρs−1 αρs−1 B1,n
= ρn−1 B1,n
αB1,n
ρn−1
= ρn−1 ηv1 . . . vn−2 w1 . . . wn−1 ρn−1 ρn−1
ρ
ρn−1 ρn−1
ρn−1 ρn−1
= η ρn−1 v1 n−1 . . . vn−2
w1 . . . wn−2
wn−1 ρn−1
Since the braid η does not involve the strands n − 1, n, then η ρn−1 = η. Also
by Lemma 1 for j = 1, . . . , n − 2 we have
ρ
ρ
vj n−1 ∈ hλj,n , λn,j i,
wj n−1 ∈ hλj,n−1 , λn−1,j i
ρ
n−1
and wn−1
∈ hλn,n−1 , λn−1,n i. Therefore
ρ
ρ
ρ
ρ
ρ
n−1
n−1
n−1
β1 = ηw1 n−1 . . . wn−2
v1 n−1 . . . vn−2
wn−1
ρn−1 ,
ρ
ρ
ρ
n−1
where the braid ηw1 n−1 . . . wn−2
does not involve the strand n, the braid vj n−1
ρn−1
belongs to hλj,n , λn,j i for j = 1, . . . , n − 2 and wn−1
∈ hλn,n−1 , λn−1,n i.
ρ
ρ
ρ
ρ
n−1
n−1
ρn−1
v1 n−1 . . . vn−2
Step 4. β2 = ηw1 n−1 . . . wn−2
Step 5. The braid %(β) has the following form
ρ
ρ
n−1
v1 . . . vn−2
%(β) = ηw1 n−1 . . . wn−2
ρn−1
ρn−1
ρ
ρn−1
, v1 . . . vn−2 ]
= ηv1 . . . vn−2 w1 . . . wn−2
[w1 n−1 . . . wn−2
21
Therefore by the remark 2 the braids %∗ (β) and %∗ (α) are equal modulo U V Pm0 , i. e.
%∗ (α) = %∗ (β) and these braids are certainly conjugated.
Case 3.1.3.2.2. ι(α)(s − 1) = ks−1 ≤ s − 2. By the equalities (7) and (8) we have
α1 = γw1 . . . wn−1 ρn−1 Bkn ,n−1 ,
%(α) = γ(w1 . . . wn−2 )ρn−1 Bkn ,n−1 .
Let us rewrite these equalities in more details. Since ι(α1 )(n) = kn = n − 1, then
Bkn ,n−1 = 1. Also we have ι(α1 )(n − 1) = n − s + ks−1 = kn−1
s−n
B1,n
n−s
B1,n
α
n − 1 −−−−→ s − 1 −−−→ ks−1 −−−−→ n − s + ks−1 = kn−1 ,
Since γ = α1 (Bkn ,n w1 . . . wn−1 )−1 , then ι(γ) maps n − 1 to kn−1
−1
(Bkn ,n w1 ...wn−1 )
−−−−−−−−−−−−−→ kn−1 ,
α1
n − 1 −−−→ kn−1
therefore by Lemma 2 we have
γ = ηv1 . . . vn−2 Bkn−1 ,n−1 ,
where η does not involve the strand n − 1, n and vj belongs to hλj,n−1 , λn−1,j i for
j = 1, . . . , n − 2. Therefore we have
α1 = ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 ρn−1 ,
%(α) = ηv1 . . . vn−2 Bkn−1 ,n−1 (w1 . . . wn−2 )ρn−1 .
Let us count %2 (α).
Step 1. Since η does not involve the strand n − 1, then ι(%(α))(n − 1) = kn−1 , then
n − 1 is a maximal number which is not fixed by %(α).
Step 2. %(α)1 = %(α) = ηv1 . . . vn−2 Bkn−1 ,n−1 (w1 . . . wn−2 )ρn−1 .
Step 3. Let us rewrite the braid %(α)1 in the required form.
%(α)1 = ηv1 . . . vn−2 Bkn−1 ,n−1 (w1 . . . wn−2 )ρn−1
ρn−1 Bk−1
= ηv1 . . . vn−2 (w1 . . . wn−2 )
n−1 ,n−1
Bkn−1 ,n−1
= ηv1 . . . vn−2 (w1 . . . wkn−1 −1 wkn−1 wkn−1 +1 . . . wn−2 )
= η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
· v1 . . . vn−3 vn−2 wkn−1
n−1 ,n−1
ρn−1 Bk−1 ,n−1
n−1
Bkn−1 ,n−1
22
ρn−1 Bk−1
n−1 ,n−1
Bkn−1 ,n−1
ρ
B −1
By Theorem 3 the braid η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 ) n−1 kn−1 ,n−1 does not
involve the strand n − 1, the braid vj ∈ hλj,n−1 , λn−1,j i for j = 1, . . . , n − 3 and
ρn−1 Bk−1
vn−2 wkn−1
n−1 ,n−1
∈ hλn−1,n−2 , λn−2,n−1 i.
ρn−1 Bk−1
Step 4. %(α)2 = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
v1 . . . vn−3 ρn−2 Bkn−1 ,n−2
Step 5. The braid %2 (α) follows
%(α)2 = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
n−1 ,n−1
(v1 . . . vn−3 )ρn−2 Bkn−1 ,n−2
Let us count %(β)
Step 1. Since ks = s − 1 and ι(α)(s − 1) = ks−1 ≤ s − 2, then the image of s under the
permutation β = ρks αρks = ρs−1 αρs−1 is equal to ks−1 .
ρs−1
ρs−1
α
s −−−−→ s − 1 −−−→ ks−1 −−−−→ ks−1 ,
then s is a maximal number which is not fixed by β. The image of s − 1 under
the permutation β = ρks αρks = ρs−1 αρs−1 is equal to s.
ρs−1
ρs−1
α
s − 1 −−−−→ s −−−→ s − 1 −−−−→ s,
s−n
n−s
Step 2. We have β1 = Bn,1
βBn,1
and ι(β1 )(n) = ks−1 + n − s = kn−1 , ι(β1 )(n − 1) = n
s−n
Bn,1
n−s
B1,n
β
n −−−−→ s −−−→ ks−1 −−−−→ ks−1 + n − s = kn−1 ,
s−n
Bn,1
s−n
Bn,1
β
n − 1 −−−−→ s − 1 −−−→ s −−−−→ n,
n−s
n−s
Step 3. Since ks = s − 1 < s then ρs−1 B1,n
= B1,n
ρn−1 , therefore we have
s−n
n−s
s−n
n−s
s−n
n−s
β1 = B1,n
βB1,n
= B1,n
ρs−1 αρs−1 B1,n
= ρn−1 B1,n
αB1,n
ρn−1
= ρn−1 ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 ρn−1 ρn−1
ρn−1
ρn−1
= η ρn−1 (v1 . . . vn−2 )ρn−1 Bkn−1
ρn−1
,n−1 (w1 . . . wn−1 )
=η
ρn−1
(v1 . . . vn−2 )
ρn−1
ρ
−1
ρn−1 Bk n−1,n−1
(w1 . . . wn−1 )
n−1
Bk−1
= η ρn−1 (v1 . . . vn−2 )ρn−1 (w1 . . . wn−1 )
n−1 ,n−1
Bk−1 ,n
n−1
= η ρn−1 (v1 . . . vn−2 )ρn−1 (w1 . . . wn−1 )
ρn−1
ρ
n−1
Bkn−1
,n−1 ρn−1
Bkn−1 ,n
Bkn−1 ,n
Bk−1
= η ρn−1 (v1 . . . vn−2 )ρn−1 (w1 . . . wkn−1 −1 wkn−1 wkn−1 +1 . . . wn−1 )
Bk−1
= η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
23
n−1 ,n
n−1 ,n
Bkn−1 ,n
Bk−1 ,n
n−1
(v1 . . . vn−2 )ρn−1 wkn−1
Bkn−1 ,n
Bk−1
By Theorem 3 the braid η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
n−1 ,n
Bk−1
ρ
does not
,n
n−1
involve the strand n, vj n−1 ∈ hλj,n , λn,j i for j = 1, . . . n−2 and wkn−1
belongs
to hλn−1,n , λn,n−1 i.
Bk−1
Step 4. β2 = η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
n−1 ,n
(v1 . . . vn−2 )ρn−1 ρn−1 Bkn−1 ,n−1
Step 5. The braid %(β) has the following form
%(β) = η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
Bk−1
n−1 ,n
(v1 . . . vn−2 )Bkn−1 ,n−1 (9)
Let us count %2 (β)
Step 1. From the equality (9) the image of n − 1 under the permutation ι(%(β)) is
equal to kn−1 , therefore n − 1 is a maximal number which is not fixed by %(β).
Step 2. %(β)1 = %(β).
Step 3. Rewrite the braid %(β)1 in the required form
Bk−1
%(β)1 = η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
n−1 ,n
Bk−1
= η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n
(v1 . . . vn−2 )Bkn−1 ,n−1
Bk−1
,n
v1 . . . vn−3 wn−1n−1 vn−2 Bkn−1 ,n−1
Since η does not involve the strands n − 1, n, then η ρn−1 = η does not involve the strand n − 1.
Bk−1
By Lemma 1 the braid vn−2 wn−1n−1
,n
belongs to
Bk−1 ,n
n−1
hλn−1,n−2 , λn−2,n−1 i, the braid (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
belongs to hλ1,n−1 , λn−1,1 i × · · · × hλn−1,n−2 , λn−2,n−1 i.
Step 4. %(β)2 = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bk−1
n−1 ,n
Bk−1
Step 5. %2 (β) = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n
v1 . . . vn−3
v1 . . . vn−3 ρn−2 Bkn−1 ,n−2
ρn−2
(v1 . . . vn−3 )ρn−2 Bkn−1 ,n−2
Using simple calculations in symmetric group it is easy to show that Bk−1
ρ
=
n−1 ,n n−2
−1
ρn−1 Bkn−1 ,n−1 ρn−1 , therefore
B −1
ρ
(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 ) kn−1 ,n n−2 =
ρn−1
ρn−1 Bk−1 ,n−1
n−1
= (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
24
B −1
ρ
and since (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 ) n−1 kn−1 ,n−1 does not involve the strands
n − 1, n, then
ρn−1
ρ
B −1
=
(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 ) n−1 kn−1 ,n−1
= (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
n−1 ,n−1
Therefore %2 (α) = %2 (β) and by the induction hypothesis the braids %∗ (α) and
%∗ (β) are conjugated by the element from Sn .
Case 3.1.4. ks + 1 ≤ i ≤ s − 2. Let us count the braid %(β) = %(ρi αρi ).
Step 1. Since ks ≤ s − 3 and ks + 1 ≤ i ≤ s − 2, then the braid ρi fixes s and ks , and
therefore the image of s under the permutation ι(β) = ι(ρi αρi ) is equal to ks .
ρi
ρi
α
s −−−→ s −−−→ ks −−−→ ks
Therefore s is a maximal number which is not fixed by β.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = kn .
s−n
B1,n
n−s
B1,n
β
n −−−−→ s −−−→ ks −−−−→ kn ,
n−s
n−s
Step 3. Since i ≤ s − 2, then ρi B1,n
= B1,n
ρn−s+i , and therefore the braid β1 follows.
s−n
n−s
s−n
n−s
s−n
n−s
β1 = B1,n
βB1,n
= B1,n
ρi αρi B1,n
= ρi+n−s B1,n
αB1,n
ρi+n−s
= ρi+n−s γw1 . . . wn−1 ρn−1 Bkn ,n−1 ρi+n−s
ρi+n−s ρi+n−s ρi+n−s
= γ ρi+n−s (w1 . . . wn−2 )ρi+n−s wn−1
ρn−1 Bkn ,n−1
Using simple calculation in the permutation group it is easy to show that
ρi+n−s ρi+n−s
ρ
ρn−1
Bkn ,n−1 = Bkni+n−s
= ρi+n−s ρi+n−s−1 Bkn ,n , hence
,n
ρ
i+n−s
β1 = γ ρi+n−s (w1 . . . wn−2 )ρi+n−s wn−1
ρi+n−s ρi+n−s−1 Bkn ,n
ρi+n−s−1
ρi+n−s
=γ
ρi+n−s ρi+n−s−1 (w1 . . . wn−2 )ρi+n−s−1 wn−1
Bkn ,n
Since i ≤ s − 2, then i + n − s − 1 ≤ n − 3 and the braid γ ρi+n−s ρi+n−s ρi+n−s−1
does not involve the strand n. Also by Lemma 1 (w1 . . . wn−2 )ρi+n−s−1 belongs
ρi+n−s−1
to hλ1,n , λn,1 i × · · · × hλn−2,n , λn,n−2 i and wn−1
= wn−1 ∈ hλn−1,n , λn,n−1 i.
Step 4. β2 = γ ρi+n−s ρi+n−s ρi+n−s−1 (w1 . . . wn−2 )ρi+n−s−1 ρn−1 Bkn ,n−1
Step 5. The braid β has the following form
%(β) = γ ρi+n−s ρi+n−s ρi+n−s−1 (w1 . . . wn−2 )ρi+n−s−1 ρn−1 Bkn ,n−1
= γ ρi+n−s (w1 . . . wn−2 )ρi+n−s−1 ρn−1 ρi+n−s−1 ρi+n−s ρi+n−s ρi+n−s−1 Bkn ,n−1
ρ
ρi+n−s
= γ ρi+n−s (w1 . . . wn−2 )ρn−1 ρi+n−s Bkni+n−s
,n−1 = %(α)
25
Therefore %(β) = %(α)ρi+n−s and by the induction hypothesis the braids %∗ (α) and
%∗ (β) are conjugated by the element from Sn .
Case 3.1.5. i = s − 1. In this case we can consider that ks < s − 1 since the case
when ks = s − 1 = i is already solved in the case 3.1.3.2.
Case 3.1.5.1. ι(α)(s − 1) = s. By the equalities (7) and (8) the braids α1 and
%(α) have the following forms
α1 = γw1 . . . wn−1 ρn−1 Bkn ,n−1 ,
%(α) = γ(w1 . . . wn−2 )ρn−1 Bkn ,n−1 .
Let us rewrite this equalities in more details. Since ι(α)(s − 1) = s, then the
permutation ι(α1 ) maps n − 1 to n.
s−n
B1,n
α
B1,nn−s
n − 1 −−−−→ s − 1 −−−→ s −−−−−−→ n
Therefore the permutation ι(γ) = ι(α1 (w1 . . . wn−1 ρn−1 Bkn ,n−1 )−1 ) fixes n − 1
−1
(w1 ...wn−1 ρn−1 Bkn ,n−1 )
n − 1 −−−→ n −−−−−−−−−−−−−−−−−→ n − 1,
α1
and by Lemma 2 we have
γ = ηv1 . . . vn−2 ,
where η does not involve the strands n − 1, n and vj ∈ hλj,n−1 , λn−1,j i. Therefore
the braids α1 , %(α) can be rewritten.
α1 = ηv1 . . . vn−2 w1 . . . wn−1 ρn−1 Bkn ,n−1
%(α) = ηv1 . . . vn−2 (w1 . . . wn−2 )ρn−1 Bkn ,n−1
Let us count %2 (α)
Step 1. Since ks < s − 1, then kn < n − 1, therefore Bk1 ,n−1 6= 1, then n − 1 is a
maximal number which is not fixed by %(α).
Step 2. %(α)1 = %(α).
Step 3. The braid %(α)1 has the following form.
%(α)1 = %(α) = ηv1 . . . vn−2 (w1 . . . wn−2 )ρn−1 Bkn ,n−1
ρn−1
= ηv1 . . . vn−3 (w1 . . . wn−3 )ρn−1 vn−2 wn−2
Bkn ,n−1
Here the braid η does not involve the strand n − 1, by Lemma 1 the braid
v1 . . . vn−3 (w1 . . . wn−3 )ρn−1 belongs to hλ1,n−1 , λn−1,1 i×· · ·×hλ1,n−3 , λn−3,1 i and
ρn−1
vn−2 wn−2
belongs to hλn−1,n−2 , λn−2,n−1 i.
26
Step 4. %(α)2 = ηv1 . . . vn−3 (w1 . . . wn−3 )ρn−1 ρn−2 Bkn ,n−2
Step 5. The braid %2 (α) follows.
%2 (α) = η(v1 . . . vn−3 )ρn−2 (w1 . . . wn−3 )ρn−1 ρn−2 Bkn ,n−2
(10)
Let us find the braid %(β) = %(ρi αρi ) = %(ρs−1 αρs−1 ).
Step 1. Since i = s − 1 and ι(α)(s − 1) = s, then the image of s under the permutation
β = ρs−1 αρs−1 is equal to s − 1
ρs−1
α
ρs−1
s −−−−→ s − 1 −−−→ s −−−−→ s − 1,
then s is a maximal number which is not fixed by β.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = n − 1.
s−n
B1,n
β
n−s
B1,n
n −−−−→ s −−−→ s − 1 −−−−→ n − 1
n−s
n−s
Step 3. Since ρs−1 B1,n
= B1,n
ρn−1 , then then we have
s−n
n−s
s−n
n−s
s−n
n−s
β1 = B1,n
βB1,n
= B1,n
ρs−1 αρs−1 B1,n
= ρn−1 B1,n
αB1,n
ρn−1
= ρn−1 ηv1 . . . vn−2 w1 . . . wn−1 ρn−1 Bkn ,n−1 ρn−1
ρ
= η ρn−1 (v1 . . . vn−2 )ρn−1 (w1 . . . wn−1 )ρn−1 ρn−1 Bknn−1
,n−1
= η ρn−1 (v1 . . . vn−2 )ρn−1 (w1 . . . wn−1 )ρn−1 Bkn ,n−1 ρn−1
= η ρn−1 Bkn ,n−1 (v1 . . . vn−2 )ρn−1 Bkn ,n−1 (w1 . . . wn−1 )ρn−1 Bkn ,n−1 ρn−1
= η ρn−1 Bkn ,n−1 (v1 . . . vn−2 )Bkn ,n (w1 . . . wn−1 )Bkn ,n ρn−1
B
B
B
B
kn ,n
kn ,n
= η ρn−1 Bkn ,n−1 (v1 . . . vn−3 )Bkn ,n (w1 . . . wn−2 )Bkn ,n wn−1
vn−2
ρn−1
kn ,n
kn ,n
= η ρn−1 Bkn ,n−1 (w1 . . . wn−2 )Bkn ,n (v1 . . . vn−3 )Bkn ,n wn−1
vn−2
ρn−1
Since η does not involve the strands n − 1, n, then η ρn−1 = η and the braid
η ρn−1 Bkn ,n−1 (w1 . . . wn−2 )Bkn ,n does not involve the strand n. By Lemma 2 the
Bkn ,n
braid (v1 . . . vn−3 )Bkn ,n wn−1
belongs to hλ1,n , λn,1 i × · · · × hλn−2,n , λn,n−2 i and
Bkn ,n
vn−2 belongs to hλn−1,n , λn,n−1 i.
B
kn ,n
Step 4. β2 = η ρn−1 Bkn ,n−1 (w1 . . . wn−2 )Bkn ,n (v1 . . . vn−3 )Bkn ,n wn−1
ρn−1
B
kn ,n
Step 5. %(β) = η ρn−1 Bkn ,n−1 (w1 . . . wn−2 )Bkn ,n (v1 . . . vn−3 )Bkn ,n ρn−1 wn−1
Let us count the braid %2 (β) = %(%(β)).
27
ρn−1
Step 1. Since η does not involve the strand n − 1, n therefore η ρn−1 = η and ι(η)
Bkn ,n ρn−1
is
fixes n − 1. Moreover since (w1 . . . wn−2 )Bkn ,n (v1 . . . vn−3 )Bkn ,n ρn−1 wn−1
a pure braid, that the image of n − 1 under the permutation ι(%(β)) is equal
to ι(Bkn ,n−1 )(n − 1) = kn . Therefore n − 1 is a maximal number which is not
fixed by %(β).
Step 2. %(β)1 = %(β).
Step 3. The braid %(β)1 has the following form
B
kn ,n
%(β)1 = η ρn−1 Bkn ,n−1 (w1 . . . wn−2 )Bkn ,n (v1 . . . vn−3 )Bkn ,n ρn−1 wn−1
−1
ρn−1
−1
= η ρn−1 (w1 . . . wn−2 )Bkn ,n Bkn ,n−1 (v1 . . . vn−3 )Bkn ,n ρn−1 Bkn ,n−1
Bk
−1
,n ρn−1 Bkn ,n−1
· wn−1n
Bkn ,n−1
Since Bkn ,n Bk−1
= ρn−1 , Bkn ,n ρn−1 Bk−1
= ρn−2 ρn−1 , then we have
n ,n−1
n ,n−1
ρ
ρ
n−2 n−1
%(β)1 = η ρn−1 (w1 . . . wn−2 )ρn−1 (v1 . . . vn−3 )ρn−2 ρn−1 wn−1
Bkn ,n−1
ρ
ρ
n−1
n−2 ρn−1
= η ρn−1 (w1 . . . wn−3 )ρn−1 (v1 . . . vn−3 )ρn−2 ρn−1 wn−2
wn−1
Bkn ,n−1
Here the braid η ρn−1 = η does not involve the strand n − 1. Also by Lemma 2
the braid (w1 . . . wn−3 )ρn−1 (v1 . . . vn−3 )ρn−2 ρn−1 belongs to the group hλ1,n−1 , λn−1,1 i×
ρn−2 ρn−1
· · · × hλn−3,n−1 , λn−1,n−3 i and wn−1
∈ hλn−1,n−2 , λn−2,n−1 i.
Step 4. %(β)2 = η ρn−1 (w1 . . . wn−3 )ρn−1 (v1 . . . vn−3 )ρn−2 ρn−1 ρn−2 Bkn ,n−2
Step 5. The braid %2 (β) has the following form
%2 (β) = η ρn−1 (w1 . . . wn−3 )ρn−1 ρn−2 (v1 . . . vn−3 )ρn−2 ρn−1 ρn−2 Bkn ,n−2
%2 (β) = η ρn−1 (v1 . . . vn−3 )ρn−2 ρn−1 ρn−2 (w1 . . . wn−3 )ρn−1 ρn−2
· [(w1 . . . wn−3 )ρn−1 ρn−2 , (v1 . . . vn−3 )ρn−2 ρn−1 ρn−2 ]Bkn ,n−2
By Lemma 2 it is obvious that (v1 . . . vn−3 )ρn−2 ρn−1 ρn−2 = (v1 . . . vn−3 )ρn−2 and
therefore by the remark 2 the braids %∗ (β) and %∗ (α) are equal modulo U V Pm0 ,
i. e. %∗ (α) = %∗ (β).
Case 3.1.5.2. ι(α)(s − 1) = s − 1. Since β = ρs−1 αρs−1 then ι(β) fixes s and maps
s − 1 to ks .
ρs−1
ρs−1
α
s −−−−→ s − 1 −−−→ s − 1 −−−−→ s
ρs−1
ρs−1
α
s − 1 −−−−→ s −−−→ ks (≤ s − 2) −−−−→ ks
28
Therefore s − 1 is maximal number which is not fixed by β, and β1 has the following
form.
s−1−n
n−s+1
−1 s−n
n−s
B1,n
β1 = B1,n
βB1,n
= B1,n
B1,n ρs−1 αρs−1 B1,n
−1
s−n
n−s
−1
= B1,n
ρn−1 B1,n
αB1,n
ρn−1 B1,n = B1,n−1
α1 B1,n−1
ρ
n−2
= α1 , then it is obvious
If we denote by δ1 = β1 , δ2 = δ1ρ1 , δ3 = δ2ρ2 , . . . , δn−1 = δn−2
that n is not fixed by ι(δj ) for every j. Since δj+1 is obtained from δj conjugating
by ρj (j < n − 1), then by the cases 3.1.1–3.1.4 the braids %∗ (δj+1 ) and %∗ (δj ) are
µj
conjugated by the element from Sn , i. e. %∗ (δj ) = %∗ (δj+1 ) . Therefore we have
µ1
%∗ (β1 ) = %∗ (δ1 ) = %∗ (δ2 )
= %∗ (δ3 )
µ2 µ1
µn−2 ...µ1
= · · · = %∗ (δn−1 )
µn−2 ...µ1
= %∗ (α1 )
,
so the braids %∗ (β) = %∗ (β1 ) and %∗ (α) = %∗ (α1 ) are conjugated by the element from
Sn .
Case 3.1.5.3. ι(α)(s − 1) ≤ s − 2.
Case 3.1.5.3.1. ι(α)(s − 1) ≥ ι(α)(s) + 1. By the equalities (7) and (8) the braids
α1 and %(α) have the following forms.
α1 = γw1 . . . wn−1 ρn−1 Bkn ,n−1
%(α) = γ(w1 . . . wn−2 )ρn−1 Bkn ,n−1
Let us rewrite this equalities in more details. Since ι(α)(s − 1) ≤ s − 2, then
ι(α1 )(n − 1) = n − s + ι(α)(s − 1).
s−n
B1,n
n−s
B1,n
α
n − 1 −−−−→ s − 1 −−−→ ι(α)(s − 1) −−−−→ n − s + ι(α)(s − 1)
Therefore since ι(α)(s−1) ≥ ι(α)(s)+1, then the braid γ = α1 (w1 . . . wn−1 ρn−1 Bkn ,n−1 )−1
maps n − 1 to kn−1 = n − s + ι(α)(s − 1) − 1 ≥ n − s + ι(α)(s) = kn
−1
(w1 ...wn−1 ρn−1 Bkn ,n−1 )
n − 1 −−−→ n − s + ι(α)(s − 1) −−−−−−−−−−−−−−−−−→ n − s + ι(α)(s − 1) − 1,
α1
and by Lemma 2 we have
γ = ηv1 . . . vn−2 Bkn−1 ,n−1 ,
where η does not involve the strands n − 1, n and vj belongs to hλj,n−1 , λn−1,j i for
j = 1, . . . , n − 1. Therefore
α1 = ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 ρn−1 Bkn ,n−1
%(α) = ηv1 . . . vn−2 Bkn−1 ,n−1 (w1 . . . wn−2 )ρn−1 Bkn ,n−1
and kn−1 ≥ kn .
Let us count %2 (α)
29
Step 1 The permutation ι(%(α)) maps n − 1 to kn−1 + 1 = n − s + ι(α)(s − 1) − 1 + 1 =
n − s + ι(α)(s − 1) 6= n, therefore n − 1 is a maximal number which is not
foxed by %(α).
Step 2 %(α)1 = %(α).
Step 3 The braid %(α)1 = %(α) can be rewritten.
%(α1 ) = ηv1 . . . vn−2 Bkn−1 ,n−1 (w1 . . . wn−2 )ρn−1 Bkn ,n−1
ρn−1 Bk−1
= ηv1 . . . vn−2 (w1 . . . wn−2 )
n−1 ,n−1
Bkn−1 ,n−1 Bkn ,n−1
Since kn−1 ≥ kn , then Bkn−1 ,n−1 Bkn ,n−1 = Bkn ,n−2 Bkn−1 +1,n−1 and
ρn−1 Bk−1
%(α1 ) = ηv1 . . . vn−2 (w1 . . . wn−2 )
n−1 ,n−1
= η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bkn ,n−2 Bkn−1 +1,n−1
ρn−1 Bk−1
n−1 ,n−1
ρn−1 Bk−1
v1 . . . vn−2 wkn−1
n−1 ,n−1
· Bkn ,n−2 Bkn−1 +1,n−1
= η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
n−1 ,n−1
ρn−1 Bk−1 ,n−1 Bkn ,n−2
n−1
· (v1 . . . vn−2 )Bkn ,n−2 wkn−1
Bkn ,n−2
Bkn−1 +1,n−1
ρn−1 Bk−1 ,n−1
n−1
= η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wkn−1 )
Bkn ,n−2
ρn−1 Bk−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 wkn−1
Bkn ,n−2 B
kn ,n−2
n−1 ,n−1
vn−3
Bkn−1 +1,n−1
ρn−1 Bk−1
By Lemma 1 the braid η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bkn ,n−2
ρn−1 Bk−1 ,n−1 Bkn ,n−2
n−1 ,n−1
does not involve the strand n−1, the braid (v1 . . . vn−4 vn−2 )Bkn ,n−2 wkn−1
B
n−1
kn ,n−2
belongs to hλ1,n−1 , λn−1,1 i×· · ·×hλn−3,n−1 , λn−1,n−3 i and vn−3
∈ λn−2,n−1 , λn−1,n−2 i.
Step 4 The braid %(α)2 has the following form
ρn−1 Bk−1
%(α)2 = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1 ,n−1 Bkn ,n−2
n−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 wkn−1
30
n−1 ,n−1
Bkn ,n−2
ρn−2 Bkn−1 +1,n−2
Step 5 Finally, the braid %2 (α) follows
%2 (α) = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
n−1 ,n−1
Bkn ,n−2
ρn−1 Bk−1 ,n−1 Bkn ,n−2 ρn−2
n−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 wkn−1
= η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bkn ,n−2 ρn−2 Bk−1
n ,n−2
· (v1 . . . vn−4 vn−2 )
Bkn−1 +1,n−2
ρn−1 Bk−1 ,n−1
n−1
ρn−1 Bk−1
n−1 ,n−1
wkn−1
Bkn ,n−2 ρn−2 Bk−1
n ,n−2
· Bkn ,n−2 Bkn−1 +1,n−2
Let us count %(β) = %(ρs−1 αρs−1 ).
Step 1 An image of s under the braid β = ρs−1 αρs−1 is equal to ι(α)(s − 1).
ρs−1
ρs−1
α
s −−−−→ s − 1 −−−→ α(s − 1) −−−−→ α(s − 1)
Therefore s is a maximal number which is not fixed by β.
s−n
n−s
Step 2 β1 = B1,n
βB1,n
.
n−s
n−s
Step 3 Since ρs−1 B1,n
= B1,n
ρn−1 , then we have
s−n
n−s
s−n
n−s
s−n
n−s
β1 = B1,n
βB1,n
= B1,n
ρs−1 αρs−1 B1,n
= ρn−1 B1,n
αB1,n
ρn−1 = ρn−1 α1 ρn−1
= ρn−1 ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 ρn−1 Bkn ,n−1 ρn−1
ρn−1
ρn−1 ρn−1
= η ρn−1 (v1 . . . vn−2 )ρn−1 Bkn−1
Bkn ,n
,n−1 (w1 . . . wn−1 )
= η ρn−1 (v1 . . . vn−2 )ρn−1 (w1 . . . wn−1 )
Bk−1
n−1 ,n−1
ρn−1
ρ
ρ
n−1
n−1
Bkn−1
,n−1 Bkn ,n
Using simple calculations in the symmetric group, it is easy to see that
ρ
ρ
n−1
n−1
Bkn−1
,n−1 Bkn ,n = Bkn ,n−1 ρn−2 Bkn−1 +1,n ,
(11)
therefore
Bk−1
β1 = η ρn−1 (v1 . . . vn−2 )ρn−1 (w1 . . . wn−1 )
n−1 ,n−1
ρn−1
Bkn ,n−1 ρn−2 Bkn−1 +1,n
Bk−1 ,n−1 ρn−1
n−1
= η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
Bk−1
n−1
· (v1 . . . vn−2 )ρn−1 wkn−1
,n−1 ρn−1
Bkn ,n−1 ρn−2 Bkn−1 +1,n
Bk−1
= η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
n−1 ,n−1
ρn−1
Bkn ,n−1 ρn−2
Bk−1 ,n−1 ρn−1 Bkn ,n−1 ρn−2
n−1
· (v1 . . . vn−2 )ρn−1 Bkn ,n−1 ρn−2 wkn−1
Bk−1 ,n−1 ρn−1
n−1
= η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
Bkn−1 +1,n
Bkn ,n−1 ρn−2
Bk−1 ,n−1 ρn−1 Bkn ,n−1 ρn−2
n−1
· (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 wkn−1
31
ρ
n−1
vn−3
Bkn ,n−1 ρn−2
Bkn−1 +1,n
By Theorem 3 the braid
η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
Bk−1
n−1 ,n−1
ρn−1
Bkn ,n−1 ρn−2
does not involve the strand n, the braid
Bk−1
n−1
(v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 wkn−1
,n−1 ρn−1 Bkn ,n−1 ρn−2
ρ
n−1
belongs to hλ1,n , λn,1 i × · · · × hλn−2,n , λn,n−2 i and the braid vn−3
belongs to hλn−1,n , λn,n−1 i.
Bkn ,n−1 ρn−2
Step 4 The braid β2 follows
β2 = η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
Bk−1
n−1 ,n−1
ρn−1
Bkn ,n−1 ρn−2
Bk−1 ,n−1 ρn−1 Bkn ,n−1 ρn−2
n−1
· (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 wkn−1
ρn−1 Bkn−1 +1,n−1
Step 5 The braid %(β) is the following.
Bk−1
%(β) = η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
n−1 ,n−1
ρn−1
Bkn ,n−1 ρn−2
Bk−1 ,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1
n−1
· (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 wkn−1
Bkn−1 +1,n−1
Let us find the braid %2 (β).
Step 1 Since η ρn−1 does not involve the strands n − 1, n, then the image of n − 1 under
the permutation ι(%(β)) is equal to the image of n − 1 under the permutation
ι(Bkn ,n−1 ρn−2 Bkn−1 +1,n−1 ) and is equal to kn . Then n − 1 is a maximal number
which is not fixed by %(β).
Step 2 %(β)1 = %(β).
Step 3 The braid %(β)1 can be rewritten
%(β)1 = η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
Bk−1
n−1 ,n−1
ρn−1
−1
· (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1
Bk−1
n−1
· wkn−1
−1
,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1
· Bkn ,n−1 ρn−2 Bkn−1 +1,n−1
The following equality is faithful in the symmetric group
Bkn ,n−1 ρn−2 Bkn−1 +1,n−1 = ρn−3 Bkn−1 ,n−2 Bkn ,n−1 = Bkn−1 ,n−3 Bkn ,n−1 ,
32
(12)
therefore
Bk−1
%(β)1 = η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−1 )
n−1 ,n−1
ρn−1
−1
· (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1
Bk−1
n−1 ,n−1
· wkn−1
ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bk−1
n ,n−1
· Bkn−1 ,n−3 Bkn ,n−1
−1
= η ρn−1 (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1
Bk−1
· (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
Bk−1
−1
,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1
Bk−1
,n−1 ρn−1
n−1
· wkn−1
· wn−1n−1
ρn−1
Bkn−1 ,n−3 Bkn ,n−1
−1
= η ρn−1 (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 Bkn−1 ,n−3
Bk−1
· (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
ρn−1 Bkn−1 ,n−3
Bk−1
−1
,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 Bkn−1 ,n−3
Bk−1
,n−1 ρn−1 Bkn−1 ,n−3
n−1
· wkn−1
· wn−1n−1
Bkn ,n−1
−1
Here the braid η ρn−1 (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 Bkn−1 ,n−3 does
not involve the strand n − 1, the braid
(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bk−1
n−1 ,n−1
ρn−1 Bkn−1 ,n−3
·
Bk−1 ,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bk−1
B
n ,n−1 kn−1 ,n−3
n−1
· wkn−1
Bk−1
belongs to hλn−1,1 , λ1,n−1 i×· · ·×hλn−3,n−1 , λn−1,n−3 i and wn−1n−1
belongs to hλn−2,n−1 , λn−1,n−2 i.
,n−1 ρn−1 Bkn−1 ,n−3
Step 4 The braid %(β)2 follows.
−1
%(β)2 = η ρn−1 (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 Bkn−1 ,n−3
Bk−1
· (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bk−1
n−1
· wkn−1
n−1 ,n−1
ρn−1 Bkn−1 ,n−3
−1
,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 Bkn−1 ,n−3
33
ρn−2 Bkn ,n−2
Step 5 The braid %2 (β) has the following form.
−1
%2 (β) = η ρn−1 (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 Bkn−1 ,n−3
Bk−1
· (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bk−1
n−1 ,n−1
· wkn−1
n−1 ,n−1
ρn−1 Bkn−1 ,n−3 ρn−2
ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bk−1
B
ρ
n ,n−1 kn−1 ,n−3 n−2
Bkn ,n−2
ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bk−1
n ,n−1
= η ρn−1 (v1 . . . vn−4 vn−2 )
Bk−1
· (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bk−1
n−1
· wkn−1
n−1 ,n−1
ρn−1 Bkn−1 ,n−3 ρn−2 Bk−1
n−1 ,n−3
−1
−1
,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 Bkn−1 ,n−3 ρn−2 Bkn−1 ,n−3
· Bkn−1 ,n−3 Bkn ,n−2
Bk−1
= η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
ρn−1 Bkn−1 ,n−3 ρn−2 Bk−1
n−1 ,n−3
−1
· (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1
−1
· [(v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 ,
Bk−1
(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
ρn−1 Bkn−1 ,n−3 ρn−2 Bk−1
n−1 ,n−3
Bk−1 ,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bk−1
B
ρ
B −1
n ,n−1 kn−1 ,n−3 n−2 kn−1 ,n−3
n−1
· wkn−1
· Bkn−1 ,n−3 Bkn ,n−2
Now we have
Bk−1
%2 (β) = η ρn−1 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
ρn−1 Bkn−1 ,n−3 ρn−2 Bk−1
n−1 ,n−3
−1
· (v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1
−1
· [(v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 ,
Bk−1
(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
ρn−1 Bkn−1 ,n−3 ρn−2 Bk−1
n−1 ,n−3
Bk−1 ,n−1 ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bk−1
B
ρ
B −1
n ,n−1 kn−1 ,n−3 n−2 kn−1 ,n−3
n−1
· wkn−1
· Bkn−1 ,n−3 Bkn ,n−2
ρn−1 Bk−1
%2 (α) = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
−1
n−1 ,n−1
ρn−1 Bk−1
(v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 Bkn ,n−2 wkn−1
Bkn ,n−2 Bkn−1 +1,n−2
34
n−1 ,n−1
Bkn ,n−2 ρn−2 Bk−1
n ,n−2
]
]
Note that the following equalities are faithful in the symmetric group
Bkn−1 ,n−3 Bkn ,n−2 = Bkn ,n−2 Bkn−1 +1,n−2
ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bk−1
= ρn−1 ρn−2 ρn−3 ρn−2 ρn−1 ρn−2 ρn−3 ρn−2
n ,n−1
= ρn−1 ρn−3 ρn−2 ρn−3 ρn−1 ρn−2 ρn−3 ρn−2
= ρn−1 ρn−3 ρn−2 ρn−1 ρn−3 ρn−2 ρn−3 ρn−2
= ρn−1 ρn−3 ρn−2 ρn−1 ρn−3 ρn−3 ρn−2 ρn−3
= ρn−1 ρn−3 ρn−2 ρn−1 ρn−2 ρn−3
= ρn−3 ρn−1 ρn−2 ρn−1 ρn−2 ρn−3
= ρn−3 ρn−1 ρn−1 ρn−2 ρn−1 ρn−3
= ρn−3 ρn−2 ρn−1 ρn−3
Bkn ,n−2 ρn−2 Bk−1
n ,n−2
= ρn−3 ρn−2 ρn−3 ρn−1
= ρn−3 ρn−2 ρn−3
Therefore we have
−1
(v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 = ((v1 . . . vn−4 vn−2 )ρn−3 ρn−2 ρn−3 )ρn−1 ,
and since (v1 . . . vn−4 vn−2 )ρn−3 ρn−2 ρn−3 does not involve the strands n − 1, n, then
((v1 . . . vn−4 vn−2 )ρn−3 ρn−2 ρn−3 )ρn−1 =
= (v1 . . . vn−4 vn−2 )ρn−3 ρn−2 ρn−3 =
−1
= (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 Bkn ,n−2
and we have
−1
−1
(v1 . . . vn−4 vn−2 )ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bkn ,n−1 = (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 Bkn ,n−2 .
Also, since
Bk−1
ρ Bkn−1 ,n−3 ρn−2 Bk−1
= Bk−1
ρ ρ
n−1 ,n−1 n−1
n−1 ,n−3
n−1 ,n−1 n−1 n−2
= Bk−1
ρ ρ
n−1 ,n−1 n−1 n−2
= ρkn−1 . . . ρn−3 ρn−2 ρn−1 ρn−2
= ρkn−1 . . . ρn−3 ρn−1 ρn−2 ρn−1
= ρn−1 ρkn−1 . . . ρn−3 ρn−2 ρn−1 = ρn−1 Bk−1
ρ ,
n−1 ,n−1 n−1
then
B −1
ρ
B
ρ
B −1
(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 ) kn−1 ,n−1 n−1 kn−1 ,n−3 n−2 kn−1 ,n−3 =
ρn−1
ρ
B −1
= (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 ) n−1 kn−1 ,n−1
,
35
ρn−1 Bk−1
and since the braid (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
strands n − 1, n, then
Bk−1
(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
n−1 ,n−1
ρn−1 Bkn−1 ,n−3 ρn−2 Bk−1
does not involve the
n−1 ,n−3
=
= (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bk−1
n−1 ,n−1
wkn−1
ρn−1 Bkn ,n−1 ρn−2 ρn−1 ρn−2 Bk−1
B
ρ
B −1
n ,n−1 kn−1 ,n−3 n−2 k
n−1 ,n−3
ρn−1 Bk−1
n−1 ,n−1
=
ρn−1 Bk−1 ,n−1 Bkn ,n−2 ρn−2 Bk−1
n ,n−2
n−1
= wkn−1
Finally since η does not involve the strands n − 1, n, then η ρn−1 = η. Therefore the
0
braids %2 (α) and %2 (β) are equal modulo U V Pn−2
and by the remark 2 the braids
∗
∗
0
∗
% (β) and % (α) are equal modulo U V Pm , i. e. % (α) = %∗ (β).
Case 3.1.5.3.2. α(s − 1) < α(s) + 1. This case literally repeats the case 3.1.5.3.1
using the fact that in this case kn−1 < kn and then using the equalities
Bkn−1 ,n−1 Bkn ,n−1 = Bkn −1,n−2 Bkn−1 ,n−1
ρ
ρ
n−1
n−1
Bkn−1
,n−1 Bkn ,n = Bkn −1,n−1 ρn−2 Bkn−1 ,n
instead of equalities (11) and (12).
Case 3.1.6. i = s. In this case the braid β = ρs αρs maps s + 1 to ks .
ρs
α
ρs
s + 1 −−−→ s −−−→ ks −−−→ ks
Therefore s + 1 is a maximal number, which is not fixed by β and we have
s+1−n
n−s−1
s−n
n−s −1
β1 = B1,n
βB1,n
= B1,n B1,n
ρs−1 αρs−1 B1,n
B1,n
s−n
n−s
−1
= B1,n ρn−1 B1,n
αB1,n
ρn−1 B1,n
−1
−1
= B1,n ρn−1 α1 ρn−1 B1,n
= ρn−2 B1,n α1 B1,n
ρn−2
ρ
ρ
B −1
n−2
n−1
If we denote by δ1 = α1 , δ2 = δ1ρ1 , δ3 = δ2ρ2 , . . . , δn−1 = δn−2
, δn = δn−1
= α1 1,n ,
δn+1 = δnρn−2 = β1 , then it is obvious, that ι(δi )(n) 6= n for i = 1 . . . n − 1. Therefore
by the previous cases %∗ (δi ) and %∗ (δi+1 ) are conjugated by the element from Sn for
every i = 1 . . . n − 1.
%(δi ) = %(δi+1 )µi
Also since ι(δn )(n − 1) 6= n − 1
B1,n
B1,n
α
1
n − 1 −−−−→ n −−−
→ kn 6= n −−−−→ kn − 1 (mod n) 6= n − 1
36
then by the previous cases %∗ (δn ) = %∗ (δn+1 )
µ1
%∗ (β1 ) = %∗ (δ1 ) = %∗ (δ2 )
= %∗ (δ3 )
µ2 µ1
µn
and we have
µn−2 ...µ1
= · · · = %∗ (δn−1 )
µn−2 ...µ1
= %∗ (α1 )
Therefore the braids %∗ (β) = %∗ (β1 ) and %∗ (α) = %∗ (α1 ) are conjugated by the
element from Sn .
Case 3.1.7. i ≥ s + 1. In this case the maximal number which is not fixed by the
braid β = ρi αρi is s and we have
ρi
α
ρi
s −−−→ s −−−→ ks −−−→ ks
Since i ≥ s + 2, then we have
i−s
n−i−1
i−s
n−i−1
n−s
ρn−1 B1,n B1,n
= B1,n
B1,n B1,n
= ρi B1,n
ρi B1,n
n−i−1
i−s
n−i−1
i−s−1
= B1,n
B1,n−1 B1,n
= B1,n
B1,n−1 B1,n B1,n
n−i−1
i−s−1
n−i−1
i−s−1
= B1,n
B1,n B2,n B1,n
= B1,n
B1,n B2,n ρ1 ρ1 B1,n
n−i+1
i−s−1
n−i+1 i−s−1
n−s
= B1,n
ρ1 B1,n
= B1,n
B1,n ρi−s = B1,n
ρi−s
Therefore the braid β1 can be rewritten
s−n
n−s
s−n
n−s
s−n
n−s
β1 = B1,n
βB1,n
= B1,n
ρi αρi B1,n
= ρi−s B1,n
αB1,n
ρi−s = ρi−s α1 ρi−s
Since ι(α1 )(n) 6= n and i − s < n, then %∗ (α1 ) and %∗ (β1 ) are conjugated by the
previous cases.
Case 3.2. The braid β is obtained from α conjugating by λ±1
i,i+1 . If β is obtained
−1
from α conjugating by λi,i+1 , then α is obtained from β conjugating by λi,i+1 and
we can consider that β is obtained from α conjugating by λi,i+1 .
As we already found in the case 3.1 (equalities (8) and (7)) the braids %(α) and
α1 has the following forms
α1 = γw1 . . . wn−1 ρn−1 Bkn ,n−2 ,
%(α) = γ(w1 . . . wn−2 )ρn−1 Bkn ,n−1 ,
where kn = n − s + ks and ks is a maximal number which is not fixed by α.
Case 3.2.1. i ≤ ks − 2. Let us count the braid %(β) = %(λ−1
i,i+1 αλi,i+1 ).
Step 1. Since λi,i+1 is a pure braid, then the maximal number which is not fixed by β
is equal to the maximal number which is not fixed by β and is equal to s.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
37
n−s
n−s
Step 3. By Lemma 1 we have λi,i+1 B1,n
= B1,n
λi+n−s,i+n−s+1 . If we denote by j =
i + n − s, then the braid β1 can be rewritten in more details
s−n −1
n−s
n−s
s−n
λi,i+1 αλi,i+1 B1,n
= B1,n
βB1,n
β1 = B1,n
−1
s−n
n−s
= λ−1
j,j+1 B1,n αB1,n λj,j+1 = λj,j+1 α1 λj,j+1
= λ−1
j,j+1 γw1 . . . wn−1 Bkn ,n λj,j+1
B −1
kn ,n
= λ−1
j,j+1 γw1 . . . wn−1 λj,j+1 Bkn ,n
B −1
kn ,n
Since i < ks − 1, then j = i + n − s < kn − 1 and therefore λj,j+1
= λj,j+1 .
Therefore we have
−1
β1 = λj,j+1
γλj,j+1 w1 . . . wn−1 Bkn ,n ,
where the braid λ−1
j,j+1 γλj,j+1 does not involve the strand n and wr belongs to
hλr,n , λn,r i for r = 1, . . . , wn−1 .
Step 4. β2 = λ−1
j,j+1 γλj,j+1 w1 . . . wn−2 ρn−1 Bkn ,n−1 .
Step 5. The braid %(β) has the following form.
ρn−1
%(β) = λ−1
Bkn ,n−1
j,j+1 γλj,j+1 (w1 . . . wn−2 )
ρn−1
= λ−1
Bkn ,n−1 λj,j+1 = %(α)λj,j+1
j,j+1 γ(w1 . . . wn−2 )
Therefore %(α) and %(β) are conjugated and by the induction hypothesis the braids
%∗ (α) and %∗ (β) are conjugated by the element from Sn .
Case 3.2.2. i = ks − 1. Let us find the braid %(β).
Step 1. The maximal number which is not fixed by α is equal to s.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
Step 3. Similarly to the case 3.2.1 we conclude that β1 = λ−1
kn −1,kn α1 λkn −1,kn and therefore we have
β1 = λ−1
kn −1,kn γw1 . . . wn−1 Bkn ,n λkn −1,kn
B −1
kn ,n
= λ−1
kn −1,kn γw1 . . . wn−1 λkn −1,kn Bkn ,n
= λ−1
kn −1,kn γw1 . . . wn−1 λkn −1,n Bkn ,n
= λ−1
kn −1,kn γw1 . . . wn−2 λkn −1,n wn−1 Bkn ,n
Since kn < n, then the braid λ−1
kn −1,kn γ does not involve the strand n.
38
Step 4. β2 = λ−1
kn −1,kn γw1 . . . wn−2 λkn −1,n ρn−1 Bkn ,n−1 .
Step 5. The braid %(β) has the following form.
ρn−1
Bkn ,n−1
%(β) = λ−1
kn −1,kn γ(w1 . . . wn−2 λkn −1,n )
ρn−1
λkn −1,n−1 Bkn ,n−1
= λ−1
kn −1,kn γ(w1 . . . wn−2 )
B
ρn−1
n ,n−1
Bkn ,n−1 λknk−1,n−1
= λ−1
kn −1,kn γ(w1 . . . wn−2 )
ρn−1
= λ−1
Bkn ,n−1 λkn −1,kn = %(α)λkn −1,kn .
kn −1,kn γ(w1 . . . wn−2 )
Therefore the braids %(α) and %(β) are conjugated and by the induction hypothesis
the braids %∗ (α) and %∗ (β) are conjugated by the element from Sn .
Case 3.2.3. i = ks .
Case 3.2.3.1. ks ≤ s − 2. Let us count the braid %(β).
Step 1. The maximal number which is not fixed by α is equal to s.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
Step 3. Similarly to the case 3.2.1 we have β1 = λ−1
kn ,kn +1 α1 λkn ,kn +1 and therefore
β1 = λ−1
kn ,kn +1 γw1 . . . wn−1 Bkn ,n λkn ,kn +1
B −1
kn ,n
= λ−1
kn ,kn +1 γw1 . . . wn−1 λkn ,kn +1 Bkn ,n
= λ−1
kn ,kn +1 γw1 . . . wn−1 λn,kn Bkn ,n
= λ−1
kn ,kn +1 γw1 . . . wn−2 λn,kn wn−1 Bkn ,n
Since ks < s − 1, then kn , kn + 1 < n and the braid λ−1
kn ,kn +1 γ does not involve
the strand n.
Step 4. β2 = λ−1
kn ,kn +1 γw1 . . . wn−2 λn,kn ρn−1 Bkn ,n−1 .
Step 5. The braid %(β) has the following form.
ρn−1
%(β) = λ−1
Bkn ,n−1
kn ,kn +1 γ(w1 . . . wn−2 λn,kn )
ρn−1
λn−1,kn Bkn ,n−1
= λ−1
kn ,kn +1 γ(w1 . . . wn−2 )
B
kn ,n−1
ρn−1
= λ−1
Bkn ,n−1 λn−1,k
kn ,kn +1 γ(w1 . . . wn−2 )
n
ρn−1
Bkn ,n−1 λkn ,kn +1 = %(α)λkn ,kn +1
= λ−1
kn ,kn +1 γ(w1 . . . wn−2 )
Therefore the braids %(α) and %(β) are conjugated and by the induction hypothesis
the braids %∗ (α) and %∗ (β) are conjugated by the element from Sn .
Case 3.2.3.2 ks = s − 1.
Case 3.2.3.2.1. ι(α)(s − 1) = s. Let us find the braid %(β) = %(λ−1
ks ,ks +1 αλks ,ks +1 ).
39
Step 1. The maximal number which is not fixed by α is equal to s.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
Step 3. Similarly to the case 3.2.1 we have β1 = λ−1
kn ,kn +1 α1 λkn ,kn +1 . Since ks = s − 1,
−1
then kn = n − 1, Bkn ,n = ρn−1 and β1 = λn−1,n α1 λn−1,n . Then we have
β1 = λ−1
n−1,n γw1 . . . wn−1 ρn−1 λn−1,n .
Since ι(α)(s − 1) = s, then ι(α1 )(n − 1) = n, therefore the braid γ =
α1 (w1 . . . wn−1 ρn−1 )−1 maps n − 1 to n − 1. Hence the braids γ and λ−1
n−1,n
commute and we have
β1 = γλ−1
n−1,n w1 . . . wn−1 ρn−1 λn−1,n
= γw1 . . . wn−2 λ−1
n−1,n wn−1 ρn−1 λn−1,n
= γw1 . . . wn−2 λ−1
n−1,n wn−1 λn,n−1 ρn−1
where the braid γ does not involve the strand n, the braid wr belongs to
hλn,r , λr,n i for r = 1, . . . , n−2 and λ−1
n−1,n wn−1 λn,n−1 belongs to hλn,n−1 , λn−1,n i.
Step 4. β2 = γw1 . . . wn−2 ρn−1 .
Step 5. %(α) = %(β).
Therefore the braids %(α) and %(β) are conjugated and by the induction hypothesis
the braids %∗ (α) and %∗ (β) are conjugated by the element from Sn .
Case 3.2.3.2.2. ι(α)(s − 1) = ks−1 ≤ s − 2. As already founded in the case
3.1.3.2.2. for ks = s − 1 and ι(α)(s − 1) = ks−1 ≤ s − 2 we have
α1 = ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 ρn−1 ,
%(α) = ηv1 . . . vn−2 Bkn−1 ,n−1 (w1 . . . wn−2 )ρn−1 ,
ρn−1 Bk−1
%2 (α) = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
(v1 . . . vn−3 )ρn−2 Bkn−1 ,n−2 ,
where the braid η does not involve the strands n − 1, n the braid vr belongs to
hλr,n−1 , λn−1,r i for r = 1, . . . , n − 2 and the braid wr belongs to hλr,n , λn,r i for
r = 1, . . . , n − 1. Let us count the braid %(β) = %(λs−1,s αλs−1,s ).
Step 1. The maximal number which is not fixed by β is equal to s.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
40
Step 3. Similarly to the case 3.2.3.2.1. β1 = λ−1
n−1,n α1 λn−1,n and we have
β1 = λ−1
n−1,n ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 ρn−1 λn−1,n
= η λn−1,n λ−1
n−1,n v1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 λn,n−1 ρn−1
Bkn−1 ,n−1
w1 . . . wn−1 λn,n−1 ρn−1
= η λn−1,n v1 . . . vn−2 Bkn−1 ,n−1 (λ−1
n−1,n )
= η λn−1,n v1 . . . vn−2 Bkn−1 ,n−1 λ−1
kn−1 ,n w1 . . . wn−1 λn,n−1 ρn−1
Since the braid η does not involve the strands n − 1, n, then η λn−1,n = η.
Step 4. β2 = ηv1 . . . vn−2 Bkn−1 ,n−1 λ−1
kn−1 ,n w1 . . . wn−2 ρn−1 .
ρn−1
Step 5. %(β) = ηv1 . . . vn−2 Bkn−1 ,n−1 (λ−1
.
kn−1 ,n w1 . . . wn−2 )
Let us count the braid %2 (β).
Step 1. Since ks−1 ≤ s−2, then kn−1 ≤ n−2, therefore Bkn−1 ,n−1 6= 1 and the maximal
number which is not fixed by %(β) is equal to n − 1.
Step 2. %(β)1 = %(β).
Step 3. The braid %(β) can be rewritten
ρn−1
%(β) = ηv1 . . . vn−2 Bkn−1 ,n−1 (λ−1
kn−1 ,n w1 . . . wn−2 )
ρn−1 Bk−1
= ηv1 . . . vn−2 (λ−1
kn−1 ,n w1 . . . wn−2 )
n−1 ,n−1
= η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1 ,n−1
n−1
· vn−2 λ−1
n−1,n−2 wkn−1
ρn−1 Bk−1
n−1 ,n−1
ρn−1 Bk−1
Step 5. %2 (β) = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
v1 . . . vn−3
Bkn−1 ,n−1
ρn−1 Bk−1
Step 4. %(β)2 = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bkn−1 ,n−1
n−1 ,n−1
v1 . . . vn−3 ρn−2 Bkn−1 ,n−2 .
(v1 . . . vn−3 )ρn−2 Bkn−1 ,n−2 .
Therefore %2 (α) = %2 (β) and by the induction hypothesis the braids %∗ (α) and %∗ (β)
are conjugated by the element from Sn .
Case 3.2.4. ks + 1 ≤ i ≤ s − 2. Let us find the braid %(β) = %(λ−1
i,i+1 αλi,i+1 ).
Step 1. The maximal number which is not fixed by β is equal to s.
n−s
s−n
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
41
n−s
n−s
Step 3. By Lemma 1 we have λi,i+1 B1,n
= B1,n
λj,j+1 , where j = i + n − s. Since
ks + 1 ≤ i ≤ s − 2 then kn + 1 ≤ j ≤ n − 2 and the braid β can be rewritten
in more details.
s−n
n−s
s−n −1
n−s
β1 = B1,n
βB1,n
= B1,n
λi,i+1 αλi,i+1 B1,n
= λ−1
j,j+1 α1 λj,j+1
= λ−1
j,j+1 γw1 . . . wn−1 Bkn ,n λj,j+1
B −1
kn ,n
= λ−1
j,j+1 γw1 . . . wn−1 λj,j+1 Bkn ,n
= λ−1
j,j+1 γw1 . . . wn−1 λj−1,j Bkn ,n
= λ−1
j,j+1 γλj−1,j w1 . . . wn−1 Bkn ,n ,
Where the braid λ−1
j,j+1 γλj−1,j does not involve the strand n.
Step 4. β2 = λ−1
j,j+1 γλj−1,j w1 . . . wn−2 ρn−1 Bkn ,n−1 .
Step 5. The braid %(β) follows
ρn−1
%(β) = λ−1
Bkn ,n−1
j,j+1 γλj−1,j (w1 . . . wn−2 )
ρn−1
= λ−1
λj−1,j Bkn ,n−1
j,j+1 γ(w1 . . . wn−2 )
B
kn ,n−1
ρn−1
= λ−1
Bkn ,n−1 λj−1,j
j,j+1 γ(w1 . . . wn−2 )
ρn−1
= λ−1
Bkn ,n−1 λj,j+1 = %(α)λj,j+1
j,j+1 γ(w1 . . . wn−2 )
Therefore the braids %(α) and %(β) are conjugated and by the induction hypothesis
the braids %∗ (α) and %∗ (β) are conjugated by the element from Sn .
Case 3.2.5. i = s − 1. In this case we can consider that ks < s − 1 since the case
when ks = s − 1 = i is already solved in the case 3.2.3.2.
Case 3.2.5.1. ι(α)(s − 1) = s. From the case 3.1.5.1. we have
α1 = ηv1 . . . vn−2 w1 . . . wn−1 ρn−1 Bkn ,n−1 ,
%(α) = ηv1 . . . vn−2 (w1 . . . wn−2 )ρn−1 Bkn ,n−1 ,
%2 (α) = η(v1 . . . vn−3 )ρn−2 (w1 . . . wn−3 )ρn−1 ρn−2 Bkn ,n−2 ,
where the braid η does not involve the strands n − 1, n. Let us count %(β).
Step 1. The maximal number which is not fixed by β is equal to s.
n−s
s−n
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
42
n−s
n−s
Step 3. By Lemma 1 we have λs−1,s B1,n
= B1,n
λn−1,n , therefore
s−n
n−s
s−n −1
n−s
β1 = B1,n
βB1,n
= B1,n
λs−1,s αλs−1,s B1,n
= λ−1
n−1,n α1 λn−1,n
= λ−1
n−1,n ηv1 . . . vn−2 w1 . . . wn−1 Bkn ,n λn−1,n
B −1
kn ,n
= ηv1 . . . vn−2 λ−1
n−1,n w1 . . . wn−1 λn−1,n Bkn ,n
= ηv1 . . . vn−2 λ−1
n−1,n w1 . . . wn−1 λn−2,n−1 Bkn ,n
= ηv1 . . . vn−2 λn−2,n−1 w1 . . . wn−2 λ−1
n−1,n wn−1 Bkn ,n ,
where the braid ηv1 . . . vn−2 λn−2,n−1 does not involve the strand n.
Step 4. β2 = ηv1 . . . vn−2 λn−2,n−1 w1 . . . wn−2 ρn−1 Bkn ,n−1 .
Step 5. %(β) = ηv1 . . . vn−2 λn−2,n−1 (w1 . . . wn−2 )ρn−1 Bkn ,n−1 .
Let us count %2 (β).
Step 1. Since the braid η does not involve the strand n − 1, then the image of n − 1
under the permutation ι(%(β)) is equal to ι(Bkn ,n−1 )(n − 1) = kn 6= n − 1.
Step 2. %(β)1 = %(β).
Step 3. The braid %(α)1 has the following form.
%(β)1 = %(β) = ηv1 . . . vn−2 λn−2,n−1 (w1 . . . wn−2 )ρn−1 Bkn ,n−1
ρn−1
= ηv1 . . . vn−3 (w1 . . . wn−3 )ρn−1 vn−2 λn−2,n−1 wn−2
Bkn ,n−1
Step 4. %(β)2 = ηv1 . . . vn−3 (w1 . . . wn−3 )ρn−1 ρn−2 Bkn ,n−2
Step 5. %2 (β) = ηv1 . . . vn−3 (w1 . . . wn−3 )ρn−1 ρn−2 Bkn ,n−2 = %2 (α).
By the induction hypothesis the braids %∗ (α) and %∗ (β) are conjugated by the element from Sn .
Case 3.2.5.2. ι(α)(s − 1) = s − 1. In this case we have
α1 = ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 ρn−1 Bkn ,n−1
%(α) = ηv1 . . . vn−2 Bkn−1 ,n−1 (w1 . . . wn−2 )ρn−1 Bkn ,n−1 ,
where the braid η does not involve the strands n − 1, n and kn−1 ≥ kn . Since
ι(α)(s − 1) = s − 1, then ι(α1 )(n − 1) = n − 1, therefore kn−1 = n − 2 and
Bkn−1 ,n−1 = ρn−2 .
α1 = ηv1 . . . vn−2 ρn−2 w1 . . . wn−1 ρn−1 Bkn ,n−1
%(α) = ηv1 . . . vn−2 ρn−2 (w1 . . . wn−2 )ρn−1 Bkn ,n−1 ,
Let us count the braid %(β).
43
Step 1. The maximal number which is not fixed by β is equal to s.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
Step 3. Similarly to the case 3.2.5.1. we have β1 = λ−1
n−1,n α1 λn−1,n and therefore
β1 = λ−1
n−1,n ηv1 . . . vn−2 ρn−2 w1 . . . wn−1 Bkn ,n λn−1,n
B −1
kn ,n
= ηv1 . . . vn−2 λ−1
n−1,n ρn−2 w1 . . . wn−1 ρn−1 λn−1,n Bkn ,n
ρn−2
w1 . . . wn−1 λn−2,n−1 Bkn ,n
= ηv1 . . . vn−2 ρn−2 (λ−1
n−1,n )
= ηv1 . . . vn−2 ρn−2 λn−2,n−1 λ−1
n−2,n w1 . . . wn−1 Bkn ,n ,
where the braid = ηv1 . . . vn−2 ρn−2 λn−2,n−1 does not involve the strand n.
Step 4. β1 = ηv1 . . . vn−2 ρn−2 λn−2,n−1 λ−1
n−2,n w1 . . . wn−2 ρn−1 Bkn ,n−1 .
Step 5. The braid %(β) follows.
ρn−1
%(β) = ηv1 . . . vn−2 ρn−2 λn−2,n−1 (λ−1
Bkn ,n−1
n−2,n w1 . . . wn−2 )
ρn−1
%(β) = ηv1 . . . vn−2 ρn−2 λn−2,n−1 λ−1
Bkn ,n−1 = %(α).
n−2,n−1 (w1 . . . wn−2 )
By the induction hypothesis the braids %∗ (α) and %∗ (β) are conjugated by the element from Sn .
Case 3.2.5.3. ι(α)(s − 1) ≤ s − 2.
Case 3.2.5.3.1. ι(α)(s − 1) ≥ ι(α)(s) + 1. From the case 3.1.5.3.1. we have
α1 = ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 ρn−1 Bkn ,n−1
%(α) = ηv1 . . . vn−2 Bkn−1 ,n−1 (w1 . . . wn−2 )ρn−1 Bkn ,n−1
%2 (α) = η(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
n−1 ,n−1
Bkn ,n−2
ρn−1 Bk−1 ,n−1 Bkn ,n−2 ρn−2
n−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 wkn−1
Bkn−1 +1,n−2
Where the braid η does not involve the strands n − 1, n and kn−1 ≥ kn . Let us count
the braid %(β) = %(λ−1
n−1,n βλn−1,n ).
Step 1. The maximal number which is not fixed by β is equal to s.
s−n
n−s
Step 2. We have β1 = B1,n
βB1,n
and ι(β1 )(n) = ks + n − s = kn 6= n.
44
Step 3. Similarly to the case 3.2.5.1. we have β1 = λ−1
n−1,n α1 λn−1,n and therefore
β1 = λ−1
n−1,n ηv1 . . . vn−2 Bkn−1 ,n−1 w1 . . . wn−1 Bkn ,n λn−1,n
B −1
kn ,n
= ηv1 . . . vn−2 λ−1
n−1,n Bkn−1 ,n−1 w1 . . . wn−1 λn−1,n Bkn ,n
Bkn−1 ,n−1
w1 . . . wn−1 λn−2,n−1 Bkn ,n
= ηv1 . . . vn−2 Bkn−1 ,n−1 (λ−1
n−1,n )
= ηv1 . . . vn−2 Bkn−1 ,n−1 λ−1
kn−1 ,n w1 . . . wn−1 λn−2,n−1 Bkn ,n
= ηv1 . . . vn−2 Bkn−1 ,n−1 λn−2,n−1 λ−1
kn−1 ,n w1 . . . wn−1 Bkn ,n
where the braid ηv1 . . . vn−2 Bkn−1 ,n−1 λn−2,n−1 does not involve the strand n.
Step 4. β2 = ηv1 . . . vn−2 Bkn−1 ,n−1 λn−2,n−1 λ−1
kn−1 ,n w1 . . . wn−2 ρn−1 Bkn ,n−1 .
ρn−1
Step 5. %(β) = ηv1 . . . vn−2 Bkn−1 ,n−1 λn−2,n−1 (λ−1
Bkn ,n−1 .
kn−1 ,n w1 . . . wn−2 )
Let us count %2 (β).
Step 1 The maximal number which is not fixed by the braid %(β) is n − 1 and
ι(%(β))(n − 1) = kn−1 + 1.
Step 2 %(β)1 = %(β).
Step 3 The braid %(β)1 can be rewritten
ρn−1
%(α1 ) = ηv1 . . . vn−2 Bkn−1 ,n−1 λn−2,n−1 (λ−1
Bkn ,n−1
kn−1 ,n w1 . . . wn−2 )
Bk−1
ρn−1 Bk−1
n−1
(λ−1
= ηv1 . . . vn−2 λn−2,n−1
kn−1 ,n w1 . . . wn−2 )
,n−1
n−1 ,n−1
ρn−1 Bk−1 ,n−1
n−1
= ηv1 . . . vn−2 λn−3,n−2 (λ−1
kn−1 ,n w1 . . . wn−2 )
Bkn−1 ,n−1 Bkn ,n−1
Bkn−1 ,n−1 Bkn ,n−1
Since kn−1 ≥ kn , then
Bkn−1 ,n−1 Bkn ,n−1 = Bkn ,n−2 Bkn−1 +1,n−1
(13)
and and therefore
%(α1 ) = ηv1 . . . vn−2 λn−3,n−2 (λ−1
kn−1 ,n w1 . . . wn−2 )
ρn−1 Bk−1
n−1 ,n−1
Bkn ,n−2 Bkn−1 +1,n−1
ρn−1 Bk−1 ,n−1
n−1
= ηλn−3,n−2 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
· v1 . . . vn−2 (λ−1
kn−1 ,n wkn−1 )
n−1 ,n−1
Bkn ,n−2 Bkn−1 +1,n−1
ρn−1 Bk−1
= ηλn−3,n−2 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
· (v1 . . . vn−2 )Bkn ,n−2 (λ−1
kn−1 ,n wkn−1 )
n−1 ,n−1
ρn−1 Bk−1 ,n−1 Bkn ,n−2
n−1
Bkn−1 +1,n−1
ρn−1 Bk−1 ,n−1
n−1
= ηλn−3,n−2 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 (λ−1
kn−1 ,n wkn−1 )
45
Bkn ,n−2
Bkn ,n−2
Bkn ,n−2 Bkn ,n−2
n−1 ,n−1
vn−3 Bkn−1 +1,n−1 ,
where the braid ηλn−3,n−2 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
does not involve the strand n − 1.
ρn−1 Bk−1
n−1 ,n−1
Bkn ,n−2
Step 4 The braid %(β)2 has the following form
ρn−1 Bk−1
%(β)2 = ηλn−3,n−2 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 (λ−1
kn−1 ,n wkn−1 )
n−1 ,n−1
Bkn ,n−2
ρn−1 Bk−1 ,n−1 Bkn ,n−2
n−1
ρn−2 Bkn−1 +1,n−2 ,
Step 5 Finally, the braid %2 (α) follows
ρn−1 Bk−1
%2 (β) = ηλn−3,n−2 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
Bkn ,n−2
ρn−1 Bk−1 ,n−1 Bkn ,n−2 ρn−2
n−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 (λ−1
kn−1 ,n wkn−1 )
ρn−1 Bk−1
B
kn ,n−2
(w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
= ηBkn ,n−2 λn−3,n−2
ρn−1 Bk−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 (λ−1
kn−1 ,n wkn−1 )
n−1 ,n−1
= ηBkn ,n−2 λn−2,kn (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
ρn−1 Bk−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 λ−1
n−2,kn wkn−1
n−1 ,n−1
n−1 ,n−1
Bkn−1 +1,n−2
Bkn ,n−2
Bkn ,n−2 ρn−2
Bkn−1 +1,n−2
ρn−1 Bk−1 ,n−1 Bkn ,n−2
n−1
Bkn ,n−2 ρn−2
Bkn−1 +1,n−2
ρn−1 Bk−1 ,n−1 Bkn ,n−2
n−1
= ηBkn ,n−2 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 λn−2,kn
ρn−1 Bk−1
· [λn−2,kn , (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
Bkn ,n−2 ρn−2
· (v1 . . . vn−4 vn−2 )
n−1 ,n−1
Bkn ,n−2
ρn−1 Bk−1 ,n−1 Bkn ,n−2 ρn−2
−1
]λn−2,kn wkn−1 n−1
Bkn−1 +1,n−2
ρn−1 Bk−1
Bkn ,n−2
ρn−1 Bk−1
Bkn ,n−2
= ηBkn ,n−2 (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
n−1 ,n−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2
· [λn−2,kn , (w1 . . . wkn−1 −1 wkn−1 +1 . . . wn−2 )
−1
n−1 ,n−1
ρn−1 Bk−1
· (v1 . . . vn−4 vn−2 )Bkn ,n−2 ρn−2 ]λn−2,kn wkn−1
n−1 ,n−1
Bkn ,n−2 ρn−2
Bkn−1 +1,n−2
0
Therefore the braids %2 (α) and %2 (β) are equal modulo U V Pn−2
and by the remark
∗
∗
0
2 the braids % (α) and % (β) are equal modulo U V Pn , therefore %∗ (α) = %∗ (β).
Case 3.2.5.3.2. ι(α)(s − 1) ≤ ι(α)(s). This case literally repeats the case 3.2.5.3.1
using the fact that in this case kn−1 < kn and then using the equalities
Bkn−1 ,n−1 Bkn ,n−1 = Bkn −1,n−2 Bkn−1 ,n−1
46
instead of equality (13).
n−s
n−s
Case 3.2.6. i = s. By Lemma 1 we have λs,s+1 B1,n
= B1,n
λn,1 , therefore β1 =
−1
−1
λn,1 α1 λn,1 . By the definition λ1,n = B2,n λ1,2 B2,n , therefore
−1
−1
−1
−1
β1 = B2,n λ−1
1,2 B2,n α1 B2,n λ1,2 B2,n = B2,n λ1,2 δ1 λ1,2 B2,n ,
−1
where δ1 = B2,n
α1 B2,n . By the case 3.1 the braids %∗ (α1 ) and %∗ (δ1 ) are conjugated
by the element from Sn . Since α is not a pure braid, then α1 and δ1 are not pure
braids. Therefore the maximal number which is not fixed by δ1 is greater then or
equal to 2. Therefore, by the cases 3.2.1-3.2.5 the braids %∗ (δ1 ) and %∗ (λ−1
1,2 δ1 λ1,2 )
are conjugated by the element from Sn . Also by the case 3.1 the braids %∗ (λ−1
1,2 δ1 λ1,2 )
−1
∗
and %∗ (B2,n λ−1
1,2 δ1 λ1,2 B2,n ) = % (β1 ) are conjugated by the element from Sn , therefore
the braids %∗ (α1 ) = %∗ (α) and %∗ (β1 ) = %∗ (β) are conjugated by the element from
Sn .
Case 3.2.7. i ≥ s + 1. In this case β1 = λ−1
j,j+1 α1 λj,j+1 , where
n−s
j = ι(B1,n
)(i) = i + n − s (mod n) < n.
Since the maximal number which is not fixed by α1 is equal to n, then %∗ (α1 )
and %∗ (β1 ) are conjugated by the element from Sn by the cases 3.2.1-3.2.5. The
proposition is proved.
The following main result of the paper follows immediately from Corollary 1 and
Proposition 2.
Theorem 4. Let α and β be unrestricted virtual braids. Then their closures α
b and
∗
∗
b
β are equivalent as fused links if and only if % (α) and % (β) are conjugated by the
element from Sn < U V Bn .
References
[1] B. Audoux, P. Bellingeri, J.-B. Meilhan, E. Wagner, On forbidden moves and the
Delta move, ArXiv: Math/1510.04237.
[2] J. Baez, Link invariants of finite type and perturbation theory, Lett. Math. Phys., V. 26, N. 1,
1992, 43-51.
[3] V. Bardakov, P. Bellingeri, C. Damiani, Unrestricted virtual braids, fused links and
other quotients of virtual braid groups, accepted to J. Knot Theory Ramifications, V. 24,
N. 12, 2015.
[4] V. Bardakov, The virtual and universal braids, Fund. Math., V. 184, 2004, 1-18.
[5] J. Birman, New points of view in knot theory, Bull. Am. Math. Soc., New Ser., V. 28, N. 2,
1993, 253-287.
[6] J. Birman, Braids, links and mapping class group, Princeton–Tokyo: Univ. press, 1974.
47
[7] S. Chmutov, S. Duzhin, J. Mostovoy, Introduction to Vassiliev Knot Invariants, Cambridge Univ. Press, 2012.
[8] R. Fenn, R. Rimanyi, C. Rourke, The braid–permutation group, Topology, V. 36, N. 1,
1997, 123-135.
[9] A. Fish, E. Keyman, Classifying links under fused isotopy, ArXiv: math/0606198.
[10] C. Gordon, J. Luecke, Knots are determined by their complements. J. Amer. Math. Soc.,
V. 2, N. 2, 1989, 371-415.
[11] M. Goussarov, M. Polyak, O. Viro, Finite-type invariants of classical and virtual knots,
Topology, V. 39, N. 5, 2000, 1045-1068.
[12] D. Joyce, A classifying invariant of knots, the knot quandle, J. Pure and Appl. Alg., V. 23,
1982, 37-65.
[13] Z. Kadar, P. Martin, E. Rowell, Z. Wang, Local representations of the loop braid
group, ArXiv: Math/1411.3768.
[14] S. Kamada, Braid presentation of virtual knots and welded knots, Osaka J. Math, V. 44,
2007, 441-458.
[15] T. Kanenobu, Forbidden moves unknot a virtual knot, J. Knot Theory Ramifications, V. 10,
N. 1, 2001, 89-96.
[16] L. Kauffman, S. Lambropoulou, Virtual braids and the L-move, J. Knot Theory Ramifications, V. 15, N. 6, 2006, 1-39.
[17] L. Kauffman, S. Lambropoulou, Virtual braids. Fund. Math., V. 184, 2004, 159-186.
[18] L. Kauffman, Virtual knot theory, Eur. J. Comb., V. 20, N. 7, 1999, 663-690.
[19] V. Manturov, Knot theory, A CRC Press Company, London, New York, Washington, 2014.
[20] V. Vassiliev, Complements of discriminants of smooth maps, Topology and applications,
Translations of Mathematical Monographs, V. 98, Amer. Math. Soc., Providence, RI, 1992.
[21] V. Vershinin, On homology of virtual braids and Burau representation, J. Knot Theory
Ramifications, V. 10, N. 5, 2001, 795-812.
48
| 4math.GR
|
arXiv:1802.10578v1 [math.AC] 28 Feb 2018
Rings of constants of linear derivations on
Fermat rings
Marcelo Veloso
e-mail: [email protected]
Ivan Shestakov
e-mail: [email protected]
Abstract
In this paper we characterize all the linear C-derivations of the Fermat
ring. We show that the Fermat ring has linear C-derivations with trivial
ring of constants and construct some examples.
Keywords: Derivations, Fermat ring, ring of constants.
2010 AMS MSC: 13N15, 13A50, 16W25.
Introduction
The present paper deals with C-derivations of the Fermat ring
Bnm =
C[X1 , . . . , Xn ]
,
(X1m1 + · · · + Xnmn )
where C[X1 , . . . , Xn ] is the polynomial ring in n variables over the complex
numbers C, n ≥ 3, m = (m1 , . . . , mn ), mi ∈ Z and mi ≥ 2 for i = 1, . . . , n.
It is well known the difficulty to describe the ring of constants of an arbitrary
derivation (see [1, 3, 4, 5]). It is also difficult to decide if the ring of constants of
a derivation is trivial (see [5, 6, 7]). In this work we study the ring of constants
of linear derivations of Fermat rings and its locally nilpotent derivations.
In [5], Andrzej Nowicki presents a description of all linear C-derivations of
the polynomial ring C[X1 , . . . , Xn ] which do not admit any nontrivial rational
constant.
In a recent paper [1], P. Brumatti and M. Veloso show that for m = (2, . . . , 2)
the ring Bnm has nonzero irreducible locally nilpotent derivations. Furthermore,
whenever m1 = · · · = mn , They show that certain classes of derivations of
C[X1 , . . . , Xn ] do not induce derivations of Bnm or are not locally nilpotent if
they do.
In this work we obtain some similar results to [1], considering more general
Fermat rings. We present a description of all the linear C-derivations of Bnm
1
when m = (m1 , . . . , mn ) and mi ≥ 3 (Theorem 5) and m = (2, . . . , 2) (Theorem 6). We also provide examples of linear derivations of Bnm with trival ring
constants.
The text is organized as follows: Section 1 gathers the basic definitions and
notations. Further, we discuss several properties of the ring Bnm are discussed
and a set of generators for Der(Bnm ) is presented. Section 2 is dedicated to
the study of the linear derivations of the Fermat ring. The set of all locally
nilpotent C-derivations of Bnm is studied in Section 3. Finally, Sections 4 and 5
are devoted to the study of the rings of constants of linear derivations of Fermat
rings.
1
Preliminaries and Some Results
In this paper the word “ring" means a commutative ring with unit and characteristic zero. Furthermore, we denote the group of units of a ring R by R∗
and the polynomial ring in n variables over R by R[X1 , . . . , Xn ]. A “domain”
is an integral domain.
An additive mapping D : R → R is said to be a derivation of R if it
satisfies the Leibniz rule: D(ab) = aD(b) + D(a)b, for all a, b ∈ R. If A is a
subring of R and D is a derivation of R satisfying D(A) = 0 is called D an
A-derivation. The set of all derivations of R is denoted by Der(R), the set of
all A-derivations of R by DerA (R) and by ker(D), the ring of constants of
D, that is ker(D) = {a ∈ R | D(r) = 0}.
In this paper, the word "derivation" implicitly means a derivation which
is C-derivation and therefore we will use the notation Der(Bnm ) to denote
DerC (Bnm ). The residue classes of variables X, Y , Z, ... module an ideal
are represented by x, y, z, respectively. The symbol C is reserved to indicate
the field of complex numbers.
A derivation D is locally nilpotent if for each r ∈ R there is an integer
n ≥ 0 such that Dn (r) = 0. We denote by LN D(R) the set of all locally
nilpotent derivations of R.
We say that a element b ∈ R is a Darboux element of D ∈ Der(R) if
b 6= 0, b is not invertible in R and D(b) = λb for some λ ∈ R. In other words, a
nonzero element b of R is a Darboux element of D if, and only if, the principal
ideal (b) = {rb | r ∈ R} is different from R and it is invariant with respect to
D, that is D((b)) ⊂ (b). If b is a Darboux element of D, then every λ ∈ R, such
that D(b) = λb, is said to be an eigenvalue of b. In particular, every element
nonzero and noninvertible element belonging to the ring of constants, ker D, is
a Darboux element of D. If R is a domain and D(b) = λb, then it is easy to see
such the eigenvalue λ is unique.
Lemma 1 Let Bnm where m = (m1 , . . . , mn ). Then, for each f ∈ Bnm , there is
a unique F ∈ C[X1 , . . . , Xn ] such that degXn < mn and f = F (x1 , . . . , xn ).
Proof. It follows directly from the Euclidean division algorithm by considering
the polynomial X1m1 + · · · + Xnmn as a monic polynomial in Xn with coefficients
2
in C[X1 , . . . , Xn−1 ].
♦
Theorem 2 ([3, Theorem 4]) If n ≥ 5 and mi ≥ 2 for all 1 ≤ i ≤ n, then Bnm
is a unique factorization domain.
♦
mn
1
We also can write Bnm = C[x1 , . . . , xn ], where xm
= 0. Here
1 + · · · + xn
x1 , x2 , . . . , xn are the images of X1 , X2 , . . . , Xn under the canonical epimormn−1
1
phism C[X1 , . . . , Xn ] → Bnm .
An element of form axm
or
1 · · · xn−1
mn−1 j
m1
bx1 · · · xn1 xn , for 1 ≤ j ≤ mn − 1, is called monomial . A nonzero element f ∈ C[x1 , . . . , xn ] is said to be homogeneous element of degree k if f
is of the form
X
f=
a(i1 ···in ) xi11 · · · xinn
i1 +···+in =k
where 1 ≤ in ≤ mn − 1 and a(i1 ···in ) ∈ C for all (i1 · · · in ). We assume that the
zero element is a homogeneous element of any degree. Furthermore, we denote
by Vk the set of all homogeneous elements of degree k. Clearly Vk is a subspace
of Bnm .
1.1
A set of generators for Der(Bnm )
Now we will present a set of generators for the Bnm -module Der(Bnm ).
First some notation will be established. Given H ∈ S = C[X1 , . . . , Xn ]
and 1 ≤ i ≤ n, the partial derivative ∂(H)
∂Xi is denoted by HXi . For all pairs
∂
∂
−HXj
i, j ∈ {1, . . . , n} with i 6= j, we define the derivation DHij = HXi
∂Xj
∂Xi
on S. Observe that DHij (H) = 0.
[n]
Let A = CI be a finitely generated C-algebra. Consider the C[n] -submodule
DI = {D ∈ DerC (C[n] ) | D(I) ⊆ I} of the module DerC (C[n] ). It is well known
that the C[n] -homomorfism ϕ : DI → DerC (A) given by ϕ(D)(g + I) = D(g) + I
induces a C[n] -isomorfism of IDerDC I(C[n] ) in DerC (A).
The Theorem 3 will be needed, its proof can be found in [2, P roposition 1].
Theorem 3 Let F ∈ C[n] = C[X1 , . . . , Xn ] (n ≥ 2) be such that {FX1 , ..., FXn }
is a regular sequence in S. If there exists a derivation ∂ on S such that ∂(F ) =
αF for some α ∈ C, then the C[n] -module
DF := {D ∈ Der(S) | D(F ) ∈ F · C[n] }
F
is generated by the derivation ∂ and the derivations Dij = Dij
for i < j.
From now on, the derivations DFij , where F = X1m1 + · · · + Xnmn , will be
denoted by Dij. Since
m −1
if
k=i
−mj Xj j
Dij (Xk ) =
mi Ximi −1
if
k=j
0
if k ∈
/ {i, j}
3
so Dij (F ) = 0. Then Dij ∈ Der(S) induces dij = mi ximi −1
in Der(Bnm ). Consider the derivation
E=
∂
m −1 ∂
− mj xj j
∂xj
∂xi
∂
∂
1
1
X1
+ ··· +
Xn
.
m1
∂X1
mn
∂Xn
Note that E satisfies E(F ) = F . Hence, E ∈ Der(S) induces
ε=
1
∂
∂
1
x1
+ ··· +
xn
∈ Der(Bnm )
m1 ∂x1
mn ∂xn
As a consequence of Theorem 3 the following result is obtained:
Proposition 4 If F = X1m1 + · · · + Xnmn then DF := {D ∈ Der(S) | D(F ) ∈
F · S} is generated by the derivation E and the derivations Dij , i < j. In
particular the Bnm -module Der(Bnm ) is generated by the derivation ε and by the
derivations dij , for i < j.
Proof. Since {m1 X m1 −1 , . . . , mn X mn −1 } is a regular sequence and E(F ) = F
the result following by Theorem 3.
♦
2
Linear derivations
This section is dedicated to the study of the linear derivations of the Fermat
ring
Bnm = C[x1 , . . . , xn ],
mn
1
where xm
1 + · · · + xn = 0.
A derivation d of the ring Bnm is called linear if
d(xi ) =
n
X
aij xj for i = 1, . . . , n, where aij ∈ C.
j=1
The matrix [d] = [aij ] is called the associated matrix of the derivation d.
Theorem 5 Let d ∈ Der(Bnm ) be linear. If m = (m1 , . . . , mn ) with mi ≥ 3
for all i = 1, . . . , n, then its associated matrix [d] is a diagonal matrix and has
the following form
α
m1
for some α ∈ C.
α
m2
..
.
α
mn
4
.
Proof. Let [d] = [aij ] be the associated matrix of d. Then d(xi ) =
n
X
aij xj , for
j=1
mn
1
all i. Since xm
1 + · · · + xn = 0,
m1 x1m1 −1 d(x1 ) + · · · + mn xnmi −1 d(xn ) = 0.
Then,
n
n
n
X
X
X
n −1
anj xj )
a2j xj ) + · · · + mn xm
(
a1j xj ) + m2 x2m2 −1 (
0 = m1 x1m1 −1 (
n
j=1
j=1
j=1
(2.1)
Now note that
n
n
X
X
1
a1j xj ) =m1 a11 (xm
)
+
m
a1j xj x1m1 −1
m1 x1m1 −1 (
1
1
j6=1
j=1
mn
2
=m1 a11 (−xm
2 − · · · − xn ) + m1
n
X
a1j xj x1m1 −1
j6=1
and
n
n
X
X
2
a2j xj ) = m2 a22 xm
+
m
a2j xj x2m2 −1
m2 x2m2 −1 (
2
2
j6=2
j=1
..
.
n
n
X
X
mn
n −1
n −1
+
m
a
x
)
=
m
a
x
anj xj xm
mn xm
(
n
nj j
n nn n
n
n
j6=n
j=1
replacing in the Equation (2.1) we obtain
mn
2
0 = (m2 a22 − m1 a11 )xm
2 + · · · + (mn ann − m1 a11 )xn + m1
n
X
a1j xj x1m1 −1 +
j6=1
m2
n
X
a2j xj x2m2 −1 + · · · + mn
j6=2
n
X
n −1
anj xj xm
.
n
j6=n
(2.2)
Observe that if mi ≥ 3, then
mi −1
mn
1
{xm
| 1 ≤ i < j ≤ n, } ∪ {xj ximi −1 | 1 ≤ j < i ≤ n}
2 , . . . , xn } ∪ {xj xi
is a linearly independent set over C. Thus, we conclude that
mn ann = · · · = m2 a22 = m1 a11 = α and aij = 0 if i 6= j,
5
i.e.
aij =
if
if
0
α
mi
i 6= j
i=j
♦
This theorem shows that for m = (m1 , . . . , mn ) and mi ≥ 3 linear derivations
of B2m are what is called diagonal derivations.
The next result characterizes linear derivations of Bnm whenever
m = (2, . . . , 2). Previously, remember that a square matrix with complex elements A is said to be skew-symmetric matrix if AT = −A (here AT stands,
of course, for the transpose of the matrix A).
Theorem 6 Let d ∈ Der(Bnm ) be linear. If m = (2, . . . , 2), then there exist a
scalar derivation dα ([dα ] is a scalar matrix) and a skew-symmetric derivation
ds ([ds ] a skew-symmetric matrix) such that d = dα + ds . This decomposition is
unique.
Proof. Let d ∈ Der(Bnm ) be a linear derivation and A = [aij ] its associated
matrix. Using the same arguments used in Theorem 5 we obtain
X
(aij + aji )xi xj
0 = (a22 − a11 )x22 + · · · + (ann − a11 )x2n +
i<j
{x22 , . . . , x2n }
Since the set
C, it follows that
∪ {xi xj ; 1 ≤ i < j ≤ n} is linearly independent over
a11 = a22 = · · · = ann = α and aij = −aji if i < j,
then its associated matrix [d] has the following form
α
a12 . . . a1n
−a12
α
a2n
..
..
.. .
.
.
.
.
.
.
−a1n −a2n . . . α
where α, aij ∈ C. Now define dα by dα (xi ) = αxi , i = 1, . . . , n and ds = d − dα .
♦
3
Locally Nilpotent Derivations
In this section we proof that the unique locally nilpotent derivation linear of
Bnm for m = (m1 , . . . , mn ) and mi ≥ 3 is the zero derivation. Further, we show
that a certain class of derivations of C[X1 , . . . , Xn ] do not induce nonzero locally
nilpotent derivation of Bnm .
[n]
Let S = CI be a finitely generated C-algebra. Consider the C[n] -submodule
DI = {D ∈ DerC (C[n] ) | D(I) ⊆ I} of the module DerC (C[n] ). It is well known
6
that the C[n] -homomorfism ϕ : DI → DerC (S) given by ϕ(D)(g + I) = D(g) + I
induces
a
C[n] -isomorfism of IDerDC I(C[n] ) in DerC (S). From this fact the following result is
obtained.
Proposition 7 Let d be a derivation of the Bnm . If d(x1 ) = a ∈ C and for each
i, 1 < i ≤ n, d(xi ) ∈ C[x1 , . . . , xi−1 ], then d is the zero derivation.
Proof.
Let F be the polynomial X1m1 + · · · + Xnmn . We know that exists D ∈ Der(C[n] ) such that D(F ) ∈ F C[n] and that d(xi ) = D(Xi ) +
F C[n] , ∀i. Thus D(X1 ) − a ∈ F C[n] , and for each i > 1 there exists Gi =
Gi (X1 , . . . , Xi−1 ) ∈ C[X1 , . . . , Xi−1 ] such that D(Xi ) − Gi ∈ F C[n] . Since
n
X
mi Ximi −1 D(Xi ) ∈ F C[n] and
D(F ) =
i=1
D(F ) =
n
X
mi Ximi −1 (D(Xi ) − Gi ) +
mi Ximi −1 Gi ,
i=1
i=1
where G1 = a, we obtain
n
X
n
X
mi Ximi −1 Gi ∈ F C[n] and then obviously Gi = 0
i=1
for all i. Thus d is the zero derivation.
Lemma 8 Let d be a linear derivation of Bnm and [aij ] its associated matrix.
Then d is locally nilpotent if and only if [aij ] is nilpotent.
Proof. The following equality can be verified by induction over s.
ds (x1 )
..
.
s
d (xn )
x1
s
= [aij ] ... .
xn
(3.3)
We know that d is locally nilpotent if and only if there exists r ∈ N such that
dr (xi ) = 0 for all i. As {x1 , . . . , xn } is linearly independent over C by the (3.3),
the result follows.
♦
Theorem 9 If d ∈ LN D(Bnm ) is linear and m = (m1 , . . . , mn ) wich mi ≥ 3,
then d is the zero derivation.
Proof. Since d is locally nilpotent, [d] is nilpotent (by Lemma 8) and diagonal
(by Theorem 5). Thus, the matrix [d] is null and d is the zero derivation.
♦
In the case m = (2, . . . , 2), linear locally nilpotent derivations of the ring
Bnm were characterized as follows.
Theorem 10 [1, Theorem 1] If d ∈ Der(Bnm ) is linear and m = (2, . . . , 2),
then d ∈ LN D(Bn2 ) if, and only if, its associated matrix is nilpotent and skewsymmetric.
♦
7
4
Ring of constants
In this section we show that the ring of constants of all nonzero linear derivations
of Bnm , where m = (m1 , . . . , mn ) and mi ≥ 3, is trivial, that is ker(d) = C.
During all this section we always consider m = (m1 , . . . , mn ) with mi ≥ 3.
The next result ensures the existence of Darboux elements for every nonzero
linear derivation of Bnm .
Proposition 11 Let d be a nonzero linear derivation of Bnm . If d(xi ) = mαi xi ,
i = 1, . . . , n, for some α ∈ C∗ , then f = bxi11 · · · xinn is a Darboux element of d,
d(f ) = λf , and
i1
i2
in
λ=α
.
+
+ ···+
m1
m2
mn
Proof. Let f = bxi11 · · · xinn . Then
d(f ) =d(bxi11 · · · xinn )
=bd(xi11 · · · xinn )
n
X
=b
ik xi11 · · · xikk −1 · · · xinn d(xk )
=b
k=1
n
X
ik xi11
k=1
n
X
=αb
k=1
=bα
=λf
· · · xikk −1
· · · xinn
α
xk
mk
ik i1
x · · · xinn
mk 1
i1
i2
in
+
+ ··· +
m1
m2
mn
xi11 · · · xinn
♦
Corollary 12 Let d be a nonzero linear derivation of Bnm . If f is a homogeneous element of degree k, then f is a Darboux element of Bnm with eigenvalue
k
λ= m
.
Proof. Let f =
X
a(i1 ···in ) xi11 · · · xinn be a homogeneous element of degree
i1 +···+in =k
k, where 0 ≤ in < m and a(i1 ···in ) ∈ C.
8
X
d(f ) =d
a(i1 ···in ) xi11
· · · xinn
i1 +···+in =k
=
X
a(i1 ···in ) d(xi11 · · · xinn )
X
a(i1 ···in )
X
a(i1 ···in )
k i1
x · · · xinn
m 1
!
i1 +···+in =k
=
i1 +···+in =k
=
i1 +···+in =k
k
=
m
X
i1
i2
in
+
+ ···+
m m
m
a(i1 ···in ) xi11
i1 +···+in =k
=λf
· · · xinn
xi11 · · · xinn
!
♦
The main result this section is:
Theorem 13 Let d be a nonzero linear derivation of Bnm . Then ker(d) = C.
Proof. By Theorem 5, d(xi ) =
α 6= 0. Let 0 6= f ∈
Bnm
α
mi xi
for i = 1, . . . , n and α ∈ C. Since d 6= 0, so
X
b(i1 ,...,in ) xi11 · · · xinn
such that d(f ) = 0. Thus f =
(i1 ,...,in )∈I
where 0 6= b(i1 ,...,in ) ∈ C for all (i1 , . . . , in ) ∈ I. Then
X
b(i1 ···in ) d(xi11 · · · xinn )
X
i2
in
i1
xi11 · · · xinn
+
+ ···+
=
b(i1 ···in ) α
m1
m2
mn
0 = d(f ) =
i1
i2
It follows from Lemma 1 that b(i1 ···in ) α( m
+ m
+ · · · + minn ) 6= 0 for all
1
2
(i1 , . . . , in ) ∈ I, because b(i1 ···in ) α 6= 0 for all (i1 , . . . , in ) ∈ I. So mi11 + mi22 +
· · · + minn = 0 for all (i1 , . . . , in ) ∈ I. This implies that (i1 , . . . , in ) = (0, . . . , 0)
for all (i1 , . . . , in ) ∈ I. Therefore f ∈ C.
♦
Theorem 14 Let d ∈ Der(Bnm ) be given by d(xi ) = mαi , i = 1, . . . , n, for some
α ∈ C. If
X
a(i1 ,...,in ) xi11 · · · xinn
f=
(i1 ,...,in )∈I
is a darboux element of d, this is, d(f ) = λf for some λ ∈ Bnm , then
i1
+ mi22 + · · · + minn ) for all (i1 , . . . , in ) ∈ I.
λ = α( m
1
9
X
Proof. Let f =
a(i1 ,...,in ) xi11 · · · xinn . It follows from Theorems 13 and 5
that
X
a(i1 ···in ) b(i1 ...in ) xi11 · · · xinn
d(f ) =
(i1 ,...,in )∈I
where b(i1 ···in ) =
So
X
α( mi11
+
i2
m2
+ ··· +
in
mn )
∈ C. Then
a(i1 ···in ) b(i1 ...in ) xi11 · · · xinn = d(f ) = λf =
X
λa(i1 ···in ) xi11 · · · xinn
a(i1 ...in ) b(i1 ···in ) = λa(i1 ...in )
for all (i1 , . . . , in ) ∈ I, by Lemma 1. Therefore λ = b(i1 ...in ) = α( mi11 +
· · · + minn ) for all (i1 , . . . , in ) ∈ I.
5
i2
m2
+
♦
The case m = (2, . . . , 2)
In this section we focus on the Fermat rings
Bnm =
C[X1 , . . . , Xn ]
,
(X12 + · · · + Xn2 )
where m = (2, . . . , 2) and n ≥ 3. For simplicity we denote Bnm by Bn2 .
We study linear derivations of Bn2 , their rings of constants, and we show how
to construct examples of linear derivations with trivial ring of constants.
The next result will be useful for this purpose.
Proposition 15 Let d be a nonzero linear derivation of Bnm , with d = dα + ds ,
where dα is the scalar derivation and is skew-symmetric derivation. Let dα
definid by dα (xi ) = αxi for i = 1, . . . , n and α ∈ C. If f ∈ Bn2 is homogeneous
element of degree k then d(f ) = λf if, and only if, ds (f ) = (λ − kα)f .
Proof. It is easy to see that dα (f ) = kaf . Suppose that d(f ) = λf . Then
λf = d(f ) = dα (f ) + ds (f ) = −αkf + ds (f ).
Hence,
ds (f ) = (λ − kα)f.
Now suppose d1 (f ) = (λ − ka)f . Then
d(f ) = dα (f ) + ds (f ) = kαf + (λ − kα)f = λf.
♦
Corollary 16 Let d = dα + ds be a nonzero linear derivation of Bn2 , where
dα is the scalar derivation and ds is skew-symmetric derivation. If f ∈ Bn2 is
homogeneous element of degree k then d(f ) = 0 if, and only if, ds (f ) = −kαf .
10
Proof. Consider λ = 0 in Proposition 15.
♦
The next Theorem shows that every skew-symmetric derivation has a nontrivial ring of constants and every nonzero scalar derivation has trivial ring of
constants.
Theorem 17 Let d = dα + ds be a nonzero linear derivation of Bn2 , where dα
is the scalar derivation and ds is skew-symmetric derivation. Then
1. If ds is zero the derivation, then ker(d) = ker(dα ) is trivial.
2. If dα is the zero derivation, then ker(d) = ker(ds ) is nontrivial.
Proof. 1) Observe that dα (f ) = kaf for all homogeneous element of degree k
and dα (Vk ) ⊂ Vk .
2) It suffices to prove that there is f ∈ Bn2 such that ds (f ) = 0 and f 6∈ C. Let
f a homogeneous element of degree 2 of Bn2 , then f = XAX T where A = [aij ]
is a symmetric matrix and X = (x1 , . . . , xn ). Observe that for B = [ds ] we have
d(f ) = d(XAX T ) =(XB)AX T + XA(XB)T
=XBAX T + XA(−BX T )
=XBAX T − XABX T
=X(BA − AB)X T .
If B 2 6= 0 consider the symmetric matrix B 2 . It follows from the above remark
that for f = XB 2 X T we have ds (f ) = 0 and f 6∈ C, because A = B 2 6= 0. If
B 2 = 0, then λ = 0 is eigenvalue of B and B T . In this case, choose a nonzero
element f = a1 x1 + · · · + an xn such that the nonzero vector (a1 , . . . , an )T is an
eigenvector of B T . So d(f ) = 0 and f 6∈ C. Therefore, ker(ds ) 6= C.
♦
We also show that there are linear derivations with dα 6= 0, ds 6= 0, and
trivial ring of constants.
Theorem 18 Let ds be a nonzero skew-symmetric derivation of Bn2 . Then
exists a scalar derivation dα of Bn2 such that the derivation d = dα + ds satisfies
ker(d) = C.
Proof. First note that the vector space Vk (the set of homogeneous elements
of degree k of Bn2 ) is invariant with respect to ds . Hence ds (f ) = 0 if only if
ds (fk ) = 0 for all homogeneous components fk of f . As a consequence of this
fact we assume that f is a homogeneous element of degree k. Let α be a nonzero
complex number that satisfies the conditions
1. α ∈
/ Spec(ds ),
2. for all positive integer k, −kα ∈
/ Spec(ds | Vk ).
11
This number exists because C is uncountable and the set of the numbers that
satisfies 1) and 2) are countable. Let α be a number that satisfies the conditions
1) and 2), then ds (f ) 6= αf and ds (f ) 6= −kαf , for all f ∈
/ C and for all positive
integer k . Let dα be a scalar derivation defined by dα (xi ) = αxi , i = 1, . . . , n.
Finally, by considering the derivation d = dα + ds , we show that ker(d) = C.
In order to do that, if g ∈ Bn2 is a nonzero homogeneous element of degree k,
then d(g) = 0 if, and only if, ds (g) = −kαg, by Corollary 16, which implies
k = 0. Therefore, g ∈ C∗ .
♦
We now provide and explicit example of such derivation:
Example 19 Let d = d1 + ds be the linear derivation of B32 = C[x, y, z] given
by
1 0
0
1 0 0
0 0
0
[d] = [d1 ] + [ds ] = 0 1 −1 = 0 1 0 + 0 0 −1 .
0 1
1
0 0 1
0 1
0
We claim that ker(d) = C. To be more precise, let Vk be the set of all homogeneous elements of degree k in C[y, z]. The set
Sk = {y k , y k−1 z, . . . , yz k−1 , z k }
is a basis for Vk . The matrix
k
1
0
−k
k
2
0 −(k − 1)
k
0
−(k − 2)
[d | Vk ] =
.
.
.
.
.
.
0
..
0
0
.
0
0
0
...
...
..
.
...
...
0
0
..
.
..
.
k−2
0
..
.
k
k−1
−2
0
k
−1
...
0
0
..
.
0
0
k
k
is the matrix the linear derivation d restrict to subspace Vk in the basis Sk . It
is easy check that Det([d | Vk ]) 6= 0 for all k ≥ 1, by the principle of induction. Then d(f ) 6= 0 for all homogeneous elements of degree k ≥ 1. Therefore,
ker(d) = C.
Theorem 20 Let d = dα + ds be a nonzero linear derivation of Bn2 , where dα
is a scalar derivation and ds is a skew-symmetric derivation. If ds is a locally
nilpotent derivation then ker(d) = C, for all nonzero scalar derivation dα .
Proof. Let 0 6= f ∈ Bn2 such that d(f ) = 0. It suffices to show that f ∈ C. We
may assume that f is a nonzero homogeneous element of degree k, because Vk is
invariant by d. Let m be the smallest positive integer such that g = dm−1
(f ) 6= 0
s
and dm
(f
)
=
0,
this
m
exists
because
d
is
a
locally
nilpotent
derivation.
This
s
s
12
implies that g is a nonzero homogeneous element of degree k, because Vk is
invariant by d. One easily verifies that dα ds = ds dα and dds = ds d. Note that
d(g) = dα (g) + ds (g) = kαg, because g is an homogeneous element of degree k
and dm
s (f ) = 0. Now observe that
d(g) = d(dm−1
(f )) = dm−1
(d(f )) = dm−1
(0) = 0.
s
s
s
We thus get kαg = 0. And so k = 0. Therefore, f ∈ C.
♦
To conclude this section we construct two families of examples which illustrate this theorem.
Example 21 Let n ≥ 3 be an odd number
defined by the skew-symmetric matrix n × n
0 0 ... 0
0 0 ... 0
.. .. . .
.
. ..
[ds ] = . .
0 0 ... 0
0 0 ... 0
1 i ... 1
and ds a linear derivation of Bn2
0
0
..
.
0
0
i
−1
−i
..
.
.
−1
−i
0
It is easy to check that [ds ]3 = 0, which implies that [ds ] is nilpotent. Then ds
is a locally nilpotent linear derivation of Bn2 , by Theorem 10. Now consider the
linear derivation d = d1 + ds , where d1 (xi ) = xi for i = 1, . . . , n. It follows from
Theorem 20 that ker(ds ) = C.
Example 22 Let n ≥ 4 be an even number and ε ∈ C a primitive
(n − 1)-th root of unity. Set ds a linear derivation of Bn2 by the skew-symmetric
matrix n × n
0 0 ... 0 ...
0
−1
0 0 0 ... 0
0
−ε
.. .. . .
.. . .
..
..
. .
.
.
.
.
.
k
0
0
.
.
.
0
.
.
.
0
−ε
[ds ] =
. . .
.
.
.
.
. . ..
..
..
..
.. ..
n−2
0 0 ... 0 ...
0
−ε
1 ε . . . εk . . . εn−2
0
Again, [ds ] is nilpotent ([de ]3 = 0). Thus, ds is a locally nilpotent derivation of
Bn2 , by Theorem 10. Now considering the linear derivation d = d1 + de , where
d1 is the same as in the previous example, we conclude that ker(ds ) = C, by
Theorem 20.
Remark: In the Example 19 it easy see that [ds ] is not nilpotent and,
consequently ds is not locally nilpotent (Theorem 10). This shows that ds locally
nilpotent is not a necessary condition in Theorem 20 for ker(dα + ds ) = C.
13
References
[1] P. Brumatti and M. Veloso, On locally nilpotent derivations of Fermat
Rings, Algebra and Discrete Mathematics 16 (1), 33–41, (2013).
[2] P. Brumatti and M. Veloso, A note on Nakai’s conjecture for the ring
K[X1 ,...,Xn ]
, Colloquium Mathematicum 123 (2), 277–283, (2011).
m
(a X 1 +···+a X mn )
1
1
n
n
[3] D. Fiston and S. Maubach, Constructing (almost) rigid rings and a UFD
having infinitely generated Derksen and Makar-Limanov invariant, Canadian Mathematical Bulletin 53 (1), 77-86, (2010).
[4] G. Freudenberg, Algebraic Theory of Locally Nilpotent Derivations, Encyclopaedia of Mathematical Sciences 136, Springer-Verlag Berlin Heidelberg,
(2006).
[5] A. Nowicki, On the nonexistence of rational first integrals for systems of
linear differential equations, Linear Algebra and Its Applications 235, 107120, (1996).
[6] J. M. Ollagnie and A. Nowicki, Derivations of polynomial algebras without
Darboux polynomials, Journal of Pure and Applied Algebra 212, 1626-1631,
(2008).
[7] J. Zielinski, Rational Constants of Generic LV Derivations and of Monomial Derivations, Bulletin of the Polish Academy of Sciences. Mathematics
61, Issue: 3, 201-208, (2013).
14
| 0math.AC
|
1
On the Feasibility of Generic Deep
Disaggregation for Single-Load Extraction
arXiv:1802.02139v1 [cs.LG] 5 Feb 2018
Karim Said Barsim and Bin Yang
{karim.barsim,bin.yang}@iss.uni-stuttgart.de
Institute of Signal Processing and System Theory, University of Stuttgart
Abstract— Recently, and with the growing development of big
energy datasets, data-driven learning techniques began to represent a potential solution to the energy disaggregation problem
outperforming engineered and hand-crafted models. However,
most proposed deep disaggregation models are load-dependent
in the sense that either expert knowledge or a hyper-parameter
optimization stage is required prior to training and deployment
(normally for each load category) even upon acquisition and
cleansing of aggregate and sub-metered data. In this paper,
we present a feasibility study on the development of a generic
disaggregation model based on data-driven learning. Specifically,
we present a generic deep disaggregation model capable of
achieving state-of-art performance in load monitoring for a
variety of load categories. The developed model is evaluated on
the publicly available UK-DALE dataset with a moderately low
sampling frequency and various domestic loads.
Index Terms— Energy/Load disaggregation, Non-Intrusive
Load Monitoring (NILM), Convolutional Neural Networks
(CNN), UNet, SegNet, UK-DALE
I. I NTRODUCTION
Energy disaggregation (or Non-Intrusive Load Monitoring
NILM) is the process of inferring individual load profiles
at the end-use level from a single or a limited number of
sensing points. Promising applications of disaggregated data
have motivated a growing research community to reach a
widely acceptable and scalable solution. Energy disaggregation
proved to be a challenging source separation problem in which
a considerably large number of parameters are to be estimated
from a limited set of measurements with little constraints.
In the last decade, energy disaggregation has witnessed
an unprecedented wide-spreading research which is easily
observed from the wide variety of learning techniques applied
to this problem alongside with the growing number of energy
datasets developed specifically for this research field. More
recently, and analogous to the current breakthrough in datadriven learning, deep neural networks have re-gained their
interest in addressing the energy disaggregation problem,
especially alongside with the recently developed large energy
datasets required for training such complex models [1–8].
The progress in this trend, however, is relatively slow when
compared to the development in either field separately. This
is sometimes attributed to the high risk of over-fitting in
neural network models [9], insufficiency or low diversity of
publicly available energy datasets [4], or limited insights and
understanding of the learning behavior of these models [5].
In this paper, we first present a feasibility study on the
development of a generic data-driven model suitable for enduse load monitoring. The proposed disaggregation model
exploits a fully convolutional neural network architecture
and is generic in the sense that none of the model hyperparameters is dependent on the load category. We assess the
feasibility of such a model through empirical evaluation of
the monitoring performance across various load categories in
a publicly available energy dataset.
II. R ELATED WORK
In this section, we briefly describe some of the most recent
works on energy disaggregation and load monitoring that
adopted data-driven learning techniques.
Mauch and Yang [2] exploited a generic two-layer bidirectional Recurrent Neural Network (RNN) architecture featuring
Long Short Term Memory (LSTM) [10] units in extracting
single load profiles. They tested their models on the Refrence
Energy Disaggregation Dataset (REDD) [11] in a de-noised
scheme [12]. Additionally, they validated the generalization of
their architecture to previously unseen loads in new buildings.
In a later work, Mauch and Yang [3] used a combination of
discriminative and generative models in a two-stage eventless
extraction of load profiles. Kelly and Knottenbelt [4] evaluated
and compared three neural network architectures on domestic
loads from the UK-Domestic Appliance Level Energy (UKDALE) [7]. The first is a bidirectional RNN architecture with
LSTM units similar to the one in [2], the second follows the
architecture of a de-noising Auto-Encoder (dAE) [13], and the
last is a regression-based disaggregator whose objective is to
estimate the main key points of an activation cycle of the target
load within a given window.
Similarly, He and Chai [6] applied two architectures, namely
a convolutional dAE and an RNN, to the same problem. In
their architectures, they also applied parallel convolutional
layers with different kernel sizes analogous to the Inception
module in GoogLeNet [14]. Zhang et al. [5] simplified the
objective of the dAE architecture in [4] to predict a single
time instance of the target load profile for a given window
of the aggregate signal. Likewise, Nascimento [15] applied
three neural network architectures, namely basic convolutional
dAE, an RNN, and a ResNet-based model [16] to the same
problem but on three target loads in the REDD dataset. He
introduced several improvements such as redefining the loss
function, exploiting batch normalization [17], and applying
residual connections [16].
Additionally, Lange et al. [18] adopted a deep neural
network with constrained binary and linear activation units
in the last two layers. Their first objective was to retrieve
subcomponents of the input signal that sum up linearly to the
aggregate active and reactive powers. Finally, they estimate
the on-off activation vector of each load. Their approach,
however, was applied on very high frequency current and
voltage measurements (12 kHz) from the Building-Level fUlly
labeled dataset for Electricity Disaggregation (BLUED) [19].
In many of these previous works [4–6, 15], each disaggregator is a neural network whose disaggregation window length
(and consequently the width of subsequent layers) depends
on the load being monitored. The disaggregation window of
each load is manually adjusted in a per-load basis to fully
capture a single activation cycle of the load. Moreover, the
disaggregation performance widely differs amongst variant
2
load categories and a model that achieves remarkably well
on one load might drastically fail for other load types.
III. L OAD M ONITORING
In this work, we focus essentially on single-load extraction
of activation profiles, of which we give a detailed description
in the following.
A. Activation profiles: definition and estimation
In the simplest case, a load is modeled as a two-state
machine which is assumed to be in the on-state whenever the
load is consuming energy from the main power source, and in
the off -state otherwise. Accordingly, load monitoring becomes
a binary classification task. Note that in contrast to previous
works, the consumption profile of a load during its on-state
need not be a defined [20] nor a piecewise-defined function
in time.
The desired signal (i.e. ground truth) of a load m in
a window of N time instances is the binary-valued signal
ω (m) ∈ {0, 1}N whose element ω (m) (n) is set (i.e. to indicate
an on-state) whenever the load is operating in one of its
activation states at time instance n and unset otherwise. In
this work, we refer to this signal as the activation profile. Applications that benefit from activation profiles include mainly
activity monitoring and occupancy detection in which timeof-use information dominates the value of energy consumed.
We define the true activation profile of a load ω (m) via
a threshold-based approach applied to the sub-metered real
power signals and similar to the one used in [4] as follows.
The sub-metered real-power x(m) of a load m is compared
against predefined thresholds to detect the operation intervals
of the load. In order to avoid anomalies and false activations or
deactivations, the load is assumed to be in an activation state
(i.e. on) if its power draw x(m) (n) exceeds a given threshold
(m)
(m)
Pon for a minimum period of time Non . Similarly, if the
(m)
power draw drops below a predefined threshold Poff for a
(m)
given period Noff , the load is assumed to be disconnected.
Otherwise, the load keeps its last observed state. Thus, the
estimated activation profile is defined as
(m)
(m)
if x(m) (k) > Pon , for n 6 k < n + Non
1,
(m)
(m)
(m)
ω (n) = 0,
if x(m) (k) 6 Poff , for n 6 k < n + Noff
(m)
ω (n − 1),
otherwise
with the initial state assumed to be off (i.e. ω (m) (0) = 0) for
(m)
(m)
(m)
(m)
all loads. Note that Pon , Non , Poff , and Noff are the only
load-dependent parameters in this work, and they are used
merely in estimating the ground truth signals. Values of these
parameters are similar or close to those adopted in [4] and are
listed in Table I for the sake of completeness.
B. Single load extraction
In single-load extraction, each disaggregator targets exclusively a single load in the monitored circuit and normally
ignores dependencies amongst loads. While exploiting loads’
dependencies is expected to improve the performance of a
disaggregation system in a given building [9, 21, 22], it is also
likely to reduce the generalization capability of such a system
to new, previously unseen buildings. This is because such
dependencies originate not only from the physical architecture
of the power line network and the assumed signal model but
also from the usage behavior of end-consumers which varies
widely from one building to another, especially within the
residential sector [23].
TABLE I: Load-dependent parameters for estimating the activation profiles.
Load
Pon = Poff [W]
Non [min.]
Noff [min.]
5
10
10
20
20
5
25
1000
50
300
1
1
30
30
1
3
5
1/3
1/6
1/6
1
1
5
5
1
3
5
1/6
1/6
1/20
Fridge (FR)
Lights (LC)
Dishwasher (DW)
Washing machine (WM)
Solar pump (SP)
TV
Boiler (BL)
Kettle (KT)
Microwave (MC)
Toaster (TS)
The load monitoring problem is modeled as K-separate binary classification tasks. Given a window of K samples of the
k=K−1
starting
aggregate real power signal x(n) = [x(n + k)]k=0
(m)
at the time instance n, the model g (x(n), θ) ∈ [0, 1]K
estimates the posterior probabilities of the activation profile
for the analogous K time instances of mth load where θ is the
model parameters (e.g. weights and biases in a neural network)
p ω (m) (n) = 1 x(n) = g (m) (x(n); θ)
(1)
g (m) (x(n); θ) = σ g̃ (m) (x(n); θ)
(2)
X = ( x(0), x(K), x(2K), . . . , x((NK − 1) · K))
(3)
where the disaggregator’s output is bound to the valid range
of a probability function via a logistic sigmoid activation in
the output layer of the network
where g̃ represents the sub-network from the input layer to
activation signals of the output layer and σ is the logistic
sigmoid function Eq. 5 applied element-wise to g̃.
In the training phase, we refer to the pair (x(n), ω(n))
as a single training segment with K data samples. Training
segments are extracted from the whole time series signals
(x, ω) using non-overlappling windows which results in a
training set whose inputs segments are
and the corresponding activation segments
Ω = ( ω(0), ω(K), ω(2K), . . . , ω((NK − 1) · K))
(4)
where NK = ⌊N/K⌋ is number of training segments (with the
·(m) notation omitted for brevity). Assuming all segments (and
the K outputs of each segment) are conditionally independent
given the input vector x(n) and identically distributed (i.i.d),
then the likelihood function becomes
NK −1 K−1
p (Ω| X, θ) =
Y Y
n=0
p (ω(k + n · K) | x(n · K))
k=0
with ω being a Bernoulli distributed random variable
NK −1 K−1
p (Ω| X, θ) =
Y Y
n=0
gk (x(n))ωk · (1 − gk (x(n)))1−ωk
k=0
where gk (x(n)) is the k th output of the disaggregator. The
negative log-likelihood -LL then becomes
NK −1 K−1
-LL = −
X X
n=0
ωk · ln gk (x(n)) + (1 − ωk ) ln(1 − gk (x(n)))
k=0
which is known as binary cross-entropy and it is the adopted
loss function in all our experiments. The choice of a logistic
sigmoid activation function in the output layer together with
the binary cross-entropy as the objective function is a standard
combination in binary classification problems [24].
Finally, the following decision rule is used to estimate the
final class labels
3
ω̂(n) =
(
0 if p(ω(n) = 0 | x(n)) > p(ω(n) = 1 | x(n))
Aggregate signal x(n)
1 otherwise
We point out that the concept of load activation cycles, a
complete cycle of operation off → on→ off, is not considered.
In other words, an activation cycle of a load can extend
over several K-length segments (such as lighting circuits
and dishwashers) or arise more than once within the same
window segment (as in fridge and kettle activations). This is
an important property since a disaggregator need not wait till
the deactivation of a load (i.e. switch-off event) but rather can
provide near real-time feedback from a partial segment of the
activation, normally with some delay.
+
Encoder Sub-Net
IV. M ODEL A RCHITECTURE
Figure 1 shows the architecture of the proposed fully
convolutional neural network model. The model consists of 46
layers in five parts (an input layer, 40 encoding and decoding
layers, 4 representation layers, and an output layer) reaching
41M trainable parameters. Each layer includes a sequence
of elementary operations shown in the figure and briefly
introduced in the sequel.
Dilated Convolutions CONV(d, k): The core operation
of each layer is the cross-correlation defined as
+
+
k=⌊k/2⌋
def
CONV(d, k): f (x) = b(n)+
X
+
+
x(n + d · k) · κ(k)
where d is the dilation rate [25], k is the kernel size, b is the
bias vector, and κ is the layer’s kernel.
Batch Normalization BN [17]: is a composition of two
affine transformations applied to the output of each layer based
on mini-batch statistics
def
BN: f (x) = γ x̂ + β = γ
Representation
k=-⌊k/2⌋
x − µB
+β
σB
+
+
2
where x is the original output of a unit, µB and σB
are the
sample mean and variance of all outputs of this neuron over
the mini-batch B, and γ and β are two learnable parameters.
Leaky Rectified Linear Units LReLU [26]: is a non-linear
activation function defined as
+
α≤1
Activation noise (noise injection) GN [27]: is a regularization technique applied during the training phase only
and consists of injecting small additive Gaussian noise (with
variance σ 2 ) to the output of the layer to avoid over-fitting
def
Decoder Sub-Net
LReLU: f (x) = max(αx, x)
+
2
GN: f (x) = x + z ∼ N(0, σ )
Sigmoidal activations LogSg: is a bounded activation
function applied to the first hidden layer and the output layer
of the model
def
LogSg: f (x) = (1 + exp(−x))−1
(5)
Down- and up-sampling: take place only across blocks
where down-sampling is performed using MaxPooling while
up-sampling is applied using forward-filling.
Parameter initialization and updates: model parameters
are initialized from a zero-mean uniform distribution [28] and
learned using a gradient-based stochastic optimization [29]
with an update rule based on the ADAM Algorithm [30] with
Nesterov momentum [31].
V. P ERFORMANCE M EASURES
Early works on energy disaggregation tended to adopt the
simple accuracy index in evaluating the performance of a
+
+
64x
10800
CONV(7, 1), BN, LogSg, GN
32x
10800
CONV(7, 3), LReLU, BN, GN
32x
10800
CONV(7, 3), LReLU, BN, GN
32x
10800
CONV(7, 3), LReLU, BN, GN
32x
10800
CONV(7, 3), LReLU, BN, GN
64x
3600
CONV(7, 3), LReLU, BN, GN
64x
3600
CONV(7, 3), LReLU, BN, GN
64x
3600
CONV(7, 3), LReLU, BN, GN
64x
3600
CONV(7, 3), LReLU, BN, GN
128x
1200
CONV(7, 3), LReLU, BN, GN
128x
1200
CONV(7, 3), LReLU, BN, GN
128x
1200
CONV(7, 3), LReLU, BN, GN
128x
1200
CONV(7, 3), LReLU, BN, GN
256x
400
CONV(7, 3), LReLU, BN, GN
256x
400
CONV(7, 3), LReLU, BN, GN
256x
400
CONV(7, 3), LReLU, BN, GN
256x
400
CONV(7, 3), LReLU, BN, GN
512x
80
CONV(7, 3), LReLU, BN, GN
512x
80
CONV(7, 3), LReLU, BN, GN
512x
80
CONV(7, 3), LReLU, BN, GN
512x
80
CONV(7, 3), LReLU, BN, GN
1024x
16
CONV(7, 3), LReLU, BN, GN
1024x
16
CONV(7, 3), LReLU, BN, GN
1024x
16
CONV(7, 3), LReLU, BN, GN
1024x
16
CONV(7, 3), LReLU, BN, GN
512x
80
CONV(7, 3), LReLU, BN, GN
512x
80
CONV(7, 3), LReLU, BN, GN
512x
80
CONV(7, 3), LReLU, BN, GN
512x
80
CONV(7, 3), LReLU, BN, GN
256x
400
CONV(7, 3), LReLU, BN, GN
256x
400
CONV(7, 3), LReLU, BN, GN
256x
400
CONV(7, 3), LReLU, BN, GN
256x
400
CONV(7, 3), LReLU, BN, GN
128x
1200
CONV(7, 3), LReLU, BN, GN
128x
1200
CONV(7, 3), LReLU, BN, GN
128x
1200
CONV(7, 3), LReLU, BN, GN
128x
1200
CONV(7, 3), LReLU, BN, GN
64x
3600
CONV(7, 3), LReLU, BN, GN
64x
3600
CONV(7, 3), LReLU, BN, GN
64x
3600
CONV(7, 3), LReLU, BN, GN
64x
3600
CONV(7, 3), LReLU, BN, GN
32x
10800
CONV(7, 3), LReLU, BN, GN
32x
10800
CONV(7, 3), LReLU, BN, GN
32x
10800
CONV(7, 3), LReLU, BN, GN
32x
10800
CONV(7, 3), LReLU, BN, GN
1x
10800
/3
/3
/3
/5
/5
x5
x5
x3
x3
x3
CONV(7, 3), BN, LogSg
Activation profile g(x(n))
Fig. 1: Architecture of the proposed energy monitoring model. Dashed
and dotted lines to the right represent outer and inner skip connections,
respectively. Solid lines to the left represent residual connections [16]. Skip
connections use channel aggregation while residual connections use elementwise addition. Green-shaded layers are followed by a pooling step, while the
red-yellow shaded ones are preceded by an un-pooling operation.
4
disaggregation system [32–34]. Later works, however, realized
the misleading interpretation of this measure (resulting from
its bias towards the prevailing class) and proposed precision,
recall, and f1 -score as alternative measures for assessing the
disaggregation performance [12, 21, 35, 36].
We, however, believe that these measures represent a onesided rigorous solution to the biasness of the accuracy index.
In fact, these metrics are fused to the assumption of scarce
load usage and fail to provide valuable interpretation of
performance if this assumption is violated.
Given the raw-count contingency table
Classes
Predictions
ω̂ +
ω̂ ω+
TP
FN
RP
-
FP
TN
RN
PP
PN
N
ω
TP:
TN:
FP:
FN:
RP:
RN:
PP:
PN:
N:
True Positives
True Negatives
False Positive
False Negatives
Real Positives
Real Negatives
Positive Predictions
Negative Predictions
Num. of samples/events
FR
LC
DW
WM
SP
TV
BL
KT
MC
TS
KL
rn
%-NM
TPA
TPR
B
M
f1 -s
MCC
0.55
0.69
0.98
0.94
0.78
0.90
0.91
0.99
0.99
0.99
0.87
0.87
0.92
0.95
0.91
0.97
0.97
0.95
0.95
0.97
0.98
0.94
0.92
0.52
0.85
0.97
0.46
0.74
0.34
0.87
0.62
0.67
0.46
0.88
0.67
0.49
0.99
0.24
0.69
0.75
0.87
0.46
0.72
0.55
0.81
0.39
0.49
0.99
0.16
0.67
0.60
0.87
0.45
0.72
0.46
0.82
0.35
0.84
0.96
0.27
0.71
0.31
0.87
0.62
0.67
0.39
0.896
0.589
0.623
0.979
0.312
0.716
0.468
0.870
0.526
0.698
0.502
0.815
0.373
0.641
0.978
0.204
0.686
0.431
0.869
0.529
0.697
0.422
TABLE III: Performance comparison between the proposed model AE and the
rectangles architecture in [4] Regr on same load instances (left) and unseen
instances from new buildings (right). All values represent the f -measure.
the aforementioned measure are defined as
accuracy = (TP + TN) / (TP + TN + FP + FN)
precision = TPA = TP / (TP + FP)
recall = TPR = TP / (TP + FN)
f1 -s = (2 x TP) / (2 x TP + FN + FP)
TABLE II: Performance comparison of 11 loads from the first building in
UK-DALE [7]. rn is the probability of the negative class in the evaluation
fold and %-NM is the percent-noisy measure [12].
(6)
(7)
In the case of scarce load usage, the probability of negative
samples becomes relatively high and the accuracy index becomes a single-sided measure, namely the true negative rate. In
this case, a trivial disaggregator (one that always predicts the
prevailing class) ambiguously yields near optimal accuracy.
The information retrieval approach to alleviate this bias is to
simply ignore the prevailing term in Eq. 6, namely TN, which
results in either the Jaccard index or the f -measure f1 -s
Eq. 7. We find this to be an extreme and ill-argued solution,
especially in assessing energy monitoring performance. First,
the scarce load usage is not always valid and is usually
violated in commercial buildings or some residential loads
such as refrigerators, air conditioners, space heaters, or electric
vehicles. Additionally, the class of always-on loads suffers
from the exact opposite situation where the class imbalance
is due to the prevailing positive class and a trivial system in
this case yields misleading near-optimal score for both the
accuracy and the f -measure.
Second, when the scarce usage assumption is valid (e.g.
for various miscellaneous appliances such as kettles, irons,
vacuum cleaners ... etc), the extent of class imbalance varies
widely amongst loads as well as users. These variations are not
reflected by any means in either of the information retrieval
measures. For these reasons, we claimed that precision and
recall are inflexible measures since they are fused to a onesided assumption regardless of the real distribution of classes.
Powers [37] introduced informedness B, markedness M, and
their geometric mean Matthews Correlation Coefficient MCC
as alternative, unbiased evaluation measures
B = TPR + TNR - 1
(8)
M = TPA + TNA - 1
(9)
√
(10)
MCC = B · M
where TNA = TN / (TN + FN) is the inverse-precision
and TNR = TN / (TN + FP) is the inverse recall. Similar
to the information retrieval measures, these alternatives were
proposed and adopted in similar application domains such as
medical diagnostics [38, 39] and recommender system evaluations [40]. We believe that the requirements of performance
Load
FR
DW
MC
WM
KT
Same instances
Accross buildings
Regr. [4]
AE
Regr. [4]
AE
0.810
0.720
0.620
0.490
0.710
0.879
0.796
0.705
0.960
0.783
0.820
0.740
0.210
0.270
0.700
0.927
0.804
0.366
0.410
0.819
evaluation in these applications are more similar to those in
energy disaggregation.
VI. E XPERIMENTS AND R ESULTS
The developed model is evaluated on the freely available
UK-DALE dataset [7], an energy dataset acquired from five
residential buildings. In this work, the 1 Hz real power
measurements represent the input signals to disaggregate while
the reference ones are the 1/6 Hz measurements up-sampled
(using fill-forward) to 1 Hz.
Table II shows the performance measures of the proposed
model evaluated on 11 loads from the first building in the
adopted dataset with a 3-hour monitoring window for all
load categories. Data folds are real power measurements from
January and February of 2015 for training and validation,
respectively, while the remaining 10 months of the 2015
represents the evaluation fold. While we provide these results
as benchmarking ones, assessment of feasibility is observed
in the following experiment.
In Table III, we compare the monitoring performance of
our model AE with the previous work in [4], specifically
the regression-based model Regr. (referred to as rectangles
architecture). We use the exact data folds adopted in [4] for
training and evaluation and define two test cases. The first
trains and evaluates on the same load instances but future
periods of operation while the second evaluates on new load
instances (from new buildings). In both cases, the proposed
model outperformed previous works in all load categories.
VII. C ONCLUSION AND FUTURE WORK
In this paper, we assessed the feasibility of a generic deep
disaggregation model for end-use load monitoring using a
fully convolutional neural network evaluated on a variety of
load categories. The proposed model (with a fixed architecture and set of hyper-parameters) outperforms previous work
and showed relatively acceptable performance across different
loads.
5
R EFERENCES
[1] K. K. Kaman, M. Faramarzi, S. Ibrahim, and M. A. M.
Yunus, “Artificial Neural Network for Non-Intrusive Electrical
Energy Monitoring System”. Indonesian Journal of Electrical
Engineering and Computer Science 6 (1):124–131, Apr. 2017.
[2] L. Mauch and B. Yang, “A new approach for supervised power
disaggregation by using a deep recurrent LSTM network”. In
proceedings of the 3rd IEEE Global Conference on Signal and
Information Processing (GlobalSIP):63–67, Dec. 2015.
[3] ——, “A novel DNN-HMM-based approach for extracting
single loads from aggregate power signals”. In proceedings of
the IEEE International Conference on Acoustics, Speech and
Signal Processing (ICASSP):2384–2388, Mar. 2016.
[4] J. Kelly and W. J. Knottenbelt, “Neural NILM: Deep
Neural Networks Applied to Energy Disaggregation”. CoRR
abs/1507.06594 , Aug. 2015.
[5] C. Zhang, M. Zhong, Z. Wang, N. Goddard, and
C. Sutton, “Sequence-to-point learning with neural networks
for nonintrusive load monitoring”. arXiv, Dec. 2016.
[6] W. He and Y. Chai, “An Empirical Study on Energy Disaggregation via Deep Learning”. The 2016 2nd International
Conference on Artificial Intelligence and Industrial Engineering
(AIIE2016), Beijing, ChinaNov. 2016.
[7] J. Kelly and W. J. Knottenbelt, “The UK-DALE dataset,
domestic appliance-level electricity demand and whole-house
demand from five UK homes”. Scientific Data, Feb. 2015.
[8] O. Parson, G. Fisher, A. Hersey, N. Batra, J. Kelly, A. Singh,
W. Knottenbelt, and A. Rogers, “”Dataport and NILMTK: A
Building Data Set Designed for Non-intrusive Load Monitoring””. In proceedings of the 3rd IEEE Global Conference
on Signal and Information Processing (GlobalSIP), Orlando,
Florida, USADec. 2015.
[9] S. Makonin, , “Real-Time Embedded Low-Frequency Load
Disaggregation”. .”
[10] S. Hochreiter and J. Schmidhuber, “Long short-term memory”.
Neural Comput. 9 (8):1735–1780, Nov. 1997.
[11] J. Z. Kolter and M. J. Johnson, “REDD: A public data set for
energy disaggregation research”. proceedings of the SustKDD
Workshop on Data Mining Applications in Sustainability, San
Diego, CA, USAApr. 2011.
[12] S. Makonin and F. Popowich, “Nonintrusive Load Monitoring
(NILM) Performance Evaluation: A unified approach for accuracy reporting”. Energy Efficiency 8 (4):809–8142015
[13] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.A. Manzagol, “Stacked Denoising Autoencoders: Learning
Useful Representations in a Deep Network with a Local
Denoising Criterion”. Journal of Machine Learning Research
11 :3371–3408, Dec. 2010.
[14] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov,
D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going deeper
with convolutions”. 2015 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR):1–9, Jun. 2015.
[15] P. Nascimento, , “Applications of Deep Learning Technologies
on NILM”. , Brazil, , Apr. 2016.
[16] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning
for Image Recognition”. , Dec. 2015.
[17] S. Ioffe and C. Szegedy, “Batch Normalization: Accelerating
Deep Network Training by Reducing Internal Covariate Shift”.
, Mar. 2015.
[18] H. Lange and M. Bergés, “BOLT: Energy Disaggregation by
Online Binary Matrix Factorization of Current Waveforms”.
Proceedings of the 3rd ACM International Conference on
Systems for Energy-Efficient Built Environmentsser. BuildSys
’16:11–20New York, NY, USA, 2016.
[19] K. Anderson, A. Ocneanu, D. Benitez, D. Carlson, A. Rowe,
and M. Berges, “BLUED: a fully labeled public dataset
for Event-Based Non-Intrusive load monitoring research”.
Proceedings of the 2nd KDD Workshop on Data Mining
Applications in Sustainability (SustKDD), Beijing, ChinaAug.
2012.
[20] M. Zeifman and K. Roth, “Viterbi algorithm with sparse
transitions (VAST) for nonintrusive load monitoring”. 2011
IEEE Symposium on Computational Intelligence Applications
In Smart Grid (CIASG):1–8, Apr. 2011.
[21] H. Kim, M. Marwah, M. F. Arlitt, G. Lyon, and J. Han,
“Unsupervised Disaggregation of Low Frequency Power Mea-
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
surements”. proceedings of the 11th International Conference
on Data Mining:747–758. Arizona: SIAM, 2010.
J. Z. Kolter, S. Batra, and A. Y. Ng, “Energy Disaggregation
via Discriminative Sparse Coding”. Proceedings of the 23rd
International Conference on Neural Information Processing
Systemsser. NIPS’10:1153–1161USA, 2010.
N. Batra, O. Parson, M. Berges, A. Singh, and A. Rogers,
“A Comparison of Non-Intrusive Load Monitoring Methods
for Commercial and Residential Buildings”. arXiv preprint,
arXiv:1408.6595
C. M. BishopPattern Recognition and Machine Learning (Information Science and Statistics). Secaucus, NJ, USA: SpringerVerlag New York, Inc., 2006.
F. Yu, V. Koltun, and T. Funkhouser, “Multi-scale Context
Aggregation by Dilated Convolutions”. arXiv:1511.07122v3,
Apr. 2016.
A. L. Maas, A. Y. Hannun, and A. Y. Ng, “Rectifier nonlinearities improve neural network acoustic models”. Proceedings of
the 30th International Conference on Machine LearningAtlanta,
Georgia, USA, 2013.
V. Nair and G. E. Hinton, “Rectified Linear Units Improve
Restricted Boltzmann Machines”. Proceedings of the 27th
International Conference on Machine Learning (ICML10)J. Fürnkranz and T. Joachims, Eds.:807–814, 2010.
X. Glorot and Y. Bengio, “Understanding the difficulty of
training deep feedforward neural networks”. proceedings of the
13th International Conference on Artificial Intelligence and
Statisticsser. Proceedings of Machine Learning ResearchY. W.
Teh and M. Titterington, Eds. 9:249–256, Chia Laguna Resort,
Sardinia, Italy13–15 May 2010.
L. Bottou, Stochastic Gradient Descent Tricks. Berlin,
Heidelberg: Springer, 2012., pp. 421–436.
D. P. Kingma and J. L. Ba, “Adam: A Method for Stochastic
Optimization”. , Jan. 2017.
T. Dozat, , “Incorporating Nesterov Momentum into Adam”. ,
Standford University, Tech. Rep. 54, , May 2015.
H.-H. Chang, C.-L. Lin, and J.-K. Lee, “Load Identification
in Nonintrusive Load Monitoring using Steady-state and Turnon Transient Energy Algorithms”. In proceedings of the 14th
International Conference on Computer Supported Cooperative
Work in Design:27–32, Apr. 2010.
,
“Belkin Energy
Disaggregation
Competition”.
,
https://www.kaggle.com/c/belkin-energy-disaggregation-competition,
accessed: 2017.06.16.
S. Makonin, F. Popowich, L. Bartram, B. Gill, and I. V.
Bajic, “AMPds: A Public Dataset for Load Disaggregation
and Eco-Feedback Research”. Electrical Power and Energy
Conference (EPEC), 2013 IEEE:1–6, Aug. 2013.
B. Christian, W. Kleiminger, R. Cicchetti, T. Staake, and S. Santini, “The ECO Data Set and the Performance of Non-Intrusive
Load Monitoring Algorithms”. proceedings of the 1st ACM
International Conference on Embedded Systems for EnergyEfficient Buildings (BuildSys):80–89, Memphis, TN, USANov.
2014.
E. Holmegaard and M. B. Kjaergaard, “NILM in an Industrial
Setting: A Load Characterization and Algorithm Evaluation”.
2016 IEEE International Conference on Smart Computing
(SMARTCOMP):1–8, May 2016.
D. M. W. Powers, “Evaluation: From precision, recall and Fmeasure to ROC, informedness, markedness, and correlation”.
Journal of Machine Learning Technologies 2 :37–632011
W. J. Youden, “Index for Rating Diagnostic Tests”. Cancer
Vol. 3 (1):32–351950
C. E. Metz, “Basic principles of ROC analysis”. Seminars in
Nuclear Medicine 8 (4):283–2981978
G. Schröder, M. Thiele, and W. Lehner, “Setting Goals and
Choosing Metrics for Recommender System Evaluations”.
UCERSTI2 Workshop at the 5th ACM Conference on Recommender Systems 23Chicago, USA, 2011.
| 1cs.CV
|
THE HILBERT SERIES OF SL2 -INVARIANTS
arXiv:1710.02606v2 [math.RA] 16 Jan 2018
PEDRO DE CARVALHO CAYRES PINTO, HANS-CHRISTIAN HERBIG, DANIEL HERDEN, AND CHRISTOPHER SEATON
Abstract. Let V be a finite dimensional representations of the group SL2 of 2×2 matrices with complex coefficients
and determinant one. Let R = C[V ]SL2 be the algebra of SL2 -invariant polynomials on V . We present a calculation
P
of the Hilbert series HilbR (t) = n≥0 dim(Rn ) tn as well as formulas for the first two coefficients of the Laurent
expansion of HilbR (t) at t = 1.
Contents
1. Introduction
Acknowledgements
2. Background and Definitions
3. Computation of the Hilbert Series
3.1. The Multivariate Hilbert Series
3.2. Analytic Continuation and the Univariate Hilbert Series
4. The Coefficients of the Laurent Expansion
4.1. Discussion of Cases
4.2. The Exceptions
4.3. The First Coefficient
4.4. The Second Coefficient
5. The Coefficients of the Laurent Expansion in Terms of Schur polynomials
6. An Algorithm to Compute the Hilbert Series
6.1. Partial Fraction Decomposition
6.2. Description of the Algorithm
References
1
3
3
4
4
5
6
6
7
8
9
10
14
15
16
18
1. Introduction
Let V be a finite
Lr dimensional representation of SL2 . It is well-known that V is isomorphic to a sum of irreducible
representations k=1 Vdk . Here, Vdk stands for the (dk + 1)-dimensional
irreducible representation of SL2 which is
Lr
given by binary forms of degree dk ∈ N. In the decomposition k=1 Vdk it is not assumed that the dk are pairwise
distinct. The algebra of polynomial SL2 -invariants R := C[V ]SL2 is a finitely generated C-algebra and carries a
natural N-grading R = ⊕n≥0 Rn . In fact it is generated by a complete system of homogeneous invariants which
obey some homogeneous relations. For a more detailed discussion, the reader may consult [12, 28].
In this paper we study the Hilbert series HilbR (t) of R, i.e. the generating function that counts the dimensions
of the homogeneous components Rn :
HilbR (t) =
∞
X
dim(Rn ) tn .
n=0
It is a classical result that HilbR (t) is rational. The degree of HilbR (t), i.e. the difference of the degree of the
numerator and the degree of the denominator, is referred to as the a-invariant a(R) of R, see [10, Definitions 3.6.13
and 4.4.4] and [15, Section 3]. The a-invariant of invariant rings has been studied for example in [11, 24, 25, 27];
note that some references use q to denote the negative a-invariant.
2010 Mathematics Subject Classification. Primary 13A50; Secondary 13H10, 05E05.
Key words and phrases. Hilbert series, special linear group, a-invariant, Schur polynomial.
C.S., D.H., and H.-C.H. were supported by a Collaborate@ICERM grant from the Institute for Computational and Experimental Research in Mathematics (ICERM). C.S. was supported by the E.C. Ellett Professorship in Mathematics and the Instituto de
Matemática Pura e Aplicada (IMPA); H.-C.H. was supported by CNPq through the Plataforma Integrada Carlos Chagas.
1
2
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
It is well-known that HilbR (t) has a pole at t = 1 of order equal the Krull dimension dim R. We use the notation
γm , m ≥ 0, to denote the coefficients in the Laurent expansion
(1.1)
HilbR (t) =
∞
X
γm
.
dim R−m
(1
−
t)
m=0
Some authors systematically use the notation γ or deg(R) for γ0 and τ or ψ(R) for γ1 . The coefficients γ0 and γ1
have clear interpretations in the case of invariants of a finite group, see [32, Lemma 2.4.4] as well as [8, Section
3.13] or [9], and their meaning in more general contexts has been investigated, e.g. in [27, Chapter 3] and [1]. For
the case of invariants of SL2 , David Hilbert [19] published in 1893 a formula for γ0 in the case that V = Vd is
irreducible for d ≥ 5:
d−3
⌊d/2⌋
X d
d
−1
−n
.
(1.2)
γ0 =
(3 − (−1)d ) d! n=0 n
2
Since then, computations of HilbR (t) and γ0 have been taken up by numerous authors, e.g. [2, 3, 4, 5, 6, 7, 21, 22,
26, 30].
The main result of this paper is Theorem 1.1, presenting formulas for γ0 and γ1 that generalize Hilbert’s formula
in Equation (1.2) to encompass all finite-dimensional representation of SL2 up to a few low dimensional exceptions
that can easily be computed directly. As a corollary, we reproduce the computation of the a-invariant given
by F. Knop and P. Littelmann [25]. On the way, we produce formulas in Proposition 3.1 and Theorem 3.2 for
the multivariate and univariate Hilbert series of C[V ]SL2 , respectively, that in particular indicate an algorithm for
computing the Hilbert series. This algorithm is described in Section 6. It has been implemented using Mathematica
[33] and is available from the authors by request.
This paper is the third in a series that uses the methods described in [12, Section 4.6.1 and 4.6.4] to systematically
compute Hilbert series of rings of invariants. The techniques were first laid out in [18], where the Hilbert series
of algebras of regular functions on linear symplectic circle quotients were investigated. As explained in that
reference, that computation is equivalent to the computation corresponding to the invariant ring of a cotangent-lifted
representation of the circle, and the extension of these techniques to arbitrary circle representations was recently
presented in [11]. The key observation is related to weights of the Cartan torus that occur with multiplicity in
the representation, which in the case considered here occurs whenever V contains two irreducible representations
Vdk whose dimensions have the same parity (and hence must occur whenever r > 2). Though these degeneracies
impose difficulties in the computation of the Hilbert series, they can be circumvented by taking advantage of
certain analytic continuations, viewing some instances of the integer weights as real parameters and perturbing
them to avoid degeneracies. This in particular can be used to show that the bare expressions for the γm in
terms of the weights have removable singularities along the diagonals. After removing these singularities, the
resulting expressions can be expressed in terms of Schur polynomials, yielding succinct expressions for the first two
γm , see Equation (1.3). In principle, this method can be used to deduce similar (but clumsier) formulas for the
higher coefficients such as γ2 and γ3 ; see [11, Theorems 6.4 and 6.5]. The potential usefulness of our technique to
representations of reductive Lie groups of higher rank is currently being explored, in particular for the case of a
torus of dimension ℓ > 1.
We briefly describe the notation required toL
state Theorem 1.1, which is adopted and explained in more detail
r
throughout the rest of the paper. For V =
k=1 Vdk , let D = dim V . Let Λ = {(k, i) ∈ Z × Z : 1 ≤ k ≤
r, ⌊dk /2⌋ + 1 ≤ i ≤ dk } with C := |Λ|, and for each (k, i) ∈ Λ, let ak,i := 2i − dk . Let a := a(k,i) : (k, i) ∈ Λ ,
and let σV = 2 if each dk is even and 1 otherwise. Finally, for an integer partition ρ = (ρ1 , . . . , ρn ) ∈ Zn
with ρ1 ≥ ρ2 ≥ · · · ρn ≥ 0, let sρ (a) denote the corresponding Schur polynomial in the variables ak,i ordered
lexicographically in (k, i) ∈ Λ (details can be found at the beginning of Section 5). We then have the following.
Lr
Theorem 1.1. Let V = k=1 Vdk be an SL2 -representation with V SL2 = {0}, and assume V is not isomorphic to
2V1 , nor Vd for d ≤ 4. The degree 3 − D coefficient γ0 of the Laurent series of HilbV (t) is given by
(1.3)
γ0 = σV
sC−3,C−3,C−3,C−4,...,1,0 (a)
.
sC−1,C−2,...,1,0 (a)
The degree 4 − D coefficient γ1 of the Laurent series is given by 3γ0 /2. Hence the a-invariant −(D − 3) − 2γ1 /γ0
of C[V ]SL2 is equal to −D.
After briefly discussing the relevant background in Section 2, we turn to the computation of the Hilbert series in
Section 3. We first compute an expression for the multivariate Hilbert series, which in this case has no degeneracies,
in Section 3.1, and then demonstrate in Section 3.2 the analytic continuation used to state the univariate Hilbert
THE HILBERT SERIES OF SL2 -INVARIANTS
3
series. We then turn to the computation of the Laurent coefficients γ0 and γ1 . The naive formulas for these, which
only apply in the cases without degeneracies, are computed in Section 4; the removal of the singularities using
Schur polynomials is explained in Section 5.
The proof of Theorem 1.1 is given in Section 5 as Theorems 5.4 and 5.5. The computation of the a-invariant
is an immediate corollary. Specifically, note that C[V ]SL2 is Gorenstein by [20, Corollary 1.9], implying that
−2γ1 /γ0 = a(C[V ]SL2 ) + dim(C[V ]SL2 ), see [28, Equation (3.32)]. All representations satisfying the hypotheses of
Theorem 1.1 are 1-large by [17, Theorem 3.4] and hence dim(C[V ]SL2 ) = D − 3, see [29, Remark 9.2(3)]. This
computation of the a-invariant agrees with that given in [25, Satz 1].
Acknowledgements
We would like to thank Gerald Schwarz for bringing to our attention the work of Friedrich Knop on the ainvariant of invariant rings. Furthermore, we would like to thank Leonid Bedratyuk for pointing out references
related to this project. Herbig, Herden, and Seaton express appreciation to the Institute for Computational and
Experimental Research in Mathematics (ICERM), Herbig and Seaton express appreciation to Baylor University,
and Herden and Seaton express appreciation to the Instituto de Matemática Pura e Aplicada (IMPA) for hospitality
during the work contained in this manuscript.
2. Background and Definitions
Let Vd denote the irreducible representation of SL2 of dimension d + 1 on binary forms of degree d. Let V be an
arbitrary SL2 -representation such that V SL2 = {0}, and then V is of the form
V =
r
M
Vdk
k=1
where each dk ≥ 1. Note that the dk need not be distinct, but throughout this paper, we will assume for convenience
that they are ordered non-decreasingly, i.e. dk ≤ dk+1 . Let D denote the dimension of V , which is given by
D := dim V = r +
r
X
dk .
k=1
Let C[V ]SL2 denote the algebra of SL2 -invariant polynomial functions on V with its usual N-grading by degree,
and let HilbV (t) = Hilb(d1 ,...,dr ) (t) denote the univariate Hilbert series of C[V ]SL2 . In Section 3.1, we will also
Lr
consider C[V ]SL2 with the Nr -grading inherited from the decomposition V =
k=1 Vdk . That is, a monomial of degree (p1 , . . . , pr ) is the product of monomials on Vdk , each of degree pk . We use HilbrV (t1 , . . . , tr ) =
Hilbr(d1 ,...,dr ) (t1 , . . . , tr ) to denote the corresponding r-variate Hilbert series.
Using the Molien-Weyl formula [12, Section 4.6.1] and Weyl’s Integration formula [14, Equation (26.19)], the
Hilbert series of C[V ]SL2 can be expressed as an integral over the Cartan torus of SL2 . It will be helpful to define
the constants ak,i := 2i − dk for k = 1, . . . , r and 0 ≤ i ≤ dk , and then the Hilbert series is given by the integral
Z
Z
(1 − z 2 ) dz
(1 − z 2 ) dz
1
1
√
√
(2.1)
HilbV (t) =
=
.
dk
dk
r Q
r Q
Q
Q
2π −1
2π −1
−a
d
−2i
k,i
k
1
1
(1 − tz
)
(1 − tz
)
S z
S z
k=1 i=0
k=1 i=0
We will often use ck to denote a real parameter that is near dk , in the sense that we will consider the limit as the
ck → dk . Similarly, we will use bk,i to denote a real parameter near ak,i .
For the case of the multivariate Hilbert series HilbrV (t1 , . . . , tr ), a simple modification to the proof of the MolienWeyl formula yields
Z
(1 − z 2 ) dz
1
.
(2.2)
HilbrV (t1 , . . . , tr ) = √
dk
r
Q Q
2π −1
−a
k,i
(1 − tk z
)
S1 z
k=1 i=0
See [31, Equation (13)], where this extension is given for the case of finite groups, as well as [13, Section IV], where
it is described for the special bigraded case of a real representation, where the bigrading considers the holomorphic
and anti-holomorphic parts separately.
It will be convenient for us to use a few different methods to index the factors in the denominator of the integral
in Equation (2.1). First, let us define
Θ := {(k, i) ∈ Z × Z : 1 ≤ k ≤ r, 0 ≤ i ≤ dk },
4
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
and then the integral in Equation (2.1) can be expressed as
Z
1
(1 − z 2 ) dz
Q
√
(2.3)
.
(1 − tz −ak,i )
z
2π −1
S1
(k,i)∈Θ
Note that Θ has D elements.
We will sometimes wish to take advantage of the grouping of the nonzero ak,i into positive and negative pairs.
Hence, define Λ to be the subset of Θ consisting of pairs (k, i) such that ak,i > 0, i.e.
Λ = {(k, i) ∈ Z × Z : 1 ≤ k ≤ r, ⌊dk /2⌋ + 1 ≤ i ≤ dk }.
Let C denote the cardinality of Λ,
C :=
r
X
i=1
⌈dk /2⌉,
and let e denote the number of ak,i = 0, corresponding to the number of k such that dk is even. Then we can
express Equation (2.3) as
Z
1
(1 − z 2 ) dz
Q
√
(2.4)
.
(1 − tz −ak,i )(1 − tz ak,i )
z(1 − t)e
2π −1
S1
(k,i)∈Λ
3. Computation of the Hilbert Series
3.1. The Multivariate Hilbert Series. We first consider the computation of the multivariate Hilbert series and
prove the following; compare [6].
L
Proposition 3.1. Let V = rk=1 Vdk be an SL2 -representation with V SL2 = {0}. The Nr -graded Hilbert series
HilbrV (t1 , . . . , tr ) is given by
2/aK,I
X
(3.1)
X
(K,I)∈Λ ζ aK,I =1
aK,I (1 − t2K )
Q
dk ∈2Z
(1 − tk )
Q
1 − ζ 2 tK
(k,i)∈Λ
(k,i)6=(K,I)
−ak,i /aK,I
(1 − ζ −ak,i tk tK
a
)(1 − ζ ak,i tk tKk,i
/aK,I
)
.
Proof. Using Equation (2.2), with the factors in the denominator indexed as in Equation (2.4), we have
Z
1
(1 − z 2 ) dz
Q
Q
√
Hilbr(d1 ,...,dr ) (t1 , . . . , tr ) =
(1 − tk )
(1 − tk z −ak,i )(1 − tk z ak,i )
z
2π −1
S1
=
1
√
2π −1
Z
dk ∈2Z
(k,i)∈Λ
P
−1+ (k,i)∈Λ ak,i
2
Q
(1 − z )z
dz
Q
.
a
k,i
(1 − tk )
(z
− tk )(1 − tk z ak,i )
S1 dk ∈2Z
(k,i)∈Λ
Assume that each |tk | < 1 for each k. Then the poles in z inside the unit disk occur at points such that z ak,i = tk ,
1/a
1/a
i.e. points of the form z = ζtk k,i where ζ is an ak,i th root of unity and the tk k,i are defined using a suitably
chosen, fixed value of the logarithm. We assume that these poles are distinct, which is true for a generic choice of
the tk . Fix a (K, I) ∈ Λ and an aK,I th root of unity ζ0 , and then we express
z
Q
dk ∈2Z
=
=
(1 − tk )
(z aK,I
Q
(1 − z 2 )
(1 − tk z −ak,i )(1 − tk z ak,i )
(k,i)∈Λ
− tK )(1 − tK
z aK,I )
z aK,I −1 (1 − z 2 )
Q
Q
(1 − tk )
dk ∈2Z
1/aK,I
(1 − tK z aK,I )(z − ζ0 tK
)
Q
(k,i)∈Λ
(k,i)6=(K,I)
(1 − tk z −ak,i )(1 − tk z ak,i )
z aK,I −1 (1 − z 2 )
Q
1/a
(1 − tk )
(z − ζtK K,I )
ζ aK,I =1
ζ6=ζ0
dk ∈2Z
Q
(k,i)∈Λ
(k,i)6=(K,I)
(1 − tk z −ak,i )(1 − tk z ak,i )
.
THE HILBERT SERIES OF SL2 -INVARIANTS
1/aK,I
Hence, we have a simple pole at z = τ := ζ0 tK
(1 − tK τ aK,I )
Q
ζ aK,I =1
ζ6=ζ0
=
(1 − t2K )τ aK,I −1
, and the residue at z = τ is given by
τ aK,I −1 (1 − τ 2 )
Q
Q
1/a
(1 − tk )
(τ − ζtK K,I )
dk ∈2Z
Q
(k,i)∈Λ
(k,i)6=(K,I)
τ aK,I −1 (1 − τ 2 )
Q
Q
(1 − tk )
(1 − ζ)
ζ aK,I =1
ζ6=1
5
dk ∈2Z
(k,i)∈Λ
(k,i)6=(K,I)
(1 − tk τ −ak,i )(1 − tk τ ak,i )
(1 − tk τ −ak,i )(1 − tk τ ak,i )
2/aK,I
=
aK,I (1 − t2K )
Q
(1 − tk )
dk ∈2Z
Q
1 − ζ02 tK
−ak,i
(k,i)∈Λ
(k,i)6=(K,I)
(1 − ζ0
−ak,i /aK,I
tk tK
a
a
)(1 − ζ0 k,i tk tKk,i
/aK,I
)
.
Summing over each choice of (K, I) and ζ0 completes the proof.
3.2. Analytic Continuation and the Univariate Hilbert Series. By the definition of the multivariate Hilbert
series, it is clear that HilbrV (∆t ) = HilbV (t), where ∆t := (t, . . . , t) ∈ Cr . However, the expression for HilbrV (t1 , . . . , tr )
given by Proposition 3.1 is not defined after this substitution unless r = 1 or 2 and, when r = 2, one element of
{d1 , d2 } is even and the other is odd. One checks that in all other cases, factors of the form (1 − tk t−1
K ) appear in
the denominator, e.g. when dk and dK have the same parity so that for some choice of i and I, ak,i = aK,I and
ζ −ak,i = ζ −aK,I = 1. While it is again clear from the definitions that lim(t1 ,...,tr )→∆t HilbrV (t1 , . . . , tr ) = HilbV (t),
we demonstrate explicitly in this section that the corresponding singularities in Equation (3.1) are removable, yielding an expression for the univariate Hilbert series that is sufficiently explicit to compute the Laurent coefficients.
Note that Equation (3.1) has singularities at tk = tK = t in the open unit disk only where ζ −ak,i t(aK,I −ak,i )/aK,I = 1,
which only occur in factors where aK,I = ak,i . The argument in this section is similar to that of [18, Section 3.3]
and [11, Theorem 3.3], the point here being that the same techniques extend to the case of G = SL2 with very
little modification.
To simplify the argument, we re-index as follows. Let a be a positive value of ai,j that occurs with multiplicity,
set
Λa := {(k, i) ∈ Λ : ak,i 6= a},
and let N be the cardinality of Λa . We consider the integral
Z
1
(1 − z 2 ) dz
√
,
N
Q
Q
2π −1
−a
a
−a
a
e
k,i
k,i
1
(1 − xj z )(1 − xj z )
(1 − xk,i z
)(1 − xk,i z )
S z(1 − t)
j=1
(k,i)∈Λa
where the xj and xk,i are assumed distinct of modulus less than 1 and contained in a fixed branch of the logarithm.
Because the integrand is defined and continuous and hence bounded for the xj and xk,i sufficiently close to t, an
application of the Dominated Convergence theorem demonstrates that the limit of this integral as the xj → t and
xk,i → t, provided it exists, is equal to HilbV (t).
1/a
At a pole of the form z = ζxJ where ζ is an ath root of unity, a computation identical to that in Proposition 3.1
yields that the residue is given by
2/a
1 − ζ 2 xJ
.
N
Q
Q
−ak,i /a
ak,i /a
−1
2
−a
a
e
k,i
k,i
(1 − ζ
xk,i xJ
)(1 − ζ xk,i xJ
)
(1 − xj xJ )(1 − xj xJ )
a(1 − t) (1 − xJ )
j=1
j6=J
(k,i)∈Λa
Rewrite
2/a
−1
xN
(1 − ζ 2 xJ )
J
,
N
Q
Q
−ak,i /a
ak,i /a
2
−a
a
e
k,i
k,i
(x
−
x
)(1
−
x
x
)
(1
−
ζ
x
x
)(1
−
ζ
x
x
)
a(1 − t) (1 − xJ )
J
j
j J
k,i J
k,i J
j=1
j6=J
(k,i)∈Λa
6
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
1/a
where J ranges from 1 to N and ζ remains fixed, i.e.
and consider the sum of the residues at z = ζxJ
N
X
2/a
−1
xN
(1 − ζ 2 xJ )
J
N
Q
Q
−a /a
a /a
J=1 a(1 − t)e (1 − x2 )
(xJ − xj )(1 − xj xJ )
(1 − ζ −ak,i xk,i xJ k,i )(1 − ζ ak,i xk,i xJk,i )
J
j=1
j6=J
N
P
(k,i)∈Λa
2/a
−1
(−1)J−1 xN
(1 − ζ 2 xJ )
J
J=1
=
a(1 −
t)e
N
Q
p=1
(1 −
x2p )
Q
1≤p<q≤N
N
Q
p=1
p6=J
(1 − x2p )
(xp − xq )(1 − xp xq )
Q
(k,i)∈Λa
Q
1≤p<q≤N
p,q6=J
(1 −
(xp − xq )(1 − xp xq )
−a /a
ζ −ak,i xk,i xJ k,i )(1
.
−
a /a
ζ ak,i xk,i xJk,i )
The numerator
is easily seen to be alternating in the xJ , implying that it is divisible by the Vandermonde deterQ
minant 1≤p<q≤N (xp − xq ) in the denominator. That is, for each fixed ζ, the sum of residues at poles of the form
1/a
z = ζxJ has removable singularities in the xJ at points where each xJ = t. Applying this argument to each value
of ak,i that occurs in Λ with multiplicity, it follows that the limit of Equation (3.1) as each tj → t exists. Finally,
by a series of simple substitutions identical to those used in [18, pages 52–53] and [11, proof of Theorem 3.3], we
have the following.
Lr
Theorem 3.2. Let V = k=1 Vdk be an SL2 -representation with V SL2 = {0}. Let a := ak,i : (k, i) ∈ Λ , and let
b := bk,i : (k, i) ∈ Λ where the bk,i are real parameters. The N-graded Hilbert series HilbV (t) is given by
(3.2)
lim
b→a
X
X
(K,I)∈Λ ζ aK,I =1
bK,I (1 − t2 )(1 − t)e
Q
(k,i)∈Λ
(k,i)6=(K,I)
1 − ζ 2 t2/bK,I
.
(1 − ζ −ak,i t(bK,I −bk,i )/bK,I )(1 − ζ ak,i t(bK,I +bk,i )/bK,I )
4. The Coefficients of the Laurent Expansion
In this section, we compute the first two coefficients of the Laurent expansion at t = 1 of the expression given
in Equation (3.2). We will see in Section 5 that, after taking the limit b → a, these coefficients correspond to the
first and second Laurent coefficients γ0 and γ1 of HilbV (t), see Equation (1.1). Throughout this section, it will be
convenient to index the factors of the integral as in Equation (2.3). Hence, for (K, I) ∈ Λ and ζ an aK,I th root of
unity, we define
(4.1)
HV,K,I,ζ (b, t) :=
1 − ζ 2 t2/bK,I
(1 − ζ −ak,i t(bK,I −bk,i )/bK,I )
Q
bK,I
(k,i)∈Θ
(k,i)6=(K,I)
so that by Theorem 3.2,
(4.2)
HilbV (t) = lim
b→a
X
X
HV,K,I,ζ (b, t).
(K,I)∈Λ ζ aK,I =1
Our method will be to consider the Laurent expansions of each of the terms HV,K,I,ζ (b, t) separately.
4.1. Discussion of Cases. Assume V is 1-large, which is true unless V is isomorphic to V1 , 2V1 , or V2 by [17,
Theorem 3.4]; we refer the reader to this reference for the definition of 1-large. Then C[V ]SL2 has Krull dimension
D − 3, see [29, Remark 9.2(3)]. Hence, the first nontrivial Laurent coefficient occurs in degree 3 − D.
Now, any term of the form HV,K,I,1 (b, t) for (K, I) ∈ Λ is of the form
(4.3)
HV,K,I,1 (b, t) =
bK,I
Q
1 − t2/bK,I
;
(1 − t(bK,I −bk,i )/bK,I )
(k,i)∈Θ
(k,i)6=(K,I)
Each such term has a pole of order |Θ| − 2 = D − 2 at t = 1. If each dk is even, then each ak,i is even, and then
for each (K, I) ∈ Λ, we have HV,K,I,1 (b, t) = HV,K,I,−1 (b, t). To simplify notation, we define
(
1, if any dk is odd,
(4.4)
σV :=
2, if all dk are even.
THE HILBERT SERIES OF SL2 -INVARIANTS
7
If d1 = 1 and all other dk are even, then for K > 1 and any I,
(4.5)
HV,K,I,−1 (b, t) =
bK,I (1 + t(bK,I −b1,0 )/bK,I )(1 +
1 − t2/bK,I
t(bK,I −b1,1 )/bK,I )
Q
(k,i)∈Θ
(k,i)6=(K,I)
k6=1
(1 − t(bK,I −bk,i )/bK,I )
,
which has a pole of order D − 4. In particular, in Equation (4.2) any term of the form HV,K,I,−1 would have a pole
of order D − 4.
In any other case, i.e. if two or more dk are odd or one dk > 1 is odd, then for all K the term HV,K,I,−1 (b, t) has
a pole of order at most D − 5, and any HV,K,I,−1 appearing in Equation (4.2) has a pole of order at most D − 6.
Now, if ζ 6= ±1 is an aK,I = (2I − dK )th root of unity, then it must be that |aK,I | ≥ 3, and hence dK ≥ 3. The
numerator 1 − ζ 2 t2/bK,I of HV,K,I,ζ (b, t) no longer has a zero at t = 1. For each odd dk , we have ak,(dk ±1)/2 = ±1,
while for each even dk , we have ak,(dk /2)±1 = ±2, and ζ ±1 , ζ ±2 6= 1. Hence, if r ≥ 2, then in Equation (4.2) each
such term has a pole of order at most D − 5. If r = 1 and d1 = 3 or 4, i.e. V ∼
= V4 , then some terms
= V3 or V ∼
HV,K,I,ζ (b, t) in Equation (4.2) corresponding to such a ζ may have pole order D − 3. If r = 1 and d1 ≥ 5, it is
easy to see that at least four of the ak,i satisfy ζ ak,i 6= 1 so that the pole order of such a term is at most D − 5.
Hence, in the computation of γ0 , the coefficient of the Laurent series at t = 1 of degree − dim C[V ]SL2 , and γ1 ,
the coefficient of degree 1 − dim C[V ]SL2 , we must consider V1 , 2V1 , V2 , V3 , and V4 separately. In all other cases,
only terms corresponding to ζ = ±1 contribute to γ0 ; terms with ζ = −1 contribute to γ0 only if each dk is even,
in which case they are identical to the corresponding terms with ζ = 1 and hence contribute to γ0 and γ1 in the
same way. If there are at least two odd dk or one odd dk > 1, then only terms corresponding to ζ = 1 contribute
to γ1 . If d1 = 1 and all other dk are even, then terms with ζ = −1 contribute to γ1 separately.
It is of interest to note that the Laurent series of HV,K,I,1 (b, t) has a term of degree 2 − D with coefficient
D−3
2bK,I
Q
(k,i)∈Θ
(k,i)6=(K,I)
(bK,I − bk,i )
.
Hence, by the above observations, we have the following.
Lr
Corollary 4.1. Let V = k=1 Vdk be an SL2 -representation with V SL2 = {0}, and assume that V is not isomorphic
to V1 , 2V1 , nor V2 . Then
(4.6)
lim
b→a
X
(K,I)∈Λ
D−3
2bK,I
Q
(k,i)∈Θ
(k,i)6=(K,I)
(bK,I − bk,i )
= 0.
Note that the quantity on the left side of Equation (4.6) is equal to 1 when V ∼
= V1 , equal to −1/4 when V ∼
= V2 ,
2V
.
and equal to −1 when V ∼
= 1
4.2. The Exceptions. For the sake of completeness, we give the Hilbert series and Laurent coefficients corresponding to the representations V1 , 2V1 , V2 , V3 , and V4 , to which the results in this section do not apply. The
Hilbert series for these cases are known and can easily be computed directly using the above methods; similarly,
generating invariants are classical and can easily be computed, e.g. using the algorithms described in [5] and [12,
Sections 4.1–2].
V1 : For the representation V1 , only the constants are invariant. Therefore,
HilbV1 (t) = 1,
γ0 = 1,
γ1 = 0,
D = 2,
and
a(C[V ]SL2 ) = 2 − D = 0.
V2 : For V2 , there is one quadratic invariant; in terms of the variables x0 , x1 , x2 given by coefficients of binary
forms of degree 2, the generating invariant is x21 − 4x0 x2 . Hence,
HilbV2 (t) =
1
,
1 − t2
γ0 =
1
,
2
γ1 =
1
,
4
D = 3,
and
a(C[V ]SL2 ) = 1 − D = −2.
2V1 : In the case of 2V1 , using coordinates (x0 , x1 ) for the first V1 and (x2 , x3 ) for the second, both with respect
to the canonical representation of SL2 , there is again a single quadratic invariant x1 x2 − x0 x3 . We again
have
1
1
1
Hilb2V1 (t) =
,
γ0 = ,
γ1 = ,
D = 4,
and
a(C[V ]SL2 ) = 2 − D = −2.
2
1−t
2
4
8
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
V3 : For the case of V3 , there is one quartic invariant; in terms of the variables x0 , x1 , x2 , x3 given by coefficients
of binary forms of degree 3, the generating invariant is x21 x22 − 4x0 x32 − 4x31 x3 + 18x0 x1 x2 x3 − 27x20 x23 .
Therefore,
1
3
1
,
γ0 = ,
γ1 = ,
D = 4,
and
a(C[V ]SL2 ) = −D = −4.
HilbV3 (t) =
1 − t4
4
8
V4 : Finally, for the case of V4 , there are two algebraically independent invariants. In the variables x0 , . . . , x4 as
above, the generating invariants are x22 − 3x1 x3 + 12x0 x4 and 2x32 − 9x1 x2 x3 + 27x0 x23 + 27x21 x4 − 72x0 x2 x4 .
Hence,
1
1
1
HilbV4 (t) =
,
γ0 = ,
γ1 = ,
D = 5,
and
a(C[V ]SL2 ) = −D = −5.
(1 − t2 )(1 − t3 )
6
4
4.3.
First Coefficient. In this section, we compute the degree 3 − D coefficient of the Laurent series of
P The P
(K,I)∈Λ
ζ aK,I =1 HV,K,I,ζ (b, t). In Section 5, we will use that γ0 is the limit of the expression computed here to
as b → a.
Lr
V is not isomorphic
Proposition 4.2. Let V = k=1 Vdk be an SL2 -representation
with V SL2 = {0}, and assume
to 2V1 , nor Vd for d ≤ 4. Let a := ak,i : (k, i) ∈ Λ , and letP
b := bk,iP: (k, i) ∈ Λ where the bk,i are real
parameters. The degree 3 − D coefficient of the Laurent series of (K,I)∈Λ ζ aK,I =1 HV,K,I,ζ (b, t) is given by
D−3
P
D−4
D−4
bK,I − 2bK,I
−
bK,I
bκ,λ
P
bk,i
2bK,I − 2 −
(κ,λ)∈Θ
X
X
(k,i)∈Θ
(κ,λ)6=(K,I)
D−4
Q
Q
bK,I
.
(4.7)
σV
= σV
(bK,I − bk,i )
(bK,I − bk,i )
(K,I)∈Λ
(K,I)∈Λ
(k,i)∈Θ
(k,i)6=(K,I)
(k,i)∈Θ
(k,i)6=(K,I)
Proof. As explained in Section 4.1, the only terms that contribute to the degree 3 − D coefficient are of the form
given in Equation (4.3) corresponding to HV,K,I,1 (b, t) in each case and HV,K,I,−1 (b, t) only if each dk is even,
yielding the σV factor. The series expansion of the numerator 1 − t2/bK,I begins
(4.8)
1 − t2/bK,I =
2
bK,I
2(b2K,I − 3bK,I + 2)
bK,I − 2
2
(1
−
t)
+
(1 − t)3 + · · · ,
b2K,I
3b3K,I
(1 − t) +
and each factor of the denominator has a series expansion beginning
bk,i
(2bK,I − bk,i )bk,i
1
bK,I
(1 − t)−1 +
(1 − t) + · · · .
+
(4.9)
=
bK,I − bk,i
2(bk,i − bK,I ) 12(bk,i − bK,I )bK,I
1 − t(bK,I −bk,i )/bK,I
Hence, for fixed K and I, the term of degree 3 − D comes from the Cauchy product of these factors in two different
ways. First, from the first term of the first (holomorphic) series, the first term in each of the Laurent series except
for one value of (k, i) (say (κ, λ)), and the second term from the series corresponding to (k, i) = (κ, λ). Second,
from the second term of the first (holomorphic) series and the first term in each of the Laurent series.
The first of these combinations yields
D−4
Y
−bK,I
bκ,λ
1
bK,I
bκ,λ
2
Q
,
=
(bK,I − bk,i )
bK,I
bK,I
bK,I − bk,i
2(bκ,λ − bK,I )
(k,i)∈Θ
(k,i)6=(K,I),(κ,λ)
(k,i)∈Θ
(k,i)6=(K,I)
while the second yields
1
bK,I
bK,I − 2
b2K,I
!
completing the proof.
to
Y
(k,i)∈Θ
(k,i)6=(K,I)
bK,I
=
bK,I − bk,i
D−4
bD−3 − 2bK,I
QK,I
,
(bK,I − bk,i )
(k,i)∈Θ
(k,i)6=(K,I)
Note that if V = Vd is irreducible with d ≥ 5, setting each b1,i = 2i − d, the expression in Equation (4.7) reduces
2σV
d
X
(2I − d)d−3 (2I − d − 1)
.
d
Q
I=⌊d/2⌋+1
(2I − 2i)
i=0
i6=I
THE HILBERT SERIES OF SL2 -INVARIANTS
9
A simple computation demonstrates that this is equal to the expression in Equation (1.2) given by Hilbert in [19].
4.4. The Second Coefficient. We now turn to the computation of the second coefficient of the Laurent expansion
of HV,K,I,ζ (b, t).
Lr
V is not isomorphic
Proposition 4.3. Let V = k=1 Vdk be an SL2 -representation
with V SL2 = {0}, and assume
to 2V1 nor Vk for k ≤ 4. Let a := ak,i : (k, i) ∈ Λ , and let b := bk,i : (k, i) ∈ Λ where the bk,i are real
parameters.
If all dk are even,
P at least
Ptwo dk are odd, or at least one odd dk > 1, then the degree 4 − D coefficient of the
Laurent series of (K,I)∈Λ ζ aK,I =1 HV,K,I,ζ (b, t) is given by
(4.10)
σV
X
D−5
bK,I
(K,I)∈Λ
Q
(k,i)∈Θ
(k,i)6=(K,I)
2 2
(b
− 3bK,I + 2)
(bK,I − bk,i ) 3 K,I
bκ,λ
bκ,λ − 5bK,I + 6 + 3
6
X
+
(κ,λ)∈Θ
(κ,λ)6=(K,I)
X
(κ′ ,λ′ )∈Θr
{(K,I),(κ,λ)}
bκ′ ,λ′
.
If d1 = 1 and all other dk are even, then the degree 4 − D coefficient of the Laurent series is given by
X
(K,I)∈Λ
D−5
bK,I
Q
(k,i)∈Θ
(k,i)6=(K,I)
(4.11)
2 2
(b
− 3bK,I + 2)
(bK,I − bk,i ) 3 K,I
X
+
(κ,λ)∈Θ
(κ,λ)6=(K,I)
+
X
(K,I)∈Λ
K6=1
bκ,λ
bκ,λ − 5bK,I + 6 + 3
6
D−7
bK,I
Q
2
(k,i)∈Θ
(k,i)6=(K,I)
k6=1
bK,I − bk,i
X
(κ′ ,λ′ )∈Θr
{(K,I),(κ,λ)}
bκ′ ,λ′
.
Proof. We first consider the case where at least two dk are odd or at least one dk > 1 is odd. Once again, it was
explained in Section 4.1 that the only contributing terms are of the form given in Equation (4.3). To compute the
degree 4 − D coefficient, we consider the factors of the term in Equation (4.3) and use the expansions given by
Equations (4.8) and (4.9). A term of degree 4 − D can arise from the Cauchy product formula from these factors
in one of four ways:
(1) The degree 1 term from the numerator, the degree −1 term from each factor of the denominator except
two, say (κ, λ), (κ′ , λ′ ) 6= (K, I), and the degree 0 term from the factors of the denominator corresponding
to (κ, λ) and (κ′ , λ′ );
(2) The degree 1 term from the numerator, the degree −1 term from each factor of the denominator except
one, say (κ, λ) 6= (K, I), and the degree 1 term from the of the denominator corresponding to (κ, λ);
(3) The degree 2 term from the numerator, the degree −1 term from each factor of the denominator except
one, say (κ, λ) 6= (K, I), and the degree 0 term from the factor of the denominator corresponding to (κ, λ);
and
(4) The degree 3 term from the numerator and the degree −1 term from each factor of the denominator.
In the case of (1), we compute
Y
1
2
bK,I
bK,I
(k,i)∈Θr
{(K,I),(κ,λ),(κ′ ,λ′ )}
bK,I
bK,I − bk,i
bκ,λ
2(bκ,λ − bK,I )
bκ′ ,λ′
2(bκ′ ,λ′ − bK,I )
=
2
D−5
bK,I
bκ,λ bκ′ ,λ′
Q
.
(bK,I − bk,i )
(k,i)∈Θ
(k,i)6=(K,I)
10
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
In (2), we have
1
bK,I
2
bK,I
Y
(k,i)∈Θr
{(K,I),(κ,λ)}
bK,I
bK,I − bk,i
(2bK,I − bκ,λ )bκ,λ
12(bκ,λ − bK,I )bK,I
=
=
D−5
−bK,I
(2bK,I − bκ,λ )bκ,λ
Q
.
6
(bK,I − bk,i )
(k,i)∈Θ
(k,i)6=(K,I)
For (3),
1
bK,I
bK,I − 2
b2K,I
!
Y
(k,i)∈Θr
{(K,I),(κ,λ)}
bK,I
bK,I − bk,i
bκ,λ
2(bκ,λ − bK,I )
D−5
−bK,I
(bK,I − 2)bκ,λ
Q
,
2
(bK,I − bk,i )
(k,i)∈Θ
(k,i)6=(K,I)
and (4) yields
1
bK,I
2(b2K,I − 3bK,I + 2)
3b3K,I
!
Y
(k,i)∈Θr{(K,I)}
2bD−5 (b2 − 3bK,I + 2)
bK,I
= K,I Q K,I
.
bK,I − bk,i
3
(bK,I − bk,i )
(k,i)∈Θ
(k,i)6=(K,I)
Combining these and summing over (K, I) ∈ Λ completes the proof of Equation (4.10).
We now assume d1 = 1 and each dk for k > 1 is even, and then σV = 1. The terms with ζ = 1 are identical to
the previous case, while terms with K > 1 and ζ = −1 are of the form giving in Equation (4.5). The degree 4 − D
coefficient of the Laurent series of such a latter term is given by
2
bK,I
1
4bK,I
Y
(k,i)∈Θ
(k,i)6=(K,I)
k6=1
bK,I
=
bK,I − bk,i 2
D−7
bK,I
Q
(k,i)∈Θ
(k,i)6=(K,I)
k6=1
bK,I − bk,i
.
Summing over all (K, I) with K > 1 yields Equation (4.11).
5. The Coefficients of the Laurent Expansion in Terms of Schur polynomials
In this section, we use Propositions 4.2 and 4.3 to give explicit formulas for γ0 and γ1 in terms of Schur
polynomials in the variables ak,i . First, let us briefly recall the definition of Schur polynomials for the convenience
of the reader.
Recall that if ρ = (ρ1 , . . . , ρn ) ∈ Zn is an integer partition, i.e. ρ1 ≥ ρ2 ≥ · · · ρn ≥ 0, then the alternant
associated to ρ in the variables x = (x1 , . . . , xn ) is defined by
ρ
Aρ (x) = det xi j .
The alternant is an alternating polynomial in the xi and hence divisible by the Vandermonde determinant
Y
Aδ (x) =
(xi − xj ),
1≤i<j≤n
where δ = (n − 1, n − 2, . . . , 1, 0). The Schur polynomial associated to ρ in the variables xi is defined to be
(5.1)
sρ (x) :=
Aδ+ρ (x)
.
Aδ (x)
Remark 5.1. For simplicity, we will sometimes refer to the Schur polynomial associated to ρ ∈ Zn that fail to be
partitions in the sense that ρ1 < ρ2 . In these cases, we mean the polynomial (or Laurent polynomial if ρ1 < 0)
defined in the same way by Equation (5.1). Note that the alternant Aδ+ρ (x) is still alternating so that sρ (x) is a
symmetric (Laurent) polynomial; however, such a polynomial may be zero for nontrivial ρ.
THE HILBERT SERIES OF SL2 -INVARIANTS
11
It is easy to see that the formulas for γ0 and γ1 can be broken down into linear combinations of sums of the form
X
X
X
bR bS bT′ ′
QK,I κ,λ κ ,λ
ΦR,S,T (b) :=
(5.2)
,
(bK,I − bk,i )
′
′
(K,I)∈Λ (κ,λ)∈Θr (κ ,λ )∈Θr
{(K,I)} {(K,I),(κ,λ)}
X
ΦR,S (b) :=
(5.3)
X
(K,I)∈Λ (κ,λ)∈Θr
{(K,I)}
(5.4)
ΦR (b) :=
X
(K,I)∈Λ
Q
(k,i)∈Θ
(k,i)6=(K,I)
S
bR
K,I bκ,λ
,
(bK,I − bk,i )
(k,i)∈Θ
(k,i)6=(K,I)
bR
K,I
,
(bK,I − bk,i )
Q
(k,i)∈Θ
(k,i)6=(K,I)
for integers R, S, and T , which may be equal to 0. Hence, we will first indicate how such sums can be expressed
in terms of Schur polynomials.
We give Θ the lexicographic ordering so that (k, i) ≤ (k ′ , i′ ) if k < k ′ or k = k ′ and i ≤ i′ . This gives a total
ordering on Θ and hence Λ, and we use ι(k, i) to denote the position of (k, i) with respect to this ordering. That
is, ι(k, i) = 1 for the first element of Λ, ι(k, i) = 2 for the second, etc.
We define PS (bΘ ) to be the power sum of degree S in the variables bΘ = {bk,i : (k, i) ∈ Θ}, i.e.
X
bSk,i .
PS (bΘ ) :=
(k,i)∈Θ
Lr
Lemma 5.2. Let V = k=1 Vdk be an SL2 -representation with V SL2 = {0}. Choose bk,i = 0 for each (k, i) ∈ Θ
such that ak,i = 0, and assume bK,dK −I = −bK,I for each (K, I) ∈ Λ. For R, S, T ∈ Z, and ordering the variables
bk,i as described above, we have
1
ΦR,S,T (b) =
(5.5)
PS (bΘ )PT (bΘ )sR−e−C,C−2,C−3,...,1,0 (b) − PT (bΘ )sR+S−e−C,C−2,C−3,...,1,0 (b)
2sδ (b)
− PS (bΘ )sR+T −e−C,C−2,C−3,...,1,0 (b) + sR+S+T −e−C,C−2,C−3,...,1,0 (b) ,
(5.6)
(5.7)
PS (bΘ )sR−e−C,C−2,C−3,...,1,0 (b) − sR+S−e−C,C−2,C−3,...,1,0 (b)
,
2sδ (b)
sR−e−C,C−2,C−3,...,1,0 (b)
ΦR (b) =
.
2sδ (b)
ΦR,S (b) =
and
Note that aK,dK −I = −aK,I so that the required relation holds in the limit b → a. Depending on the values of
R, S, T , and e, it may be that the partitions appearing in Equations (5.5), (5.6), and (5.7) are not non-increasing
in the first two entries so that the Schur polynomials are non-standard in the sense described in Remark 5.1.
Proof. Throughout the proof, we let PS and PT denote PS (bΘ ) and PT (bΘ ), respectively. We express
P
bSκ,λ bTκ′ ,λ′
bR
K,I
ΦR,S,T (b) =
X
(K,I)∈Λ
=
X
(K,I)∈Λ
=
Q
(k,i)∈
Θr{(K,I)}
(bK,I − bk,i )
S
T
bR
K,I (PS − bK,I )(PT − bK,I )
Q
(bK,I − bk,i )
(k,i)∈
Θr{(K,I)}
S
T
bR
K,I (PS − bK,I )(PT − bK,I )
Q
(bK,I − bk,i )(bK,I + bk,i )
− 0)e
X
2bK,I (bK,I
X
bR−e−1
(PS − bSK,I )(PT − bTK,I )
K,I
Q
,
(b2K,I − b2k,i )
2
(K,I)∈Λ
=
(κ,λ),(κ′ ,λ′ )∈Θr{(K,I)}
(κ,λ)6=(κ′ ,λ′ )
(K,I)∈Λ
(k,i)∈
Λr{(K,I)}
(k,i)∈
Λr{(K,I)}
12
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
where the fact that bK,dK −I = −bK,I implies that each factor bK,I − bk,dK −I = 2bK,I . Using the ordering of Θ
described above, we can express
X
(K,I)∈Λ
=
bR−e−1
(PS − bSK,I )(PT − bTK,I )
K,I
Q
2
(b2K,I − b2k,i )
(k,i)∈
Λr{(K,I)}
X
(−1)ι(K,I)−1 bR−e−1
(PS − bSK,I )(PT − bTK,I )
K,I
Q
Q
2
(b2k,i − b2K,I )
(b2K,I − b2k,i )
X
(−1)ι(K,I)−1 bR−e−1
(PS − bSK,I )(PT − bTK,I )
K,I
Q
Q
2
(b2k,i − b2K,I )
(b2K,I − b2k,i )
(K,I)∈Λ
=
(k,i)∈Λ
(k,i)<(K,I)
(K,I)∈Λ
(k,i)∈Λ
(k,i)>(K,I)
(k,i)∈Λ
(k,i)<(K,I)
(k,i)∈Λ
(k,i)>(K,I)
(−1)ι(K,I)−1 bR−e−1
(PS − bSK,I )(PT − bTK,I )
K,I
=
X
P
(k,i),(κ,λ)∈
Λr{(K,I)}
(k,i)<(κ,λ)
(b2k,i − b2κ,λ )
(b2k,i − b2κ,λ )
Q
2
(K,I)∈Λ
Q
(k,i),(κ,λ)∈Λ
(k,i)<(κ,λ)
−e
−e
(−1)ι(K,I)−1 (bR−e−1
PS PT − bR+S−e
PT − bR+T
PS + bR+S+T
)
K,I
K,I
K,I
K,I
(K,I)∈Λ
(5.8)
=
2
(k,i),(κ,λ)∈
Λr{(K,I)}
(k,i)<(κ,λ)
(b2k,i − b2κ,λ )
Q
Q
(b2k,i − b2κ,λ )
.
(k,i),(κ,λ)∈Λ
(k,i)<(κ,λ)
Expanding the numerator of this expression and factoring out the power sums, it can be written as a sum of similar
expressions, which we consider simultaneously. Recall that C = |Λ|. For any integer M , we have
P
Q
(−1)ι(K,I)−1 bM
(b2k,i − b2κ,λ )
K,I
(K,I)∈Λ
(k,i),(κ,λ)∈Λr{(K,I)}
(k,i)<(κ,λ)
Q
2
(bk,i − b2κ,λ )
(k,i),(κ,λ)∈Λ
(k,i)<(κ,λ)
P
(−1)ι(K,I)−1 bM
K,I
(K,I)∈Λ
=
Q
(k,i),(κ,λ)∈Λ
(k,i)<(κ,λ)
=
Q
(k,i),(κ,λ)∈Λr{(K,I)}
(k,i)<(κ,λ)
(bk,i + bκ,λ )(bk,i − bκ,λ )
ρj +C−j
det bk,i
(k,i)∈Λ
1≤j≤C
Q
(k,i),(κ,λ)∈Λ
(k,i)<(κ,λ)
(b2k,i − b2κ,λ )
(bk,i + bκ,λ )(bk,i − bκ,λ )
,
ρj +n−j
where ρ = (M − C + 1, C − 2, C − 3, C − 4, . . . , 1, 0), and where the matrix bk,i
is interpreted as indexing
rows by (k, i) ∈ Λ (in terms of the order described above) and columns j = 1, . . . , C. To see this last step, notice
that the numerator of the previous equation can be seen as the cofactor expansion of the determinant along the
first row. Hence,
ρj +C−j
det bk,i
(k,i)∈Λ
s (b)
sρ (b)
1≤j≤C
Q
Q ρ
.
=
=
sδ (b)
(bk,i + bκ,λ )(bk,i − bκ,λ )
(bk,i + bκ,λ )
(k,i),(κ,λ)∈Λ
(k,i)<(κ,λ)
(k,i),(κ,λ)∈Λ
(k,i)<(κ,λ)
Applying this observation for several values of M to Equation (5.8), we express ΦR,S,T (b) in the form given in
Equation (5.5).
THE HILBERT SERIES OF SL2 -INVARIANTS
13
The proofs of the remaining equations are identical but with one or both of the factors (PS −bSK,I ) and (PT −bTK,I )
removed.
Remark 5.3. Note that one may also consider the expressions in Equations (5.2), (5.3), and (5.7) as polynomials
in the variables (bk,i : (k, i) ∈ Θ . The resulting expressions are partial Laurent-Schur polynomials as defined in
[11, Section 5], where the two sets of variables correspond to the b(k,i) with (k, i) ∈ Λ and (k, i) ∈
/ Λ.
With this, we can now complete the computations of γ0 and γ1 and hence the proof of Theorem 1.1. We claim
that as b → a, the expressions given by Propositions 4.2 and 4.3 tend to γ0 and γ1 , respectively. This can be
seen by noting that the integrands in the definitions of the Laurent coefficients are continuous on the circle and
hence can be bounded as b → a, so the Dominated Convergence Theorem allows one to exchange the limit with
the integral. See [18, end of Section 5-2], where this argument is given in detail in a very similar case.
We first consider γ0 .
Lr
Theorem 5.4. Let V = k=1 Vdk be an SL2 -representation
with V SL2 = {0}, and assume V is not isomorphic
to 2V1 , nor Vd for d ≤ 4. Let a := a(k,i) : (k, i) ∈ Λ . The degree 3 − D coefficient γ0 of the Laurent series of
HilbV (t) is given by
(5.9)
γ0 =
σV sC−3,C−3,C−3,C−4,...,1,0 (a)
.
sC−1,C−2,...,1,0 (a)
Note that if C = 2, e.g. if V = V1 ⊕V2 or V = 2V2 , then the partition (C−3, C−3, C−3, C−4, . . . , 1, 0) = (−1, −1)
and hence corresponds to a Laurent-Schur polynomial; see Remark 5.1. In all other cases under consideration, the
partition appearing in Equation (5.9) is a standard integer partition.
Proof. As in Theorem 3.2 and Section 4.3, we let b := b(k,i) : (k, i) ∈ Λ where the b(k,i) denote real parameters.
Using Equations (5.6) and (5.7), we rewrite Equation (4.7) as
σV
σV ΦD−3 (b) − 2ΦD−4 (b) − ΦD−4,1 (b) =
(5.10)
sD−e−C−3,C−2,C−3,...,1,0 (b)
2sδ (b)
− 2sD−e−C−4,C−2,C−3,...,1,0 (b) − P1 (bΘ )sD−e−C−4,C−2,C−3,...,1,0 (b)
+ sD−e−C−3,C−2,C−3,...,1,0 (b) .
Recall that
sδ (b) =
Y
bk,i + bκ,λ ,
(k,i),(κ,λ)∈Λ
(k,i)<(κ,λ)
where the order < is that described before Lemma 5.2. As ak,i > 0 for each (k, i) ∈ Λ, sδ (a) 6= 0. Hence, the limit
as b → a of the expression in Equation (5.10) is continuous at b = a. Moreover, as the elements of aΘ are either
zero or come in positive and negative pairs, P1 (aΘ ) = 0. Noting that D − e − C is the number of negative elements
of Θ, which is equal to C, the number of elements of Λ, yields
σV
γ0 =
2sC−3,C−2,C−3,...,1,0 (a) − 2sC−4,C−2,C−3,...,1,0 (a) .
2sδ (a)
However, note that sC−3,C−2,C−3,...,1,0 (a) is defined in Equation (5.1) in terms of the alternant associated to
δ + (C − 3, C − 2, C − 3, . . . , 1, 0) = (2C − 4, 2C − 4, 2C − 6, . . . , 2, 0). Provided C ≥ 2, which is true for all cases
under consideration, the repetition of 2C − 4 implies that sC−3,C−2,C−3,...,1,0 (a) = 0. Now sC−4,C−2,C−3,...,1,0 (a) is
defined in terms of the alternant associated to δ + (C − 4, C − 2, C − 3, . . . , 1, 0) = (2C − 5, 2C − 4, 2C − 6, . . . , 2, 0),
which is not in standard form. Switching the first two entries, we have (2C − 4, 2C − 5, 2C − 6, . . . , 2, 0) =
δ + (C − 3, C − 3, C − 3, C − 4, . . . , 1, 0). Hence,
sC−4,C−2,C−3,...,1,0 (a) = −sC−3,C−3,C−3,C−4,...,1,0 (a),
completing the proof.
We now turn to the computation of γ1 , which completes this section.
Lr
with V SL2 = {0}, and assume
V is not isomorphic to
Theorem 5.5. Let V = k=1 Vdk be an SL2 -representation
2V1 , nor Vd for d ≤ 4. Let a := a(k,i) : (k, i) ∈ Λ and a1 := a(k,i) : (k, i) ∈ Λ, k 6= 1 . Then
(5.11)
γ1 =
3γ0
3σV sC−3,C−3,C−3,C−4,...,1,0 (a)
=
.
2sC−1,C−2,...,1,0 (a)
2
14
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
Proof. We continue to let b := b(k,i) : (k, i) ∈ Λ where the b(k,i) denote real parameters. We first assume all dk
are even, at least two dk are odd, or at least one odd dk > 1. Using Equations (5.5), (5.6), and (5.7), and expressing
the Schur polynomial sM,C−2,C−3,...,1,0 (b) as sM,...,0 (b) for brevity, Equation (4.10) is equal to
h2
i
4
1
5
1
σV ΦD−3 (b) − 2ΦD−4 (b) + ΦD−5 (b) + ΦD−5,2 (b) − ΦD−4,1 (b) + ΦD−5,1 (b) + ΦD−5,1,1 (b)
3
3
6
6
2
σV h 2
4
=
sD−e−C−3,...,0 (b) − 2sD−e−C−4,...,0 (b) + sD−e−C−5,...,0 (b)
2sδ (b) 3
3
5
1
+ P2 (bΘ )sD−e−C−5,...,0 (b) − sD−e−C−3,...,0 (b) − P1 (bΘ )sD−e−C−4,...,0 (b) − sD−e−C−3,...,0 (b)
6
6
1
+ P1 (bΘ )sD−e−C−5,...,0 (b) − sD−e−C−4,...,0 (b) +
P1 (bΘ )2 sD−e−C−5,...,0 (b)
2
i
− 2P1 (bΘ )sD−e−C−4,...,0 (b) + sD−e−C−3,...,0 (b)
.
Then using the fact that D − e − C = C noted in the proof of Theorem 5.4, this is equal to
σV h 2
4
sC−3,...,0 (b) − 2sC−4,...,0 (b) + sC−5,...,0 (b)
2sδ (b) 3
3
5
1
+ P2 (bΘ )sC−5,...,0 (b) − sC−3,...,0 (b) − P1 (bΘ )sC−4,...,0 (b) − sC−3,...,0 (b)
6
6
1
+ P1 (bΘ )sC−5,...,0 (b) − sC−4,...,0 (b) +
P1 (bΘ )2 sC−5,...,0 (b) − 2P1 (bΘ )sC−4,...,0 (b)
2
i
+ sC−3,...,0 (b)
i
σV h
11sC−3,...,0 (b) − 11P1 (bΘ ) + 18 sC−4,...,0 (b) + P2 (bΘ ) + 3P1 (bΘ )2 + 6P1 (bΘ ) + 8 sC−5,...,0 (b) .
=
12sδ (b)
As in the proof of Theorem 5.4, sδ (a) 6= 0 so that this expression is continuous at b = a. Similarly, P1 (aΘ ) = 0,
the non-standard Schur polynomial sC−3,...,0 (a) = 0, and sC−4,...,0 (a) = −sC−3,C−3,C−3,C−4,...,1,0 (a). The nonstandard Schur polynomial sC−5,...,0 (a) is defined in terms of the alternant δ + (C − 5, C − 2, C − 3, . . . , 1, 0) =
(2C − 6, 2C − 4, 2C − 6, . . . , 2, 0) and hence vanishes. This completes the proof in this case.
Now assume d1 = 1 and all other dk are even. We need only deal with the additional sum in Equation (4.11),
X
(5.12)
D−7
bK,I
(K,I)∈Λ
K6=1
Q
2
(k,i)∈Θ
(k,i)6=(K,I)
k6=1
bK,I − bk,i
.
Define
Θ1 = {(k, i) ∈ Θ : k 6= 1}, Λ1 = {(k, i) ∈ Λ : k 6= 1}, and b1 := bk,i : (k, i) ∈ Λ1 ; then a1 := ak,i : (k, i) ∈
Lr
Λ1 . Note that we can treat Θ1 and Λ1 as associated to the representation k=2 Vdk , which has dimension D − 2;
the value of e is unchanged, and Λ1 has cardinality C − 1. Then we can rewrite Equation (5.12) as
1
2
X
(K,I)∈Λ1
D−7
bK,I
Q
(k,i)∈Θ1
(k,i)6=(K,I)
bK,I − bk,i
=
1
ΦD−7 (b1 ).
2
Applying Equation (5.7) and recalling that D − e − C = C, this is equal to
sC−6,C−3,C−4,...,1,0 (b1 )
.
4sδ (b1 )
We again note that sδ (a1 ) 6= 0 so that this function is continuous at b1 = a1 . The Schur polynomial associated to the
non-standard partition (C −6, C −3, C −4, . . . , 1, 0) is defined by the alternant δC−1 +(C −6, C −3, C −4, . . . , 1, 0) =
(2C − 8, 2C − 6, 2C − 8, . . . , 2, 0) and hence vanishes, completing the proof.
6. An Algorithm to Compute the Hilbert Series
L
In this section, we describe an algorithm to compute HilbV (t) for an arbitrary representation V = rk=1 Vdk of
SL2 . This algorithm is similar to those given in [18, Section 4] and [11, Section 4] for circle actions. Note, however,
that those algorithms consider only the generic cases with no degeneracies caused by repeated weights. In the case
of SL2 -invariants, this hypothesis is very restrictive; as was explained in the introduction, it implies that r ≤ 2
and, when r = 2, the degrees d1 and d2 have opposite parities. Hence, we begin by presenting a partial fraction
THE HILBERT SERIES OF SL2 -INVARIANTS
15
decomposition in Section 6.1 that allows us to extend to the general case. Note that this decomposition can be
used to extend the algorithms of [11, 16] to the degenerate cases as well.
6.1. Partial Fraction Decomposition. The main partial fraction decomposition we consider is the following.
Proposition 6.1. For t ∈ C, distinct values x1 , . . . , xn ∈ C, and positive integers m1 , . . . , mn , we have
n m
i −1
X
X
Gi,j (x1 , . . . , xn )
1
,
=
m
i
(1 − txi )
(1 − txi )mi −j
i=1
i=1 j=0
n
Y
(6.1)
where
j
Gi,j (x1 , . . . , xn ) =
1
d
j!(−xi )j dtj
n
Y
k=1
k6=i
1
(1 − txk )mk
.
t= x1
i
Proof. Consider
f (t) :=
n
Y
i=1
1
(1 − txi )mi
as a function of t. Clearly, a partial fraction decomposition
f (t) = C(x1 , . . . , xn ) +
n m
i −1
X
X
Gi,j (x1 , . . . , xn )
i=1 j=0
(1 − txi )mi −j
is possible. Observe that C(x1 , . . . , xn ) = limt→∞ f (t) = 0, so we need only evaluate the Gi,j (x1 , . . . , xn ). We have
n
Y
k=1
k6=i
1
= (1 − txi )mi f (t)
(1 − txk )mk
=
m
i −1
X
ℓ=0
(6.2)
=
j−1
X
ℓ=0
(6.3)
Gi,ℓ (x1 , . . . , xn )(1 − txi )ℓ +
n mX
k −1
X
Gk,ℓ (x1 , . . . , xn )(1 − txi )mi
k=1 ℓ=0
k6=i
(1 − txk )mk −ℓ
Gi,ℓ (x1 , . . . , xn )(1 − txi )ℓ + Gi,j (x1 , . . . , xn )(1 − txi )j
+
m
i −1
X
ℓ=j+1
Gi,ℓ (x1 , . . . , xn )(1 − txi )ℓ +
n mX
k −1
X
Gk,ℓ (x1 , . . . , xn )(1 − txi )mi
k=1 ℓ=0
k6=i
(1 − txk )mk −ℓ
.
The first sum in (6.2) clearly has degree at most j − 1 as a polynomial in t, while, noting that j + 1 ≤ mi , the
expression (6.3) evidently has a zero at t = 1/xi of multiplicity at least j + 1. Hence,
n
j
Y
d
1
j
= j!(−xi ) Gi,j (x1 , . . . , xn )
dtj
(1 − txk )mk
k=1
k6=i
j
+
d
dtj
m
i −1
X
ℓ=j+1
n mX
k −1
mi
X
Gk,ℓ (x1 , . . . , xn )(1 − txi )
Gi,ℓ (x1 , . . . , xn )(1 − txi )ℓ +
,
(1 − txk )mk −ℓ
k=1 ℓ=0
k6=i
where t = 1/xi is a zero of the expression on the second line so that
n
j
Y
d
1
= j!(−xi )j Gi,j (x1 , . . . , xn ).
dtj
(1 − txk )mk
k=1
k6=i
t= x1
i
We observe two interesting special cases, beginning with the case where each mi = 1.
16
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
Corollary 6.2. For t ∈ C and distinct x1 , . . . , xn ∈ C, we have
n
Y
n
X
1
=
1 − txi
i=1
i=1
1
.
n
Q
(1 − xk /xi )
(1 − txi )
k=1
k6=i
Restricting to the case n = 2 yields the following pleasing formula that we have come to refer to as the Yin-Yang
formula.
Corollary 6.3. For t ∈ C, x, y ∈ C distinct, and positive integers α and β, we have
j+α−1
β−1
α−1
i+β−1
X
X
1
(−y/x)i
(−x/y)j
1
j
i
=
+
.
(6.4)
(1 − tx)α (1 − ty)β
(1 − tx)α−i (1 − y/x)β+i j=0 (1 − ty)β−j (1 − x/y)α+j
i=0
6.2. Description of the Algorithm. Let V be a reducible representation of SL2 . For simplicity we will assume
that V has no trivial subrepresentations. The gist of the algorithm is formula (6.7) below.
To describe the algorithm, it will be convenient to introduce a new notation for the decomposition of V into
irreducible representations as follows. Decompose the representation into V = Veven ⊕ Vodd , where Veven consists
of those representations whose irreducible components have even degree and Vodd consists of those representations
whose irreducible components have odd degree. Let d1 > d2 > . . . > dr > 0 denote the (even) degrees of the
irreducible components of Veven and e1 > e2 > . . . > es > 0 the (odd) degrees of the irreducible components of
Vodd . Then we can write
r
M
i
Vdm
,
Veven =
i
i=1
where mi is the multiplicity of Vdi and
Vodd =
s
M
Venj j ,
j=1
where nj is the multiplicity of Vej .
We now determine the weights of the Cartan torus P
and their corresponding multiplicities. The even weights
d1 − 2i for i = 0, 1, . . . , d1 occur with multiplicity µi := k: |d1 −2i|≤dk mk . Similarly, the odd weights that occur in
P
V are e1 − 2j for j = 0, 1, . . . , e1 and occur with multiplicity νj := ℓ: |e1 −2j|≤eℓ nℓ .
With this notation we rewrite the Hilbert series in Equation (2.1) as follows:
Z
1
(1 − z 2 ) dz
√
HilbV (t) =
.
e1
d
Q
2π −1 S1 Q1
ν
e
−2j
d
−2i
µ
j
1
1
i
(1 − tz
)
(1 − tz
)
z
j=0
i=0
We introduce variables x = (x0 , x1 , . . . , xd1 ) and y = (y0 , y1 , . . . , ye1 ) corresponding to the even and odd weights
of the Cartan torus. Moreover, we introduce the function
Φ(t, x, y) :=
1
d1
Q
(1 − txi )µi
i=0
e1
Q
(1 − tyj )νj
j=0
and the exceptional set E of points z ∈ S1 such that z d1 , z d1 −2 , . . . , z −d1 , z e1 , z e1 −2 , . . . , z −e1 are not pairwise distinct. Noting that E ⊂ S1 is finite and defining g : C → Cd1 +e1 +2 , g(z) := (z d1 , z d1 −2 , . . . , z −d1 , z e1 , z e1 −2 , . . . , z −e1 ),
we can write
Z
1
dz
(6.5)
HilbV (t) = √
(1 − z 2 )Φ(t, g(z)) .
z
1
2π −1 S \E
At this point we make use of the results of Section 6.1. Namely, if x0 , x1 , . . . , xd1 , y0 , y1 , . . . , ye1 are pairwise
distinct we use the partial fraction decomposition
e1 νX
d1 µX
i −1
k −1
X
X
Gi,j (x, y)
Hk,ℓ (x, y)
+
,
Φ(t, x, y) =
µi −j
(1
−
tx
)
(1
− tyk )νk −ℓ
i
i=0 j=0
k=0 ℓ=0
THE HILBERT SERIES OF SL2 -INVARIANTS
17
where
j
Gi,j (x, y) =
Hk,ℓ (x, y) =
1
d
j
j!(−xi ) dtj
e1
d1
Y
Y
−µ
(1 − tym )−νk
(1 − txl ) l
,
m=0
l=0
l6=i
t=1/xi
d1
e1
Y
dℓ Y
1
−µl
(1 − tym )−νk
(1
−
tx
)
l
ℓ!(−yk )ℓ dtℓ
m=0
l=0
m6=k
.
t=1/yk
Substituting this into Equation (6.5) we find
Z
e1 νX
d1 µX
i −1
k −1
X
X
1
Gi,j (g(z))
Hk,ℓ (g(z))
dz .
√
HilbV (t) =
+
(1 − z 2 )
d1 −2i )µi −j
e1 −2k )νk −ℓ
(1
−
tz
(1
−
tz
z
2π −1 S1 \E
i=0 j=0
k=0 ℓ=0
Introducing Φi,j (z) := (1 − z 2 )Gi,j (g(z)) and Ψk,ℓ (z) := (1 − z 2 )Hk,ℓ (g(z)) and observing that the integrands have
no singularities along E yields
d1 µX
e1 νX
i −1 Z
k −1 Z
X
X
dz
dz
1
1
Φi,j (z)
Ψk,ℓ (z)
(6.6)
HilbV (t) = √
+ √
.
d1 −2i )µi −j z
e1 −2k )νk −ℓ z
(1
−
tz
(1
−
tz
1
1
2π −1 i=0 j=0 S
2π −1 k=0 ℓ=0 S
Now, for a non-negative P
integer a, recall [30], [18, Section 4] the operation Ua : Q[[z]] → Q[[t]] that assigns to a
∞
formal power series F (z) = i=0 Fi z i the series
(Ua F )(t) :=
∞
X
i=0
Fia ti ∈ Q[[t]].
By [18, Lemma 4.1], if F (t) is the power series of a rational function, then (Ua F )(t) is as well. Similarly, if a 6= 0
Ua can be described in terms of averaging over ath roots of unity:
√
1 X
a
(Ua F )(t) =
F (ζ t).
a a
ζ =1
We have the following.
n
d
n
n
followed
Proposition 6.4. For n ≥ 0, define the differential operator Dn := dt
n ◦ (t ·), i.e. multiplication by t
P∞
i
by n-fold differentiation with respect to t. If F (z) = i=0 Fi z is convergent near z = 0, then
Z
1
F (z)
dz
1
√
=
(Dm−1 (Ua F )) (t).
(m − 1)!
2π −1 S1 (1 − tz −a )m z
P∞
i
Proof. It is easy to see that both sides are equal to i=0 m−1+i
m−1 Fia t .
Applying Proposition 6.4 to Equation (6.6), we obtain
d1
(6.7)
HilbV (t) =
i −1
2 µ
X
X
(Dµ
e1 −1
νX
2
k −1
X
(Ud1 −2i Φi,j )) (t)
(Dνk −ℓ−1 (Ue1 −2k Ψk,ℓ )) (t)
+
.
(µi − j − 1)!
(νk − ℓ − 1)!
i −j−1
i=0 j=0
k=0 ℓ=0
To compute each (Ud1 −2i Φi,j )(t) and (Ue1 −2k Ψk,ℓ )(t), we follow the process described in [18, Section 4]. Specifically, when computing (Ua F )(t) for a rational function F (z) whose denominator consists of factors of the form
(1 − z b ), each such factor transforms by the rule
(1 − z b ) 7−→ (1 − tlcm(a,b)/a )gcd(a,b)
to yield the denominator of (Ua F )(t). Then we can determine the numerator num((Ua F )(t)) of (Ua F )(t) via
num((Ua F )(t)) = (Ua F )(t) denom((Ua F )(t)).
Writing HilbV (t) = P (t)/Q(t), we have by Kempf’s bound [23, Theorem 4.3] that deg(P ) ≤ deg(Q). Hence for
each F , as any terms in the numerator with degree larger than that of the denominator will cancel in the complete
expression, we need only determine the Taylor expansion up to a deg(Q).
This algorithm has been implemented on Mathematica and is available from the authors by request. We have
been able to use it to compute the Hilbert series of large irreducible representations on a PC; as an example,
18
P. DE CARVALHO CAYRES PINTO, H.-C. HERBIG, D. HERDEN, AND C. SEATON
Q49
HilbV50 (t) was computed in 52 hours. It has denominator (1 − t2 )2 i=3 (1 − ti ) and a numerator of degree 1175
with largest coefficient approximately 1.6996 × 1052 .
For a simple reducible example with multiplicities, V2 ⊕ V3 ⊕ V3 is computed in a few seconds; the Hilbert series
is given by
−t21 − t18 − 3t17 − 4t16 − 5t15 − 8t14 − 7t13 − 3t12 − 2t11 + 2t10 + 3t9 + 7t8 + 8t7 + 5t6 + 4t5 + 3t4 + t3 + 1
.
(1 − t2 )2 (1 − t3 )2 (1 − t4 )3 (1 − t5 )2
References
1. Luchezar L. Avramov, Ragnar-Olaf Buchweitz, and Judith D. Sally, Laurent coefficients and Ext of finite graded modules, Math.
Ann. 307 (1997), no. 3, 401–415.
2. L. P. Bedratyuk, Poincaré series of the multigraded algebras of SL2 -invariants, Ukrainian Math. J. 63 (2011), no. 6, 880–890.
3. Leonid Bedratyuk, The Poincare series for the algebra of covariants of a binary form, Int. J. Algebra 4 (2010), no. 25-28, 1201–1207.
4.
, Bivariate Poincaré series for the algebra of covariants of a binary form, ISRN Algebra (2011), Art. ID 312789, 11.
5.
, The MAPLE package for SL2 -invariants and kernel of Weitzenböck derivations, (2011), arXiv:1101.0622 [math.AG].
6. Leonid Bedratyuk and Lyubomyr Bedratyuk, Multivariate Poincaré series for algebras of SL2 -invariants, C. R. Acad. Bulgare
Sci. 64 (2011), no. 6, 807–814.
7. Leonid Bedratyuk and Nadia Ilash, The degree of the algebra of covariants of a binary form, J. Commut. Algebra 7 (2015), no. 4,
459–472.
8. D. J. Benson, Polynomial invariants of finite groups, London Mathematical Society Lecture Note Series, vol. 190, Cambridge
University Press, Cambridge, 1993.
9. D. J. Benson and W. W. Crawley-Boevey, A ramification formula for Poincaré series, and a hyperplane formula for modular
invariants, Bull. London Math. Soc. 27 (1995), no. 5, 435–440.
10. Winfried Bruns and Jürgen Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics, vol. 39, Cambridge
University Press, Cambridge, 1993.
11. L. Emily Cowie, Hans-Christian Herbig, Daniel Herden, and Christopher Seaton, The Hilbert series and a-invariant of circle
invariants, (2017), arXiv:1707.03128 [math.RA].
12. Harm Derksen and Gregor Kemper, Computational invariant theory, Invariant Theory and Algebraic Transformation Groups, I,
Springer-Verlag, Berlin, 2002, Encyclopaedia of Mathematical Sciences, 130.
13. Michael Forger, Invariant polynomials and Molien functions, J. Math. Phys. 39 (1998), no. 2, 1107–1141.
14. William Fulton and Joe Harris, Representation theory, Graduate Texts in Mathematics, vol. 129, Springer-Verlag, New York, 1991,
A first course, Readings in Mathematics.
15. Shiro Goto and Keiichi Watanabe, On graded rings. I, J. Math. Soc. Japan 30 (1978), no. 2, 179–213.
16. Hans-Christian Herbig, Daniel Herden, and Christopher Seaton, On compositions with x2 /(1 − x), Proc. Amer. Math. Soc. 143
(2015), no. 11, 4583–4596.
17. Hans-Christian Herbig and Gerald W. Schwarz, The Koszul complex of a moment map, J. Symplectic Geom. 11 (2013), no. 3,
497–508.
18. Hans-Christian Herbig and Christopher Seaton, The Hilbert series of a linear symplectic circle quotient, Exp. Math. 23 (2014),
no. 1, 46–65.
19. David Hilbert, Ueber die vollen Invariantensysteme, Math. Ann. 42 (1893), no. 3, 313–373.
20. Melvin Hochster and Joel L. Roberts, Rings of invariants of reductive groups acting on regular rings are Cohen-Macaulay, Advances
in Math. 13 (1974), 115–175.
21. Nadia Ilash, The Poincaré series for the algebras of joint invariants and covariants of n linear forms, C. R. Acad. Bulgare Sci. 68
(2015), no. 6, 715–724.
, Poincaré series for the algebras of joint invariants and covariants of n quadratic forms, Carpathian Math. Publ. 9 (2017),
22.
no. 1, 57–62.
23. George Kempf, The Hochster-Roberts theorem of invariant theory, Michigan Math. J. 26 (1979), no. 1, 19–32.
24. Friedrich Knop, Der kanonische Modul eines Invariantenrings, J. Algebra 127 (1989), no. 1, 40–54.
25. Friedrich Knop and Peter Littelmann, Der Grad erzeugender Funktionen von Invariantenringen, Math. Z. 196 (1987), no. 2,
211–229.
26. P. Littelmann and C. Procesi, On the Poincaré series of the invariants of binary forms, J. Algebra 133 (1990), no. 2, 490–499.
27. V. L. Popov, Groups, generators, syzygies, and orbits in invariant theory, Translations of Mathematical Monographs, vol. 100,
American Mathematical Society, Providence, RI, 1992, Translated from the Russian by A. Martsinkovsky.
28. V. L. Popov and È. B. Vinberg, Invariant theory, Algebraic geometry. IV, Encyclopaedia of Mathematical Sciences, vol. 55,
Springer-Verlag, Berlin, 1994, Linear algebraic groups. Invariant theory, A translation of ıt Algebraic geometry. 4 (Russian), Akad.
Nauk SSSR Vsesoyuz. Inst. Nauchn. i Tekhn. Inform., Moscow, 1989 [ MR1100483 (91k:14001)], Translation edited by A. N. Parshin
and I. R. Shafarevich, pp. vi+284.
29. Gerald W. Schwarz, Lifting differential operators from orbit spaces, Ann. Sci. École Norm. Sup. (4) 28 (1995), no. 3, 253–305.
30. T. A. Springer, On the invariant theory of SU2 , Nederl. Akad. Wetensch. Indag. Math. 42 (1980), no. 3, 339–345.
31. Richard P. Stanley, Invariants of finite groups and their applications to combinatorics, Bull. Amer. Math. Soc. (N.S.) 1 (1979),
no. 3, 475–511.
32. Bernd Sturmfels, Algorithms in invariant theory, Texts and Monographs in Symbolic Computation, Springer-Verlag, Vienna, 1993.
33. Wolfram Research, Mathematica edition: Version 7.0, (2008), http://www.wolfram.com/mathematica/.
THE HILBERT SERIES OF SL2 -INVARIANTS
19
PEE/COPPE, Universidade Federal do Rio de Janeiro, Av. Athos da Silveira Ramos 149, Centro de Tecnologia Bloco H, Caixa postal 68504, CEP: 21941-972, Rio de Janeiro, Brazil
E-mail address: [email protected]
Departamento de Matemática Aplicada, Av. Athos da Silveira Ramos 149, Centro de Tecnologia - Bloco C, CEP:
21941-909 - Rio de Janeiro, Brazil
E-mail address: [email protected]
Department of Mathematics, Baylor University, One Bear Place #97328, Waco, TX 76798-7328, USA
E-mail address: Daniel [email protected]
Department of Mathematics and Computer Science, Rhodes College, 2000 N. Parkway, Memphis, TN 38112
E-mail address: [email protected]
| 0math.AC
|
Ontology-based Fuzzy Markup Language Agent for
Student and Robot Co-Learning
Chang-Shing Lee, Mei-Hui Wang, Tzong-Xiang Huang
Li-Chung Chen, Yung-Ching Huang, Sheng-Chi Yang
Dept. of Computer Science and Information Engineering
National University of Tainan
Tainan, Tawain
[email protected]
Chien-Hsun Tseng, Pi-Hsia Hung
Dept. of Education
National University of Tainan
Tainan, Tawain
[email protected]
Abstract-An intelligent robot agent based on domain ontology,
machine learning mechanism, and Fuzzy Markup Language
(FML) for students and robot co-learning is presented in this
paper. The machine-human co-learning model is established to
help various students learn the mathematical concepts based on
their learning ability and performance. Meanwhile, the robot acts
as a teacher’s assistant to co-learn with children in the class. The
FML-based knowledge base and rule base are embedded in the
robot so that the teachers can get feedback from the robot on
whether students make progress or not. Next, we inferred students’
learning performance based on learning content’s difficulty and
students’ ability, concentration level, as well as teamwork sprit in
the class. Experimental results show that learning with the robot
is helpful for disadvantaged and below-basic children. Moreover,
the accuracy of the intelligent FML-based agent for student
learning is increased after machine learning mechanism.
Keywords—Ontology, Fuzzy Markup Language, Intelligent
Agent, Student Learning, Robot
I. INTRODUCTION
Ontology model can provide knowledge representation and
reasoning capabilities for machines to solve a task as well as to
allow semantic interoperability between systems or agents [1].
Owning to the rapid advance in artificial intelligence (AI),
Sophia, a social humanoid robot, came to the world in 2015. She
was programmed to give pre-written responses to specific
questions and also became the first ever to be granted a full
Saudi Arabian citizenship in 2017 [2]. Additionally, Liu et al. [3]
proposed a fuzzy ontology representation model to express
common fuzzy knowledge. Meditskos and Kompatsiaris [4]
presented iKnow to capitalize on the use of OWL ontological
knowledge to capture domain relationships between low-level
observations and high-level activities. Lee et al. used the
ontology to represent the knowledge of patent technology
requirement evaluation and recommendation [5] as well as
proposed a type-2 fuzzy ontology for personal diabetic-diet
recommendation [6].
Fuzzy Markup Language (FML) is a specific purpose
markup language based on XML to describe the structure and
behavior of a fuzzy system independently of the hardware
architecture [13-14]. Since May 2016, FML has become one of
the IEEE standards [15-16] and has been applied to a lot of
researches like game of Go [11, 17], diet assessment [14, 18, 22],
and so on. Nowadays, human and machine co-learning is an
important topic for current societies. One way to provide a robot
The authors would like to thank the financially support sponsored by the Ministry of
Science and Technology of Taiwan under the grants MOST 106-3114-E-024-001 and 1062221-E-024-019.
XXX-X-XXXX-XXXX-X/XX/$XX.00 © 20XX IEEE
Naoyuki Kubota
Dept. of System Design
Tokyo Metropolitan University
Tokyo, Japan
[email protected]
with such learning capability is to use machine learning [8].
Learning explores and understands the learning process of
humans, and machine learning studies how algorithms learn
from data [9]. Jain et al. [10] proposed an artificial-based student
learning evaluation tool to test with students from undergraduate
courses. There exists some FML-based real-world applications
to persons’ learning. For example, Lee et al. proposed a FMLbased intelligent adaptive assessment platform for learning
materials recommendation [20], and an online self-learning
platform construction based on genetic FML (GFML) and item
response theory (IRT) agent [21]. They also proposed a FMLbased dynamic assessment agent for human-machine
cooperative system on game of Go [11]. By combining particle
swarm optimization (PSO) with FML, called PFML, Lee et al.
applied to human and machine co-learning on game of Go [12]
and student learning performance evaluation [19].
Advancement in technology is bringing robots into
interpersonal aspects of student’s learning [7]; therefore,
including the robot into the class to co-learn with humans has
been a trend for recent years. This paper brings the robots Palro,
developed by FUJISOFT Japan and Zenbo, developed by ASUS,
Taiwan, into an elementary school to co-learn mathematics with
four-grade children. The objective of this paper is to represent
the knowledge of the robot for student and robot co-learning.
We first construct the student learning performance ontology for
the robot agent to predict their learning performance. Then, we
construct the student and robot co-learning ontology to
recommend students for suitable learning contents. After that,
we use FML to describe the knowledge base and rule base for
the constructed ontologies. Finally, we apply the developed
robot agent to the four-grade students for learning mathematical
concepts of number line and groups of numbers. We also use
machine learning techniques to optimize the involved student
learning performance using GFML [17-18] and PFML [19]. The
experiments show that the proposed ontology-based fuzzy
markup language agent is feasible for student and robot colearning.
The remainder of this paper is organized as follows: Section
II introduces the ontology model for student and robot colearning. Section III describes the fuzzy markup language agent,
including the proposed system structure as well as knowledge
base, rule base, and optimization model for the proposed student
and robot co-learning. The experimental results are shown in
Section IV and conclusions are given in Section V.
II. ONTOLOGY MODEL FOR STUDENT AND ROBOT CO-LEARNING
A. Student Learning Performance Ontology for Robot Agent
The structure of student learning performance ontology
based on Fuzzy Markup Language (FML) for robot agent
reasoning is shown in Fig.1. The domain name is defined as
FML output variable which connects to the FML linguistic
concepts of the output fuzzy variable. There are some FML input
variables, for example, FML input variable 1, FML input
variable 2, FML input variable 3, …, and FML input variable M,
to connect to the linguistic concepts of FML input variables.
Each FML input variable contains some linguistic concepts. The
linguistic concept of FML input variable contains some
attributes, such as Area, Grade, Subject, and so on. In addition,
some operations are also defined in the linguistic concept, for
example, operation Recommend.
Generalization
FML Output Variable
Aggregation
…
FML Linguistic Concept 1
FML Linguistic Concept 2
…
Learning Content Difficulty (LCD)
Association
FML Linguistic Concept 3
…
FML Input Variable 1
C. Student and Robot Co-Learning Ontology
Fig. 3 shows student and robot co-learning ontology. The
developed robot agent predicts the involved students’ learning
performance in the class according to the constructed student
learning performance ontology shown in Fig. 2. Next, according
to the feedback of the students’ learning performance and
students’ ability, the robot agent provides students with suitable
learning contents for their next study. Fig. 3 also shows partial
learning content ontology about the concepts of number line and
groups of numbers. Categories Number and Calculation and
Quantity and Measurement have concepts Number Line,
Distance, Four fundamental operations of arithmetic, …, etc.
For example, before knowing the concepts of Number Line,
students should know the concepts of Positive Number which
includes the concept of Positive Integer.
…
FML Input Variable 2
FML Input Variable 3
VeryEasy
FML Linguistic Concept N
…
…
…
Easy
Student Learning Performance
(SLP)
Student Ability (SA)
FML Input Variable M
Area: Countryside
Grade: Fourth
Subject: Math
FML Linguistic Concept 1
Area:
Grade:
Subject:
…
Area:
Grade:
Subject:
Rec omm end:
Proficient
Area: Countryside
Grade: Fourth
Subject: Math
Area: Countryside
Grade: Fourth
Subject: Math
...
...
...
...
...
...
Rec omm end:
Rec omm end:
Rec omm end:
Rec omm end:
Fig. 1. Structure of FML-based student learning performance ontology.
B. An Instance of Student Learning Performance Ontology
Fig. 2 shows an instance of FML-based student learning
performance ontology. The domain name for FML output
variable is Student Learning Performance (SLP), there are four
FML input variables in the ontology, including: Student Ability
(SA), Learning Content Difficult (LCD), Student Concentration
Level (SCL), and Student Teamwork Spirit (STS). There are five
linguistic concepts defined in FML output variable, including
FallBehind, Insufficient, Basic, Good, and Excellent. Each
FML input variable contains four linguistic concepts in this
paper. For instance, LCD has an association relations with
VeryEasy, Easy, Average, and Hard. Each linguistic concept
contains some attributes like Area, Grade, and Subject, as well
as operations like Recommend.
(a)
Fourth-Grade Mathematical
Course Materials Ontology for
Elementary School
Number and
Calculation
Know
Know
Good
…
Know
Know
...
Positive
Number
Include
Sense of
Volume
Positive
Integer
Student Concentration Level
Student Teamwork Spirit
(SA)
(LCD)
(SCL)
(STS)
Recommend =
[VeryEasy, Easy, Average, Hard]
Area: Countryside
… Grade: form third to fourth
Subject: Mathematics
...
Recommend =
[VeryEasy, Easy, Average, Hard]
Area: Countryside
… Grade: form third to fourth
Subject: Mathematics
...
...
...
Recommend =
[VeryEasy, Easy, Average, Hard]
...
...
...
...
Recommend =
[VeryEasy, Easy, Average, Hard]
Area: Countryside
… Grade: form third to fourth
Subject: Mathematics
Positive
Distracted
VeryEasy
calculation of
Multiplication
...
calculation of
addition
Mix calculation of
addition and multiplication
Conception of equalization
Concepts of groups of numbers
…
Learning Content Difficulty
Area: Countryside
Grade: form third to fourth
Subject: Mathematics
...
Excellent
Student Ability
BelowBasic
Four fundamental
operations of aritgmetic
...
Distance
Algebra
...
Basic
…
…
Statistics and
Probability
Concepts of Number Line
(SLP)
Insufficient
Quantity and
Measurement
...
Number
Line
Student Learning Performance
FallBehind
Rec omm end:
BelowBasic
Good
Area: Countryside
Grade: Fourth
Subject: Math
...
Recommend:
Recommend:
Rec omm end:
Rec omm end:
Insufficient
Area: Countryside
Grade: Fourth
Subject: Math
...
Recommend:
Area:
Grade:
Subject:
...
...
...
...
Recommend:
…
Advanced
Area: Countryside
Grade: Fourth
Subject: Math
...
…
Basic
Area: Countryside
Grade: Fourth
Subject: Math
...
...
Rec omm end:
Area:
Grade:
Subject:
...
FML Linguistic Concept NM
Exc ellent
Area: Countryside
Grade: Fourth
Subject: Math
Basic
Area: Countryside
Grade: Fourth
Subject: Math
...
FML Linguistic Concept N1
Hard
…
…
FallBehind
FML Linguistic Concept 1
Average
…
Fig. 2. An instance of FML-based student learning performance ontology.
(b)
Fig. 3.
(a) Student and robot co-learning ontology and (b) an example of
learning content ontology of the fourth-grade number line and groups of
numbers.
III. FML-BASED INTELLIGENT AGENTS FOR STUDENT AND ROBOT
CO-LEARNING
A. System Structure for Student and Robot Co-Learning
We propose the FML-based intelligent agents, including
an ontology agent, a teaching assistant agent, a co-learning
learning, an assessment agent, and a recommendation agent,
for student and robot co-learning in this section. Fig. 4 shows
the structure of FML-based intelligent robot agents for student
learning performance assessment and learning content
recommendation. The publisher finds domain experts to write
the textbook for teacher’s teaching and student’s learning after
the government, for example, Ministry of Education, defines
the learning outline for different grades of students. The
ontology agent constructs the domain ontology based on the
textbook for the teaching assistant agent. In addition, the colearning agent helps teacher teach students in the class and the
assessment agent classifies the student learning performance
into five categories, including FallBehind, Insufficient, Basic,
Good, and Excellent. Finally, the recommendation agent helps
teachers and students choose suitable learning contents for their
further study and learning. Fig. 5 shows the communication
structure among the developed learning contents, students, and
the robot Palro which communicates with the server via the
developed robot socket client.
Nonfocused, Focused, Absorbed}; (4) Student Teamwork Spirit
(STS) = {Passive, Normal, Initiative, Positive}. In addition, we
define the knowledge base for FML output variable Student
Learning Performance (SLP) = {FallBehind, Insufficient,
Basic, Good, Excellent}. Figs. 6(a)-6(e) show the fuzzy sets for
fuzzy variables SA, LCD, SCL, STS, and SLP, respectively.
(a)
(b)
(c)
Math
...
Teaching Assistant Agent
(d)
Co-learning Agent
...
Text book
...
Ontology Agent
Learning Outline
Excellent
Good
Basic
Insufficient
Recommendation Agent
Assessment Agent
FML-based Intelligent Agents
Fig. 4. Structure of FML-based intelligent robot agent for student learning
performance assessment and learning content recommendation.
Client
Server
RobotSocketClient
NCHC
WebSocket
Database
Mathematical
Learning Contents
Local Notebook
Students
(e)
FallBehind
Fig. 6. FML input variables (a) SA, (b) LCD, (c) SCL, (d) STS, and (e) FML
output variable SLP.
C. Rule Base for Human and Robot Co-Learning
The FML robot agent first predicts students’ learning
performance based on the knowledge base, and then provides
suitable learning contents to students for next study according to
the students’ learning performance. In 2017, we developed the
mathematical learning contents using PHP and Java languages
to construct the learning-content server which allows students to
surf on it to learn in the class. Table I shows partial fuzzy rules
and Table II shows partial knowledge base and rule base of FML
that predicts students’ learning performance.
TABLE I.
Palro
Fig. 5. Communication structure among the developed learning contents,
students, and Palro.
B. Knowledge Base for Human and Robot Co-Learning
In this paper, we propose a FML Robot Agent for Student
Learning Performance Prediction, the knowledge base for FML
input variables are defined as follows: (1) Student Ability (SA) =
{BelowBasic, Basic, Proficient, Advanced}; (2) Learning
Content Difficulty (LCD) = {VeryEasy, Easy, Average, Hard};
(3) Student Concentration Level (SCL) = {Distracted,
No
1
2
3
4
5
6
7
8
9
10
SA
BelowBasic
BelowBasic
BelowBasic
BelowBasic
BelowBasic
BelowBasic
BelowBasic
BelowBasic
BelowBasic
BelowBasic
LCD
VeryEasy
VeryEasy
VeryEasy
VeryEasy
VeryEasy
VeryEasy
VeryEasy
VeryEasy
VeryEasy
VeryEasy
PARTIAL FUZZY RULES
SCL
Distracted
Distracted
Distracted
Distracted
Nonfocused
Nonfocused
Nonfocused
Nonfocused
Focused
Focused
⋮
STS
Passive
Normal
Initiative
Positive
Passive
Normal
Initiative
Positive
Passive
Normal
SLP
FallBehind
FallBehind
FallBehind
FallBehind
FallBehind
FallBehind
FallBehind
Insufficient
FallBehind
FallBehind
250
251
252
253
254
255
256
Advanced
Advanced
Advanced
Advanced
Advanced
Advanced
Advanced
TABLE II.
Hard
Hard
Hard
Hard
Hard
Hard
Hard
Focused
Focused
Focused
Absorbed
Absorbed
Absorbed
Absorbed
Normal
Initiative
Positive
Passive
Normal
Initiative
Positive
Excellent
Excellent
Excellent
Excellent
Excellent
Excellent
Excellent
PARTIAL KNOWLEDGE BASE AND RULE BASE OF FML
and one FML output variable represent the position of the
particle in 5-dimensional space where are optimized by
adjusting the moving velocity in order to reach convergence.
The domain of the particle in each dimension is bounded in
[domain left, domain right] of the FML variable [19]. In this
paper, the domains of from the first to the fifth dimension are
[−4, 4], [−4, 4], [0, 10], [0, 10], and [0, 1] to optimize the
parameters of FML variables SA, LCD, SCL, STS, and SLP,
respectively.
IV. EXPERMENTAL RESULTS
There are three parts in the experimental results: (1) Parts 1
and 2 are to test the behavior for FML-based intelligent agent,
including an assessment agent and a recommendation agent for
student learning performance assessment and learning content
recommendation, respectively. (2) Part 3 is to deploy the FMLbased intelligent agent with different robots, Palro and Zenbo,
to an elementary school for four-grade students that co-learned
with the robots about mathematical concepts of number line and
types of numbers in Nov. and Dec. 2017, respectively.
⋮
⋮
⋮
D. Machine Learning for knowledge base optimalization
This subsection describes the machine learning methods,
including GFML [17, 18, 22] and PFML [19] to optimize the
knowledge base of Fuzzy Markup Language. The former is to
combine genetic algorithm with FML and the latter is to
combine PSO with FML. Fig. 7 shows the FML-based
intelligent agents with a machine learning mechanism for
students learning. For GFML, three types of genes are defined,
including knowledge-based genes (composed of the FML
variables’ names and the objects with the linguistic terms of
their own fuzzy variables), rule-based genes (a collection of the
weight of the fuzzy rules), and fuzzy-hedged genes (linguistic
terms’ hedge of the fuzzy variables) [22]. In this paper, the
encoded chromosome has 266 genes, including the knowledgebased genes (G1 to G5), the rule-based genes (G6 to G261), and
the fuzzy-hedged genes (G262 to G266).
FML KB / RB
Learning Data
Machine Learning
Co-Learning Agent
N
Termination ?
Y
A. Part 1: Student Learning Performance Assessment
In Part 1 of the experiments, we propose a FML robot agent
for student learning performance assessment. The knowledge
base for FML input variables are defined as follows: (1) Student
Ability (SA) = {BelowBasic, Basic, Proficient, Advanced}={[4, -4, -1.11, -0.6], [-1.11, -0.6, 0.05, 0.4], [0.05, 0.4, 0.95, 1.5],
[0.95, 1.5, 4, 4]}; (2) Learning Content Difficulty (LCD) =
{VeryEasy, Easy, Average, Hard} ={[-4, -4, -1.11, -0.6], [1.11, -0.6, 0.05, 0.4], [0.05, 0.4, 0.95, 1.5], [0.95, 1.5, 4, 4]}; (3)
Student Concentration Level (SCL) = {Distracted, Nonfocused,
Focused, Absorbed} ={[0, 0, 2, 3], [2, 3, 4, 5], [4, 5, 6, 7], [6,
7, 10, 10]}; (4) Student Teamwork Spirit (STS) = {Passive,
Normal, Initiative, Positive}={[0, 0, 2, 3], [2, 3, 4, 5], [4, 5, 6,
7], [6, 7, 10, 10]}. In addition, we define the knowledge base
for FML output variable Student Learning Performance (SLP)
= {FallBehind, Insufficient, Basic, Good, Excellent}={[0.0,
0.0, 0.2, 0.3], [0.2, 0.3, 0.4, 0.5], [0.4, 0.5, 0.6, 0.7], [0.6, 0.7,
0.8, 0.9], [0.8, 0.9, 1, 1]}.
We first simulate 400 records and then use K-fold cross
validation method to evaluate the performance. The fitness
function is mean square error (MSE). In this paper, K = 5 which
means that 80% of data for training and 20% for testing. Figs.
8 and 9 show the learned fuzzy sets for fuzzy variables SA, LCD,
SCL, STS, and SLP, by applying GA (crossover rate / mutation
rate = 0.9 / 0.1) and PSO with 84 particles learning mechanisms
to learn 1000 and 3000 generations, respectively. Fig. 10 shows
that learning 3000 generations performs better than the others
for both GA and PSO. Additionally, the proposed PSO learning
method has a better performance than GA learning.
BelowBasic
Teaching Field
Basic
Advanced
Proficient
1
Teaching Assistant Agent
After-Learning
FML KB / RB
SA
-3.98
-3.95
Fig. 7. FML-based intelligent agents with a machine learing mechanism for
students learning.
-2.65
-1.11 -0.71
-0.15 0.05
0.61
0.4
2.23
0.95
3.58
4
(a)
VeryEasy
Easy
Average
Hard
1
LCD
For PFML, the total number of parameters for each particle
is 84 in this paper. The parameters of four FML input variables
-4 -3.77
-1.16 -0.92
-1.08 -0.75
0.05 0.16
0.4
0.58
(b)
1.15
1.58 1.75
3.18
4
Distracted
Focused
Nonfocused
Absorbed
1
SCL
0
2.87 2.99
3 3.03
0.8
5.63 5.8
5.45
4.35
6.61
7
7.52
10
(c)
Passive
Normal
Initiative
Positive
1
STS
0.6
0
3.65
2.85
1.81
5.5 5.68
4.04 4.53
4.33
6.23
7.51
6.94
10
(d)
FallBehind
Basic
Insufficient
Good
1
Excellent
SLP
0 0.01
0.259
0.267
0.14
0.09
0.46 0.48
0.38
0.37
0.56
0.54
0.61
0.71
0.73
0.8
0.85
1
0.99
(e)
Fig. 8. FML variables after GA learning (a) SA, (b) LCD, (c) SCL, (d) STS,
and (e) SLP.
0
Basic
BelowBasic
1
-4
-3.25
-2.36
-0.15
Proficient
Advanced
1.37 1.46
-0.11 -0.01
4 SA
3.84
(a)
Average
Easy
VeryEasy
Hard
1
0
-4 -3.85
-1.38
0.31
4
2.43
0.8
LCD
(b)
Distracted
Nonfocused
1
Focused
Absorbed
Performance (SLP) and the output fuzzy variable is
Recommended Learning Content Rank (RLCR) with 8 linguistic
terms including last-grade high intermediate level (LGHIL),
last-grade advanced level (LGAL), current-grade elementary
level (CGEL), current-grade intermediate level (CGIL),
current-grade high intermediate level (CGHIL), current-grade
advanced level (CGAL), next-grade elementary level (NGEL),
and next-grade intermediate level (NGIL). The range of RLCR
is between -4 and +4 and it is the same as student’s ability [20].
Table III shows partial knowledge base and rule base of
learning content recommendation which is constructed
according to the learned knowledge of PSO learning
mechanism. The total number of fuzzy rules is 20 listed in Table
IV. Table V lists partial input data that recommend the learning
content rank and column RLCRDO is the desired output (DO).
Columns RLCRBLKB and RLCRALKB show partial inferred
results when we extracted the parameters of input fuzzy
variables SA and SLP from the before-learning and afterlearning knowledge of PSO learning mechanism, respectively.
Table V indicates that the robot agent with the learned
knowledge recommends more suitable learning contents owing
to an increase in accuracy from 78.75% to 87%.
TABLE III.
0
2.96
1.09
0
3.76
5.52
KNOWLEDGE BASE OF PART-2 EXPERIMENT.
10 SCL
7.99 8.26
(c)
0
Normal
Passive
1
1.62
0
2.05
Initiative
5.82
4.01
Positive
6.18
9.4
9.88
10 STS
(d)
FallBehind
Insufficient
Basic
1
0
0
0.28
0.08 0.1
0.33
0.48
Good
0.53
0.65
0.68
Excellent
0.94
0.72
1
SLP
(e)
Fig. 9. FML variables after PSO learning (a) SA, (b) LCD, (c) SCL, (d) STS,
and (e) SLP.
0.008
Before Learning
After GA Learning
0.007
After PSO Learning
MSE
0.006
0.0055
0.0055
0.0055
0.005
0.004
0.003
0.003
0.0028
0.0028
0.0027
0.0024
0.0022
0.002
0.001
TABLE IV.
0
1000 Generations
2000 Generations
FUZZY RULES OF RECOMMENDING LEARNING CONTENT.
3000 Generations
Fig. 10. MSE values of before learning, after GA learning, and after PSO
learning with evolving 1000, 2000, and 3000 generations.
B. Part 2: Learning Content Recommendation
The purpose of Part 2 of the experiments is to recommend
learning contents for next students’ learning by feeding the
learned knowledge from Part 1 into the robot. We categorize
the learning contents into four levels, including elementary,
intermediate, high intermediate, and advanced levels. The input
fuzzy variables are Student Ability (SA) and Student Learning
No.
1
2
3
4
5
6
7
8
9
10
SA
Below Basic
Below Basic
Below Basic
Below Basic
Below Basic
Basic
Basic
Basic
Basic
Basic
SLP
Fall Behind
Insufficient
Basic
Good
Excellent
Fall Behind
Insufficient
Basic
Good
Excellent
RLCR
LGHIL
LGAL
LGAL
CGEL
CGIL
LGAL
CGEL
CGIL
CGHIL
CGAL
No.
11
12
13
14
15
16
17
18
19
20
SA
Proficient
Proficient
Proficient
Proficient
Proficient
Advanced
Advanced
Advanced
Advanced
Advanced
SLP
Fall Behind
Insufficient
Basic
Good
Excellent
Fall Behind
Insufficient
Basic
Good
Excellent
RLCR
CGIL
CGHIL
CGAL
CGAL
NGEL
CGAL
CGAL
NGEL
NGIL
NGIL
SA
-1.43
-1.03
-2.23
-1.88
-3.74
-2.87
-1.68
-0.97
-1.5
-2.65
SLP
0.111
0.167
0.098
0.11
0.113
0.116
0.153
0.117
0.105
0.112
396
397
398
399
400
Accuracy
2.87
3.71
1.43
1.61
1.57
0.903
0.902
0.803
0.85
0.907
PARTIAL INPUT DATA.
RLCRDO
-1.99067
-1.57467
-2.55867
-2.29333
-3.52533
-2.93733
-2.04533
-1.668
-2.05333
-2.80133
RLCLBLKB
-3.611
-2.997
-2.985
-3.611
-3.611
-3.611
-2.791
-3.024
-3.48
-3.611
RLCRALKB
-2.248
-2.073
-2.67
-2.472
-3.611
-3.599
-2.369
-2.049
-3.072
-2.75
2.988
3.545333
1.761333
2.006667
2.132
3.314
3.367
3.079
3.367
3.339
78.75%
2.439
2.735
1.25
1.441
1.5533
87%
⋮
Moreover, the students cannot do next item until they correctly
answer the current item. Fig. 12 shows the average distance
between two points (students’ response data and correct
answer), on a number line from Item 1 to Item 22. If the average
distance is small, then the students’ learning performance is
good. We observe that Group B co-learns better with the robots
than Group A.
4.83
5
4.5
Average Distance
between response data and correct answer
TABLE V.
No
1
2
3
4
5
6
7
8
9
10
4.5
4
3.5
Group A
3.25
Group B
3
3
2.6
2.5
2
1.5
1.5
1
1
0.5
0.5 0.5
0.5
0.5
0.5
0 00
0
0 00
0
0
0
6
7
8
0 00 0
00 0
00 00
0 00 00 00
0 0
00
0
1
C. Part 3: FML-based Intelligent Agent for Student Learning
In Part 3 of the experiments, we deployed two different
robots, Palro and Zenbo, with the proposed FML-based
intelligent agent to an elementary school for students and robot
co-learning. The involved fourth-grade elementary students
were divided into two groups. Each group contains four
students with different levels of ability. The purpose of group
learning is to hope that students can learn together with the
robots by teamwork and that students with stronger learning
ability can guide weaker students along with the robot’s
assistance. Fig 11 shows the grouping learning diagram and
actual teaching situation that the involved four-grade students
and robot co-learning mathematics in the class in Kaohsiung,
Taiwan on Nov. 25 and Dec. 23, 2017. Group A learns
mathematics together with Palro and Zenbo but Group B does
with Palro and iPads.
Group B
Group A
2
3
First deploy is to learn the concepts of the number line on
Nov. 25, 2017 through play. The total number of the items is
22. The bigger the item number, the harder the item. The
students input their response data to shoot the target of the
number line shown on the screen of Zenbo (Group A) or iPad
(Group B). Meanwhile, Palro provided students with some hints
when they failed to hit the target but cheered for them when
they made it. In addition, Palro will provide different levels of
hints according to the number of incorrect answer. The more
number of incorrect answer, the more detailed the hints.
5
9
10 11 12 13 14 15 16 17 18 19 20 21 22
Item No.
Fig. 12. Average distance between two points (students’ response data and
correct answer) on a number line for Items 1−22 for the study on Nov. 25, 2017.
The second study was about learning the concept of groups
of numbers on Dec. 23, 2017 through play. Palro and Zenbo
also play the similar role to the first study on Nov. 25, 2017.
We used three monsters, including MonsterA, MonsterB, and
MonsterC, to represent the scores 8, 12, and –4, respectively.
The involved students hit the exact number of monsters to
complete their mission. For example, if their mission is to get
score 56 by hitting three kinds of monsters, they can hit three
MonsterAs, one MonsterB, and one MonsterC to get score 3×8
+ 3×12 – 1 × 4 = 56. Their challenged difficulty is divided into
three levels, including Easy, Average, and Hard. Each
challenge has different numbers of missions. However, there is
an upper bound of the number of making a response to each
mission. Table VI shows the obtained score for each level. The
score is calculated according to how many times students make
a response to the mission and whether they successfully
complete the mission or not in the end. The total score is
bounded in [0, 29].
TABLE VI.
SCORE BASED ON HOW MANY TIMES STUDENTS TRY TO MAKE
A RESPONSE.
Times
Fig. 11. Grouping learning diagram and actual teaching situation in the class.
4
1
2
3
4
Easy
2
1
0
0
Challenge Level
Average
3
2
1
0
Hard
3
2
1
0
Fig. 13 shows the number of making a response of each
mission for Groups A and B on Dec. 23, 2017. The first mission
of each challenge is to test if students fully understand their
mission for the current challenge. Observe Fig. 13 that
understanding their mission of the Average and Hard
challenges is the most difficult one for both Groups A and B
when we compare to the other missions. The symbol “×” of Fig.
13 denotes that the involved students failed this mission in the
end. The total acquired score of Groups A and B is 23 and 22,
respectively, which means that two-group students perform
equally and grouping learning with the robot is feasible for the
involved students.
[7]
4
Number of making a response
3.5
Mission 1
3
Mission 2
3
3
[8]
3
Mission 3
2.5
Mission 4
2
2
2
2
2
[9]
1.5
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
[10]
0.5
0
Easy (Group A)
Easy (Group B)
Average (Group A)
Average (Group B)
Hard (Group A)
Hard (Group B)
Fig. 13. Number of making a response of each mission for Groups A and B on
Dec. 23, 2017.
[11]
V. CONCULSION AND FUTURE WORK
This paper presents an FML-based intelligent agent for
students and robot co-learning. The student learning
performance ontology and the student with robot co-learning
ontology are proposed for the intelligent agent. In addition, the
machine learning mechanism, including GA and PSO, are also
adopted for the knowledge base refinement. The machinehuman co-learning model is established to help various students
learn the mathematical concepts based on their learning ability
and performance. Experimental results show that learning with
the robot is helpful for the involved students. In the future, the
intelligent agent with different robots will be deployed in
various learning environments to help more students’ learning.
[12]
[13]
[14]
[15]
ACKNOWLEDGMENT
The authors would like to thank Han-Ru Liu, Jou-Te Tsai,
master students of Dept. of Education, National University of
Tainan (NUTN), faculty of San Pi Elementary School in
Kaohsiung for their help with introducing the developed
learning contents into the teaching field. Finally, we also would
like to thank the involved students of San Pi Elementary School.
[16]
[17]
[18]
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
E. Prestes, J. L. Carbonera, S. R. Fiorini, V. A. M. Jorge, M. Abel, R.
Madhavan, A. Locoro, P. Goncalves, M. E. Barreto, M. Habib, A. Chibani,
S. Gerard, Y. Amirat, and C. Schlenoff, “Towards a core ontology for
robotics and automation,” Robotics and Autonomous Systems,” vol. 61,
pp. 1193-1204, 2013.
H. Taylor, “Could you fall in love with this robot?,” Mar. 2016, [Online]
Available:
https://www.cnbc.com/2016/03/16/could-you-fall-in-lovewith-this-robot.html.
J. Liu, B. J. Zheng, L. M. Luo, J. S. Zhou, Y. Zhang, and Z. T. Yu,
“Ontology representation and mapping of common fuzzy knowledge,”
Neurocomputing, vol. 215, pp. 184-195, 2016.
G. Meditskos and I. Kompatsiaris, “iKnow: Ontology-driven situational
awareness for the recognition of activities of daily living,” Pervasive and
Mobile Computing, vol. 40, pp. 17-41, 2017.
C. S. Lee, M. H. Wang, Y. C. Hsiao, B. H. Tsai, “Ontology-based GFML
agent for patent technology requirement evaluation and recommendation,”
Soft-Computing, 2017. (DOI: 10.1007/s00500-017-2859-1)
C. S. Lee, M. H. Wang, and H. Hagras, “A type-2 fuzzy ontology and its
application to personal diabetic-diet recommendation,” IEEE
Transactions on Fuzzy Systems, vol. 18, no. 2, pp. 374-395, Apr. 2010.
[19]
[20]
[21]
[22]
A. Edwards, C. Edwards, P. R. Spence, C. Harris, and A. Gambino,
“Robots in the classroom: differences in students’ perceptions of
credibility and learning between teacher as robot and robot as teachers,”
Compuers in Human Behavior, vol. 65, pp. 627-634, 2016.
E. Senft, P. Baxter, J. Kennedy, S. Lemaignan, and T. Belpaeme,
“Supervised autonomy for online learning in human-robot interaction,”
Pattern Recognition Letters,” vol. 99, pp. 77-86, 2017.
M. Vahdat, L. Oneto, D. Anguita, M. Funk, and M. Rauterberg, “Can
machine learning explain human learning?” Neurocomputing, vol. 192,
pp. 14-28, 2016
G. P. Jain, V. P. Gurupur, J. L. Schroeder, and E. D. Faulkenberry,
“Artificial intelligence-based student learning evaluation: a concept mapbased approach for analyzing a student’s understanding of a topic,” IEEE
Transactions on Learning Technologies, vol. 7, no. 3, pp. 267-279, 2014.
C. S. Lee, M. H. Wang, S. C. Yang, P. H. Hung, S. W. Lin, N. Shuo, N.
Kubota, C. H. Chou, P. C. Chou, and C. H. Kao, “FML-based dynamic
assessment agent for human-machine cooperative system on game of Go,”
International Journal of Uncertainty, Fuzziness and Knowledge-Based
Systems, vol. 25, no. 5, pp. 677-705, 2017.
C. S. Lee, M. H. Wang, S. C. Yang, and C. H. Kao, “From T2 FS-based
MoGoTW system to DyNaDF for human and machine co-learning on Go,”
in R. John, H. Hagras, and O. Castillo (editors), Type-2 Fuzzy Logic and
Systems Dedicated to Professor Jerry Mendel for his Pioneering
Contribution, Berlin, Springer, Mar. 2018, pp. 1-24.
G. Acampora and V. Loia, Vincenzo, “Using FML and fuzzy technology
in adaptive bmbient intelligence environments,” International Journal of
Computational Intelligence Research, vol. 1, no. 2.. pp. 171-182, 2005.
C. S. Lee, M. H. Wang, G. Acampora, C. Y. Hsu, and H. Hagras, “Diet
assessment based on type-2 fuzzy ontology and fuzzy markup language,”
International Journal of Intelligent System, vol. 25, no. 12, pp. 1187-1216,
2010.
G. Acampora, B. D. Stefano, and A. Vitiello, IEEE 1855TM: The first
IEEE standard sponsored by IEEE Computational Intelligence Society,
IEEE Computational Intelligence Magazine vol. 11, no. 4, pp. 4-6, 2015.
IEEE Computational Intelligence Society, 1855-2016-IEEE Standard for
Fuzzy
Markup
Language,
2016
[Online].
Available:
http://ieeexplore.ieee.org/servlet/opac?punumber=7479439.
C. S. Lee, M. H. Wang, M. J. Wu, O. Teytaud, and S. J. Yen, “T2FSbased adaptive linguistic assessment system for semantic analysis and
human performance evaluation on game of Go,” IEEE Transactions on
Fuzzy Systems, vol. 23, no. 2, pp. 400-420, Apr. 2015.
C. S. Lee, M. H. Wang, and S. T. Lan, “Adaptive personalized diet
linguistic recommendation mechanism based on type-2 fuzzy sets and
genetic fuzzy markup language,” IEEE Transactions on Fuzzy Systems,
vol. 23, no. 5, pp. 1777-1802, Oct. 2015.
S. Lee, M. H. Wang, C. S. Wang, O. Teytaud, J. L. Liu, S. W. Lin, and P.
H. Hung, “PSO-based fuzzy markup language for student learning
performance evaluation and educational application,” IEEE Transactions
on Fuzzy Systems, 2017. (Revised)
C. S. Lee, M. H. Wang, J. L. Yu, K. H. Lin, T. T. Lin, S. C. Yang, and S.
L. Cho, “FML-based intelligent adaptive assessment platform for learning
materials recommendation,” 2015 IEEE International Conference on
Fuzzy Systems (FUZZ-IEEE 2015), Istanbul, Turkey, Aug. 2-5, 2015.
C. S. Lee, M. H. Wang, K. H. Lin, S. C. Yang, and T. T. Lin, “GFMLbased IRT agent for online self-learning platform construction,” 2016
World Congress on Computational Intelligence (IEEE WCCI 2016),
Vancouver, Canada, Jul. 24-29, 2016.
C. S. Lee, M. H. Wang, H. Hagas, Z. W. Chen, S. T. Lan, S. E. Kuo, H.
C. Kuo, and H. H. Cheng, “A novel genetic fuzzy markup language and
its application to healthy diet assessment,” International Journal of
Uncertainty, Fuzziness, and Knowledge-Based Systems, vol. 20, no. 2, pp.
247-278, Oct. 2012.
| 2cs.AI
|
arXiv:1702.01345v1 [math.AC] 4 Feb 2017
Local dimension theory of tensor products of algebras
over a ring
Samir Bouchiba
∗
Department of Mathematics, Faculty of Sciences, University Moulay Ismail,
Meknes, Morocco
Abstract
Our main goal in this paper is to set the general frame for studying the
dimension theory of tensor products of algebras over an arbitrary ring R. Actually, we translate the theory initiated by A. Grothendieck and R. Sharp and
subsequently developed by A. Wadsworth on Krull dimension of tensor products
of algebras over a field k into the general setting of algebras over an arbitrary
ring R. For this sake, we introduce and study the notion of a fibred AF-ring
over a ring R. This concept extends naturally the notion of AF-ring over a field
introduced by A. Wadsworth in [14] to algebras over arbitrary rings. We prove
that Wadsworth theorems express local properties related to the fibre rings of
tensor products of algebras over a ring. Also, given a triplet of rings (R, A, B)
consisting of two R-algebras A and B such that A⊗R B 6= {0}, we introduce the
inherent notion to (R, A, B) of a B-fibred AF-ring which allows to compute the
Krull dimension of all fiber rings of the considered tensor product A ⊗R B. As
an application, we provide a formula for the Krull dimension of A ⊗R B when
A and B are R-algebras with A is zero-dimensional as well as for the Krull
dimension of A ⊗Z B when A is a fibred AF-ring over the ring of integers Z with
nonzero characteristic and B is an arbitrary ring. This enables us to answer a
question of Jorge Matinez on evaluating the Krull dimension of A ⊗Z B when
A is a Boolean ring. Actually, we prove that if A and B are rings such that
B
.
A ⊗Z B is not trivial and A is a Boolean ring, then dim(A ⊗Z B) = dim
2B
MSC (2000): 13D02; 13D05; 13D07; 16E05; 16E10.
∗
Email: [email protected]
2010 Mathematics Subject Classification: 13C15; 13A15.
Key words and phrases: Krull dimension; fibre ring; AF-ring; Fibred AF-ring; height; Boolean
ring.
1
Local dimension theory of algebras
1
2
Introduction
All rings considered in this paper are commutative with identity element and all
ring homomorphisms are unital. Here and subsequently, R stands for an arbitrary
ring and k stands for a field. Let A be a ring and p be a prime ideal of A. Then
Spec(A) denotes the set of all prime ideals of A and kA (p) denotes the quotient field
A
of . Also, if n ≥ 0 is a positive integer, A[n] denotes the polynomial ring in n
p
indeterminates A[X1 , X2 , · · · , Xn ] and ht(p[n]) stands for the height of the extended
ideal p[n] := p[X1 , X2 , · · · , Xn ] of p. Further, if A is an algebra over a field k, then
t.d.(A : k) denotes the transcendence degree of A over k. Any unreferenced material
is standard as in [11], [7] and [15].
It is a paper of R. Sharp on Krull dimension of tensor products of two field
extensions of a field k which gave the initial impetus to study the Krull dimension
of tensor products. Actually, in [12], Sharp proved that, for any two extension fields
K and L of k, dim(K ⊗k L) = min(t.d.(K : k), t.d.(L : k)) (actually, this result
appeared ten years earlier in Grothendieck’s EGA [9, Remarque 4.2.1.4, p. 349]).
This formula is rather surprising since, as one may expect, the structure of the
tensor product should reflect the way the two components interact and not only the
structure of each component. This fact affords motivation to Wadsworth to work on
this subject in [14]. He aimed at seeking geometric properties of primes of A ⊗k B
and to widen the scope of k-algebras A and B whose tensor product Krull dimension,
dim(A ⊗k B), shows exclusive dependence on individual characteristics of A and B.
The algebras which proved to be tractable for Krull dimension computations turned
out to be those rings A which satisfy the altitude formula over k (AF-rings for short),
that is,
A
: k = t.d.(Ap : k)
ht(p) + t.d.
p
for all prime ideals p of A. The class of AF-rings contains the most basic rings
of algebraic geometry, including finitely generated k-algebras. Wadsworth proved
through [14, Theorem 3.7] that if A is an AF-domain and B is any k-algebra, then
n
B
:k
:
dim(A ⊗k B) = max ht(q[t.d.(A : k)]) + min t.d.(A : k), ht(p) + t.d.
q
o
p ∈ Spec(A) and q ∈ Spec(B) .
As a consequence of this, [14, Theorem 3.8] states that if A1 and A2 are both AFdomains, then
dim(A1 ⊗k A2 ) = min dim(A1 ) + t.d.(A2 : k), t.d.(A1 : k) + dim(A2 ) .
Local dimension theory of algebras
3
Further, he gave a result which yields a classification of the prime ideals of A ⊗k B
according to their contractions to A and B (cf. [14, Proposition 2.3]). In [1], we
continued the work of Wadsworth and transferred all his theorems in [14] on AFdomains to AF-rings. This passage from domains to rings with zero-divisors is well
reflected in new formulas for the Krull dimension of tensor products involving AFrings. As it turns out from the present work, it is these formulas that are relevant
in our treatment of the general setting of tensor products over a ring R. We refer
the reader to [1, 2, 3, 4, 5, 6, 12, 13, 14] for basics and recent investigations on the
dimension theory of tensor products of algebras over a field.
The main goal of this paper is to set the general frame to study the dimension theory of tensor products of algebras over an arbitrary ring R. Actually, we
translate the theory initiated by A. Grothendieck and R. Sharp and subsequently
developed by A. Wadsworth on Krull dimension of tensor products of algebras over
a field k into the general setting of algebras over an arbitrary ring R. It turns out
that Wadsworth theorem express local properties related to the fibre rings of the
tensor products over an arbitrary ring R. For this sake, we introduce and study
the notion of a fibred AF-ring over a ring R. Actually, we say that an R-algebra
A is a fibred AF-ring over R if the fibre ring kR (p) ⊗R A is an AF-ring over kR (p)
for any prime ideal p of R such that kR (p) ⊗R A 6= {0}. When restricted to tensor
products over a field the notion of a fibred AF-ring boils down to the classical one
of an AF-ring. It is notable that all finitely generated algebras over R proved to be
fibred AF-rings as well as all zero-dimensional rings which are R-algebras. We prove
that the fibred AF-rings inherit all properties of Wadsworth introduced AF-rings.
Moreover, given a triplet of rings (R, A, B) consisting of R-algebras A and B such
that A ⊗R B 6= {0}, we introduce and study the notion of a B-fibred AF-ring over R
which is a somewhat inherent concept to the given triplet (R, A, B). So, when A is
a B-fibred AF-ring over R, we can explicit the Krull dimension of all fiber rings of
A ⊗R B and in various cases this enables us to determine dim(A ⊗R B). As an application, we compute the Krull dimension of A ⊗R B when A and B are R-algebras
with A is zero-dimensional. Also, we provide a formula for the Krull dimension of
A ⊗R B when A and B are R-algebras with A is zero-dimensional as well as for the
Krull dimension of A ⊗Z B when A is a fibred AF-ring over the ring of integers Z
with nonzero characteristic and B is an arbitrary ring. This allows us to answer a
question of Jorge Matinez on evaluating the Krull dimension of A ⊗Z B when A is
a Boolean ring. Actually, we prove that if A and B are rings such that A ⊗Z B is
B
.
not trivial and A is a Boolean ring, then dim(A ⊗Z B) = dim
2B
Acknowledgement
I would like to thank Professor Jeorge Martinez for the discussion led with him on
4
Local dimension theory of algebras
the possible connections between the dimension theory of tensor products of algebras
and the dimension theory of frames. His questions on this issue were the source of
motivation to write this paper.
2
Local spectrum and effective spectrum
This section introduces the effective spectrum of a ring R with respect to an Ralgebra as well as local notions of well known concepts of dimension theory of rings
such that the height of a prime ideal, the Krull dimension and the spectrum of a
ring.
Let R be an arbitrary ring and let A be an R-algebra. Denote by fA : R −→ A,
with fA (r) = r.1A for any r ∈ R and where 1A is the unit element of A, the ring
homomorphism defining the structure of algebra of A over R. Let K := Ker(fA ). It
is easily seen that for each prime ideal P of A, K ⊆ p := fA−1 (P ) and that the induced
R
R
A
−→ A and ff
−→ , defined by fA (r) = fA (r) and
homomorphisms fA :
A :
K
p
P
^
e
fA (r̃) = fA (r) for each r ∈ R, are injective. Let S be a multiplicative subset of R.
Recall that the localization of A by S is the S −1 R-algebra S −1 A := S −1 R⊗R A. Our
first result proves that S −1 A is isomorphic to a localization of A by a multiplicative
subset S of A. By virtue of this lemma, we deduce that, for any multiplicative
subset S of R, S −1 A enjoys all properties satisfied by the well known localization
by a multiplicative subset of A.
Lemma 2.1. Let R be a ring and A an R-algebra. Let S be a multiplicative subset
of R. Let S := fA (S) denote the the correspondingmultiplicativesubset of A. Then
X ri
X r i ai
−1
the natural map ϕ : S −1 A −→ S A such that ϕ
⊗ R ai =
, for
si
fA (si )
i∈Λ
i∈Λ
any finite set Λ, any {ri : i ∈ Λ} ⊆ R, {si : i ∈ Λ} ⊆ S and {ai : i ∈ Λ} ⊆ A, is an
isomorphism of R-algebras.
−1
Proof.
to see that the mapping f : S −1 R × A −→ S A such that
r It is easy
ra
is well defined and is R-biadditive. This yields the existence of
f ,a =
s
fA (s)
the assigned homomorphism ϕ of R-algebras.
Also,
to check that the
a
it isroutine
a 1
−1
−1
map g : S A −→ S R ⊗R A defined by g
:= g
= ⊗R a for each
fA (s)
s.1A
s
a ∈ A and each s ∈ S is an homomorphism of R-algebras. Then observe that ϕ ◦ g =
idS −1 A and g ◦ ϕ = idS −1 A . Hence ϕ is an isomorphism of R-algebras.
The above discussion allows us to announce the following lemma which collects
certain properties and facts about fibre rings. These properties were stated in [10,
5
Local dimension theory of algebras
page 84] in the Noetherian setting but in fact they hold in the general case.
Lemma 2.2. Let R be a ring and let A be an R-algebra. Let p be a prime ideal of
R
R. Let Sp := \ {0}. Then
p
A
A
P
1) kR (p) ⊗R A ∼
:= Sp−1
and kR (p) ⊗R P ∼
for each prime
= Sp−1
= Sp−1
pA
fA (p)A
pA
ideal P of A such that fA−1 (P ) = p.
2) Spec(kR (p) ⊗R A) = {kR (p) ⊗R P : P ∈ Spec(A) such that fA−1 (P ) = p}.
3) There exists an order-preserving bijective correspondence between the spectrum of
kR (p) ⊗R A and the set of prime ideals of A which contract to p over R.
4) Let P be a prime ideal of A and let p := fA−1 (P ). Then
(kR (p) ⊗R A)kR (p)⊗R P ∼
= kR (p) ⊗R AP .
5) Let P be a prime ideal of A and let p := fA−1 (P ). Then
A
kR (p) ⊗R A ∼
= kR (p) ⊗R .
kR (p) ⊗R P
P
Proof. In view of [10, p. 84], it remains to give a proof of (4) and (5).
4) Let P be a prime ideal of A and p := fA−1 (P ). Consider the multiplicative subset
T := R \ p of R. Then, by (1),
(kR (p) ⊗R A)kR (p)⊗R P
∼
=
∼
=
∼
=
∼
=
A
Sp−1
P
pA Sp−1 pA
A
P
pA pA
AP
pAP
R
⊗R AP .
p
Also, notice that, on the one hand,
T −1 R ⊗R
R
⊗R AP
p
R
∼
= T −1 ⊗R AP
p
R
= Sp−1 ⊗R AP
p
= kR (p) ⊗R AP
6
Local dimension theory of algebras
and, on the other,
T −1 R ⊗R
R
⊗R AP
p
R
⊗R (T −1 R ⊗R AP )
p
R
∼
⊗R T −1 AP
=
p
R
=
⊗R AP as fA (T ) ⊆ A \ P and thus each element of
p
fA (T ) is invertible in AP .
∼
=
It follows that (kR (p) ⊗R A)kR (p)⊗R P ∼
= kR (p) ⊗R AP , as desired.
5) Let P be a prime ideal of A and p := fA−1 (P ). Then, by (1),
kR (p) ⊗R A
kR (p) ⊗R P
∼
=
∼
=
∼
=
Sp−1 (A/pA)
Sp−1 (P/pA)
A
Sp−1
P
A
A
kR (p) ⊗R
as p
= (0).
P
P
This completes the proof.
Given an R-algebra A, it is to be noted that not all the prime ideals of R are
essential in capturing the prime ideal structure of A over R. Actually, there are
prime ideals and chains of prime ideals of R that have no effect on the structure of
the spectrum of A (see Example 2.5). That is the reason why we introduce in what
follows the notion of effective spectrum of R with respect to A and effective Krull
dimension of R with respect to A.
Definition 2.3. Let A be an R-algebra.
1) A prime ideal p of R is said to be an effective prime ideal of R with respect to A
if the fibre ring kR (p) ⊗R A 6= {0}.
2) We define the effective spectrum of R with respect to A to be the set denoted by
SpecA
e (R) consisting of all effective prime ideals p of R with respect to A, namely,
SpecA
e (R) = {p ∈ Spec(R) : kR (p) ⊗R A 6= {0}}.
A
Also, we denote by MaxA
e (R) the subset of maximal elements of Spece (R), that is,
the set of maximal effective prime ideals of R with respect to A.
3) Let p ∈ SpecA
e (R). We define the effective height of p with respect to A, denoted
A
by hte (p), to be the supremum of lengths of chains of effective prime ideals p0 ⊂
p1 ⊂ · · · ⊂ pn = p of R terminating at p.
4) We define the effective Krull dimension of R with respect to A to be the invariant
7
Local dimension theory of algebras
denoted by dimA
e (R) which is the supremum of effective heights of effective prime
ideals of R with respect to A, that is,
A
A
dimA
e (R) = sup{hte (p) : p ∈ Spece (R)}.
.
The following result determines the effective spectrum of a ring R with respect
to various constructions. Its proof is easy and thus omitted.
−1
Proposition 2.4. 1) Let A be an R-algebra. Then SpecA
e (R) = fA (Spec(A))
R[X ,X ,··· ,Xn ]
2) Let X1 , X2 , · · · , Xn be indeterminates over R. Then Spece 1 2
(R) =
Spec(R).
3) Let R ֒→ A be an integral extension of rings. Then SpecA
e (R) = Spec(R).
−1 R
S
(R) = {p ∈ Spec(R) : p∩S =
4) Let S be a multiplicative subset of R. Then Spece
∅}.
5) Let A be an R-algebra and S be a multiplicative subset of A. Then
−1
−1
SpeceS A (R) = fA−1 {P ∈ Spec(A) : P ∩S = ∅} = fA−1 (SpeceS A (A)) ⊆ SpecA
e (R).
R
6) Let I be an ideal of R. Then SpeceI (R) = {p ∈ Spec(R) : I ⊆ p}.
Next, for any positive integer n ≥ 2, we exhibit an example of a ring R and an Ralgebra A such that there exists a chain of distinct prime ideals p0 ⊂ p1 ⊂ · · · ⊂ pn in
R with both ends p0 , pn ∈ SpecA
e (R) while the intermediate elements p2 , · · · , pn−1 6∈
SpecA
(R).
It
would
be
interesting
to afford a ring T issued from R and A such that
e
A
Spec(T ) = Spece (R). This task is not easy and it turns out from the next example
that T is neither a localization of R nor a factor ring of R.
Example 2.5. Let k be a field and t be an indeterminate over k. It is known,
by [10, Lemma 1], that there exists an infinite number of formal power series
g1 (t), g2 (t), · · · , gm (t), · · · of k[[t]] which are algebraically independent over k. In
fact, we can choose the gi in the maximal ideal tk[[t]]. Actually, assume that
g1 (t) = 1 + a1 t + a2 t2 + · · · + am tm + · · · ∈ k[[t]] \ tk[[t]]. Then, observe that
h1 (t) := g1 (t) − 1 ∈ tk[[t]] and, for each integer i ≥ 2, hi (t) := gi (t) − gi (0)g1 (t) ∈
tk[[t]], and the formal power series h1 (t), h2 (t), · · · , hm (t), · · · are algebraically independent over k as k[h1 (t), hi1 (t), hi2 (t), · · · , hin (t)] = k[g1 (t), gi1 (t), gi2 (t), · · · , gin (t)]
is a polynomial ring in n + 1 indeterminates for any finite subset {i1 , i2 , · · · , in }
of N \ {0}. Therefore let g1 (t), g2 (t), · · · , gm (t), · · · ∈ tk[[t]] be algebraically independent elements over k. Let n ≥ 2 be an integer and X1 , X2 , · · · , Xn be
Local dimension theory of algebras
8
indeterminates over k. Let R := k[X1 , X2 , · · · , Xn ] and A := k[[t]] be the formal power series ring over k which is a rank one discrete valuation ring and thus
Spec(A) = {(0), tk[[t]]}. We endow A with the R-algebra structure induced by the
ring homomorphism fA : R −→ A such that fA (Xi ) = gi (t) for each i = 1, 2, · · · , n.
Therefore fA−1 (tk[[t]]) = (X1 , X2 , · · · , Xn ) is a maximal ideal of R of height n. Also,
as g1 , g2 , · · · , gn are algebraically independent over k, it is readily checked that fA is
injective. Hence fA−1 ((0)) = (0). It follows that SpecA
e (R) = {(0), (X1 , X2 , · · · , Xn )}
and, in particular, any prime ideal of R properly between (0) and the maximal ideal
(X1 , X2 , · · · , Xn ) is not effective with respect to A, as desired .
Remark 2.6. 1) Let A be an R-algebra. Then, for any A-algebra B, the natural
−1
B
map fA−1 : SpecB
e (A) −→ Spece (R) is surjective while, in general, fA : Spec(A) −→
Spec(R) is not so.
2) Let A be an R-algebra. Then dimA
e (R) ≤ dim(R) and this inequality might be
strict as proved by Example 2.5 which shows that for any positive integer n ≥ 2
there exists a ring R and an R-algebra A such that dim(R) = n while dimA
e (R) = 1,
as desired.
To get prepared for the general setting of tensor products, we next introduce
local notions of well known concepts of the dimension theory of rings, namely the
height of a prime ideal of a ring A, the spectrum of A and the Krull dimension of
A.
Definition 2.7. Let R be a ring and A be an R-algebra.
1) Let p ∈ SpecA
e (R). Then
a) Specp (A) := {I ∈ Spec(A) : fA−1 (I) = p} denotes the set of all prime ideals of
A which contract to p over R.
b) If I ∈ Specp (A), then the height of I at p, denoted by htp (I), is the maximum
of lengths of chains I0 ⊂ I1 ⊂ · · · ⊂ In = I of prime ideals of A which contract to p
over R.
c) The Krull dimension of A at p is the invariant
dimp (A) := max{htp (I) : I ∈ Specp (A)}.
2) We define the fibre Krull dimension of A with respect to R to be the maximal
length of chains of prime ideals of A lying over a common (effective) prime ideal of
R (with respect to A), that is the invariant
f-dimR (A) = sup{dimp (A) : p ∈ SpecA
e (R)}.
By virtue of Lemma 2.2, we get the following result which connects the above
local data of A with those relative to fibre rings issued from A.
Local dimension theory of algebras
9
Corollary 2.8. Let R be a ring and A be an R-algebra. Let p ∈ Spec(R). Then
1) There exists an order-preserving bijective correspondence between Spec(kR (p) ⊗R
A) and Specp (A).
2) If I ∈ Specp (A), then htp (I) = ht(kR (p) ⊗R I).
3) dimp (A) = dim(kR (p) ⊗R A).
4) f-dimR (A) = sup{dim(kA (p) ⊗R A) : p ∈ SpecA
e (R)}.
Next, given an R-algebra A, we give lower and upper bounds of the Krull dimension of A in terms of the Krull dimension of its fibre rings and the effective Krull
dimension of R with respect to A. Observe that the formulas given in the following theorem are reminiscent of Seidenberg’s inequalities for the Krull dimension of
polynomial rings.
Recall that a ring homomorphism f : A −→ B is said to satisfy the Going-Down
property (GD for short) if for any prime ideals p ⊆ q of A such that there exists
Q ∈ Spec(B) with Q ∩ A = q, then there exists P ∈ Spec(B) such that P ∩ A = p
and P ⊆ Q. It is then easy to see that if a ring homomorphism f : R −→ A satisfies
GD and if p ∈ SpecA
e (R), then any prime ideal q of R such that q ⊆ p is an effective
prime ideal of R with respect to A, and thus htA
e (p) = ht(p).
Theorem 2.9. Let R be a ring and let A be an R-algebra. Then
1) f-dimR (A) ≤ dim(A) ≤ f-dimR (A) + (1 + f-dimR (A))dimA
e (R).
2) If the homomorphism fA : R −→ A satisfies the Going-down property, then,
A
sup{ht(p)+dimp (A) : p ∈ SpecA
e (R)} ≤ dim(A) ≤ f-dimR (A)+(1+f-dimR (A))dime (R).
Proof. 1) It suffices to prove the second inequality. If either dimA
e (R) = +∞ or fdimR (A) = +∞, then we are done. Assume that dimA
(R)
<
+∞
and f-dimR (A) <
e
+∞. Let P0 ⊂ P1 ⊂ · · · ⊂ Pn be a chain of distinct prime ideals of A. Then the
corresponding chain of contractions p0 ⊂ p1 ⊂ · · · ⊂ pr is composed of effective
prime ideals of R with respect to A. Observe that the number of the Pi ’s lying over
a fixed prime pj is inferieur than or equal to the Krull dimension of the fibre ring
kR (pj ) ⊗R A plus one, that is, dimpj (A) + 1, and dimpj (A) ≤ f-dimR (A). Further,
as p0 ⊂ p1 ⊂ · · · ⊂ pr is a chain composed of 1 + r effective prime ideals of R with
respect to A and as r ≤ dimA
e (R), we get
n ≤ r(f-dimR (A) + 1) + f-dimR (A)
≤ f-dimR (A) + (f-dimR (A) + 1)dimA
e (R).
This yields the desired inequality.
2) Assume that fA satisfies GD. Let n := htA
e (p) and p0 ⊂ p1 ⊂ · · · ⊂ pn = p be
a chain of distinct effective primes of R with respect to A. Fixing P ∈ Specp (A)
and applying the Going-down property yields the existence of a chain P0 ⊂ P1 ⊂
10
Local dimension theory of algebras
· · · ⊂ Pn = P such that each Pi ∈ Specpi (A) for i = 0, 1, · · · , n − 1. Then htA
e (p) =
n ≤ ht(P ) for each P ∈ Specp (A). Let Q0 ⊂ Q1 ⊂ · · · ⊂ Qr be a chain of
Specp (A) such that r := dim(kR (p) ⊗R A) = dimp (A). Therefore, as by the first step
htA
e (p) = n ≤ ht(Q0 ), we get
htA
e (p) + dimp (A) ≤ ht(Q0 ) + r ≤ ht(Qr ) ≤ dim(A)
for each effective prime ideal p of R. Hence, by (1), as htA
e (p) = ht(p), we get the
desired inequalities completing the proof.
Next, we list various applications and consequences of Theorem 2.9. The first
result gives a condition for coincidence of the Krull dimension of A and its fiber
Krull dimension with respect to R.
Corollary 2.10. Let A be an R-algebra. If dimA
e (R) = 0, then
dim(A) = f-dimR (A).
We aim via the following corollaries to recover Seidenberg’s inequalities for polynomial rings. The next result might be termed Seidenberg’s inequalities for algebras
over an arbitrary ring.
Corollary 2.11. Let A be an R-algebra such that fA satisfies GD. Assume that
f-dimR (A) = dimm (A) for each m ∈ MaxA
e (R). Then
A
f-dimR (A) + dimA
e (R) ≤ dim(A) ≤ f-dimR (A) + (1 + f-dimR (A))dime (R).
Proof. Observe, by Theorem 2.9, that, in particular, sup{htA
e (m)+dimm (A) : m ∈
A
Maxe (R)} ≤ dim(A). Then, as f-dimR (A) = dimm (A) for each m ∈ MaxA
e (R),
A (R)}+f-dim (A) ≤ dim(A). Therefore
sup{htA
(m)
:
m
∈
Max
R
e
e
dimA
e (R)+f-dimR (A) ≤ dim(A) establishing the desired inequalities.
Next, we recover Seidenberg’s inequalities for polynomial rings.
Corollary 2.12. Let R be a ring and let X1 , X2 , · · · , Xn be indeterminates over R.
Then
n + dim(R) ≤ dim(R[X1 , X2 , · · · , Xn ]) ≤ n + (n + 1)dim(R).
Proof. Observe that the homomorphism R −→ R[X1 , X2 , · · · , Xn ] satisfies GD and
R[X ,··· ,Xn ]
R[X ,··· ,Xn ]
Spece 1
(R) = Spec(R), thus dime 1
(R) = dim(R). Also,
dimp (R[X1 , X2 , · · · , Xn ]) =
=
dim(kR (p) ⊗R R[X1 , X2 , · · · , Xn ])
dim(kR (p)[X1 , X2 , · · · , Xn ]) = n
for each prime ideal p of R. Then
f-dimR (R[X1 , X2 , · · · , Xn ]) = dimp (R[X1 , X2 , · · · , Xn ]) = n
for each prime ideal p of R. Now, Corollary 2.11 completes the proof.
Local dimension theory of algebras
3
11
Local and global Transcendence degree over an arbitrary ring
In this section we introduce the local and global transcendence degree of algebras
over an arbitrary ring.
Recall that if k is a field, then it is customary to denote by
n
A
o
t.d.(A : k) := sup t.d.
: k : p ∈ Spec(A)
p
=
sup{t.d.(kA (p) : k) : p ∈ Spec(A)}
the transcendence degree of a k-algebra A over k. This section aims at giving a
definition of the notion of the transcendence degree of an R-algebra A over R in
accordance with the field case.
In the same spirit of Definition 2.7 and in order to prepare the ground for the
general case of tensor products over an arbitrary ring, we next introduce a local
notion of the transcendence degree of an R-algebra A over R as well as the “general”
transcendence degree of A over R which turns out to be in total accordance with
the well known notion of the transcendence degree over a field.
Definition 3.1. Let R be a ring and A an R-algebra.
1) Let p ∈ SpecA
e (R). We define the transcendence degree at p of A over R to be the
the transcendence degree of the fibre ring kR (p) ⊗R A (over p) over the field kR (p),
that is,
t.d.p (A : R) := t.d. kR (p) ⊗R A : kR (p) .
2) We define the transcendence degree of A over R to be the supremum of the transcendence degrees of A over R at effective prime ideals p of R with respect to A, that
is,
A
t.d.(A : R) := sup{t.d.
(R)}
n p(A : R) : p ∈ Spece
o
=
sup t.d. kR (p) ⊗R A : kR (p) : p ∈ SpecA
e (R) .
We will next prove that the transcendence degree of an R-algebra A over R
depends on the endowing structure of algebra of A over R, namely on the ring
homomorphism fA .
The following proposition shows that the transcendence degree over a ring R
at an effective prime ideal of R shares all known properties of the transcendence
degree over a field k. Notice that, when P is a prime ideal of an R-algebra A and
A
R
and a subring of
which
p := fA−1 (P ), then fA induces an isomorphism between
p
P
R
A
means that
might be identified with a subring of .
p
P
Local dimension theory of algebras
Proposition 3.2. Let A be an R-algebra.
1) If P is a prime ideal of A and p := fA−1 (P ), then
A R
A
= t.d. kA (P ) : kR (p) .
: R = t.d.
:
t.d.p
P
P p
2) Let p ∈ SpecA
(R).
Then,
e
A
n
o
t.d.p (A : R) = sup t.d.p
: R : P ∈ Specp (A) .
P
3) If P is a prime ideal of A with p := fA−1 (P ), then
A
n
o
t.d.p (AP : R) = sup t.d.p
: R : Q ∈ Specp (A) such that Q ⊆ P
n
Q
o
= sup t.d. kA (Q) : kR (p) : Q ∈ Specp (A) with Q ⊆ P
4) t.d.(A : R) =
=
A
n
o
sup t.d.p
: R : p ∈ Spec(R) and P ∈ Specp (A)
o
n AP R
: p ∈ Spec(R) and P ∈ Specp (A) .
:
sup t.d.
P p
R
\ {0} for each prime ideal p of R.
p
1) Let P be a prime ideal of A and let p := fA−1 (P ). Then,
A
A
:R
= t.d. kR (p) ⊗R
: kR (p)
t.d.p
P
P
A/P
−1
= t.d. Sp
: kR (p) (Lemma 2.2(1))
p(A/P )
A
A
A
−1
: kR (p) as p
= fA (p) = {0}
= t.d. Sp
P
P
A P
R
,
:
= t.d.
P p
Proof. Let Sp =
as desired.
2) Let p ∈ SpecA
e (R). Then, by (1) and Lemma 2.2((2) and (5)),
t.d.p (A : R) := t.d. kR (p) ⊗R A : kR (p)
n
k (p) ⊗ A
o
R
R
=
sup t.d.
: kR (p) : P ∈ Specp (A)
o
n
kR (p) ⊗R P
A
: kR (p) : P ∈ Specp (A)
=
sup t.d. kR (p) ⊗R
A
n
P
o
=
sup t.d.p
: R : P ∈ Specp (A)
n
AP R
o
= sup t.d.
: P ∈ Specp (A) .
:
P p
12
Local dimension theory of algebras
13
3) Let P be a prime ideal of A and p := fA−1 (P ). Then, by (1) and (2),
n
A
o
R
P
t.d.p (AP : R) = sup t.d.
: Q ∈ Specp (A) such that Q ⊆ P
:
p
P
o
n
Q
A R
: Q ∈ Specp (A) such that Q ⊆ P
:
= sup t.d.
QA p
n
o
= sup t.d.p
: R : Q ∈ Specp (A) such that Q ⊆ P .
Q
4) It follows easily from (1) and (2) completing the proof.
It is notable that the introduced notion of transcendence degree of A over R
depends on the structure of R-algebra over A, namely, on the ring homomorphism
fA , as shown by the following simple example.
Example 3.3. Let k be a field and X an indeterminate over k. Let R = k[X] × k
and let A = k(X). Consider the following ring homomorphisms f1 , f2 : R −→ A such
that f1 (g(X), α) = g(X) and f2 (g(X), α) = α. These two homomorphisms define
two different R-algebra structures over A. Moreover, observe that f1−1 ((0)) = (0)×k
(A,f )
(A,f )
and f2−1 ((0)) = k[X] × (0). Then Spece 1 (R) = {(0) × k} and Spece 2 (R) =
{k[X] × (0)}. Hence, by Proposition 3.2(4),
k[X] × k
t.d.(A :f1 R) = t.d. k(X) :
(0) × k
= t.d.(k(X) : k[X]) = 0
while
4
k[X] × k
t.d.(A :f2 R) = t.d. k(X) :
k[X] × (0)
= t.d.(k(X) : k) = 1.
Effective spectrum with respect to tensor products
Let R be a ring and let A and B be R-algebras. First, it is worth to note that
the tensor product A ⊗R B over R might be trivial even if A and B are not so.
Of course, the interesting case is when A ⊗R B 6= {0} which makes it legitimate to
introduce the notion of a triplet of rings (R, A, B) consisting of a given ring R and
two R-algebras A and B such that A ⊗R B 6= {0}.
Let A and B be R-algebras. We denote by µA : A −→ A ⊗R B and µB : B −→
A ⊗R B the canonical algebra homomorphisms over A and B, respectively, such that
µA (a) = a ⊗R 1 and µB (b) = 1 ⊗R b for each a ∈ A and each b ∈ B. Observe that
the following diagram (D) is commutative:
14
Local dimension theory of algebras
A
fA
µA
ր
ց
fA⊗R B
−→
R
fB
A ⊗R B
µB
ց
ր
B
From this section onward, given ideals I, J, H of A, B and A ⊗R B, respectively, we
adopt the following notation for easiness: I ∩ R := fA−1 (I), J ∩ R := fB−1 (J) and
−1
H ∩ A := µ−1
A (H), H ∩ B := µB (H).
We begin by recording the following isomorphisms related to the fiber rings of
the tensor products over an arbitrary ring R.
Lemma 4.1. Let R be a ring. Let A and B be algebras over R. Then
B
A
⊗kR (p) Sp−1
kR (p)⊗R (A⊗R B) ∼
= (kR (p)⊗R A)⊗kR (p) (kR (p)⊗R B) ∼
= Sp−1
pA
pB
where Sp :=
R
\ {0}.
p
Proof. It is direct by Lemma 2.2(1).
Remark. Let R be a ring and A, B be two R-algebras. Let I be a prime ideal
of A. Notice that, by considering the ring homomorphism µA : A −→ A ⊗R B,
kA (I) ⊗R B ∼
= kA (I) ⊗A (A ⊗R B) stands for the fibre ring of A ⊗R B over I. Thus
f-dimA (A ⊗R B) = sup{dim(kA (I) ⊗R B) : I ∈ SpeceA⊗R B (A)}.
The next theorem examines the effective spectrum of a ring with respect to
tensor products.
Theorem 4.2. Let R be a ring. Let A and B be two R-algebras. Let p ∈ Spec(R),
I ∈ Spec(A) and J ∈ Spec(B). Then
B
1) SpeceA⊗R B (R) = SpecA
e (R)∩ Spece (R).
2) There exists a prime ideal P of A ⊗R B such that P ∩ A = I and P ∩ B = J if
and only if I ∩ R = J ∩ R.
3)
SpeceA⊗R B (A) = {I ∈ Spec(A) : I ∩ R ∈ SpecB
e (R)}
= {I ∈ Spec(A) : ∃J ∈ Spec(B) such that I ∩ R = J ∩ R}.
Local dimension theory of algebras
15
Proof. 1) Let p ∈ SpeceA⊗R B (R). Then there exists a prime ideal P of A ⊗R B
such that p = P ∩ R. Let I = P ∩ A and J = P ∩ B. Then, by the above
commutative diagram (D), I and J are prime ideals of A and B, respectively, such
B
that I ∩ R = J ∩ R = P ∩ R = p, that is, p ∈ SpecA
e (R)∩ Spece (R). Conversely,
A
B
assume that p ∈ Spece (R)∩ Spece (R). Hence, by Lemma 4.1, the fibre ring
kR (p) ⊗R (A ⊗R B) ∼
= (kR (p) ⊗R A) ⊗kR (p) (kR (p) ⊗R B) 6= {0}
as the fibre rings kR (p) ⊗R A and kR (p) ⊗R B are not trivial. It follows that p ∈
SpeceA⊗R B (R), as desired.
2) See [8, Corollaire 3.2.7.1.(i)].
3) First, let I ∈ Spec(A) such that I ∩ R ∈ SpecB
e (R). Then, there exists J ∈
Spec(B) such that I ∩ R = J ∩ R so that by (2), there exists P ∈ Spec(A ⊗R B)
R B (A). Conversely, let
such that P ∩ A = I (and P ∩ B = J). Thus I ∈ SpecA⊗
e
A⊗
B
I ∈ Spece R (A). Then there exists P ∈ Spec(A ⊗R B) such that P ∩ A = I, so
that, using the above commutative diagram (D), I ∩ R = P ∩ R = (P ∩ B) ∩ R ∈
SpecB
e (R) completing the proof.
The following corollary totally characterizes when two algebras A and B over a
ring R constitute a triplet (R, A, B) of rings.
Corollary 4.3. Let R be a ring and A, B be two R-algebras. Then the following
assertions are equivalent:
1) (R, A, B) is a triplet of rings;
B
2) SpecA
e (R)∩ Spece (R) 6= ∅;
3) There exists a prime ideal I of A and a prime ideal J of B such that I ∩ R =
J ∩ R.
Proof. 1) ⇔ 2) It suffices to observe that, by Proposition 2.4(1) and Theorem 4.2,
B
−1
(Spec(A ⊗R B)) = SpeceA⊗R B (R) = SpecA
fA⊗
e (R) ∩ Spece (R).
RB
2) ⇔ 3) It is direct.
It is easy to provide examples of nontrivial algebras over a ring R such that
A ⊗R B = {0} is trivial. But Corollary 4.3 characterizes when this tensor product
A ⊗R B is trivial by checking connections between the spectrum of the three components of this construction, namely Spec(R), Spec(A) and Spec(B). For instance,
given a ring R and two distinct prime ideals p and q of R, applying Corollary 4.3,
k (p)
k (q)
note that kR (p) ⊗R kR (q) = {0} since Spece R (R) = {p} while Spece R (R) = {q}
k (p)
k (q)
and thus Spece R (R)∩ Spece R (R) = ∅.
Local dimension theory of algebras
16
Corollary 4.4. Let (R, A, B) be a triplet of rings. Then
B
dimeA⊗R B (R) ≤ min(dimA
e (R), dime (R)).
A⊗R B (R) = 0.
B
In particular, if either dimA
e (R) = 0 or dime (R) = 0, then dime
R B (R) = SpecA (R)∩ SpecB (R).
Proof. It is straightforward from Theorem 4.2 as SpecA⊗
e
e
e
The next proposition allows us to give lower and upper bounds of the Krull
dimension of tensor products of algebras over a ring R in terms of the Krull dimension
of its fibre rings and the effective Krull dimension of its components.
Proposition 4.5. Let (R, A, B) be a triplet of rings. Then
1) f-dimR (A⊗R B) ≤ dim(A⊗R B) ≤ f-dimR (A⊗R B)+(1+f-dimR (A⊗R B))dimeA⊗R B (R).
2) f-dimA (A⊗R B) ≤ dim(A⊗R B) ≤ f-dimA (A⊗R B)+(1+f-dimA (A⊗R B))dimeA⊗R B (A).
Proof. It follows from Theorem 2.9(1).
In light of Proposition 4.5, when the effective Krull dimension of a component
of a tensor product with respect to this construction is zero, the Krull dimension of
the tensor product turns out to be its own fibre Krull dimension. This result will
allow us to explicit the Krull dimension of the tensor products involving the ahead
introduced notion of fibred AF-rings.
Corollary 4.6. Let (R, A, B) be a triplet of rings.
1) If dimeA⊗R B (R) = 0, then
dim(A ⊗R B) = f-dimR (A ⊗R B).
2) If dimeA⊗R B (A) = 0, then
dim(A ⊗R B) = f-dimA (A ⊗R B).
Proof. It follows easily from Proposition 4.5.
5
Fibred AF-rings
This section introduces and studies the notion of fibred AF-rings over an arbitrary
ring R. This new concept extends that of AF-ring over a field introduced by A.
Wadsworth in [14].
Local dimension theory of algebras
17
Definition 5.1. Let R be a ring.
1) Let A be an R-algebra and p ∈ SpecA
e (R). A is said to be a fibred AF-ring at p if
its fiber ring kR (p) ⊗R A is an AF-ring over kR (p).
2) An R-algebra A is said to be a fibred AF-ring over R if it is a fibred AF-ring at
each effective prime ideal p of R with respect to A, that is, if each nontrivial fibre
ring kR (p) ⊗R A is an AF-ring over kR (p).
3) Let (R, A, B) be a triplet of rings.
B
a) If p ∈ SpecA
e (R)∩ Spece (R) and A is a fibred AF-ring at p, then A is said to
be a B-fibred AF-ring at p.
b) A is said to be a B-fibred AF-ring over R if A is a B-fibred AF-ring at each
common effective prime ideal p of R with respect to A and B.
Remark. 1) From this definition, it is clear that the notion of a fibred AF-ring
extends that of an AF-ring introduced by Wadsworth as any algebra A over a field
k possesses only one fiber ring over k which is A itself.
2) Note that the notion of B-fibred AF-ring is inherent to the considered triplet of
rings (R, A, B) and it is easy to see that the following assertions are equivalent:
a) A is a fibred AF-ring over R;
c) A is a B-fibred AF-ring over R for any R-algebra B such that (R, A, B) is a
triplet;
c) A is an R-fibred AF-ring over R.
Let k be a field and A be a k-algebra. Recall that, for each prime ideal P of A,
A
: k ≤ t.d.(AP : k) and A is said to be an AF-ring if this inequality
ht(P )+t.d.
P
A
: k = t.d.(AP : k) for each
turns out to be an equality, that is, ht(P )+t.d.
P
prime ideal P of A. Next, we show that these properties translate into local data
for algebras over an arbitrary ring R.
Proposition 5.2. Let R be a ring and let p be a prime ideal of R. Let A be an
R-algebra.
1) If P is a prime ideal of A such that P ∩ R = p, then
A
: R ≤ t.d.p (AP : R).
htp (P ) + t.d.p
P
2) Let B be an R-algebra such that (R, A, B) is a triplet of rings. Assume that p ∈
A
B
SpecA
e (R) (resp., p ∈ Spece (R)∩ Spece (R)). Then A is a fibred AF-ring (resp.,
B-fibred AF-ring) at p if and only if
A
htp (P ) + t.d.p
: R = t.d.p (AP : R)
P
for each prime ideal P of A such that p = P ∩ R.
Local dimension theory of algebras
18
Proof. 1) Let P be a prime ideal of A such that P ∩ R = p. Then, by Lemma 2.2,
htp (P ) + t.d.p
A
P
:R
A
= ht(kR (p) ⊗R P ) + t.d. kR (p) ⊗R
: kR (p)
k (p) ⊗ P
R
RA
: kR (p)
= ht(kR (p) ⊗R P ) + t.d.
kR (p) ⊗R P
≤ t.d. (kR (p) ⊗R A)kR (p)⊗R P : kR (p)
= t.d. kR (p) ⊗R AP : kR (p)
=
t.d.p (AP : R),
as desired.
2) It is direct from Definition 5.1 taking into account the following equalities which
A
k (p) ⊗ A
R
R
figure in Lemma 2.2: t.d.p
: R = t.d.
: kR (p) and t.d.p (AP :
P
kR (p) ⊗R P
R) = t.d. kR (p) ⊗R A)kR (p)⊗R P : kR (p) for each prime ideal P of A with p :=
P ∩ R.
The following result exhibits various classes of fibred AF-rings.
Proposition 5.3. 1) If k is a field, then the fibred AF-rings over k are exactly the
AF-rings over k.
2) Any finitely generated R-algebra R[x1 , x2 , ..., xn ] is a fibred AF-ring over R.
3) Let (R, A, B) be a triplet of rings. If dimeA⊗R B (A) = 0, then A is a B-fibred
AF-ring over R.
4) Any zero-dimensional ring A which is an R-algebra is a fibred AF-ring over R.
Proof. 1) It is straightforward.
2) Let A = R[X1 , X2 , · · · , Xn ] be a polynomial ring in n-variables over R. Let
p ∈ Spec(R). Then kR (p) ⊗R A ∼
= kR (p)[X1 , X2 , ..., Xn ] is clearly an AF-ring over
kR (p) [14, Corollary 3.2]. Hence A is a fibred AF-ring over R. Now, let A =
R[x1 , x2 , · · · , xn ] be any finitely generated R-algebra and let p ∈ Spec(R). Then
R[X1 , X2 , · · · , Xn ]
for some ideal I of R[X1 , X2 , · · · , Xn ] and thus, by Lemma
A∼
=
I
Local dimension theory of algebras
19
2.2,
R[X1 , X2 , · · · , Xn ]
kR (p) ⊗R A ∼
= kR (p) ⊗R
I
R[X1 , X2 , · · · , Xn ]
∼
= (kR (p) ⊗R R[X1 , X2 , · · · , Xn ]) ⊗R[X1 ,X2 ,··· ,Xn ]
I
∼
=
kR (p)[X1 , X2 , · · · , Xn ]
:= kR (p)[y1 , y2 , · · · , yn ], where yi := Xi
IkR (p)[X1 , X2 , · · · , Xn ]
for each i = 1, 2, · · · , n,
is a finitely generated kR (p)-algebra which is an AF-ring over kR (p) by [14, page
395]. Hence A is a fibred AF-ring over R proving (2).
R B (A) = 0. Let
3) Let (R, A, B) be a triplet of rings such that dimA⊗
e
A⊗R B (R). Let P be a prime ideal of A such that
B
p ∈ SpecA
e (R)∩ Spece (R) = Spece
P ∩ R = p. Then, by Proposition 3.2,
o
n
A R
: Q ∈ Specp (A) with Q ⊆ P .
:
t.d.p (AP : R) = sup t.d.
Q p
R B (R), by
Let Q ⊆ P be a prime ideal of A with Q ∩ R = p. Then, as p ∈ SpecA⊗
e
A⊗
B
A⊗
B
Theorem 4.2(3), Q ∈ Spece R (A). Now, since dime R (A) = 0, we get Q = P .
Therefore, by Proposition 3.2,
A R
A
t.d.p (AP : R) = t.d.
= t.d.p
:
:R .
P p
P
P
R B (A) = 0, ht
Also, as dimA⊗
= 0 since any prime ideal Q such that pA ⊆
e
pA
Q ⊆ P is an effective prime ideal of A with respect to A ⊗R B (see Theorem 4.2(3)).
P
Then, by Corollary 2.8 , htp (P ) = ht
= 0. It follows that
pA
A
: R = t.d.p (AP : R).
htp (P ) + t.d.p
P
Then, by Proposition 5.2, A is a B-fibred AF-ring over R, as desired.
4) Note that if dim(A) = 0, then, in particular, dimeA⊗R B (A) = 0 for any triplet
(R, A, B) of rings. Hence, by (3), A is a B-fibred AF-ring over R for any R-algebra
B such that (R, A, B) is a triplet of rings. In particular for B = R, A is an R-fibred
AF-ring over R which means that A is a fibred AF-ring over R, as desired.
We next establish the stability of the fibred AF-ring notion under various type
of constructions.
Local dimension theory of algebras
20
Proposition 5.4. Let R be a ring and let A be an R-algebra. Let p be a prime ideal
of R.
1) If A is a fibred AF-ring at p and S is a multiplicative subset of A such that p ∈
−1
SpecSe A (R), then the localization S −1 A is a fibred AF-ring at p.
2) Let A1 , A2 , ..., An be fibred AF-rings at p. Then A1 ⊗R A2 ⊗R · · · ⊗R An is a fibred
AF-ring at p.
3) If A is a fibred AF-ring at p, then the polynomial ring A[X1 , X2 , ..., Xn ] is a fibred
AF-ring at p.
Proof. 1) Let A be a fibred AF-ring at p and S be a multiplicative subset of A such
that kR (p) ⊗R S −1 A 6= {0}. Then, by Proposition 2.4(5), there exists P ∈ Spec(A)
such that P ∩ S = ∅ and p = P ∩ R. Let S be the image of S via the homomorphism
P
A
. Observe that pA ∩ S = ∅ and
∩ S = ∅. Then, by Lemma 2.2,
A −→
pA
pA
S −1 A R
S −1 P
S −1 A
+
t.d.
= ht
:
htp (S −1 P ) + t.d.p −1 : R
−1
−1 P
S P
p
pS PA
S
A R
−1
= ht S
+ t.d.
:
p
A P
P pA
R
+ t.d.
:
= ht
pA
P p
A
:R
= htp (P ) + t.d.p
P
= t.d.p (AP : R)
as A is a fibred AF-ring at p. Moreover, note that t.d.p ((S −1 A)S −1 P : R) = t.d.p (AP :
R). It follows that
S −1 A
htp (S −1 P ) + t.d.p −1 : R = t.d.p ((S −1 A)S −1 P : R).
S P
Hence S −1 A is a fibred AF-ring at p.
2) Let p be a prime ideal of R. Then
kR (p) ⊗R (A1 ⊗R · · · ⊗R An ) ∼
= (kR (p) ⊗R A1 ) ⊗kR (p) (kR (p) ⊗R (A2 ⊗R · · · ⊗R An ))
∼
= (kR (p) ⊗R A1 ) ⊗kR (p) · · · ⊗kR (p) (kR (p) ⊗R An ).
First, as each kR (p) ⊗R Ai 6= {0}, kR (p) ⊗R (A1 ⊗R A2 ⊗R · · · ⊗R An ) 6= {0}.
Hence, since each kR (p) ⊗R Ai is an AF-ring over kR (p), we get, by [14, Proposition
3.1], (kR (p) ⊗R A1 ) ⊗kR (p) · · · ⊗kR (p) (kR (p) ⊗R An ) is an AF-ring over kR (p) so that
kR (p)⊗R (A1 ⊗R · · · ⊗R An ) is an AF-ring over kR (p). It follows that A1 ⊗R · · · ⊗R An
is a fibred AF-ring at p.
3) It follows easily from (2) as A[X1 , X2 , ..., Xn ] ∼
= R[X1 , X2 , ..., Xn ] ⊗R A and, by
Proposition 5.3(2), R[X1 , X2 , ..., Xn ] is a fibred AF-ring at p.
Local dimension theory of algebras
21
Corollary 5.5. Let R be a ring.
1) If A is a fibred AF-ring over R and S is a multiplicative subset of A, then the
localization S −1 A is a fibred AF-ring over R.
2) Let A1 , A2 , ..., An be fibred AF-rings over R. Then A1 ⊗R A2 ⊗R · · · ⊗R An is a
fibred AF-ring over R.
3) If A is a fibred AF-ring over R, then the polynomial ring A[X1 , X2 , ..., Xn ] is a
fibred AF-ring over R.
The following two corollaries give the B-fibred AF-ring versions of the above
Proposition 5.4 and Corollary 5.5. Their proofs are straightforward.
Corollary 5.6. Let (R, A, B) be a triplet of rings. Let p be a prime ideal of R.
1) If A is a B-fibred AF-ring at p and S is a multiplicative subset of A such that
−1
p ∈ SpecSe A (R), then the localization S −1 A is a B-fibred AF-ring at p.
2) Let A1 , A2 , ..., An be B-fibred AF-rings at p. Then A1 ⊗R A2 ⊗R · · · ⊗R An is a
B-fibred AF-ring at p.
3) If A is a B-fibred AF-ring at p, then the polynomial ring A[X1 , X2 , ..., Xn ] is a
B-fibred AF-ring at p.
Corollary 5.7. Let (R, A, B) be a triplet of rings.
1) If A is a B-fibred AF-ring over R and S is a multiplicative subset of A, then the
localization S −1 A is a B-fibred AF-ring over R.
2) Let A1 , A2 , ..., An be B-fibred AF-rings over R. Then A1 ⊗R A2 ⊗R · · · ⊗R An is
a B-fibred AF-ring over R.
3) If A is a B-fibred AF-ring over R, then the polynomial ring A[X1 , X2 , ..., Xn ] is
a B-fibred AF-ring over R.
6
Krull dimension of tensor products involving fibred
AF-rings
The goal of this section is to discuss and compute the Krull dimension of the tensor
product of algebras over R involving fibred AF-rings in various settings.
The following theorem allows to compute the Krull dimension of all fibre rings
of the tensor product of algebras A and B over a ring R in the case when A is a
fibred AF-ring over R. This result translates Wadsworth theorem [14, Theorem 3.7]
into the general setting of tensor products over an arbitrary ring R. We give the
next more general version of a triplet (R, A, B) of rings such that A is a B-fibred
AF-ring over an effective prime ideal p of R.
Local dimension theory of algebras
22
Notation. 1) Let A be a ring and P be a prime ideal of A. Let n ≥ 1 be a
positive integer. Then, for easiness of notation, we denote by A[n] the polynomial
ring in n indeterminates A[X1 , X2 , · · · , n] and by P [n] the extended prime ideal
P [X1 , X2 , · · · , Xn ] of A[X1 , X2 , · · · , Xn ].
2) Let A be an algebra over a field k. Let 0 ≤ d ≤ s be positive integers. Then, in
[14], Wadsworth adopted the following notation:
o
n
A
:k
: P ∈ Spec(A) .
D(s, d, A) := sup ht(P [s]) + min s, d + t.d.
P
3) Let A be an algebra over a ring R and p be a prime ideal of R. Let 0 ≤ d ≤ s be
positive integers. Then, we adopt the following notation for a local invariant of the
above D(s, d, A):
Dp (s, d, A) := D(s, d, kR (p) ⊗R A).
We begin by expliciting the local invariant Dp (s, d, A) in terms of the local invariants of the height and transcendence degree.
Lemma 6.1. Let R be a ring. Let A be an algebra over R and p be a prime ideal
of R. Let 0 ≤ d ≤ s be positive integers. Then
o
A
n
: R : P ∈ Specp (A) .
Dp (s, d, A) = sup htp (P [s]) + min s, d + t.d.p
P
It is worth noting that if A and B are algebras over a ring R and X1 , X2 , · · · , Xn
are indeterminates, then
(A ⊗R B)[X1 , X2 , · · · , Xn ] ∼
= A ⊗R B ⊗R R[X1 , X2 , · · · , Xn ]
∼
= A[X1 , X2 , · · · , Xn ] ⊗R B
∼
= A ⊗R B[X1 , X2 , · · · , Xn ].
Proof. Observe that, using Lemma 2.2 and Corollary 2.8,
Dp (s, d, A) = D(s, d, kR (p) ⊗R A)
n
k (p) ⊗ A
R
R
= sup ht(kR (p) ⊗R P )[s] + min s, d + t.d.
: kR (p) :
kR (p) ⊗R P
o
P ∈ Specp (A)
n
A
: kR (p) :
= sup ht(kR (p) ⊗R (P [s])) + min s, d + t.d. kR (p) ⊗R
P
o
P ∈ Specp (A)
A
n
o
= sup htp (P [s]) + min s, d + t.d.p
: R : P ∈ Specp (A) ,
P
as desired.
Local dimension theory of algebras
23
Recall that Wadsworth proved in [14] that, given a field k, if A is an AF-domain
and B is any k-algebra, then
dim(A ⊗k B) = D(t.d.(A), dim(A), B)[14, Theorem 3.7].
We generalized this result in [1] to AF-rings by proving that if A is an AF-ring and
B is any k-algebra, then,
n
o
dim(A⊗k B) = sup D t.d.(AP : k), dim(AP ), B : P ∈ Spec(A) [1, Theorem 1.4].
Our first main result gives a new version of the above-cited Wadsworth theorem
in the general setting of tensor products of algebras over an arbitrary ring R.
B
Theorem 6.2. Let (R, A, B) be a triplet of rings and let p ∈ SpecA
e (R)∩ Spece (R).
Assume that A is a B-fibred AF-ring at p. Then
o
n
dimp (A ⊗R B) = sup Dp t.d.p (AI : R), htp (I), B : I ∈ Specp (A) .
Proof. Observe that
kR (p) ⊗R (A ⊗R B) ∼
= (kR (p) ⊗R A) ⊗kR (p) (kR (p) ⊗R B)
and that kR (p) ⊗R A is an AF-ring over the field kR (p). Then, using [1, Theorem
1.4], we get, by Lemma 2.2(4),
dim(kR (p) ⊗R (A ⊗R B)) = dim (kR (p) ⊗R A) ⊗kR (p) (kR (p) ⊗R B)
n
= sup D t.d. (kR (p) ⊗R A)kR (p)⊗R I : kR (p) ,
o
ht(kR (p) ⊗R I), kR (p) ⊗R B : I ∈ Specp (A)
n
= sup D t.d.p (AI : R), htp (I), kR (p) ⊗R B :
o
I ∈ Specp (A)
n
= sup Dp t.d.p (AI : R), htp (I), B :
o
I ∈ Specp (A) ,
as desired.
The following corollaries compute the Krull dimension of tensor products involving algebras whose (effective) Krull dimension is zero. It is clear that if dim(A) = 0,
then for any nontrivial A-algebra C, dimC
e (A) = 0. Also, in Example 6.6, we
record the existence of various cases of triplets of rings (R, A, B) such that either
dimeA⊗R B (R) = 0 or dimeA⊗R B (A) = 0.
Local dimension theory of algebras
24
Corollary 6.3. Let (R, A, B) be a triplet of rings such that A is a B-fibred AF-ring
and dimeA⊗R B (R) = 0 (in particular, dim(R) = 0). Then,
n
B
dim(A ⊗R B) = sup Dp t.d.p (AI : R), htp (I), B : p ∈ SpecA
e (R) ∩ Spece (R) and
o
I ∈ Specp (A) .
R B (R) = 0, by Corollary 4.6(1),
Proof. As dimA⊗
e
dim(A ⊗R B) = f-dimR (A ⊗R B)
= sup{dimp (A ⊗R B) : p ∈ SpeceA⊗R B (R)}.
Then, Theorem 6.2 completes the proof.
Corollary 6.4. Let (R, A, B) be a triplet of rings such that A is a B-fibred AF-ring
B
and either dimA
e (R) = 0 or dime (R) = 0. Then,
n
B
dim(A ⊗R B) = sup Dp t.d.p (AI : R), htp (I), B : p ∈ SpecA
e (R) ∩ Spece (R) and
o
I ∈ Specp (A) .
Proof. It is straightforward by Corollary 4.4 and Corollary 6.3.
We devote the following theorem to the case where one component of a tensor
product is a zero-dimensional ring.
Theorem 6.5. Let (R, A, B) be a triplet of rings such that dimeA⊗R B (A) = 0 (in
particular, dim(A) = 0). Then
n
B
dim(A ⊗R B) = sup Dp t.d.p (kA (I) : R), 0, B : p ∈ SpecA
e (R) ∩ Spece (R) and
o
I ∈ Specp (A) .
R B (A) = 0, by Corollary 4.6(2),
Proof. As dimA⊗
e
dim(A ⊗R B) = f-dimA (A ⊗R B)
.
= sup{dim(kA (I) ⊗R B) : I ∈ SpeceA⊗R B (A)}
Let I ∈ SpeceA⊗R B (A) and p := I ∩ R. Note that, by Proposition 5.3, kA (I), being
R B (A), then p ∈
zero-dimensional, is a fibred AF-ring over R. Also, as I ∈ SpecA⊗
e
A⊗
B
R
Spece
(R). Therefore, Theorem 6.2 yields
dimp (kA (I) ⊗R B) = Dp (t.d.p (AI : R), htp (I), B).
25
Local dimension theory of algebras
R B (R), by Theorem 4.2(3), any J ∈ Spec (A) is an effective
Now, since p ∈ SpecA⊗
p
e
R B (A) = 0, we get
prime ideal of A with respect to A ⊗R B. Therefore, since dimA⊗
e
htp (I) = ht(
k (p)
I
) = 0.
pA
k (p)⊗ B
R
Furthermore, as dime A (R) = 0, we get, by Corollary 4.4, dime A
(R) = 0. It
follows, by Corollary 4.6(1) and as kA (I) ⊗R B possesses only one fibre ring which
is kR (p) ⊗R (kA (I) ⊗R B), that
dim(kA (I) ⊗R B) = f-dimR (kA (I) ⊗R B)
= dimp (kA (I) ⊗R B)
= Dp (t.d.p (AI : R), htp (I), B)
= Dp (t.d.p (AI : R), 0, B).
Consequently,
dim(A⊗R B) = sup{Dp (t.d.p (AI : R), 0, B) : p ∈ SpeceA⊗R B (R) and I ∈ Specp (A)}
R B (R) = SpecA (R)∩ SpecB (R).
completing the proof as, by Theorem 3.3, SpecA⊗
e
e
e
Example 6.6. 1) Let R := Z and A, B be rings such that (Z, A, B) is a triplet
of rings. Let char(A) = n and char(B) = m such that n 6= 0 and m 6= 0.
Then SpeceA⊗Z B (Z) = {pZ : p is a common prime divisor of n and m} and thus
Z B (Z) = 0.
dimA⊗
e
2) Let R := Z. Let n ≥ 1 be an integer. Let A = ZpZ +XQ[[X]] be a D+m construcZ
tion issued from the local ring Q[[X]] = Q + XQ[[X]]. Let B be a
-algebra and
nZ
let p be a prime divisor of n. Then, Spec(A) = {(0), XQ[[X]], pZpZ + XQ[[X]]} and
A⊗Z B (Z) = {pZ}
SpecB
e (Z) = {qZ : q is a positive prime divisor of n}. Therefore Spece
A⊗
B
A⊗
Z
and Spece
(A) = {pZpZ + XQ[[X]]}. It follows that dime Z B (Z) = 0 and
A⊗
B
dime Z (A) = 0.
Next, we deal with tensor products over the ring of integers Z. This allows us
to answer a question rised by Jorge Martinez on evaluating the Krull dimension of
the tensor product over Z of two rings one of which is a Boolean ring.
Corollary 6.7. Let (Z, A, B) be a triplet of rings such that char(A) =: n 6= 0. Assume that A is a B-fibred AF-ring over Z. Let n = pα1 1 pα2 2 ...pαr r be the decomposition
of n into prime factors. Then
o
n
dim(A⊗Z B) = sup Dpi t.d.pi (AI : Z), htpi (I), B : i = 1, 2, · · · , n and I ∈ Specpi (A) .
Local dimension theory of algebras
26
Z
is identified to a subring of A and that
Proof. Observe that, as char(A) = n,
nZ
o
Z np Z p Z
pj Z
pr Z
1
2
=
. Then, for each j = 1, 2, · · · , r,
,
,··· ,
is a minimal
Spec
nZ
nZ nZ
nZ
nZ
pj Z
Z
Z
prime ideal of
and thus there exists Ij ∈ Spec(A) such that Ij ∩
=
.
nZ
nZ
nZ
Hence, for each j = 1, 2, · · · , r, there exists Ij ∈ Spec(A) such that Ij ∩ Z = pj Z.
A
Therefore SpecA
e (Z) = {p1 Z, p2 Z, · · · , pr Z}). Thus dime (Z) = 0. Now, Corollary
6.4 completes the proof.
We close with the following corollary which presents an answer to a question of
Jorge Martinez on evaluating the Krull dimension of the tensor product over the
ring of integers Z of two rings one of which is Boolean.
First, we record the following well known characteristics of Boolean rings.
Lemma 6.8. Let R be a Boolean ring. Then
1) R is commutative.
2) char(R) = 2.
3) dim(R) = 0.
R∼ Z
4)
for each prime ideal p of R.
=
p
2Z
Z
R
is an algebraic field extension of
for each prime ideal p of R.
5)
p
2Z
Corollary 6.9. Let (Z, A, B) be a triplet of rings such that A is a Boolean ring.
Then
B
.
dim(A ⊗Z B) = dim
2B
Proof. Using the proof of Corollary 6.7, we get SpecA
e (Z) = {2Z}. Also, as dim(A) =
0, by Proposition 5.3, A is a fibred AF-ring over Z. Moreover,
n
A Z
o
t.d.2Z (A : Z) = sup t.d.
: I ∈ Spec(A) = 0
:
I 2Z
Z
A
is algebraic over
, by Lemma 6.8(5). It follows, by Theorem 6.5, Lemma
as
I
2Z
6.1 and Lemma 6.8, that
dim(A ⊗Z B) = = D2Z (0, 0, B)
o
n J
: J ∈ Spec2Z (B)
= sup ht
2B
B
= dim
2B
completing the proof.
Local dimension theory of algebras
27
References
[1] S. Bouchiba, F. Girolami, S. Kabbaj, The dimension of tensor products of AFrings, Commutative ring theory, 141-153, Lecture Notes in Pure and Appl.
Math., 185, Dekker, New York, 1997.
[2] S. Bouchiba, F. Girolami, and S. Kabbaj, The dimension of tensor products of
k-algebras arising from pullbacks, J. Pure Appl. Algebra 137 (1999), 125-138.
[3] S. Bouchiba, D.E. Dobbs and S. Kabbaj, On the prime ideal structure of tensor
products of algebras. J. Pure Appl. Algebra 176 (2002), no. 2-3, 89-112
[4] S. Bouchiba and S. Kabbaj, Tensor products of Cohen-Macaulay rings: Solution
to a problem of Grothendieck, J. Algebra 252 (2002) 65-73.
[5] S. Bouchiba, On Krull dimension of tensor products of algebras arising from
AF-domains, J. Pure Appl. Algebra 203 (2005) 237-251.
[6] S. Bouchiba, Chains of prime ideals in tensor products of algebras, J. Pure Appl.
Algebra 209 (2007), no. 3, 621-630.
[7] N. Bourbaki, Algèbre, Chap 1 à 3, Diffusion C.C.L.S, Paris (1970).
[8] A. Grothendieck; J. Dieudonné. Élements de géométrie algébrique, I.
Grundlehren Math. Wiss., 166. Springer-Verlag, Berlin, 1971.
[9] A. Grothendieck, Eléments de géométrie algébrique, Vol. IV, Institut des Hautes
Etudes Sci. Publ. Math. No. 24, Bures-sur-yvette, 1965.
[10] S. MacLane and O. F. G. Schilling, Zero-dimensional branches of rank one on
algebraic varieties, Ann. of Math. vol. 40 (1939), 507-520.
[11] H. Matsumura, Commutative Ring Theory, Cambridge University Press, 1986.
[12] R.Y. Sharp, The dimension of the tensor product of two field extensions, Bull.
London Math. Soc. 9 (1977), 42-48.
[13] R.Y. Sharp, Simplifications in the theory of tensor products of field extensions,
J. London Math. Soc. (2), 15(1) (1977), 48-50.
[14] A.R. Wadsworth, The Krull dimension of tensor products of commutative algebras over a field, J. London Math. Soc. 19 (1979), 391-401.
[15] O. Zariski and P. Samuel, Commutative Algebra Vol. I, Van Nostrand, Princeton, 1960.
| 0math.AC
|
arXiv:1712.08730v1 [cs.CV] 23 Dec 2017
Combining Weakly and Webly Supervised Learning
for Classifying Food Images
Parneet Kaur
Rutgers University
New Brunswick, NJ
[email protected]
Karan Sikka
SRI International
Princeton, NJ
[email protected]
Ajay Divakaran
SRI International
Princeton, NJ
[email protected]
Abstract
Food classification from images is a fine-grained classification problem. Manual
curation of food images is cost, time and scalability prohibitive. On the other
hand, web data is available freely but contains noise. In this paper, we address the
problem of classifying food images with minimal data curation. We also tackle a
key problems with food images from the web where they often have multiple cooccuring food types but are weakly labeled with a single label. We first demonstrate
that by sequentially adding a few manually curated samples to a larger uncurated
dataset from two web sources, the top-1 classification accuracy increases from
50.3% to 72.8%. To tackle the issue of weak labels, we augment the deep model
with Weakly Supervised learning (WSL) that results in an increase in performance
to 76.2%. Finally, we show some qualitative results to provide insights into the
performance improvements using the proposed ideas.
1
Introduction
Increasing use of smartphones has generated interest in developing tools for monitoring food intake
and trends [27, 34, 24]. Estimate of calorie intake can help users to modify their food habits to
maintain a healthy diet. Current food journaling applications like Fitbit App [1], MyFitnessPal [3] and
My Diet Coach [2] require users to enter their meal information manually. A study of 141 participants
in [12] reports that 25% of the participants stopped food journaling because of the effort involved while
16% stopped because they found it to be time consuming. Capturing images of meals is easier, faster
and convenient than manual data entry. An automated algorithm for measuring calories from images
should be able to solve several sub-problems − classify, segment and estimate 3D volume of the
given food items. In this paper we focus on the first task of classification of food items in still images.
This is a challenging task due to a large number of food categories, high intra-class variation and
low inter-class variation among different food classes. Further, in comparison to standard computer
vision problems such as object detection [21] and scene classification [36], present datasets for food
classification are limited in both quantity and quality to train deep networks (see section 2). Prior
works try to resolve this issue by collecting training data using human annotators or crowd-sourcing
platforms [14, 9, 17, 34, 24]. Such data curation is expensive and limits the scalability in terms of
number of training categories as well as number of training samples per category. Moreover, it is
challenging to label images for food classification tasks as they often have co-occurring food items,
partially occluded food items, and large variability in scale and viewpoints. Accurate annotation
of these images would require bounding boxes, making data curation even more time and cost
prohibitive. Thus, it is important to build food datasets with minimal data curation so that they can be
scaled for novel categories based on the final application.
Unlike data obtained by human supervision, web data is freely available in abundance but contains
different types of noise [10, 32, 29]. Web images collected via search engines may include images
of processed and packaged food items as well as ingredients required to prepare the food items as
Figure 1: Proposed pipeline for food image classification. We use inexpensive but noisy web data and
sequentially add manually curated data to the weakly supervised uncurated data. We also propose to
augment the deep model with Weakly Supervised learning (WSL) to tackle the cross-category noise
present in web images, and to identify discriminative regions to disambiguate between fine-grained
classes.
Figure 2: Noise in web data. Cross-domain Noise: Along with the images of specific food class,
web image search also include images of processed and packaged food items and their ingredients.
Cross-category Noise: An image may have multiple food items but it has only one label as its ground
truth, resulting in cross-domain noise.
shown in Figure 2. We refer to this noise as cross-domain noise as it is introduced by the bias due to
specific search engine and user tags. In addition, the web data may also include images with multiple
food items while being labeled for a single food category (cross-category noise). For example, in
images labeled as Guacamole, Nachos can be predominant (Figure 2). Further, the web results may
also include images not belonging to any particular class.
We address the problem of food image classification by combining webly and weakly supervised
learning ( Figure 1). We first propose to overcome the issues associated with obtaining clean training
data for food classification by using inexpensive but noisy web data. In particular we demonstrate
that by sequentially adding manually curated data to the uncurated data from web search engines,
the classification performance improves linearly. We show that by augmenting a smaller curated
dataset with larger uncurated web data the classification accuracy increases from 50.3% to 72.8%,
which is at par with the performance obtained with the manually curated dataset (63.3%). We also
propose to augment the deep model with weakly supervised learning (WSL) for for two reasons
(1) tackle the cross-category noise present in web images, and (2) identify discriminative regions to
disambiguate between fine-grained classes. We are able to approximately localize food items using
the activation maps provided by WSL. We show that by using WSL, the classification accuracy on
test data further increases to 76.2%. We finally show qualitative results and provide useful insights
into the two proposed strategies and discuss the reasons for performance improvements.
2
Related Work
Traditional computer vision feature vectors such has HOG, SIFT, bag-of-features, Gabor filters and
color histograms have been used for classifying food images in [34, 27, 8, 5, 16] Recent state-of-theart deep learning methods for food recognition and localization have led to significant improvement
in performance [23, 24, 22, 33, 28, 7]. However, the proposed methods use training data with only
one food item in the image [23] or have labels for multiple food items in images [17, 24]. The
preparation of training data requires manual curation. The Food-101 dataset [8] is often used for food
classification. It is collected from a food discovery website foodspotting.com and generally
contains less cross-domain noise as compared to images obtained from search engines such as
Google.com. However, this website relies on images sent by users and thus has limited images for
2
unique food categories, limiting expansion to new categories. In [31], food data is collected from
the web but also relies on textual information along with the images. CNNs have also been used to
classify food vs. non-food items in [28, 7]. In addition, [7] also provides food activation maps on the
input image to generate bounding boxes for localization. We address the problem of classifying food
items by using the noisy web data and incorporating weakly supervised learning for training CNNs.
Recent approaches of webly supervised learning in computer vision leverage from the noisy web data,
which is easy and inexpensive to collect. Prior work uses web data to train CNNs for classification
and object detection. [19] use noisy data collected from web for fine-grained classification. They also
use active learning-based approach for collecting data when only limited examples are available from
web. They demonstrate that even if the classification task at hand has small number of categories,
using a network trained with more categories gives better performance. Motivated by curriculum
learning, [10], propose an algorithm to first train a model on simple images from Google and estimate
a relationship graph between different classes. The confusion matrix is integrated with the model
and is fine-tuned on harder Flickr images. The confusion matrix makes the network robust to noise
and improves performance. Similarly, [26] modified the loss function by using the noise distribution
from the noisy images.
Food images often consist of multiple food items instead of a single food item and require bounding
boxes for annotation. To avoid expensive curation, weakly supervised learning (WSL) utilizes imagelevel labels instead of pixel-level labels or bounding boxes. In [35, 25], the network architecture
is modified to incorporate WSL by adding a global pooling layer. Along with image classification,
these architectures are able to localize the discriminative image parts. In [13], the authors include top
instances (most informative regions) and negative evidences(least informative regions) in the network
architecture to identify discriminative image parts more accurately. To address object detection,
the authors in [6] modify the deep network using a spatial pyramid pooling layer and use region
proposals to simultaneously select discriminative regions and perform classification. In [11], the
authors present a multi-fold multiple instance learning approach that detects object regions using
CNN and fisher vector features while avoiding convergence to local optima.
In this paper, we combine the webly and weakly supervised learning to address the problem of
food classification. We sequentially add curated data to the weakly labeled uncurated web data and
augment the deep model with WSL. We report improved performance as well as gain insights by
visualizing the qualitative results.
3
Approach
We first describe the datasets used to highlight the benefits of using uncurated data with manually
curated data for the task of food classification. Thereafter, we briefly discuss weakly supervised
learning to train the deep network.
3.1
Datasets
We first collect food images from the web and augmented it with both curated and additional uncurated
images, and test our method on a separate clean test set. The datasets are described below:
1. Food-101 [8]: This dataset consists of 101 food categories with 750 training and 250 test
images per category. The test data was manually cleaned by the authors whereas the training
data consists of cross-category noise i.e. images with multiple food items labeled with a
single class. We use the manually cleaned test data as the curated dataset (25k images),
Food-101-CUR, which is used to augment the web dataset. We use 10% of the uncurated
training data for validation and 90% of uncurated data (referred to as Food-101-UNCUR)
for data augmentation for training the deep model.
2. Food-Web-G: We collect the web data using Google image search for food categories from
Food-101 dataset [8]. The restrictions on public search results limited the collected data
to approximately 800 images per category. We removed images smaller than 256 pixels in
height or width from the dataset. As previously described, the web data is weakly labeled
and consists of both cross-domain and cross-category noise as shown in Figure 2. We refer
to this dataset as Food-Web-G
3
3. UEC256 [17]: This dataset consists of 256 food categories, including Japanese and international dishes and each category has at least 100 images with bounding box indicating
the location of its category label. Since this dataset provides the advantage of complete
bounding box level annotations, we use this dataset for testing. We construct the test set by
selecting 25 categories in common with the Food-101 dataset and extract cropped images
using the given bounding boxes.
3.2
Weakly Supervised Learning (WSL)
The data collected from web using food label tags is weakly labeled i.e. an image is labeled with a
single label when it contains multiple food objects. We observe that most uncurated food images
were unsegmented with images containing either items from co-occurring food classes or background
objects such as kitchenware. We propose to tackle this problem by augmenting the deep network
with WSL that explicitly grounds the discriminative parts of an image for the given training label
[35], resulting in a better model for classification.
As shown in Figure 1, we incorporate discriminative localization capabilities into the deep model
by adding a 1 × 1 convolution layer and a spatial pooling layer to a pretrained CNN [25, 35]. The
convolution layer generates N × N × K class-wise score maps from previous activations. The spatial
pooling layer in our architecture is a global average pooling layer which has recently been shown to
outperform the global max pooling step for localization in WSL [25, 35]. Max pooling only identifies
the most discriminative region by ignoring lower activations, while average pooling finds the extent
of the object by recognizing all discriminative regions and thus giving better localization. The spatial
pooling layer returns class-wise score for each image which are then used to compute cross-entropy
loss. During test phase, we visualize the heats maps for different classes by overlaying the predicted
score maps on the original image.
Additionally, food classification is a fine-grained classification problem [19] and we later show that
discriminative localization also aids in correctly classifying visually similar classes. Compared to
Krause et al. [19], who show the benefits of noisy data for fine-grained tasks such as bird classification,
we also highlight the benefits of WSL for learning with noisy data for food classification.
3.3
Implementation Details
We use Inception-Resnet [30] as the base architecture and fine-tune the weights of a pre-trained
network (ImageNet). During training, we use Adam optimizer with a learning rate 10−3 for the last
fully-connected (classification) layer and 10−4 for the pre-trained layers. We use batch size of 50.
For WSL, we initialize the network with the weights obtained by training the base model and only
fine-tune the layers added for weak localization with learning rate of 10−3 . For WSL we obtain
localized score maps for different classes by adding a 1 × 1 convolutional layer to map the input
feature maps into classification score maps [35]. For an input image of 299 × 299, we get a score
map of 8 × 8 from the output of this convolutional layer, which gives approximate localization when
resized to the size of input image. The average pooling layer is of size 8 × 8, stride 1 and padding 1.
4
4.1
Experiments
Quantitative Results
We report top-1 classification accuracy for different combinations of datasets (see section 3) and
WSL in Table 1. We first discuss the performance without WSL, where the baseline performance
using Google images (Food-Web-G) is 55.3%. We observe that augmenting Food-Web-G (66.9k
samples) with a small proportion of curated data (25k samples) improves the performance to 69.7%,
whereas augmentation with additional uncurated data (67.5k samples from foodspotting.com)
results in 70.1%. The performance of both combinations is higher compared to the curated data alone
(63.3%) clearly highlighting the performance benefits of using noisy web data. We also observe that
different sources of web images i.e. Google versus Foodspotting results in different performance
(55.3% versus 70.5% respectively) for similar number of training samples. As previously mentioned,
Foodspotting is crowdsourced by food enthusiasts, who often compete for ratings, and thus has
less cross-domain noise and better quality compared to Google images. By combining all the three
4
Dataset
Food-Web-G
Food-101-CUR
Food-101-UNCUR
Food-Web-G + Food-101-CUR
Food-Web-G + Food-101-UNCUR
Food-Web-UNCUR + Food-101-CUR
All datasets
No. of
images
66.9k
25k
67.5k
92.5k
134.4k
92.5k
159.3k
Type
w/o WSL
with WSL
N
C
N
N +C
N
N +C
N +C
55.3%
63.3%
70.5%
69.7%
70.1%
71.4%
72.8%
61.6%
64.0%
73.2%
73.0%
74.0%
75.1%
76.2%
Table 1: Classification accuracy for different combinations of datasets with and without Weakly
Supervised training. The number of images for each combination (k = 1000) and type of dataset (N :
noisy and C:clean) are also shown.
(a)
(b)
Figure 3: Classification Accuracy using (a) Inception Resnet, (b) Inception Resnet with localization
layer. As the curated data (Food-101 CUR) is added to the web data, the classification accuracy on
the test data (UEC256-test) increases. Increasing the web data results in further improvement. The
red line shows the baseline performance with individual datasets.
datasets, we observe a classification accuracy of 72.8%, which outperforms the performance obtained
by either curated and uncurated datasets alone.
We also wanted to study the variation in performance on using different proportions of clean and
unclean images. As shown in Figure 3, by sequentially adding manually curated data (Food-101
CUR) to the web data (Food-Web-G), the classification performance improves linearly from 50.3%
to 69.0%. By adding the uncurated data from foodspotting, it further increases to 72.8%. We also
observe significant improvements by adding discriminative localization to the deep model, where the
classification accuracy further increases to 76.2%. In particular we observe a consistent improvement
across all data splits by using WSL e.g. for the combination of both uncurated datasets from Google
and foodspotting, the hike in performance by using WSL is 4% absolute points. This performance
trend highlights the advantages of WSL in tackling the noise present in food images by implicitly
performing foreground segmentation and also focusing on the correct food item in the case when
multiple food items are present (cross-category noise).
4.2
Qualitative Results
We show the heat maps indicating the approximate localization of the top-1 predicted label for few
training images with multiple food items in Figure 4. We see that for some training images (Figure
4a) the network learns to correctly localize the correct food type among co-occurring food classes
e.g. it is able to identify rice in “fried rice” example. This ability could explain the reasons for
performance benefits especially when training data is not completely labeled. However, we also
observe that for frequently co-occurring food items, sometimes the network learns to localize multiple
food types together. As shown in Figure 4b), network learns “chicken” and “rice” as one category
because they co-occur in many training examples. The network also learns wrong food item for some
co-occurring food items. For example, Figure 4c shows some examples where the network learns
5
(a)
(b)
(c)
Figure 4: Heat maps showing approximate pixel-wise predicted probabilities obtained by Weakly
Supervised training for few training images. We show three cases where (a) the food items are
localized correctly, (b) network localizes frequently co-occurring food items due to weak labels for
training, and (c) network localizes a frequently co-occurring food item instead of the labeled food
item due to incomplete and noisy training data.
Figure 5: Test images that are misclassified without any localization but correctly classified with
weak localization. We also show the heat map, predicted label with the two approaches, and the true
label.
to recognize “sauce” instead of Gyoza. This is a drawback with standard WSL methods where the
algorithm generally tends to focus on the most discriminative part and overfits. We can overcome this
aspect by either leveraging additional clean training data or using recent advances in WSL [20, 18].
We show the heat maps for test images that are misclassified without localization but are correctly
classified with localization in Figure 5. Food classification is a fine-grained classification problem
and we can see that WSL helps by identifying discriminative parts for different food items. For e.g.,
the model grounds the noodle pieces in “miso soup” image in Figure 5 that makes it possible to
differentiate it from “chocolate cake” class, both of which are generally dark brown in color.
We observe that the properties of training data and quality of labeling influences the test performance.
There are unique ways of cooking a food item in different cuisines, resulting in variability in
appearance. The UEC256 test data mainly contains Japanese cuisines that may not be seen by the
network during training phase. We found that some test images are misclassified if their appearance
varies from the training images. Figure 6a shows an example of the category “omelette” that has high
variability for training and test data. We observe that the performance on test data is also influenced
by the weak/incomplete labeling of training data. For example, as shown in Figure 6b, the training
dataset contains these two categories: “french fries” and “fish and chips”. “Fish and chips” always
contains french fries, however this information is not used during the training phase resulting in high
confusion between these classes during testing.
Misclassification on test images also occurs due to the presence of multiple food items. Localization
heatmaps show that the network also focuses on the partially occluded food items in the images.
Figure 7 shows some examples where the test images with true label “french fries” are misclassified
because the network focuses on the other partial food items in the image. Even though the top-most
6
(a)
(b)
Figure 6: Misclassification in test data. (a) Examples for category ‘Omelette’ in training data (top
row) and test data (bottom row). Data distribution in training and test data differs because they are
collected from different sources, resulting in misclassification of some test images. (b) Examples
from training data (top row) and test data (bottom row). Inter-class similarity in training data causes
confusion and results in misclassification of test data.
Figure 7: The test images are misclassified in presence of occluded food items because the network
learns to localize co-occurring food items during training. For these images, the top-5 predicted
labels includes the ground truth.
predicted label corresponds to the partially occluded food item, the correct label is often found in the
top-5 predictions (top-5 accuracy is 90.8%).
We also generated bounding boxes from the heatmaps as shown in Figure 8 and will evaluate the
localization performance in the future.
5
Conclusion
In this paper, we leverage the freely available web data to address the problem of food classification.
By augmenting the abundantly available uncurated web data with limited manually curated dataset and
using weakly supervised learning, we achieve a classification accuracy of 76.2%. The performance
improves linearly as the amount of curated data for training is increased. We examine the localization
maps and observe that WSL aids the network by learning to approximately localize a food item
even in presence of multiple food items. Additionally, we examine some cases where discriminative
localization helps to disambiguate visually similar classes. Although we chose to focus on WSL
in this work, additional performance improvement can also be obtained by other complementary
approaches such as cost sensitive loss [10, 26] and domain adaptation [4].
7
Figure 8: UEC256 test images. [Top row] Heatmaps for the top-1 predicted class. [Bottom row]
Bounding boxes obtained using the heatmaps.
6
Acknowledgments
We thank Carter Brown, Ankan Bansal, Kilho Son and Anirban Roy for many helpful discussions.
References
[1] Fitbit app. https://www.fitbit.com/app. Accessed: 2017-11-14.
[2] My diet coach. https://play.google.com/store/apps/details?id=com.
dietcoacher.sos. Accessed: 2017-11-14.
[3] Myfitnesspal. https://www.myfitnesspal.com. Accessed: 2017-11-14.
[4] Alessandro Bergamo and Lorenzo Torresani. Exploiting weakly-labeled web images to improve object classification: a domain adaptation approach. In Advances in neural information
processing systems, pages 181–189, 2010.
[5] Vinay Bettadapura, Edison Thomaz, Aman Parnami, Gregory D Abowd, and Irfan Essa. Leveraging context to support automated food recognition in restaurants. In Applications of Computer
Vision (WACV), 2015 IEEE Winter Conference on, pages 580–587. IEEE, 2015.
[6] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2846–2854, 2016.
[7] M. Bolaños and P. Radeva. Simultaneous food localization and recognition. In 2016 23rd
International Conference on Pattern Recognition (ICPR), pages 3140–3145, Dec 2016.
[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101–mining discriminative
components with random forests. In European Conference on Computer Vision, pages 446–461.
Springer, 2014.
[9] Mei-Yun Chen, Yung-Hsiang Yang, Chia-Ju Ho, Shih-Han Wang, Shane-Ming Liu, Eugene
Chang, Che-Hua Yeh, and Ming Ouhyoung. Automatic chinese food identification and quantity
estimation. In SIGGRAPH Asia 2012 Technical Briefs, page 29. ACM, 2012.
[10] Xinlei Chen and Abhinav Gupta. Webly supervised learning of convolutional networks. In
Proceedings of the IEEE International Conference on Computer Vision, pages 1431–1439,
2015.
[11] Ramazan Gokberk Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object
localization with multi-fold multiple instance learning. IEEE transactions on pattern analysis
and machine intelligence, 39(1):189–203, 2017.
[12] Felicia Cordeiro, Elizabeth Bales, Erin Cherry, and James Fogarty. Rethinking the mobile food
journal: Exploring opportunities for lightweight photo-based capture. In Proceedings of the
33rd Annual ACM Conference on Human Factors in Computing Systems, pages 3207–3216.
ACM, 2015.
[13] Thibaut Durand, Nicolas Thome, and Matthieu Cord. Weldon: Weakly supervised learning
of deep convolutional neural networks. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pages 4743–4752, 2016.
8
[14] Giovanni Maria Farinella, Dario Allegra, Marco Moltisanti, Filippo Stanco, and Sebastiano
Battiato. Retrieval and classification of food images. Computers in biology and medicine,
77:23–39, 2016.
[15] Armand Joulin, Laurens van der Maaten, Allan Jabri, and Nicolas Vasilache. Learning visual
features from large weakly supervised data. In European Conference on Computer Vision, pages
67–84. Springer, 2016.
[16] Taichi Joutou and Keiji Yanai. A food image recognition system with multiple kernel learning.
In Image Processing (ICIP), 2009 16th IEEE International Conference on, pages 285–288.
IEEE, 2009.
[17] Yoshiyuki Kawano and Keiji Yanai. Automatic expansion of a food image dataset leveraging
existing categories with domain adaptation. In ECCV Workshops (3), pages 3–17, 2014.
[18] Dahun Kim, Donghyeon Cho, Donggeun Yoo, and In So Kweon. Two-phase learning for
weakly supervised object localization. In The IEEE International Conference on Computer
Vision (ICCV), Oct 2017.
[19] Jonathan Krause, Benjamin Sapp, Andrew Howard, Howard Zhou, Alexander Toshev, Tom
Duerig, James Philbin, and Li Fei-Fei. The unreasonable effectiveness of noisy data for finegrained recognition. In European Conference on Computer Vision, pages 301–320. Springer,
2016.
[20] Krishna Kumar Singh and Yong Jae Lee. Hide-and-seek: Forcing a network to be meticulous
for weakly-supervised object and action localization. In The IEEE International Conference on
Computer Vision (ICCV), Oct 2017.
[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr
Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European
conference on computer vision, pages 740–755. Springer, 2014.
[22] Chang Liu, Yu Cao, Yan Luo, Guanling Chen, Vinod Vokkarane, and Yunsheng Ma. Deepfood: Deep learning-based food image recognition for computer-aided dietary assessment. In
International Conference on Smart Homes and Health Telematics, pages 37–48. Springer, 2016.
[23] Renfeng Liu. Food recognition and detection with minimum supervision. 2016.
[24] Austin Meyers, Nick Johnston, Vivek Rathod, Anoop Korattikara, Alex Gorban, Nathan Silberman, Sergio Guadarrama, George Papandreou, Jonathan Huang, and Kevin P Murphy.
Im2calories: towards an automated mobile vision food diary. In Proceedings of the IEEE
International Conference on Computer Vision, pages 1233–1241, 2015.
[25] Maxime Oquab, Léon Bottou, Ivan Laptev, and Josef Sivic. Is object localization for free?weakly-supervised learning with convolutional neural networks. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pages 685–694, 2015.
[26] Giorgio Patrini, Alessandro Rozza, Aditya Menon, Richard Nock, and Lizhen Qu. Making neural
networks robust to label noise: a loss correction approach. arXiv preprint arXiv:1609.03683,
2016.
[27] Manika Puri, Zhiwei Zhu, Qian Yu, Ajay Divakaran, and Harpreet Sawhney. Recognition and
volume estimation of food intake using a mobile device. In Applications of Computer Vision
(WACV), 2009 Workshop on, pages 1–8. IEEE, 2009.
[28] Ashutosh Singla, Lin Yuan, and Touradj Ebrahimi. Food/non-food image classification and
food categorization using pre-trained googlenet model. In Proceedings of the 2nd International
Workshop on Multimedia Assisted Dietary Management, pages 3–11. ACM, 2016.
[29] Sainbayar Sukhbaatar and Rob Fergus. Learning from noisy labels with deep neural networks.
arXiv preprint arXiv:1406.2080, 2(3):4, 2014.
[30] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4,
inception-resnet and the impact of residual connections on learning. In AAAI, pages 4278–4284,
2017.
[31] Xin Wang, Devinder Kumar, Nicolas Thome, Matthieu Cord, and Frederic Precioso. Recipe
recognition with large multimodal food dataset. In Multimedia & Expo Workshops (ICMEW),
2015 IEEE International Conference on, pages 1–6. IEEE, 2015.
9
[32] Xin-Jing Wang, Lei Zhang, Xirong Li, and Wei-Ying Ma. Annotating images by mining image
search results. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(11):1919–
1932, 2008.
[33] Keiji Yanai and Yoshiyuki Kawano. Food image recognition using deep convolutional network
with pre-training and fine-tuning. In Multimedia & Expo Workshops (ICMEW), 2015 IEEE
International Conference on, pages 1–6. IEEE, 2015.
[34] Weiyu Zhang, Qian Yu, Behjat Siddiquie, Ajay Divakaran, and Harpreet Sawhney. “snap-n-eat”
food recognition and nutrition estimation on a smartphone. Journal of diabetes science and
technology, 9(3):525–533, 2015.
[35] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning
deep features for discriminative localization. In The IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), June 2016.
[36] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A
10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 2017.
10
| 1cs.CV
|
Submitted to Management Science
manuscript MS-17-01173
arXiv:1607.07306v2 [cs.GT] 20 Jun 2017
Authors are encouraged to submit new papers to INFORMS journals by means of
a style file template, which includes the journal title. However, use of a template
does not certify that the paper has been accepted for publication in the named journal. INFORMS journal templates are for the exclusive purpose of submitting to an
INFORMS journal and should not be used to distribute the papers in print or online
or to submit the papers to another publication.
The Costs and Benefits of Sharing: Sequential
Individual Rationality and Fairness
Ragavendran Gopalakrishnan
Conduent Labs India (formerly Xerox Research Centre India), [email protected]
Koyel Mukherjee
IBM Research India, [email protected]
Theja Tulabandhula
University of Illinois Chicago, [email protected]
In designing dynamic shared service systems that incentivize customers to opt for shared rather than exclusive service, the traditional notion of individual rationality may be insufficient, as a customer’s estimated
utility could fluctuate arbitrarily during their time in the shared system, as long as their realized utility at
service completion is not worse than that for exclusive service. In this work, within a model that explicitly considers the “inconvenience costs” incurred by customers due to sharing, we introduce the notion of
sequential individual rationality (SIR) that requires that the disutility of existing customers is nonincreasing
as the system state changes due to new customer arrivals. Next, under SIR, we observe that cost sharing
can also be viewed as benefit sharing, which inspires a natural definition of sequential fairness (SF)—the
total incremental benefit due to a new customer is shared among existing customers in proportion to the
incremental inconvenience suffered.
We demonstrate the effectiveness of these notions by applying them to a ridesharing system, where
unexpected detours to pick up subsequent passengers inconvenience the existing passengers. Imposing SIR
and SF reveals interesting and surprising results, including: (a) natural limits on the incremental detours
permissible, (b) exact characterization of “SIR-feasible” routes, which boast sublinear upper and lower
bounds on the fractional detours, (c) exact characterization of sequentially fair cost sharing schemes, which
includes a strong requirement that passengers must compensate each other for the detour inconveniences
that they cause, and (d) new algorithmic problems related to and motivated by SIR.
Key words : shared service system; ridesharing; cost sharing; sequential individual rationality; sequential
fairness; algorithmic mechanism design; graph algorithms
1
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
2
1.
Introduction
Sharing of resources and services is ubiquitous in today’s economy, driven by increasing costs
to the individual of enjoying exclusive access. For example, increased congestion on roads has
pushed more commuters towards public transportation and ridesharing (McKenzie 2015, Sivak
and Schoettle 2016). Funding shortages of governments and thin profit margins of businesses force
more people to share a smaller set of customer-serving resources both physically (contact centers (Aksin et al. 2007), airport security checks (Cole 2015, Zou et al. 2015)) and virtually (online
services (Armbrust et al. 2010)). The popularity of cloud computing services (e.g., Amazon Web
Services, Microsoft Azure) has increased because they drastically bring down the computing infrastructure costs (Kondo et al. 2009, Ahn et al. 2012). Even more examples include spectrum sharing
in wireless networks (Peha 2009), and shared logistics in supply chain distribution networks (Bowersox et al. 2000).
Of particular interest are shared service systems where arriving customers spend a finite amount
of time and money in the system getting served, and leave the system upon service completion.
However, because the service is shared, the time taken for a customer to be served, as well as
the monetary cost of service, can change depending upon the arrival/departure of other customers
into/from the system. Usually, the more the customers that share the same amount of resources, the
more time and less money each individual customer spends in the system. In ridesharing, picking up
an additional passenger involves a detour which increases passengers’ commute times, but brings
down their shares of the total cost. In a priority queue, there are multiple service levels; those
with higher priority (and hence, shorter average waiting times) cost more (Katta and Sethuraman
2005). Due to resource pooling in cloud computing services, the response times for the same load
can be higher if other tenants sharing the same set of resources place concurrent loads, and so,
offerings with such performance variability are priced lower (Jackson 2011).
When designing the pricing or cost sharing scheme for a shared service system, two major
factors that influence a customer’s choice of shared service must be considered: (a) Individual
Rationality (IR), where a customer compares her own utility between the shared and exclusive
service options, and asks herself if the cheaper monetary cost of the shared service is worth the
longer waiting/response time, and (b) Fairness, where a customer compares her utility with those
of other customers in the system, and evaluates whether everyone is “equitably” better off in
opting for the shared service. While traditional notions of IR and fairness capture the requirement
that customers who opt for a shared service are “happier” than they would be had they availed
an exclusive service, they fail to address a customer’s experience during the time they are in the
system due to subsequent arrivals (which could be unexpected), e.g., a sudden burst of arrivals
into a higher-priority queue could frustrate a customer waiting in a lower-priority queue at that
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
3
moment, even if she is later offered a discount for having waited longer. In ridesharing, a frequent
source of frustration are the (often unexpected) detours taken to pick up and/or drop off additional
passengers, which inconvenience existing passengers.
Our goal in this paper is to address this concern by defining appropriate new notions of IR and
fairness, and demonstrate their effectiveness by investigating their consequences (characteristic and
algorithmic) within the context of a specific application, namely ridesharing.
1.1.
Our Contributions
First, in Section 2, we introduce the concepts of sequential IR (SIR) and sequential fairness (SF),
that extend their traditional counterparts to be applicable at a finer granularity to a dynamic
shared service system. This involves invoking appropriate IR and fairness constraints every time the
state of the system changes due to a new customer arrival.1 We model the disutility of a customer
as the sum of the monetary cost of service and an “inconvenience cost” due to the presence of other
customers in the system. While IR simply requires the disutility when opting for the shared service
to be not greater than that for an exclusive service, SIR additionally requires the disutility to be
nonincreasing throughout the time spent in the shared system. The companion fairness concept,
SF, requires that the marginal decrease in disutility (the benefit of sharing) every time the state of
the system changes is “equitably” experienced by all the customers in the system. We also briefly
discuss a concern that these strong concepts could be too restrictive to allow feasible practical
policies in certain scenarios.
Next, we apply these concepts to a ridesharing system (motivated by its impact on promoting
sustainable behavior), where cost sharing among passengers is a major design component.2 SIR
targets the oft-lamented pain points due to detours experienced by passengers during the ride, by
ensuring that existing passengers are progressively better off every time an additional passenger is
picked up. In addition, SIR also ensures a certain degree of robustness, e.g., in a dynamic/online
setting, passengers would remain satisfied even if a future pickup is canceled. We show that imposing
SIR and SF on the routing and cost sharing schemes of a ridesharing system is not restrictive; to
the contrary, it brings out several interesting and surprising consequences:
(a) In Section 3.1, we provide an exact characterization (Theorem 1) for any route to be “SIRfeasible”, that is, there exists some budget-balanced cost sharing scheme that is SIR on that
route. These SIR-feasibility constraints, though fairly straightforward to derive, are necessarily
1
The concepts can be naturally extended to other situations which affect the customer experience, e.g.,
adding/removing service capacity, customer departures, etc.
2
Other elements are effective ride matching, routing, and, in the case of commercial ridesharing, profitability; we
explain how our focus on the costs influences these other factors, e.g., by adding the SIR and SF constraints to a
matching/routing algorithm that is already in use.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
4
complex, so in Section 4, we consider a simplified scenario where all the passengers are travelling
to a common destination. For this “single dropoff” scenario, we show that the SIR-feasibility
constraints simplify to natural upper bounds on the incremental detours, that keep shrinking
as the ride progresses towards the destination and as more passengers are picked up.
(b) In a series of theorems (Theorems 2-5) in Section 4.1, we show that the above bounds on
incremental detours can be aggregated to establish upper and lower bounds on the total detour
endured by a passenger as a fraction of their shortest distance to their destination. These
bounds depend on how sensitive the passengers are to detours; in realistic scenarios, these
bounds are sublinear in the number of passengers.
(c) In Section 5, we present an exact characterization of sequentially fair cost sharing schemes for
the single dropoff scenario (Theorem 6), which exposes several practical structural properties
of such schemes, including a surprisingly strong requirement that passengers must compensate
each other for the detour inconveniences that they cause.
(d) Finally, in Section 6, we explore some important algorithmic questions motivated by SIR. In
particular, it is unknown, even for the single dropoff scenario, whether there exists a polynomial
time algorithm to check for the existence of SIR-feasible routes, when restricted to an arbitrary
metric space (we show that it is NP-hard otherwise). Even if so, we show that optimizing for
total distance traveled over SIR-feasible routes is NP-hard (through a reduction from a variant
of Metric-TSP). We then consider a variant of the vehicle routing problem where passengers
are allocated to vehicles such that the total vehicle-miles traveled is minimized. While this
problem is known to be NP-hard in general, we show that it can be solved in polynomial time
given a fixed ordering on the pickup points.
1.2.
Related Work
The cost sharing problem for ridesharing has garnered relatively little attention in literature (compared to the ride matching and route optimization problems)—in most existing schemes, individual
passengers are asked to post what they are willing to pay in advance (Cao et al. 2015), share
the total cost proportionately according to the distances travelled (Geisberger et al. 2010, Agatz
et al. 2011), or negotiate their cost shares on their own during/after the ride. Such methods ignore
the real-time costs and delays incurred during the ride (as in the first instance), are insensitive
to the disproportionate delays encountered during the ride (as in the second instance), or lead to
a complicated and often uncomfortable negotiation process between possible strangers (as in the
third instance).
Recent work has studied cost sharing when passengers have significant autonomy in choosing
rides or forming ridesharing groups, e.g., cost sharing schemes based on the concept of kernel in
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
5
cooperative game theory (Bistaffa et al. 2015), second-price auction based solutions (Kleiner et al.
2011), and market based ride-matching models with deficit control (Zhao et al. 2014). Fair cost
sharing in ridesharing has also been studied under a mechanism design framework by Kamar and
Horvitz (2009), where an individually rational VCG-based payment scheme is modified to recover
budget-balance at the cost of incentive compatibility, and by Nguyen (2013), where customers are
offered an additive, detour-based discount, and the allocations and pricing are determined through
an auction. Our work differs from all the above in that we do not make any assumptions about
the mechanics of ride matching; our cost sharing model is independent of the routing framework
(static or dynamic), and is applicable to community carpooling and commercial ridesharing alike.
Moreover, we define a monotonic form of individual rationality for dynamic ridesharing.
Our work is different from the problem of pricing in ridesharing (see, e.g., Banerjee et al. (2015),
Bimpikis et al. (2016)); our focus is on sharing the resulting cost among the passengers.
Previous works on ridesharing that address individual rationality and detour limits treat them
as independent constraints, e.g., Kamar and Horvitz (2009), Santos and Xavier (2013), Pelzer
et al. (2015). In contrast, in our model, requiring (sequential) individual rationality induces natural
bounds on detours experienced by the ridesharing passengers.
There is a plethora of work when it comes to optimization problems in ridesharing (Agatz et al.
2012, Furuhata et al. 2013, Pelzer et al. 2015, Ozkan and Ward 2016, Alonso-Mora et al. 2017).
While the detour constraints that SIR induces can augment any routing optimization problem,
in this paper, we focus on finding an allocation of passengers to vehicles that minimizes the total
vehicle-miles, which is a variant of the vehicle routing problem (Cordeau et al. 2006), with the
additional constraint of SIR.
Variations of individual rationality involving temporal aspects are well studied in the economics
literature, e.g., ex-ante, interim, and ex-post individual rationality in mechanism design (Narahari
et al. 2009), and sequential individual rationality in bargaining and repeated games (Esteban 1991).
However, we are the first to explore its applicability to dynamic shared service systems and fairness
properties of the resulting outcomes.
An extensive literature on cooperative game theory and fair division (Moulin 2004, Jain and
Mahdian 2007) offers various cost sharing schemes that can be analyzed in our framework. Our
view of fairness relies on how the total incremental benefit due to ridesharing is allocated among the
passengers during each stage of the ride (sequential fairness). While we believe the two approaches
are not independent, exploring the connections is beyond the scope of this work.
We now outline some of the related work in cost sharing in a dynamic shared resource allocation setting, in which our contributions can potentially discover alternate solutions with different
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
6
desiderata. For instance, Elmachtoub and Levi (2014) study online allocation problems where customers arrive sequentially, and decisions regarding whether to accept or reject customers must be
made upon their arrival, with the goal of minimizing the sum of production costs (for accepted
customers) and rejection costs.
Additionally, the trade-off of (general forms of) fairness with efficiency for allocation problems
(in different domains, ranging from online advertisement portals (Bateni et al. 2016) to manufacturing and retail (Haitao Cui et al. 2007)) has been widely studied in the literature, and recently
conceptualized and studied as the “price of fairness” (Bertsimas et al. 2011, Dickerson et al. 2014,
Heydrich and van Stee 2015). We hope that the concepts of individual rationality and fairness that
our work introduces would inspire similar studies (theoretical and empirical) on their trade-offs
with different notions of efficiency such as social welfare maximization and profit optimization.
2.
Sequential Individual Rationality and Sequential Fairness
In this section, we formally define the notions of sequential individual rationality and sequential
fairness for shared service systems. We introduce the necessary notation first.
Let N = {1, 2, . . . , n} denote the set of customers, ordered according to their arrival times. Let
ti > 0 denote the time at which customer i ∈ N arrives into the system, and let T = {t0 , t1 , t2 , . . .},
where t0 = 0. Let `(i) denote the last customer to arrive into the system before i leaves the system.
Let su denote the system state at time u, which encodes the necessary information about all the
customers that are in the system at time u. Let Sij = {su | u ∈ T and ti ≤ u ≤ tj } denote the
information set of states from time ti to tj .
For any j ∈ N , let S(j) = {1, 2, . . . , j } denote the set of customers who have arrived until tj , and
let OC (S1j ) denote the operating cost of the system, conditional on there being no more arrivals
after tj . f denotes the cost sharing scheme according to which the operating cost is shared among
the customers. In particular, the monetary cost of service to customer i ∈ S(j), conditional on there
being no more arrivals after tj , is f (i, Sij ).
Definition 1. A cost sharing scheme f is budget balanced if
X
f (i, Sij ) = OC (S1j )
∀ j ∈ N.
(1)
i∈S(j)
For any j ∈ N , the inconvenience cost incurred by customer i ∈ S(j) due to all the other customers
she encounters in the system, assuming no customers arrive after time tj , is denoted by IC i (Sij ).
Several factors may affect the monetary and inconvenience costs, whose exact functional forms
would depend on the mechanics of the system being modeled.3
3
For example, one way to model the inconvenience cost could be to measure the additional time the customer spends
in the system compared to when she is the sole customer served, scaled by how much she values a unit of her time.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
2.1.
7
Disutility and Individual Rationality (IR)
The disutility of a customer i, assuming no customers arrive after time tj , is defined as the the
sum of their monetary and inconvenience costs, that is,
0
0≤j <i
DU i ,
DU i (tj ) = f (i, Sij ) + IC i (Sij ), i ≤ j ≤ `(i)
DU (t ),
j > `(i)
i `(i)
(2)
where DU 0i denotes the disutility corresponding to the exclusive service to customer i.
Definition 2. A cost sharing scheme f is Individually Rational (IR) if
DU i (tn ) ≤ DU i (t0 )
2.2.
∀ i ∈ N.
(3)
Sequential Individual Rationality (SIR)
While IR ensures that a customer’s disutility at the time of service completion in a shared system is
not greater than that of an exclusive service, it still allows for the disutility to fluctuate arbitrarily
during the time spent in the system, which can negatively affect the customer experience.
Definition 3. A cost sharing scheme f is Sequentially Individually Rational (SIR) if
DU i (tj ) ≤ DU i (tj−1 )
2.3.
∀ 1≤j≤n
∀ i ∈ N.
(4)
The Benefit of Sharing and Sequential Fairness
Under a cost sharing scheme that is IR, the decrease in disutility to a customer due to her participation in a shared service system (the difference between the right and left hand sides of her
IR constraint (3)) can be viewed as her benefit of sharing. Further, it can be seen that the total
benefit of sharing, obtained by summing the individual benefits, is independent of the cost sharing
scheme, as long as it is budget-balanced. This observation exposes an underlying “duality” – a cost
sharing scheme can, in fact, be viewed as a benefit sharing scheme. Such a view invites defining
cost sharing schemes based on traditional notions of fairness, e.g., a fair cost sharing scheme should
distribute the total benefit among the service-sharing customers suitably proportionately.
We extend this notion to budget balanced cost sharing schemes that are SIR by looking into
how they distribute the total incremental benefit due to each subsequent customer arriving into
the system, leading to a natural definition of sequential fairness.
Definition 4. The incremental benefit to customer i ∈ N due to the arrival of an incoming
customer j ∈ N is given by
IB i (Sij ) = DU i (tj−1 ) − DU i (tj ).
(5)
Definition 5. The total incremental benefit due to the arrival of an incoming customer j ∈ N
is given by
T IB (Sij ) =
X
i∈S(j)
IB i (Sij ).
(6)
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
8
We take a very general, but minimal, approach to defining sequential fairness. All that is required
of a cost sharing scheme to be sequentially fair is that, when an incoming customer j arrives into
the system, the portion of the total incremental benefit that is enjoyed by a previous customer
i (1 ≤ i ≤ j − 1), is proportional to the incremental inconvenience cost to i due to the incoming
customer. This is formalized in the following definition.
Definition 6. Given a vector β~ = (β2 , β3 , . . . , βn ), where 0 ≤ βj ≤ 1 for 2 ≤ j ≤ n, a budget
~
balanced, SIR cost sharing scheme f is β-sequentially
fair if, for all 2 ≤ j ≤ n,
βj P IC i (Sij )−IC i (Si(j−1) )
, 1≤i≤j −1
IB i (Sij )
j−1
m=1 (IC m (Smj )−IC m (Sm(j−1) ))
=
T IB (Sij ) 1 − βj ,
i = j.
(7)
Here, 1 − βj is the fraction of the total incremental benefit enjoyed by the incoming customer j
as a result of joining the service system, and βj , the remaining fraction, is split among the previous
IC (S
)
j jj
corresponds to a special case where the
(
)
incoming customer is treated just the same as everyone else.
customers. Setting 1 − βj =
2.4.
Pj
m=1 IC m (Smj )−IC m (Sm(j−1) )
Weaker Notions of SIR
There may be situations where SIR is so strong that no feasible practical policy can be expected
to satisfy it. For example, in shared service systems where the cost of service is fixed, e.g., priority
queues at airports, banks, and hospitals, SIR-compliant policies would require dynamically adding
service capacity to counter the inconvenience to low-priority customers every time a high-priority
customer arrives into the system. However, it may not be practical to do so each and every time such
an arrival occurs. Therefore, weaker notions of SIR could be proposed, e.g., approximate SIR, where
only a bounded increase in disutility is allowed (which would induce threshold staffing policies). In
addition, when arrivals and service times are modeled probabilistically, as in a queueing system,
appropriate probabilistic notions of SIR may be needed, since requiring SIR on every sample path
could be too restrictive. Future work should explore such interesting extensions.
3.
A Model for Cost Sharing in Ridesharing
In this section, we instantiate the above general model for a ridesharing system, where N denotes
the set of passengers. For each passenger i ∈ N , let Si and Di denote their pickup and dropoff
points, which are assumed to belong to an underlying metric space.
Additional Notation: We assume access to a routing algorithm that, given any subset S ⊆ N ,
computes a valid route rS (an ordered sequence of pickup/dropoff points) that serves all the
passengers in S. Thus, we define the following distance functions for any subset S ⊆ N : (a) d(S; rS )
denotes the total distance traveled along route rS , and (b) di (S; rS ), for i ∈ S, denotes the total
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
9
distance traveled along route rS from Si to Di . For simplicity, we assume that the costs are
completely determined by the traversed distances.4
Accordingly, the operational cost (or the “meter fare”), and the inconvenience cost are
OC (S; rS ) = αop d(S; rS ), and
(8)
IC i (S; rS ) = αi di (S; rS ) − di ({i}; r{i} ) ,
(9)
where αop > 0 is the price (in commercial ridesharing) or operating cost (in community carpooling)
per unit distance, and, for each i ∈ N , αi ≥ 0 is her inconvenience cost per unit distance, known as
her “detour sensitivity”. For simplicity, we denote di ({i}; r{i} ) by Si Di .
The cost sharing scheme f is such that, for any subset S ⊆ N , f (i, S; rS ) denotes the portion of
OC (S; rS ) allocated to passenger i ∈ S. We set f (i, S; rS ) = 0 whenever i ∈
/ S.
Disutility, IR and SIR: The definitions of disutility, IR, and SIR from equations (2)-(4) carry
over in a straightforward manner to the ridesharing scenario, with the dependence on the route
emphasized. Thus, DU i (tj ) is the disutility of passenger i along a route rN (tj ) which is identical
to rN up to time tj , but thereafter does not pick up any more passengers, proceeding only to drop
off the remaining passengers at their respective destinations. Also, the disutility corresponding to
exclusive service is DU 0i = f (i, {i}; r{i} ) = αop Si Di .
Definition 7. A route rN is IR-feasible (respectively, SIR-feasible) if there exists a budgetbalanced cost sharing scheme f that is IR (respectively, SIR) on rN .
From here on, whenever it is understood from context, we drop the explicit dependence on the
route to simplify notation. Next, we present an illustrative example.
Example 1. Consider n = 3 passengers, picked up from their sources S1 , S2 , S3 (in that
order),and travelling to a common destination D. The progression of the route rN (t), as the
passengers are picked up one by one, is depicted in Fig. 1. Given the final route rN , the total
distances traveled by passengers 1, 2 and 3 are d1 (N ) = S1 S2 + S2 S3 + S3 D, d2 (N ) = S2 S3 + S3 D
and d3 (N ) = S3 D. The total distance traveled is d(N ) = S1 S2 + S2 S3 + S3 D. The operational cost
is thus OC (N ) = αop (S1 S2 + S2 S3 + S3 D). Therefore, if f is a budget balanced cost sharing scheme,
f (1, N ) + f (2, N ) + f (3, N ) = αop (S1 S2 + S2 S3 + S3 D).
The inconvenience costs incurred by each passenger due to other passengers are:
IC 1 (N ) = α1 (S1 S2 + S2 S3 + S3 D − S1 D),
IC 2 (N ) = α2 (S2 S3 + S3 D − S2 D),
IC 3 (N ) = α3 (S3 D − S3 D) = 0.
4
It is straightforward to extend our model and results to costs that depend on a combination of distance and time.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
10
Figure 1
Route progress while picking up passengers traveling to a common destination.
Thus, a budget-balanced cost sharing scheme f is IR on route rN if
f (1, N ) + α1 (S1 S2 + S2 S3 + S3 D − S1 D) ≤ αop S1 D,
f (2, N ) + α2 (S2 S3 + S3 D − S2 D) ≤ αop S2 D,
f (3, N ) ≤ αop S3 D.
The SIR constraints are stronger, since they require IR at every stage of the ride:
f (1, N ) + α1 (S1 S2 + S2 S3 + S3 D − S1 D) ≤ f (1, N \ {3}) + α1 (S1 S2 + S2 D − S1 D) ≤ αop S1 D,
f (2, N ) + α2 (S2 S3 + S3 D − S2 D) ≤ f (2, N \ {3}) ≤ αop S2 D,
f (3, N ) ≤ αop S3 D.
A necessary condition for the route to be SIR-feasible is therefore obtained by summing up these
inequalities (at each stage), using budget-balance of f , and simplifying:
S2 S3 + S3 D − S2 D ≤
αop
S3 D
αop + α1 + α2
and S1 S2 + S2 D − S1 D ≤
αop
S2 D.
αop + α1
These “triangle inequalities” can be interpreted as imposing upper bounds on the incremental
detours at every stage of the ride. We discuss this in more detail in Section 4.
3.1.
Characterizing SIR-Feasibile Routes
The intuition gained from Example 1 suggests that routes with “large” detours are unlikely to
be SIR-feasible, that is, no budget-balanced cost sharing scheme would be SIR on such routes.
Theorem 1 provides a formal characterization of SIR-feasible routes.5
Theorem 1. A route rN is SIR-feasible if and only if
αop
j−1
X
d(S(j)) − d(S(j − 1)) +
αi di (S(j)) − di (S(j − 1))
(10)
i=1
≤ αop d({j }) − αj dj (S(j)) − d({j }) ,
5
∀ 2 ≤ j ≤ n.
Such a characterization would be useful to augment a routing algorithm in suggesting SIR-feasible routes (when
grouping ridesharing requests and assigning them to vehicles).
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
11
The proof is from expanding SIR constraints and algebraic manipulation (Appendix A.1).
Note that the SIR-feasibility equation (10) only guarantees that there exists a budget balanced
cost sharing scheme that is SIR on route rN . To check whether a specific cost sharing scheme is
SIR, we would have to go back to the individual constraints (4).
The recursive nature of the SIR-feasibility equation (10) makes it particularly easy to be incorporated into practical routing algorithms that involve sequential decision making, especially in
dynamic ridesharing (see Section 6).
4.
The Single Dropoff Scenario
The previous section illustrates the complexity of the most general case, where the route rN consists
of multiple pickup and dropoff points. Unfortunately, this complexity makes it difficult to infer a
useful interpretation of the SIR-feasibility constraints (10). Thus, in this section, we consider the
special case where all passengers 1 ≤ j ≤ n are traveling to a common destination Dj = D, which
exposes an interesting property of SIR, namely, that SIR translates to “natural” bounds on the
incremental detours. We begin this section by simplifying the general expressions introduced in
Section 3 to the single dropoff scenario.
We denote the distance between any two locations A and B in the underlying metric space by
AB. Recall that S(j) = {1, 2, . . . , j }, and that we hide the explicit dependence on the routes to
simplify notation. Thus, the distance functions become
d(S(j)) =
j−1
X
Sk Sk+1 + Sj D,
1 ≤ j ≤ n,
and di (S(j)) =
k=1
j−1
X
Sk Sk+1 + Sj D,
1 ≤ i ≤ j ≤ n.
k=i
(11)
The cost functions then become
OC (S(j)) = αop d(S(j)) = αop
j−1
X
!
Sk Sk+1 + Sj D ,
1 ≤ j ≤ n.
k=1
IC i (S(j)) = αi (di (S(j)) − di ({i})) = αi
j−1
X
!
Sk Sk+1 + Sj D − Si D ,
1 ≤ i ≤ j ≤ n.
k=i
The disutilities are given by
DU i (S(j)) = f (i, S(j)) + αi
j−1
X
!
Sk Sk+1 + Sj D − Si D ,
1 ≤ i ≤ j ≤ n.
k=i
The IR constraints for any budget-balanced cost sharing scheme simplify to
!
n−1
X
f (i, N ) + αi
Sk Sk+1 + Sn D − Si D ≤ f (i, {i}) = αop Si D, 1 ≤ i ≤ n.
k=i
(12)
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
12
The SIR constraints (18) and (19) for a budget-balanced cost sharing scheme simplify to
f (i, S(j)) − f (i, S(j − 1)) + αi (Sj−1 Sj + Sj D − Sj−1 D) ≤ 0, 1 ≤ i < j ≤ n.
f (j, S(j)) ≤ αop Sj D, 2 ≤ j ≤ n.
Finally, the SIR-feasibility constraints (10) from Theorem 1 simplify to
Sj−1 Sj + Sj D − Sj−1 D ≤
1+
Sj D
Pj−1
1
αop
k=1 αk
,
2 ≤ j ≤ n.
(13)
Notice that when restricted to the single dropoff scenario, the SIR-feasibility constraints assume a
much simpler form. For each j, the constraint has terms involving only j and j − 1. This “Markovian” nature could prove useful when studying algorithmic problems relating to SIR (see Section 6).
Upon closer inspection, we note that the left hand side of the SIR-feasibility constraints (13) are
nothing but the incremental detours due to picking up subsequent passengers j. Thus, (13) can be
viewed as imposing an upper bound on the permissible incremental detour involved in picking up
passenger j. This bound diminishes with increasing j and increasing proximity to the destination,
which means that as more passengers are picked up, the permissible additional detour to pick up yet
another passenger keeps shrinking, which is natural. For the passengers in Example 1, Fig. 2 shows
the evolution of the “SIR-feasible region” (points from which the next passenger can be picked up
so that the resultant route is SIR-feasible) in Euclidean space, when αj = αop for j = 1, 2, 3. The
shape resembles that of a rotated teardrop.
Figure 2
Evolution of the SIR-feasible region (dark shade) while picking up passengers that are traveling to a
common destination. Note that the region diminishes rapidly with every subsequent pickup.
4.1.
Bounds on Total Distance Traveled along SIR-Feasible Routes
The bounds on incremental detours given by the SIR-feasibility constraints (13) can be combined
to obtain bounds on the total distance traveled by a passenger i ∈ N along any SIR-feasible route,
as a fraction of their direct travel distance Si D. We call this measure the “starvation factor” of
passenger i. The starvation factor of a route is the maximum starvation factor among all the
passengers. Intuitively, the starvation factor of a route is a decreasing function of the ratios
αk
,
αop
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
13
since the permissible detours are, from (13). That is, passengers that are more sensitive to detours
should suffer smaller starvation factors. Our goal in this section is to quantify this intuition.
Let I (n) denote the space of all single dropoff instances of size n (consisting of n pickup points
and a common dropoff point from an underlying metric space). Given an instance p ∈ I (n), let
R(p) denote the set of all SIR-feasible routes for this instance.
;r)
Given an SIR-feasible route r ∈ R(p), let γr (i) = diS(N
denote the starvation factor of passenger
iD
Pn−1
i along route r, where di (N ; r) = k=i Sk Sk+1 + Sn D, from (11), and let γr = maxi∈N γr (i) denote
the starvation factor of the route r.
Definition 8. The SIR-starvation factor over all single dropoff instances of size n is
γ(n) = max min γr .
p∈I(n) r∈R(p)
We show the following bounds for γ(n):
1. Upper Bounds: (Theorems 2-4) The worst starvation factor among SIR-feasible routes,
√
maxp∈I(n) maxr∈R(p) γr , is (i) Θ(2n ) when ααopi → 0, (ii) Θ( n) when ααopi = 1, and (iii) 1 when
αi
αop
→ ∞, for all i ∈ N . As upper bounds for γ(n), these are not necessarily tight, since an
instance for which an SIR-feasible route has the worst starvation factor may also admit other
SIR-feasible routes with smaller starvation factors.
2. Lower Bounds: (Theorem 5) γ(n) is no smaller than (i) Θ(n) when
when
αi
αop
αi
αop
→ 0, and (ii) Θ(log n)
= 1, for all i ∈ N . These lower bounds are tight.
It is interesting to note that the gap between the upper and lower bounds narrows down and
vanishes as
αi
αop
increases to ∞.6 The proofs are complex, and can be found in Appendix A.
We begin by establishing an almost obvious result that when passengers are infinitely inconvenienced by even the smallest of detours,7 the only SIR-feasible routes (indeed, even IR-feasible
routes) are those with zero detours, which implies a starvation factor of 1.
Theorem 2. If
αi
αop
→ ∞ for all i ∈ N , then γr = 1 for any SIR-feasible route r.
Next, we consider passengers who value their time more than αop , and show that the worst they
√
would have to endure is a sublinear starvation factor, in particular, Θ( n). This is tight, that
√
is, there exists an SIR-feasible route with Θ( n) starvation factor, when αi = αop for all i ∈ N .
√
However, as the αi keep increasing beyond αop , this bound becomes looser, culminating in a Θ( n)
gap when αi → ∞, as evidenced by Theorem 2.
Theorem 3. If
αi
αop
√
≥ 1 for all i ∈ N , then γr ≤ 2 n for any SIR-feasible route r.
6
Note that, by definition, 1 is always a trivial lower bound for the starvation factor of any route, since the points are
from an underlying metric space.
7
Frankly, why would such passengers even consider ridesharing?
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
14
Even though it may be unrealistic, as an academic exercise, we investigate an upper bound on
γr when the passengers are completely unaffected by detours, that is,
αi
αop
→ 0 for all i ∈ N . Not
surprisingly, it turns out that the starvation factor can be exponentially large in such a scenario,
as the next theorem shows.
Theorem 4. If
αi
αop
→ 0 for all i ∈ N , then γr ≤ 2n for any SIR-feasible route r.
The upper bounds of Theorems 3-4 on γr are tight, as discussed next; however, by Definition 8,
they also serve as upper bounds on γ(n), in which capacity, they may not necessarily be tight. This is
because, an instance for which an SIR-feasible route has the worst starvation factor may also admit
better SIR-feasible routes. For example, Fig. 3 depicts an instance in one-dimensional Euclidean
space for which the route (S1 , S2 , . . . , Sn , D) is SIR-feasible (satisfying (21) with equality) and has a
√
starvation factor of Θ( n). (The same instance with the distances appropriately modified illustrates
the Θ(2n ) starvation factor of Theorem 4.) However, note that the reverse route (Sn , Sn−1 , . . . , S1 , D)
is also SIR-feasible and has a starvation factor of 1.
Figure 3
√
Single dropoff instance with a route (S1 , S2 , . . . , Sn , D) whose starvation factor is Θ( n). If the
distances Si D, 1 ≤ i ≤ n, were 2i−1 ` instead, then the starvation factor of the same route would be Θ(2n ).
Finally, we establish a tight lower bound on γ(n) for arbitrary αi > 0, by exhibiting an instance
with a unique SIR-feasible route with the desired starvation factor.
Pn
Pj−1 −1
Theorem 5. γ(n) ≥ j=1 1 + α1op k=1 αk
.
It is easy to observe that the lower bound of Theorem 5 simplifies to Θ(log n) when
Θ(n) when
5.
αi
αop
αi
αop
= 1, and
→ 0, for all i ∈ N .
The Benefit of Ridesharing and Sequential Fairness
In this section, we explore the consequences of sequential fairness (defined in Section 2.3) on the
design of cost sharing schemes for ridesharing, in the single dropoff scenario. First, for 2 ≤ j ≤ n,
1 ≤ i ≤ j, the expression for incremental benefit to passenger i due to the addition of passenger j
is
(
f (i, S(j − 1)) − f (i, S(j)) − αi (Sj−1 Sj + Sj D − Sj−1 D) , 1 ≤ i < j
IB i (S(j)) =
αop Sj D − f (j, S(j)),
i = j.
(14)
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
15
Thus, for 2 ≤ j ≤ n, the total incremental benefit due to the addition of passenger j is given by
!
j
j−1
X
X
T IB (S(j)) =
IB k (S(j)) = αop Sj D − αop +
αk (Sj−1 Sj + Sj D − Sj−1 D) ,
(15)
k=1
k=1
where the dependence on f vanishes due to budget-balance.
For the single dropoff scenario, the incremental inconvenience cost to i due to the detour caused
by j is given by IC i (Sij ) − IC i (Si(j−1) ) = αi (Sj−1 Sj + Sj D − Sj−1 D); hence, Definition 6 simplifies
to the following equivalent definition.
Definition 9. Given a vector β~ = (β2 , β3 , . . . , βn ), where 0 ≤ βj ≤ 1 for 2 ≤ j ≤ n, a budget
~
balanced cost sharing scheme f is β-sequentially
fair if, on any SIR-feasible route,
(
αi
βj Pj−1
, 1≤i≤j −1
IB i (S(j))
αm
m=1
=
(∀ 2 ≤ j ≤ n)
T IB (S(j))
1 − βj ,
i = j.
Note that 1 − βj denotes the fraction of the total incremental benefit enjoyed by the new passenger
j as a result of having them join the ride, and βj denotes the remaining fraction, which is split
among the existing passengers in proportion to their αi values.
It turns out that the requirements imposed by Definition 9, while perhaps appearing to be quite
lenient, are sufficient for a strong and meaningful characterization of sequentially fair cost sharing
schemes, as we discuss next.
5.1.
Characterizing Sequentially Fair Cost Sharing Schemes
We begin this section with a theorem that provides an exact characterization of budget balanced
sequentially fair cost sharing schemes for single dropoff scenarios.
Theorem 6. Given a vector β~ = (β2 , β3 , . . . , βn ), where 0 ≤ βj ≤ 1 for 2 ≤ j ≤ n, a budget~
balanced cost sharing scheme f is β-sequentially
fair if and only if, for 2 ≤ j ≤ n,
• The cost to the incoming passenger j is given by
"
f (j, S(j)) = βj αop Sj D + (1 − βj )
αop +
j−1
X
!
#
αm (Sj−1 Sj + Sj D − Sj−1 D) .
(16)
m=1
• The incremental “discount” to each existing passenger 1 ≤ i ≤ j − 1 is given by
"
#
αi
f (i, S(j − 1)) − f (i, S(j)) = βj Pj−1
m=1
αm
(αop Sj D − αop (Sj−1 Sj + Sj D − Sj−1 D))
(17)
+ (1 − βj ) αi (Sj−1 Sj + Sj D − Sj−1 D) .
We omit the proof, since it is simply a straightforward substitution of equations (14)-(15) in
Definition 9 and rearrangement of the terms. The characterization of Theorem 6 reveals elegant
structural properties of sequentially fair cost sharing schemes:
16
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
(a) Online Implementability. When a passenger j is picked up, their estimated cost is given
by f (j, S(j)), which is their final payment if there are no more pickups. At the same time,
each existing passenger i is offered a “discount” in the amount of f (i, S(j − 1)) − f (i, S(j))
that brings down their earlier cost estimates. This suggests a novel “reverse-meter” design for
a ridesharing application on each passenger’s smartphone that keeps track of their estimated
final payment, as the ride progresses. Starting with f (i, S(i)) when passenger i begins their
ride, it would keep decreasing every time a detour begins to pick up a new passenger. Such a
visually compelling interface would encourage wider adoption of ridesharing.
(b) Convex Combination of Extreme Schemes. For each j, 2 ≤ j ≤ n, the cost sharing scheme
is a convex combination of the following two extreme schemes:
• The total incremental benefit is fully enjoyed by the incoming passenger j, i.e., βj =
0. Here, from (16)-(17), the incoming passenger j (a) pays the service provider an amount
αop (Sj−1 Sj + Sj D − Sj−1 D) that corresponds to the increase in the operational cost, and
(b) pays each existing passenger 1 ≤ i ≤ j − 1 an amount αi (Sj−1 Sj + Sj D − Sj−1 D) that
corresponds to the incremental inconvenience cost they suffered.
• The total incremental benefit is fully enjoyed by the existing passengers 1 ≤ i ≤ j − 1, i.e.,
βj = 1. Here, from (16)-(17), the incoming passenger j pays αop Sj D, the same as they would
have paid for a private ride. From this, the service provider recovers αop (Sj−1 Sj + Sj D − Sj−1 D)
that corresponds to the increase in the operational cost, and what is left is split among the
existing passengers proportional to their αi values.
Note that the incoming passenger j pays the least in the former scheme (βj = 0) and the most
in the latter scheme (βj = 1).
(c) Transfers Between Passengers. From the previous observation, it follows that incoming
passengers must, at minimum, fully compensate existing passengers for the incremental inconvenience costs that resulted from the detour to pick them up, which can be viewed as internal
transfers between passengers. Even though it may be reasonable to expect this from a fair cost
sharing scheme, it is remarkable that sequential fairness mandates this property.
In designing a sequentially fair cost sharing scheme, β~ can be chosen strategically to incentivize
commuters to rideshare. A commonly used incentive is to guarantee a minimum discount on the
cost of a private ride. In our framework of sequentially fair cost sharing schemes, it corresponds to
setting βj so that f (j, S(j)) is a desired fraction of αop Sj D.8 We end this section with an example.
8
The SIR-feasibility constraints would have to be appropriately tightened to guarantee such a discount.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
17
Example 2. Consider the single dropoff scenario, where we also assume that αi = αop = 1 for
all i ∈ N . For 1 ≤ i ≤ j ≤ n, we define the cost sharing scheme f XC as:
j
X
Sk−1 Sk Sj D
f (i, S(j)) =
+
k−1
j
k=i+1
XC
!
+ (i − 1) (Si−1 Si + Si D − Si−1 D) −
j
X
!
(Sk−1 Sk + Sk D − Sk−1 D) ,
k=i+1
The first terms correspond to dividing the operational cost of each segment equally among the
ridesharing passengers traveling along that segment. The second terms correspond to the passenger
i compensating each of the i − 1 passengers that were picked up earlier, for the incremental detour
they suffered. The last terms correspond to the net compensation received by passenger i from all
passengers that were picked up later, for the incremental detours that i suffered.
Intuitively, f XC is a “fair” cost sharing scheme. In fact, it can be shown that for β~ =
~
it is a β-sequentially
fair cost sharing scheme:
1 1
, , . . . , n1
2 3
,
From (16)-(17), we get
IB j (S(j))
Sj D − f XC (j, S(j))
=
T IB(S(j)) Sj D − j (Sj−1 Sj + Sj D − Sj−1 D)
S D
Sj D − jj + (j − 1) (Sj−1 Si + Sj D − Sj−1 D)
j −1
1
=
=
=1− ,
Sj D − j (Sj−1 Sj + Sj D − Sj−1 D)
j
j
as desired. Also, for 1 ≤ i ≤ j − 1, we get
IB i (S(j))
f XC (i, S(j − 1)) − f XC (i, S(j)) − (Sj−1 Sj + Sj D − Sj−1 D)
=
T IB(S(j))
S D − j (Sj−1 Sj + Sj D − Sj−1 D)
j
Sj−1 Sj
Sj−1 D
S D
− j−1 + jj + (Sj−1 Sj + Sj D − Sj−1 D) − (Sj−1 Sj + Sj D − Sj−1 D)
j−1
=
Sj D − j (Sj−1 Sj + Sj D − Sj−1 D)
=
6.
Sj D
j (j−1)
− j−1 1 (Sj−1 Sj + Sj D − Sj−1 D)
Sj D − j (Sj−1 Sj + Sj D − Sj−1 D)
=
1 1
.
j j −1
New Algorithmic Problems
As discussed in Section 3.1, the SIR-feasibility constraints (10) or (13), can be considered as
additional constraints to the routing optimization problem. For instance, vehicle routing problems
with various operational objectives, ridesharing with multiple pickups and dropoff points, online
routing problems can all benefit from incorporating SIR-feasibility constraints while performing
route optimization. As a concrete example, consider the following single dropoff ride matching and
routing problem:
Given n pickup points and a common dropoff point in a metric space, (a) does there exist an
n
allocation of pickup points to 1 ≤ m ≤ n vehicles, each with capacity d m
e ≤ c ≤ n, such that there
exists an SIR-feasible route for each vehicle? And (b) if so, what is the allocation and corresponding
routes that minimize the total “vehicle-miles” traveled?
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
18
We do not know whether the feasibility problem (a) can be solved in polynomial time, even when
m = 1 and αi = αj for all 1 ≤ i, j ≤ n, where it reduces to finding a sequence of the pickup points
that satisfies the inequalities (13). The “Markovian” nature of these inequalities (each inequality
only depends on adjacent pickup points in the route) suggests that it may be worth trying to
come up with a polynomial time algorithm for the feasibility problem. In Section 6.1, we show that
this problem is NP-hard when not restricted to a metric space, which implies that any poly-time
algorithm, if one exists, must necessarily exploit the properties of a metric space. However, even if
one succeeds in this endeavor, we show in Section 6.2 that the optimization (b) over all SIR-feasible
routes is NP-hard.
Like SIR-feasibility, there might be other constraints on the ordering of the pickup points (for
instance, due to hard requirements on pickup times). Studying such variants might help understand
how to tackle SIR-feasibility constraints. For example, it is known that finding the optimal allocation (minimizing the total vehicle-miles traveled) of passengers to vehicles without any restriction
on the order of pickups is NP-hard Cordeau et al. (2006). On the other hand, as we show in
Section 6.3, the problem is polynomial time solvable if a strict total ordering is imposed and the
capacity of each vehicle is unrestricted. It then becomes an interesting future direction to investigate
what kinds of order constraints retain polynomial time solvability of the problem.
6.1.
Determining Existence of SIR-Feasible Routes is Hard
In this section, we present Theorem 7, which shows that, for the single-dropoff scenario, determining
whether an SIR-feasible route exists is NP-hard in general, by a reduction from the undirected
Hamiltonian path problem.9 The proof is deferred to Appendix B.
Definition 10. Given a set N of n pickup points, and a common dropoff point in an underlying
(possibly non-metric) space, and positive coefficients αop , α1 , α2 , . . . , αn , SIR-Feasibility is the
problem of determining whether an SIR-feasible route of length n exists, that is, whether there
exists a sequence of the pickup points that satisfies the SIR-feasibility constraints (13).
Theorem 7. SIR-Feasibility is NP-hard.
However, it can be easily seen that SIR-Feasibility is not hard in certain special cases and
in certain metric spaces. Consider an input graph, where the pickup points and the dropoff point
are embedded on a line, and αi = αop for all i ∈ N . Without loss of generality, we assume that the
pickup points {S1 , . . . , Sn } appear in the same order on the line, so that S1 and Sn are the two
end points. Clearly, if the destination D occurs before S1 (respectively, after Sn ), the instance is
9
Given an undirected graph, a Hamiltonian path is a path in the graph that visits each vertex exactly once. The
undirected Hamiltonian path problem is to determine, given an undirected graph, whether a Hamiltonian path exists.
It is known to be NP-hard.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
19
SIR-feasible. This is because the route starting from Sn (respectively, S1 ) and ending at D, visiting
all the pickup points along the way incurs zero detour for everyone, and is thus SIR-feasible. In fact,
such a route also traverses the minimum distance among all feasible routes. However, consider the
case where D is located at some intermediate location. Such an instance will never be SIR-feasible.
To see this, first consider an instance where n = 2, and S1 < D < S2 . Let S1 D = x, S2 D = y; hence
S1 S2 = x + y. We analyze the SIR-feasiblity constraints (13) for each of two cases. If S1 is visited
before S2 , then SIR-feasibility requires that x + y + y − x ≤ y2 , which is impossible. Similarly, if S2
is visited before S1 , then SIR-feasibility requires that x + y + x − y ≤ x2 , which is also impossible.
Now, when n > 2 and D is located at an intermediate point, any feasible route must, at some
point, “jump over” D from some Si to another Sj , at which stage the analysis would be the same
as that for n = 2, and is therefore not SIR-feasible. A similar phenomenon can be observed when
the underlying metric is a tree rooted at D and the pickup points are located at the leaves, and
αi = αop for all i ∈ N . It can be shown that instances where the pickup points are spread across
more than one subtree rooted at D cannot be SIR-feasible, and when the pickup points are all part
of a single subtree rooted at D, SIR-feasibility can be checked in polynomial time. We leave open
the problem of determining whether SIR-Feasibility is hard in general metric spaces.
6.2.
Optimizing over SIR-Feasible Routes is Hard
Given an undirected weighted graph, the problem of determining an optimal Hamiltonian cycle10
(one that minimizes the sum of the weights of its edges) is a well known problem called the Traveling
Salesperson Problem, abbreviated as TSP. A slight variant of this problem, known as Path-TSP, is
when the traveling salesperson is not necessarily required to return to the starting point or depot, in
which case we only seek an optimal Hamiltonian path. These problems are NP-hard (Papadimitriou
1994). Special cases of the above problems arise when the graph is complete and the edge weights
correspond to distances between vertices from a metric space. These variants, which we call MetricTSP and Metric-Path-TSP, respectively, are also NP-hard, e.g., Papadimitriou (1977) showed the
hardness for the Euclidean metric.
Definition 11. Given a set N of n pickup points, a common dropoff point in an underlying
metric space, and positive coefficients αop , α1 , α2 , . . . , αn , Opt-SIR-Route is the problem of finding
an SIR-feasible route of length n of minimum total distance.
Theorem 8. Opt-SIR-Route is NP-hard.
The proof is via a reduction from Metric-Path-TSP ; see Appendix B.
10
A Hamiltonian cycle is a Hamiltonian path that is a cycle. In other words, it is a cycle in the graph that visits each
vertex exactly once.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
20
6.3.
Optimal Allocation of Totally Ordered Passengers to Uncapacitated Vehicles
In this section, we present a polynomial time algorithm for optimal allocation of passengers to
vehicles (minimizing the total vehicle-miles traveled), given a total order on the pickups, and when
the capacity of any vehicle is unrestricted. To the best of our knowledge, this result is new; see Prins
et al. (2014) for a survey on related problem variants.
Our result relies on reducing the allocation problem to a minimum cost flow problem on a flow
network with integral capacities. We are given the set N of passengers (that is, the set of n ordered
pickup locations) traveling to a single dropoff location D. Without loss of generality, we let the
indices in N reflect the position in the pickup order, that is, u ∈ N is the u-th pick up from location
Su . For convenience, we index the destination D as n + 1. Let the unknown optimal assignment use
1 ≤ m0 ≤ n vehicles (we address how to find it later). A directed acyclic flow network (see Figure 4)
is then constructed as follows:
(1) s and t denote the source and sink vertices, respectively.
(2) For each passenger/pickup location u ∈ N , we create two vertices and an edge: an entry vertex
uin , an exit vertex uout , and an edge of cost 0 and capacity 1 directed from uin to uout . We also
create a vertex n + 1 corresponding to the dropoff location.
(3) We create n edges, one each of cost 0 and capacity 1 from the source vertex s to each of the
entry vertices uin , u ∈ N .
(4) We create n edges, one each of cost Su D and capacity 1 from each of the exit vertices uout ,
u ∈ N , to the dropoff vertex n + 1.
(5) To encode the pickup order, for each 1 ≤ u < v ≤ n we create an edge of cost (Su Sv − L)
and capacity 1 directed from uout to vin , where L is a sufficiently large number satisfying
L > 2 maxu,v∈N ∪{n+1} Su Sv .
(6) We add a final edge of cost 0 and capacity m0 from the dropoff vertex n + 1 to the sink vertex
t, thereby limiting the maximum flow in the network to m0 units.
Since all the edge capacities are integral, the integrality theorem guarantees an integral minimum
cost maximum flow, and we assume access to a poly-time algorithm to compute it in a network
with possibly negative costs on edges. Notice that we do have negative edge costs (step (5) of the
above construction); however, our network is a directed acyclic graph, owing to the fact that there
is a total ordering on the pickup locations. Hence, there are no negative cost cycles.
We defer the full proof to Appendix B.3; however, we briefly outline the steps involved:
• Any integral maximum flow from s to t must be comprised of m0 vertex-disjoint paths between
the source vertex s and the dropoff vertex n + 1.
• Any integral minimum cost flow must cover all the 2n pickup vertices, that is, a unit of flow
enters every entry vertex uin , and a unit of flow exits each exit vertex uout , u ∈ N .
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
Figure 4
21
Illustration of the directed acyclic flow network, a minimum cost maximum flow on which corresponds
to an assignment of n totally ordered passengers to m0 uncapacitated vehicles. Each of the edge labels
correspond to a tuple consisting of edge cost and edge capacity.
• The partition of N according to the m0 vertex-disjoint paths between s and n + 1 in an
integral minimum cost maximum flow corresponds to the optimal allocation of the n totally ordered
passengers among m0 uncapacitated vehicles.
Finally, we argue that the overall optimal assignment can be obtained by computing the optimal
assignments using the above reduction for each 1 ≤ m0 ≤ n and choosing the one with the overall
minimum cost, which completes the reduction.
7.
Concluding Remarks
In addition to the discussion in Section 2.4, and the open algorithmic questions raised in the
previous section, there are a few other important aspects that we believe future work should
address. We conclude the paper with brief discussions of these issues.
Throughout, we have assumed knowledge of the passengers’ αi values; but in reality, they are
most likely private information, especially in commercial ridesharing. One can either attempt to
learn these values over time from passenger feedback, or, one can ask the passenger for this information. In the latter case, truthful reporting is a concern. Since our framework explicitly takes into
account the inconvenience costs of passengers while considering SIR-feasibility as well as sequential
fairness, no sequentially fair cost sharing scheme can be dominant strategy incentive compatible. It
would be interesting to study the trade-offs between efficiency, fairness, budget-balance, and incentive compatibility. The best framework in which to study these questions (including in more general
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
22
settings where there is uncertainty about future pickup requests) is perhaps online mechanism
design (Parkes 2007, Zhao et al. 2015, Shen et al. 2016).
The duality between cost sharing and benefit sharing in our framework is worth a deeper analysis.
While the space of cost sharing schemes that the two views accommodate are no different from each
other, there is a crucial difference in approaching their design. In particular, notice that a budget
balanced cost sharing scheme need only recover the operational costs; see (1). The inconvenience
costs experienced by the passengers are a separate artifact of our framework, which only explicitly
affect the design of cost sharing schemes when viewed through the lens of benefit sharing and
sequential fairness. What “traditional” fairness properties does a sequentially fair cost sharing
scheme possess? Under what conditions, if any, is it a (generalized) Shapley value, or is in the core?
Appendix A:
A.1.
Proofs from Section 4
Proof of Theorem 1
The proof follows from expanding the SIR constraints (4). First, for any 2 ≤ j ≤ n, and 1 ≤ i ≤ j − 1, the SIR
constraint can be expanded as
f (i, S(j)) + IC i (S(j)) ≤ f (i, S(j − 1)) + IC i (S(j − 1))
=⇒ f (i, S(j)) − f (i, S(j − 1)) + IC i (S(j)) − IC i (S(j − 1)) ≤ 0.
(18)
For i = j, the SIR constraint can be expanded as
f (j, S(j)) + IC j (S(j)) ≤ f (j, {j})
(19)
=⇒ f (j, S(j)) ≤ αop d({j}) − IC j (S(j)),
where, it follows from budget-balance that f (j, {j}) = αop d({j}).
The “only if” direction can be seen to hold by adding all the j inequalities given by (18) and (19), for all
2 ≤ j ≤ n.
j
X
i=1
f (i, S(j)) −
j−1
X
j−1
X
f (i, S(j − 1)) +
IC i (S(j)) − IC i (S(j − 1)) ≤ αop d({j}) − IC j (S(j)).
i=1
i=1
Using the budget-balance property (1) to simplify the first two terms, we get
j−1
X
OC(S(j)) − OC(S(j − 1)) +
IC i (S(j)) − IC i (S(j − 1)) ≤ αop d({j}) − IC j (S(j)).
(20)
i=1
Equation (10) then follows by substituting for OC(·) and IC i (·) from (8) and (9) respectively, and simplifying.
Next, we prove the “if” direction. Assuming that (10) holds, or, alternatively, assuming that (20)
holds, it suffices to exhibit a budget-balanced cost sharing scheme f , under which all the SIR constraints
given by (18) and (19) are satisfied. For 1 ≤ j ≤ n, and 1 ≤ i ≤ j, we construct f (i, S(j)) recursively, so
that (18) and (19) are satisfied. The base case follows from budget-balance, that is, f (i, {i}) = αop d({i}) for
all i ∈ N . Assume that for some 2 ≤ j ≤ n, we have defined f (i, S(j − 1)) for all 1 ≤ i ≤ j − 1. Then, we set
f (i, S(j)) = f (i, S(j − 1)) − IC i (S(j)) − IC i (S(j − 1)) , 1 ≤ i ≤ j − 1
f (j, S(j)) = OC(S(j)) −
j−1
X
i=1
f (i, S(j)).
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
23
By construction, it follows that (18) is satisfied, and f is budget-balanced. It remains to be shown that (19)
is also satisfied. By budget balance,
f (j, S(j)) = OC(S(j)) −
j−1
X
f (i, S(j))
i=1
= OC(S(j)) −
j−1
X
f (i, S(j − 1)) − IC i (S(j)) − IC i (S(j − 1))
i=1
j−1
X
= OC(S(j)) − OC(S(j − 1)) +
IC i (S(j)) − IC i (S(j − 1))
i=1
≤ αop d({j}) − IC j (S(j)),
where, the last step follows from the assumption that (20) holds, and the previous step follows from the
budget-balance property. This completes the proof.
A.2.
Proof of Theorem 2
First, we note that in the limit, when
αi
αop
→ ∞ for all i ∈ N , the SIR-feasibility constraints (13) reduce to
Sj−1 Sj + Sj D − Sj−1 D ≤ 0,
2 ≤ j ≤ n.
Since the points are from an underlying metric space, distances satisfy the triangle inequality, which means
Sj−1 Sj + Sj D − Sj−1 D ≥ 0,
2 ≤ j ≤ n.
Sj−1 Sj + Sj D − Sj−1 D = 0,
2 ≤ j ≤ n.
Therefore, it must be that
By summing up the last n − i equations, i.e., i + 1 ≤ j ≤ n, we get
n−1
X
Sj Sj +1 + Sn D − Si D = 0,
j =i
from which we obtain
Pn−1
γr = max
i∈N
j =i
Sj Sj +1 + Sn D
Si D
!
= 1.
This completes the proof.
A.3.
Proof of Theorem 3
First, we note that under the constraint
αi
αop
≥ 1 for all i ∈ N , the SIR-feasibility constraints (13) imply
Sj−1 Sj + Sj D − Sj−1 D ≤
Sj D
,
j
2 ≤ j ≤ n.
(21)
We begin by deriving an upper bound on the starvation factor of the i-th passenger, 1 ≤ i < n, along any
SIR-feasible route. (Note that, in any single dropoff instance, the starvation factor of the last passenger to
be picked up is always 1.) First, we sum up the last n − i inequalities of (21), i.e., i + 1 ≤ j ≤ n, to obtain
n−1
X
j =i
Sj Sj +1 + Sn D − Si D ≤
n
X
Sj D
j =i
j
.
(22)
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
24
Next, we derive upper bounds for each Sj D, i < j ≤ n, in terms of Si D. The j-th SIR-feasibility constraint
from (21) can be rewritten as
Sj D −
Sj D
≤ Sj−1 D − Sj−1 Sj .
j
We know that Sj−1 Sj + Sj−1 D ≥ Sj D, since all points are from an underlying metric space and therefore,
distances are symmetric and satisfy the triangle inequality. Using this inequality above, we get
Sj D
≤ Sj−1 D − (Sj D − Sj−1 D)
j
=⇒ (2j − 1)Sj D ≤ 2jSj−1 D
2j
=⇒ Sj D ≤
Sj−1 D.
2j − 1
Sj D −
Unraveling the recursion yields
j
Y
2k
Sj D ≤
2k
−1
k=i+1
where, for m ≥ 1, Cm =
Qm
2k
k=1 2k−1
Cj =
!
Si D =
Cj
Si D,
Ci
. We can evaluate Cj as follows:
j
Y
j
Y (2k)2
22j (j!)2
22j
2k
=
=
= 2j .
2k − 1 k=1 2k(2k − 1)
(2j)!
j
k=1
We then use a known lower bound for the central binomial coefficient,
2j
j
≥
22j−1
√
j
√
, to obtain Cj ≤ 2 j. This
√
yields Sj D ≤
2 j
Ci
Si D. Substituting in (22), we get
√
n
n
X
2 j
2 X 1
√
Sj Sj +1 + Sn D − Si D ≤
Si D =
Ci j
Ci j =i j
j =i
j =i
!!
n−1
n
X
2 X 1
√
=⇒
Sj Sj +1 + Sn D ≤ 1 +
Si D.
Ci j =i j
j =i
n−1
X
!
Si D
This results in the desired upper bound for the starvation factor of the i-th passenger along any SIR-feasible
route:
2
γr (i) ≤ 1 +
Ci
n
X
1
√
j
j =i
!
.
The starvation factor of a route is the maximum starvation factor of all its passengers:
!!
!
n
n
n
X
2 X 1
2 X 1
1
√
√
√ ,
γr = max γr (i) ≤ max 1 +
=1+
=1+
i∈N
1≤i<n
Ci j =i j
C1 j =1 j
j
j =1
since Ci is increasing in i and C1 = 2. The final step is to show that for all n ≥ 1,
Pn
j =1
1
√
j
√
≤ 2 n − 1. The
proof is by induction. The base case (for n = 1) is satisfied with equality. Assume that the statement is true
√
√
√
Pk+1
4k(k+1)+1
4k(k+1)+1+1
√
√
for some k ≥ 1. Then, for k + 1, we have, j =1 √1j ≤ 2 k − 1 + √k1+1 =
−
1
≤
−1 =
k+1
k+1
√
√
(2k+1)+1
√
− 1 = 2 k + 1 − 1, which completes the inductive step. Using this bound, we get γr ≤ 2 n, as
k+1
desired. This completes the proof.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
A.4.
25
Proof of Theorem 4
First, we note that in the limit, when
αi
αop
→ 0 for all i ∈ N , the SIR-feasibility constraints (13) reduce to
Sj−1 Sj + Sj D − Sj−1 D ≤ Sj D,
2 ≤ j ≤ n.
(23)
Our proof technique is exactly the same as that for Theorem 3. We begin by deriving an upper bound on
the starvation factor of the i-th passenger, 1 ≤ i < n, along any SIR-feasible route, by summing up the last
n − i inequalities of (23) to obtain
n−1
X
Sj Sj +1 + Sn D − Si D ≤
j =i
n
X
Sj D.
(24)
j =i
Next, we derive upper bounds for each Sj D, i < j ≤ n, in terms of Si D. The j-th SIR-feasibility constraint
from (23) can be rewritten as Sj−1 Sj ≤ Sj−1 D. Using this in the triangle inequality Sj D ≤ Sj−1 Sj + Sj−1 D,
we get Sj D ≤ 2Sj−1 D. Unraveling this recursion then yields Sj D ≤ 2j−i Si D. Substituting this in (24),
n−1
X
Sj Sj +1 + Sn D − Si D ≤
j =i
=⇒
n−1
X
n
X
2n−i Si D =
j =i
n−i
X
2j Si D = 2n−i+1 − 1 Si D
j =0
Sj Sj +1 + Sn D ≤ 2n−i+1 Si D.
j =i
Thus, the starvation factor of the i-th passenger along any SIR-feasible route is upper bounded as γr (i) ≤
2n−i+1 . Finally,
γr = max γr (i) ≤ max 2n−i+1 = 2n .
i∈N
1≤i<n
This completes the proof.
A.5.
Proof of Theorem 5
Pj−1 −1
To reduce notational clutter, we let zj = 1 + α1op k=1 αk
, for 1 ≤ j ≤ n. We exhibit a single dropoff
Pn
instance of size n for which there is a unique SIR-feasible path whose starvation factor is exactly j =1 zj .
This instance is depicted in Fig. 5. Here, Sj D = ` for 1 ≤ j ≤ n, and Sj−1 Sk > Sj−1 Sj = zj ` for 2 ≤ j < k ≤ n.
It is straightforward to see that the route (S1 , S2 , . . . , Sn , D) is SIR-feasible from (13), since for 2 ≤ j ≤ n,
we have Sj−1 Sj + Sj D − Sj−1 D = zj ` + ` − ` = zj Sj D, by construction. Thus, the starvation factor for this
Pn
Pn
route is given by j =2 zj + 1 = j =1 zj , as desired.
It remains to be shown that no other route is SIR-feasible. First, we note that the SIR-feasibility constraints (13) for this example simplify to
Sj−1 Sj ≤ zj `,
2 ≤ j ≤ n,
(25)
where z2 > z3 > . . . > zn , and Sj refers to the j-th pickup point along the route. The proof is by induction.
First, consider the pickup point S1 , whose distance from S2 is z2 `, and from any other pickup point is strictly
greater than z2 `, by construction. From (25), it can be seen that no two pickup points that are more than
z2 ` apart can be visited in succession, and that the only way to visit two pickup points that are exactly z2 `
apart is to visit them first and second. Thus, any SIR-feasible route must begin by visiting S1 and S2 first.
This logic can be extended to build the unique SIR-feasible route that we analyzed above. This completes
the proof.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
26
Figure 5
Appendix B:
B.1.
Single dropoff instance to establish lower bound on the SIR-starvation factor.
Proofs from Section 6
Proof of Theorem 7
Given an instance of the Hamiltonian path problem in the form of a simple, undirected graph G = (V, E),
where V = {v1 , v2 , . . . , vn }, we construct an instance of SIR-Feasibility as follows. Let Pj denote a pickup
point corresponding to vertex vj ∈ V . Let N = {P1 , P2 , . . . , Pn } denote the set of pickup points, and D denote
the common dropoff point. Then, we set the pairwise distances to
(
`
, (vi , vj ) ∈ E
Pi Pj = n
`, otherwise,
where ` > 0 is any constant. We also set Pi D = ` for all i, and αop = α1 = α2 = . . . = αn , so that the
SIR-feasibility constraints are given by (21). Then, there is a one-to-one correspondence between the set of
Hamiltonian paths in G and the set of SIR-feasible routes in the corresponding instance of SIR-Feasibility,
as follows:
1. Given a Hamiltonian path through a sequence of vertices (u1 , u2 , . . . , un ) in G, let the corresponding
sequence of pickup points be (S1 , S2 , . . . , Sn ). Then, the route (S1 , S2 , . . . , Sn , D) is SIR-feasible, since
the SIR-feasibility constraints (21) reduce to Sj−1 Sj ≤
`
j
for 2 ≤ j ≤ n, which are true, by construction.
2. Given an SIR-feasible route (S1 , S2 , . . . , Sn , D), let the corresponding sequence of vertices in G be
(u1 , u2 , . . . , un ). Since the route is SIR-feasible, it must be that Sj−1 Sj ≤
`
j
for 2 ≤ j ≤ n. By construction,
`
n
this means that Sj−1 Sj = , implying that (uj−1 , uj ) ∈ E for 2 ≤ j ≤ n. Thus, the corresponding path
is Hamiltonian.
Hence, any algorithm for SIR-Feasibility can be used to solve the undirected Hamiltonian path problem
with a polynomial overhead in running time. Since the latter is NP-hard, so is the former. This completes
the proof.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
B.2.
27
Proof of Theorem 8
Given an instance of Metric-Path-TSP in the form of a complete undirected graph G = (V, E) and distances
d(vi , vj ) for each vi , vj ∈ V from a metric space, we construct an instance of Opt-SIR-Route as follows. Let
Pj denote a pickup point corresponding to vertex vj ∈ V . Let N = {P1 , P2 , . . . , Pn } denote the set of pickup
points, and D denote the common dropoff point.
We set the pairwise distances Pi Pj to be equal to d(vi , vj ) for all vi , vj ∈ V . We also set Pi D = L for all i,
where
L>n
max Pi Pj
1≤i<j≤n
is any constant. We also set αop = α1 = α2 = . . . = αn , so that the SIR-feasibility constraints are given by (21).
It is easy to see that for any route (S1 , S2 , . . . , Sn , D), these SIR-feasibility constraints reduce to Sj−1 Sj ≤
`
j
for 2 ≤ j ≤ n, which are true, by construction and our choice of L. Thus, all n! routes in our constructed
instance of Opt-SIR-Route are SIR-feasible. Moreover, by construction, the distance traveled along any route
is exactly L more than the weight of the path determined by the corresponding sequence of vertices in G.
This implies that any optimal SIR-feasible route is given by a sequence of pickup points corresponding to
an optimal Hamiltonian path in G, followed by a visit to D. Hence, any algorithm for Opt-SIR-Route can
be used to solve Metric-Path-TSP with a polynomial overhead in running time. Since the latter is NP-hard,
so is the former. This completes the proof.
B.3.
Proof of Reduction from Section 6.3
Lemma 1. Any integral maximum flow from s to t must be comprised of m0 vertex-disjoint paths between
the source vertex s and the dropoff vertex n + 1.
Proof.
First, we observe that any integral feasible flow from s to t in the network is comprised of vertex-
disjoint paths between the source vertex s and the dropoff vertex n + 1, each carrying one unit of flow. This
is because, every entry vertex uin has only one outgoing edge, namely, the one directed to its corresponding
exit vertex uout , which has unit capacity. (Similarly, every exit vertex only has one incoming edge, of unit
capacity.) Thus, once a unit of flow is routed through uin and uout by some path, another path cannot route
any additional flow through these vertices. Since the maximum flow on the network is m0 units, any integral
feasible maximum flow would have to have m0 such vertex-disjoint paths between s and n + 1, each carrying
one unit of flow. This completes the proof.
Lemma 2. In any integral minimum cost flow, for every u ∈ N , there is exactly one unit of flow entering
uin and exactly one unit of flow leaving uout .
Proof.
From the proof of Lemma 1, any integral feasible flow from s to t in the network is comprised of
vertex-disjoint paths between the source vertex s and the dropoff vertex n + 1. Suppose by way of contradiction, an integral minimum cost flow does not route any flow through vin for some v ∈ N . Let Gv denote
the set of passengers z ∈ N such that z < v and a unit of flow is routed via (zin , zout ). Consider two cases:
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
28
1. Case 1: Gv 6= ∅. Let u = max Gv , and let Pu be the path that carries a unit of flow from s to n + 1
through uin and uout . The first vertex in Pu after uout is either an entry vertex win for some w ∈ N
(with w > v), or the dropoff vertex n + 1. Then, we construct a new flow where Pu is modified to route
its unit of flow from uout first to vin to vout and then to win or n + 1, as the case may be. (Note that
this new flow is feasible, since u < v < w < n + 1.) If M and M 0 denote the costs of the original flow
and the new flow, then, we show that M 0 < M , contradicting the optimality of M :
• If the original flow took the route uout → win , and consequently, the new flow takes the route
uout → vin → vout → win , then, M 0 = M + Su Sv − L + Sv Sw − L − (Su Sw − L) < M by our choice of L.
• If the original flow took the route uout → n + 1, and consequently, the new flow takes the route
uout → vin → vout → n + 1, then, M 0 = M + Su Sv − L + Sv Sn+1 − Su Sn+1 < M by our choice of L.
2. Case 2: Gv = ∅. Let w ∈ N be such that a unit of flow is routed from s to win , Pw denoting the
corresponding path. There may be more than one choice for win as defined, but all of them satisfy
v < w, since Gv = ∅, so it does not matter which one is picked. As before, we construct a new flow
where Pw is modified to route its unit of flow from s first to vin to vout and then to win . (Note that
this new flow is feasible, since v < w.) If M and M 0 denote the costs of the original flow and the new
flow, M 0 = M + Sv Sw − L < M by our choice of L, contradicting the optimality of M .
This completes the proof.
Lemma 3. The partition of N according to the m0 vertex-disjoint paths between s and n + 1 in an integral
minimum cost maximum flow corresponds to the optimal allocation of the n totally ordered passengers among
m0 uncapacitated vehicles in the single dropoff scenario.
Proof.
From Lemma 1 and Lemma 2, we know that any integral minimum cost maximum flow F is
comprised of m0 vertex-disjoint paths between s and n + 1 that cover all n pickup points between them,
by routing a unit of flow along (uin , uout ) for all u ∈ N . We adopt a simplified representation of a path by
removing the edges from the source vertex s, as well as the edges between uin and uout , the entry and exit
vertices corresponding to pickup points u ∈ N . For example, a path s → uin → uout → vin → vout → n + 1
would be contracted to u → v → n + 1. Note that this does not affect the cost computation, since only zero
cost edges are removed. For any u, v ∈ N , the cost of any edge (u, v) in the new representation is simply the
cost of the edge (uout , vin ) in the old representation. Similarly, for any u ∈ N , the cost of any edge (u, n + 1)
in the new representation is simply the cost of the edge (uout , n + 1) in the old representation. Let the set of
these m0 paths be denoted as PF . Thus, we have established a one-to-one correspondence between (a) the set
of all integral flows F comprised of m0 vertex-disjoint paths PF that collectively cover all n pickup locations,
and (b) the set of all allocations of n totally ordered passengers (traveling to a single dropoff location n + 1)
to m0 uncapacitated vehicles.
For any path P ∈ PF , let |P | denote the length of the path, that is, the number of edges in the path. The
cost of path P is then given by
c(P ) =
XX
1≤u<v≤n
(u,v )∈P
(Su Sv − L) +
X
1≤u≤n
(u,n+1)∈P
Su Sn+1 .
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
29
Since all paths end with vertex n + 1, there are |P | − 1 terms in the first sum and 1 term in the last sum.
Thus, c(P ) can be equivalently written as
c(P ) =
XX
Su Sv − (|P | − 1)L.
1≤u<v≤n+1
(u,v )∈P
The cost of flow F is simply the sum of the costs of the paths in PF , given by
X
c(F) =
c(P ) =
P ∈PF
XX
Su Sv −
1≤u<v≤n
S +1
(u,v )∈
PF
X
(|P | − 1)L.
P ∈PF
Since |P |, the length of path P , also denotes the number of pickup points covered by P , and all the m0 paths
are vertex-disjoint (except for n + 1), the summation in the second term is simply n − m0 , independent of
the flow F. Thus,
c(F) =
XX
Su Sv − (n − m0 )L = c(AF ) − (n − m0 )L,
(26)
1≤u<v≤n
S +1
(u,v )∈
PF
where c(AF ) denotes the cost (total vehicle-miles traveled) of the corresponding allocation of n totally
ordered passengers (traveling to a single dropoff location n + 1) to m0 uncapacitated vehicles. From (26), it
is clear that the set of integral minimum cost maximum flows arg minF c(F) also corresponds to the set of
optimal allocations of n totally ordered passengers among m0 uncapacitated vehicles in the single dropoff
scenario. This completes the proof.
Theorem 9. There exists a poly-time algorithm to find an optimal allocation of totally ordered passengers
to uncapacitated vehicles in the single dropoff scenario.
Proof.
Using the one-to-one correspondence established in Lemma 3, for each “guess” 1 ≤ m0 ≤ n, we
find the corresponding optimal allocation by solving a minimum cost maximum flow problem in poly-time,
finally choosing a guess with the overall least cost allocation.
References
Agatz N, Erera A, Savelsbergh M, Wang X (2012) Optimization for dynamic ride-sharing: A review. European
Journal of Operational Research 223(2):295–303.
Agatz NA, Erera AL, Savelsbergh MW, Wang X (2011) Dynamic ride-sharing: A simulation study in metro
atlanta. Transportation Research Part B: Methodological 45(9):1450–1464.
Ahn J, Kim C, Han J, Choi Yr, Huh J (2012) Dynamic virtual machine scheduling in clouds for architectural shared resources. Proceedings of the 4th USENIX conference on Hot Topics in Cloud Computing,
HotCloud’12.
Aksin Z, Armony M, Mehrotra V (2007) The modern call center: A multi-disciplinary perspective on operations management research. Production and Operations Management 16(6):665–688.
Alonso-Mora J, Samaranayake S, Wallar A, Frazzoli E, Rus D (2017) On-demand high-capacity ride-sharing
via dynamic trip-vehicle assignment. Proceedings of the National Academy of Sciences 462467.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
30
Armbrust M, Fox A, Griffith R, Joseph AD, Katz R, Konwinski A, Lee G, Patterson D, Rabkin A, Stoica I,
et al. (2010) A view of cloud computing. Communications of the ACM 53(4):50–58.
Banerjee S, Johari R, Riquelme C (2015) Pricing in ride-sharing platforms: A queueing-theoretic approach.
Proceedings of the Sixteenth ACM Conference on Economics and Computation, 639–639.
Bateni MH, Chen Y, Ciocan DF, Mirrokni V (2016) Fair Resource Allocation in a Volatile Marketplace .
Bertsimas D, Farias VF, Trichakis N (2011) The price of fairness. Operations research 59(1):17–31.
Bimpikis K, Candogan O, Daniela S (2016) Spatial pricing in ride-sharing networks. SSRN .
Bistaffa F, Farinelli A, Chalkiadakis G, Ramchurn S (2015) Recommending fair payments for large-scale
social ridesharing. Proceedings of the 9th ACM Recommender Systems Conference (RecSys), 139–146.
Bowersox DJ, Closs DJ, Stank TP (2000) Ten mega-trends that will revolutionize supply chain logistics.
Journal of Business Logistics 21(2):1.
Cao B, Alarabi L, Mokbel MF, Basalamah A (2015) SHAREK: A scalable dynamic ride sharing system.
Proceedings of the 16th IEEE International Conference on Mobile Data Management (MDM), 4–13.
Cole A (2015) Improving Airport Funding to Meet the Needs of Passengers. Tax Foundation Fiscal Fact
(466).
Cordeau JF, Laporte G, Savelsbergh MW, Vigo D (2006) Vehicle routing. Handbooks in operations research
and management science: Transportation, volume 14, chapter 6, 367–428 (Elsevier).
Dickerson JP, Procaccia AD, Sandholm T (2014) Price of fairness in kidney exchange. Proceedings of the 2014
international conference on Autonomous agents and multi-agent systems, 1013–1020 (International
Foundation for Autonomous Agents and Multiagent Systems).
Elmachtoub AN, Levi R (2014) From cost sharing mechanisms to online selection problems. Mathematics of
Operations Research 40(3):542–557.
Esteban J (1991) The social viability of money (competitive equilibria and the core of overlapping generations
economies), volume 372 (Springer-Verlag).
Furuhata M, Dessouky M, Ordóñez F, Brunet ME, Wang X, Koenig S (2013) Ridesharing: The state-of-theart and future directions. Transportation Research Part B: Methodological 57:28–46.
Geisberger R, Luxen D, Neubauer S, Sanders P, Volker L (2010) Fast detour computation for ride sharing.
Proceedings of the 10th Workshop on Algorithmic Approaches for Transportation Modelling, Optimization, and Systems, OASIcs-OpenAccess Series in Informatics, volume 14, 88–99 (Schloss DagstuhlLeibniz-Zentrum fuer Informatik).
Haitao Cui T, Raju JS, Zhang ZJ (2007) Fairness and channel coordination. Management Science 53(8):1303–
1314.
Heydrich S, van Stee R (2015) Dividing connected chores fairly. Theoretical Computer Science 593:51–61.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
31
Jackson KL (2011) The economic benefit of cloud computing. Technical report, NJVC.
Jain K, Mahdian M (2007) Cost sharing. Nisan N, Roughgarden T, Tardos É, Vazirani VV, eds., Algorithmic
Game Theory, chapter 15, 385–410 (Cambridge University Press).
Kamar E, Horvitz E (2009) Collaboration and shared plans in the open world: Studies of ridesharing. IJCAI,
187–194.
Katta A, Sethuraman J (2005) Pricing strategies and service differentiation in queues—A profit maximization
perspective. Working paper, Columbia University.
Kleiner A, Nebel B, Ziparo VA (2011) A mechanism for dynamic ride sharing based on parallel auctions.
Proceedings of the Twenty-Second International Joint Conference on Artificial Intelligence (IJCAI),
266–272.
Kondo D, Javadi B, Malecot P, Cappello F, Anderson DP (2009) Cost-benefit analysis of cloud computing versus desktop grids. IEEE International Symposium on Parallel & Distributed Processing, 1–12
(IEEE).
McKenzie B (2015) Who drives to work? Commuting by automobile in the United States: 2013. American
Community Survey Reports .
Moulin H (2004) Fair division and collective welfare (MIT press).
Narahari Y, Narayanam R, Garg D, Prakash H (2009) Foundations of mechanism design. Game Theoretic
Problems in Network Economics and Mechanism Design Solutions, Part of the Series on Advanced
Information and Knowledge Processing, 1–131 (Springer).
Nguyen DT (2013) Fair cost sharing auction mechanisms in last mile ridesharing. Ph.D. thesis, Singapore
Management University.
Ozkan E, Ward AR (2016) Dynamic matching for real-time ridesharing. SSRN .
Papadimitriou CH (1977) The euclidean travelling salesman problem is NP-complete. Theoretical Computer
Science 4(3):237–244.
Papadimitriou CH (1994) Computational Complexity (Addison-Wesley).
Parkes DC (2007) Online mechanisms. Nisan N, Roughgarden T, Tardos É, Vazirani VV, eds., Algorithmic
Game Theory, chapter 15, 411–439 (Cambridge University Press).
Peha JM (2009) Sharing spectrum through spectrum policy reform and cognitive radio. Proceedings of the
IEEE 97(4):708–719.
Pelzer D, Xiao J, Zehe D, Lees MH, Knoll AC, Aydt H (2015) A partition-based match making algorithm
for dynamic ridesharing. IEEE Transactions on Intelligent Transportation Systems 16(5):2587–2598.
Prins C, Lacomme P, Prodhon C (2014) Order-first split-second methods for vehicle routing problems: A
review. Transportation Research Part C: Emerging Technologies 40:179–200.
Gopalakrishnan, Mukherjee, and Tulabandhula: The Costs and Benefits of Sharing
Article submitted to Management Science; manuscript no. MS-17-01173
32
Santos DO, Xavier EC (2013) Dynamic taxi and ridesharing: A framework and heuristics for the optimization
problem. Proceedings of the Twenty-Third International Joint Conference on Artificial Intelligence
(IJCAI), volume 13, 2885–2891.
Shen W, Lopes CV, Crandall JW (2016) An online mechanism for ridesharing in autonomous mobility-ondemand systems. Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, 475–481.
Sivak M, Schoettle B (2016) Recent decreases in the proportion of persons with a driver’s license across all
age groups. Transportation Research Institute Report, UMTRI-2016-4, University of Michigan, Ann
Arbor .
Zhao D, Ramchurn SD, Jennings NR (2015) Incentive design for ridesharing with uncertainty. CoRR
abs/1505.01617.
Zhao D, Zhang D, Gerding EH, Sakurai Y, Yokoo M (2014) Incentives in ridesharing with deficit control.
Proceedings of the 2014 International Conference on Autonomous Agents and Multi-Agent Systems,
1021–1028.
Zou B, Kafle N, Chang YT, Park K (2015) US airport financial reform and its implications for airport
efficiency: An exploratory investigation. Journal of Air Transport Management 47:66–78.
| 8cs.DS
|
Coordinated Motion Planning: Reconfiguring a Swarm of Labeled
Robots with Bounded Stretch∗
Erik D. Demaine1 , Sándor P. Fekete2 , Phillip Keldenich2 , Henk Meijer3 , and Christian
Scheffer2
arXiv:1801.01689v1 [cs.CG] 5 Jan 2018
1
MIT Computer Science and Artificial Intelligence Laboratory, Cambridge, MA, USA,
[email protected]
2
Department of Computer Science, TU Braunschweig Braunschweig, Germany,
{s.fekete,p.keldenich,c.scheffer}@tu-bs.de
3
Science Department, University College Roosevelt Middelburg, The Netherlands,
[email protected]
Abstract
We present a number of breakthroughs for coordinated motion planning, in which the objective is
to reconfigure a swarm of labeled convex objects by a combination of parallel, continuous, collision-free
translations into a given target arrangement. Problems of this type can be traced back to the classic work
of Schwartz and Sharir (1983), who gave a method for deciding the existence of a coordinated motion
for a set of disks between obstacles; their approach is polynomial in the complexity of the obstacles, but
exponential in the number of disks. Other previous work has largely focused on sequential schedules, in
which one robot moves at a time.
We provide constant-factor approximation algorithms for minimizing the execution time of a coordinated, parallel motion plan for a swarm of robots in the absence of obstacles, provided some amount of
separability.
Our algorithm achieves constant stretch factor: If all robots are at most d units from their respective
starting positions, the total duration of the overall schedule is O(d). Extensions include unlabeled robots
and different classes of robots. We also prove that finding a plan with minimal execution time is NP-hard,
even for a grid arrangement without any stationary obstacles. On the other hand, we show that for
densely packed disks that cannot be well separated, a stretch factor Ω(N 1/4 ) may be required. On the
positive side, we establish a stretch factor of O(N 1/2 ) even in this case.
1
Introduction
Since the beginning of computational geometry, robot motion planning has been at the focus of algorithmic
research. Planning the relocation of a geometric object among geometric obstacles leads to intricate scientific
challenges, requiring the combination of deep geometric and mathematical insights with algorithmic techniques.
With the broad and ongoing progress in robotics, the increasing importance of intelligent global planning
with performance guarantees requires more sophisticated algorithmic reasoning, in particular when it comes
to the higher-level task of coordinating the motion of many robots.
From the early days, multi-robot coordination has received attention from the algorithmic side. Even in
the groundbreaking work by Schwartz and Sharir [52] from the 1980s, one of the challenges was coordinating
∗ This
work was partially supported by the DFG Research Unit Controlling Concurrent Change, funding number FOR 1800,
project FE407/17-2, Conflict Resolution and Optimization.
1
the motion of several disk-shaped objects among obstacles. Their algorithms run in time polynomial in the
complexity of the obstacles, but exponential in the number of disks. This illustrates the significant challenge
of coordinating many individual robots. In addition, a growing number of applications focus primarily on
robot interaction in the absence of obstacles, such as air traffic control or swarm robotics, where the goal is
overall efficiency, rather than individual navigation.
With the challenges of multi-robot coordination being well known, there is still a huge demand for positive
results with provable performance guarantees. In this paper, we provide significant progress in this direction,
with a broad spectrum of results.
1.1
Our Results
• For the problem of minimizing the total time for reconfiguring a system of labeled circular robots
in a grid environment, we show that it is strongly NP-complete to compute an optimal solution; see
Theorem 1.
• We give an O(1)-approximation for the long-standing open problems of parallel motion-planning with
minimum makespan in a grid setting. This result is based on establishing an absolute performance
guarantee: We prove that for any labeled arrangement of robots, there is always an overall schedule
that gets each robot to its target destination with bounded stretch, i.e., within a constant factor of
the largest individual distance. See Theorem 3 for the base case of grid-based configurations, which is
extended later on.
• For our approach, we make use of a technique to separate planar (cyclic) flows into so-called subflows
whose thickness can be controlled by the number of subflows, see Definition 7 and Lemma 8. This is of
independent interest for the area of packet routing with bounded memory: Our Theorem 4 implies that
O(D) steps are sufficient to route any permutation of dilation D on the grid, even with a buffer size of
1, resolving an open question by Scheideler [51] dating back to 1998.
• We extend our approach to establish constant stretch for the generalization of colored robot classes, for
which unlabeled robots are another special case; see Theorem 13.
• We extend our results to the scenario with continuous motion and arbitrary coordinates, provided
the distance between a robot’s start and target positions is at least one diameter; see Theorem 15.
This implies that efficient multi-robot coordination is always possible under relatively mild separability
conditions; this includes non-convex robots.
1/4
• For the continuous case of N
√ unit disks and weaker separability, we establish a lower bound of Ω(N )
and an upper bound of O( N ) on the achievable stretch, see Theorem 14 and Theorem 15.
We also highlight the geometric difficulty of computing optimal trajectories even in seemingly simple
cases; due to limited space, this can be found in Appendix E.
1.2
Related Work
Different variants of multiple-object motion planning problems have received a large amount of attention
from researchers in various areas of computer science and engineering; see [18] for a survey. Their practical
relevance is reflected by the fact that there are industrial solutions used in automated warehouses for certain
restricted forms of these problems [73]. There are different orthogonal criteria by which these problems can
be characterized. A very important distinction is between discrete and continuous scenarios. In the discrete
case, the input is a graph in which no two objects may use a vertex or edge at the same time; depending
on the scenario, it may be allowed to rotate fully populated cycles. In the continuous or geometric setting,
the objects are shapes in some geometric space which must be moved to a given target position in such a
way that their interiors do not intersect at any time. Depending on the scenario, the shapes may or may
not touch. Moreover, the objects may be confined to a certain region and there may be stationary obstacles.
2
Under these restrictions, it is unclear whether the target configuration is reachable at all. Aronov et al. [4]
demonstrate that, for up to three robots of constant complexity, a path can be constructed efficiently if
one exists. Ramanathan and Alagar [45] as well as Schwartz and Sharir [52] consider the case of several
disk-shaped objects moving amongst polygonal obstacles. They both find algorithms deciding whether a given
target configuration is reachable. Their algorithms run in time polynomial in the complexity of the obstacles,
but exponential in the number of disks. Hopcroft et al. [32] and Hopcroft and Wilfong [33] demonstrate
the reachability of a given target configuration is PSPACE-complete to decide; this already holds when
restricted to rectangular objects moving in a rectangular region. Their proof was later generalized by Hearn
and Demaine [29, 30], who proved that rectangles of size 1 × 2 and 2 × 1 are sufficient and introduced a
more general framework to prove PSPACE-hardness of certain block sliding games. Moreover, this problem
is similar to the well-known Rush Hour Problem, which was shown to be PSPACE-complete by Flake and
Baum [23]. For moving disks, Spirakis and Yap [62] have proven strong NP-hardness of the same problem;
however, their proof makes use of disks of varying size. Bereg et al. [6] as well as Abellanas et al. [1] consider
minimizing the number of moves of a set of disks into a target arrangement without obstacles. They provide
simple algorithms and establish upper and lower bounds on the number of moves, where a move consists of
sliding one disk along some curve without intersecting other disks. These bounds were later improved on by
Dumitrescu and Jiang [19], who also prove that the problem remains NP-hard for congruent disks even when
the motion is restricted to sliding.
Kirkpatrick and Liu [37] consider the case of moving two disks of arbitrary radius from a start into a
target configuration in an otherwise obstacle-free plane, minimizing the sum of distances travelled by the
disks. They provide optimal solutions for two disks moving from an arbitrary initial configuration into an
arbitrary goal configuration. Their arguments do not seem to generalize to the makespan.
Dı́az-Báñez et al. [17] considered the task of extracting a single object from a group of convex objects,
moving a minimal number of objects out of the way. They present an algorithm that finds the optimal
direction for extracting the object in polynomial time.
On the practical side, there are several approaches to solving multi-object motion planning problems,
both optimally and heuristically. For discrete instances with a moderate number of objects, optimal solutions
can be found using standard search strategies like A∗ [28] in the high-dimensional search space of possible
configurations. Numerous techniques can be used to improve the efficiency of these strategies [22, 26, 63].
Moreover, there is some work employing SAT solvers [34, 36] to solve multi-object motion planning problems
to optimality. More recently, Yu and LaValle [74] present an IP-based exact algorithm for minimizing the
makespan that works for hundreds of robots, even for challenging configurations with densities of up to 100%.
For larger instances, one has to resort to heuristic solutions. In priority planning [9,16,21,25,47,67,69], the
paths are planned one-by-one by assigning priorities to the objects and planning the movement in decreasing
order of priority, treating all objects with higher priority as moving obstacles. Kant and Zucker [35] decompose
the problem into planning the paths for all objects and avoiding collisions by adapting the velocity of the
objects appropriately, an approach which several papers are based on [11, 41, 42, 44, 56]. Another approach is
to compute paths for the objects individually and resolve collisions locally [24].
Between these simple decoupled heuristics which only consider individual objects at a time and highdimensional coupled search algorithms lie dynamically-coupled algorithms [3, 5, 53–55, 57, 68] which aim for
better solutions at the price of higher computational costs. These algorithms typically consider individual
objects and only increase the dimension of the search space once a non-trivial interaction between objects is
discovered. Recently, Wagner and Choset [71] provided a complete algorithm based on a similar principle.
With the advent of robot swarms, practical solutions to these problems became more important and the
robotics community started to develop practical sampling-based algorithms [31, 49, 50, 59, 60, 64, 70] which,
while working well in practice, are not guaranteed to find an (optimal) solution. In another recent work, Yu
and Rus [75] present a practical algorithm based on a fine-grained discretization combined with an IP for
the resulting discrete problem to provide near-optimal solutions even for densely populated environments.
Other related work includes Rubenstein et al. [48], who demonstrated how to reconfigure a large swarm of
simple, disk-shaped Kilobots; however, their method is sequential, relocating one robot at a time, so a full
reconfiguration of 1000 robots takes about a day, highlighting the relevance of truly parallel motion planning.
3
Further extensions to higher-dimensional problems (with a wide range of additional motion constraints) are
swarms of drones (e.g., the work by Kumar [66]) and even air traffic control (see Delahaye et al. [12] for a
recent survey).
In both discrete and geometric variants of the problem, the objects can be labeled, colored or unlabeled.
In the labeled case, the objects are all distinguishable and each object has its own, uniquely defined target
position. This is the most extensively studied scenario among the three. In the colored case, the objects
are partitioned into k groups and each target position can only be covered by an object with the right
color. This case was recently considered by Solovey and Halperin [57], who present and evaluate a practical
sampling-based algorithm. In the unlabeled case, the objects are indistinguishable and each target position
can be covered by any object. This scenario was first considered by Kloder and Hutchinson [38], who
presented a practical sampling-based algorithm. In this situation, Turpin et al. [65] prove that it is possible
to find a solution in polynomial time, if one exists. This solution is optimal with respect to the longest
distance traveled by any one robot. However, their results only hold for disk-shaped robots under additional
restrictive assumptions on the free space. For unit disks and simple polygons, Adler et al. [2] provide a
polynomial-time algorithm under the additional assumption that the start and target positions have some
minimal distance from each other. Under similar separability assumptions, Solovey et al. [61] provide a
polynomial time-algorithm that produces a set of paths that is no longer than OPT + 4m, where m is the
number of robots. However, they do not consider the makespan, but only the total path length. On the
negative side, Solovey and Halperin [58] prove that the unlabeled multiple-object motion planning problem is
PSPACE-hard, even when restricted to unit square objects in a polygonal environment.
Regarding discrete multiple-object motion planning, Cǎlinescu et al. [8] consider the non-parallel motion
planning problem on graphs, where each object can be moved along an unoccupied path in one move. They
prove that both in the unlabeled and in the labeled case, minimizing the number of moves required is
APX-hard. They provide 3-approximation algorithms for the unlabeled case on general graphs. Moreover,
they prove that the problem remains NP-complete on the infinite rectangular grid. Their results are different
from our results because the objective they consider is not closely related to the makespan. For other work,
see [7, 14, 15, 27] for particular examples.
On grid graphs, the problem can be cast as a very restrictive variant of mesh-connected routing, where
each processor can only hold one packet at any time. However, approaches developed for this problem (see
Kunde [40] and Cheung and Lau [10]) typically assume that at least a constant number of packets can be
held at any processor. On the other hand, on grid graphs, the problem resembles the generalization of the
15-puzzle, for which Wagner [72] and Kornhauser et al. [39] have given an efficient algorithm that decides
reachability of a target configuration and provided both lower and upper bounds on the number of moves
required. However, Ratner and Warmuth [46] proved finding a shortest solution for this puzzle remains
NP-hard. Demaine et al. [13] also consider various grids. For the triangular grid, they give efficiently verifiable
conditions for checking whether a solution exists.
2
Preliminaries
In the grid setting of Section 3 we consider an n1 × n2 -grid G = (V, E), which is dual to an n1 × n2 -rectangle
P in which the considered robots are arranged. A configuration of P is a mapping C : V → {1, . . . , N, ⊥},
which is injective w.r.t. the labels {1, . . . , N } of the N ≤ |P | robots to be moved, where ⊥ denotes the empty
square. The inverse image of a robot’s label ` is C −1 (`). In the following, we consider a start configuration
Cs and target configuration Ct ; for i ∈ {1, . . . , N }, we call Cs−1 (i) and Ct−1 (i) the start and target position of
the robot i. Given the (minimum) Manhattan distance between each robot’s start/target positions for each
robot, we denote by d the maximum such distance over all robots.
A configuration C1 : V → {1, . . . , N, ⊥} can be transformed within one single transformation step into
another configuration C2 : V → {1, . . . , N, ⊥}, denoted C1 → C2 , if C1−1 (`) = C2−1 (`) or (C1−1 (`), C2−1 (`)) ∈ E
holds for all ` ∈ {1, . . . , N }, i.e., if each robot does not move or moves to one of the at most four adjacent
squares. Furthermore, two robots cannot exchange their squares in one transformation step, i.e., for all
occupied squares v 6= w ∈ V , we require that C2 (v) = C1 (w) implies C2 (w) 6= C1 (v). For M ∈ N, a schedule
4
is a sequence C1 → · · · → CM of transformations. The number of steps in a schedule is called its makespan.
Given a start configuration Cs and a target configuration Ct , the optimal makespan is the minimum number
of steps in a schedule starting with Cs and ending with Ct . Let n > 1. Note that for the 2 × 2-, 1 × n- and
n × 1-rectangles, there are pairs of start and target configurations where no such sequence exists. For all other
rectangles, such configurations do not exist; we provide an O(1)-approximation of the makespan in Section 3.
For the continuous setting of Section 5, we consider N robots R := {1, . . . , N } ⊆ N. The Euclidean
distance between two points p, q ∈ R2 is |pq| := ||p − q||2 . Every robot r has a start and target position
sr , tr ∈ R2 with |si sj |, |ti tj | ≥ 2 for all i 6= j. In the following, d := maxr∈R |sr tr | is the maximum distance
a robot has to cover. A trajectory of a robot r is a curve mr : [0, Tr ] → R2 , where Tr ∈ R+ denotes the
travel time of r. This curve mr does not have to be totally differentiable, but must be totally left- and
right-differentiable. Intuitively, at any point in time, a robot has a unique past and future direction that
are not necessarily identical. This allows the robot to make sharp turns, but does not allow jumps. We
bound the speed of the robot by 1, i.e., for each point in time, both left and right derivative of mr have
Euclidean length at most 1. Let mi : [0, Ti ] → R2 and mj : [0, Tj ] → R2 be two trajectories; w.l.o.g., all
travel times are equal to the maximum travel time Tmax by extending mr with mr (t) = mr (Tr ) for all
Tr < t ≤ Tmax . The trajectories mi and mj are compatible if the corresponding robots do not intersect
at any time, i.e., if |mi (t)mj (t)| ≥ 2 holds for all t ∈ [0, Ti ]. A trajectory set of R is a set of compatible
trajectories {m1 , . . . , mN }, one for each robot. The (continuous) makespan of a trajectory set {m1 , . . . , mN }
is defined as maxr∈R Tr . A trajectory set {m1 , . . . , mN } realizes a pair of start and target configurations
S := ({s1 , . . . , sN }, {t1 , . . . , tN }) if mr (0) = sr and mr (Tr ) = tr hold for all r ∈ R. We are searching for a
trajectory set {m1 , . . . , mN } realizing S with minimal makespan.
3
Labeled Grid Permutation
Let n1 ≥ n2 ≥ 2, n1 ≥ 3 and let P be an n1 × n2 -rectangle. In this section, we show that computing the
optimal makespan of arbitrarily chosen start and target configurations Cs and Ct of k robots in P is strongly
NP-complete. This is followed by a O(1)-approximation for the makespan.
Theorem 1. The minimum makespan parallel motion planning problem on a grid is strongly NP-hard.
We prove hardness using a reduction from Monotone 3-Sat. Intuitively speaking, given a formula, we
construct a parallel motion planning instance with a variable robot for each variable in the formula. To encode
a truth assignment, each variable robot is forced to move on one of two paths. This is done by employing two
groups of auxiliary robots that have to move towards their goal in a straight line in order to realize the given
makespan. These auxiliary robots form moving obstacles whose position is known at any point in time.
The variable robots cross paths with checker robots, one for each literal of the formula, forcing the checker
to wait for one time step if the assignment does not satisfy the literal. The checker robots then cross paths
with clause robots; each clause robot has to move to its goal without delay and can only do so if at least one
of the checkers did not wait. In order to ensure that the checkers meet with the clauses at the right time,
further auxiliary robots force the checkers to perform a sequence of side steps in the beginning. Figure 1
gives a rough overview of the construction; full details of the proof are given in Section 6.
In the proof of NP-completeness, we use a pair of start and target configurations in which the corresponding
grids are not fully occupied. However, for our constant-factor approximation, we assume in Theorem 3 that
the grid is fully occupied. This assumption is without loss of generality; our approximation algorithm works
for any grid population, see Theorem 4.
Our constant-factor approximation is based on an algorithm that computes a schedule with a makespan
upper-bounded by O(n1 + n2 ) described by Lemma 2. Based on Lemma 2, we give a constant factor
approximation of the makespan, see Theorem 3. Finally, we embed the algorithm of Theorem 3 into a more
general approach to ensure simultaneously a polynomial running time w.r.t. the number N of input robots
and a constant approximation factor, see Theorem 4.
Lemma 2. For a pair of start and target configurations Cs and Ct of an n1 × n2 -rectangle, we can compute
in polynomial time w.r.t. n1 and n2 a sequence of O(n1 + n2 ) steps transforming Cs into Ct .
5
clauses
aux.
variables
aux.
side steps
checkers
Figure 1: A sketch of the parallel motion planning instance resulting from the reduction.
The high-level idea of the algorithm of Lemma 2 is the following. We apply a sorting algorithm called
RotateSort [43] that computes a corresponding permutation of an n1 × n2 (orthogonal) grid within
O(n1 + n2 ) parallel steps. Each parallel step is made up of a set of pairwise disjoint swaps, each of which
causes two neighbouring robots to exchange their positions. Because in our model direct swaps are not
allowed, we simulate one parallel step by a sequence of O(1) transformation steps. This still results in a
sequence of O(n1 + n2 ) transformation steps. A detailed description of the algorithm used in the proof of
Lemma 2 is given in Section 6.2.
Based on the algorithm of Lemma 2, we can give a constant-factor approximation algorithm.
Theorem 3. There is an algorithm with running time O(dn1 n2 ) that, given an arbitrary pair of start and
target configurations of an n1 × n2 -rectangle with maximum distance d between any start and target position,
computes a schedule of makespan O(d), i.e., an approximation algorithm with constant stretch.
For the algorithm of Theorem 3, Lemma 2 is repeatedly applied to rectangles of side length O(d),
resulting in O(d) transformation steps in total. Because d is a lower bound on the makespan, this yields an
O(1)-approximation of the makespan.
At a high level, the algorithm of Theorem 3 first computes the maximal Manhattan distance d between a
robot’s start and target position. Then we partition P into a set T of pairwise disjoint rectangular tiles, where
each tile t ∈ T is an n01 × n02 -rectangle for n01 , n02 ≤ 24d. We then use an algorithm based on flows to compute
a sequence of O(d) transformation steps, ensuring that all robots are in their target tile. Once all robots are
in the correct tile, we use Lemma 2 simultaneously on all tiles to move each robot to the correct position
within its target tile. The details of the algorithm of Theorem 3 are given further down in this section.
The above mentioned tiling construction ensures that each square of P belongs to one unambiguously
defined tile and each robot has a start and target tile.
Based on the approach of Theorem 3 we give a O(1)-approximation algorithm for the makespan with a
running time polynomial w.r.t. the number N of robots to be moved.
Theorem 4. There is an algorithm with running time O(N 5 ) that, given an arbitrary pair of start and target
configurations of a rectangle P with N robots to be moved and maximum distance d between any start and
target position, computes a schedule of makespan O(d), i.e., an approximation algorithm with constant stretch.
Intuitively speaking, the approach of Theorem 4 distinguishes two cases.
(1) Both b n41 c and the maximum distance d between the robots’ start and target positions, are lowerbounded by the number N of input robots.
6
(2) N > b n41 c or N > d.
In case (1), the grid is populated sparsely enough such that the robots’ trajectories in northern, eastern,
southern, and western direction can be done sequentially by four individual transformation sequences, see
Figure 12.
In order to ensure that each robot has locally enough space, we consider a preprocessed start configuration
Co in which the robots have odd coordinates. We ensure that Cs can be transformed into Co within O(d)
steps. Analogously, we ensure that the outcome of the northern, eastern, southern, and western trajectories is
a configuration Ce with even coordinates, such that Ce can be transformed into Ct within O(d) transformation
steps.
In the second case, we apply the approach of Theorem 3 as a subroutine to a union of smallest rectangles
that contain the robots’ start and target configurations, see Figure 13.
The full detailed version of the proof of Theorem 4 can be found in Section 6.3.
In the rest of Section 3, we give the proof of Theorem 3, i.e. we give an algorithm that computes a
schedule with makespan linear in the maximum distance between robots’ start and target positions. The
remainder of the proof of Theorem 3 is structured as follows. In Section 3.1 we give an outline of our flow
algorithm that ensures that each robot reaches its target tile in O(d) transformation steps. Section 3.2 gives
the full intuition of this algorithm and its subroutines. (For full details, we refer to Section 6).
3.1
Outline of the Approximation Algorithm of Theorem 3
We model the trajectories of robots between tiles as a flow fT , using the weighted directed graph GT =
(T, ET , fT ), which is dual to the tiling T defined in the previous section. In GT , we have an edge (v, w) ∈ ET
if there is at least one robot that has to move from v into w. Furthermore, we define the weight fT ((v, w)) of
an edge as the integer number of robots that move from v to w. As P is fully occupied, fT is a circulation,
i.e., a flow with no sources or sinks, in which flow conservation has to hold at all vertices. Because the side
lengths of the tiles are greater than d, GT is a grid graph with additional diagonal edges and thus has degree
at most 8.
The maximum edge value of fT is Θ(d2 ), but only O(d) robots can possibly leave a tile within a single
transformation step. Therefore, we decompose the flow fT of robots into a partition consisting of O(d)
subflows, where each individual robot’s motion is modeled by exactly one subflow and each edge in the subflow
has value at most d. Thus we are able to realize each subflow in a single transformation step by placing the
corresponding robots adjacent to the boundaries of its corresponding tiles before we realize the subflow. To
facilitate the decomposition into subflows, we first preprocess GT . In total, the algorithm consists of the
following subroutines, elaborated in detail in Section 3.2.
• Step 1: Compute d, the tiling T and the corresponding flow GT .
• Step 2: Preprocess GT in order to remove intersecting and bidirectional edges.
• Step 3: Compute a partition into O(d) d-subflows.
• Step 4: Realize the O(d) subflows using O(d) transformation steps.
• Step 5: Simultaneously apply Lemma 2 to all tiles, moving each robot to its target
position.
3.2
Details of the Approximation Algorithm of Theorem 3
In this section we only give more detailed descriptions of Steps 1-4 because Step 5 is a trivial application of
Lemma 2 to all tiles in parallel.
7
3.2.1
Step 1: Compute d, the Tiling T , and the corresponding Flow GT
The maximal distance between robots’ start and target positions can be computed in a straightforward
manner.
For the tiling, we assume that the rectangle P is axis aligned and that its bottom-left corner is (0, 0).
n1
c vertical lines `v1 , . . . , `vkv with x-coordinate modulo 12d equal to 0. Analogously, we
We consider kv := b 12d
n2
consider kh := b 12d c horizontal lines `h1 , . . . , `hkh with y-coordinate modulo 12d to 0. Finally, we consider the
tiling of P that is induced by the arrangement induced by `v1 , . . . , `vkv −1 , `h1 , . . . , `hhv −1 and the boundary of P ,
see Figure 11. This implies that the side length of a tile is upper-bounded by 24d − 1.
Finally, computing the flow GT is straightforward by considering the tiling T and the robots’ start and
target positions.
3.2.2
Step 2: Ensuring Planarity and Unidirectionality
After initialization, we preprocess GT , removing edge intersections and bidirectional edges by transforming
the start configuration Cs into an intermediate start configuration Cs0 , obtaining a planar flow without
bidirectional edges. This transformation consists of two steps: (1) ensuring planarity and (2) ensuring
unidirectionality.
1
v1
w2
v1
1
2
3
v2
w2
1
w1
v2
(a)
1
w1
(b)
3
2
(c)
(d)
Figure 2: Illustration of the preprocessing (step (1): before and after removing crossing edges (a)+(b) and
step (2): before and after removing bidirectional edges (c)+(d)). The red arrows indicate how robots change
their positions during the preprocessing steps.
Step (1): We observe that edge crossings only occur between two diagonal edges with adjacent source
tiles, as illustrated in Figure 2(a)+(b). To remove a crossing, it suffices to eliminate one of the diagonal edges
by exchange robots between the source tiles. To eliminate all crossings, each robot is moved at most once,
because after moving, the robot does no longer participate in a diagonal edge. Thus, all necessary exchanges
can be done in O(d) steps by Lemma 2, covering the tiling T by constantly many layers, similar to the proof
of Lemma 2.
Step (2): We delete a bidirectional edge (v, w), (w, v) by moving min{fT ((v, w)), fT ((w, v))} robots with
target tile w from v to w and vice versa which achieves that min{fT ((v, w)), fT ((w, v))} robots achieve their
target tile w and min{fT ((v, w)), fT ((w, v))} robots achieve their target tile v, thus eliminating the edge
with lower flow value. This process is depicted in Figure 2(c)+(d). Like step (1), this can be done in O(d)
parallel steps by Lemma 2. As we do not add any edges, we maintain planarity during step (2). Observe that
during the preprocessing, we do not destroy the grid structure of GT .
Step (1) and step (2) maintain the flow property of fT without any other manipulations to the flow fT ,
because both preprocessing steps can be represented by local circulations.
3.2.3
Step 3: Computing a Flow Partition
After preprocessing, we partition the flow GT into d-subflows.
Definition 5. A subflow of GT is a circulation G0T = (T, E 0 , fT0 ), such that E 0 ⊆ ET , and 0 ≤ fT0 (e) ≤ fT (e)
for all e ∈ E 0 . If fT0 (e) ≤ z for all e ∈ E 0 and some z ∈ N, we call G0T a z-flow.
8
The flow partition relies on an upper bound on the maximal edge weight in GT . By construction, tiles
have side length at most 24d; therefore, each tile consists of at most 576d2 unit squares. This yields the
following upper bound; a tighter constant factor can be achieved using a more sophisticated argument.
Observation 6. We have fT (e) ≤ 576d2 for all e ∈ ET .
Definition 7. A (z, `)-partition of GT is a set of ` z-subflows {G1 = (V1 , E1 , f1 ), . . . , G` = (V` , E` , f` )} of
GT , such that G1 , . . . , G` sum up to GT .
Lemma 8. We can compute a (d, O(d))-partition of GT in polynomial time.
Proof sketch. In a slight abuse of notation, throughout this proof, the elements in sets of cycles are not
necessarily unique. A (d, O(d))-partition can be constructed using the following steps.
• We start by computing a (1, h)-partition C of GT consisting of h ≤ n1 n2 cycles. This is possible
because GT is a circulation. If a cycle C intersects itself, we subdivide C into smaller cycles that are
intersection-free. Furthermore, h is clearly upper bounded by the number of robots n1 n2 , because
every robot can contribute only 1 to the sum of all edges in GT . As the cycles do not self-intersect, we
can partition the cycles C by their orientation, obtaining the set C of clockwise and the set C of
counterclockwise cycles.
• We use C and C to compute a (1, h0 )-partition C1 ∪ C2 ∪ C1 ∪ C2 with h0 ≤ n1 n2 , such that
two cycles from the same subset C1 , C2 , C1 , or C2 share a common orientation. Furthermore, we
guarantee that two cycles from the same subset are either edge-disjoint or one lies nested in the other.
A partition such as this can be constructed by applying a recursive peeling algorithm to C and C as
depicted in Figure 3, yielding a decomposition of the flow induced by C into two cycle sets C1 and
C2 , where C1 consists of clockwise cycles and C2 consists of counterclockwise cycles, and a similar
partition of C , see the appendix for details.
Figure 3: Recursive peeling of the area bounded by the cycles from C , resulting in clockwise cycles (thick
black cycles). Cycles constituting the boundary of holes are counterclockwise (thick red cycles). Note that an
edge e vanishes when fT (e) cycles containing that edge are removed by the peeling algorithm described above.
• Afterwards, we partition each set C1 , C2 , C1 , and C2 into O(d) subsets, each inducing a d-subflow of
GT , see the appendix for details.
3.2.4
A Subroutine of Step 4: Realizing a Single Subflow
In this section, we present a procedure for realizing a single d-subflow G0T of GT .
Definition 9. A schedule t := C1 → · · · → Ck+1 realizes a subflow G0T = (T, E 0 , fT0 ) if, for each pair v, w of
tiles, the number of robots moved by t from their start tile v to their target tile w is fT0 ((v, w)), where we let
fT0 ((v, w)) = 0 if (v, w) ∈
/ E0.
9
Lemma 10. Let G0T = (T, ET0 , fT0 ) be a planar unidirectional d-subflow. There is a polynomial-time algorithm
that computes a schedule C1 → · · · → Ck+1 realizing G0T for a constant k ∈ O(1).
Proof sketch. We give a high-level description of the proof and refer to Section 6.5 for details.
54 321
u
w
2
2
1
2
3
v
2
3
5
4
4
2
3
678
3
3
(a) Preprocessing of diagonal
edges.
(b) Configuration and flow after preprocessing.
(c) A crossing-free matching of incoming
and outgoing robots and the connecting
paths inside the corresponding tile, for
d = 3.
Figure 4: Procedure for computing transformation steps that realize a d-subflow. Figures (a) and (b) illustrate
how we preprocess G0T such that ET0 consists of horizontal and vertical edges only. Figure (c) illustrates the
main approach. White disks illustrate start positions and black disks illustrate target positions.
Our algorithm uses k = O(d) preprocessing steps C1 → · · · → Ck , as depicted in Figure 4(a)+(b), and
one final realization step Ck → Ck+1 , shown in Figure 4(c), pushing the robots from their start tiles into
their target tiles. The preprocessing eliminates diagonal edges and places the moving robots next to the
border of their target tiles. For the final realization step we compute a pairwise disjoint matching between
incoming and outgoing robots, such that each pair is connected by a tunnel inside the corresponding tile in
which these tunnels do not intersect, see Figure 4(a). The final realization step is given via the robots’ motion
induced by pushing each robot into the interior of the tile and by pushing this one-step motion through the
corresponding tunnel into the direction of the corresponding outgoing robot.
3.2.5
Step 4: Realizing All Subflows
Next we extend the idea of Lemma 10 to ` ≤ d subflows instead of one and demonstrate how this can be
leveraged to move all robots to their target tile using O(d) transformation steps.
Lemma 11. Let S := hG1 = (V1 , E1 , f1 ), . . . , G` = (V` , E` , f` )i be a sequence of ` ≤ d unidirectional
planar d-subflows of GT . There is a polynomial-time algorithm computing O(d) + ` transformation steps
C1 → · · · → Ck+` realizing S.
Proof sketch. We give a high level description of the proof and refer for details to Section 6.6.
Let t be an arbitrary tile. Similar to the approach of Lemma 10, we first apply a preprocessing step
guaranteeing that the robots to be moved into or out of t are in the right position close to the boundary of t,
see Figure 5. Thereafter we move the robots into their target tiles, using ` applications of the algorithm from
Lemma 10 without the preprocessing phase. In particular, we realize a sequence of ` d-subflows by applying `
times the single realization step of Lemma 10.
10
(a)
(b)
(c)
(d)
(e)
Figure 5: Stacking robots in lines induced by flows of the edges of the subflows to be realized.
Lemma 12. There is a polynomial-time algorithm computing O(d) transformation steps moving all robots
into their target tiles.
Proof. By Lemma 8, we can compute a (d, cd)-partition of GT for c ∈ O(1). We group the corresponding
d-subflows into cd
d = c sequences, each consisting of at most d d-subflows. We realize each sequence by
applying Lemma 11, using O(d) transformation steps for each sequence. This leads to O(cd) = O(d) steps
for realizing all sequences of d-subflows.
For the proof of Theorem 3, we still need to analyze the time complexity of our approach, for which we
refer to Section 6.7.
4
Variants on Labeling
A different version is the unlabeled variant, in which all robots are the same. A generalization of both this
and the labeled version arises when robots belong to one of k color classes, with robots from the same color
class being identical.
We formalize this problem variant by using a coloring c : {1, . . . , n1 n2 } → {1, . . . , k} for grouping the
robots. By populating unoccupied cells with robots carrying color k + 1, we may assume that each unit
square in the environment P is occupied. The robots draw an image I = I 1 , . . . , I k , where I i is the set of
cells occupied by a robot with color i. We say that two images Is and It are compatible if in Is and It the
number of cells colored with color i are equal for each color i = 1, . . . , k. By moving the robots, we want to
transform a start image Is into a compatible target image It , minimizing the makespan.
Theorem 13. There is an algorithm with running time O(k(N )1.5 log(N ) + N 5 ) for computing, given start
and target images Is , It with maximum distance d between start and target positions, an O(1)-approximation
of the optimal makespan M and a corresponding schedule.
The basic idea is to transform the given unlabeled problem setting into a labeled problem setting by
solving a geometric bottleneck matching problem, see the appendix for details.
5
Continuous Motion
The continuous case considers N unit disks that have to move into a target configuration; the velocity of
each robot is bounded by 1, and we want to minimize the makespan. For arrangements of disks that are not
well separated, we show that constant stretch is impossible.
Theorem 14. There is an instance with optimal makespan M ∈ Ω(N 1/4 ), see Figure 16.
The basic proof idea is as follows. Let {m1 , . . . , mN } be an arbitrary trajectory set with makespan M . We
show that there must be a point in time t ∈ [0, M ] where the area of Conv(m1 (t), . . . , mN (t)) is lower-bounded
by cN + Ω(N 3/4 ), where
cN is the area of the convex hull Conv(m1 (0), . . . , mN (0)) of m1 (0), . . . , mN (0).
Assume M ∈ o N 1/4 and consider the area of Conv(m1 (t0 ), . . . , mN (t0 )) at some point t0 ∈ [0, M ]. This
√
area is at most cN + O( N ) · o N 1/4 which is a contradiction. Proof details are given in Section 8.1.
Conversely, we give a non-trivial upper bound on the stretch, as follows.
11
right auxilliaries
xn−1
false
true
...
...
...
...
...
...
clauses
...
left auxilliaries
xn−1
false
x1 : (0, 6)
true
aux.
variables
false
x0 : (0, 0)
...
x0 : (M − 2, 0)
(1, −M + 1)
...
true
...
aux.
side steps
x1 : (M − 2, 6)
...
(M − 3, −M + 1)
checkers
Figure 6: Left: The structure of the resulting parallel motion planning problem instance. Right: Start
configuration (disks) and target configuration (crosses) of variable robots and their auxiliaries. The left
auxiliary robot for xj starts at position (1, 6j + 1) and has to move down towards its target position
(1, 6j + 1 − M ) in each time step. The right auxiliary robot for xj starts at (M − 3, −M + 6j + 1) and has to
move up towards its target position (M − 3, 6j + 1).
√
Theorem 15. There is an algorithm
that computes a trajectory set with continuous makespan of O(d + N ).
√
If d ∈ Ω(1), this implies a O( N )-approximation algorithm.
√
The approach of Theorem 15 applies an underlying grid with mesh size 2 2. Our algorithm (1) moves
the robots to vertices of the grid, (2) applies our O(1)-approximation for the discrete case, and (3) moves the
robots from the vertices of the grid to their targets. For a detailed description of the Algorithm of Theorem 15
see Section 8.2.
6
Details for Labeled Grid Permutation
In this section, we give the details omitted in the high-level description of our results for the problem variant
of labeled grid permutations that we considered in Section 3.
6.1
Details for the NP-Completeness of the Grid Case
Theorem 1. The minimum makespan parallel motion planning problem on a grid is strongly NP-hard.
Proof. The proof is based on a reduction from the NP-hard problem Monotone 3-Sat, which asks to decide
whether a Boolean 3-CNF formula ϕ is satisfiable, where in each clause the literals are either all positive or
all negative. All coordinates and the makespan are constructed to be polynomial in the input size, implying
strong NP-hardness. Thus, there is no FPTAS for the problem, unless P = NP.
For the remainder of the proof, let ϕ have n variables {x0 , . . . , xn−1 } and m clauses {C1 , . . . , Cm }. From
ϕ, we construct an instance of the minimum makespan parallel motion planning problem that has optimum
makespan M if ϕ is satisfiable and M + 1 otherwise. During the description of the construction, we keep M
variable, fixing its value once the construction is complete. The structure of the resulting instance is sketched
in Figure 6.
Each variable xj is represented by a variable robot. Additionally, for each variable there are two auxiliary
robots that force the variable robot to take one of two different paths to its goal in any solution with makespan
M , see Figure 6. The left auxiliary robots start at positions (1, 6j + 1) and move down towards their target
positions (1, 6j + 1 − M ) in each time step. The right auxiliary robots start at positions (M − 3, −M + 6j + 1)
12
and have to move up towards their target positions (M − 3, 6j + 1). The variable robot for variable xj starts
at position (0, 6j) and has to travel M − 2 units to the right towards its goal position (M − 2, 6j). In the
first time step, each variable robot can either wait or move upwards. Afterwards, it must move to the right in
every time step until passing the right group of auxiliary robots at x = M − 3. It cannot wait or move down
before this point, as this would lead to a collision with the corresponding right auxiliary robot. Therefore in
any schedule with makespan M , after the kth time step, each variable robot has x-coordinate k − 1 for any
1 ≤ k ≤ M − 3.
For each clause Ci = {xj1 , xj2 , xj3 } with j1 < j2 < j3 , we have three checker robots c1i , c2i , c3i checking
whether their corresponding literal satisfies the clause. The
checkers for clause Ci start at positions
αi1 := 6(ni + j1 ), −6ni − fi , . . . , αi3 := 6(ni + j3 ), −6ni − fi , where fi = 1 iff Ci is negative and fi = 0
otherwise. As depicted in Figure 7, a checker has to wait one time step for the corresponding variable iff the
checked literal is not true.
}
true
{z
}
6(ni + j)
...
{z
6(ni + j)
false
...
6(ni + j) + 1
...
c`i
(a)
|
...
c`i
...
}
xj |
true
...
{z
{z
}
6(ni + j)
|
...
xj |
false
...
(b)
Figure 7: (a): A checker c`i for variable xj in a positive clause Ci . (b): A checker c`i for variable xj in a
negative clause Ci . Checkers must wait iff the variable assignment does not match.
Checker c3i has to move M − 1 units up to its target position t3i := αi3 + (0, M − 1). Let d1 := 6(j3 − j1 ) be
the horizontal distance between the initial positions of c1i and c3i , and let d2 := 6(j3 − j2 ) analogously. Both
d1 and d2 are always even and at least six; therefore s1 := d21 + 2 < d1 and s2 := d22 + 1 < d2 are integer. We
force c1i to take s1 steps to the right towards its target position t1i := αi1 + (s1 , M − 1 − s1 ). Analogously, c2i
has target position t2i := αi2 + (s2 , M − 1 − s2 ). Each checker travels a total distance of M − 1; thus they are
allowed to wait for one time step, but have to move on an xy-monotone path towards their target position.
Because moves to the right do not change the position of a checker relative to the variables, we may
assume the checkers to move to the right from their initial position before moving up. In fact, we enforce this
behavior using auxiliary robots as depicted in Figure 8. Moreover, each clause Ci also has a clause robot
ensuring that there is at least one satisfied literal. The clause robots start to the right of the checkers and
above the variables and have to move M − 2 units to the left and two units downwards, and therefore have to
move towards their target in every round without waiting for the checkers. The clause robot of each clause is
placed such that checkers for other clauses cannot interfere with its path, see Figure 6. To be more precise,
as shown in Figure 8, the clause robot stops at position t1i − (3, 3) and starts at position t1i + (−1, M − 5).
The vertical offset between the checkers introduced by the side steps that c1i and c2i perform is chosen such
that the clause robot can pass through the checkers without waiting iff one of the checkers did not wait. This
is the case iff at least one literal of the clause is satisfied.
It remains to determine the critical makespan M . This critical makespan M must be large enough to
allow the checkers of the last clause Cm to pass through the variable robots and their clause robot. Moreover,
it must also allow the variable robots to cross paths with all checkers. The checkers of the last variable travel
left of the line x = 6n(m + 1) − 6. Therefore, a makespan M ≥ 6n(m + 1) suffices for the variable robots.
Regarding the clauses, if the last clause is negative, the starting points of its checkers are located on the line
y = −6nm − 1. The topmost variable robot travels below the line y = 6(n − 1) + 1. To keep our argument
13
d1
−
2 }|
z
z
s1 − 1
}|
t1i
{
d2
2
−1
}| {
z
z
3rd literal true
2nd literal true
}|
d1
2
+2
d2
2
+1
1st literal true
{
xj3
}|
xj2
{
z
s −1
z 2}| {
xj1
2
{
Figure 8: Left: A group of auxiliary robots is used to force the first two checkers of each clause to perform
their side steps before moving up. Each auxiliary robot has to move downwards M units. Right: A clause
robot (orange) meeting the corresponding checkers (black for satisfied checkers, red for non-satisfied checkers).
simple, we want to make sure that the clauses stay strictly above all variables. Due to the position of the
clauses, this means that we have to ensure that the checker for the first literal of the last clause has target
position above the line y = 6(n − 1) + 5. Therefore, not accounting for the side steps of the checkers, we have
to set M ≥ (6nm + 1) + (6(n − 1) + 5) = 6n(m + 1). Clearly, the number of side steps performed by each
checker is less than 6n. Therefore, in total, a critical makespan of M := 6n(m + 2) is sufficient.
In our construction, a makespan of M is feasible iff for every clause robot there is one checker that does
not wait, which implies that each clause has a satisfied literal under the assignment induced by the variable
robots. Therefore, a makespan of M is feasible iff ϕ is satisfiable.
Finally, observe that even though our reduction uses individually labeled robots, three colors are already
sufficient. One can use color 1 for variables, color 2 for checkers and color 3 for clauses and all auxiliaries.
6.2
Details on Computing a Schedule With a Makespan of O(n1 + n2 )
Next, we give the details of an algorithm that computes a sequence of O(n1 + n2 ) steps transforming an
arbitrary start configuration Cs into an arbitrary target configuration Ct of an n1 × n2 rectangle, see Lemma 2.
This algorithm is based on a sorting algorithm, called RotateSort that uses swap operations, in which two
robots exchanging their positions within one single step, as elementary operations. As our model does not
allow swap operations, we first have to show how to simulate swap operations at the expense of increasing
the makespan by a factor upper-bounded by some constant.
In order to simulate swap operations, we first observe that LaValle and Yu [74] proved that for a 3 × 3square, each start configuration can be transformed into an arbitrary target configuration. This result is
easily established for 2 × 3-rectangles; see Figure 9 for how to realize a transposition.
14
1
4
2
5
3
6
2
1
5
4
3
6
2
1
3 6
5 4
1
5
2 3
4 6
Figure 9: Using three moves for swapping two positions in a 2 × 3-arrangement.
Lemma 16. For a pair of start and target configurations Cs and Ct of a 2 × 3-rectangle, we can compute a
sequence of at most seven steps transforming Cs into Ct .
Lemma 16 is the building block for permuting n1 × n2 rectangles within makespan O(n1 + n2 ).
Lemma 2. For a pair of start and target configuration Cs and Ct of an n1 × n2 -rectangle, we can compute in
polynomial time a sequence of O(n1 + n2 ) steps transforming Cs into Ct .
Proof. The straightforward proof relies on covering the rectangle by a set of disjoint 2 × 3- and 3 × 2-rectangles,
on which swap operations are performed in parallel, with each swap operation exchanging the position of
two adjacent robots. We say that two swap operations are disjoint if all four positions of the two swaps
are distinct. Although direct swap operations of adjacent robots are not possible, Lemma 16 allows us to
perform an arbitrary number of pairwise disjoint swap operations within each 2 × 3-rectangle with O(1)
transformation steps. As illustrated in Figure 10, we cover P by twelve different layers of rectangles, such that
each pair of adjacent unit squares from P lies in one of the 2 × 3-rectangles or in one of the 3 × 2-rectangles.
In particular, we distinguish between 2 × 3- and 3 × 2-rectangles inside the n1 × n2 -rectangle. Furthermore,
we distinguish between different positions of 2 × 3-rectangles w.r.t. line numbers modulo 2 and w.r.t. column
numbers modulo 3; see Figures 10a)-f). Analogously, we distinguish between different positions of 3 × 2rectangles w.r.t. line numbers modulo 3 and w.r.t. column numbers modulo 2; see the Figures 10g)-l). This
results in twelve different classes of rectangles.
a)
b)
c)
d)
e)
f)
g)
h)
i)
j)
k)
l)
Figure 10: Covering of P by pairwise disjoint 2 × 3- and 3 × 2-rectangles in twelve layers.
Given a set S of pairwise disjoint swap operations, we subdivide S into these twelve layers, such that
the two robots of each swap operation lie in the same small rectangle of the corresponding layer. Lemma 16
implies that all swap operations of one layer can be done in parallel with O(1) transformation steps. Therefore,
all swap operations in S can be done in O(1) transformation steps.
This allows us to apply a sorting algorithm for n1 × n2 -meshes, called Rotatesort [43], whose only
elementary steps are swap operations of adjacent cells. We employ Rotatesort by labeling the robots in the
target configuration based on the snake-like ordering guaranteed by Rotatesort. Applying Rotatesort
to the start configuration with the robots labeled in this way, we obtain the required target configuration.
Marberg and Gafni [43] show that Rotatesort needs O(n1 + n2 ) phases, where each phase consists of
pairwise disjoint swap operations. This leads to O(n1 + n2 ) transformation steps in our model.
Figure 11, shows an example of a pair of start and target configuration and the resulting flow.
6.3
Details on the Approach of Theorem 4
15
1
1
1
1
1
1
2
2
1
Figure 11: A tiling of an 26 × 32-rectangle by four tiles with d = 1. Robots not in their target tile are
illustrated by small dots. Their target positions are depicted as white disks. The dual graph of the tiling is
illustrated by large dots and directed edges between them. The edges of the dual graph are annotated with
the value of the flow on the corresponding edge. In general, it is not guaranteed that robots that have to
change the tile lie adjacent to the border between their start and target tile. However, this is the case for
d = 1, as illustrated in this figure.
Theorem 4. There is an algorithm with running time O(N 5 ) that, given an arbitrary pair of start and
target configurations of a rectangle P with N robots to be moved and maximum distance d between any start
and target position, computes a schedule of makespan O(d), i.e., an approximation algorithm with constant
stretch.
Proof. Our algorithm considers the two cases (1) N ≤ d n41 e, d and (2) N > d n41 e or N > d separately as
follows:
In case (1), we apply the following approach whose steps, described next, are all realizable because
N ≤ d n41 e, d. We assume w.l.o.g. that n1 and n2 are even. Otherwise, starting from the start configuration,
we move all robots from the last line into the second-to-last line and all robots from the last column into
the second-to-last column within O(d) transformation steps. The reversed argument implies that there is a
sequence of O(d) transformation steps leading from an even-sized configuration to the target configuration.
Thus, from now on, we restrict our considerations to even-sized rectangles.
For each pair of start and target configurations Cs and Ct of P , there are two configurations Co and Ce ,
such that the two following conditions are fulfilled: (1) The coordinates of the robots in Co are odd and the
coordinates of the robots in Ce are even and (2) Cs and Ce can be transformed into Co and Ct within O(d)
transformation steps. Thus, we still have to give an approach for how Co can be transformed into Ce within
O(d) transformation steps.
First of all, we ensure in parallel for all robots that they achieve the position that is induced by the
x-coordinate of their position in Ce and the y-coordinate of their position in Co . We call the corresponding
configuration intermediate configuration Ci with intermediate positions and coordinates. In order to obtain
the intermediate configuration, starting from Co , we first push in parallel all robots, that have to move to the
right, one position upwards, then move them simultaneously to the right until they achieve their intermediate
x-coordinate, and, push a robot immediately one position downwards when it reaches its intermediate
x-coordinate, see Figure 12.
After that, we apply the analogous approach for robots that have to move to the left, resulting in the
intermediate configuration.
Secondly, starting from the intermediate configuration, we apply the above described two-stepped approach
for horizontal movements in an analogous version in order to ensure that the y-coordinate of each robot r is
equal to the y-coordinate of r in the configuration Ce , while guaranteeing that the x-coordinate of r stays
the same. This results in the configuration Ce .
16
The start and
and target
configuration
Cs and Ct
6
5
4
Cs → . . . → Ct
3
2
1
1
Preprocessing
to ensure odd
coordinates
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
6
5
4
Cs → . . . → Co
3
2
1
Moving
robots
in eastern
direction
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
6
5
4
3
2
1
Moving
robots
in western
direction
6
Co → . . . → Ci
5
4
3
2
1
Moving
robots
in northern
direction
6
5
4
3
2
1
Moving
robots
in southern
direction
6
C i → . . . Ce
5
4
3
2
1
Postprocessing
to generate
target
configuration
from even
coordinates
6
5
4
Ce → . . . → Ct
3
2
1
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
Figure 12: A stepwise illustration of the approach for case (1) of Theorem 4.
The transformation steps leading from Cs to Co and leading from Ce to Ct can be computed in O(N · N )
time by making use of the fact that the robots’ positions are explicitly given via their coordinates. The same
reasoning implies that the sequences of transformation steps leading from Co to the intermediate configuration
and leading from the intermediate configuration to Ce can be computed in O(N · N ) time.
In case (2), we apply the approach of Theorem 3 as a subroutine in the following approach: For each
robot we consider the smallest rectangle that contains the robot’s start and target positions. If the rectangle
has a height or width of 1, we extend the height or width to 2. Now we iteratively replace two rectangles R1
and R2 intersecting each other by the smallest rectangle that contains R1 and R2 .
This results in a set of rectangles that are pairwise intersection free, allowing us to apply the approach of
Theorem 3 to each resulting rectangle in parallel, while ensuring that each robot is involved in at most one
application of the approach of Theorem 3.
As the side lengths of the initial rectangles are upper-bounded by d, we conclude that the sum of the
lengths of the finally computed rectangles is upper bounded by N · d, which in turn is upper bounded by N 2
in that case. This implies a running time of O(d · N 2 ) ≤ O(N 3 ).
17
Figure 13: An illustration how the approach for case (2) of Theorem 4 clusters the pairs of robots’ start and
target positions.
6.4
Details on Step 3: Computing a Flow Partition
Lemma 8. We can compute a (d, O(d))-partition of GT in polynomial time.
Proof. In a slight abuse of notation, throughout this proof, the elements in sets of cycles are not necessarily
unique. A (d, O(d))-partition can be constructed using the following steps.
• We start by computing a (1, h)-partition C of GT consisting of h ≤ n1 n2 cycles. This is possible
because GT is a circulation. If a cycle C intersects itself, we subdivide C into smaller cycles that are
intersection-free. Furthermore, h is clearly upper bounded by the number of robots n1 n2 , because
every robot can contribute only 1 to the sum of all edges in GT . As the cycles do not self-intersect, we
can partition the cycles C by their orientation, obtaining the set C of clockwise and the set C of
counterclockwise cycles.
• We use C and C to compute a (1, h0 )-partition C1 ∪ C2 ∪ C1 ∪ C2 with h0 ≤ n1 n2 , such that
two cycles from the same subset C1 , C2 , C1 , or C2 share a common orientation. Furthermore, we
guarantee that two cycles from the same subset are either edge-disjoint or one lies nested in the other.
A partition such as this can be constructed by applying a recursive peeling algorithm to C and C as
depicted in Figure 3, yielding a decomposition of the flow induced by C into two cycle sets C1 and
C2 , where C1 consists of clockwise cycles and C2 consists of counterclockwise cycles, and a similar
partition of C . In particular, we apply the following approach iteratively to C : We consider the
union A of the area bounded by the cycles from C . We remove a flow value of 1 from all edges of
the outer boundary component of A. In particular, we add the corresponding 1-subflow G1 to C1 and
remove G1 from C . Analogously, we remove 1-subflows from C that are induced by inner boundary
components and add these 1-subflows to C2 .
• Afterwards, we partition each set C1 , C2 , C1 , and C2 into O(d) subsets, each inducing a d-subflow of
GT . This can be done as follows. Let C ∈ {C1 , C2 , C1 , C2 }. Recall that every pair of cycles from C
either consists of one cycle nested inside the other or of edge-disjoint cycles. The cycles induce a dual
forest D = (C, ED ), where a cycle v has a child w iff w lies inside v and there is no other cycle lying in
v that w lies in. We label the cycles by their depth in D modulo 576d and let Gi be the flow induced
by all cycles carrying label i, thus obtaining O(d) subflows Gi .
Finally, we show that each subflow Gi obtained in this way is a d-subflow of GT . To this end, we observe the
following. Let e ∈ ET be an arbitrarily chosen edge and let v, w ∈ C be two cycles sharing e. This implies
18
Figure 14: Recursive peeling of the area bounded by the cycles from C , resulting in clockwise cycles (thick
black cycles). Cycles constituting the boundary of holes are counterclockwise (thick red cycles). Note that an
edge e vanishes when fT (e) cycles containing that edge are removed by the peeling algorithm described above.
that v and w lie nested inside of each other; w.l.o.g., assume that w lies inside v. Thus, in D, v lies on the
path from w to its root, and e is contained in all cycles on the path between v and w. On the other hand,
due to Observation 6, all cycles containing e lie on a path of length at most 576d2 in D. Therefore, e has a
2
weight of at most 576d
576d = d in each Gi , and Gi is a d-subflow.
6.5
Details on a Subroutine of Step 4: Realizing a Single Subflow
By Lemma 10, we give an approach that computes a schedule of constant length for a given d-subflow.
Lemma 10. Let G0T = (T, ET0 , fT0 ) be a planar unidirectional d-subflow. There is a polynomial-time algorithm
that computes a schedule C1 → · · · → Ck+1 realizing G0T for a constant k ∈ O(1).
Proof. Our algorithm uses k = O(d) preprocessing steps C1 → · · · → Ck , as depicted in Figure 4(a)+(b),
and one final realization step Ck → Ck+1 , shown in Figure 4(c), moving the robots from their start tiles
into their target tiles. The preprocessing replaces diagonal edges by pairs of orthogonal edges, see the red
arrows in Figure 4(a), and places the moving robots next to the border of their target tiles. Note that the
replacements of the diagonal edges cannot be done as part of the preprocessing of Step 2. This is because the
replaced diagonal edges may be part of circular flows that cannot be realized locally, as it is done for crossing
or bidirectional edges in Step 2 of our algorithm.
For the final realization step we compute a pairwise disjoint matching between incoming and outgoing
robots, such that each pair is connected by a tunnel inside the corresponding tile in which these tunnels do
not intersect each other, see Figure 4(a). The final realization step is given via the robots’ motion induced by
moving each robot into the interior of the tile and by moving this one-step motion through the corresponding
tunnel into the direction of the corresponding outgoing robot.
The preprocessing steps C1 → · · · → Ck : Let v be an arbitrary tile. We place all robots corresponding to
horizontal and vertical edges (v, w) of G0T in a row adjacent to the side shared by v and w. We can do this
for all tiles using O(d) parallel steps by applying Lemma 2.
Next, we eliminate diagonal edges (w, v) ∈ ET0 as follows. There are two tiles sharing a side with both w
and v; let u be one of them. First we place the fT0 ((w, v)) robots with start tile w and target tile v in a row
next to the side between w and u. Then, we move them to u by exchanging them with fT0 ((w, v)) robots with
start and target tile u that lie next to the side between u and v, as shown in Figure 4(a). In the resulting
flow, the diagonal edge (w, v) with weight fT0 ((w, v)) is replaced by adding a flow of value fT0 ((w, v)) on the
edges (w, u), (u, v).
We process all tiles as described above in two parallel phases by applying Lemma 2 twice: first on all rows
with even index and then on all rows with odd index, thus ensuring that parallel applications of Lemma 2 do
not interfere with each other.
19
The realization step Ck → Ck+1 : Let t be an arbitrary tile. For the transformation step Ck → Ck+1 , we
need a matching between incoming and outgoing robots of t, such that there is a set of non-intersecting paths
in t connecting each incoming robot with its corresponding outgoing robot. As illustrated in Figure 4(c),
these paths induce the required transformation Ck → Ck+1 .
We compute this matching by selecting an incoming robot rin and matching it to a robot rout , such that
there is a path p ⊆ ∂t between rin and rout that does not touch another incoming or outgoing robot. We
remove the matched robots from consideration and repeat the matching procedure until no further unmatched
robots exist.
The non-intersecting paths between the positions of the matched robots are constructed as follows. For
i ≥ 1, the ith hull of t is the union of all squares on the boundary of the rectangle remaining after the hulls
1, . . . , i − 1 are removed. The path between rin and rout consists of three pieces, as shown in Figure 4(c). For
the ith matched pair of robots, the initial and the last part of the path are straight line segments orthogonal
to ∂t, from the position of rin to the d + ith hull and from the d + ith hull towards the position of rout . The
main part of the path lies on the d + ith hull, connecting the end of the initial part to the beginning of the
last part.
6.6
Details on Step 4: Realizing All Subflows
Lemma 11. Let S := hG1 = (V1 , E1 , f1 ), . . . , G` = (V` , E` , f` )i be a sequence of ` ≤ d unidirectional
planar d-subflows of GT . There is a polynomial-time algorithm computing O(d) + ` transformation steps
C1 → · · · → Ck+` realizing S.
Proof. Let t be an arbitrary tile. Similar to the approach of Lemma 10, we first apply a preprocessing step
guaranteeing that the robots to be moved into or out of t are in the right position close to the boundary of t.
Thereafter we move the robots into their target tiles, using ` applications of the algorithm from Lemma 10
without the preprocessing phase. In particular, we realize a sequence of ` d-subflows by applying ` times the
single realization step of the algorithm from Lemma 10.
In order to ensure that a sequence of ` realization steps from Lemma 10 without intermediate preprocessing
steps realizes a sequence of ` d-subflows, we apply the following O(d) preprocessing steps for all ` realization
steps in advance: For each side of the tile t, we place all leaving or entering robots that belong to the same
subflow in a common row and stack these rows in the order which is induced by the sequence of the subflows
to be realized, see Figure 15(a). Finally, pushing all stacked robots downwards into the direction of the
boundary ∂t of the tile ensures, that processing one realization step implies that all robots involved in the
following realization step lie in a row adjacent to ∂t, see Figure 15.
In the following we describe how we place the robots in their start tiles as a preprocessing step. First, we
use the same preprocessing step as in Lemma 10 to eliminate diagonal edges. For a simplified illustration,
we describe the remainder of the preprocessing in two steps that can be realized by just one application of
Lemma 2. After elimination of diagonal edges, we proceed by stacking the rows of robots moving out of t in
the order in which the subflows are to be processed, see Figure 15(a). Then we push the robots towards the
boundary ∂t of their start tile until they meet either ∂t or another moving robot. See Figure 15(a), image 2
for an example.
This preprocessing ensures that, after each application of the algorithm of Lemma 10, all robots moving
out of t in the next transformation step lie in a row adjacent to ∂t. Therefore this preprocessing can be used
to replace the preprocessing done in Lemma 10. For an example, see Figure 15(a), images 3–9.
As ` ≤ d, the stacked rows have a height of at most d. Thus, they are contained in hulls 1 to d. Therefore,
and because the flows are unidirectional and diagonals are eliminated, the structure of the stacks is not
damaged by the applications of Lemma 10, allowing us to realize ` ≤ d subflows in O(d) transformation steps
instead of one.
20
45 321
678
(b) Our preprocessing step applied to the example
(a) Stacking the rows of robots corresponding to the flow values of Figure 4 and the path matching (red) of the first
on the edges of the subflows to be realized.
step.
Figure 15: Realizing a sequence of subflows by stacking the rows of robots to be moved onto each other in
the order the subflows are realized in.
6.7
Runtime Analysis of the Algorithm of Theorem 3
Theorem 3. There is an algorithm with running time O(dn1 n2 ) that, given an arbitrary pair of start and
target configurations of an n1 × n2 -rectangle with maximum distance d between any start and target position,
computes a schedule of makespan O(d), i.e., an approximation algorithm with constant stretch.
Proof of Theorem 3. The steps of our algorithm have the following time complexity:
Initialization step 1: Computing d, T and GT is possible in O(n1 n2 ) time.
Step 2 & 5: The application of Lemma 2 requires O(d3 ) time for each tile, so these steps can be done in
O(dn1 n2 ) time.
Step 3: All subroutines of Step P
3 can be done in an overall time of O(n1 n2 ). In particular, the (1, h)-partition
C of GT can be computed in e∈ET fT (e) ∈ O(n1 n2 ) time by a simple greedy algorithm. The number of
edges in all cycles from C combined is at most n1 n2 , which is the number of robots in P . Thus, resolving
self-intersections of cycles in C can be done in O(n1 n2 ) time. As |C | ∈ O(n1 n2 ), the partition of C into
C1 , C2 , C1 , and C2 takes time O(n1 n2 ). Furthermore, the partitioning of C1 , C2 , C1 , and C2 into O(d)
d-subflows can be done in time O(n1 n2 ).
Step 4: The parallel applications of Lemma 2 to disjoint rectangles can be computed in O(dn1 n2 ). Furthermore, the construction of all connecting paths between incoming and outgoing robots for all tiles needs
O(dn1 n2 ) time per application of the algorithm of Lemma 11. By applying Lemma 11 constantly many times,
Step 4 needs O(dn1 n2 ) time.
21
7
Details for Variants on Labeling
In this section we show how to extend our approach of Section 3 to the unlabeled and, more generally, the
colored variant of the parallel robot motion-planning problem.
Theorem 13. There is an algorithm with running time O(k(N )1.5 log(N ) + N 5 ) for computing, given start
and target images Is , It with maximum distance d between start and target positions, an O(1)-approximation
of the optimal makespan M and a corresponding schedule.
Proof. We transform the input into an instance of the labeled variant, such that an O(1)-approximation for
the labeled instance provides an O(1)-approximation for the colored instance. For each color i, we consider
the two point sets Ai , B i ⊂ R2 , where Ai contains the center points aiv of all unit squares v ∈ Isi and B i
contains the center points biv of all v ∈ Iti .
A bottleneck matching between Ai and B i is a perfect matching between Ai and B i that minimizes the
maximal distance. The cost of an optimal bottleneck matching between Ai and B i is in O(M ), because a
transformation sequence induces a bottleneck matching on all color classes. Efrat et al. [20] show that the
geometric bottleneck matching problem can be solved in O(|A + B|1.5 log |A + B|) time.
A set of k bottleneck matchings between the sets Ai and B i induces labeled start and target configurations
Cs , Ct . Applying the algorithm from Section 3 to these yields a sequence of transformation steps of length
O(M ).
8
Details for Continuous Motion
In this section, we consider the continuous geometric case in which the robots are identical geometric objects
that have to move into a target configuration in the plane without overlapping at any point in time. We want
to minimize the makespan under these conditions, where the velocity of each robot is bounded by 1.
8.1
A Lower Bound for Unbounded Environments
In this section we give a worst-case lower bound of Ω(N 1/4 d) for the continuous makespan where N is the
number of robots. To be more precise, we construct a pair of start and target configurations of N robots
as illustrated in Figure 16(a). In this instance, we have d = 2. In Theorem 14, we show that the optimal
continuous makespan of this instance is in Ω(N 1/4 ), yielding the worst-case lower bound stated above.
(a) Start and target positions of
the robots.
(b) Voronoi diagram in the start
and target configuration.
(c) Bounding polygon for the moving robots.
Figure 16: The start and target configurations of our lower-bound construction where an arrow points from a
start position to the corresponding target position.
22
More formally, let {m1 , . . . , mN } be an arbitrary trajectory set with makespan M , realizing the start and
target configurations as illustrated in Figure 16(a). By applying a simple continuity argument, we show that
there must be a point in time t ∈ [0, M ] such that the area of the convex hull Conv(m1 (t), . . . , mN (t)) of
3/4
m1 (t), . . . , mN (t) is lower
is the area of the of Conv(m1 (0), . . . , mN (0)).
bounded by cN + Ω(N ), where cN
1/4
Assume M ∈ o N
and consider the area of Conv(m1 (t0 ), . . . , mN (t0 )) at some point t0 ∈ [0, M ]. This
√
area is at most cN + O( N ) · o N 1/4 , because asymptotically, the area gained during the movement is
bounded by the product of makespan and circumference. This contradicts the lower bound stated above.
A key ingredient for the construction of the time point t ∈ [0, M ] is the fact that the distance between the
centers of two robots change continuously. In fact, we know that the Euclidean distance between two centers
is 2-Lipschitz, because the velocity of the robots is bounded by 1.
Definition 17. A function f : R → R is λ-Lipschitz (continuous) if |f (x) − f (y)| ≤ λ|x − y| holds for all
x, y ∈ R.
Observation 18. For all i, j ∈ R, the distance between the centers mi (·) and mj (·) of robots i and j is
2-Lipschitz.
Let V be the Voronoi diagram of the centers {m1 (M ), . . . , mN (M )} restricted to
Conv(m1 (M ), . . . , mN (M )) in the target configuration, as illustrated in Figure 16(b). For m ∈ {m1 , . . . , mN }
and t ∈ [0, M ], we denote the Voronoi region of m(t) w.r.t. {m1 (t), . . . , mN (t)} by V (m(t)). Let p be
the trajectory of an arbitrary robot not on the convex hull in the target configuration. Furthermore,
let p1 , . . . , p6 ∈ {m1 , . . . , mN } be the trajectories of the six robots 1, . . . , 6 adjacent to p in the target
configuration.
1
In the following, we show that there is a time interval I = [t0 , t0 + 20
] such that the area of V (p(t00 ))
00
is lower bounded by 3.479 for all t ∈ I, see Lemma 21. This is larger than the area of V (p(0)) and
V (p(M )) by a constant factor. Based on that, we construct the time point t ∈ [0, M ] such that the area of
Conv(m1 (t), . . . , mN (t)) is lower bounded by cN + Ω(N 3/4 ), see Lemma 22. To this end, we need to relate
the area of a Voronoi region to the length of the corresponding Delaunay edges.
0
0
0
0
Lemma 19. Let t0 ∈ [0, M
] and p(t◦) ∈ {m1 (t ), . . . , mN (t0 )}. If the maximal distance between p(t ) and its
Voronoi neighbors is λ ∈ 2, 4 cos(50 ) , the area of V (p(t )) is at least
λ
λ
λ
λ
4
3 sin arccos
− tan 90◦ − arccos
+√ ,
4
4
4
4
3
which is√at least 3.479 for λ ∈ [2.1, 2.2]. Furthermore, the area of V (p(M )) in the target configuration is
√6 = 2 3 ≤ 3.465.
3
Proof. Let p1 (t0 ) ∈ {m1 (t0 ), . . . , mN (t0 )} be the center of a robot with |p(t0 )p1 (t0 )| = λ. Because all Voronoi
neighbors of p(t0 ) have distance less than 4 cos(50◦ ), the angle between two Voronoi neighbors of p(t0 ) in p(t0 )
is greater than 50◦ . Thus, p := p(t0 ) has at most six Voronoi neighbors and the area of V (p(t0 )) is minimized if
p(t0 ) has five further Voronoi neighbors p2 (t0 ), . . . , p6 (t0 ). We can assume |p(t0 )p2 (t0 )| = · · · = |p(t0 )p6 (t0 )| = 2
because this does not increase the area of V (p(t0 )). W.l.o.g., let p1 := p1 (t0 ), . . . , p6 := p6 (t0 ) be in
counterclockwise order around p. This situation is depicted in Figure 17.
We find a lower bound on V (p) by lower bounding the intersections of V (p) with the Delaunay triangles
that are adjacent to p, i.e., with the triangles built by the edges p1 p2 , . . . , p5 p6 and p6 p1 with p, see Figure 17.
The area of the two triangles 41 and 46 built by p1 p2 and p6 p1 with p are minimized by assuming the
configuration of Figure 17(a)+(b), i.e., for |p1 p2 | = |p1 p6 | = 2.
In the configuration of Figure 17(a)+(b), we lower bound the area of 46 ∩ V (p) as follows: We subdivide
the area of 46 ∩ V (p) into three subsets A, B, and C, and lower bound the area of 46 ∩ V (p) by the sum of
lower bounds for |A|, |B|, and |C|.
Let u, v, and c be the mid points of pp6 , pp1 , and 46 , see Figure 17(b). Furthermore, let h be
the vertical side length of A and ` be the vertical
side length of B. The interior angle
of 46 at p is
arccos( λ4 ). Thus, we obtain h = sin arccos λ4 , which implies |A| = 12 · λ4 sin arccos λ4 . The interior
23
p3
43
p2 42
41
p1
p
p1
p2
p4
44
46
p6
46
v
c C
B
p
h
`
A
p1
p3
43
p
44
46
p6
45 p5
p6
(b) Lower bound on the area
of 46 ∩ V (p).
p2
42
p3
p4
V (p)
u
(a) Configuration with minimal area of 41 ∩ V (p) and
46 ∩ V (p).
41
42
45 p5
(c) Configuration with minimal area of 4i ∩ V (p) for
i ∈ {2, . . . , 5}.
V (p)
p
(d) Lower bound on
the area of 4i ∩ V (p)
for i ∈ {2, . . . , 5}.
Figure 17: Lower bounding the area of V (p) by lower bounding the sum of the areas of the intersections of
V (p) with the Delaunay triangles 4i for i ∈ {1, . . . , 6} for a maximal distance of 2.5 between p1 and p.
angle of B at c is 90◦ − arccos λ4 . Hence, we get ` = λ4 tan 90◦ − arccos λ4 because the length of B’s
horizontal side is λ4 . Therefore, |B| = 12 · λ4 · λ4 tan 90◦ − arccos λ4 . Finally, we have |C| = λ4 (h − `) =
λ
λ
− λ4 tan 90◦ − arccos λ4
. As 46 ∩ V (p) and 41 ∩ V (p) are symmetric, this gives us a
4 sin arccos 4
lower bound of 2(|A| + |B| + |C|) on |(41 ∪ 46 ) ∩ V (p)|.
Furthermore, the area of (42 ∪ · · · ∪ 45 ) ∩ V (p) is minimized by the configuration, implying the highest
possible packing density, illustrated in Figure 17(c)+(d)
for |p2 p| = · · · = |p6 p| = |p2 p3 | = · · · = |p5 p6 | = 2.
√
Therefore, this area is at least 4· 13 ·|4i | = 4· 31 · 12 ·2· 3 = √43 . All in all, we obtain |V (p)| ≥ 2(|A|+|B|+|C|)+ √43 .
For λ ∈ [2.1, 2.2], this is at least 1.17046 + √43 ≥ 3.479. In the target configuration, we have |p(M )pi (M )| = 2
√
for i ∈ {1, . . . , 6}. Therefore the area of V (p(M )) is √63 = 2 3 ≤ 3.465.
1
Next, we prove that there is a time t0 with an interval I := [t0 , t0 + 20
] during which the area of
Conv(p, p1 , . . . , p6 ) is greater by a constant factor than the area of Conv(p, p1 , . . . , p6 ) in the target configuration. To this end, we use the following observation that is an immediate consequence of the intermediate
value theorem.
Observation 20. There is a time t0 ∈ [0, M ] for which the maximal distance between p(t0 ) and p1 (t0 ), . . . , p6 (t0 )
is 2.2.
Lemma 21. There is a time t0 ∈ [0, M ] such that for all t00 ∈ [t0 , t0 +
3.479 ≥ 1.004 · |V (p(M ))|.
1
20 ],
the area of V (p(t00 )) is at least
Proof. Let λ(t) be the maximal distance between p(t) and p1 (t), . . . , p6 (t). √
By Observation 20, there is a
maximal time t0 with λ(t0 ) = 2.2. Therefore, and because 2.2 < 4 cos(50◦ ) < 2 2, the points p1 (t0 ), . . . , p6 (t0 )
are the Voronoi neighbors of p(t0 ). By Observation 18, λ(t) is 2-Lipschitz. This, together with the maximality
1
of t0 , implies 2.1 ≤ λ(t00 ) ≤ 2.2 for t00 ∈ [t0 , t0 + 20
]. Thus, Lemma 19 applies and yields |V (p(t00 ))| ≥ 3.479 ≥
1
00
0 0
1.004 · |V (p(M ))| for all t ∈ [t , t + 20 ].
Lemma
is a time t ∈
[0, M ] for which the area of Conv(m1 (t), . . . , mN (t)) is lower-bounded by
22. There
√ √
√
N
N
3.479 20M − 2π 2 N + M
+ 2 3 N − 20M
.
Proof. By Lemma 21, for each robot i that does not lie on the boundary of the start configuration, there
1
is a point in time t00 ∈ [0, M ] such that the area of V (p(t00 )) is at least 3.479 for all t00 ∈ [t0 , t0 + 20
]. The
N
N
continuous pigeonhole principle yields a time point t ∈ [0, M ] such that the area of k := 20M ∈ Θ( M
)
Voronoi regions
V
(q
(t)),
.
.
.
,
V
(q
(t))
is
at
least
3.479.
For
all
the
remaining
Voronoi
regions,
the
area
1
k
√
is at least 2 3 corresponding to the largest possible packing density as achieved in the start and target
configurations.
24
√
√
We give an upper bound N ≤ 2π(2 N + M ) on the number of robots whose Voronoi regions are not
contained in Conv(m1 (t), . . . , mN (t)). W.l.o.g., we assume that all these regions are Voronoi regions whose
area we lower bounded by 3.479. Moreover, we can assume all of these regions have zero area, i.e., ignoring
when lower bounding the area of Conv(m1 (t), . . . , mN (t)). Thus, the area of Conv(m1 (t), . . . , mN (t)) is at
least
√
√
√ √
N
N
3.479(k − N ) + 2 3(N − k) = 3.479 20M
− 2π 2 N + M
+ 2 3 N − 20M
.
Figure 18: An upper-bound construction for the number of robots whose Voronoi regions may intersect the
boundary of smallest enclosing ball for {m1 (t), . . . , mN (t)}. The radius of the smallest enclosing ball is upper
bounded by the distance from the center to the boundary in the start configuration plus the considered
makespan illustrated by the dashed circles.
√
√
It still remains to prove the upper bound N ≤ 2π(2 N + M ) on the number of robots whose Voronoi
regions are not contained in Conv(m1 (t), . . . , mN√(t)). First, we observe
√ that the length of the boundary
of Conv(m1 (t), . . . , mN (t)) is at most B := 2π(2 N + M ), because 2 N + M is an upper bound on the
radius of the smallest ball containing m1 (t), . . . , mN (t). In order to estimate N , we consider the maximal
number of points from [0, B] × R≥0 whose Voronoi regions intersect √
the x-axis.
This number is achieved
√
for the configuration as illustrated in Figure 18, implying N ≤ √B2 = 2π(2 N + M ), thus concluding the
proof.
√
√
Lemma 23. For each t ∈ [0, M ], |Conv(m1 (t), . . . , mN (t))| ≤ 2 3N + 2π( N + M )M .
Proof. In the√start configuration, the intersection of each Voronoi cell with Conv(m1 (0), . . . , m√N (0)) has
an area of 2 3. Thus, the convex hull of the start configuration has an area of at most 2 3N . We
give an upper bound on the area A gained during the motion, i.e., the area of Conv(m1 (t), . . . , mN (t)) \
Conv(m1 (0), . . . , mN (0)), corresponding √
to the gray region in Figure √
18. The length of the boundary
∂Conv(m1 (t), . . . , mN (t)) is at most 2π( N + M ), implying A ≤ 2π( N + M )M , thus concluding the
proof.
Theorem 14. There is an instance with optimal makespan M ∈ Ω(N 1/4 ), see Figure 16.
25
Proof. Combining the bounds from Lemma 22 and Lemma 23 yields black the following
√
√ √
N
N
+2 3 N −
3.479
− 2π 2 N + M
20M
20M
√
√
≤ 2 3N + 2π( N + M )M.
√
√ √
N
− 3, 479 2π 2 N + M ≤ 2π( N + M )M.
⇔
0, 014
20M
√
√
√
N
If M ∈ Ω( N ) holds, we are done. Otherwise we obtain M
∈ O(M N ) ⇔ N ∈ O(M 2 ), and thus
M ∈ Ω(N 1/4 ), concluding the proof.
8.2
An Upper Bound for Unbounded Environments
Next we give upper bounds on the stretch and makespan for moving disks in unbounded environments. First,
we show that we can achieve constant stretch for well-separated robots.
Theorem 24. If the distance between the centers of two robots of radius 1 is at least 4 in the start and target
configurations, we can achieve a makespan in O(d), i.e., constant stretch.
4
1
2
√
2 2
√
Figure 19: A mesh size of 2 2 avoids robot collisions, and the cell diagonals have length 4. Note that robots
may have arbitrary shape, as the separation argument applies to their circumcircles.
√
Proof. We consider a grid D with mesh size 2 2. In this way, as shown in Figure 19, two robots starting
simultaneously from different cells and traveling along two incident edges can touch when they reach the
midpoints, but do not collide. Moreover, the diagonals have length 4. By choosing a grid that has no
robot center on a grid line, every cell of D contains at most one start and one target position of a robot.
Additionally, we can move each robot in the start and target configuration to the center of its own cell,
allowing us to to use our algorithm from Section 3. Overall, we achieve a set of trajectories with makespan in
O(d).
√
In the remainder of this section we give an O( N )-approximation algorithm for the continuous makespan
for these kind of well-separated arrangements, by extending the
√ approach for discrete grids.
Again, we make use of an underlying grid with mesh size 2 2. Our algorithm proceeds in three phases. (1)
Moving the robots to vertices of the grid, (2) applying our O(1)-approximation for the discrete
case, and (3)
√
moving the robots from the vertices of the grid to their target positions.
To
ensure
a
O(
N
)-approximation,
√
we move each robot center to a grid vertex within a distance of O( N ). Phases (1) and (3) are symmetric in
the following sense. By applying the steps of phase (1) to the target configuration in reverse, we compute a
grid configuration that serves as target configuration for phase (2).
26
Phase (1) works as follows. (1.1) We
(x, y)-lexicographical
m by sorting the N robots according to thel√
m
l√begin
order. Then we subdivide them into
N vertical slices, each containing at most
N robots. To the
√
√
right of every slice, we add a vertical buffer slice of width 4 2 by moving all robots not yet considered by√4 2
units to the right. These trajectories are used in parallel, the distance covered by each robot is in O( N ).
The buffer slices guarantee that in all following steps, the robots in each vertical slice are independent of each
other.
For (1.2), we continue by sorting the robots within the vertical slices√according to the (y, x)-lexicographical
order. We separate the robots by ensuring vertical distance at least 4 2 between every pair of robots. This
can be done by moving the robots upwards, starting from the second-to-lowest
one. These trajectories can be
√
done in parallel and the distance covered by each robot is in O( N ). For (1.3), we finally move each robot
to the bottom-left vertex of the grid cell containing its center.
computes a trajectory plan with continuous makespan in
Theorem
√ 15. The algorithm described above
√
O(d + N ). If d ∈ Ω(1), this implies a O( N )-approximation algorithm.
Proof. Phase
√ (1) guarantees that either the horizontal or the vertical distance between each pair of robots is
at least 4 2. Therefore, in each grid cell, there is at most one robot and each robot is moved to its own grid
vertex. In phase (2), each robot is moved by O(d0 ) units, where d0 is the maximal distance between√
a robot’s
start and target position in the grid. As the
distance
each
robot
covers
in
phases
(1)
and
(3)
is
O(
N ), the
√
distance traveled in phase (2) is√in O(d + N ). Therefore, the trajectory set computed by the algorithm has
continuous makespan in O(d + N ). The running time as described above is pseudopolynomial; it becomes
polynomial by using standard compression techniques, e.g., by compressing large empty rectangles.
8.3
Colored and Unlabeled Disks
We can combine the positive results of the previous section with the technique of Theorem 24 to achieve the
same result for colored (and in particular, unlabeled) disks.
Corollary 25. There is an algorithm with running time O(k(mn)1.5 log(mn) + dmn) that computes, given
start and target images Is , It , an O(1)-approximation of the optimal makespan M and a corresponding set of
trajectories.
Proof. The proof proceeds analogous to Theorem 13: After computing an optimal bottleneck matching, apply
Theorem 3 in the setting of Theorem 24.
9
Geometric Difficulties
From a practical point of view, it is also desirable to compute provably optimal trajectories for specific
instances of moderate size, instead of solutions for large instances that are within a provable constant factor
of the optimum. This also plays a role as a building block for other purposes, e.g., providing a formal proof
of NP-hardness for parallel geometric motion planning for disks: We need to be able to establish the shape of
optimal trajectories when building gadgets for a hardness construction.
However, the geometry involved in this goal is far from easy, even for an example as small as the one
shown in Figure 20. This is closely related to recent work by Kirkpatrick and Liu [37], who devote a whole
paper to computing optimal trajectories for two disks in an arbitrary initial and target configuration, with
the objective of minimizing the total distance travelled instead of the makespan. A key insight is that optimal
trajectories consist of a limited number of circular arcs. This is not necessarily the case for trajectories that
minimize the makespan. Even for the seeming simplicity of our example, we do not have a proof of optimality
of the trajectory T3 shown on the right. This illustrates the difficulty of characterizing and establishing
optimal trajectories that minimize the total duration of a parallel schedule, highlighting the special role of
geometry for the problem.
27
A
B
T1
T2
a
b=b’
m1
a’
m2
b"
Figure 20: Moving the left unit disk A from position a to position a0 at distance 4, with disk B starting and
ending at b = b0 . (Left) Trajectory T1 rotates disk A around the stationary disk B resulting in makespan
2π
√ = 6.28 . . .. Trajectories T2 rotate both disks around the centers m1 and m2 , resulting in makespan
2π = 4.44 . . .. (Right) Choosing a circular arc through (−2, 0), (2, 0) and the (numerically optimized) point
(0, 0.493 . . .) for disk A (with B moving accordingly at distance 2) yields the trajectory T3 with makespan
4.16 . . ..
10
Conclusion
We have presented progress on a number of algorithmic problems of parallel motion planning, also shedding
light on a wide range of interesting open problems described in the following.
The first set of problems consider complexity. The labeled problem of Section 3 is known to be NP-complete
for planar graphs. It is natural to conjecture that the geometric version is also hard. It seems tougher to
characterize the family of optimal trajectories: As shown above, their nature is unclear, so membership in
NP is doubtful.
A second set of questions considers the relationship between
stretch factor and disk separability in the
√
continuous setting. We believe that the upper bound of O( N ) on the worst-case stretch factor for dense
arrangements is tight. What is the critical separability of disks for which constant stretch can be achieved?
How does the stretch factor increase as a function of N below this threshold? For sparse arrangements of
disks, simple greedy, straight-line trajectories between the origins and destinations of disks encounter only
isolated conflicts, resulting in small stretch factors close to 1, i.e., 1 + o(1). What is the relationship between
(local) density and the achievable stretch factor along the whole density spectrum?
Finally, practical motion planning requires a better handle on characterizing and computing optimal
solutions for specific instances, along with lower bounds, possibly based on numerical methods and tools.
Moreover, there is a wide range of additional objectives and requirements, such as accounting for acceleration
or deceleration of disks, turn cost, or multi-stop tour planning. All these are left for future work.
Acknowledgements. We thank anonymous reviewers of a preliminary version of the paper for helping
to improve the overall presentation.
References
[1] M. Abellanas, S. Bereg, F. Hurtado, A. G. Olaverri, D. Rappaport, and J. Tejel. Moving coins.
Computational Geometry: Theory and Applications, 34(1):35–48, 2006.
[2] A. Adler, M. de Berg, D. Halperin, and K. Solovey. Efficient multi-robot motion planning for unlabeled
discs in simple polygons. IEEE Transactions on Automation Science and Engineering, 12(4):1309–1317,
2015.
28
[3] K. M. Al-Wahedi. A hybrid local-global motion planner for multi-agent coordination. Master’s thesis,
Case Western Reserve University, 2000.
[4] B. Aronov, M. de Berg, A. F. van der Stappen, P. Švestka, and J. Vleugels. Motion planning for multiple
robots. Discrete & Computational Geometry, 22(4):505–525, 1999.
[5] M. Barer, G. Sharon, R. Stern, and A. Felner. Suboptimal variants of the conflict-based search algorithm
for the multi-agent pathfinding problem. In Proc. 7th Ann. Symp. Combinatorial Search, pages 19–27,
2014.
[6] S. Bereg, A. Dumitrescu, and J. Pach. Sliding disks in the plane. International Journal of Computational
Geometry & Applications, 18(5):373–387, 2008.
[7] P. Berman, E. D. Demaine, and M. Zadimoghaddam. O(1)-approximations for maximum movement
problems. In Proc. 14th Int. W. Approximation Algorithms for Combinatorial Optimization Problems
(APPROX 2011), pages 62–74, Princeton, New Jersey, August 17–19 2011.
[8] G. Calinescu, A. Dumitrescu, and J. Pach. Reconfigurations in graphs and grids. SIAM Journal on
Discrete Mathematics, 22(1):124–138, 2008.
[9] C. E. Campbell and J. Y. S. Luh. A preliminary study on path planning of collision avoidance for
mechanical manipulators. Technical report, Purdue University, School of Electrical Engineering, 1980.
[10] S. Cheung and F. C. M. Lau. Mesh permutation routing with locality. Information Processing Letters,
43(2):101–105, 1992.
[11] R. Cui, B. Gao, and J. Guo. Pareto-optimal coordination of multiple robots with safety guarantees.
Autonomous Robots, 32(3):189–205, 2012.
[12] D. Delahaye, S. Puechmorel, P. Tsiotras, and E. Féron. Mathematical models for aircraft trajectory
design: A survey. In Air Traffic Management and Systems, pages 205–247. Springer, 2014.
[13] E. D. Demaine, M. L. Demaine, and H. Verrill. Coin-moving puzzles. In R. Nowakowski, editor, More
Games of No Chance, Mathematical Sciences Research Institute Publications, pages 405–431. Cambridge
University Press, 2011.
[14] E. D. Demaine, M. T. Hajiaghayi, H. Mahini, A. S. Sayedi-Roshkhar, S. Oveisgharan, and M. Zadimoghaddam. Minimizing movement. ACM Transactions on Algorithms, 5(3):Article 30, July 2009.
[15] E. D. Demaine, M. T. Hajiaghayi, and D. Marx. Minimizing movement: Fixed-parameter tractability.
ACM Transactions on Algorithms, 11(2):Paper 14, November 2014.
[16] V. R. Desaraju and J. P. How. Decentralized path planning for multi-agent teams with complex
constraints. Autonomous Robots, 32(4):385–403, 2012.
[17] J. M. Dı́az-Báñez, M. A. Heredia, C. Peláez, J. A. Sellarès, J. Urrutia, and I. Ventura. Convex blocking
and partial orders on the plane. Computational Geometry: Theory and Applications, 51:55–66, 2016.
[18] A. Dumitrescu. Motion planning and reconfiguration for systems of multiple objects. In S. Kolski, editor,
Mobile Robots: Perception & Navigation, pages 1–20. InTech, 2007.
[19] A. Dumitrescu and M. Jiang. On reconfiguration of disks in the plane and related problems. Computational
Geometry: Theory and Applications, 46:191–202, 2013.
[20] A. Efrat, A. Itai, and M. J. Katz. Geometry helps in bottleneck matching and related problems.
Algorithmica, 31(1):1–28, 2001.
[21] M. Erdmann and T. Lozano-Pérez. On multiple moving objects. Algorithmica, 2(1):477–521, 1987.
29
[22] A. Felner, M. Goldenberg, G. Sharon, R. Stern, T. Beja, N. R. Sturtevant, J. Schaeffer, and R. Holte.
Partial-expansion A∗ with selective node generation. In Proc. AAAI Conf. Artificial Intelligence, pages
471–477, 2012.
[23] G. W. Flake and E. B. Baum. Rush Hour is PSPACE-complete, or “Why you should generously tip
parking lot attendants”. Theoretical Computer Science, 270(1):895–911, 2002.
[24] E. Freund and H. Hoyer. On the on-line solution of the findpath problem in multi-robot systems. In
O. Faugeras and G. Giralt, editors, 3rd Int. Symp. Robotics Research, pages 253–262, 1985.
[25] A. Geramifard, P. Chubak, and V. Bulitko. Biased cost pathfinding. In Proc. 2nd Conf. Artificial
Intelligence and Interactive Digital Entertainment, pages 112–114, 2006.
[26] M. Goldenberg, A. Felner, R. Stern, G. Sharon, N. R. Sturtevant, R. C. Holte, and J. Schaeffer. Enhanced
partial expansion A∗ . J. Artificial Intelligence Research, 50:141–187, 2014.
[27] M. T. Hajiaghayi, R. Khandekar, M. R. Khani, and G. Kortsarz. Approximation algorithms for movement
repairmen. ACM Transactions on Algorithms, 12(4):54:1–54:38, Sept. 2016.
[28] P. E. Hart, N. J. Nilsson, and B. Raphael. A formal basis for the heuristic determination of minimum
cost paths. IEEE Trans. Systems Science and Cybernetics, 4(2):100–107, 1968.
[29] R. A. Hearn and E. D. Demaine. PSPACE-completeness of sliding-block puzzles and other problems
through the nondeterministic constraint logic model of computation. Theoretical Computer Science,
343(1):72–96, 2005.
[30] R. A. Hearn and E. D. Demaine. Games, puzzles, and computation. CRC Press, 2009.
[31] S. Hirsch and D. Halperin. Hybrid motion planning: Coordinating two discs moving among polygonal
obstacles in the plane. In Algorithmic Foundations of Robotics V, pages 239–255. Springer, 2004.
[32] J. E. Hopcroft, J. T. Schwartz, and M. Sharir. On the complexity of motion planning for multiple
independent objects; PSPACE-hardness of the warehouseman’s problem. Int. J. Robotics Research,
3(4):76–88, 1984.
[33] J. E. Hopcroft and G. T. Wilfong. Reducing multiple object motion planning to graph searching. SIAM
J. Comput., 15(3):768–785, 1986.
[34] R. Huang, Y. Chen, and W. Zhang. A novel transition based encoding scheme for planning as satisfiability.
In Proc. AAAI Conf. Artificial Intelligence, pages 89–94, 2010.
[35] K. Kant and S. W. Zucker. Toward efficient trajectory planning: The path-velocity decomposition. Int.
J. Robotics Research, 5(3):72–89, 1986.
[36] H. Kautz and B. Selman. Unifying SAT-based and graph-based planning. In Int. J. Conf. Artificial
Intelligence, volume 99, pages 318–325, 1999.
[37] D. Kirkpatrick and P. Liu. Characterizing minimum-length coordinated motions for two discs. In Proc.
28th Canadian Conf. Computational Geometry (CCCG 2016), pages 252–259, August 3–5 2016. Full
version at https://arxiv.org/abs/1607.04005.
[38] S. Kloder and S. Hutchinson. Path planning for permutation-invariant multi-robot formations. In IEEE
Trans. Robotics, volume 22, pages 650–665. IEEE, 2006.
[39] D. Kornhauser, G. Miller, and P. Spirakis. Coordinating pebble motion on graphs, the diameter of
permutation groups, and applications. In Annual Symposium on Foundations of Computer Science, 1984,
SFCS ’84, pages 241–250, 1984.
30
[40] M. Kunde. Routing and sorting on mesh-connected arrays. In VLSI Algorithms and Architectures: 3rd
Aegean Workshop on Computation (AWOC 88), pages 423–433. Springer, 1988.
[41] S. LaValle and S. A. Hutchinson. Optimal motion planning for multiple robots having independent goals.
In IEEE Trans. Robotics and Automation, volume 14, pages 912–925, 1998.
[42] S. Leroy, J.-P. Laumond, and T. Siméon. Multiple path coordination for mobile robots: A geometric
algorithm. In Proc. 16th Int. Joint Conf. Artificial Intelligence, pages 1118–1123. Morgan Kaufmann
Publishers Inc., 1999.
[43] J. M. Marberg and E. Gafni. Sorting in constant number of row and column phases on a mesh.
Algorithmica, 3:561–572, 1988.
[44] J. Peng and S. Akella. Coordinating multiple robots with kinodynamic constraints along specified paths.
Int. J. Robotics Research, 24(4):295–310, 2005.
[45] G. Ramanathan and V. Alagar. Algorithmic motion planning in robotics: Coordinated motion of several
disks amidst polygonal obstacles. In Proc. 1985 IEEE Int. Conf. Robotics and Automation, volume 2,
pages 514–522, 1985.
[46] D. Ratner and M. K. Warmuth. Finding a shortest solution for the N × N extension of the 15-puzzle is
intractable. In Proc. AAAI Conf. Artificial Intelligence, pages 168–172, 1986.
[47] R. Regele and P. Levi. Cooperative multi-robot path planning by heuristic priority adjustment. In 2006
IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pages 5954–5959. IEEE, 2006.
[48] M. Rubenstein, A. Cornejo, and R. Nagpal. Programmable self-assembly in a thousand-robot swarm.
Science, 345(6198):795–799, 2014.
[49] O. Salzman, M. Hemmer, and D. Halperin. On the power of manifold samples in exploring configuration
spaces and the dimensionality of narrow passages. IEEE Trans. Automation Science and Engineering,
12(2):529–538, 2015.
[50] G. Sanchez and J.-C. Latombe. Using a PRM planner to compare centralized and decoupled planning
for multi-robot systems. In Proc. IEEE Int. Conf. Robotics and Automation, pages 2112–2119, 2002.
[51] C. Scheideler. Universal routing strategies for interconnection networks, volume 1390 of Lecture Notes in
Computer Science. Springer, 1998.
[52] J. T. Schwartz and M. Sharir. On the piano movers’ problem: III. Coordinating the motion of several
independent bodies: the special case of circular bodies moving amidst polygonal barriers. Int. J. Robotics
Research, 2(3):46–75, 1983.
[53] G. Sharon, R. Stern, A. Felner, and N. R. Sturtevant. Meta-agent conflict-based search for optimal
multi-agent path finding. In Proc. Symp. Combinatorial Search, pages 97–104, 2012.
[54] G. Sharon, R. Stern, A. Felner, and N. R. Sturtevant. Conflict-based search for optimal multi-agent
pathfinding. Artificial Intelligence, 219:40–66, 2015.
[55] G. Sharon, R. Stern, M. Goldenberg, and A. Felner. The increasing cost tree search for optimal
multi-agent pathfinding. Artificial Intelligence, 195:470–495, 2013.
[56] T. Siméon, S. Leroy, and J.-P. Laumond. Path coordination for multiple mobile robots: A resolutioncomplete algorithm. IEEE Trans. Robotics and Automation, 18(1):42–49, 2002.
[57] K. Solovey and D. Halperin. k-color multi-robot motion planning. Int. J. Robotics Research, 33(1):82–97,
2014.
31
[58] K. Solovey and D. Halperin. On the hardness of unlabeled multi-robot motion planning. The International
Journal of Robotics Research, 35(14):1750–1759, 2016.
[59] K. Solovey and D. Halperin. Sampling-Based Bottleneck Pathfinding with Applications to Fréchet
Matching. In P. Sankowski and C. Zaroliagis, editors, 24th Annual European Symposium on Algorithms
(ESA 2016), volume 57 of Leibniz International Proceedings in Informatics (LIPIcs), pages 76:1–76:16,
Dagstuhl, Germany, 2016. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik.
[60] K. Solovey, O. Salzman, and D. Halperin. Finding a needle in an exponential haystack: Discrete RRT for
exploration of implicit roadmaps in multi-robot motion planning. The International Journal of Robotics
Research, 35(5):501–513, 2016.
[61] K. Solovey, J. Yu, O. Zamir, and D. Halperin. Motion planning for unlabeled discs with optimality
guarantees. In Robotics: Science and Systems (RSS), 2015.
[62] P. Spirakis and C. K. Yap. Strong NP-hardness of moving many discs. Information processing letters,
19(1):55–59, 1984.
[63] T. Standley. Finding optimal solutions to cooperative pathfinding problems. In Proc. 24th AAAI Conf.
Artificial Intelligence, pages 173–178, 2010.
[64] P. Švestka and M. H. Overmars. Coordinated path planning for multiple robots. Robotics and autonomous
systems, 23(3):125–152, 1998.
[65] M. Turpin, N. Michael, and V. Kumar. Trajectory planning and assignment in multirobot systems. In
Algorithmic foundations of robotics X, pages 175–190. Springer, 2013.
[66] M. Turpin, K. Mohta, N. Michael, and V. Kumar. Goal assignment and trajectory planning for large
teams of interchangeable robots. Autonomous Robots, 37(4):401–415, December 2014.
[67] J. P. van den Berg and M. H. Overmars. Prioritized motion planning for multiple robots. In 2005
IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pages 430–435. IEEE, 2005.
[68] J. P. van den Berg, J. Snoeyink, M. C. Lin, and D. Manocha. Centralized path planning for multiple
robots: Optimal decoupling into sequential plans. In Robotics: Science and systems, volume 2, pages
2–3, 2009.
[69] M. Čáp, P. Novák, M. Seleckỳ, J. Faigl, and J. Vokřı̀nek. Asynchronous decentralized prioritized planning
for coordination in multi-robot system. In 2013 IEEE/RSJ Int. Conf. Intelligent Robots and Systems,
pages 3822–3829. IEEE, 2013.
[70] G. Wagner and H. Choset. M ∗ : A complete multirobot path planning algorithm with performance
bounds. In 2011 IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pages 3260–3267. IEEE, 2011.
[71] G. Wagner and H. Choset. Subdimensional expansion for multirobot path planning. Artificial Intelligence,
219:1–24, 2015.
[72] R. M. Wilson. Graph puzzles, homotopy, and the alternating group. Journal of Combinatorial Theory,
Series B, 16(1):86–96, 1974.
[73] P. R. Wurman, R. D’Andrea, and M. Mountz. Coordinating hundreds of cooperative, autonomous
vehicles in warehouses. AI magazine, 29(1):9, 2008.
[74] J. Yu and S. M. LaValle. Optimal multirobot path planning on graphs: Complete algorithms and
effective heuristics. IEEE Trans. Robotics, 32(5):1163–1177, 2016.
[75] J. Yu and D. Rus. An effective algorithmic framework for near optimal multi-robot path planning. In
Proc. Int. Symp. Robot. Res., 2015.
32
| 8cs.DS
|
arXiv:1802.04205v2 [cs.RO] 15 Feb 2018
Efficient Hierarchical Robot Motion Planning
Under Uncertainty and Hybrid Dynamics
Ajinkya Jain
Scott Niekum
Department of Mechanical Engineering,
University of Texas at Austin,
Austin, TX 78712, USA
Email: [email protected]
Department of Computer Science,
University of Texas at Austin,
Austin, TX 78712, USA
Email: [email protected]
Abstract—Noisy observations coupled with nonlinear dynamics
pose one of the biggest challenges in robot motion planning.
By decomposing the nonlinear dynamics into a discrete set of
local dynamics models, hybrid dynamics provide a natural way
to model nonlinear dynamics, especially in systems with sudden
“jumps” in the dynamics, due to factors such as contacts. We propose a hierarchical POMDP planner that develops locally optimal
motion plans for hybrid dynamics models [1]. The hierarchical
planner first develops a high-level motion plan to sequence the
local dynamics models to be visited. The high-level plan is then
converted into a detailed cost-optimized continuous state plan.
This hierarchical planning approach results in a decomposition
of the POMDP planning problem into smaller sub-parts that
can be solved with significantly lower computational costs. The
ability to sequence the visitation of local dynamics models also
provides a powerful way to leverage the hybrid dynamics to
reduce state uncertainty. We evaluate the proposed planner for
two navigation and localization tasks in simulated domains, as
well as an assembly task with a real robotic manipulator.
I. I NTRODUCTION
One of the biggest challenges in robot motion planning is
to develop feasible motion plans for systems having highly
nonlinear dynamics in the presence of partial or noisy observations. Often, these nonlinearities are caused by sudden
transitions or “jumps” in the dynamics (for example, due to
contacts in a robot manipulation task). When the dynamics
of a task can change suddenly in state space, even small
state estimation errors can lead to large deviations and plan
failure. Therefore, reasoning about the uncertainty over states
becomes crucial in order to develop robust motion plans.
Planning problems under uncertainty are often represented
as a partially observable Markov decision process (POMDP)
[2]. POMDP problems have been shown in literature to be
PSPACE-complete [3], making exact planning intractable.
To make planning tractable, POMDP planners typically
leverage various types of approximations [4, 5, 6, 7, 8] or
structural assumptions [7, 9, 10, 11, 12] that simplify the
problem. In this work, we propose to leverage a natural,
simplifying assumption that the nonlinear dynamics of robot
motion planning tasks can be decomposed into a discrete
set of simpler local dynamics models, of which only one
is active at any given time (e.g. a change in dynamics due
to contact). Note that these local dynamics models may be
approximate, especially when they are learned from data or are
a simplification of a complex underlying model. A complete
dynamics model can then be defined as a hybrid dynamics
model having hybrid states comprised of the continuous states
of the system along with a discrete state denoting the active
local dynamics model.
The primary contribution of this work is a novel POMDP
planner that plans in a hybrid belief space, allowing for
efficient information gathering and planning under uncertainty
with hybrid dynamics. We define the hybrid belief to be composed of a mixture of Gaussians and a discrete distribution,
which represent the uncertainty over the state of the robot
and the active dynamics model respectively. Using the hybrid
belief representation, a hierarchical POMDP motion planner
is presented that solves the POMDP problem by dividing it
into two levels: at the higher level, discrete state plans are
generated to find a sequence of local models that should be
visited during the task, and at the lower level, these discrete
state plans are converted into cost-optimized continuous state
belief-space plans.
The biggest advantage of dividing the planning problem into
two levels is that it breaks long-horizon planning problems
into multiple smaller segments that can be sequenced to
find a complete solution. Since POMDP planning becomes
exponentially more difficult with longer horizons ([3]), a
hierarchical approach breaks problem into chunks that can
be solved with significantly less effort. Another major benefit
of discrete state planning is that the planner can chose to
leverage a specific local dynamics model in order to improve
the effectiveness of the generated plans. For example, if it is
known a priori that in the k-th local dynamics model, motion
is allowed only along a particular vector (e.g. due to presence
of a wall), it can be used to reduce the state uncertainty along
the dimensions orthogonal to the allowed motion vector. This
indirect feedback for uncertainty reduction is critical for tasks
in which the observations are either extremely noisy or not
available at all.
The proposed POMDP motion planner is evaluated on two
simulated domains: autonomous robot navigation under uncertainty with spatially varying dynamics and planar navigation
while localizing using walls. Finally, our algorithm is evaluated with a physical robotic manipulator that is tasked with
partially assembling a toy airplane from the YCB dataset [13].
II. R ELATED W ORKS
A. POMDP Planning
Broadly, POMDP solving approaches can be divided into
two categories based on whether their state, action and observation spaces are discrete or continuous. Discrete space
POMDP solvers, in general, either approximate the value
function using point-based methods [5, 6] or use Monte-Carlo
sampling in the belief space [14, 15] to make the POMDP
problem tractable. Continuous space POMDP solvers often
approximate the belief over states as a distribution having finite
parameters (typically Gaussian) and either solve the problem
analytically using gradients [10, 11, 12] or use random sampling in the belief space [7, 8]. Other approaches have also
extended point-based methods to continuous domains [6].
Discrete space POMDP solvers have been shown to be
able to successfully plan for large discrete space domains,
however, continuous space domains are infinite-dimensional,
and discrete space solvers often fail to find feasible solutions
for planning horizons longer than a few steps [15]. Among
continuous space POMDP solvers, Agha-Mohammadi et al.
[7] and Hollinger and Sukhatme [8] have proposed sampling based methods that can find effective solutions even in
complex domains. However, they suffer from the problem of
obtaining sub-optimal solutions which can only be probabilistically optimal at best [16]. Gradient-based POMDP solvers
[10, 11, 12] form another class of very powerful POMDP
solvers which can find locally optimal solutions, but in the
context of manipulation planning, sudden changes in dynamics
due to contacts result in non-finite gradients at the transition
points and restrict the applicability of such methods.
POMDP solvers for hybrid domains, such as the one discussed in this work, have been previously discussed by Brunskill et al. [4], Sreenath et al. [9] and Agha-mohammadi et al.
[17]. Brunskill et al. [4] proposed a point-based POMDP planning algorithm, SM-POMDP planner, for solving continuousstate POMDPs based on the hybrid system dynamics. They
approximated the complex nonlinear system dynamics using
a hybrid multi-modal dynamics model with continuous statedependent discrete mode switching conditions. However, unlike our POMDP planner, SM-POMDP planner plans only in
the continuous domain and the discrete states are obtained
“passively” using the switching conditions. While this approach can be used to find feasible motion plans, it is not
leveraging some of the major natural advantages of the hybrid
dynamics representation such as shorter planning horizons and
a structured way to leverage dynamics for state uncertainty
reduction. Sreenath et al. [9] discussed the problem of bipedal
walking on a varying terrain by formulating it as a POMDP
problem defined on a continuous-time hybrid system. They
proposed a bi-level POMDP controller to track the transitions
in the terrain as a set of discrete states and were able to show
stable bipedal walking in simulated domains. However, this is
a passive approach as well, as it uses hybrid dynamics only
to capture the transitions in the terrain and not to simplify
the POMDP problem. Agha-mohammadi et al. [17] discussed
a POMDP solver with hybrid states to solve health-aware
stochastic motion planning problem for quadrotors, however,
the proposed solution is restricted only to the domains in which
the discrete and continuous states evolve independently.
B. Manipulation Planning and Control
As hybrid dynamics models are very effective in modeling nonlinearities that are due to sudden transitions in the
dynamics, a natural application domain for the proposed
POMDP solver is contact-rich robot manipulation. One of the
current approaches for solving the robot manipulation planning
problem is to search for an optimal sequence of parameterized
manipulation actions or primitives to perform the task [18, 19].
Kroemer et al. [19] have proposed to represent primitives
for different phases (modes) of a multi-phase manipulation
task using dynamic movement primitives (DMPs) and learn
a library of such manipulation skills which can be optimally
sequenced to perform a task. Unfortunately, a lack of a task
dynamics model prevents these methods from generalizing to
novel manipulation tasks, e.g. having different cost functions,
even if it involves the same objects.
In more recent works by Levine et al. [20] and Fu et al. [21],
authors have used deep learning techniques to develop endto-end control policies directly from vision; however, under
sparse availability of training data (especially in robotics),
these approaches tend to fail to develop generalized control
policies for all system states or initial conditions.
III. P RELIMINARIES AND D EFINITIONS
A. POMDPs
Partially Observable Markov Decision Processes (POMDPs)
provide a mathematical framework for the problem of sequential decision making under uncertainty [2]. Let X ⊂ Rn be
the space of all possible states x of the robot, U ⊂ Rm be
the space of all possible control inputs u and Z ⊂ Rk be the
space of all possible sensor measurements z the robot may
receive. To account for state uncertainty, a distribution of the
state xt of the robot given all past control inputs and sensor
measurements is defined as the belief b[xt ], given as
b[xt ] = p[xt |u0 , ..., ut−1 , z1 , ..., zt ]
(1)
where xt ∈ X , ut ∈ U and zt ∈ Z are the robot’s
state, control input and received measurement at time step t,
respectively and B ⊂ {X →
− R} represent the space of all
possible beliefs.
In a general case, considering a stochastic dynamics and
observation model for the process given as
xt+1 ∼ p[xt+1 |xt , ut ],
zt ∼ p[zt |xt ]
(2)
for a given control input ut and a measurement zt+1 , the
belief can be propagated using Bayesian filtering as
Z
b[xt+1 ] = ηp[zt+1 |xt+1 ] p[xt+1 |xt , ut ]b[xt ]dxt (3)
where η is a normalizing constant independent of xt+1 .
In this work, we consider hybrid beliefs to represent the
distribution of hybrid states, with the beliefs over continuous
states represented as a mixture of Gaussians and a discrete
distribution that represents confidence over the active local
dynamics model. Further details on belief propagation are
discussed in the section IV.
B. Trajectory Optimization using Direct Transcription
Direct Transcription is a trajectory optimization method
in which a constrained nonlinear optimization problem is
set up with the user-defined objective function over a set
of knot-points {xi , ui }) chosen to discretize the continuous
space trajectory into a set of decision variables. The system
dynamics are imposed as the constraints on the optimization
problem. For discrete-time systems, these knot-points can
be taken as the system state xt and the control input ut
at each time step t. However, planning for longer horizons
will then require specifying a high number of knot-points
(xi , ui ) which can result in very high computational costs.
This can be resolved by approximately parameterizing the
space of possible trajectories by a series of M segments and
solving the optimization problem for a knot points only at
the start and end points of segments. The intermediate points
on the trajectory can be obtained via numerical integration.
Let x01:M and u01:M −1 be sets of state and action variables
that parameterize the trajectory in terms of segments. The ith
segment can be assumed to start at time iδ and ends at time
T
for a time horizon T .
iδ + δ − 1, where δ =
M
A general objective function for trajectory optimization can
be given as
ˆ 0 1 : M , u01:M )
J(x1:T , u1:T ) ≈ J(x
=
M
X
0
0T
0
x̃0T
i Qx̃i + ũi Rũi
j=1
where Q and R represent the cost matrices associated with
the state and the input respectively. The system dynamics
incorporated as constraints can be defined as:
x02 =φ(x01 , u01 )
..
.
x0k =φ(x0k−1 , u0k−1 )
where the function φ(x0i , u0i ) can be seen as performing numerical integration of the current state variable x0i till the next
state variable x0i+1 . The function φ is given as
+
[F (xt+1 , ui ) − F (xt , ui )]
where x ∈ Rn , u ∈ Rm and z ∈ Rl are the continuous state,
control input and observation variables, respectively and q ∈ Q
represents the active discrete state of the system. Evolution of
the discrete state of the system can be modeled by a finite
state Markov chain as
µt+1 = Πµt
(6)
where Π = {πij } is the discrete state transition matrix and
µ ∈ RQ is the probability distribution over discrete states at
time t.
IV. A PPROACH
We propose to solve the problem of motion planning under
uncertainty for tasks governed by highly nonlinear dynamics as
a POMDP problem defined on a hybrid dynamics model. Different local dynamics models constituting the task dynamics
are represented as distinct discrete states of the hybrid model.
Under uncertainty over the robot state, a separate discrete
distribution needs to be maintained to represent our confidence
over the active local dynamics model at each time step. Jointly,
a hybrid belief over the hybrid state of the system can be
defined with a continuous part representing uncertainty over
the robot state and a discrete part representing uncertainty in
the active local dynamics model.
In this work, we assume that the continuous part of hybrid
belief is represented by a mixture of L Gaussian distributions
given as
L
X
bxt =
αl N (µl , Σl )
(7)
l=1
x0i+1 = φ(x0i ) =F (x0i , ui )
iδ+δ−1
X
states q ∈ Q ⊂ W [1]. Each discrete state of the system
corresponds to a separate dynamics model that governs the
evolution of continuous states. These types of dynamical models are sometimes referred to as switched dynamical systems
in the literature [22].
In a hybrid model, discrete state transitions of the system
can be represented as a directed graph with each possible
discrete state q corresponding to a node and edges (e ∈
E ⊆ Q × Q) marking possible transitions between the nodes.
These transitions are conditioned on the continuous states. A
transition from the discrete state q to another state q 0 happens
if the continuous states x are in the guard set G(q, q 0 ) of the
0
0
edge eqq where eqq = {q, q 0 }, G(·) : E → P (X) and P (X)
is the power set of X. Thus, a hybrid dynamics model H can
be defined as
xt+1 = fq (xt , ut )
(5)
zt = hq (xt )
where αl represents the mixing weight of l-th Gaussian.
(4)
t=iδ
where F (xt , ut ) represents the system dynamics.
C. Hybrid Dynamics
A hybrid dynamics model of a system is a dynamics model
in which the states of the system evolve with time over both
continuous space x ∈ X = RN and a finite set of discrete
A. Belief Propagation under Hybrid Dynamics
A hybrid belief can be defined as B = {bx , bq }, where bx
and bq correspond to the belief over the continuous robot state
x and the discrete states (local dynamics models) q respectively. Propagation of hybrid belief using Bayesian filtering can
be separated into two steps: prediction based on the dynamics
model to obtain a belief prior and an update step using the
received observation to find belief posterior.
1) Belief Prior: At each time step t, we can propagate the
current belief bxt through the system dynamics of each discrete
0
state, F q (xt , ut ), individually and then take a weighted sum
of the propagated belief set to obtain a belief prior for the next
time step b̂xt+1
X 0
q 0
b̂xt+1 =
F q (bxt , ut ) bt q
(8)
q0
q 0
bt q
0
bqt+1 = γMt+1 ◦ b̂qt+1
0
(12)
0
T
where Mt+1 = [P (zt+1 |qt+1 = q )]
∀q ∈ Q, ◦ is the
element-wise multiplication operator, γ is a normalization
constant and
0
0
bqt ,
where
= p(qt = q |xt ) is q -th component of
and xt ,
qt and ut represent the continuous states, discrete state and
continuous control input to the system at time t and b̂[x(t+1)]
is denoted as b̂xt+1 .
Under stochastic continuous state dynamics, the definition
of the discrete state transition matrix as given in Equation 6
needs to be extended. Assuming the transitions of discrete
states are given by a directed graph with self-loops, we can
define the extended discrete state transition matrix Π at time
j
t as Πt = {πt (i, j) = p(qt+1
|qti , b̂xt+1 ) ∀q i , q j ∈ Q} as
Z
j
i
πt (i, j) = η
1qqi (x)b̂xt+1 (x)dx, if ∃ eqqj ,
N
(9)
R
= ,
otherwise
j
where 1qqi (x) is an indicator function defined as
(
1, if x ∈ G(q i , q j )
qj
1qi (x) =
0, otherwise
(10)
P|Q|
η is a normalization constant, given as η = k=1 π(i, k)
and is a small probability to handle scenarios in which the
received observations do not correspond to any legal discrete
transition based on the current belief.
Calculating the extended discrete state transition matrix Πt
at each time step using Eq. 9 can be computationally expensive. An approximation of Πt can be obtained by sampling n
random points from the belief over continuous states bxt+1 and
calculating ratio of points lying in the guard set G(q i , q j ) to
the total number of sampled points for each discrete state q j .
2) Belief Posterior: We use a hybrid estimation algorithm
based on Bayesian filtering to reduce the uncertainty over
states using noisy continuous state observations. The proposed
algorithm consists of two layers of filters: first to estimate the
continuous states of the system and second to estimate the
discrete states of the system. Upon receiving observation zt+1 ,
the continuous state prior is updated by taking a weighted sum
of a bank of extended Kalman filters running independently,
with each discrete mode having an individual filter. The
weights for the sum is determined using the prior for the
discrete mode b̂qt+1 . The complete update step for continuous
states can be written as
q
X q0
q0
q0
bxt+1 = b̂xt+1 +
Kt+1 (zt+1 − Ht+1
(b̂xt+1 )) b̂t+1
(11)
q0
0
the discrete state can be obtained by using a Bayesian filter
update given as
where Kqt+1 is the Kalman Gain for discrete state q 0 at time
qq0
t + 1 and b̂t+1
is q 0 -th component of b̂qt+1 . The update for
q
P (zt+1 |qt+1 = q 0 ) = zt+1 ∼ Ht+1
(bxt+1 )
0
q
where Ht+1
(.) is the observation function for state q 0 .
Mixing weights for the mixture of Gaussians are also
updated based on the received observations as
l
αt+1
= N (ν|0, Σlt+1 )
(13)
l
where innovation ν is given as ν = zt+1 − ẑt+1
and
X q0
qq0
l
Ht+1 µlt+1 × b̂t+1
ẑt+1
=
q0
A new mixture of L Gaussians is then chosen to represent
the continuous belief bxt+1 at time step t + 1.
B. Direct Planning
With the hybrid belief propagation equations defined, we
can now use the trajectory optimization technique to solve
the POMDP problem. We assume maximum likely observations (MLO) obtained by propagating the current belief over
continuous states through the system dynamics (Eqn. 8) as
true observations for developing locally optimal motion plans.
This is a standard assumption to make while solving POMDP
problems and has been discussed previously by Platt et al. [23].
In this work, the nonlinear optimization problem set up for
trajectory optimization is posed as a sequential least squares
programming (SLSQP) problem and solved using the SNOPT
software package [24, 25]. We denote this approach as the
direct planning approach.
C. Hierarchical Planner
Although the direct planning approach can be used to solve
the POMDP problem, planning for longer horizons in complex
tasks, such as contact-rich manipulation tasks, can result in
infeasible computational costs [3]. To tackle this challenge, we
propose a hierarchical planner that decomposes the POMDP
problem into smaller subproblems which can be solved with
significantly less effort. The proposed hierarchical planner has
two levels: a higher level to find the best sequence of local
dynamics models that should be visited along the path (by
visiting the region in the continuous state space corresponding
to its guard set, G(q 0 , ·)) and a lower level that is similar
to the direct planning approach discussed above. The higher
level planner starts by generating a set of high-level plans
consisting of all possible permutations of the discrete states of
the task. A high-level plan is then converted into a sequence of
continuous state goals which represent the most likely points
in the continuous state space for activating the corresponding
set of discrete states (see Algorithm [1]). The lower level
planner is then called for each of these continuous state
goals and a complete continuous state path for the highlevel plan is generated by combining the outputs of lower
level planner. An additional discrete state is added to each
high-level plan which represents the desired goal of the task
and is considered to be active within an −neighbourhood
of the actual task goal. High-level plans are then ranked by
calculating a divergence cost on the distribution of planner’s
confidence on the active discrete state at the final point
of the plan and the desired confidence distribution (all the
probability mass within the −neighbourhood of the goal).
Continuous state plan corresponding to the high-level plan
with the minimum cost is chosen as the best path for the task.
In this work, we have used Hellinger distance to calculate
the divergence cost [26] between the discrete distributions as
it forms a symmetric bounded metric with a value between
0 and 1, and was found to be more numerically stable than
the Bhattacharya distance, KL-divergence, and its symmetric
form on the tested application domains. Radial basis functions
were used to interpolate the divergence costs throughout the
domain and the differential evolution method was used to find
the approximately globally optimal solutions of the generated
cost map [27].
Algorithm 1: High-Level Plan −
→ Continuous State Goals
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Function ds plan to cs goals (high-level plan)
k
for each qgoal
in high-level plan do
Find equivalent
( confidence distribution Wgoal :
k
1, if q = qgoal
Wgoal (q) =
0, else
Sample n random points:
Xsample = {x1 , ..., xn } ∼ X ;
for each xi ∈ Xsample do
Find confidence distributions on discrete
states wi ∈ Wsample :
Sample a random set X 0 ∼ X ;
for each q 0 ∈ Q do
|x0 ∈ X 0 ∩ G(q 0 , q 00 ) ∀q 00 |
wi (q 0 ) =
;
|X 0 |
Find cost of divergence ci ∈ C 0 ⊂ R:
ci (xi ) = Hellinger(wi , Wgoal );
Define cost map on complete domain X :
Ccomplete (x) = Interpolate(C 0 );
Find best representative point in continuous state:
xkbest = global optimization(x, Ccomplete );
Append xkbest to Xcs goals ;
return Xcs
goals ;
D. Trajectory Stabilization
With the MLO assumption, it is very likely that during execution the belief over robot state will diverge from the nominal
trajectory planned. To ensure that the execution phase belief
follows the plan, a belief space LQR (B-LQR) controller can be
defined around the nominal trajectory. B-LQR controllers were
introduced by Platt et al[23] and can be seen as belief-space
extension of Linear-Quadratic Regulators (LQR). For systems
with linear-Gaussian process and observation dynamics, a BLQR controller is optimal and equivalent to a linear-Quadratic
Gaussian (LQG) controller. In B-LQR, each point in the
nominal trajectory is defined as a set point and quadratic costs
are defined for distance from the set point and the control effort
required to converge to the set point. Closed form solutions
exist to ensure convergence to the set point within a finite
time horizon. While stabilizing the trajectory, the most likely
active discrete state is taken to define the governing dynamics
of the system. However, it may happen that B-LQR controller
is unable to stabilize the execution phase (actual) belief around
the nominal trajectory. If the planned belief for the next step
deviates more than a δ-threshold from the actual belief after
the observation update, a replanning call to the planner is
triggered.
V. E XPERIMENTS
The proposed POMDP solver for hybrid dynamics was
tested on three tasks: autonomous navigation for informationgathering in a benchmark light-dark domain with spatially
varying dynamics, navigation and localization in a walleddomain with extremely poor observations, and a real manipulation task of partially assembling a toy airplane [13] under noisy
observations by leveraging contacts to reduce uncertainty.
A. Domain-I: Light-Dark Domain
The first task aims to test the capability of the lower-level
planner to develop effective motion plans under spatiallyvarying dynamics. A 2D light-dark domain [23] was considered for the task in which observation quality is proportional
to the degree of light at a given state. The task objective was to
reach the pre-specified goal with minimum state uncertainty.
The 2D domain ({x, y} ∈ [−10, 10]) was considered to contain
three different local dynamics functions, given as
if x < −1
xt + 0.5u,
f (xt , u) = xt + u,
if x ∈ [−1, 4]
T
xt + [2u1 , u2 ] , if x > 4
where xt = {xt , yt }T . Belief over continuous states was
considered to be a Gaussian distribution bx = {µ, σ}. The
observation function was taken as h(xt ) = xt + w with
zero-mean Gaussian observation noise w ∼ N (·|0, W (x))
where W (x) = 21 (5 − xx )2 + const. Matrices defining the
cost function over error in states, control input and additional cost for final state error and covariance were taken as
Q = diag(0.5, 0.5), R = diag(0.5, 0.5), Qlarge = 30 and
Λ = 400 respectively.
We compare the performance of the direct-planning approach with a baseline approach which, instead of having a
probabilistic belief over the active dynamics model, determines
the active dynamics model based on the maximum likelihood
state of the continuous belief. Trajectories planned and the
actual executions are shown in Figure 1. Figure 1 shows that
the direct planning approach was able to develop effective
belief space plans that move into the light to reduce uncertainty before heading to the goal even under changing local
dynamics, while the baseline approach fails to do so. It can
be seen from Figure 1 that if the uncertainty over continuous
states is high, having a completely deterministic approach over
active discrete states while planning can cause the belief to
diverge completely from the actual robot state. Hence, it is
critical to maintain a belief over the discrete states as well in
order to develop meaningful belief space plans.
Fig. 1: Plots showing planned and actual robot trajectories for
the baseline and the direct planning approach. Initial belief
mean µ = {2, 2}, cov = diag(5.0, 5.0) (yellow cross on
red trajectory), True start position:={2.5,0, 0} (yellow cross
on white trajectory), goal position:={0,0} (shown by green
square). Brightness reflects quality of observation. Dashed
lines separate different dynamics modes
B. Domain-II: Walled Domain
In the second domain, we compared the performance of
the hierarchical planner with the direct planning approach.
Note that the direct planning approach is similar in principle
to the SM-POMDP planner proposed by Brunskill et al. [4]
and hence, provides a comparison of the proposed hierarchical
planner with a similarly passive planning approach. The 2D
domain ({x, y} ∈ [−2, 15]) consisted of two perpendicular
walls parallel to the x and y axis respectively. Observations
were considered to be extremely noisy and had zero-mean
Gaussian observation noise w given as w ∼ N (·|0, 15). As
the motion along the wall is constrained to be only parallel to
the wall, the robot can use it to efficiently localize itself in a
direction orthogonal to the wall. The hybrid dynamics model
of the domain can be given as
xt + "
u,
if x > −2, y > −2
#
x + 0 0 u, if x < −2
t
0 1
f (xt , u) =
"
#
1 0
u, if x > −2, y < −2,
xt +
0 0
where xt = {xt , yt }T . Beliefs over continuous states were
considered to be a Gaussian distribution bx = {µ, σ}. Observation function was taken as h(xt ) = xt + w. Matrices
defining the cost function over error in states, control input,
additional cost for final state error and covariance were taken
as Q = diag(0.5, 0.5), R = diag(10.0, 10.0), Qlarge = 1e4
and Λ = 1e7 respectively.
Sample trajectories planned by the direct planning and the
hierarchical planner are shown in Figure 2. It is evident from
the figures that the hierarchical planner plans to selectively
visit the two discrete states representing the walls, in contrast
to the direct method. Also, the hierarchical planner is able to
converge to the goal faster and with a much lower uncertainty
than the direct planning approach. As the direct planner
does not leverage the knowledge of local dynamics models
in a structured way, it needs to plan longer trajectories to
gather more information. However, due to high noise in the
observations, it still fails to converge to the goal with high
accuracy.
Additional statistical analysis to compare the two approaches in terms of total planning time, final error and final
belief uncertainty are presented in Table I. It can be seen
from the table I that, for comparable final error and final
belief uncertainty, the hierarchical planner is able to find a
solution approximately 5 times faster than the direct planning
approach. The planning horizon was set to 20 time steps.
Optimized planning parameters for both approaches were first
obtained by conducting multiple test runs in the preparation
phase. Belief and actual robot start conditions were taken as
[5, 5]T and [3.5, 2.0]T respectively. The termination condition
was triggered when the maximum likelihood point of the belief
converged within a ball of 0.2 unit radius around the set goal
([0, 0]T ) with a maximum covariance of 1 unit.
Metric
Total time (in seconds)
Final Error
Final Max. Belief Uncertainty
Direct
51.908
[−0.168, 0.172]T
0.696
Hierarchical
10.695
[0.086, 0.198]T
0.625
TABLE I: Comparison of direct and hierarchical planning.
Values are averaged over 5 runs for both methods.
C. Domain-III: Airplane assembly
Finally, we experimentally demonstrate that the hierarchical
POMDP planner can be used to tractably solve a real world
manipulation task—the partial assembly of a toy airplane from
the YCB dataset [13] (shown in Figure 3). We considered
the first step of inserting the landing gear into the wing as a
test case for our planner. The task requires high precision,
as the maximum final state error margin for a successful
execution is ±0.2 cm. As an added challenge, no direct
feedback on the location of the hole on the wing was made
available to the planner. Only noisy feedback on the location
of the airplane in the world frame was provided (average
estimation error ±2.0 cm; obtained by doing an online object
cluster extraction, using multi-plane segmentation from the
Point Cloud Library (PCL) on the point cloud data of a
Microsoft Kinect v2 sensor). This experiment demonstrates
two important features of the proposed planner: first, the planner can be scaled to solve real-world manipulation planning
(a) Direct Planning
(b) Hierarchical Planning
Fig. 2: A comparison of planned and actual trajectories using the direct planning and hierarchical planning approaches on the
walled domain. For both cases, Initial belief mean µ = {5, 5}, cov = diag(11.5, 11.5) , True start position:={3.5, 2.0}. Gray
circles represent belief covariance
under uncertainty problems and second, due to the hierarchical
planning approach, the planner essentially enables the robot to
plan and “feel around” to localize itself even when the direct
visual observations aren’t available, similar to what a human
might do.
In a robot manipulation task involving contacts, based
on the type of contact between the bodies, the number of
state-dependent local dynamics models can be large, or even
infinite. We simplify the problem by assuming an approximate
hybrid dynamics model, in which the local dynamics models
correspond to possible motion constraints that the robot can
encounter while executing the task. For example, the task of
placing a cup on a table can be considered to be approximately
made of two local dynamics models: one when the two objects
are not in contact and the other when the cup is in contact
with the table plane. The second dynamics model represents
the motion constraint placed on the cup by the table by
restricting its motion to be only along its plane and not
penetrating it. This approximation helps in having a succinct
and effective representation of the task dynamics; as under this
Fig. 3: Toy airplane: assembled (left) and unassembled (right)
approximation, for a specific set of inputs, the relative motion
between the two objects in contact will always be the same
independent of the type of contact between them. In this case,
the specific set of inputs would be the set of all inputs which
do not result in moving the cup away from the table plane,
resulting in breaking the contact between them.
In this experiment, we consider the domain to be made up of
four distinct local dynamics models: two corresponding to the
linear motions along the wing plane edges, one corresponding
to the corner of the plane and one to represent free-body
motion elsewhere in the domain. At the highest level, the
planning problem can be broken down into two steps: first,
to localize the gear at a point in a plane parallel to the wing
and second, to insert the gear into the hole. A hybrid dynamics
model in a plane parallel to the wing can be given as
"
#
0 0
xt +
u, if x ∈ [4, 4.5], y > −13.5
0 1
"
#
1 0
f (xt , u) = xt +
u, if x < 4, y ∈ [−14, −13]
0 0
xt + 0 ∗ u,
if x ∈ [4, 4.5], y ∈ [−14, −13.5]
xt + u,
otherwise
(14)
where 1 unit in continuous space = 1 cm. The observation
function was given as h(xt ) = xt + w with zero-mean
Gaussian observation noise w ∼ N (·|0, 2I2 ))
Experiments were conducted using a bi-manual manipulator
robot with two Kinova Jaco2 7-dof arms. The planner took
14.682 seconds to plan the path. Figure 5 shows the trajectory
planned by the hierarchical planner and the actual trajectory
taken by the robot in a plane parallel to the wing. Figure 4
Fig. 4: Snapshots of the robot assembling the toy airplane
Fig. 5: Planned and Actual trajectories for the airplane assembly task. Bold black lines represents the edges of the airplane
wing
shows snapshots of the trajectory executed by the robot during
the task from two perpendicular angles. It can be see from the
second panel of the Fig. 4 that planner plans to activate the
motion constraint parallel to the wing in order to reduce its
uncertainty. Once localized in the plane parallel to the wing,
the robot changes planes to move to a point directly above
the hole and then proceeds to insert the landing gear into the
airplane.
VI. C ONCLUSION
Nonlinear task dynamics, especially due to sudden changes
in the dynamics, can be effectively modelled using a hybrid
dynamics model. A hybrid dynamics model consists of a
set of local dynamics models with only one of them being
active at a time. In this work, we propose a hierarchical
POMDP planner for hybrid dynamics which can develop
locally optimal motion plans for tasks involving nonlinear
dynamics under noisy observations. The proposed planner
generates hierarchical motion plans at two levels: first, a
high-level motion plan that sequences the local dynamics
models to be visited and second, based on the best high-level
plan, a detailed continuous state motion plan to be followed
by the robot. The hierarchical planning approach breaks the
large POMDP problem into multiple smaller segments with
shorter planning horizons, which significantly increases the
computational efficiency of the planner. High-level planning
also enables the robot to leverage task dynamics to improve
its performance—for example, reducing uncertainty using the
task motion constraints in order to develop motion plans which
are more robust to state uncertainty.
In the present work, a hybrid model of the task dynamics
needs to be provided to the planner by an expert. Hence, a
natural extension of this work is to autonomously learn the
hybrid dynamics model of the task. For example, Niekum et
al. have proposed methods [28, 29] to learn the articulation
motion models encountered while manipulating an object. In
the future, the proposed POMDP planner may be combined
with these methods to develop an end-to-end approach for
learning hybrid dynamics models for the manipulation tasks
and use them to generate motions plans that are robust to state
uncertainty.
R EFERENCES
[1] John Lygeros, Shankar Sastry, and Claire Tomlin. Hybrid
systems: Foundations, advanced topics and applications.
under copyright to be published by Springer Verlag,
2012.
[2] Sebastian Thrun. Probabilistic robotics. Communications
of the ACM, 45(3):52–57, 2002.
[3] Christos H Papadimitriou and John N Tsitsiklis. The
complexity of markov decision processes. Mathematics
of operations research, 12(3):441–450, 1987.
[4] Emma Brunskill, Leslie Kaelbling, Tomas Lozano-Perez,
and Nicholas Roy. Continuous-State POMDPs with
Hybrid Dynamics. Symposium on Artificial Intelligence
and Mathematics, pages 13–18, 2008.
[5] Hanna Kurniawati, David Hsu, and Wee Sun Lee. Sarsop:
Efficient point-based pomdp planning by approximating
optimally reachable belief spaces. In Robotics: Science
and Systems, volume 2008. Zurich, Switzerland, 2008.
[6] Guy Shani, Joelle Pineau, and Robert Kaplow. A survey
of point-based pomdp solvers. Autonomous Agents and
Multi-Agent Systems, 27(1):1–51, 2013.
[7] Ali-Akbar Agha-Mohammadi, Suman Chakravorty, and
Nancy M Amato.
Firm: Sampling-based feedback
motion-planning under motion uncertainty and imperfect
measurements. The International Journal of Robotics
Research, 33(2):268–304, 2014.
[8] Geoffrey A Hollinger and Gaurav S Sukhatme.
Sampling-based robotic information gathering algorithms. The International Journal of Robotics Research,
33(9):1271–1287, 2014.
[9] Koushil Sreenath, Connie R Hill Jr, and Vijay Kumar.
A partially observable hybrid system model for bipedal
locomotion for adapting to terrain variations. In Proceedings of the 16th international conference on Hybrid
systems: computation and control, pages 137–142. ACM,
2013.
[10] Jur Van Den Berg, Sachin Patil, and Ron Alterovitz.
Motion planning under uncertainty using iterative local
optimization in belief space. The International Journal
of Robotics Research, 31(11):1263–1278, 2012.
[11] Vadim Indelman, Luca Carlone, and Frank Dellaert.
Planning in the continuous domain: A generalized belief
space approach for autonomous navigation in unknown
environments. The International Journal of Robotics
Research, 34(7):849–882, 2015.
[12] Anirudha Majumdar and Russ Tedrake. Funnel libraries
for real-time robust feedback motion planning. The
International Journal of Robotics Research, 36(8):947–
982, 2017.
[13] Berk Calli, Arjun Singh, James Bruce, Aaron Walsman,
Kurt Konolige, Siddhartha Srinivasa, Pieter Abbeel, and
Aaron M Dollar. Yale-cmu-berkeley dataset for robotic
manipulation research. The International Journal of
Robotics Research, 36(3):261–268, 2017.
[14] David Silver and Joel Veness. Monte-carlo planning
in large pomdps. In Advances in neural information
processing systems, pages 2164–2172, 2010.
[15] Hanna Kurniawati and Vinay Yadav. An online pomdp
solver for uncertainty planning in dynamic environment.
In Robotics Research, pages 611–629. Springer, 2016.
[16] Mohamed Elbanhawi and Milan Simic. Sampling-based
robot motion planning: A review. IEEE Access, 2:56–77,
2014.
[17] Ali-akbar Agha-mohammadi, N Kemal Ure, Jonathan P
How, and John Vian. Health aware stochastic planning for persistent package delivery missions using
quadrotors. In Intelligent Robots and Systems (IROS
2014), 2014 IEEE/RSJ International Conference on,
pages 3389–3396. IEEE, 2014.
[18] Mehmet R Dogar and Siddhartha S Srinivasa. A planning
framework for non-prehensile manipulation under clutter
and uncertainty. Autonomous Robots, 33(3):217–236,
2012.
[19] Oliver Kroemer, Christian Daniel, Gerhard Neumann,
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
Herke Van Hoof, and Jan Peters. Towards learning
hierarchical skills for multi-phase manipulation tasks. In
Robotics and Automation (ICRA), 2015 IEEE International Conference on, pages 1503–1510. IEEE, 2015.
Sergey Levine, Nolan Wagener, and Pieter Abbeel.
Learning contact-rich manipulation skills with guided
policy search. In 2015 IEEE international conference on
robotics and automation (ICRA), pages 156–163. IEEE,
2015.
Justin Fu, Sergey Levine, and Pieter Abbeel. One-shot
learning of manipulation skills with online dynamics
adaptation and neural network priors. In Intelligent
Robots and Systems (IROS), 2016 IEEE/RSJ International Conference on, pages 4019–4026. IEEE, 2016.
Zoubin Ghahramani and Geoffrey E Hinton. Variational
learning for switching state-space models. Neural computation, 12(4):831–864, 2000.
R Platt Jr, Russ Tedrake, Leslie Kaelbling, and Tomas
Lozano-Perez. Belief space planning assuming maximum
likelihood observations. Robotics: Science and Systems,
2010. ISSN 2330765X.
Philip E. Gill, Walter Murray, and Michael A. Saunders.
SNOPT: An SQP algorithm for large-scale constrained
optimization. SIAM Rev., 47:99–131, 2005.
Philip E. Gill, Walter Murray, Michael A. Saunders, and
Elizabeth Wong. User’s guide for SNOPT 7.6: Software
for large-scale nonlinear programming. Center for Computational Mathematics Report CCoM 17-1, Department
of Mathematics, University of California, San Diego, La
Jolla, CA, 2017.
Sung-Hyuk Cha.
Comprehensive survey on distance/similarity measures between probability density
functions. City, 1(2):1, 2007.
Rainer Storn and Kenneth Price. Differential evolution–
a simple and efficient heuristic for global optimization
over continuous spaces. Journal of global optimization,
11(4):341–359, 1997.
Scott Niekum, Sarah Osentoski, Christopher G Atkeson,
and Andrew G Barto. Online bayesian changepoint
detection for articulated motion models. In 2015 IEEE
International Conference on Robotics and Automation
(ICRA), pages 1468–1475. IEEE, 2015.
Karol Hausman, Scott Niekum, Sarah Osentoski, and
Gaurav Sukhatme. Active articulation model estimation
through interactive perception. In IEEE International
Conference on Robotics and Automation, 2015.
| 2cs.AI
|
Some results on generalized local cohomology modules
arXiv:1107.5079v4 [math.AC] 11 Sep 2013
Alireza Vahidi
Department of Mathematics, Payame Noor University, I. R of IRAN
E-mail: [email protected]
Moharram Aghapournahr
Department of Mathematics, Faculty of Science, Arak University, Arak, 38156-8-8349, IRAN
E-mail: [email protected]
Abstract
Let R be a commutative Noetherian ring with non-zero identity, a an ideal of R, M a finite
R–module and X an arbitrary R–module. Here, we show that, in the Serre subcategories of
the category of R–modules, how the generalized local cohomology modules, the ordinary local
cohomology modules and the extension modules behave similarly at the initial points. We conclude some Artinianness and cofiniteness results for Hna (M, X), and some finiteness results for
SuppR (Hna (M, X)) and AssR (Hna (M, X)).
Keywords: Generalized local cohomology modules; Serre subcategories.
2010 Mathematics Subject Classification: 13D07; 13D45.
1
Introduction
Let R be a commutative Noetherian ring with non-zero identity. We use symbols a, M , and X as
an ideal of R, a finite (i.e. finitely generated) R–module, and an arbitrary R–module which is not
necessarily finite. For basic results, notations and terminologies not given in this paper, the reader is
referred to [4] and [5].
The ith generalized local cohomology module
Hia (M, X) ∼
ExtiR (M/an M, X),
= lim
−→
n∈N
which is a generalization of the ith ordinary local cohomology module
Hia (X) ∼
ExtiR (R/an , X),
= lim
−→
n∈N
was introduced by Herzog in his habilitation [12] and then continued by Suzuki [21], Bijan-Zadeh [3],
Yassemi [22] and some other authors. They studied some basic duality theorems, vanishing and other
properties of generalized local cohomology modules which also generalize several known facts about
extension modules and ordinary local cohomology modules.
In Section 2, we present the main results of this paper which determine that, for non-negative
n
integers m and n, when R–modules Hna (M, X), HomR (R/a, Hna (M, X)) and Extm
R (M, Ha (X)) are in
1
2
A. Vahidi and M. Aghapournahr
a Serre subcategory of the category of R–modules (i.e. the class of R–modules which is closed under
n
taking submodules, quotients and extensions), and when Hm+n
(M, X) ∼
= Extm
a
R (M, Ha (X)) holds
(Theorems 2.6, 2.10, 2.13 and 2.21). We use these theorems to show that, in the Serre subcategories
of the category of R–modules, how the generalized local cohomology modules and the ordinary local
cohomology modules behave similarly at the initial points (Corollaries 2.9, 2.14 and 2.15). We also find,
in Corollaries 2.17 and 2.18, the relation of regular sequences with respect to Serre classes, introduced
in [1, Definition 2.6], and the membership of generalized local cohomology modules, extension modules
and Koszul cohomology modules in Serre subcategories. Note that, one can apply our results to the
Serre subcategories of Examples 2.2 and 2.8 to deduce more properties of generalized local cohomology
modules.
Section 3 consists of applications. We first, in Corollaries 3.1 and 3.2, study Artinian generalized
local cohomology modules and show that if dimR (R/a) = 0, then the generalized local cohomology
modules Hna (M, X) are Artinian and (a+AnnR M )–cofinite (i.e. SuppR (Hna (M, X)) ⊆ V(a+AnnR M )
and ExtiR (R/a + AnnR M, Hna (M, X)) is finite for all i). Then we present the relation between length,
annihilator and support of generalized local cohomology modules, and those of ordinary local cohomology modules (Corollaries 3.3, 3.4 and 3.5). We also prove that
S
S
S
SuppR (ExtiR (M/aM, X)),
SuppR (Hia+AnnR M (X)) =
SuppR (Hia (M, X)) =
i<n
i<n
i<n
and if Hia (M, X) = 0 for all i < n, then AssR (Hna (M, X)) = AssR (ExtnR (M/aM, X)) (Corollaries
3.6 and 3.8). This implies that if SuppR (Hia (M, X)) is finite for all i < n, then the finiteness of
AssR (Hna (M, X)) is equivalent to the finiteness of AssR (ExtnR (M/aM, X)). Finally, in the study of
finiteness of the set of associated prime ideals of generalized local cohomology modules, we point out
the proof of [14, Theorem 2.3] contains a flaw, but we show that the statements of [14, Corollaries 2.4
trough 2.7] are true (Remark 3.7 and Corollaries 3.11 trough 3.14).
Even though we can show some of our results by using spectral sequences, we are avoiding the use
of this technique completely in this work and we provide more elementary proofs for the results.
2
Main Results
Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative integer. We first
present sufficient conditions which convince us the R–modules Hna (M, X) and HomR (R/a, Hna (M, X))
are in a Serre subcategory of the category of R–modules.
Definition 2.1. Recall that a Serre subcategory S of the category of R–modules is a subclass of
R–modules such that for any short exact sequence
0 −→ X 0 −→ X −→ X 00 −→ 0,
the module X is in S if and only if X 0 and X 00 are in S.
Example 2.2. The following classes are Serre subcategories of the category of R–modules.
(a) The class of zero R–modules.
(b) The class of finite length R–modules.
(c) The class of finite R–modules.
Some results on generalized local cohomology
3
(d) The class of Artinian R–modules.
(e) The class of R–modules with finite support.
(f) The class of R–modules with Krull dimension less than n, where n is a non-negative integer.
(g) The class of R–modules with finite Krull dimension.
(h) The class of minimax R–modules (An R–module X is said to be minimax if there is a finite
submodule X 0 of X such that X/X 0 is Artinian [23]).
Definition 2.3. Let λ : S −→ T be a function from a Serre subcategory of the category of R–modules
S to a partially ordered Abelian monoid (T , F, ). We say that λ is a subadditive function if λ(0) = 0
and for any short exact sequence
0 −→ X 0 −→ X −→ X 00 −→ 0,
in which all the terms belong to S, λ(X 0 ) λ(X), λ(X 00 ) λ(X) and λ(X) λ(X 0 )Fλ(X 00 ).
Example 2.4. The following functions are subadditive.
(a) The function λ(X) = lR (X), length of X, from the class of finite length R–modules to the
partially ordered Abelian monoid (Z, +, ≤).
(b) The function λ(X) = (0 :R X), annihilator of X, from the category of R–modules to the partially
ordered Abelian monoid (Ideals(R), ., ⊇).
(c) The function λ(X) = SuppR (X), support of X, from the category of R–modules to the partially
ordered Abelian monoid (P(Spec R), ∪, ⊆).
In this paper, S is a Serre subcategory of the category of R–modules, (T , F, ) is a partially
ordered Abelian monoid and λ : S −→ T is a subadditive function.
Our method to prove the main results of the paper is based on the induction argument and we
need the following useful lemmas for the base cases and inductive steps. Note that, for all i, we have
the isomorphism
Hia (M, X) ∼
= Hi (Γa (HomR (M, E • ))),
where E • is an injective resolution of X.
Lemma 2.5. Let M be a finite R–module, X be an arbitrary R–module and p be a prime ideal of R.
Then the following statements hold true.
(a) Γa (M, X) ∼
= HomR (M, Γa (X)).
(b) Hia (M, X)p ∼
= HiaRp (Mp , Xp ) for all i.
(c) If SuppR (M ) ∩ SuppR (X) ⊆ V(a), then Hia (M, X) ∼
= ExtiR (M, X) for all i.
Proof. This is easy and left to the reader.
Theorem 2.6. Let M be a finite R–module, X be an arbitrary R–module, and n be a non-negative
r
n
integer such that Extn−r
R (M, Ha (X)) is in S for all r, 0 6 r 6 n. Then Ha (M, X) ∈ S, and
n
r
λ(Hna (M, X)) F λ(Extn−r
R (M, Ha (X))).
r=0
4
A. Vahidi and M. Aghapournahr
Proof. We prove by using induction on n. The case n = 0 is clear from Lemma 2.5 (a). Suppose that
n > 0 and that n − 1 is settled. Let X = X/Γa (X) and L = E(X)/X where E(X) is an injective hull
of X. Since Γa (X) = 0 = Γa (E(X)), Γa (M, X) = 0 = Γa (M, E(X)) by Lemma 2.5 (a). Applying the
derived functors of Γa (−) and Γa (M, −) to the short exact sequence
0 → X → E(X) → L → 0,
we obtain, for all i > 0, the isomorphisms
i−1
∼ i
∼ i
∼ i
Hi−1
a (L) = Ha (X) (= Ha (X)) and Ha (M, L) = Ha (M, X).
From the above isomorphisms, for all r, 0 6 r 6 n − 1, we have
(n−1)−r
n−(r+1)
ExtR
(M, Hra (L)) ∼
(M, Hr+1
= ExtR
a (X))
which is in S by assumptions. Thus, from the induction hypothesis on L,
Hn−1
(M, L) ∈ S
a
n−1
(n−1)−r
and λ(Hn−1
(M, L)) F λ(ExtR
a
Therefore
Hna (M, X) ∈ S
r=0
(M, Hra (L))).
n
n−r
(M, Hra (X))).
and λ(Hna (M, X)) F λ(ExtR
r=1
Now, by the short exact sequence
0 → Γa (X) → X → X → 0
and Lemma 2.5 (c), we get the long exact sequence
· · · −→ ExtnR (M, Γa (X)) −→ Hna (M, X) −→ Hna (M, X) −→ · · ·
which shows that
n
n−r
(M, Hra (X)))
Hna (M, X) ∈ S and λ(Hna (M, X)) F λ(ExtR
r=0
as we desired.
Definition 2.7. ([1, Definition 2.1] and [2, Definition 3.1]) Recall that, a Serre subcategory of the
category of R–modules M is said to be Melkersson subcategory with respect to the ideal a if for any a–
torsion R–module X, 0 :X a is in M implies that X is in M. Also, M is called Melkersson subcategory
when it is Melkersson with respect to all ideals of R.
Example 2.8. The following classes of modules are Melkersson subcategories by Example 2.2 and [1,
Lemma 2.2].
(a)
(b)
(c)
(d)
(e)
The
The
The
The
The
class
class
class
class
class
of
of
of
of
of
zero R–modules.
Artinian R–modules.
R–modules with finite support.
R–modules with Krull dimension less than n, where n is a non-negative integer.
R–modules with finite Krull dimension.
In this paper, Ma stands as a Melkersson subcategory with respect to the ideal a, Ma+AnnR M as
a Melkersson subcategory with respect to the ideal a + AnnR M and M as a Melkersson subcategory.
The second author and Melkersson in [1, Theorem 2.9 (i) ↔ (vi)] proved the following corollary
for Melkersson subcategories, while it was a simple conclusion of Theorem 2.6 for any arbitrary Serre
subcategories. This also generalizes [19, Theorem 2.2 and Corollary 2.3] for an arbitrary R–module
X when we consider S as the class of minimax R–modules and the class of Artinian R–modules,
respectively.
Some results on generalized local cohomology
5
Corollary 2.9. Suppose that X is an arbitrary R–module and that n is a non-negative integer. Then
the following statements are equivalent.
(i) Hia (X) is in S for all i 6 n (for all i).
(ii) Hia (M, X) is in S for any finite R–module M and for all i 6 n (for all i).
Proof. (i) ⇒ (ii). Assume that i is an integer such that i 6 n. Since Hra (X) is in S for all r, 0 6 r 6 i,
i−r
ExtR
(M, Hra (X)) is in S for all r, 0 6 r 6 i. Thus, by Theorem 2.6, Hia (M, X) is in S.
Theorem 2.10. Let M be a finite R–module, X be an arbitrary R–module, and n be a non-negative
integer. Then the following statements hold true.
(a) If Hra (X) ∈ S for all r, 0 6 r < n, then HomR (R/a, Hna (M, X)) ∈ S whenever ExtnR (M/aM, X) ∈
S.
(b) If Hra (X) = 0 for all r, 0 6 r < n, then HomR (R/a, Hna (M, X)) ∼
= ExtnR (M/aM, X).
Proof. We prove by using induction on n. From
HomR (R/a, Γa (M, X)) ∼
=
∼
=
∼
=
∼
=
Lemma 2.5 (a), we get
HomR (R/a, HomR (M, Γa (X)))
HomR (R/a ⊗R M, Γa (X))
HomR (M/aM, Γa (X))
HomR (M/aM, X)
because HomR (M/aM, X/Γa (X)) = 0. Thus the assertion follows in the case that n = 0. Suppose
that n > 0 and that n − 1 is settled. To complete the induction argument, one can use the short exact
sequence
0 → X → E(X) → L → 0
and employ the induction hypothesis with a similar method as in the proof of Theorem 2.6.
Remark 2.11. Theorem 2.6, Corollary 2.9 and Theorem 2.10 can be applied to each Serre subcategory mentioned in Example 2.2 resulting in each case in a number of facts about generalized local
cohomology modules. One can also use the Serre subcategories and Melkersson subcategories of Examples 2.2 and 2.8 in the results that follow to deduce more properties of generalized local cohomology
modules.
As an application of the above theorem, we can state the following corollary.
Corollary 2.12. Let M be a finite R–module, X be an arbitrary R–module, and n be a non-negative
integer such that Hia (X) ∈ Ma for all i < n. Then Hna (M, X) ∈ Ma whenever ExtnR (M/aM, X) ∈ Ma .
Proof. Since Hna (M, X) is an a–torsion R–module, the assertion follows from Theorem 2.10 (a).
Now, for non-negative integers m and n, we present sufficient conditions which ensure us the
n
R–module Extm
R (M, Ha (X)) is in a Serre subcategory of the category of R–modules.
Theorem 2.13. Let M be a finite R–module, X be an arbitrary R–module, and m, n be non-negative
integers. Assume also that
(i) Hm+n
(M, X) is in S,
a
6
A. Vahidi and M. Aghapournahr
(ii) Extm+1+r
(M, Hn−r
(X)) is in S for all r, 1 ≤ r ≤ n, and
a
R
m−1−r
n+r
(iii) ExtR
(M, Ha (X)) is in S for all r, 1 ≤ r ≤ m − 1.
n
Then Extm
R (M, Ha (X)) ∈ S, and
n
λ(Extm
R (M, Ha (X)))
n
m−1
r=1
r=1
m+1+r
m−1−r
λ(Ham+n (M, X))F( F λ(ExtR
(M, Han−r (X))))F( F λ(ExtR
(M, Hn+r
(X)))).
a
Proof. We prove by induction on n. Let n = 0 and set X = X/ Γa (X). By hypothesis (iii),
(m−1)−r
ExtR
(M, Hra (X)) is in S for all r, 0 6 r 6 m − 1. Thus, from Theorem 2.6,
Hm−1
(M, X) ∈ S
a
m−1
(m−1)−r
and λ(Hm−1
(M, X)) F λ(ExtR
a
r=1
(M, Hra (X))).
By considering Lemma 2.5 (c) and applying the derived functors of Γa (M, −) to the short exact sequence
0 → Γa (X) → X → X → 0,
we obtain the long exact sequence
m
· · · −→ Hm−1
(M, X) −→ Extm
a
R (M, Γa (X)) −→ Ha (M, X) −→ · · ·
which shows that, by hypothesis (i), Extm
R (M, Γa (X)) ∈ S and
m−1
m−1−r
m
(M, Hra (X)))).
λ(Extm
R (M, Γa (X))) λ(Ha (M, X))F( F λ(ExtR
r=1
Thus the assertion follows in this case.
Now, assume that n > 0 and that n − 1 is settled. Let X = X/Γa (X) and L = E(X)/X where
E(X) is an injective hull of X. By the short exact sequence
0 → X → E(X) → L → 0,
the proof is sufficiently similar to that of Theorem 2.6 to be omitted. We leave the proof to the
reader.
The next corollary shows that, in Melkersson subcategories, the generalized local cohomology
modules Hia (M, X) and the ordinary local cohomology modules Hia+AnnR M (X) behave similarly at the
initial points.
Corollary 2.14. Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative
integer. Then the following statements are equivalent.
(i) Hia+AnnR M (X) is in Ma+AnnR M for all i 6 n (for all i).
(ii) Hia (M, X) is in Ma+AnnR M for all i 6 n (for all i).
Proof. (i) ⇒ (ii). Since Hia (M, X) ∼
= Hia+AnnR M (M, X) for all i, the assertion holds from Corollary
2.9.
(ii) ⇒ (i). We use induction on n. Let n = 0. By considering the exact sequence
0 −→ HomR (M/aM, Γa (X)) −→ HomR (M, Γa (X)),
HomR (M/aM, Γa (X)) is in Ma+AnnR M from Lemma 2.5 (a). Thus, by [1, Theorem 2.9 (iv) → (i)],
Γa+AnnR M (Γa (X)) is in Ma+AnnR M . Therefore Γa+AnnR M (X) is in Ma+AnnR M .
Assume that n > 0 and that n − 1 is settled. By the induction hypothesis, Hia+AnnR M (X) is in
Ma+AnnR M for all i 6 n − 1. Apply Theorem 2.13 with m = 0 to see that HomR (M, Hna+AnnR M (X))
Some results on generalized local cohomology
7
is in Ma+AnnR M . Thus HomR (M/aM, Hna+AnnR M (X)) is in Ma+AnnR M by the exact sequence
0 −→ HomR (M/aM, Hna+AnnR M (X)) −→ HomR (M, Hna+AnnR M (X)).
Again from [1, Theorem 2.9 (iv) → (i)], we have Γa+AnnR M (Hna+AnnR M (X)) ∈ Ma+AnnR M which
shows that Hna+AnnR M (X) ∈ Ma+AnnR M .
Corollary 2.15. Let M be a finite R–module and X be an arbitrary R–module. Then we have
(a)
(b)
(c)
(d)
inf{i : Hia (M, X) ∈
/ S} ≥ inf{i : Hia (X) ∈
/ S}.
i
inf{i : Ha (M, X) ∈
/ Ma+AnnR M } = inf{i : Hia+AnnR M (X) ∈
/ Ma+AnnR M }.
i
i
inf{i : Ha (M, X) ∈
/ Ma+AnnR M } = inf{i : Ha (X) ∈
/ Ma+AnnR M } whenever ΓAnnR M (X) = X.
i
i
inf{i : Ha (M, X) ∈
/ Ma } = inf{i : Ha (X) ∈
/ Ma } whenever AnnR M ⊆ a (e.g. M is faithful).
Proof. Follows from Corollaries 2.9 and 2.14.
In the next corollary, we state the membership of the generalized local cohomology modules with
respect to different ideals in Melkersson subcategories of the category of R–modules.
Corollary 2.16. Suppose that M is a finite R–module and X is an arbitrary R–module. Assume also
that n is a non-negative integer and b is an ideal of R such that a ⊆ b. Then Hib (M, X) is in M for
all i ≤ n (for all i) whenever Hia (M, X) is in M for all i ≤ n (for all i).
Proof. Follows from Corollary 2.14 and [2, Proposition 3.4].
In [1, Definition 2.6 and Example 2.8], the second author and Melkersson introduced the concept of
S–regular sequences on a module that recovered poor sequences, filter-regular sequences, generalized
regular sequences and sequences in dimension> n, where n is a non-negative integer, on a module.
They also found, in [1, Theorem 2.9 (i) ↔ (vii)], the relation of this notion on a finite module and
the membership of local cohomology modules in Melkersson subcategories. In the next corollary, we
state a similar characterization for generalized local cohomology modules. Coung and Hoang in [8,
Theorem 3.1] proved part [(i) ↔ (iv)] of the following corollary for the class of Artinian R–modules
in the case that R was a local ring.
Corollary 2.17. Suppose that M is a finite R–module such that a + AnnR M = (x1 , . . . , xr ). Assume also that X is an arbitrary R–module and that n is a non-negative integer. Then the following
statements are equivalent.
(i) Hia (M, X) is in Ma+AnnR M for all i 6 n (for all i).
(ii) ExtiR (M/aM, X) is in Ma+AnnR M for all i 6 n (for all i).
(iii) Hi (x1 , . . . , xr ; X) is in Ma+AnnR M for all i 6 n (for all i).
When X is finite, these conditions are also equivalent to:
(iv) There is a sequence of length n + 1 in a + AnnR M that is Ma+AnnR M –regular on X.
Proof. This follows from Corollary 2.14 and [1, Theorem 2.9].
8
A. Vahidi and M. Aghapournahr
Suppose that X is a finite R–module such that X/aX is not in Ma . The second author and
Melkersson, in [1, Lemma 2.14], proved that every sequence in a which is Ma –regular on X can
be extended to a maximal one and all maximal Ma –regular sequences on X in a have the same
length. They denoted this common length by Ma –deptha (X), in [1, Definition 2.15], and proved, in
[1, Theorem 2.18], that it is the least integer such that Hia (X), ExtiR (R/a, X) or Koszul cohomology
modules with respect to a are not in Ma . Using the Melkersson subcategories of Example 2.8, this
notion gives ordinary depth, filter-depth, generalized depth and n-depth, where n is a non-negative
integer. In the following, we prove that Ma+AnnR M − deptha+AnnR M (X) is the least integer such
that Hia (M, X), ExtiR (M/aM, X) or Koszul cohomology modules with respect to a + AnnR M are
not in Ma+AnnR M . This generalizes the result of Bijan-Zadeh [3, Proposition 5.5] when we consider
Ma+AnnR M as the class of zero R–modules. It also recovers [6, Theorem 2.2], [8, Theorem 3.1], [7,
Theorem 4.1] and [16, Theorem 2.8] if we put Ma+AnnR M the class of Artinian R–modules or the
class of R–modules with finite support. Note that, all of these theorems are in the local case while
our corollary is in general.
Corollary 2.18. Suppose that M is a finite R–module with a + AnnR M = (x1 , . . . , xr ) and X is a
finite R–module with X/(a + AnnR M )X ∈
/ Ma+AnnR M . Then
(a) Ma+AnnR M − deptha+AnnR M (X) = inf{i : Hia (M, X) ∈
/ Ma+AnnR M }.
i
(b) Ma+AnnR M − deptha+AnnR M (X) = inf{i : ExtR (M/aM, X) ∈
/ Ma+AnnR M }.
i
(c) Ma+AnnR M − deptha+AnnR M (X) = inf{i : H (x1 , . . . , xr ; X) ∈
/ Ma+AnnR M }.
Proof. Follows from [1, Lemma 2.14] and Corollary 2.17.
As applications of Theorems 2.6 and 2.13, we can state the following corollaries.
Corollary 2.19. Let M be a finite R–module, X be an arbitrary R–module, and n be a non-negative
i
integer such that Extj−i
R (M, Ha (X)) is in S for all i, j with 0 ≤ i ≤ n − 1 and j = n, n + 1. Then
Hna (M, X) is in S if and only if HomR (M, Hna (X)) is in S.
Corollary 2.20. Let X be an R–module and m, n be non-negative integers such that Hia (X) is in S
n
for all i, 0 6 i 6 n−1 or n+1 6 i 6 m+n. Then Hm+n
(M, X) is in S if and only if Extm
a
R (M, Ha (X))
is in S.
In the following theorem, for non-negative integers m and n, we find some sufficient conditions for
n
validity of the isomorphism Hm+n
(M, X) ∼
= Extm
a
R (M, Ha (X)).
Theorem 2.21. Let M be a finite R–module, X be an arbitrary R–module, and m, n be non-negative
integers. Assume also that
(i) Extm+n−r
(M, Hra (X)) = 0 for all r, 0 ≤ r ≤ n − 1 or n + 1 ≤ r ≤ m + n,
R
(ii) Extm+1+r
(M, Hn−r
(X)) = 0 for all r, 1 ≤ r ≤ n, and
a
R
m−1−r
n+r
(iii) ExtR
(M, Ha (X)) = 0 for all r, 1 ≤ r ≤ m − 1.
n
Then we have Hm+n
(M, X) ∼
= Extm
a
R (M, Ha (X)).
Some results on generalized local cohomology
9
Proof. We prove by using induction on n. Let n = 0. We have Hm−1
(M, X/Γa (X)) = 0 =
a
m
Ha (M, X/Γa (X)) from hypothesis (iii) and (i), and Theorem 2.6 with S = 0. Now, the assertion
follows by the exact sequence
m
m
Hm−1
(M, X/Γa (X)) −→ Extm
a
R (M, Γa (X)) −→ Ha (M, X) −→ Ha (M, X/Γa (X))
obtained from the short exact sequence
0 −→ Γa (X) −→ X −→ X/Γa (X) −→ 0
and Lemma 2.5 (c).
Assume that n > 0 and that n − 1 is settled. By considering the short exact sequence
0 → X → E(X) → L → 0,
the proof is similar to that of Theorem 2.6.
Yassemi, in [22, Example 3.6], has given an example to show that the R–modules Hna (M, X) and
HomR (M, Hna (X)) are not always equal. We show that, with some conditions, they are isomorph.
Corollary 2.22. (cf. [13, Proposition 2.3 (ii)]) Let M be a finite R–module, X be an arbitrary R–
i
module and n be a non-negative integer such that Extj−i
R (M, Ha (X)) = 0 for all i, j with 0 6 i 6 n − 1
and j = n, n + 1. Then we have Hna (M, X) ∼
= HomR (M, Hna (X)).
Proof. Apply Theorem 2.21 with m = 0.
In consistence with Corollary 2.20, one can state the following corollary which shows that if X is a
finite module and a is an ideal generated by an X-regular sequence of length n, then the generalized
local cohomology modules are exactly extension modules of ordinary local cohomology modules.
Corollary 2.23. Suppose that M is a finite R–module, X is an arbitrary R–module, and n, m are nonnegative integers such that n ≤ m. Assume also that Hia (X) = 0 for all i, i 6= n (resp. 0 ≤ i ≤ n − 1 or
i
n
∼
n + 1 ≤ i ≤ m). Then we have Hi+n
a (M, X) = ExtR (M, Ha (X)) for all i, i ≥ 0 (resp. 0 ≤ i ≤ m − n).
Proof. For all i, i ≥ 0 (resp. 0 ≤ i ≤ m − n), apply Theorem 2.21 with m = i.
3
Applications
Recall that, an R–module X is said to be a–cofinite if SuppR (X) ⊆ V (a) and ExtiR (R/a, X) is finite
for all i. Note that, by [18, Proposition 4.1], the class of Artinian a–cofinite modules is a Melkersson
subcategory with respect to the ideal a.
Corollary 3.1. Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative
integer. Then the following statements are equivalent.
(i) Hia (M, X) is Artinian and (a + AnnR M )–cofinite for all i 6 n (for all i).
(ii) ExtiR (M/aM, X) has finite length for all i 6 n (for all i).
Proof. (i) ⇒ (ii). From Corollary 2.14, Hia+AnnR M (X) is Artinian and (a + AnnR M )–cofinite for all
i 6 n. Thus ExtiR (R/a + AnnR M , X) has finite length for all i 6 n by [2, Corollary 4.12]. Therefore,
from [13, Proposition 3.4], ExtiR (M/aM, X) has finite length for all i 6 n.
10
A. Vahidi and M. Aghapournahr
(ii) ⇒ (i). Since every finite length (a + AnnR M )–torsion module is Artinian and (a + AnnR M )–
cofinite, the assertion follows from Corollary 2.17.
Chu and Tang in [6, Proposition 2.4] proved the part [(i) ↔ (ii)] of the following corollary in the
local case (see also [8, Corrollary 3.2], [9, Corollary 3.3] and [11, Theorem 2.2]).
Corollary 3.2. Suppose that M, X are finite R–modules and that n is a non-negative integer. Then
the following statements are equivalent.
(i) dimR (Hia (M, X)) ≤ 0 for all i 6 n (for all i).
(ii) Hia (M, X) is Artinian for all i 6 n (for all i).
(iii) Hia (M, X) is Artinian and (a + AnnR M )–cofinite for all i 6 n (for all i).
In particular, if dimR (R/a) = 0, then Hia (M, X) is Artinian and (a + AnnR M )–cofinite for all i.
Proof. (i) ⇒ (iii). Since every finite module with zero dimension is of finite length, the assertion
follows from Corollary 2.17 [(i) → (ii)] (where Ma+AnnR M is taken the class of R–modules with Krull
dimension less than 1) and Corollary 3.1 [(ii) → (i)].
In [20, Theorem 3.2] and for a non-negative integer n, Schenzel proved that
• ExtnR (M, X) is of finite length, and
n
X
i
• lR (ExtnR (M, X)) ≤
lR (Extn−i
R (M, Hm (X)))
i=0
when (R, m) is a local ring and M, X are finite R–modules such that M ⊗R X is of finite length.
As an application of Theorem 2.6, by considering Lemma 2.5 (c) and [20, Lemma 3.1], the following
corollary extends [20, Theorem 3.2].
Corollary 3.3. Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative
i
integer such that Extn−i
R (M, Ha (X)) is of finite length for all i 6 n. Then
(a) Hna (M, X) is of finite length, and
n
X
i
n
(b) lR (Ha (M, X)) ≤
lR (Extn−i
R (M, Ha (X))).
i=0
Proof. Since the class of finite length R–modules is a Serre subcategory of the category of R–modules
and λ(X) = lR (X) is a subadditive function from the class of finite length R–modules to the partially
ordered Abelian monoid (Z, +, ≤), the assertion follows form Theorem 2.6.
As another application of Theorem 2.6, we find the relation between annihilator of generalized
local cohomology modules and annihilator of ordinary local cohomology modules.
Corollary 3.4. Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative
integer. Then we have
n
Y
i
n
(a)
(0 :R Extn−i
R (M, Ha (X))) ⊆ (0 :R Ha (M, X)).
i=0
Some results on generalized local cohomology
(b)
n
Y
(0
i=0
:R Hia (X))
⊆
n
\
11
(0 :R Hia (M, X)).
i=0
Proof. (a) Since λ(X) = (0 :R X) is a subadditive function from the category of R–modules to the
partially ordered Abelian monoid (Ideals(R), ., ⊇), the assertion follows form Theorem 2.6.
i
(b) For all i 6 j 6 n, we have (0 :R Hia (X)) ⊆ (0 :R Extj−i
R (M, Ha (X))). Thus the assertion follows
from part (a).
In the course of the remaining parts of the paper for an ideal a of R and for an arbitrary R–module
X, by cdR (a, X) (cohomological dimension of X with respect to a), we mean the largest integer i in
which Hia (X) is non-zero. The next result presents the relation between support of generalized local
cohomology modules and support of ordinary local cohomology modules.
Corollary 3.5. Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative
integer. Then we have
[
n−i
(M, Hia (X))).
SuppR (ExtR
(a) SuppR (Hna (M, X)) ⊆
(b)
i≤n [
[
SuppR (Hia (M, X)) ⊆
i≤n
SuppR (Hia (X)).
i≤n
In particular, SuppR (Hna (M, X)) ⊆
[
SuppR (Hia (X)).
i≤cdR (a,X)
Proof. (a) Since λ(X) = SuppR (X) is a subadditive function from the category of R–modules to the
partially ordered Abelian monoid (P(Spec R), ∪, ⊆), the assertion follows form Theorem 2.6.
(b) Follows from the first part.
The vanishing of generalized local cohomology modules from upper bounds needs special conditions
and in all of them M must have finite projective dimension (see [22, Theorems 2.5 and 3.7], [7, Theorem
3.1] and [13, Proposition 2.8]). However, in the following corollary, we show that there is a union of
finitely many supports of generalized local cohomology modules such that the other supports can be
viewed as its subset even if M has infinite projective dimension. Parts (a) and (b) of the following
corollary in the local case has been proven in [7, Lemma 2.8 and Corollary 2.9] by Coung and Hoang
when X is a finite R–module but we prove it without assuming that X is finite and with no restrictions
on R.
Corollary 3.6. Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative
integer. Then the following statements hold true.
[
[
(a)
SuppR (Hia (M, X)) =
SuppR (Hia+AnnR M (X)).
(b)
(c)
(d)
i≤n
[
i≤n
[
i≤n
[
SuppR (Hia (M, X)) =
i≤n
[
SuppR (ExtiR (M/aM, X)).
i≤n
SuppR (Hia (M, X)) is a closed set when X is a finite R–module.
SuppR (Hia (M, X)) is a closed set when X is a finite R–module.
i
12
A. Vahidi and M. Aghapournahr
[
In particular, SuppR (Hna (M, X)) ⊆
SuppR (Hia (M, X)).
i≤cdR (a+AnnR M,X)
Proof. (a) By Lemma 2.5 (b) and Corollary 2.14, we have
S
p∈
/
SuppR (Hia (M, X)) ⇔ ∀i ≤ n;
Hia (M, X)p = 0
i≤n
⇔ ∀i ≤ n;
HiaRp (Mp , Xp ) = 0
⇔ ∀i ≤ n;
HiaRp +AnnRp Mp (Xp ) = 0
⇔ ∀i ≤ n;
Hi
(X)p = 0
S a+AnnR M i
⇔
p∈
/
SuppR (Ha+AnnR M (X))
i≤n
as we desired.
(b) From Lemma 2.5 (b) and Corollary 2.17 [(i) ↔ (ii)], we get
S
p∈
/
SuppR (Hia (M, X)) ⇔ ∀i ≤ n;
Hia (M, X)p = 0
i≤n
⇔ ∀i ≤ n;
HiaRp (Mp , Xp ) = 0
⇔ ∀i ≤ n;
ExtiRp (Mp /(aRp )Mp , Xp ) = 0
⇔ ∀i ≤ n;
ExtiR (M/aM, X)p = 0
S
SuppR (ExtiR (M/aM, X))
⇔
p∈
/
i≤n
as desired.
(c) This is clear from the second part.
(d) By the first part, we have
[
SuppR (Hia (M, X)) =
i
[
SuppR (Hia (M, X)).
i≤cdR (a+AnnR M,X)
Thus the assertion follows from part (c).
Remark 3.7. Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative
integer. In [14, Theorem 2.3], Mafi proved
S
n−i
(M, Hia (X)))
AssR (Hna (M, X)) ⊆ ni=0 AssR (ExtR
and, in [14, Corollaries 2.4 through 2.7], used it to deduce some results about finiteness of the set of associated prime ideals of generalized local cohomology modules. Although Corollaries 2.4 through 2.7 in
i,t−i
[14] are true, the proof of [14, Theorem 2.3] holds a flaw. In its proof, even though (E∞
=) ker di,t−i
t+2
i,t−i
is a subquotient of ker di,t−i
(⊆
E
),
2
2
i,t−i
i,t−i
(E∞
=) ker di,t−i
(⊆ E2i,t−i )
t+2 ⊆ ker d2
is not necessarily true and so dose not assert that
i,t−i
AssR (E∞
) ⊆ AssR (E2i,t−i ).
In the followings, we state some results about finiteness of the set of associated prime ideals of generalized local cohomology modules which, among other things, establish the statements of [14, Corollaries
2.4 through 2.7].
Coung and Hoang in [8, Theorem 2.4] proved the following corollary when X is a finite module.
Corollary 3.8. Let M be a finite R–module, X be an arbitrary R–module and n be a non-negative
integer such that Hia (M, X) = 0 for all i < n. Then we have
AssR (Hna (M, X)) = AssR (ExtnR (M/aM, X)).
Some results on generalized local cohomology
13
Proof. Since, by Corollary 2.14, Hia+AnnR M (X) = 0 for all i < n,
HomR (R/a + AnnR M, Hna (M, X)) ∼
= ExtnR (M/(a + AnnR M )M, X)
from Theorem 2.10 (b). Thus we have
AssR (Hna (M, X)) = AssR (Hna+AnnR M (M, X))
T
= V (a + AnnR M ) AssR (Hna+AnnR M (M, X))
= AssR (HomR (R/a + AnnR M, Hna (M, X)))
= AssR (ExtnR (M/(a + AnnR M )M, X))
= AssR (ExtnR (M/aM, X)),
as desired.
In [7, Theorem 4.5], part (a) of the following corollary has been proven when X is a finite module
and R is a local ring.
Corollary 3.9. Suppose that M is a finite R–module, X is an arbitrary R–module and n is a nonnegative integer. Assume also that
[
[
Pn =
SuppR (Hia (M, X)) (=
SuppR (ExtiR (M/aM, X))).
i<n
i<n
Then the following statements hold true.
(a)
(b)
(c)
(d)
AssR (Hna (M, X)) ∪ Pn = AssR (ExtnR (M/aM, X)) ∪ Pn .
AssR (Hna (M, X)) ⊆ AssR (ExtnR (M/aM, X)) ∪ Pn .
AssR (ExtnR (M/aM, X)) ⊆ AssR (Hna (M, X)) ∪ Pn .
If Hia (M, X) has finite support for all i < n, then AssR (Hna (M, X)) is a finite set if and only if
AssR (ExtnR (M/aM, X)) is a finite set.
Proof. (a) If p ∈
/
S
SuppR (Hia (M, X)), then
i<n
AssRp (HnaRp (Mp , Xp )) = AssRp (ExtnRp (Mp /(aRp )Mp , Xp ))
from Lemma 2.5 (b) and Corollary 3.8. Thus, p is not in the left side if and only if it is not in the
right side.
Recall that, an R–module X is said to be weakly Laskerian if the set of associated prime ideals of
any quotient module of X is finite ([10, Definition 2.1]). The category of weakly Laskerian R–modules
is a Serre subcategory of the category of R–modules ([10, Lemma 2.3 (i)]) and is denoted by Cw.l (R).
Corollary 3.10. (cf. [15, Lemma 3.1]) Let M be a finite R–module, X be an arbitrary R–module,
i
and n be a non-negative integer such that Extn−i
R (M, Ha (X)) is weakly Laskerian for all i 6 n. Then
(a) Hna (M, X) is weakly Laskerian.
(b) AssR (Hna (M, X)) is finite.
Proof. (a) Apply Theorem 2.6 with S = Cw.l (R).
(b) This is clear from part (a).
In the following corollary, we generalize [14, Corollary 2.4] and [15, Theorem 3.3]. Note that, for
a finite R–modules M and a non-negative integer n, X ∈ S implies that ExtnR (M/aM, X) ∈ S and,
by [13, Proposition 3.4], we have ExtnR (M/aM, X) ∈ S when ExtiR (R/a, X) ∈ S for all i ≤ n.
14
A. Vahidi and M. Aghapournahr
Corollary 3.11. (cf. [14, Corollary 2.4] and [15, Theorem 3.3]) Let M be a finite R–module, X be
an arbitrary R–module, and n be a non-negative integer such that ExtnR (M/aM, X) and Hia (X), for
all i < n, are weakly Laskerian. Then
(a) HomR (R/a, Hna (M, X)) is weakly Laskerian.
(b) AssR (Hna (M, X)) is finite.
Proof. (a) Apply Theorem 2.10 (a) with S = Cw.l (R).
T
(b) Since AssR (HomR (R/a, Hna (M, X))) = V (a) AssR (Hna (M, X)) = AssR (Hna (M, X)), the assertion follows from part (a).
Corollary 3.12. (cf. [14, Corollary 2.5]) Suppose that R is a local ring with maximal ideal m and
dim R ≤ 2. Assume also that M is a finite R–module and X is an arbitrary R–module such that
Γa (X) is weakly Laskerian. Then AssR (Hia (M, X)) is finite for all i.
Proof. This follows from Corollary 3.10 and [17, Corollaries 2.4 and 2.5].
Corollary 3.13. (cf. [14, Corollary 2.6]) Suppose that R is a local ring with maximal ideal m and
dim R = n. Assume also that M is a finite R–module and X is an arbitrary R–module such that
Hja (X) = 0 for all j 6= n − 1, n. Then AssR (Hia (M, X)) is finite for all i.
Proof. Follows from Corollary 3.10 and [17, Corollaries 2.4 and 2.5].
Corollary 3.14. (cf. [14, Corollary 2.7]) Suppose that R is a local ring with maximal ideal m and
dimR R/a = 1. Assume also that M is a finite R–module and X is an arbitrary R–module. Then
AssR (Hia (M, X)) is finite for all i.
Proof. It follows from Corollary 3.10 or Corollary 3.11.
References
[1] Aghapournahr, M., Melkersson, L. (2008). Local cohomology and Serre subcategories. J. Algebra
320:1275–1287.
[2] Aghapournahr, M., Taherizadeh, A. J., Vahidi, A. (2011). Extension functors of local cohomology
modules. Bull. Iran. Math. Soc. 37:117–134.
[3] Bijan-Zadeh, M. H. (1980). A commen generalization of local cohomology theories. Glasgow Math.
J. 21:173–181.
[4] Brodmann, M. P., Sharp, R. Y. (1998). Local cohomology: an algebraic introduction with geometric
applications. Cambridge, Cambridge University Press.
[5] Bruns, W., Herzog, J. (1998). Cohen-Macaulay rings. Cambridge, Cambridge University Press.
[6] Chu, L., Tang, Z. (2007). On the artinianness of generalized local cohomology. Commun. Algebra
35:3821–3827.
Some results on generalized local cohomology
15
[7] Coung, N. T., Hoang, N. V. (2001). On the vanishing and the finiteness of supports of generalized
local cohomology modules. Manuscripta Math. 104:519–525.
[8] Coung, N. T., Hoang, N. V. (2005). Some finite properties of generalized local cohomology modules.
East-West J. Math. 2:107–115.
[9] Dibaei, M. T., Vahidi, A. (2011). Artinian and non-Artinian local cohomology modules. Can. Math.
Bull. 54:619–629.
[10] Divaani-Aazar, K., Mafi, A. (2005). Associated primes of local cohomology modules. P. Am.
Math. Soc. 133:655–660.
[11] Divaani-Aazar, K., Sazeedeh, R., Tousi, M. (2005). On vanishing of generalized local cohomology
modules. Algebr. Colloq. 2:213–218.
[12] Herzog, J. (1970). Komplexe, auflösungen und dualität in der lokalen algebra. Invent. Math.
9:145–164.
[13] Hasanzadeh, S. H., Vahidi, A. (2009). On vanishing and cofinitness of generalized local cohomology modules. Commun. Algebra 37:2290–2299.
[14] Mafi, A. (2006). On the associated primes of generalized local cohomology modules. Commun.
Algebra 34:2489–2494.
[15] Mafi, A. (2009). A generalization of the finiteness problem in local cohomology modules. P. Indian
A. S. (Math. Sci.) 119:159–164.
[16] Mafi, A. (2009). On the finiteness results of the generalized local cohomology modules. Algebr.
Colloq. 16:325–332.
[17] Marley, T. (2001). The associoated primes of local cohomology modules of small dimension.
Manuscripta Math. 104:519–525.
[18] Melkersson, L. (2005). Modules cofinite with respect to an ideal. J. Algebra 285:649–668.
[19] Saremi, H. (2009). On minimax and generalized local cohomology modules. Acta. Math. Vietnamica 34:269–273.
[20] Schenzel, P. (1998). On the use of local cohomology in algebra and geometry. Prog. Math. 166:241–
292.
[21] Suzuki, N. (1978). On the generalized local cohomology and its duality. J. Math. Kyoto U. 18:71–
85.
[22] Yassemi, S. (1994). Generalized section functors. J. Pure Appl. Algebra 95:103–119.
[23] Zöschinger, H. (1986). Minimax moduln. J. Algebra 102:1–32.
| 0math.AC
|
arXiv:1410.3735v1 [cs.PL] 14 Oct 2014
The Foundational Cryptography Framework
Adam Petcher
Greg Morrisett
Harvard University and MIT Lincoln Laboratory
[email protected]
Harvard University
[email protected]
Abstract
We present the Foundational Cryptography Framework (FCF) for
developing and checking complete proofs of security for cryptographic schemes within a proof assistant. This is a general-purpose
framework that is capable of modeling and reasoning about a wide
range of cryptographic schemes, security definitions, and assumptions. Security is proven in the computational model, and the proof
provides concrete bounds as well as asymptotic conclusions. FCF
provides a language for probabilistic programs, a theory that is used
to reason about programs, and a library of tactics and definitions
that are useful in proofs about cryptography. The framework is designed to leverage fully the existing theory and capabilities of the
Coq proof assistant in order to reduce the effort required to develop
proofs.
Categories and Subject Descriptors F.3.1 [LOGICS AND MEANINGS OF PROGRAMS]: Specifying and Verifying and Reasoning
about Programs; D.3.1 [PROGRAMMING LANGUAGES]: Formal Definitions and Theory
Keywords
Coq
Cryptography, Proof Assistant, Mechanized Proof,
1. Introduction
Cryptographic algorithms and protocols are becoming more numerous, specialized, and complicated. As a result, it is likely that security vulnerabilities will slip by peer review. To address this problem, some cryptographers [6, 16] have proposed an increased level
of rigor and formality for cryptographic proofs. It is our hope that
eventually, cryptographers will be able to describe cryptographic
schemes and security proofs using a formal language, and the
proofs can be checked automatically by a highly trustworthy mechanized proof checker.
To enable such mechanically-verified proofs, we have developed The Foundational Cryptography Framework (FCF). This
framework embeds into the Coq proof assistant [18] a simple
probabilistic programming language to allow the specification of
cryptographic schemes, security definitions, and assumptions. The
framework also includes useful theory, tactics, and definitions that
assist with the construction of proofs of security. Once complete,
the proof can be checked by the Coq proof checker. Facts proven
in FCF include the security of El Gamal encryption [14], and of
the encryption scheme described in Section 4 of this paper. We
have also proven the security of the “tuple-set” construction of
[10], which is a significant portion of a practical searchable symmetric encryption scheme. This is a complex and sophisticated
construction, and the proof requires over 7000 lines of Coq code
and includes a core argument involving more than 30 intermediate
games.
FCF is heavily influenced by CertiCrypt [4], which was later
followed by EasyCrypt [5]. CertiCrypt is a framework that is built
on Coq, and allows the development of mechanized proofs of
security in the computational model for arbitrary cryptographic
constructions. Unfortunately, proof development in CertiCrypt is
time-consuming, and the developer must spend a disproportionate amount of time on simple, uninteresting goals. To address
these limitations, the group behind CertiCrypt developed EasyCrypt, which has a similar semantics and logic, and uses the Why3
framework and SMT solvers to improve proof automation. EasyCrypt takes a huge step forward in terms of usability and automation, but it sacrifices some trustworthiness due to that fact that the
trusted computing base is larger and the basis of the mechanization
is a set of axiomatic rules.
Following the release of EasyCrypt, a team of cryptographers
and programming language experts (including one of the authors
of this paper) attempted [17] to prove the security of a private
information retrieval system [13]. This effort did not produce a
complete proof because certain required facts could not be proven
in EasyCrypt. Specifically, it was impossible to prove particular
equivalences involving loop fusion and order permutation within
a loop. In order to allow these equivalences in EasyCrypt, it would
be necessary to prove them correct on paper and then modify the
EasyCrypt code to include appropriate rules. EasyCrypt has seen
significant improvements since its release, and it is possible that
these sorts of equivalence are now supported, but a developer may
encounter some other goal that EasyCrypt does not support.
FCF is a foundational framework like CertiCrypt, in which the
rules used to prove equivalence of programs (or any fact) are mechanized proofs derived from the semantics or other core definitions.
In such a framework, the problem of the previous paragraph can
be addressed by proving the appropriate theorem within the framework, and then by using that theorem to obatain the desired equivalence. An important difference between CertiCrypt and FCF is that
CertiCrypt uses a deep embedding of a probabilistic programming
language whereas FCF uses a shallow embedding (similar to [20]).
The shallow embedding allows us to easily extend the language,
and to make better use of Coq’s tactic language and existing automated tactics to reduce the effort required to develop proofs. The
result is a framework that is foundational and easily extensible, but
in which proof development effort is greatly reduced.
2. Design Goals
Based on our experience working with EasyCrypt, we formulated a
set of idealized design goals that a practical mechanized cryptography framework should satisfy. We believe that FCF achieves many
of these goals, though there is still some room for improvement, as
discussed in Section 5.
Familiarity Security definitions and descriptions of cryptographic
schemes should look similar to how they would appear in cryptography literature, and a cryptographer with no knowledge of programming language theory or proof assistants should be able to
understand them. Furthermore, a cryptographer should be able to
inspect and understand the foundations of the framework itself.
Proof Automation The system should use automation to reduce
the effort required to develop a proof. Ideally, this automation is
extensible, so that the developer can produce tactics for solving
new kinds of goals.
Trustworthiness Proofs should be checked by a trustworthy procedure, and the core definitions (e.g., programming language semantics) that must be trusted in order to trust a proof should be
relatively simple and easy to understand.
Extensibility It should be possible to directly incorporate any
existing theory that has been developed for the proof assistant. For
example, it should be possible to directly incorporate an existing
theory of lattices in order to support cryptography that is based on
lattices and their related assumptions.
Concrete Security The security proof should provide concrete
bounds on the probability that an adversary is able to defeat the
scheme. Concrete bounds provide more information than asymptotic statements, and they inform the selection of values for system
parameters in order to achieve the desired level of security in practice.
Abstraction The system should support abstraction over types,
procedures, proofs, and modules containing any of these items.
Abstraction over procedures and primitive types is necessary for
writing security definitions, and for reasoning about adversaries in a
natural way. The inclusion of abstraction over proofs and structures
adds a powerful mechanism for developing sophisticated abstract
arguments that can be reused in future proofs.
Code Generation The system should be able to generate code
containing the procedures of the cryptographic scheme that was
proven secure. This code can then be used for basic testing, prototyping, or as an executable model to which future implementations
will be compared during testing.
the semantics of the language, and using them to complete a proof
does not reduce the trustworthiness of the proof.
By combining all of the components described above, a developer can produce a proof relating the probability that some adversary defeats the scheme to the probability that some other adversary
is able to solve a problem that is assumed to be hard. This is a result in the concrete setting, in which probability values are given as
expressions, and certain problems are assumed to be hard for particular constructed adversaries. In such a result, it may be necessary
to inspect an expression describing a probability value to ensure it
is sufficiently “small,” or to inspect a procedure to ensure it is in the
correct complexity class. FCF provides additional facilities to obtain more traditional asymptotic results, in which these procedures
and expressions do not require inspection. A set of asymptotic definitions (Section 3.6) allows conclusions like “this probability is
negligible” or “this procedure executes a polynomial number of
queries.” In order to apply an assumption about a hard problem, it
may be necessary to prove that some procedure is efficient in some
sense. So FCF provides an extensible notion of efficiency (Section
3.7) and a characterization of non-uniform polynomial time Turing
machines.1
3.1 Probabilistic Programs
We describe probabilistic programs using Gallina, the purely functional programming language of Coq, extended with a computational monad in the spirit of Ramsey and Pfeffer [21], that supports
drawing random bit vectors from an input tape. Listing 1 contains
an example of a valid FCF program that implements a one-time
pad using bit vectors. This program accepts a bit vector argument
x, samples a random bit vector of length c (where c is a constant
declared outside of this function) and assigns the result to variable
p, then returns p ⊕ x.
Definition OTP (x : Bvector c) : Comp (Bvector c)
:= p <-$ {0, 1}ˆc; ret (p xor x)
3. Framework Components
In a typical cryptographic proof, we specify cryptographic schemes,
security definitions, and (assumed) hard problems, and then we
prove a reduction from a properly-instantiated security definition
to one or more problems that are assumed to be hard. In other
words, we assume the existence of an effective adversary against
the scheme in question, and then prove that we can construct a
procedure that can effectively solve a problem that is assumed to
be hard. This reduction results in a contradiction that allows us
to conclude that an effective adversary against the scheme cannot
exist.
The cryptographic schemes, security definitions, and hard problems are probabilistic, and FCF provides a common probabilistic
programming language (Section 3.1) for describing all three. Then
we provide a denotational semantics (Section 3.1) that allows reasoning about the probability distributions that correspond to programs in this language. This semantics assigns a numeric value to
an event in a probability distribution, and it also allows us to conclude that two distributions are equivalent and we can replace one
with the other (which supports the game-hopping style of [6]).
It can be cumbersome to work directly in the semantics, so we
provide an equational theory (Section 3.2) of distributions that can
be used to prove that distributions are related by equality, inequality or “closeness.” A program logic (Section 3.3) is also provided
to ease the development of proofs involving state or looping behavior. To reduce the effort required to develop a proof, the framework
provides a library of tactics (Section 3.4) and a library of common program elements with associated theory (Section 3.5). The
equational theory, program logic, tactics, and programming library
greatly simplify proof development, yet they are all derived from
Listing 1. An Example of a Probabilistic Program
The syntax of the language is defined by an inductive type
called Comp and is shown in Listing 2. At a high-level, Comp is an
embedded domain-specific language that inherits the host language
Gallina, and extends it with operations for generating and working
with random bits.
Inductive Comp : Set -> Type :=
| Ret : forall {A : Set}{H: EqDec A},
A -> Comp A
| Bind : forall {A B : Set},
Comp B -> (B -> Comp A) -> Comp A
| Rnd : forall n, Comp (Bvector n)
| Repeat : forall {A : Set},
Comp A -> (A -> bool) -> Comp A.
Listing 2. Probabilistic Computation Syntax
The most notable primitive operation is Rnd, which produces n
uniformly random bits. The Repeat operation repeats a computation until some decidable predicate holds on the value returned.
This operation allows a restricted form of non-termination that is
sometimes useful (e.g., for sampling natural numbers in a specified
range). The operations Bind and Ret are the standard monadic
constructors, and allow the construction of sequences of computations, and computations from arbitrary Gallina terms and functions, respectively. However, note that the Ret constructor requires
1 The
current release of the FCF code for version 8.4 of Coq is included as
auxiliary material.
a proof of decidable equality for the underlying return type, which
is necessary to provide a computational semantics as seen later in
this section. In the remainder of this paper, we will use a more natural notation for these constructors: {0, 1}n is equivalent to (Rnd
$
n), x ← c; f is the same as (Bind c (fun x ⇒ f), and
e). The framework includes an ASCII form of
ret e is (Ret
this notation as seen in Listing 1. In the case of Ret, the notation
serves to hide the proof of decidable equality, which is irrelevant to
the programmer and is usually constructed automatically by proof
search.
FCF uses a shallow embedding, in which functions in the object language are realized using functions in the metalanguage. In
contrast, CertiCrypt uses a deep embedding, in which the data type
describing the object language includes constructs for specifying
and calling functions, as well as all of the primitives such as bitvectors and xor.
We have found that there are key benefits to shallow embedding. The primary benefit is that we immediately gain all of the
capability of the metalanguage, including (in the case of Coq) dependent types, higher-order functions, modules, etc. Another benefit is that it is very simple to include any necessary theory in a
security proof, and all of the theory that has been developed in the
proof assistant can be directly utilized. One benefit that is specific
to Coq (and other proof assistants with this property) is that Gallina functions are necessarily terminating, and Coq provides some
fairly complex mechanisms for proving that a function terminates.
By combining this restriction on functions with additional restrictions on Repeat, we can ensure that a computation (eventually)
terminates, and that this computation corresponds with a distribution in which the total probability mass is 1.
On the other hand, the shallow embedding approach does have
some drawbacks. The main drawback is that a Gallina function
is opaque; we can only reason about a Gallina function based
on its input/output behavior. The most significant effect of this
limitation is that we cannot directly reason about the computational
complexity of a Gallina function. We address this issue in Section
3.7.
that any n-bit value is equal to a randomly chosen n-bit value.
The probability that Repeat c P produces x is the conditional
probability of x given P in c—which is equivalent to the function
shown in Figure 1.
It is important to note that this language is purely functional,
but the monadic style gives programs an imperative appearance.
This appearance supports the Familiarity design goal since cryptographic definitions and games are typically written in an imperative
style.
It is sometimes necessary to include some state in a cryptographic definition or proof. This can be easily accomplished by
layering a state monad on top of Comp. However, this simple approach does not allow the development of definitions in which an
adversary has access to an oracle that must maintain some hidden
state across multiple interactions with the adversary. The definition
could not simply pass the state to the adversary, because then the
adversary could inspect or modify it. So FCF provides an extension
to Comp for probabilistic procedures with access to a stateful oracle. The syntax of this extended language (Listing 3) is defined in
another inductive type called OracleComp, where OracleComp
A B C is a procedure that returns a value of type C, and has access
to an oracle that takes a value of type A and returns a value of type
B.
Inductive OracleComp :
Set -> Set -> Set -> Type :=
| OC_Query : forall (A B : Set),
A -> OracleComp A B B
| OC_Run : forall (A B C A’ B’ S : Set),
EqDec S -> EqDec B -> EqDec A ->
OracleComp A B C -> S ->
(S -> A -> OracleComp A’ B’ (B * S)) ->
OracleComp A’ B’ (C * S)
| OC_Ret : forall A B C,
Comp C -> OracleComp A B C
| OC_Bind : forall A B C C’,
OracleComp A B C ->
(C -> OracleComp A B C’) ->
OracleComp A B C’.
Listing 3. Computation with Oracle Access Syntax
Jret aK = 1{a}
$
Jx ← c; f xK = λx.
X
(Jf bK x) ∗ (JcK b)
b∈supp(JcK)
J{0, 1}n K = λx. 2−n
JRepeat c P K = λx.(1P x) ∗ (JcK x) ∗
X
b∈P
(JcK b)
!−1
Figure 1. Semantics of Probabilistic Computations
The denotational semantics of a probabilistic computation is
shown in Figure 1. The denotation of a term of type Comp A is a
function in A → Q which should be interpreted as the probability
mass function of a distribution on A. In Figure 1, 1S is the indicator
function for set S. So the denotation of ret a is a function that
returns 1 when the argument is definitionally equal to a, and 0
$
otherwise. We can view the denotation of x ← c; f c as a marginal
probability of the joint distribution formed by c and f . We know
the probability of all events in c, but we only know the probability
of events in f conditioned on events in c, so we can compute
the probability of any event in this marginal distribution using
the law of total probability. The fact that random bits are uniform
and independent is encoded in the denotation of {0, 1}n , which
is a function that ignores the argument and returns the probability
The OC Query constructor is used to query the oracle, and
OC Run is used to run some program under a different oracle
that is allowed to access the current oracle. The OC Bind and
OC Ret constructors are used for sequencing and for promoting
terms into the language, as usual. In the rest of this paper, we
overload the sequencing and ret notation in order to use them for
OracleComp as well as Comp. We use query and run, omitting
the additional types and decidable equality proofs, as notation for
the corresponding constructors of OracleComp.
Jquery aK = λo s.(o s a)
Jrun c′ o′ s′ K = λo s.Jc′ (λx y.J(o′ (f st x) y) o (snd x)K) (s′ , s)K
$
Jret cK = λo s.x ← c; ret (x, s)
$
$
Jx ← c; f xK = λo s.[x, s′ ] ← Jc o sK; J(f x) o s′ K
Figure 2. Semantics of Computations with Oracle Access
The denotation of an OracleComp is a function from an oracle
and an oracle state to a Comp that returns a pair containing the value
provided by the OracleComp and the final state of the oracle.
The type of an oracle that takes an A and returns a B is (S -> A
-> Comp(B * S)) for some type S which holds the state of the
oracle. The denotational semantics is shown in Figure 2.
3.2 (In)Equational Theory of Distributions
A common goal in a security proof is to compare two distributions with respect to some particular value (or pair of values) in
the distributions. To assist with such goals, we have provided an
(in)equational theory for distributions. This theory contains facts
that can be used to show that two probability values are equal,
that one is less than another, or that the distance between them
is bounded by some value. For simplicity of notation, equality is
overloaded in the statements below in order to apply to both numeric values and distributions. When we say that two distributions (represented by probability mass functions) are equal, as in
D1 = D2 , we mean that the functions are extensionally equal, that
is ∀x, (D1 x) = (D2 x).
Theorem 1 (Monad Laws).
$
Ja ← ret b; f aK = J(f b)K
$
Ja ← c; ret aK = JcK
$
$
$
$
Ja ← (b ← c1 ; c2 b); c3 aK = Jb ← c1 ; a ← c2 b; c3 aK
Theorem 2 (Commutativity).
$
$
$
$
Ja ← c1 ; b ← c2 ; c3 a bK = Jb ← c2 ; a ← c1 ; c3 a bK
Theorem 3 (Distribution Irrelevance). For any well-formed computation c,
$
(∀x ∈ supp(JcK), Jf xKy = v) ⇒ Ja ← c; f aKy = v
Theorem 4 (Distribution Isomorphism). For any f which is a bijection from supp(Jc2 K) to supp(Jc1 K),
∀x ∈ supp(Jc2 K), Jc1 K(f x) = Jc2 Kx
∧ ∀x ∈ supp(Jc2 K), Jf1 (f x)K v1 = Jf2 xKv2
$
$
⇒ Ja ← c1 ; f1 aK v1 = Ja ← c2 ; f2 aK v2
Theorem 5 (Identical Until Bad).
$
$
Ja ← c1 ; ret (B a)K = Ja ← c2 ; ret (B a)K ∧
$
Ja ← c1 ; ret (P a, B a)K(x, false) =
$
Ja ← c2 ; ret (P a, B a)K(x, false) ⇒
$
$
| Ja ← c1 ; ret (P a)K x − Ja ← c2 ; ret (P a)K x | ≤
$
Ja ← c1 ; ret (B a)K true
The meaning and utility of many of the above theorems is direct
(such as the standard monad properties in Theorem 1), but others
require some explanation. Theorem 3 considers a situation in which
the probability of some event y in Jf xK is the same for all x
produced by computation c. Then the distribution JcK is irrelevant,
and it can be ignored. This theorem only applies to well-formed
computations: A well-formed computation is one that terminates
with probability 1, and therefore corresponds to a valid probability
distribution.
Theorem 4 is a powerful theorem that corresponds to the common informal argument that two random variables “have the same
distribution.” More formally, assume distributions Jc1 K and Jc2 K assign equal probability to any pair of events (f x) and x for some
bijection f . Then a pair of sequences beginning with c1 and c2 are
denotationally equivalent as long as the second computations in the
sequences are equivalent when conditioned on (f x) and x. A special case of this theorem is when f is the identity function, which
allows us to simply “skip” over two semantically equivalent computations at the beginning of a sequence.
Theorem 5, also known as the “Fundamental Lemma” from
[6], is typically used to bound the distance between two games by
the probability of some unlikely event. Computations c1 and c2
produce both a value of interest and an indication of whether some
“bad” event happened. We use (decidable) predicate B to extract
whether the bad event occurred, and projection P to extract the
value of interest. If the probability of the “bad” event occurring in
c1 and c2 is the same, and if the distribution of the value of interest
is the same in c1 and c2 when the bad event does not happen, then
the distance between the probability of the value of interest in c1
and and c2 is at most the probability of the “bad” event occurring.
3.3 Program Logic
The final goal of a cryptographic proof is always some relation
on probability distributions, and in some cases it is possible to
complete the proof entirely within the equational theory described
in 3.2. However, when the proof requires reasoning about loops or
state, a more expressive theory may be needed in order to discharge
some intermediate goals. For this reason, FCF includes a program
logic that can be used to reason about changes to program state
as the program executes. Importantly, the program logic is related
to the theory of probability distributions through completeness and
soundness theorems which allow the developer to derive facts about
distributions from program logic facts, and vice-versa.
The core logic is a Probabilistic Relational Postcondition Logic
(PRPL), that behaves like a Hoare logic, except there are no preconditions. The definition of a PRPL specification is given in Definition 1. In less formal terms, we say that computations p and q
are related by the predicate Φ if both p and q are marginals of the
same joint probability distribution, and Φ holds on all values in the
support of that joint distribution.
Definition 1 (PRPL Specification). Given p : Comp A and q :
Comp B, p ∼ q{Φ} iff,
∃ (d : Comp (A * B)), ∀(x, y) ∈ supp(JdK), Φ x y ∧
$
$
JpK = Jx ← d; ret (f st x)K ∧ JqK = Jx ← d; ret (snd x)K
Using the PRPL, we can construct a Probabilistic Relational
Hoare Logic (PRHL) which includes a notion of precondition for
functions that return computations as shown in Definition 2. The
resulting program logic is very similar to the Probabilistic Relational Hoare Logic of EasyCrypt [5], and it has many of the same
properties.
Definition 2 (PRHL Specification). Given p : A -> Comp B and
q : C -> Comp D, {Ψ}p ∼ q{Φ} iff, ∀a b, Ψ a b ⇒ (p a) ∼
(q b){Φ}.
Several theorems are provided along with the program logic definitions to simplify reasoning about programs. In order to use the
program logic, one only needs to apply the appropriate theorem, so
it is not necessary to produce the joint distribution described in the
definition of a PRPL specification unless a suitable theorem is not
provided. Theorems are provided for reasoning about the basic programming language constructs, interactions between programs and
oracles, specifications describing equivalence, and the relationship
between the program logic and the theory of probability distributions. Some of the more interesting program logic theorems are
described below.
Theorem 6 (Soundness/Completeness w.r.t. Equality).
p ∼ q{λ a b.a = x ⇔ b = y} ⇔ JpK x = JqK y
Theorem 7 (Soundness/Completeness w.r.t. Inequality).
p ∼ q{λ a b.a = x ⇒ b = y} ⇔ JpK x ≤ JqK y
Theorem 8 (Sequence Rule).
p ∼ q{Φ′ } ⇒ {Φ′ }r ∼ s{Φ} ⇒
$
$
(x ← p; r x) ∼ (x ← q; s x){Φ}
Theorem 9 (Oracle Equivalence). Given an OracleComp c, and a
pair of oracles, o and p with initial states s and t,
Φ = λ x y.(f st x) = (f st y) ∧ P (snd x)(snd y) ⇒
∀a s′ t′ , P s′ t′ ⇒ (o s′ a) ∼ (p t′ a){Φ} ⇒
P s t ⇒ (JcK o s) ∼ (JcK p t){Φ}
Theorems 6 and 7 relate judgments in the program logic to relations on probability distributions. The forward direction (soundness) is typically used in a proof to transform the goal into the program logic in order to accurately reason about loops and/or state.
Once the goal is in the program logic, the backward direction (completeness) can be used to return to a goal about distributions, or
to apply an existing theorem that describes relations on probability distributions. Theorem 8 is the relational form of the standard
Hoare logic sequence rule, and it supports the decomposition of
program logic judgments.
Theorem 9 allows the developer to replace some oracle with an
observationally equivalent oracle. This theorem takes a relational
invariant P on the states of the oracles, and requires the developer
to prove that if P holds on the states of the oracles and they are
given identical input, then P holds on the resulting states and the
outputs are identical. As long as P holds on the initial state of
the oracle, this theorem concludes that the values returned by the
program interacting with the oracles are equal, and that P holds
on the final state. There is also a more general form of this theorem
(omitted for brevity) in which the state of the oracle is allowed to go
bad, and the interaction only produces equivalent results if the state
does not go bad. This more general theorem can be combined with
Theorem 5 to get “identical until bad” results for program/oracle
interactions.
3.5 Programming Library
The framework includes a library containing useful programming
structures and their related theory. For example, the library includes
several sampling routines, such as drawing a natural number from a
specified range; drawing an element from a finite list, set, or group;
or sampling an arbitrary Bernoulli distribution. These sampling
routines are all computations based on the Rnd statement provided
by the language, and each routine is accompanied by a theory
establishing that the resulting distribution is correct and has the
desired properties.
The CompFold package contains higher-order functions for
folding and mapping a computation over a list. This package uses
the program logic extensively, and many of the theorems take
a specification on a pair of computations as an argument, and
produce a specification on the result of folding/mapping those
computations over a list. The package also contains theorems about
typical list and loop manipulations such as appending, flattening,
fusion/fission and order permutation.
3.6 Asymptotic Theory
The bulk of the effort in a security proof will be spent obtaining
some result in the concrete setting. From there, a little more effort is required to produce a proof of some asymptotic fact that
one would typically encounter in cryptography literature. To enable such asymptotic definitions and proofs, FCF includes a library
of standard asymptotic definitions such as Definitions 3 and 4. The
library also includes theorems that can be used to prove that functions are polynomial or negligible based on their composition(e.g.,
the sum of polynomials is polynomial, the quotient of polynomial
and exponential is negligible).
Definition 3 (At Most Polynomial). A function f : N → N is at
most polynomial iff ∃x c1 c2 , ∀n, f (n) ≤ c1 ∗ nx + c2
Definition 4 (Negligible Function). A function f : N → Q is
negligible iff ∀c, ∃n, ∀x > n, f (x) < 1/xc
3.7 Efficient Procedures
3.4 Tactics
The framework includes several tactics that can be used to transform goals using the facts in Sections 3.2 and 3.3. For example,
the comp swap l tactic applies commutativity (Theorem 2) to
swap two independent statements at the beginning of the program
on the left (of the equality, inequality, or program logic specification) in the current goal. There are similar tactics for manipulating games based on Left Identity (comp ret), Associativity
(comp inline), Distribution Irrelevance (comp irr), and the
special case of Distribution Isomorphism in which the bijection is
the identity function (comp skip). Many of these tactics can be
applied to goals related to probability distributions as well as goals
in the program logic.
A tactic called dist compute is provided to automatically
discharge goals involving simple computations for which the corresponding distribution obviously has some desired property—
typically that the probability of some event equals some specific
value. A common proof technique is to develop a program in
which the probability value of a particular event is obvious, and
then relate other programs to this one by equivalence proofs. Then
dist compute can be used to automatically compute the desired
probability value for this program. The tactic works by producing
an arithmetic expression from the computation(s) and then performing case splits in appropriate ways in order to get goals that
can be solved automatically by existing Coq decision procedures
(such as intuition and omega).
A typical asymptotic security property states that a family of cryptographic schemes has some desirable property for all efficient adversaries. So in order to prove and apply these properties, we require some notion of “efficient” (families of) procedures. The language of computations used in FCF does not imply any particular
model of computation—it is just a mechanism to specify probability distributions in a computational manner. Any notion of “efficiency” must first fix a model of computation, and then a complexity class on that model. We want this notion of efficiency to be
flexible and extensible, so we can support several different models
of computation and complexity classes.
To accomplish this flexibility, we parameterize asymptotic security definitions by an “admissibility predicate” indicating the class
of adversaries against which a problem is assumed to be hard, or
a scheme is proven to be secure. In this setting, the adversary is a
family of procedures indexed by a natural number which indicates
the value of the security parameter. The admissibility predicate can
describe the efficiency of the adversary as well as other properties
such as well-formedness or the number of allowed oracle queries
as a function of the security parameter.
FCF includes a simple cost model and an associated admissibility predicate describing non-uniform worst-case polynomial time
Turing machines that perform a (worst case) polynomial number
of oracle queries. This admissibility predicate is constructed using a concrete cost model that assigns numeric costs to particular Coq functions, Comp values, and OracleComp values. In this
cost model, the cost of executing a function is in N, indicating the
worst-case (over all arguments) execution time. The cost of running
a Comp is in N, indicating the worst-case execution time over all
outcomes. The cost of executing an OracleComp is in N → N,
and is a function from the cost of executing the oracle to the cost
of executing the computation, including the cost of executing all
oracle queries.
The cost model for Gallina functions is axiomatic, as there is
no direct way to capture such an intensional property for these
terms. Our cost model includes axioms for primitive operations
as well as a set of combinators for building more complicated
functions. For example, the model includes an axiom stating that
the xor operation for bit vectors of length c has a cost of c. As
other examples, the model includes axioms stating that the cost of
f composed with g is the sum of the costs of f and g, and the cost
of if e1 then e2 else e3 is the cost of e1 plus the maximum of
the costs of e2 and e3 . Obviously, our cost axioms are incomplete,
but in practice, the number required is relatively small since it is
only necessary to reason about the functions used by a constructed
adversary in a proof. Of course, the axioms need to be carefully
inspected to ensure they accurately describe the desired complexity
class.2 But of course, a similar kind of inspection is needed to
ensure the faithfullness of a cost model for a deeply-embedded
language.
to unrolling Repeat statements are trivial to prove under the
operational semantics.
Another benefit of the operational semantics and proof of equivalence is that this semantics can be considered to be the basic semantics for computations, and the denotational semantics no longer
needs to be trusted. Some may prefer this arrangement, since the
operational semantics more closely resembles a typical model of
computation, and may be easier to understand and inspect. The operational semantics can also be used as a basis for a model of computation used to determine whether programs are efficient.
Now that we have an operational semantics, we can simply use
the standard Coq extraction mechanism to extract it along with the
model of interest and all supporting types and functions. Of course,
the trustworthiness of the extracted code depends on the correctness of Coq’s extraction mechanism. Gallina does not allow infinite
recursion, so the framework includes OCaml code that runs a computation under the operational semantics until a value is obtained.
The final step is instantiating any abstract types and functions with
appropriate OCaml code. This extraction mechanism does not produce production-quality code, but the code could be used for purposes related to prototyping and testing.
4. Security Proof Construction
3.8 Code Extraction
FCF provides a code extraction mechanism that includes a strong
guarantee of equivalence between a model of a probabilistic program and the code extracted from that model. The denotational semantics of probabilistic computations relates a computation to a
probability distribution, but it does not contain sufficient information to allow us to reason about the behavior of such computations
on a traditional computer. So we developed a small-step operational
semantics that describes the behavior of these computations on a
machine in which the memory contains values rather than probability distributions. The operational semantics (omitted for brevity)
is an oracle machine that is given a finite list of bits representing the
“random” input, and it describes how a computation takes a single
step to produce a new computation, a final value, or fails due to
insufficient input bits.
To show that this semantics is correct, we consider [c]n , the multiset of results obtained by running a program c under this semantics on the set of all input lists of length n. We can view [c]n as a
distribution, where the mass of some value a in the distribution is
the proportion of input strings that cause the program to terminate
with value a. In order to compare the operational semantics with the
denotational semantics, we want to view the operational semantics
as a relation between computations and distributions. So the distribution related to computation c by the operational semantics is
limn→∞ [c]n . The statement of equivalence between the semantics
is shown in Theorem 10.
Theorem 10. If c is well-formed, then
lim [c]n = JcK
n→∞
FCF contains a proof of Theorem 10 as a validation of the
operational semantics used for extraction, but this theorem also
provides other benefits. Because limits are unique, if two programs
are equivalent under the operational semantics, then they are also
equivalent under the denotational semantics. This allows us to
prove equivalence of two programs using the operational semantics
when it is more convenient to do so. For example, theorems related
2 Furthermore, a proof that a Gallina term has a cost described by these
axioms does not mean that the extracted OCaml code will have this complexity, but rather, there exists some (propositionally) equivalent term which
has the described cost. Since we are only trying to show the existence of an
effective procedure, this is sufficient for our purposes.
This section uses an example to describe the process of constructing a proof of security using the general process described at the
beginning of Section 3. We consider a simple encryption scheme
constructed from a pseudorandom function (PRF), and we prove
that ciphertexts produced by this scheme are indistinguishable under chosen plaintext attack (IND-CPA). These security definitions,
and the formal description of the construction, are provided in later
sub-sections.
This example proof is relatively simple, yet it contains many
elements that one would find in a typical cryptographic argument,
and so it allows us to exercise all of the key functionality of the
framework. A more complex mechanized proof (e.g., the proof
of [10]) may have more intermediate games and a different set
of arguments to justify game transformations, but the structure is
similar to the proof that follows.
4.1 Concrete Security Definitions
In FCF, security definitions are used to describe properties that
some construction is proven to have, as well as problems that
are assumed to be hard. In the PRF encryption proof, we use
the definition of a PRF to assume that such a PRF exists, and
we use that assumption to prove that the construction in question
has the IND-CPA property. A concrete security definition typically
contains some game and an expression that describes the advantage
of some adversary – i.e., the probability that the adversary will
“win” the game.
The game used to define the concrete security of a PRF is
shown in Listing 4. Less formally, we say that f is a PRF for some
adversary A, if A cannot effectively distinguish f from a random
function. So this means that we expect that PRF Advantage is
“small” as long as A is an admissible adversary.
The function f oracle simply puts the function f in the form
of an oracle, though a very simple one with no state and with
deterministic behavior. The procedure RndR func is an oracle
implementing a random function constructed using the provided
computation RndR. The expressions involving A use a coercion in
Coq to invoke the denotational semantics for OracleComp, and
therefore ensure that A can query the oracle but has no access to the
state of the oracle.
At a high level, this definition involves two games describing
two different “worlds” in which the adversary may find himself.
In one world (PRF G A) the adversary interacts with the PRF,
Variable
Variable
Variable
Variable
Variable
Key D R : Set.
RndKey : Comp Key.
RndR : Comp R.
A : OracleComp D R bool.
f : Key -> D -> R.
Definition f_oracle(k : Key)
(x : unit)(d : D) : Comp (R * unit) :=
ret (f k d, tt).
Definition PRF_G_A : Comp bool :=
k <-$ RndKey;
[b, _] <-$2 A (f_oracle k) tt;
ret b.
Definition PRF_G_B : Comp bool :=
[b, _] <-$2 A (RndR_func) nil;
ret b.
Definition PRF_Advantage :=
| Pr[PRF_G_A] - Pr[PRF_G_B] |.
Listing 4. PRF Concrete Security Definition
and in the other (PRF G B) the adversary interacts with a random
function. In each game, the adversary interacts with the oracle and
then outputs a bit. The advantage of the adversary is the difference
between the probability that he outputs 1 in world PRF G A and the
probability that he outputs 1 in world PRF G B. If f is a PRF, then
this advantage should be small.
The concrete security definition for IND-CPA encryption is
shown in Listing 5. In this definition, KeyGen and Encrypt are
the key generation and encryption procedures. The adversary comprises two procedures, A1 and A2 with different signatures, and the
adversary is allowed to share arbitrary state information between
these two procedures. This definition uses a slightly different style
than the PRF definition—there is one game and the “world” is chosen at random within that game. Then the adversary attempts to
determine which world was chosen.
In Listing 5, the game produces an encryption oracle from
the Encrypt function and a randomly-generated encryption key.
Then the remainder of the game, including the calls to A1 and A2,
may interact with that oracle. The code for this definition includes
some additional notation (different arrows and extra $ symbols)
that is only used to provide hints to the Coq parser and does not
change the behavior of the program.
4.2 Construction
The construction, like the security definitions, can be modeled in a
very natural way. Of course, one must take care to ensure that the
construction has the correct signature as specified in the desired
security property. The PRF encryption construction is shown in
Listing 6.
In the PRF Encryption construction, we assume a nat called
eta (η) which will serve as the security parameter. The encryption
scheme is based on a function f, and the scheme will only be secure
if f is a PRF. The type of keys and plaintexts is bit vectors of
length eta, and the type of ciphertexts is pairs of these bit vectors.
The decryption function is included for completeness, but it is not
needed for this security proof.
4.3 Sequence of Games
The sequence of games represents the overall strategy for completing the proof. In the case of PRF Encryption, we want to show
that the probability that the adversary will correctly guess the randomly chosen “world” is close to 1/2. We accomplish this by instantiating the IND-CPA security definition with the construction, and
then transforming this game, little by little, until we have a game in
Variable Plaintext Ciphertext Key State : Set.
Variable KeyGen : Comp Key.
Variable Encrypt : Key -> Ciphertext ->
Comp Plaintext.
Variable A1 : OracleComp Plaintext Ciphertext
(Plaintext * Plaintext * State).
Variable A2 : State -> Ciphertext ->
OracleComp Plaintext Ciphertext bool.
Definition EncryptOracle
(k : Key)(x : unit)(p : Plaintext) :=
c <-$ Encrypt k p;
ret (c, tt).
Definition IND_CPA_SecretKey_G :=
key <-$ KeyGen ;
[b, _] <-$2
(
[p0, p1, s_A] <--$3 A1;
b <--$$ {0, 1};
pb <- if b then p1 else p0;
c <--$$ Encrypt key pb;
b’ <--$ A2 s_A c;
$ ret eqb b b’
)
(EncryptOracle key) tt;
ret b.
Definition IND_CPA_SecretKey_Advantage :=
| Pr[IND_CPA_SecretKey_G] - 1 / 2 |.
Listing 5. IND-CPA Concrete Security Definition
Variable eta : nat.
Variable f : Bvector eta ->
Bvector eta -> Bvector eta.
Definition PRFE_KeyGen :=
{0, 1} ˆ eta.
Definition PRFE_Encrypt
(k : Key )(p : Plaintext) :=
r <-$ {0, 1} ˆ eta;
ret (r, p xor (f k r)).
Definition PRFE_Decrypt
(k : Key)(c : Ciphertext) :=
(snd c) xor (f k (fst c)).
Listing 6. Encryption using a PRF
which this probability is exactly 1/2. Each transformation may add
some concrete value to the bounds, and we want to ensure that the
sum of these values is small.
The diagram in Figure 3 shows the entire sequence of games, as
well as the relationship between each pair of games in the sequence.
In this diagram, two games are related by = if they are identical,
and by ≈ if they are close. When the equivalence is non-trivial, the
diagram gives an argument for the equivalence, which implies a
bound on the distance between the games when they are not equal.
A detailed description of each game transformation follows:
1. Instantiate the IND-CPA definition with the construction. Unfold definitions and simplify. (Listing 7)
2. Apply the PRF definition to replace the PRF with a random
function. (Listing 8)
3. Replace the random function output used to encrypt the challenge ciphertext with a bit vector selected completely at random. With overwhelming probability, the adversary does not
notice this change. (Listing 9)
IND GPA G
= G1
≈PRF Advantage
G2
≈Random List Collision
G3
Definition G3 :=
[a, o] <-$2 A1 (RF_Encrypt) nil;
[p0, p1, s_A] <-3 a;
b <-$ {0, 1};
pb <- if b then p1 else p0;
r <-$ {0, 1}ˆeta;
pad <-$ {0, 1}ˆeta;
c <- (r, pb xor pad);
[b’, o] <-$2 (A2 s_A c) RF_Encrypt o;
ret (eqb b b’).
Listing 9. Game 3
=One Time Pad
G4
= G5 =
1/2
Figure 3. Sequence of Games Diagram
4. We have modified the game to the point that encryption of the
challenge plaintext is by one-time pad. So we can replace the
ciphertext with a randomly-chosen value. (Listing 10)
5. Now the ciphertext is independent from the plaintext, and thus
independent from the random bit that was used to select the
“world.” This means we can move this coin flip to after the adversary guesses which world he is in. In this game, it is obvious
that the probability that the adversary guesses the correct outcome of the coin flip is exactly one half. (Listing 11)
Definition G1 :=
key <-$ PRFE_KeyGen;
[b, _] <-$2
(
[p0, p1, s_A] <--$3 A1;
b <--$$ {0, 1};
pb <- if b then p1 else p0;
c <--$$ PRFE_Encrypt key pb;
b’ <--$ (A2 s_A c);
$ ret (eqb b b’)
)
(PRFE_EncryptOracle key) tt;
ret b.
Listing 7. Game 1
Definition PRFE_RandomFunc := @randomFunc
(Bvector eta) (Bvector eta) ({0,1}ˆeta) _.
Definition RF_Encrypt s p :=
r <-$ {0, 1} ˆ eta;
[pad, s] <-$2 PRFE_RandomFunc s r;
ret (r, p xor pad, s).
Definition G2 :=
[a, o] <-$2 A1 (RF_Encrypt) nil;
[p0, p1, s_A] <-3 a;
b <-$ {0, 1};
pb <- if b then p1 else p0;
[c, o] <-$2 RF_Encrypt o pb;
[b’, o] <-$2 (A2 s_A c) RF_Encrypt o;
ret (eqb b b’).
Listing 8. Game 2
4.4 Equivalence Proofs
The next step is to prove the appropriate sort of equivalence between each pair of games in the sequence. In the case of PRF Encryption, the goal is to show that the distance between the IND-CPA
Definition G4 :=
[a, o] <-$2 A1 (RF_Encrypt) nil;
[p0, p1, s_A] <-3 a;
b <-$ {0, 1};
pb <- if b then p1 else p0;
r <-$ {0, 1}ˆeta;
pad <-$ {0, 1}ˆeta;
c <- (r, pad);
[b’, o] <-$2 (A2 s_A c) RF_Encrypt o;
ret (eqb b b’).
Listing 10. Game 4
Definition G5 :=
[a, o] <-$2 A1 (RF_Encrypt) nil;
[p0, p1, s_A] <-3 a;
r <-$ {0, 1}ˆeta;
pad <-$ {0, 1}ˆeta;
c <- (r, pad);
[b’, o] <-$2 (A2 s_A c) RF_Encrypt o;
b <-$ {0, 1};
ret (eqb b b’).
Listing 11. Game 5
game and 1/2 is very small, and we accomplish this by showing that
each pair of games in the sequence is either identical or “close.”
The first step is to show that Game 1 (Listing 7) really is the
IND-CPA game instantiated with this encryption scheme. This fact
(Listing 12) is obvious, and the proof can be completed using Coq’s
reflexivity tactic. A1 and A2 are parameters representing the
procedures of the adversary against the encryption scheme. In the
statement of this theorem, == is equality for rational numbers. This
equality is registered with Coq’s setoid system to enable tactics
such as reflexivity and rewriting.
Theorem G1_equiv :
Pr[IND_CPA_SecretKey_G PRFE_KeyGen PRFE_Encrypt
A1 A2] == Pr[G1].
Listing 12. Equivalence of the Security Definition and Game 1
Next we show that the distance between Games 1 and 2 is exactly the advantage of some adversary against a PRF. The adversary against the PRF (Listing 13) is constructed from A1 and A2.
PRFE Encrypt OC is an encryption oracle that interacts with the
PRF as an oracle. PRF A provides this encryption oracle to A1 and
A2 using the OC Run operation.
To prove the “closeness” of Games 1 and 2, first we prove that
the interaction between PRF A and the PRF oracle is equivalent to
Game 1 (Theorem 14). Then we prove that the interaction between
PRF A and the random function oracle is equivalent to Game 2
(Theorem 15). Finally we apply the results of parts 1 and 2 and
unify with the definition of a PRF (Theorem 16).
To prove Theorems 14 and 15, we mostly perform simple
manipulations such as applying the denotational semantics of
OracleComp, inlining, and removing identical statements at the
beginning of the game. In both of these proofs, the adversary is
Definition PRFE_Encrypt_OC (x : unit)
(p : Plaintext) : OracleComp (Bvector eta)
(Bvector eta) (Ciphertext * unit) :=
r <--$$ {0, 1} ˆ eta;
pad <--$ OC_Query r;
$ (ret (r, p xor pad, tt)).
Definition PRF_A : OracleComp (Bvector eta)
(Bvector eta) bool :=
[a, n] <--$2 OC_Run A1 PRFE_Encrypt_OC tt;
[p0, p1, s_A] <-3 a;
b <--$$ {0, 1};
pb <- if b then p1 else p0;
r <--$$ {0, 1} ˆ eta;
pad <--$ OC_Query r;
c <- (r, pb xor pad);
z <--$ OC_Run (A2 s_A c) PRFE_Encrypt_OC n;
[b’, _] <-2 z;
$ ret (eqb b b’).
Listing 13. The Constructed Adversary Against the PRF
Theorem G1_PRF_A_equiv :
Pr[k <-$ {0, 1}ˆ eta;
[b, _] <-$2 PRF_A (f_oracle k) tt;
ret b] == Pr[G1].
Listing 14. Equivalence PRF A and G1
Theorem G2_PRF_A_equiv :
Pr[[b, _] <-$2 PRF_A PRFE_RandomFunc nil;
ret b] == Pr[G2].
Listing 15. Equivalence PRF A and G2
Theorem G1_G2_close : | Pr[G1] - Pr[G2] | ==
PRF_Advantage ({0, 1}ˆeta) ({0, 1}ˆeta)
f PRF_A.
Listing 16. Closeness of Game 1 and Game 2
interacting with two different, but observationally equivalent, oracles. So we use the program logic and Theorem 9 to prove that
these interactions produce equivalent results.
Next we show that Games 2 and 3 are “close” by demonstrating
that these games are “identical until bad” in the sense of Theorem
5. The “bad” event of interest is the event that the randomlygenerated PRF input used to encrypt the challenge plaintext (r
in Game 2) is also used to encrypt some other value during the
interaction between the adversary and the encryption oracle. There
are two separate adversary procedures, and each one is capable
of encountering r during its interaction with the oracle. So we
divide this proof into two parts, one for each adversary procedure,
where each part includes an “identical until bad” argument. In the
first step, we produce the pad value randomly (without using the
random function), but then add an entry for r and pad to the
state of the random function. In the second step, we produce pad
randomly, and do not add an entry to the random function.
To get an expression for the probability of the “bad”event, we
assume natural numbers q1 and q2 , and that A1 performs at most
q1 queries and A2 performs at most q2 queries. FCF includes a
library module called RndInList that includes general-purpose
arguments related to the probability of encountering a randomly
selected value in a list of a certain length, and the probability
of encountering a certain value in a list of randomly-generated
elements of a certain length. By combining the “identical until
bad” proofs with these arguments to get expressions bounding the
probabilities of the bad events, we obtain the result of Listing 17.
The next step is to use a one-time-pad argument to replace the
challenge ciphertext with a randomly-chosen value. The library
contains a generic one-time-pad argument that we can apply here.
Theorem G2_G3_close :
| Pr[G2] - Pr[G3] | <=
q1 / (2 ˆ eta) + q2 / (2 ˆ eta).
Listing 17. Closenes of Game 2 and Game 3
We transform this game into an equivalent game that unifies with
the one-time-pad argument, then we apply the argument to get the
result shown in Listing 18.
Theorem G3_G4_equiv : Pr[G3] == Pr[G4].
Listing 18. Equivalence of Game 3 and Game 4
Now that the ciphertext is independent of the challenge bit, we
produce a new game by moving the sampling of the challenge bit
to the end of the game. To prove this fact (Listing 19), we simply
unfold the required definitions, skip over all of the identical pairs
of statements at the beginning of the proof, then swap the order of
independent statements in the game on the left in order to make
these statements align with the identical statements in the game on
the right.
Theorem G4_G5_equiv :
Pr[G4] == Pr[G5].
unfold G4, G5.
do 3 (comp_skip; comp_simp; comp_swap_l).
comp_skip; comp_simp.
reflexivity.
Qed.
Listing 19. Equivalence of Game 4 and Game 5
Finally, we develop the proof that the adversary wins Game
5 with probability exactly 1/2. This proof (Listing 20) proceeds
by discarding all of the initial statements in the game using the
comp irr l tactic. Note that this tactic produces an obligation to
prove that the statement being discarded is a well-formed computation, which can be discharged with the tactic wftac. Then what
remains is a very simple game, and dist compute can automatically compute the probability that this game returns true.
Theorem G5_one_half :
Pr[G5] == 1/2.
do 4 comp_irr_l; wftac.
dist_compute.
Qed.
Listing 20. Probability of Winning Game 5
By combining the equivalences of each pair of intermediate
games, we get the final concrete security result shown in Listing 21.
It is important to note that the statement of this theorem does not
reference any of the intermediate games. The sequence of games
was only a tool that we used to get the final result, and this sequence
does not need to be inspected in order to trust the result.
The concrete security result in Listing 21 may be sufficient for
many purposes. We have an expression describing the advantage of
the adversary, and we can inspect this expression to see whether this
advantage is sufficiently small. We also must inspect the definition
of the adversary PRF A, which appears in this result, and ensure
that this adversary is “efficient” according to the desired complexity class. Next we will show how to derive an asymptotic security
result based on this concrete result. A benefit of proving asymptotic security is that this proof removes the requirement to inspect
the constructed adversary and the expression describing the adversary’s advantage.
Theorem PRFE_IND_CPA_concrete :
IND_CPA_SecretKey_Advantage PRFE_KeyGen
PRFE_Encrypt A1 A2 <=
PRF_Advantage ({0, 1}ˆeta) ({0, 1}ˆeta)
f PRF_A + (q1 / 2ˆeta + q2 / 2ˆeta).
Listing 21. Concrete Security Result
4.5 Asymptotic Security Definitions
Now we give the asymptotic security definitions for PRFs and INDCPA encryption. These definitions are parameterized by an admissibility predicate as described in Section 3.7. The IND-CPA definition accepts two admissibility predicates – one for each adversary
procedure.
The asymptotic security definition for a PRF is given in Listing 22. In this definition, RndKey, RndR, and f are nat-indexed
families of procedures. Similarly in the IND-CPA definition (Listing 23), KeyGen and Encrypt are nat-indexed families of procedures. Both of these definitions are claims over all admissible
nat-indexed adversary families. Note that both definitions reuse
the expressions provided in the concrete security definitions. This
style provides a convenient method for developing an asymptotic
security proof from a concrete security proof.
Variable
Variable
Variable
Variable
D R Key : nat -> Set.
RndKey : forall n, Comp (Key n).
RndR : forall n, Comp (R n).
f : forall n, Key n -> D n-> R n.
We begin by assuming costs for A1 and A2. A1 cost is a function describing the cost of A 1. A2 cost 1 is a number describing how much it costs for A2 to compute an OracleComp that is
closed over a state and a ciphertext. Then A2 cost 2 is a function
describing the cost of executing this OracleComp. Given these
assumptions, we can give a cost to PRF A as shown in Listing 24. In
the statement of this theorem, oc cost, comp cost, and cost
are the cost models for OracleComp, Comp, and Coq functions,
respectively. Note that this cost model is overly conservative and
some costs are counted multiple times.
Theorem PRF_A_cost :
oc_cost cost (comp_cost cost) PRF_A
(fun x => (A1_cost (x + (5 * eta))) +
(A2_cost_2 (x + (5 * eta))) +
x + 5 * A2_cost_1 + 6 + 7 * eta).
Listing 24. Cost of Constructed Procedure PRF A
This proof is completed by repeatedly applying the rule of the
cost model that is relevant to the term in the goal, which is a highly
syntax-directed operation that can be mostly automated. Once all
these syntax-directed rules are applied, the developer is obligated
to prove that the expression obtained in this process is equal to (or
less than) the expression in the statement of the theorem. In this last
step of the proof, automated tactics such as omega are very useful.
4.7 Asymptotic Security Proof
Definition PRF :=
forall (A : \forall n, OracleComp (D n) (R n)
bool), admissible_A A ->
negligible (fun n => PRF_Advantage
(RndKey n) (RndR n) (@f n) (A n)).
Listing 22. Definition of a PRF
Variable Plaintext Ciphertext Key State :
nat -> Set.
Variable KeyGen : forall n, Comp (Key n).
Variable Encrypt : forall n, Key n ->
Ciphertext n -> Comp (Plaintext n).
Definition IND_CPA_SecretKey :=
forall (State : nat -> Set)
(A1 : forall n, OracleComp (Plaintext n)
(Ciphertext n)
(Plaintext n * Plaintext n * State n))
(A2 : forall n, State n -> Ciphertext n ->
OracleComp (Plaintext n) (Ciphertext n)
bool),
admissible_A1 A1 ->
admissible_A2 A2 ->
negligible
(fun n => IND_CPA_SecretKey_Advantage
(KeyGen n) (@Encrypt n) (A1 n) (A2 n) ).
Listing 23. Definition of IND-CPA Encryption
4.6 Efficiency of Constructed Adversaries
The first step in proving an asymptotic security result is to view
each constructed adversary in the concrete proof as a nat-indexed
family of adversaries, and prove that this family is “efficient” as
defined by some complexity class. In the PRF Encryption proof, we
use the non-uniform polynomial time complexity class described
in Section 3.7. Because this class includes a concrete cost model,
we begin with a proof of the concrete cost of each constructed
adversary procedure.
The final step in the proof is to show that the security definition shown in Listing 23 holds on this construction as long
as f is a PRF as defined in Listing 22. The statement of this
fact is shown in Listing 25. Note that admissible oc and
admissible oc func 2 are the admissibility predicates for
OracleComp and for functions with two arguments that produce
an OracleComp defined in the simple complexity class described
in Section 3.7.
Theorem PRFE_IND_CPA :
PRF Rnd Rnd f (admissible_oc cost) ->
IND_CPA_SecretKey
PRFE_KeyGen (fun n => PRFE_Encrypt (@f n))
(admissible_oc cost)
(admissible_oc_func_2 cost).
Listing 25. Asymptotic Security of PRF Encryption
The primary obligation of this proof is to show that the function
defining the advantage of any admissible family of adversaries
against this encryption scheme is a negligible function. The fact
that this adversary family is admissible allows us to use the result of
Listing 24, along with other facts, to conclude that the constructed
adversary family against the PRF is admissible. In the course of
this proof, we must show that the expression implied by Figure 24
is at most polynomial in η if x is at most polynomial in η and all
the costs related to PRF A1 and PRF A2 are at most polynomial
in η. This fact is proven using the provided theory of polynomial
functions (Section 3.6).
From the admissibility of the constructed adversary, and from
the fact the f is a PRF against all admissible adversaries, we can
conclude that the constructed adversary’s advantage against the
PRF is negligible. The advantage of this adversary against the PRF
is one of the terms that appears in the bounds of the concrete result
(Listing 21). The other term is q1 /2η + q2 /2η , where q1 and q2
are the number of oracle queries performed by the two adversary
procedures. The admissibility predicates ensure that each adversary
only performs a polynomial number of queries, so q1 and q2 must
be polynomial in η, and this expression is negligible in η. So the
advantage of the adversary against this encryption scheme is the
sum of two negligible functions, and is therefore negligible.
The entire proof of security for this encryption scheme requries
approximately 1500 lines of Coq code, of which about 700 lines
are specification (including 100 lines of cryptographic definitions
and intermediate games) and 800 lines are proof. The proof incorporates another 500 lines of code for the reusable arguments (e.g.,
the one-time pad argument). We expect that a skilled Coq developer could complete such a proof in a matter of days (though he
may require the help of a cryptographer to develop the sequence of
games and high-level arguments).
5. Evaluation
This section attempts to evaluate FCF against the design goals
listed in Section 2, and to contrast with both CertiCrypt and EasyCrypt.
All three of these frameworks provide concrete bounds, so this
criterion is not discussed further. And, all three frameworks use a
relatively familiar syntax for security definitions and constructions.
We believe that, based on our experience working with cryptographers, they can easily understand these definitions (e.g., Listing
4) after spending a few minutes familiarizing themselves with the
notation.
Regarding proof automation, FCF lies somewhere between CertiCrypt and EasyCrypt. EasyCrypt achieves a significant level of
automation by using SMT solvers to discharge simple logical goals,
but higher-level goals still need to be addressed manually by applying tactics. FCF achieves a similarly high level of automation
through the use of existing and custom Coq tactics. These tactics are not as powerful as modern SMT solvers, so the developer may need to manually address some goals in FCF that would
be discharged automatically in EasyCrypt. However, the semantics of programs in FCF is computational, so Coq is able to immediately compute an expression describing the probability distribution for any program. This allows some simple equivalences
to be discharged immediately using this computation and FCF’s
dist compute tactic.
Regarding trust in extensional properties, FCF and CertiCrypt
are foundational, meaning that the program logic is constructed
definitionally from the semantics. In contrast, EasyCrypt relies
upon a set of axioms for its program logic. EasyCrypt also relies
on the correctness of the EasyCrypt front end and the Why3 verification generator, whereas FCF and CertiCrypt only depend on
the Coq type checker. EasyCrypt provides no support for reasoning about intensional properties like execution time, whereas CertiCrypt and FCF do, though FCF provides this suport using a trusted
set of axioms.
EasyCrypt and CertiCrypt are based on simply-typed, first-order
languages. This design makes it difficult to directly support abstraction, extension, and reuse, though these frameworks include elements which support these goals to some extent. In contrast, FCF
uses a shallow embedding and the advanced features of Coq, such
as dependent types, modules, notation, and higher-order functions,
to support abstraction, extensiblity, and reuse. We believe that having such a rich language for describing games and assumptions is
critical for scaling to larger protocols.
FCF supports code generation with a semantics that is proven to
be equivalent to the semantics used to reason about the probabilistic behavior of programs. That is, a program extracted from an FCF
model is guaranteed to produce the correct probability distribution
when the input bits provided to it are uniformly distributed, assuming the extraction mechanism of Coq preserves meaning. There has
been some initial work in producing implementations that correspond to EasyCrypt models, but there is no formal relationship be-
tween the semantics of the implementation and the semantics used
to reason about the model.
6. Related Work
There has been a large amount of work in the area of verifying
cryptographic schemes in recent years. In this section we will describe some of this related work, focusing on systems that attempt
to establish security in the computational model. CertiCrypt [4] and
EasyCrypt [5] have been thoroughly discussed previously in this
paper.
There are several other examples of frameworks for cryptographic security proofs implemented within proof assistants. The
most similar work is that of Nowak [20], who was the first to develop proofs of cryptography in Coq using a shallow embedding
in which programs have probability distributions as their denotations. FCF builds on this work by adding more tools for modeling
and reasoning such as procedures with oracle access (Section 3.1),
a program logic (Section 3.3), and asymptotic reasoning (Section
3.6).
The work of [2] is a Coq library utilizing a deeply-embedded
imperative programming language. This library is a predecessor to
CertiCrypt, and it includes some important elements that were later
adopted by CertiCrypt. Notably, the probabilistic programming language in this work is given a semantics in which program states are
distributions, and the semantics describes how these distributions
are transformed by each command in the language. CertiCrypt and
EasyCrypt extended this work by adding language constructs such
as oracles and unrestricted loops, and well as reasoning tools such
as the Probabilistic Relational Hoare Logic.
Verypto [8] is a fully-featured framework built on Isabelle [19]
that includes a deep embedding of a functional programming language. To allow state information to remain hidden from adversaries, Verypto provides ML-style references, in contrast to the oracle system provided by FCF. To date, Verypto has only been used
to prove the security of simple constructions, but this work uses an
interesting approach that deserves more exploration.
CryptoVerif [9] is a tool based on a concurrent, probabilistic
process calculus that is only able to prove properties related to
secrecy and authenticity. CryptoVerif is highly automated to the
extent that it will even attempt to locate intermediate games, and so
proof development in CryptoVerif requires far less effort compared
to FCF or EasyCrypt. However, there are a large number of proofs
that could be completed in FCF or EasyCrypt that are impossible
in CryptoVerif due to its specialized nature.
Refinement types [7] have been used by Fournet et al [15] to
develop proofs of security for cryptographic schemes in the computational model. In this system, a security property is specified as
an ideal functionality (in the sense of the real/ideal paradigm), and
proofs are completed using the “sequence of games” style in the
asymptotic setting. This approach allows the proofs of security to
be fairly simple, but no concrete security claims are proved, so it
may be difficult to make practical claims based on such a proof.
Computational soundness [1] provides another mechanism for
verifying cryptographic schemes. This approach attempts to derive
security in the computational model from security in the symbolic
model by showing that any likely execution trace in the computational model also exists in the symbolic model. It is possible to
mechanize such a proof as described in [3]. This approach is limited to classes of schemes for which computational soundness results have been discovered. Another limitation with this approach
is that it can only produce proofs in the asymptotic setting—there
is no way to prove concrete security claims.
Protocol Composition Logic (PCL) [12] provides a logic and
proof system for verifying cryptographic schemes in the symbolic
model. The system is based on a process calculus and allows rea-
soning about the results of individual protocol steps. More recent
work [11] has extended this logic to allow for proofs in the computational model. In computational PCL, formulas are interpreted
against probability distributions on traces and a formula is true if
it holds with overwhelming probability. This approach is similar to
computational soundness in that low-probability traces are ignored,
and proofs of concrete security claims are impossible.
[9] B. Blanchet. Computationally sound mechanized proofs of correspondence assertions. In 20th IEEE Computer Security Foundations Symposium (CSF’07), pages 97–111, Venice, Italy, July 2007. IEEE.
[10] D. Cash, S. Jarecki, C. Jutla, H. Krawczyk, M.-C. Rou, and M. Steiner.
Highly-scalable searchable symmetric encryption with support for
boolean queries. In R. Canetti and J. Garay, editors, Advances in Cryptology CRYPTO 2013, volume 8042 of Lecture Notes in Computer
Science, pages 353–373. Springer Berlin Heidelberg, 2013. ISBN 9783-642-40040-7. .
7. Conclusion and Future Work
[11] A. Datta, A. Derek, J. C. Mitchell, V. Shmatikov, and M. Turuani.
Probabilistic polynomial-time semantics for a protocol security logic.
In Proceedings of the 32nd international conference on Automata,
Languages and Programming, ICALP’05, pages 16–29, Berlin, Heidelberg, 2005. Springer-Verlag. ISBN 3-540-27580-0, 978-3-54027580-0. .
Our contribution is a complete mechanized framework for specifying and checking cryptographic proofs within a proof assistant. Our
framework compares favorably to the current state of the art, and
provides many new benefits, such as extensibility through a foundational approach, a powerful language for describing schemes, and
the ability to extract excutable code. Of course, whether these benefits apply at scale is still an open question, and thus a key direction
for us is to prove security for an even wider range of standard constructions, as well as novel cryptographic schemes. In particular,
we have proven the security of the “tuple set” construction of [10],
and we intend to continue developing this work into a proof of a
searchable symmetric encryption scheme.
The biggest limitation of FCF is that it currently lacks definitions for many cost models and complexity classes that are commonly used in cryptography. We hope to develop more cost models and complexity classes, including a complexity class describing
(uniform) probabilistic polynomial time Turing machines.
References
[1] M. Abadi and P. Rogaway.
Reconciling two views of cryptography (the computational soundness of formal encryption).
In Proceedings of the International Conference IFIP
on Theoretical Computer Science, Exploring New Frontiers
of Theoretical Informatics, TCS ’00, pages 3–22, London,
UK, UK, 2000. Springer-Verlag. ISBN 3-540-67823-9. URL
http://dl.acm.org/citation.cfm?id=647318.723498.
[2] R. Affeldt, M. Tanaka, and N. Marti. Formal proof of provable security by game-playing in a proof assistant.
In Proceedings of the 1st International Conference on Provable Security, ProvSec’07, pages 151–168, Berlin, Heidelberg, 2007.
Springer-Verlag. ISBN 3-540-75669-8, 978-3-540-75669-9. URL
http://dl.acm.org/citation.cfm?id=1779394.1779408.
[3] M. Backes and D. Unruh. Computational soundness of symbolic zeroknowledge proofs against active attackers. In 21st IEEE Computer
Security Foundations Symposium, CSF 2008, pages 255–269, June
2008. Preprint on IACR ePrint 2008/152.
[4] G. Barthe, B. Grégoire, and S. Zanella Béguelin. Formal certification of code-based cryptographic proofs.
In 36th ACM
SIGPLAN-SIGACT Symposium on Principles of Programming
Languages, POPL 2009, pages 90–101. ACM, 2009.
URL
http://dx.doi.org/10.1145/1480881.1480894.
[12] A. Datta, A. Derek, J. C. Mitchell, and A. Roy. Protocol composition
logic (pcl). Electronic Notes in Theoretical Computer Science, 172:
311–358, 2007.
[13] E. De Cristofaro, S. Jarecki, X. Liu, Y. Lu, and G. Tsudik.
Privacy-protecting information retrieval, University of Irvine
team: Protocol and proofs. Appendix E of SPAR Program BAA:
https://www.fbo.gov/utils/view?id=32750071e5cf4afc3b7e973d6
2010.
[14] T. Elgamal. A public key cryptosystem and a signature scheme based
on discrete logarithms. Information Theory, IEEE Transactions on, 31
(4):469–472, 1985. ISSN 0018-9448. .
[15] C. Fournet, M. Kohlweiss, and P.-Y. Strub. Modular code-based
cryptographic verification. In Y. Chen, G. Danezis, and V. Shmatikov,
editors, ACM Conference on Computer and Communications Security,
pages 341–350. ACM, 2011. ISBN 978-1-4503-0948-6.
[16] S. Halevi.
A plausible approach to computer-aided cryptographic proofs. Cryptology ePrint Archive, Report 2005/181, 2005.
http://eprint.iacr.org/.
[17] J. Herzog, C. Meadows, A. Jaggard, A. Stoughton, and J. Katz.
MITLL-NRL panel: Easycrypt 0.2 feedback and opinions.
http://web.archive.org/web/20140703170052/https://easycrypt
2013. Accessed: 201407-03.
[18] The Coq development team. The Coq proof assistant reference manual. LogiCal Project, 2004. URL http://coq.inria.fr. Version 8.0.
[19] T. Nipkow, L. C. Paulson, and M. Wenzel. Isabelle/HOL — A Proof
Assistant for Higher-Order Logic, volume 2283 of LNCS. Springer,
2002.
[20] D. Nowak.
A framework for game-based security
proofs.
Cryptology ePrint Archive, Report 2007/199, 2007.
http://eprint.iacr.org/.
[21] N. Ramsey and A. Pfeffer.
Stochastic lambda calculus and
monads of probability distributions.
In Proceedings of the
29th ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, POPL ’02, pages 154–165, New York,
NY, USA, 2002. ACM.
ISBN 1-58113-450-9.
.
URL
http://doi.acm.org/10.1145/503272.503288.
[5] G. Barthe, B. Grégoire, S. Heraud, and S. Zanella Béguelin.
Computer-aided security proofs for the working cryptographer. In Advances in Cryptology – CRYPTO 2011, volume 6841 of Lecture Notes
in Computer Science, pages 71–90. Springer, 2011.
[6] M. Bellare and P. Rogaway. Code-based game-playing proofs and
the security of triple encryption. Cryptology ePrint Archive, Report
2004/331, 2004. http://eprint.iacr.org/.
[7] J. Bengtson, K. Bhargavan, C. Fournet, S. Maffeis, and A. D. Gordon.
Refinement types for secure implementations. In In 21st IEEE Computer Security Foundations Symposium (CSF08, pages 17–32. IEEE,
2008.
[8] M. Berg.
Formal Verification of Cryptographic Security Proofs.
PhD thesis, Saarland University, 2013.
URL
http://www.infsec.cs.uni-saarland.de/˜berg/publications/thesis-berg.pdf.
| 6cs.PL
|
1
Pulling back error to the hidden-node parameter
technology: Single-hidden-layer feedforward network
without output weight
arXiv:1405.1445v1 [cs.NE] 6 May 2014
Yimin Yang, Q. M. Jonathan Wu, Guangbin Huang,
and Yaonan Wang
Abstract—According to conventional neural network theories, the feature of single-hidden-layer feedforward neural networks(SLFNs) resorts
to parameters of the weighted connections and hidden nodes. SLFNs are
universal approximators when at least the parameters of the networks
including hidden-node parameter and output weight are exist. Unlike
above neural network theories, this paper indicates that in order to
let SLFNs work as universal approximators, one may simply calculate
the hidden node parameter only and the output weight is not needed
at all. In other words, this proposed neural network architecture can
be considered as a standard SLFNs with fixing output weight equal to
an unit vector. Further more, this paper presents experiments which
show that the proposed learning method tends to extremely reduce
network output error to a very small number with only 1 hidden node.
Simulation results demonstrate that the proposed method can provide
several to thousands of times faster than other learning algorithm
including BP, SVM/SVR and other ELM methods.
including the hidden-node parameters (a,b) and output weight β
are allowed adjustable[4][5].
Unlike above neural network theories that all the parameters
in networks are allowed adjustable, other researches proposed
some semi-random network theories[6][7][8]. For example, Lowe
[8] focus on a specific RBF network: The centers a in [8] can be
randomly selected from the training data instead of tuning, but
the impact factor b of RBF hidden node is not randomly selected
and usually determined by users.
Unlike above semi-random network theories, in 2006, Huang
et al[9] illustrated that iterative techniques are not required in
adjusting all the parameters of SLFNs at all. Based on this idea,
Huang et al proposed simple and efficient learning steps referred
to as extreme learning machine(ELM). In [10][11][12], Huang et
al have proved that SLFNs with randomly generated hidden node
parameter can work as universal approximators by only calculating
the output weights linking the hidden layer to the output nodes.
Recently ELM development [13] shows that ELM unifies FNNs
and SVM/LS-SVM. Compared to ELM, LS-SVM and PSVM achieve
suboptimal solutions and have a higher computational cost.
Index Terms—Bidirectional Extreme Learning Machine, Feedforward
neural network, universal approximation, number of hidden nodes,
learning effectiveness
Above neural network theories indicate that SLFNs can work
as universal approximation at least hidden-node parameters1 and
output weight should be exist, however, in this paper we indicate
that output weight do not need exist in SLFNs at all.
I. I NTRODUCTION
In [14] we proposed a learning algorithm, called bidirectional
extreme learning machine (B-ELM) in which half of hiddennode parameters are not randomly selected and are calculated
by pulling back the network residual error to input weight. The
experimental results in [14] indicated that B-ELM tends to reduce
network output error to a very small value at an extremely early
learning stage. Further more, our recent experimental results indicate that in B-ELM[14], output weight play a very minion role in
the network learning effectiveness. Inspired by these experimental
results, in this paper, we show that SLFNs without output weight
can approximate any target continuous function and classify any
disjoint regions if one using pulling back error to hidden-node
parameters. In particular, the following contributions have been
made in this paper.
The widespread popularity of neural networks in many fields
is mainly due to their ability to approximate complex nonlinear mappings directly from the input samples. In the past two
decades, due to their universal approximation capability, feedforward neural networks (FNNs) have been extensively used in
classification and regression problem[1]. According to Jaeger’s estimation[2], 95% literatures are mainly on FNNs. As a specific type
of FNNs, the single-hidden-layer feedforward network (SLFNs)
plays an important role in practical applications[3]. For N arbitrary
distinct samples (xi ,ti ), where xi = [xi1 , xi2 ,· · · , xin ]T ∈ Rn and
ti ∈ Rm , an SLFNs with L hidden nodes and activation function
h(x) are mathematically modeled as
f L (x) =
L
X
βi h(ai · x j + bi ), j = 1,· · · , N
(1)
i=1
where h(ai ,bi ,x) denotes the output of the i th hidden node with
the hidden-node parameters (ai ,bi ) ∈ Rn × R and βi ∈ R is the
output weight between the i th hidden node and the output node.
ai · x denotes the inner product of vector ai and x in Rn .
An active topic on the universal approximation capability of
SLFNs is then how to determine the parameters ai ,bi , and
βi (i = 1,· · · ,L) such that the network output f L (x) can approximate
a given target T,T = [t1 ,· · · ,tN ]. The feature of SLFNs resorts to
parameters of the output weight and hidden nodes parameters.
According to conventional neural network theories, SLFNs are
universal approximators when all the parameters of the networks
This work was supported by National Natural Science Foundation of
China (61175075,61301254,61304007), and and Hunan Provincial Innovation Foundation For Postgraduate (CX2012B147).
Y. M. Yang is with the College of Electric Engineering, Guangxi University,
Nanning 530004, China, and also with the Department of Electrical and
Computer Engineering, University of Windsor N9B 3P4, Canada.
Q. M. Jonathan Wu is with the Department of Electrical and Computer
Engineering, University of Windsor N9B 3P4, Canada.
G. B. Huang is with the School of Electrical and Electronic Engineering,
Nanyang Technological University, Singapore 639798, Singapore.
Y. N. Wang are with the College of Electrical and Information Engineering, Hunan University, Changsha 410082, China.
1) The learning speed of proposed learning method can be
several to thousands of times faster than other learning methods
including SVM, BP and other ELMs. Further more, it can provide
good generalization performance and can be applied in regression
and classification applications directly.
2) Different from conventional SLFNs in which the hidden
node parameter and output weight should be needed, in the
proposed method, we proved that SLFNs without output weight
can still approximate any target continuous function and classify
any disjoint regions. Thus the architecture of this single parameter
neural network is extremely simpler than traditional SLFNs.
3) Different from other neural networks requiring large number
of hidden nodes2 , experimental study shows that the proposed
learning method with only one hidden node can give significant
improvements on accuracy instead of maintaining a large hiddennode-numbers hidden layer.
1 hidden-node parameters can be generated randomly
2 In [13], Huang et al indicate "The generalization performance of ELM
is not sensitive to the dimensionality L of the feature space (the number
of hidden nodes) as long as L is set large enough (e.g., L > 1000 for all the
real-world cases tested in our simulations)."
2
II. P RELIMINARIES AND N OTATION
IV. SLFN S WITHOUT OUTPUT WEIGHT
A. Notations and Definitions
The sets of real, integer, positive real and positive integer
numbers are denoted by R,Z,R+ and Z+ , respectively. Similar to
[1], let ̥2 (X ) be a space of functions f on a compact subset
X in the n-dimensional
Euclidean space Rn such that | f |2 are
R
integrable, that is, X | f (x)|2 d x < ∞. Let ̥2 (Rn ) be denoted by
̥2 . For u, v ∈ ̥2 (X ), the inner product < u, v > is defined by
Z
< u, v >=
u(x)v (x)dx
(2)
X
The norm in ̥2 (X ) space will be denoted as || · ||. L denotes the
number of hidden nodes. For N training samples, x,x ∈ RN×n
denotes the input matrix of network, T ∈ RN×m denotes the
desire output matrix of network. H ∈ RN×m is called the hidden
layer output matrix of the SLFNs; the i th column of H (Hi ) is
the i th hidden node output with respect to inputs. The hidden
layer output matrix Hi is said to be randomly generated function
sequence Hri if the corresponding hidden-node parameters (ai ,bi )
are randomly generated. . en ,en ∈ RN×m denotes the residual error
function for the current network f n with n hidden nodes. I is unit
matrix and I ∈ Rm×m .
III. B IDIRECTIONAL ELM FOR REGRESSION PROBLEM
N ∈ Rn ×R
Theorem 1: [14] Given N training samples {(xi , ti )}i=1
come from the same continuous function, given the sigmoid or
sine activation function h : R → R; Given a error feedback function
sequence He2n (x,a,b) by
He2n = e 2n−1 · (β2n−1 )−1
(3)
If activation function h is sin/cos, given a normalized function u :
R → [0,1]; If activation function h is sigmoid, given a normalized
function u : R → (0,1]. Then for any continuous target function f ,
randomly generated function sequence Hr2n+1 , limn→∞ kf − (Hr1 ·
e
e
β1 + Ĥ2 (â2 , b̂2 )·β2 +· · ·+Hr2n−1 ·β2n−1 + Ĥ2n (â2n , b̂2n )·β2n k = 0 hold
with probability one if
â2n = h −1 (u(He2n )) · x−1 , â2n ∈ Rn
q
b̂2n = mse(h −1 (u(He2n )) − â2n · x), b̂2n ∈ R
e
Ĥ2n = u −1 (h(â2n · x + b̂2n ))
e 2n−1 · He2n
, β2n ∈ R
β2n = e
H2n · (He2n )T
e 2n · Hr2n+1
β2n+1 = r
, β2n+1 ∈ R
H2n+1 · (Hr2n+1 )T
(4)
Basic idea 1: our recent experimental results indicate that in
B-ELM[14], output weight play a very minion role in the network
learning effectiveness. Inspired by these experimental results, in
this proposed method, we directly set output weight equal to unit
matrix.
N ∈ Rn × Rm
Theorem 2: Given N training samples {(xi ,ti )}i=1
come from the same continuous function, given an SLFNs with
any bounded nonconstant piecewise continuous function H : R →
R for additive nodes or sine nodes, for any continues target
function f , obtained error feedback function sequence Hen ,n ∈ Z ,
βn−1 )k = 0 holds with probability one if
l i m n→∞ kf − ( f n−1 + Hen ·β
βn−1 )−1
Hen = en−1 · (β
(8)
βn ∈ Rm×m
β n = β n−1 = I,β
(9)
Proof: The validity of this theorem is obvious because β n−1 = I
βn−1 )−1 = I, Hen equal to en−1 . And we can get ken k = 0.
and (β
Remark 2: When Hen = en−1 = T, it is easy to notice that the
proposed method can reduce the network output error to 0. Thus
the learning problem has been converted into finding optimal
hidden node parameter (a,b) which lead to H(a,b,x) −→ T.
Basic idea 2: For fixed output weight β equal to unit matrix or
vector (β ∈ Rm×m ), seen from equation (8)-(9), to train an SLFN
is simply equivalent to finding a least-square solution a−1 of the
linear system H(a,x) = T. If activation function can be invertible, to
train an SLFN is simply equivalent to pulling back residual error
to input weight. For example, for N arbitrary distinct samples
{x,T}, x ∈ RN×n ,T ∈ RN×m ,T ∈ [0,1], If activation function is sine
function, to train an SLFN is simply equivalent to finding a leastsquare solution â of the linear system a · x = arcsin(T):
kH(â1 ,· · · , ân ,x) − Tk = min kH(a1 ,· · · ,an ,x) − Tk
a
According to [16], the smallest norm least-squares solution of the
above linear system is aˆn = arcsin(en−1 ) · x−1 . Based on this idea,
we give the following theorem.
Lemma 1: [1] Given a bounded nonconstant piecewise continuous function H : R → R, we have
lim
(5)
(6)
(7)
where h −1 and u −1 represent its reverse function, respectively. if
h is sine activation function, h −1 (·) = ar csi n(·); if h is sigmoid
1
activation function, h −1 (·) = −log( (·)
− 1).
Remark 1: Compared with B-ELM, In the proposed method, we
only make two changes. The first one is we set β 1 = · · · = β n−1 =
· · · = I. The second one is the pseudoinverse of input data x−1 has
been changed as x−1 = xT (I+xxT )−1 based on the ridge regression
theory. Although very small changes are made, the experimental
results show that by using this proposed learning method, one
hidden-node SLFNs without output weight (output weight β equal
to unit matrix) can achieve similar generalization performance as
other standard SLFNs with hundreds of hidden nodes. Further
more, different from B-ELM [14] which only work for regression
problem, the proposed method can be applied in regression and
multi-classification applications.
(10)
(a,b)→(a 0 ,b0 )
kH(a · x + b) − H(a0 · x + b0 )k = 0
(11)
Theorem 3: Given N arbitrary distinct samples {x,T},x ∈
RN×n ,T ∈ RN×m , given the sigmoid or sine activation function
h, for any continuous desire output T, there exist limn→∞ kT −
β1 +· · ·+Ĥn (ân , b̂n ,x)β
βn k = 0 hold with probability one
(Ĥ1 (â1 , b̂1 ,x)β
if
Hen = en−1
ân = h −1 (u(Hen )) · x−1 , ân ∈ Rn×m
q
b̂n = mse(h −1 (u(Hen )) − ân · x) , b̂n ∈ Rm
(12)
β1 = β 2 = · · · = βn = I
Ĥn = u −1 (h(ân · x + b̂n ))
(13)
where if activation function h is sin/cos, given a normalized
function u : R → [0,1]; If activation function h is sigmoid, given
a normalized function u : R → (0,1]. h −1 and u −1 represent its
reverse function, respectively. If h is sine activation function,
h −1 (·) = arcsin(u(·); if h is sigmoid activation function, h −1 (·) =
1 − 1), x−1 = xT (I + xxT )−1 .
−log( u(·)
Proof: For an activation function h(x) : R → R, Hen is given by
Hen = h(λn )
(14)
3
In order to let λ2n ∈ Rm , here we give a normalized function
u(·): u(H) ∈ [0,1] if activation function is sin/cos; u(H) ∈ (0,1) if
activation function is sigmoid. Then for sine hidden node
λ2n = h −1 (u(Hen )) = (ar csi n(u(H en ))
(15)
For sigmoid hidden node
λn = h −1 (u(Hen )) = −log(
1
− 1)
u(Hen )
(16)
let λn = an · x, for sine activation function, we have
ân = h −1 (u(Hen )) · x−1 = ar csi n(u(H en )) · x−1
1
− 1) · x −1
u(Hen )
(17)
(18)
where x−1 is the Moore-Penrose generalized inverse of the given
set of training examples[15]. Similar to [16], we have 1: ân =
ar csi n(u(H en )) · x−1 is one of the least-squares solutions of a
general linear system an · x = λn , meaning that the smallest error
can be reached by this solution:
kân · x − λn k = kân x−1 x − λn k = min kan · x − ar csi n(u(H en ))k (19)
an
2: the special solution ân = h −1 (u(Hen ))·x−1 has the smallest norm
among all the least-squares solutions of an · x = λn , which is
guarantee that an ∈ [−1,1]. Although the smallest error can be
reached by equation (17)-(18), we still can reduce its error by
adding bias bn . For sine activation function:
q
b̂n = mse(h −1 (u(Hen )) − aˆn · x)
q
(20)
= mse(ar csi n(u(H en )) − aˆn · x)
For sigmoid activation function
q
b̂n = mse(h −1 (u(Hen )) − aˆn · x)
q
= mse((−log(1/u(Hen ) − 1)) − ân · x)
(21)
According to 19 and Lemma 1, we have
min ku −1 (h(an · x)) − u −1 (h(λn ))k
an
= ku −1 (h(ân · x)) − u −1 (h(λn ))k
> ku
−1
(h(ân · x + b̂n )) − u
−1
(22)
(h(λn ))k = kσk
We consider the residual error as
=kHen k2 (
2βn 〈e n−1 ,Hen 〉
kHen k2
(23)
− β2n )
e
Ĥn = u −1 (h(ân · x + b̂n ))
en−1 − σ
(24)
βn
ên−1
=
β n−1
Because β n = β n−1 = I, we have equation (23) ≥ 0 is still valid for
e
△ = kĤn k2 (
βn k〈ên−1 , βên−1 〉
2kβ
n−1
ê
k βn−1 k2
βn k2 )
− kβ
e
2kên−1 k2
kên−1 k2
βn−1 k2
kβ
e
= kĤn k2 β 2n ≥ 0
β2n )
−β
where u is a normalized function, ai ∈ Rn×m ,bi ∈ Rm . Here, the
proposed the proposed method for SLFN can be summarized in
Algorithm 1.
Algorithm 1 the proposed method algorithm
N ⊂ Rn × Rm ,the
Initialization: Given a training set {(xi ,ti )}i=1
hidden-node output function H(a,b,x), continuous target function f , set number of hidden nodes L = 1, e = T.
Learning step:
while L < L max do
Increase by one the number of hidden nodes L;L = L + 1;
Step 1) set HeL = e;
Step 2) calculate the input weight aL , bias bL based on
equation 12;
Step 3) calculate e after adding the new hidden node L:
e = e − u −1 (h(a · x + b))
end while
Remark 4: Different from other neural network learning methods in which output weight parameter should be adjusted, in the
proposed method, the output weight of SLFNs can be equal to
unit matrix and thus the proposed neural network does not need
output weight at all. Thus the architecture and computational cost
of this proposed method are much smaller than other traditional
SLFNs.
Remark 5: Subsection V.C presents experiments which show
that the proposed method with only one hidden node can give
better generalization performance than the proposed network with
L(L > 1) hidden node. Based on this experimental results, for N arbitrary distinct samples (xi ,ti ) where xi = [xi1 , xi2 ,· · · , xi N ]T ∈ Rn
and ti ∈ Rm , the proposed network is mathematically modeled as
(27)
where u is a normalized function, a1 ∈ Rn×m ,b1 ∈ Rm . Thus
algorithm 1 can be modified as algorithm 2.
N ⊂ Rn × Rm ,the
Initialization: Given a training set {(xi ,ti )}i=1
hidden-node output function H(a,b,x), continues target function f , set number of hidden nodes L = 1.
Learning step:
Step 1) set He1 = T;
Step 2) calculate the input weight a1 , bias b1 based on equation
12;
V. E XPERIMENTAL V ERIFICATION
n−1
= kĤn k2 (
(26)
i=1
Algorithm 2 the proposed method algorithm
Let
=
L
X
u −1 (h(ai · x j + bi )), j = 1,· · · , N
f L (x) = u −1 (h(a1 · x j + b1 )), j = 1,· · · , N
△ =ke n−1 k2 − ke n−1 − Hen · βn k2
=2βn 〈e n−1 ,Hen 〉 − kHen k2 · β2n
Remark 3: According to Theorem 2-3, for N arbitrary distinct
samples (xi ,ti ) where xi = [xi1 , xi2 ,· · · , xi N ]T ∈ Rn and ti ∈ Rm , the
proposed network with L hidden nodes and activation function
h(x) are mathematically modeled as
f L (x) =
For sigmoid activation function, we have
ân = h −1 (u(Hen )) · x−1 = −log(
Now based on equation 25, we have ken−1 k ≥ ken k, so the
sequence ken k is decreasing and bounded below by zero and the
sequence ken k converges.
(25)
To examine the performance of our proposed algorithm (BELM), in this section, we test them on some benchmark regression
and classification problems. Neural networks are tested in SVR,
SVM, BP, EM-ELM,I-ELM, EI-ELM, B-ELM, ELM and proposed the
proposed method.
4
TABLE I
S PECIFICATION OF REGRESSION PROBLEMS
Datasets
#Attri
#Train
# Test
Auto MPG
Machine CPU
Fried
Wine Quality
Puma
California Housing
House 8L
Parkinsons motor
Parkinsons total
Puma
Delta elevators
Abalone
8
6
11
12
9
8
9
26
26
9
6
9
200
100
20768
2898
4500
16000
16000
4000
4000
6000
6000
3000
192
109
20000
2000
3692
4000
6784
1875
1875
2192
3000
1477
SPECIFICATION OF
TABLE II
S MALL /M EDIUM - SIZED CL ASSIFICATION PROBLEMS
Datasets
#Feature
#Train
# Test
A9a
colon-cancer
USPS
Sonar
Hill Valley
Protein
123
2000
256
60
101
357
32561
40
7291
150
606
17766
16281
22
2007
58
606
6621
TABLE III
SPECIFICATION OF L ARGE - SIZED CL ASSIFICATION PROBLEMS
Datasets
#Feature
#Train
# Test
Covtype.binary
Mushrooms
Gisette
Leukemia
Duke
Connect-4
Mnist
DNA
w3a
54
112
5000
7129
7129
126
780
180
300
300000
4000
6000
38
29
50000
40000
1046
4912
280000
4122
1000
34
15
17557
30000
1186
44837
A. Benchmark Data Sets
In order to extensively verify the performance of different algorithms, wide type of data sets have been tested in our simulations,
which are of small size, medium dimensions, large size, and/or high
dimensions. These data sets include 12 regression problems and
15 classification problems. Most of the data sets are taken from
UCI Machine Learning Repository3 and LIBSVM DATA SETS 4 .
Regression Data Sets: The 12 regression data sets(cf.Table I)can
be classified into two groups of data:
1) data sets with relatively small size and low dimensions, e.g.,
Auto MPG, Machine CPU, Puma, Wine, Abalone;
2) data sets with relatively medium size and low dimensions,
e.g., Delta, Fried, California Housing, Parkinsons;
Classification Data Sets: The 15 classification data sets(cf.Table
II and Table III) can be classified into three groups of data:
1) data sets with relative medium size and medium dimensions,
e.g., Sonar, Hill Valley, Wa3, DNA, Mushrooms, A9a, USPS;
2) data sets with relative small size and high dimensions, e.g.,
Colon-cancer, Leukemia, Duke;
3) data sets with relative large size and high dimensions, e.g.,
Protein, Covtype.binary, Gisette, Mnist, Connect-4;
In these data sets, the input data are normalized into [−1,1]
while the output data for regression are normalized into the range
[0,1]. All data sets have been preprocessed in the same way (heldout method). Ten different random permutations of the whole
data set are taken without replacement, and some(see in tables)
are used to create the training set and the remaining is used for
the test set. The average results are obtained over 50 trials for all
problems.
B. Simulation Environment Settings
The simulations of different algorithms on the data sets which
are shown in Table I and Table II are carried out in Matlab 2009a
environment running on the same Windows 7 machine with at 2
GB of memory and an i5-430 (2.33G) processor. The codes used
for SVM and SVR are downloaded from LIBSVM5 , The codes used
for B-ELM, ELM and I-ELM are downloaded from ELM6 .
For SVM and SVR, in order to achieve good generalization
performance, the cost parameter C and kernel parameter γ of SVM
and SVR need to be chosen appropriately. We have tried a wide
range of C and γ. For each data set, similar to [17], we have used 30
different value of C and γ, resulting in a total of 900 pairs of (C ,γ).
The 30 different value of C and γ are {2−15 ,2−14 ,· · · ,214 ,215 }.
Average results of 50 trials of simulations with each combination
of (C ,γ) are obtained and the best performance obtained by
SVM/SVR are shown in this paper.
For BP, the number of hidden nodes are gradually increased by
an interval of 5 and the nearly optimal number of nodes for BP are
then selected based on cross-validation method. Average results
of 50 trails of simulations for each fixed size of SLFN are obtained
and finally the best performance obtained by BP are shown in this
paper as well.
Simulations on large data sets(cf.Table III) are carried out in a
high-performance computer with Intel Xeon E3-1230 v2 processor
(3.2G) and 16-GB memory.
C. Generalization performance comparison of ELM methods with
different hidden nodes
The aim of this subsection is to show that the proposed method
with only one hidden node generally achieves better generalization
performance than other learning methods. And it is also to show
that the proposed method with one hidden node achieves the
best performance than the proposed method with L,L > 1 one
hidden node. In this subsection, I-ELM, ELM, EI-ELM and the
proposed method are compared in one regression problem and
three classification problems: Fried, DNA, USPS and Mushroom.
In these cases, all the algorithms increase the hidden nodes one by
one. More importantly, we find that the testing accuracy obtained
by proposed method is reduced to a very high value when only
one hidden node is used. And the testing accuracy obtained by
proposed method is not increased but is reduced when hidden
node added one by one. This means the proposed method only
need to calculates one-hidden-node parameter(a1 ,b1 ) once and
then SLFNs without output weight can achieve similar generalization performance as other learning method with hundreds of
hidden nodes. Thus in the following experiments, the number of
hidden node equal to one in the proposed method.
3 http://archive.ics.uci.edu/ml/datasets.html
5 http://www.csie.ntu.edu.tw/ cjlin/libsvmtools/datasets/
4 http://www.csie.ntu.edu.tw/ cjlin/libsvmtools/datasets/
6 http://www.ntu.edu.sg/home/egbhuang/elm_codes.html
5
TABLE IV
P ERFORMANCE COMPARISON ( MEAN - MEAN TESTING RMSE; TIME - TRAINING TIME )
Datasets
House 8L
Auto MPG
Machine CPU
Fried
Delta ailerons
PD motor
PD total
Puma
Delta ele
Abalone
Wine
California house
I-ELM (200 nodes)
B-ELM (200 nodes)
EI-ELM (200 nodes, p = 50)
the proposed method (1 nodes)
Mean
time(s)
Mean
time(s)
Mean
time(s)
Mean
time(s)
0.0946
0.1000
0.0591
0.1135
0.0538
0.2318
0.2178
0.1860
0.1223
0.0938
0.1360
0.1801
1.1872
0.2025
0.1909
0.8327
0.4680
0.4639
0.4678
0.5070
0.5313
0.3398
0.3516
1.1482
0.0818
0.0920
0.0554
0.0857
0.0431
0.2241
0.2137
0.1832
0.1156
0.0808
0.1264
0.1450
3.8821
0.3732
0.3469
5.5063
1.3946
4.7680
4.9278
2.1846
1.6206
1.2549
1.7098
7.2625
0.0850
0.0918
0.0551
0.0856
0.0417
0.2251
0.2124
0.1830
0.1155
0.0848
0.1266
0.1505
10.7691
1.3004
1.2633
7.4016
3.5478
3.9016
3.7854
4.2161
4.0240
2.6676
2.7126
12.0832
0.0819
0.0996
0.0489
0.0834
0.0453
0.2210
0.2136
0.1808
0.1174
0.0828
0.1250
0.1420
0.0020
<0.0001
<0.0001
0.0051
<0.0001
0.0037
0.0023
0.0012
<0.0001
0.0017
0.0031
0.0078
TABLE V
P ERFORMANCE COMPARISON ( MEAN - MEAN TESTING RMSE; TIME - TRAINING TIME )
Datasets
EM-ELM (200 nodes)
House 8L
Auto MPG
Machine CPU
Fried
Delta ailerons
PD motor
PD total
Puma
Abalone
Wine
California house
ELM (200 nodes)
the proposed method (1 nodes)
Mean
time(s)
Mean
time(s)
Mean
time(s)
0.0663
0.0968
0.0521
0.0618
0.0421
0.2196
0.2094
0.1478
0.0817
0.1216
0.1302
7.0388
0.0075
0.1385
18.0290
0.1342
0.7394
0.5944
4.8392
0.1638
0.3806
3.5574
0.0718
0.0976
0.0513
0.0619
0.0431
0.2190
0.2076
0.1602
0.0824
0.1229
0.1354
0.8369
0.0156
0.0069
1.3135
0.0616
0.2730
0.2838
0.3728
0.0761
0.1950
0.9753
0.0819
0.0996
0.0489
0.0834
0.0453
0.2203
0.2136
0.1808
0.0828
0.1250
0.1420
0.0020
<0.0001
<0.0001
0.0051
<0.0001
0.0037
0.0023
0.0012
0.0017
0.0031
0.0078
TABLE VII
P ERFORMANCE COMPARISON ( MEAN - MEAN TESTING RMSE; TIME - TRAINING TIME )
Datasets
Eplison-SVR
House 8L
Auto MPG
Machine CPU
Fried
Delta ailerons
California house
PD total
BP
the proposed method (1 nodes)
Mean
time(s)
Mean
time(s)
Mean
time(s)
0.0799
0.0985
0.0727
0.0829
0.0402
0.1529
0.2082
53.6531
0.0234
0.0187
197.9534
6.8718
35.2250
7.2540
0.0790
0.0953
0.0843
0.0591
0.0415
0.1435
0.2120
27.8462
1.6034
0.7129
81.8774
12.6735
54.3081
12.6438
0.0819
0.0996
0.0489
0.0834
0.0453
0.1420
0.2136
0.0020
<0.0001
<0.0001
0.0051
<0.0001
0.0078
0.0023
TABLE VIII
P ERFORMANCE COMPARISON ( MEAN - MEAN TESTING RMSE; TIME - TRAINING TIME )
Datasets
Covtype.binary
Mushrooms
Gisette
Leukemia
W3a
Duke
Connect-4
Mnist
DNA
SVM
ELM
the proposed method (1 nodes)
Mean
time(s)
Mean
time(s)
#node
Mean
time(s)
74.84%
86.90%
77.68%
82.58%
97.18%
86.36%
66.01%
70.85%
93.70%
413.5275
38.6247
309.3968
2.3914
4.5552
0.0156
569.6221
478.4707
0.4680
77.27%
46.97%
88.69%
76.47%
97.25%
79.27%
76.55%
91.60%
84.94%
36.5947
0.9126
6.4093
9.0340
0.9095
7.8437
7.3757
8.1651
0.2122
500
500
500
5000
500
5000
500
500
500
76.55%
88.84%
94.10%
85.29%
98.17%
92.67%
75.40%
84.20%
92.41%
1.2043
0.0047
48.2027
20.9915
0.1872
20.0352
0.7597
8.8858
0.0187
D. Real-world regression problems
The experimental results between proposed the proposed
method and some other incremental ELMs (B-ELM, I-ELM, and
EI-ELM) are given in Table IV-Table V. In these tables, the close
results obtained by different algorithms are underlined and the
apparent better results are shown in boldface. All the incremental
ELMs (I-ELM, B-ELM, EI-ELM) increase the hidden nodes one by
one till nodes-numbers equal to 200, while for fixed ELMs (ELM,
EM-ELM), 200-hidden-nodes are used. It can be seen that the
6
TABLE IX
P ERFORMANCE COMPARISON ( MEAN - MEAN TESTING RMSE; TIME - TRAINING TIME )
Datasets
A9a
Colon
USPS
Sonar
Hill Valley
Protein
SVM
ELM
the proposed method (1 nodes)
Mean
time(s)
Mean
time(s)
#node
Mean
time(s)
77.39%
76.67%
94.65%
86.29%
58.67%
51.18%
295.0603
10.0156
146.4942
0.0172
0.1295
253.5796
85.10%
80.67%
93.54%
80.86%
64.31
67.09%
4.5871
11.6283
2.0639
0.0686
0.1647
5.0919
500
5000
500
500
500
500
85.57%
85.06%
88.86%
75.69%
67.61%
68.76%
0.5714
0.9719
0.4898
<0.0001
0.0047
1.9953
TABLE VI
P ERFORMANCE COMPARISON ( MEAN - MEAN TESTING RMSE; TIME - TRAINING
TIME )
better results are shown in boldface. As seen from those simulation
results given in these tables, the proposed method can always
achieve comparable performance as SVM and ELM with much
Datasets
ELM
(1 nodes) the proposed method (1 nodes) faster learning speed. Take Covtype.binary (large number of training samples with medium input dimensions) and Gisette (medium
Mean
time(s)
Mean
time(s)
number of training samples with high input dimensions).
House 8L
0.1083
0.0009
0.0819
0.0020
1) For Covtype.binary data set, the proposed method runs 1403
Auto MPG
0.2126
< 0.0001
0.0996
<0.0001
times and 35 times faster than ELM and SVM, respectively.
Machine CPU
0.1331
<0.0001
0.0489
<0.0001
2) For Gisette data set, the proposed method runs 341 times
Fried
0.2207
0.0031
0.0834
0.0051
Delta ailerons
0.0864
<0.0001
0.0453
<0.0001
and 1.7 times faster than ELM and SVM, respectively.
PD motor
0.2620
0.0020
0.2210
0.0037
Huang et al.[1][18][16][13] have systematically investigated the
PD total
0.2548
0.0007
0.2136
0.0023
performance of ELM, SVM/SVR and BP for most data sets tested
0.1808
0.0012
Puma
0.2856
0.0012
in this work. It is found that ELM obtain similar generalization
Delta ele
0.1454
<0.0001
0.1174
<0.0001
Abalone
0.1363
0.0007
0.0828
0.0017
performance as SVM/SVR but in much simpler and faster way.
Wine
0.1750
0.0006
0.1250
0.0031
Similar to those above works, our testing results(cf. Table VII-IX)
California house
0.2496
0.0027
0.1420
0.0078
shows that the proposed the proposed method always provide
comparable performance as SVM/SVR and BP with much faster
learning speed.
proposed method can always achieve similar performance as other
On the other hand, the proposed method requires none human
ELMs with much higher learning speed. In Table IV, for Machine intervention than SVM, BP and other ELM methods. Different
CPU problem, the the proposed method runs 1900 times, 3400 from SVM which is sensitive to the combinations of parameters
times, 12000 times faster than the I-ELM, B-ELM and EI-ELM, (C ,γ), or from other ELM methods in which parameter C needs to
respectively. For Abalone problem, the proposed method runs 200 be specified by users, the proposed method have none specified
times, 700 times, 1600 times faster than I-ELM, B-ELM and EI- parameter and is ease of use in the respective implementations.
ELM, respectively. In Table V, for Wine problem, the the proposed
VI. C ONCLUSION
method runs 120 times and 60 times faster than EM-ELM and
ELM, respectively. and the testing RMSE of EI-ELM is 2 times larger
Unlike other SLFN learning methods, in our new approach,
than the testing RMSE of B-ELM. The B-ELM runs 1.5 times faster one may simply calculate the hidden node parameter once and
than the I-ELM and the testing RMSE for the obained I-ELM is 5 the output weight is not need at all. And it has been rigorously
times larger than the testing RMSE for B-ELM.
proved that the proposed method can greatly enhance the learning
If only 1-hidden-node being used, those ELM methods such effectiveness, reduce the computation cost, and eventually further
as I-ELM, ELM, EM-ELM and B-ELM can be considered as the increase the learning speed. The simulation results on sigmoid
same learning method (ELM[13]). Thus in Table VI, we carried out type of hidden nodes show that compared to other learning
performance comparisons between the proposed method and 1- methods including SVM/SVR, BP and ELMs, the new approach
hidden-node ELM. As observed from Table VI, the average testing can significantly reduce the NN training time several to thousands
RMSE obtained by the proposed method are much better than of times and can applied in regression and classification problems.
the ELM. For California house and Delta ailerons problem, the Thus this method can be used efficiently in many applications.
testing RMSE obtained by ELM runs 2 times larger than that of
However, we find an interesting phenomenon which we are not
the proposed method. In real applications, SLFNs with only 1 able to prove in this method, which should be worth pointing out.
hidden nodes is extremely small network structure, meaning that Experimental results show that this proposed learning method
after trained this small size network may response to new external with one hidden node can achieve better generalization perforunknown stimuli much faster and much more accurate than other mance than the same method with L,L > 1 hidden nodes. This
ELM algorithms in real deployment.
phenomenon of this proposed method bring about many advantages, but if researchers can find the nature of this phenomenon,
it can have far reaching consequences on the generalization ability
E. Real-world classification problems
In order to indicate the advantage of the B-ELM on classification of neural network.
performance, the testing accuracy between the proposed the
proposed method and other algorithms has also been conducted.
Table VIII and IX display the performance comparison of SVM,
ELM and the proposed method. In these tables, the close results
obtained by different algorithms are underlined and the apparent
R EFERENCES
[1] Guang-Bin Huang, Lei Chen, and Chee-Kheong Siew. Universal
approximation using incremental constructive feedforward networks
with random hidden nodes. IEEE Transactions on Neural Networks,
17(4):879–892, 2006. [1, 2, 6]
7
[2] Herbert Jaeger. A tutorial on training recurrent neural networks ,
covering BPPT , RTRL , EKF and the " echo state network " approach.
2002:1–46, 2005. [1]
[3] G. B. Huang, P. Saratchandran, and N. Sundararajan. An efficient
sequential learning algorithm for growing and pruning rbf (gap-rbf)
networks. IEEE Transactions on Systems Man and Cybernetics Part
B-cybernetics, 34(6):2284–2292, December 2004. [1]
[4] Guang-Bin Huang. Learning capability and storage capacity of twohidden-layer feedforward networks. IEEE Transactions on Neural
Networks, 14(2):274–281, 2003. [1]
[5] Rui Zhang, Yuan Lan, Guang-Bin Huang, and Zong-Ben Xu. Universal
approximation of extreme learning machine with adaptive growth of
hidden nodes. IEEE Transactions on Neural Networks and Learning
Systems, 23(2):365–371, February 2012. [1]
[6] B. Igelnik and Y. H. Pao. Stochastic choice of basis functions in
adaptive function approximation and the functional-link net. IEEE
transactions on neural networks / a publication of the IEEE Neural
Networks Council, 6(6):1320–9, 1995. [1]
[7] Y. H. PAO, G. H. PARK, and D. J. SOBAJIC. Learning and generalization
characteristics of the random vector functional-link net. Neurocomputing, 6(2):163–180, April 1994. [1]
[8] D.S.Broomhead and D.Lowe. Multivariable functional interpolation
and adaptive networks. Complex Systems, 2:321–355, 1988. [1]
[9] Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew.
Extreme learning machine. In Technical Report ICIS/03/2004 (also in
http://www.ntu.edu.sg/eee/icis/cv/egbhuang.htm), School of Electrical
and Electronic Engineering, Nanyang Technological University, Singapore, January 2004. [1]
[10] Ming-Bin Li, Guang-Bin Huang, P. Saratchandran, and Narasimhan
Sundararajan. Fully complex extreme learning machine. Neurocomputing, 68:306–314, 2005. [1]
[11] Guang-Bin Huang and Chee-Kheong Siew. Extreme learning machine
with randomly assigned RBF kernels. International Journal of Information Technology, 11(1):16–24, 2005. [1]
[12] Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew.
Realtime learning capability of neural networks.
In Technical Report ICIS/45/2003, School of Electrical and Electronic Engineering,
Nanyang Technological University, Singapore, April 2003. [1]
[13] G. B. Huang, H. M. Zhou, X. J. Ding, and R. Zhang. Extreme
learning machine for regression and multiclass classification. IEEE
Transactions on Systems Man and Cybernetics Part B-cybernetics,
42(2):513–529, April 2012. [1, 6]
[14] Y. M. Yang, Y. N. Wang, and X. F. Yuan. Parallel chaos search based
incremental extreme learning machine. Neural Processing Letters,
37(3):277–301, June 2013. [1, 2]
[15] A. E. Hoerl and R. W. Kennard. Ridge regression: Biased estimation for
nonorthogonal problems. Technometrics, 42(1):80–86, February 2000.
[3]
[16] Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme
learning machine: Theory and applications. Neurocomputing, 70:489–
501, 2006. [3, 6]
[17] C. W. Hsu and C. J. Lin. A comparison of methods for multiclass
support vector machines. IEEE Transactions on Neural Networks,
13(2):415–425, March 2002. [4]
[18] Guorui Feng, Guang-Bin Huang, Qingping Lin, and Robert Gay. Error
minimized extreme learning machine with growth of hidden nodes
and incremental learning. IEEE Transactions on Neural Networks,
20(8):1352–1357, 2009. [6]
| 9cs.NE
|
Overcoming Catastrophic Forgetting with Hard Attention to the Task
Joan Serrà 1 Dı́dac Surı́s 1 2 Marius Miron 1 3 Alexandros Karatzoglou 1
arXiv:1801.01423v2 [cs.LG] 14 Feb 2018
Abstract
Catastrophic forgetting occurs when a neural network loses the information learned in a previous
task after training on subsequent tasks. This problem remains a hurdle for artificial intelligence
systems with sequential learning capabilities. In
this paper, we propose a task-based hard attention
mechanism that preserves previous tasks’ information without affecting the current task’s learning.
A hard attention mask is learned concurrently to
every task, through stochastic gradient descent,
and previous masks are exploited to condition
such learning. We show that the proposed mechanism is effective for reducing catastrophic forgetting, cutting current rates by 45 to 80%. We also
show that it is robust to different hyperparameter
choices, and that it offers a number of monitoring
capabilities. The approach features the possibility
to control both the stability and compactness of
the learned knowledge, which we believe makes
it also attractive for online learning or network
compression applications.
1. Introduction
With the renewed interest in neural networks, old problems
re-emerge, specially if the solution is still open. That is
the case with the so-called catastrophic forgetting or catastrophic interference problem (McCloskey & Cohen, 1989;
Ratcliff, 1990). In essence, catastrophic forgetting corresponds to the tendency of a neural network to forget what
it learned upon learning from new or different information.
For instance, when a network is first trained to convergence
on one task, and then trained on a second task, it forgets
how to perform the first task.
Overcoming catastrophic forgetting is an important step
in the advancement towards more general artificial intelligence systems (Legg & Hutter, 2007). Such systems
1
Telefónica Research, Barcelona, Spain 2 Universitat
Politècnica de Catalunya, Barcelona, Spain 3 Universitat Pompeu Fabra, Barcelona, Spain. Correspondence to: Joan Serrà
<[email protected]>.
Ongoing work. Preprint.
should be able to seamlessly remember different tasks, and
to learn them sequentially, following a lifelong learning
paradigm (Thrun & Mitchell, 1995). Apart from being more
biologically plausible (Clegg et al., 1998), there are many
practical situations which require a sequential learning system (cf. Thrun & Mitchell, 1995). For instance, it may be
unattainable for a robot to retrain from scratch its underlying model upon encountering a new object/task. After
accumulating a large number of objects/tasks and their corresponding information, performing concurrent or multitask
learning at scale may be too costly.
Storing previous information and using it to retrain the
model was among the earliest attempts to overcome catastrophic forgetting; a strategy named “rehearsal” (Robins,
1995). The use of memory modules in this context has
been a subject of research until today (Rebuffi et al., 2017;
Lopez-Paz & Ranzato, 2017). However, due to efficiency
and capacity constrains, memory-free approaches were also
introduced, starting with what was termed as “pseudorehearsal” (Robins, 1995). This approach has found some
success in transfer learning situations where one needs to
maintain a certain accuracy on the source task after learning
the target task (Jung et al., 2016; Li & Hoiem, 2017). Within
the pseudo-rehearsal category, we could also consider recent
approaches that substitute the memory module by a generative network (Venkatesan et al., 2017; Shin et al., 2017;
Nguyen et al., 2017). Besides the difficulty of training a
generative network for a sequence of tasks or certain types
of data, both rehearsal and pseudo-rehearsal approaches
imply some form of concurrent learning, that is, having to
re-process ‘old’ instances for learning a new task.
The other popular strategy to overcome catastrophic forgetting is to reduce representational overlap (French, 1991).
This can be done at the output, intermediate, and also input
levels (Gutsein & Stump, 2015; He & Jaeger, 2018). A
clean way of doing that in a soft manner is through so-called
“structural regularization” (Zenke et al., 2017), either present
in the loss function (Kirkpatrick et al., 2017; Zenke et al.,
2017) or at a separate merging step (Lee et al., 2017). With
these strategies, one seeks to prevent major changes in the
weights that were important for previous tasks. Dedicating
specific sub-parts of the network for each task is another
way of reducing representational overlap (Rusu et al., 2016;
Fernando et al., 2017; Yoon et al., 2018). The main trade-
Overcoming Catastrophic Forgetting with Hard Attention to the Task
off in representational overlap is to effectively distribute
the capacity of the network across tasks while maintaining
important weights and reusing previous knowledge.
In this paper, we propose a task-based hard attention mechanism that maintains the information from previous tasks
without affecting the learning of a new task. Concurrently
to learning a task, we also learn almost-binary attention
vectors through gated task embeddings, using backpropagation and minibatch stochastic gradient descent (SGD). The
attention vectors of previous tasks are used to define a mask
and constrain the updates of the network’s weights on current tasks. Since masks are almost binary, a portion of the
weights remains static while the rest adapt to the new task.
We call our approach hard attention to the task (HAT). We
evaluate HAT in the context of image classification, using
what we believe is a high-standard evaluation protocol: we
consider random sequences of 8 publicly-available data sets
representing different tasks, and compare with a dozen of
recent competitive approaches. We show favorable results in
4 different experimental setups, cutting current rates by 45
to 80%. We also show robustness with respect to hyperparameters and illustrate a number of monitoring capabilities.
We make the code for all experiments publicly-available1 .
2. Putting Hard Attention to the Task
2.1. Motivation
The primary observation that drives the proposed approach
is that the task definition or, more pragmatically, its identifier, is crucial for the operation of the network. Consider
the task of discriminating between bird and dog images.
When training the network to do so, it may learn some
set of intermediate features. If the second task is to discriminate between brown and black animals using the same
data (assuming it only contained birds and dogs that were
either brown or black), the network may learn a new set
of features, some of them with not much overlap with the
first ones. Thus, if training data is the same in both tasks,
one important difference should be the task description or
identifier. Our intention is to learn to use the task identifier
to condition every layer, and to later exploit this learned
conditioning to prevent forgetting previous tasks.
2.2. Architecture
To condition to the current task t, we employ a layer-wise
attention mechanism (Fig. 1, top). Given the output of the
units2 of layer l, hl , we element-wise multiply h0l = atl hl .
However, an important difference with common attention
1
Code is attached in a ZIP file as Supplementary Material.
In the remaining of the paper, we will use ‘units’ to refer to
both linear units (or fully-connected neurons) and convolutional
filters. HAT can be extended to any parametric layer.
2
Figure 1. Schematic diagram of the proposed approach: forward
(top) and backward (bottom) passes.
mechanisms is that, instead of forming a probability distribution, atl is a gated version of a task embedding vector etl ,
atl = σ setl ,
(1)
where σ(x) ∈ [0, 1] is a gate function and s is a positive
scaling parameter. We use a sigmoid gate in our experiments,
but note that other gating mechanisms could be used. All
layers l = 1, . . . L − 1 operate equally except the last one,
layer L, where atL is binary hard-coded. The operation
of layer L is equivalent to a multi-head output (Bakker &
Heskes, 2003), which is routinely employed in the context
of catastrophic forgetting (for example Rusu et al., 2016; Li
& Hoiem, 2017; Nguyen et al., 2017).
The idea behind the gating mechanism of Eq. 1 is to form
hard, possibly binary attention masks which, act as “inhibitory synapses” (McCulloch & Pitts, 1943), and can thus
activate or deactivate the output of the units of every layer.
In this way, and similar to PathNet (Fernando et al., 2017),
we dynamically create and destroy paths across layers that
can be later preserved when learning a new task. However,
unlike PathNet, the paths in HAT are not based on modules,
but on single units. Therefore, we do not need to pre-assign
a module size nor to set a maximum number of modules
per task. Given some network architecture, HAT learns
and automatically dimensions individual-unit paths, which
ultimately affect individual layer weights. Furthermore, instead of learning paths in a separate stage using genetic
algorithms, HAT learns them together with the rest of the
network, using backpropagation and SGD.
Overcoming Catastrophic Forgetting with Hard Attention to the Task
2.3. Network Training
To preserve the information learned in previous tasks upon
learning a new task, we condition the gradients according
to the cumulative attention from all the previous tasks. To
obtain a cumulative attention vector, after learning task t
and obtaining atl , we recursively compute
t ≤t−1
a≤t
,
l = max al , al
using element-wise maximum and the all-zero vector for
a≤0
l . This preserves the attention values for units that were
important for previous tasks, allowing them to condition the
training of future tasks.
To condition the training of task t + 1, we modify the gradient gl,ij at layer l with the reverse of the minimum of the
cumulative attention in the current and previous layers:
h
i
≤t
0
gl,ij
= 1 − min a≤t
,
a
gl,ij ,
(2)
l,i
l−1,j
where the unit indices i and j correspond to the output
(l) and input (l − 1) layers, respectively. In other words,
≤t
we expand the vectors a≤t
l and al−1 to match the dimensions of the gradient tensor of layer l, and then perform
a element-wise minimum, subtraction, and multiplication
(Fig. 1, bottom). We do not compute any attention over the
input data if this consists of complex signals like images or
audio. However, in the case such data consisted of separate
or independent features, one could also consider them as the
output of some layer and apply the same methodology.
Note that, with Eq. 2, we create masks to prevent large updates to the weights that were important for previous tasks.
This is similar to the approach of PackNet (Mallya & Lazebnik, 2017), which was made public during the development
of HAT. In PackNet, after an heuristic selection and retraining, a binary mask is found and later applied to freeze the
corresponding network weights. In this regard, HAT differs
from PackNet in three important aspects. Firstly, our mask
is unit-based, with weight-based masks automatically derived from those. Therefore, HAT also stores and maintains
a lightweight structure. Secondly, our mask is learned, instead of heuristically- or rule-driven. Therefore, HAT does
not need to pre-assign compression ratios nor to determine
parameter importance through a post-training step. Thirdly,
our mask is not necessarily binary, allowing intermediate
values between 0 and 1. This can be useful if we want to
reuse weights for learning other tasks, at the expense of
some forgetting, or we want to work in a more online mode,
forgetting the oldest tasks to remember new ones.
2.4. Hard Attention Training
To obtain a totally binary attention vector atl , one could use
a unit step function as gate. However, since we want to
train the embeddings etl with backpropagation (Fig. 1), we
prefer a differentiable function. To construct a pseudo-step
function that allows the gradient to flow, we use a sigmoid
with a positive scaling parameter s (Eq. 1). This scaling
is introduced to control the polarization, or ‘hardness’, of
the pseudo-step function and the resulting output atl . Our
strategy is to anneal s during training, inducing a gradient
flow, and set s = smax during testing, using smax 1
such that Eq. 1 approximates a unit step function. Notice
that when s → ∞ we get atl,i → {0, 1}, and that when
s → 0 we get atl,i → 1/2. We will use the latter to start a
training epoch with all network units being equally active,
and progressively polarize them within the epoch.
During a training epoch, we incrementally linearly anneal
the value of s by
1
1
b−1
s=
+ smax −
,
(3)
smax
smax B − 1
where b = 1, . . . B is the batch index and B is the total
number of batches in an epoch. The hyperparameter smax ≥
1 controls the stability of the learned tasks or, in other words
the plasticity of the network’s units. If smax is close to
1, the gating mechanism operates like a regular sigmoid
function, without particularly enforcing the binarization of
atl . This provides plasticity to the units, with the model being
able to forget previous tasks at the backpropagation stage
(Sec. 2.3). If, alternatively, smax is a larger number, the
gating mechanism starts operating as a unit step function.
This provides stability with regard to previously learned
tasks, preventing changes in the corresponding weights at
the backpropagation stage.
2.5. Embedding Gradient Compensation
In preliminary analysis, we empirically observed that embeddings etl were not changing much, and that the magnitude
of the gradient was weak on those weights. After some investigation, we realized that the major part of the problem
was due to the introduced annealing scheme (Eq. 3). To
illustrate the effect of the annealing scheme on the gradients
of etl , consider a uniformly distributed embedding etl,i across
the active range of a standard sigmoid, etl,i ∈ [−6, 6]. If we
do not perform any annealing and set s = 1, we obtain a cumulative gradient after one epoch that has a bell-like shape
and spans the whole sigmoid range (Fig. 2). Contrastingly, if
we set s = smax , we obtain a much larger magnitude, but in
a much lower range (etl,i ∈ [−1, 1] in Fig. 2). The annealed
version of s yields a distribution in-between, with a lower
range than s = 1 and a lower magnitude than s = smax .
A desirable situation would be to have a wide range, ideally spanning the range of s = 1, and a large cumulative
magnitude, ideally proportional to the one in the active region when s = smax . To achieve that, we apply a gradient
compensation before updating etl .
Overcoming Catastrophic Forgetting with Hard Attention to the Task
Cumulative ql, i
Constant s = 1
Constant s = smax
Annealed s
Compensated
−6
−4
−2
0
el,t i
2
4
6
Figure 2. Illustration of the effect that annealing s has on the gradient q of et .
In essence, the idea of the embedding gradient compensation is to remove the effects of the annealed sigmoid and
to artificially impose the desired range and magnitude motivated in the previous paragraph. To do so, we divide the
gradient ql,i by the derivative of the annealed sigmoid, and
multiply by the desired compensation,
h
i
smax σ etl,i 1 − σ etl,i
0
h
i ql,i ,
ql,i
=
sσ setl,i 1 − σ setl,i
which, after operating, yields
h
i
smax cosh setl,i + 1
0
h
i ql,i .
ql,i
=
s cosh etl,i + 1
The use of L1 regularization to promote network sparsity
in the context of catastrophic forgetting has also been considered by Yoon et al. (2018) with dynamically expandable
networks (DEN), which were introduced while developing
HAT. In DEN, plain L1 regularization is combined with a
considerable set of heuristics such as L2-transfer, thresholding, and a measure of “semantic drift”, and is applied
to all network weights in the so-called “selective retraining”
phase. In HAT, we use an attention-weighted L1 regularization over attention values, which is an independent part of
the single training phase of the approach. Instead of considering network weights, HAT focuses on unit attention.
3. Related Work
|setl,i |
For numerical stability, we clamp
≤ 50 and constrain
etl,i to remain within the active range of the standard sigmoid,
etl,i ∈ [−6, 6]. In any case, however, ql,i → 0 when we hit
those limits. That is, we are in the constant regions of
the pseudo-step function. Notice also that, by Eq. 3, the
minimum s is never equal to 0.
2.6. Promoting Low Capacity Usage
It is important to realize that the hard attention values
atl,i that are ‘active’, that is, atl,i → 1, directly determine the units that will be dedicated to task t. Therefore, in order to have some model capacity reserved for
future tasks, we promote sparsity on the set of attention
vectors At = {at1 , . . . atL−1 }. To do so, we add a regularization term to the loss function L that takes into account
the set of cumulative attention vectors up to task t − 1,
<t
A<t = {a<t
1 , . . . aL−1 }:
L0 y, ŷ, At , A<t = L (y, ŷ) + cR At , A<t , (4)
where c is the regularization constant,
PL−1 PNl t
al,i 1 − a<t
l=1
i=1
l,i
R At , A<t =
PL−1 PNl
<t
l=1
i=1 1 − al,i
is the regularization term, and Nl corresponds to the number of units in layer l. Notice that Eq. 5 corresponds to a
weighted and normalized L1 regularization over At . Cumulative attentions over the past tasks A<t define a weight for
t
the current task, such that if a<t
l,i → 1 then al,i receives a
weight close to 0 and vice versa. This excludes the units
that were attended in previous tasks from regularization,
unconstraining their reuse in the current task. The hyperparameter c ≥ 0 controls the capacity spent on each task
(Eq. 4). In a sense, it can be thought of as a compressibility
constant, affecting the compactness of the learned models:
the higher the c, the lower the number of active attention
values atl,i and the more sparse the resulting network is. We
set c globally for all tasks and let HAT adapt to the best
compression for each individual task.
(5)
We compare the proposed approach with the conceptually
closest works, some of which appeared concurrently to the
development of HAT. A more general overview of related
work has been done in Sec. 1. A qualitative comparison
with three of the most related strategies has been done along
Sec. 2. A quantitative comparison with these and other
approaches is done in Sec. 4 and Appendix C.
Both elastic weight consolidation (EWC; Kirkpatrick et al.,
2017) and synaptic intelligence (SI; Zenke et al., 2017) approaches add a ‘soft’ structural regularization term to the
loss function in order to discourage changes to weights that
are important for previous tasks. HAT uses a ‘hard’ structural regularization, and does it both at the loss function
and gradient magnitudes explicitly. EWC measures weights’
importance after network training, while SI and HAT compute weights’ importance concurrently to network training.
EWC and SI use specific formulation while HAT learns
attention masks. Incremental moment matching (IMM; Lee
et al., 2017) is an evolution of EWC, performing a separate
model-merging step after learning a new task.
Progressive neural networks (PNNs; Rusu et al., 2016) distribute the network weights in a column-wise fashion, pre-
Overcoming Catastrophic Forgetting with Hard Attention to the Task
assigning a column width per task. They employ so-called
adapters to reuse knowledge from previous columns/tasks,
leading to a progressive increase of the number of weights
assigned to future tasks. Instead of blindly pre-assigning
column widths, HAT learns such ‘widths’ per layer, together
with the network weights, and adapts them to the difficulty
of the current task. PathNet (Fernando et al., 2017) also
pre-assigns some amount of network capacity per task but,
in contrast to PNNs, avoids network columns and adapters.
It uses an evolutionary approach to learn paths between a
constant number of so-called modules (layer subsets) that
interconnect between themselves. HAT does not maintain a
population of solutions, entirely trains with backpropagation
and SGD, and does not rely on a constant set of modules.
Together with PNNs and PathNet, PackNet (Mallya &
Lazebnik, 2017) also employs a binary mask to constrain the
network. However, such constrain is not based on columns
nor layer modules, but on network weights. Therefore, it
allows for a potentially better use of the network’s capacity. PackNet is based on heuristic weight pruning, with
pre-assigned pruning ratios. HAT also focuses on network
weights, but uses unit-based masks to constrain those, which
also results in a lightweight structure. It avoids any absolute
or pre-assigned pruning ratio, although it uses the compressibility parameter c to influence the compactness of the
learned models. Another difference between HAT and the
previous three approaches is that it does not use purely binary masks. Instead, the stability parameter smax controls
the degree of binarization.
Dynamically expandable networks (DEN; Yoon et al., 2018)
also assign network capacity depending on the task at hand.
However, they do so in a separate stage called “selective
retraining”. A complex mixture of heuristics and hyperparameters is used to identify “drifting” units, which are
duplicated and retrained in another stage. L1 regularization and L2-transfer are employed to condition learning,
together with the corresponding regularization constants
and an additional set of thresholds. HAT strives for simplicity, restricting the number of hyperparameters to two that
have a straightforward conceptual interpretation. Instead
of plain L1 regularization over network weights, HAT employs an attention-weighted L1 regularization over attention
masks. Attention masks are a lightweight structure that can
be plugged in without the need of introducing important
changes to a pre-existing network.
4. Experiments
Setups — Common setups to evaluate catastrophic forgetting in a classification context are based on permutations
of the MNIST data (Srivastava et al., 2013), label splits
of the MNIST data (Lee et al., 2017), incrementally learning classes of the CIFAR data sets (Lopez-Paz & Ranzato,
2017), or two-task transfer learning setups where accuracy
is measured on both source and target tasks (Li & Hoiem,
2017). However, there are some limitations with these setups. Firstly, performing permutations of the MNIST data
has been suggested to favor certain approaches, yielding misleading results3 in the context of catastrophic forgetting (Lee
et al., 2017). Secondly, using only the MNIST data may
not be very representative of modern computer vision tasks,
nor particularly challenging (Xiao et al., 2017). Thirdly,
incrementally adding classes or groups of classes implies
the assumption that all data comes from the same joint distribution, which is unrealistic for a real-world setting. Finally,
evaluating catastrophic forgetting with only two tasks biases
the conclusions towards transfer learning setups, and prevents the analysis of truly sequential learning with more than
two tasks. In this paper, we consider the aforementioned
MNIST and CIFAR setups (Sec. 4.2). Nonetheless, we primarily evaluate on a sequence of multiple tasks formed by
different classification data sets (Sec. 4.1).
To obtain a generic estimate, we weigh a number of tasks
and uniformly randomize their order. After training task t,
we compute the accuracies on all testing sets of tasks τ ≤ t.
We repeat 10 times this sequential train/test procedure with
10 different seed numbers, which are also used in the rest of
randomizations and initializations (see below). To compare
between different task accuracies, and in order to obtain
a general measurement of the amount of forgetting, we
introduce the forgetting ratio
ρτ ≤t =
Aτ ≤t − AτR
AτJ ≤t − AτR
− 1,
(6)
where Aτ ≤t is the accuracy measured on task τ after sequentially learning task t, AτR is the accuracy of a random,
frequency-based classifier solely trained on task τ , and AτJ ≤t
is the accuracy measured on task τ after jointly learning t
tasks in a multitask fashion. Note that ρ ≈ −1 and ρ ≈ 0
correspond to performances close to the ones of the random
and multitask classifiers, respectively. To report a single
number after learning t tasks, we take the average
t
ρ≤t =
1 X τ ≤t
ρ .
t τ =1
Data — We consider 8 common image classification data
sets and adapt them, if necessary, to an input size of
32 × 32 × 3 pixels. The number of classes goes from
10 to 100, training set sizes from 16,853 to 73,257, and
test set sizes from 1,873 to 26,032. For each task, we ran3
Essentially, the MNIST data contains many values close to
0 that allow for an easier identification of the important units or
weights which, if permuted, can then be easily frozen without
overlapping with the ones of the other tasks (see Lee et al., 2017).
Overcoming Catastrophic Forgetting with Hard Attention to the Task
Baselines — We consider 2 reference approaches plus
9 recent and competitive ones: standard SGD with
dropout (Goodfellow et al., 2014), SGD freezing all layers except the last one (SGD-F), EWC, IMM (Mean and
Mode variants), learning without forgetting (LWF; Li &
Hoiem, 2017), less-forgetting learning (LFL; Jung et al.,
2016), PathNet, and PNNs. To find the best hyperparameter
combination for each approach, we perform a grid search using a task sequence determined by a single seed. To compute
the forgetting ratio ρ (Eq. 6), we also run the aforementioned
random and multitask classifiers.
Network — Unless stated otherwise, we employ an
AlexNet-like architecture (Krizhevsky et al., 2012) with
3 convolutional layers of 64, 128, and 256 filters with 4 × 4,
3 × 3, and 2 × 2 kernel sizes, respectively, plus two fullyconnected layers of 2048 units each. We use rectified linear
units as activations, and 2 × 2 max-pooling after the convolutional layers. We also use a dropout of 0.2 for the
first two layers and of 0.5 for the rest. A fully-connected
layer with a softmax output is used as a final layer, together
with categorical cross entropy loss. All layers are randomly
initialized with Xavier uniform initialization (Glorot & Bengio, 2010) except the embedding layers, for which we use a
Gaussian distribution N (0, 1). Unless stated otherwise, our
code uses PyTorch’s defaults for version 0.2.0 (Paszke et al.,
2017). We adapt the same base architecture to all baseline
approaches and match their number of parameters to 7.1 M.
0.0
−0.1
−0.2
ρ ≤t
domly split 15% of the training set and keep it for validation purposes. The considered data sets are: CIFAR10 and
CIFAR100 (Krizhevsky, 2009), FaceScrub (Ng & Winkler,
2014), FashionMNIST (Xiao et al., 2017), NotMNIST (Bulatov, 2011), MNIST (LeCun et al., 1998), SVHN (Netzer
et al., 2011), and TrafficSigns (Stallkamp et al., 2011). For
further details on data we refer to Appendix A.
−0.3
−0.4
−0.5
−0.6
1
2
Multitask
SGD
SGD-F
3
4
t
EWC
IMM-Mode
IMM-Mean
5
6
7
LWF
LFL
PathNet
8
PNN
HAT
Figure 3. Average forgetting ratio ρ≤t for the considered approaches (10 runs).
Table 1. Average forgetting ratio after the second (ρ≤2 ) and the
last (ρ≤8 ) task for the considered approaches (10 runs, standard
deviation into parenthesis).
A PPROACH
LFL
LWF
SGD
IMM-M ODE
SGD-F
IMM-M EAN
EWC
PATH N ET
PNN
HAT
ρ≤2
ρ≤8
-0.73 (0.29)
-0.14 (0.13)
-0.20 (0.08)
-0.11 (0.08)
-0.20 (0.15)
-0.12 (0.10)
-0.08 (0.06)
-0.09 (0.16)
-0.11 (0.10)
-0.02 (0.03)
-0.92 (0.08)
-0.80 (0.06)
-0.66 (0.03)
-0.49 (0.05)
-0.44 (0.06)
-0.42 (0.04)
-0.25 (0.03)
-0.17 (0.23)
-0.11 (0.01)
-0.06 (0.01)
4.1. Results
beyond a transfer learning setup. We find LFL extremely
sensitive to the configuration of its hyperparameter, to the
point that what is a good value for one seed, turns out to
be a bad choice for another seed. Hence the poor average
performance for 10 seeds. The highest standard deviations
are obtained by LFL and PathNet (Table 1), which suggests
a high sensitivity with respect to hyperparameters, initializations, or data sets. Another thing to note is that the IMM
approaches only perform similarly or slightly better than
the SGD-F reference. We believe this is due to both the
different nature of the tasks’ data and the consideration of
more than two tasks, which complicates the choice of the
mixing hyperparameter.
We first look at the average forgetting ratio ρ≤t after learning
task t (Fig. 3). A first thing to note is that not all the considered baselines perform better than the SGD references. That
is the case of LWF and LFL. For LWF, we observe it is still
competitive in the two-task setup for which it was designed,
t = 2. However, its performance rapidly degrades for t > 2,
indicating that the approach has difficulties in extending
The best performing baselines are EWC, PathNet, and PNN.
PathNet and PNN present contrasting behaviors. Both, by
construction, never forget; therefore, the important difference is in their learning capability. PathNet starts by correctly learning the first task and progressively exhibits difficulties to do so for t ≥ 2. Contrastingly, PNNs exhibits
difficulty in the first tasks and becomes better as t increases.
Training — We train all models with backpropagation and
plain SGD, using a learning rate of 0.05, and decaying it by
a factor of 3 if there is no improvement in the validation loss
for 5 consecutive epochs. We stop training when we reach
a learning rate lower than 10−4 or we have iterated over
200 epochs (we made sure that all considered approaches
reached a stable solution before 200 epochs). Batch size is
set to 64. All methods use the same task sequence, data split,
batch shuffle, and weight initialization for a given seed.
Overcoming Catastrophic Forgetting with Hard Attention to the Task
We now move to the HAT results. First of all, we observe
that HAT consistently performs better than all considered
baselines for all t ≥ 2 (Fig. 3). For the case of t = 2, it
obtains an average forgetting ratio ρ≤2 = −0.02, while the
best baseline is EWC with ρ≤2 = −0.08 (Table 1). For
the case of t = 8, HAT obtains ρ≤8 = −0.06, while the
best baseline is PNN with ρ≤8 = −0.11. This implies
a reduction in forgetting of 75% for t = 2 and 45% for
t = 8. Notice that the standard deviation of HAT is lower
than the ones obtained by the big majority of the baselines
(Table 1). This denotes a certain stability of HAT with
respect to different task sequences, data sets, data splits, and
network initializations.
Given the slightly increasing tendency of PNN with t
(Fig. 3), one could speculate that PNN would score above
HAT for t > 8. However, our empirical analyzes suggest
that that is not the case (presumably due to the capacity
pre-assignment and parameter increase problems underlined
in Sec. 3 and above). In particular, we observe a gradual
lowering of PathNet and PNN curves with increasing sequences from t = 2 to 8. In addition, we observe PathNet
and PNN obtaining worse performances than EWC in the
case of t = 10 for the incremental class setup (see below
and Appendix C.1). In general, none of the baseline methods consistently outperforms the rest across setups and for
all t, a situation that we do observe with HAT.
4.2. Additional Results
To broaden the strength of our results, we additionally experiment with three common alternative setups. First, we
consider an incremental class learning scenario, similar
to Lopez-Paz & Ranzato (2017), using class subsets of both
CIFAR10 and CIFAR100 data. In this setup, the best baseline after t ≥ 3 is EWC, with ρ≤10 = −0.18. HAT scores
ρ≤10 = −0.09 (55% forgetting reduction). Next, we consider the permuted MNIST sequence of tasks (Srivastava
et al., 2013). In this setup, the best result we could find in the
literature was from SI, with A≤10 = 97.1%. HAT scores
A≤10 = 98.6% (52% error rate reduction). Finally, we also
consider the split MNIST task of Lee et al. (2017). In this
setup, the best result from the literature corresponds to the
conceptor-aided backpropagation approach (He & Jaeger,
2018), with A≤2 = 94.9%. HAT scores A≤2 = 99.0%
(80% error rate reduction). The detail for all these setups
and results can be found in Appendix C.
0.00
−0.05
ρ ≤8
These contrasting behaviors are due to the way the two approaches allocate the network capacity. As mentioned, they
cannot do it dynamically, and therefore need to pre-assign
a number of network weights per task. When having more
tasks but the same network capacity, this pre-assignment
increasingly harms the performance of these baselines, lowering the corresponding curves in Fig. 3.
−0.10
−0.15
−0.20
0
1
smax = 25
smax = 50
2
c
smax = 100
smax = 200
3
4
smax = 400
smax = 800
Figure 4. Effect of hyperparameters smax and c on average forgetting ratio ρ≤8 . Results for seed 0.
4.3. Hyperparameters
In any machine learning algorithm, it is important to assess
the sensitivity with respect to the hyperparameters. HAT
has two: the stability parameter smax and the compressibility parameter c (Secs. 2.4 and 2.6). A low smax provides
plasticity to the units and capacity of adaptation, but the
network may easily forget what it learned. A high smax
prevents forgetting, but the network may have difficulties in
adapting to new tasks. A low c allows to use almost all of
the network’s capacity for a given task, potentially spending
too much in the current task. A high c forces it to learn
a very compact model, at the expense of not reaching the
accuracy that the original network could have reached. We
empirically found good operation ranges smax ∈ [25, 800]
and c ∈ [0.1, 2.5]. As we can see, any variation within these
ranges results in reasonable performance (Fig. 4). Unless
stated otherwise, we use smax = 400 and c = 0.75.
4.4. Monitoring and Network Pruning
It is interesting to note that the hard attention mechanism introduced in Sec. 2 offers a number of possibilities to monitor
the behavior of our models. For instance, by computing the
conditioning mask in Eq. 2 from the hard attention vectors
a≤t
l , we can assess which weights obtain a high attention
value, binarize it, and compute an estimate of the instantaneous network capacity usage (Fig. 5). We may also inform
ourselves of the amount of active weights per layer and task
(Appendix B.2). Another facet we can monitor is the weight
reuse across tasks. By a similar procedure, comparing the
conditioning masks between tasks ti and tj , j > i, we
can asses the percentage of weights of task ti that are later
reused in task tj (Fig. 6).
Another by-product of hard attention masks is that we can
Overcoming Catastrophic Forgetting with Hard Attention to the Task
80
80
60
60
A 1 [%]
100
Used capacity [%]
100
40
20
0
20
0
200
400
Epochs
FaceScrub
MNIST
CIFAR100
0
600
NotMNIST
SVHN
CIFAR10
TrafficSigns
FashionMNIST
Figure 5. Network capacity usage with sequential task learning
(seed 0). Dashed vertical lines correspond to a task switch.
1
5
2
6
5
5
5
5
5
4
3
3
4
4
3
9
8
11
10
9
8
10
9
9
11
11
10
13
12
3
Task
40
4
5
6
0
20
CIFAR10
CIFAR100
FaceScrub
40
60
Compression [%]
FashionMNIST
MNIST
NotMNIST
80
100
SVHN
TrafficSigns
Figure 7. Validation accuracy A1 as a function of compression
percentage. Every dot corresponds to an epoch and triangles match
the accuracy of the SGD approach (no compression).
the task (Appendix B.3). Comparing these numbers with the
compression rates used by PackNet (25 or 50%), we see that
HAT generally uses a much more compact model. Comparing with DEN on the specific MNIST and CIFAR100 tasks
(18 and 52%), we observe that HAT compresses to 1 and
21%, respectively. Interestingly, and in contrast to these and
the majority of network pruning approaches, HAT learns to
prune network weights through backpropagation and SGD,
and at the same time as the network weights themselves.
13
7
2
3
4
5
Task
6
7
8
Figure 6. Percentage of weight reuse across tasks. Seed 0 sequence:
(1) FaceScrub, (2) MNIST, (3) CIFAR100, (4) NotMNIST, (5)
SVHN, (6) CIFAR10, (7) TrafficSigns, and (8) FashionMNIST.
use them to assess which of the network’s weights are important, and then prune the most irrelevant ones (LeCun et al.,
1990). This way, we can compress the network for further
deployment in low-resource devices or time-constrained
environments (cf. Han et al., 2016). If we want to focus
on such compression task, we can set c to a higher value
than the one used for catastrophic forgetting and start with
a positive random initialization of the embeddings el . The
former will promote more compression while the latter will
ensure we start learning the model by putting attention to all
weights in the first epochs (full capacity). We empirically
found that using c = 1.5 and U(0, 2) yields a reasonable
trade-off between accuracy and compression for a single
task (Fig. 7). With that, we can compress the network to
sizes between 1 and 21% of its original size, depending on
5. Conclusion
We introduce HAT, a hard attention mechanism that, by
focusing on a task embedding, is able to protect the information of previous tasks while learning new tasks. This
hard attention mechanism is lightweight, in the sense that
it adds a small fraction of weights to the base network,
and is trained together with the main model, with negligible overhead using backpropagation and vanilla SGD. We
demonstrate the effectiveness of the approach to control
catastrophic forgetting in the image classification context by
running a series of experiments with multiple data sets and
state-of-the-art approaches. HAT has only two hyperparameters, which intuitively refer to the stability and compactness
of the learned knowledge, and whose tuning we demonstrate
is not crucial for obtaining good performance. In addition,
HAT offers the possibility to monitor the used network capacity across tasks and layers, the unit reuse across tasks,
and the compressibility of a model trained for a given task.
We hope that our approach may be also useful in online
learning or network compression contexts, and that the hard
attention mechanism presented here may also find some
applicability beyond the catastrophic forgetting problem.
Overcoming Catastrophic Forgetting with Hard Attention to the Task
References
Bakker, B. and Heskes, T. Task clustering and gating for
bayesian multitask learning. Journal of Machine Learning Research, 4:83–99, 2003.
Bulatov, Y.
NotMNIST dataset.
Technical report,
2011. URL http://yaroslavvb.blogspot.it/
2011/09/notmnist-dataset.html.
Clegg, B. A., DiGirolamo, G. J., and Keele, S. W. Sequence
learning. Trends in Cognitive Sciences, 2(8):275–281,
1998.
Duchi, J., Hazan, E., and Singer, Y. Adaptive subgradient
methods for online learning and stochastic optimization.
Journal of Machine Learning Research, 12:2121–2159,
2011.
Evgeniou, T. and Pontil, M. Regularized multi-task learning.
In Proc. of the ACM SIGKDD Int. Conf. on Knowledge
Discovery and Data Mining (KDD), pp. 109–117, 2004.
Kemelmacher-Shlizerman, I., Seitz, S. M., Miller, D., and
Brossard, E. The megaface benchmark: 1 million faces
for recognition at scale. In Proc. of the IEEE Conf. on
Computer Vision and Pattern Recognition (CVPR), pp.
4873–4882, 2016.
Kingma, D. P. and Ba, J. L. Adam: a method for stochastic
optimization. In Proc. of the Int. Conf. on Learning
Representations (ICLR), 2015.
Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T.,
Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., and Hadsell, R. Overcoming catastrophic forgetting in neural networks. Proc. of the National Academy
of Sciences of the USA, 114(13):3521–3526, 2017.
Krizhevsky, A. Learning multiple layers of features from
tiny images. Msc thesis, University of Toronto, Toronto,
Canada, 2009.
Fernando, C., Banarse, D., Blundell, C., Zwols, Y., Ha, D.,
Rusu, A. A., Pritzel, A., and Wierstra, D. PathNet: evolution channels gradient descent in super neural networks.
ArXiv, 1701.08734, 2017.
Krizhevsky, A., Sutskever, I., and Hinton, G. ImageNet
classification with deep convolutional neural networks.
In Pereira, F., Burges, C. J. C., Bottou, L., and Weinberger,
K. Q. (eds.), Advances in Neural Information Processing
Systems (NIPS), volume 25, pp. 1097–1105. 2012.
French, R. M. Using semi-distributed representations to
overcome catastrophic forgetting in connectionist networks. In Proc. of the Annual Conf. of the Cognitive
Science Society (CogSci), pp. 173–178, 1991.
LeCun, Y., Denker, J. S., and Solla, S. A. Optimal brain
damage. In Touretzky, D. S. (ed.), Advances in Neural
Information Processing Systems (NIPS), volume 2, pp.
598–605. Morgan Kaufmann, 1990.
Glorot, X. and Bengio, Y. Understanding the difficulty
of training deep feedforward neural networks. In Proc.
of the Int. Conf. on Artificial Intelligence and Statistics
(AISTATS), pp. 249–256, 2010.
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. Gradientbased learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
Goodfellow, I., Mizra, M., Da, X., Courville, A., and Bengio, Y. An empirical investigation of catastrophic forgetting in gradient-based neural networks. In Proc. of the
Int. Conf. on Learning Representations (ICLR), 2014.
Gutsein, S. and Stump, E. Reduction of catastrophic forgetting with transfer learning and ternary output codes. In
Proc. of the Int. Joint Conf. on Neural Networks (IJCNN),
pp. 1–8, 2015.
Han, S., Mao, H., and Dally, W. J. Deep compression:
compressing deep neural networks with pruning, trained
quantization and Huffman coding. In Proc. of the Int.
Conf. on Learning Representations (ICLR), 2016.
He, X. and Jaeger, H. Overcoming catastrophic interference
using conceptor-aided backpropagation. In Proc. of the
Int. Conf. on Learning Representations (ICLR), 2018.
Jung, H., Ju, J., Jung, M., and Kim, J. Less-forgetting
learning in deep neural networks. ArXiv, 1607.00122,
2016.
Lee, S.-W., Kim, J.-H., Jun, J., Ha, J.-W., and Zhang, B.-T.
Overcoming catastrophic forgetting by incremental moment matching. In Guyon, I., Luxburg, U. V., Bengio, S.,
Wallach, H., Fergus, R., Vishwanathan, S., and Garnett,
R. (eds.), Advances in Neural Information Processing
Systems (NIPS), volume 30, pp. 4655–4665. Curran Associates Inc., 2017.
Legg, S. and Hutter, M. Universal intelligence: a definition
of machine intelligence. Minds and Machines, 17(4):
391–444, 2007.
Li, Z. and Hoiem, D. Learning without forgetting. IEEE
Trans. on Pattern Analysis and Machine Intelligence, PP
(99):1–1, 2017.
Lopez-Paz, D. and Ranzato, M. A. Gradient episodic memory for continuum learning. In Guyon, I., Luxburg, U. V.,
Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S.,
and Garnett, R. (eds.), Advances in Neural Information
Processing Systems (NIPS), volume 30, pp. 6449–6458.
Curran Associates Inc., 2017.
Overcoming Catastrophic Forgetting with Hard Attention to the Task
Mallya, A. and Lazebnik, S. PackNet: adding multiple
tasks to a single network by iterative pruning. ArXiv,
1711.05769, 2017.
McCloskey, M. and Cohen, N. Catastrophic interference
in connectionist networks: the sequential learning problem. Psychology of Learning and Motivation, 24:109–
165, 1989.
McCulloch, W. S. and Pitts, W. A logical calculus of the
ideas immanent in nervous activity. The Bulletin of Mathematical Biophysics, 5(4):115–133, 1943.
Netzer, Y., Wang, T., Coates, A., Bissacco, A., Wu, B.,
and Ng, A. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep
Learning and Unsupervised Feature Learning (NIPSDeepLearning), 2011.
Ng, H.-W. and Winkler, S. A data-driven approach to cleaning large face datasets. In Proc. of the IEEE Int. Conf. on
Image Processing (ICIP), pp. 343–347, 2014.
Nguyen, C., Li, Y., Bui, T. D., and Turner, R. E. Variational
continual learning. ArXiv, 1710.10628, 2017.
Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E.,
DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., and Lerer,
A. Automatic differentiation in PyTorch. In NIPS Workshop on The Future of Gradient-based Machine Learning
Software & Techniques (NIPS-Autodiff), 2017.
Ratcliff, R. Connectionist models of recognition memory:
constraints imposed by learning and forgetting functions.
Psychological Review, 97:285–308, 1990.
Rebuffi, S., Kolesnikov, A., Sperl, G., and Lampert, C.
iCaRL: incremental classifier and representation learning. In Proc. of the IEEE Conf. on Computer Vision and
Pattern Recognition (CVPR), pp. 2001–2010, 2017.
Shin, H., Lee, J. K., Kim, J., and Kim, J. Continual learning
with deep generative replay. In Guyon, I., Luxburg, U. V.,
Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S.,
and Garnett, R. (eds.), Advances in Neural Information
Processing Systems (NIPS), volume 30, pp. 2993–3002.
Curran Associates Inc., 2017.
Sprechmann, P., Jayakumar, S., Rae, J., Pritzel, A., Puigdomènech, A., Uria, B., Vinyals, O., Hassabis, D., Pascanu, R., and Blundell, C. Memory-based parameter
adaptation. In Proc. of the Int. Conf. on Learning Representations (ICLR), 2018.
Srivastava, R. K., Masci, J., Kazerounian, S., Gomez, F.,
and Schmidhuber, J. Compete to compute. In Burges,
C. J. C., Bottou, L., Welling, M., Ghahramani, Z., and
Weinberger, K. (eds.), Advances in Neural Information
Processing Systems (NIPS), volume 26, pp. 2310–2318.
Curran Associates Inc., 2013.
Stallkamp, J., Schlipsing, M., Salmen, J., and Igel, C. The
German traffic sign recognition benchmark: a multi-class
classification competition. In Proc. of the Int. Joint Conf.
on Neural Networks (IJCNN), pp. 1453–1460, 2011.
Thrun, S. and Mitchell, T. Lifelong robot learning. Robotics
and Autonomous Systems, 15:25–46, 1995.
Venkatesan, R., Venkateswara, H., Panchanathan, S., and Li,
B. A strategy for an uncompromising incremental learner.
ArXiv, 1705.00744, 2017.
Wang, H. and Raj, B. On the origin of deep learning. ArXiv,
1702.07800, 2017.
Xiao, H., Rasul, K., and Vollgraf, R. Fashion-MNIST: a
novel image dataset for benchmarking machine learning
algorithms. ArXiv, 1708.07747, 2017.
Robins, A. Catastrophic forgetting, rehearsal and pseudorehearsal. Connection Science, 7:123–146, 1995.
Yoon, J., Yang, E., Lee, J., and Hwang, S. J. Lifelong
learning with dynamically expandable networks. In Proc.
of the Int. Conf. on Learning Representations (ICLR),
2018.
Rusu, A. A., Rabinowitz, N. C., Desjardins, G., Soyer, H.,
Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., and Hadsell, R. Progressive neural networks. ArXiv, 1606.04671,
2016.
Zenke, F., Poole, B., and Ganguli, S. Improved multitask
learning through synaptic intelligence. In Proc. of the
Int. Conf. on Machine Learning (ICML), pp. 3987–3995,
2017.
Overcoming Catastrophic Forgetting with Hard Attention to the Task
APPENDIX
A. Data
The data sets used in our experiments are summarized in Table 2. The MNIST data set (LeCun et al., 1998) comprises
28 × 28 monochromatic images of handwritten digits. Fashion-MNIST (Xiao et al., 2017) comprises gray-scale images of
the same size from Zalando’s articles4 . The German traffic sign data set (TrafficSigns; Stallkamp et al., 2011) contains traffic
sign images. We used the version of the data set from the Udacity self-driving car github repository5 . The NotMNIST data
set (Bulatov, 2011) comprises glyphs extracted from publicly available fonts, making a similar data set to MNIST; we just
need to resize the images6 . The SVHN data set (Netzer et al., 2011) comprises digits cropped from house numbers in Google
Street View images. The FaceScrub data set (Ng & Winkler, 2014) is widely used in face recognition tasks (KemelmacherShlizerman et al., 2016). Because some of the images listed in the original data set were not hosted anymore on the
corresponding Internet domains, we use a version of the data set stored on the MegaFace challenge website7 (KemelmacherShlizerman et al., 2016), from which we select the first 100 people with the most appearances8 . The CIFAR10 and CIFAR100
data sets contain 32 × 32 color images (Krizhevsky, 2009).
To match the image input shape required in our experiments, some of the images in the corresponding data sets need to be
resized (FaceScrub, TrafficSigns, and NotMNIST) or padded with zeros (MNIST and FashionMNIST). In addition, for the
data sets comprising monochromatic images, we replicate the image across all RGB channels. Note that we do not perform
any sort of data augmentation; we just adapt the inputs. We provide the necessary code to perform such adaptations in the
links listed above.
Table 2. Data sets used in the study: name, reference, number of classes, and number of train and test instances.
DATA SET
CIFAR10 (K RIZHEVSKY , 2009)
CIFAR100 (K RIZHEVSKY , 2009)
FACE S CRUB (N G & W INKLER , 2014)
FASHION MNIST (X IAO ET AL ., 2017)
N OT MNIST (B ULATOV , 2011)
MNIST (L E C UN ET AL ., 1998)
SVHN (N ETZER ET AL ., 2011)
T RAFFIC S IGNS (S TALLKAMP ET AL ., 2011)
C LASSES
T RAIN
T EST
10
100
100
10
10
10
100
43
50,000
50,000
20,600
60,000
16,853
60,000
73,257
39,209
10,000
10,000
2,289
10,000
1,873
10,000
26,032
12,630
B. Raw Results
B.1. Task Mixture
We report all forgetting ratios ρ≤t for t = 1 to 8 in Table 3. A total of 10 runs with 10 different seeds are performed and the
averages and standard deviations are taken.
B.2. Layer Use
In Fig. 8 we show an example of layer capacity monitoring as the sequence of tasks evolves. As mentioned in the main
paper, we can compute a percent of active weights for a given layer and task.
B.3. Network Compression
The final results of the network compression experiment reported in the main paper (after reaching convergence) are available
in Table 4. We run HAT on isolated tasks with c = 1.5 and uniform embedding initialization U(0, 2).
4
https://github.com/zalandoresearch/fashion-mnist
https://github.com/georgesung/traffic_sign_classification_german
6
Code is attached in a ZIP file.
7
http://megaface.cs.washington.edu/participate/challenge.html
8
Code is attached in a ZIP file.
5
Overcoming Catastrophic Forgetting with Hard Attention to the Task
Table 3. Average forgetting ratio ρ≤t for the considered approaches (10 runs, standard deviation into parenthesis).
A PPROACH
LFL
LWF
SGD
IMM-M ODE
SGD-F
IMM-M EAN
EWC
PATH N ET
PNN
HAT
ρ≤1
ρ≤2
ρ≤3
ρ≤4
ρ≤5
ρ≤6
ρ≤7
ρ≤8
-0.00 (0.01)
-0.00 (0.01)
-0.00 (0.00)
-0.00 (0.01)
-0.00 (0.00)
-0.00 (0.00)
-0.00 (0.00)
-0.02 (0.03)
-0.10 (0.12)
-0.01 (0.02)
-0.73 (0.29)
-0.14 (0.13)
-0.20 (0.08)
-0.11 (0.08)
-0.20 (0.15)
-0.12 (0.10)
-0.08 (0.06)
-0.09 (0.16)
-0.11 (0.10)
-0.02 (0.03)
-0.88 (0.18)
-0.38 (0.17)
-0.41 (0.09)
-0.27 (0.12)
-0.30 (0.15)
-0.24 (0.11)
-0.15 (0.11)
-0.11 (0.19)
-0.13 (0.09)
-0.03 (0.03)
-0.89 (0.13)
-0.63 (0.11)
-0.49 (0.07)
-0.37 (0.10)
-0.38 (0.11)
-0.32 (0.06)
-0.18 (0.07)
-0.12 (0.21)
-0.14 (0.04)
-0.03 (0.02)
-0.91 (0.11)
-0.68 (0.08)
-0.54 (0.07)
-0.39 (0.07)
-0.42 (0.09)
-0.37 (0.06)
-0.21 (0.07)
-0.14 (0.22)
-0.13 (0.03)
-0.04 (0.02)
-0.90 (0.09)
-0.70 (0.03)
-0.57 (0.06)
-0.45 (0.05)
-0.44 (0.08)
-0.40 (0.06)
-0.23 (0.04)
-0.15 (0.23)
-0.13 (0.02)
-0.05 (0.02)
-0.92 (0.08)
-0.76 (0.06)
-0.62 (0.06)
-0.49 (0.06)
-0.45 (0.07)
-0.42 (0.07)
-0.25 (0.05)
-0.17 (0.23)
-0.12 (0.01)
-0.06 (0.02)
-0.92 (0.08)
-0.80 (0.06)
-0.66 (0.03)
-0.49 (0.05)
-0.44 (0.06)
-0.42 (0.04)
-0.25 (0.03)
-0.17 (0.23)
-0.11 (0.01)
-0.06 (0.01)
Used weights [%]
100
80
60
40
20
0
Conv1 Conv2 Conv3 FC1
Layer
Cumulative
FaceScrub
MNIST
CIFAR100
NotMNIST
SVHN
FC2
FC3
CIFAR10
TrafficSigns
FashionMNIST
Figure 8. Layer-wise weight usage with sequential task learning, including (lines) and excluding (bars) the cumulative attention of past
tasks. Task sequence corresponds to seed 0.
Table 4. Results for the compression experiment reported in the main paper: test accuracy A1 with SGD, test accuracy A1 after
compressing with HAT, and percentage of network weights used after compression.
DATA SET
CIFAR10
CIFAR100
FACE S CRUB
FASHION MNIST
MNIST
N OT MNIST
SVHN
T RAFFIC S IGNS
R AW A1
C OMPRESSED A1
79.9%
52.7%
82.7%
92.4%
99.5%
90.9%
94.2%
97.5%
80.8%
49.1%
82.3%
91.9%
99.4%
91.5%
93.8%
98.1%
S IZE
13.9%
21.4%
21.0%
2.3%
1.2%
5.7%
3.1%
2.9%
B.4. Training Time
To have an idea of the training time for each of the considered approaches, we report some reference values in Table 5. We
see that HAT is also quite competitive in this aspect.
Overcoming Catastrophic Forgetting with Hard Attention to the Task
Table 5. Wall-clock training time measured on a single NVIDIA Pascal Titan X GPU: total (after learning the 8 tasks), per epoch, and
per batch (batches of 64). Batch processing time is measured for a forward pass (Batch-F), and for both a forward and a backward pass
(Batch-FB).
A PPROACH
T OTAL [ H ]
PNN
PATH N ET
EWC
M ULTITASK
IMM-M EAN
IMM-M ODE
LWF
HAT
SGD
LFL
SGD-F
T RAINING TIME
E POCH [ S ] BATCH -F [ MS ]
6.0
4.5
3.9
3.4
3.2
3.1
2.2
2.2
1.4
1.3
0.5
4.1
3.6
3.1
94.8
2.6
2.5
2.2
1.6
0.9
0.9
0.9
BATCH -FB [ MS ]
10.2
10.6
7.9
3.1
6.9
6.7
5.7
4.0
2.5
4.4
2.5
27.5
23.9
19.7
15.7
17.1
16.0
14.2
11.7
6.6
9.2
6.8
C. Additional Results
C.1. Incremental CIFAR
As an additional experiment to complement our evaluation, we consider the incremental CIFAR setup, following a similar
approach as Lopez-Paz & Ranzato (2017). We divide both CIFAR10 and CIFAR100 data sets into consecutive-class subsets
and use them as tasks, presented in random order according to the seed. We take groups of 2 classes for CIFAR10 and
20 classes for CIFAR100, yielding a total of 10 tasks. We decide to take groups of 2 and 20 classes in order to have a similar
number of training instances per task. The rest of the procedure is as in the main paper. The most important results are
summarized there. The complete numbers are depicted in Fig. 9 and reported in Table 6.
0.0
−0.1
ρ ≤t
−0.2
−0.3
−0.4
−0.5
−0.6
1
2
Multitask
SGD
SGD-F
3
4
5
EWC
IMM-Mode
IMM-Mean
t
6
7
8
LWF
LFL
PathNet
9
10
PNN
HAT
Figure 9. Average forgetting ratio ρ≤t for the incremental CIFAR task (average after 10 runs).
C.2. Permuted MNIST
A common experiment is the one proposed by Srivastava et al. (2013), and later employed to evaluate catastrophic forgetting
by Goodfellow et al. (2014). It consists of taking random permutations of the pixels in the MNIST data set as tasks. Typically,
the average accuracy after sequentially training on 10 MNIST permutations is reported. To match the different number of
Overcoming Catastrophic Forgetting with Hard Attention to the Task
Table 6. Average forgetting ratio ρ≤t for the incremental CIFAR task (10 runs, standard deviation into parenthesis).
A PPROACH
LFL
LWF
SGD-F
PATH N ET
SGD
IMM-M EAN
IMM-M ODE
PNN
EWC
HAT
ρ≤1
ρ≤2
ρ≤3
ρ≤4
ρ≤5
ρ≤6
ρ≤7
ρ≤8
ρ≤9
ρ≤10
-0.00 (0.01)
-0.00 (0.02)
-0.00 (0.01)
-0.15 (0.31)
-0.00 (0.01)
-0.00 (0.02)
-0.00 (0.01)
-0.26 (0.16)
-0.00 (0.01)
-0.03 (0.04)
-0.53 (0.31)
-0.10 (0.03)
-0.25 (0.14)
-0.18 (0.20)
-0.19 (0.09)
-0.14 (0.08)
-0.14 (0.10)
-0.26 (0.08)
-0.13 (0.09)
-0.05 (0.02)
-0.63 (0.25)
-0.27 (0.05)
-0.33 (0.16)
-0.21 (0.26)
-0.27 (0.09)
-0.21 (0.10)
-0.21 (0.11)
-0.25 (0.05)
-0.15 (0.08)
-0.05 (0.02)
-0.67 (0.21)
-0.42 (0.05)
-0.35 (0.18)
-0.22 (0.28)
-0.30 (0.04)
-0.22 (0.10)
-0.23 (0.06)
-0.23 (0.04)
-0.16 (0.07)
-0.06 (0.01)
-0.70 (0.20)
-0.50 (0.06)
-0.37 (0.16)
-0.24 (0.29)
-0.30 (0.06)
-0.25 (0.10)
-0.25 (0.09)
-0.22 (0.03)
-0.17 (0.06)
-0.06 (0.01)
-0.74 (0.17)
-0.53 (0.04)
-0.40 (0.18)
-0.27 (0.29)
-0.28 (0.04)
-0.26 (0.08)
-0.23 (0.07)
-0.23 (0.03)
-0.18 (0.06)
-0.07 (0.01)
-0.77 (0.15)
-0.59 (0.06)
-0.41 (0.18)
-0.28 (0.30)
-0.31 (0.03)
-0.27 (0.08)
-0.26 (0.05)
-0.22 (0.03)
-0.19 (0.08)
-0.07 (0.01)
-0.79 (0.14)
-0.64 (0.06)
-0.41 (0.19)
-0.30 (0.30)
-0.32 (0.04)
-0.28 (0.08)
-0.27 (0.04)
-0.21 (0.02)
-0.18 (0.07)
-0.08 (0.01)
-0.79 (0.14)
-0.68 (0.05)
-0.42 (0.19)
-0.32 (0.29)
-0.30 (0.05)
-0.29 (0.07)
-0.25 (0.04)
-0.21 (0.02)
-0.18 (0.06)
-0.08 (0.01)
-0.78 (0.14)
-0.70 (0.05)
-0.43 (0.20)
-0.35 (0.28)
-0.30 (0.04)
-0.30 (0.07)
-0.25 (0.04)
-0.21 (0.02)
-0.18 (0.06)
-0.09 (0.01)
Table 7. Accuracy on the permuted MNIST task (Srivastava et al., 2013), taking the average after training 10 tasks. The only exception is
the generative replay approach, whose performance was assessed after 5 tasks. Superscripts indicate results reported by (1) Nguyen et al.
(2017) and (2) He & Jaeger (2018). An asterisk after parameter count indicates that the approach presents some additional structure not
included in such parameter count (for instance, some memory module or an additional generative network).
A PPROACH
GEM (L OPEZ -PAZ & R ANZATO , 2017)
SI (Z ENKE ET AL ., 2017)1
EWC (K IRKPATRICK ET AL ., 2017)2
M B PA + EWC – 1000 EX . (S PRECHMANN ET AL ., 2018)
VCL (N GUYEN ET AL ., 2017)
HAT – S MALL
G ENERATIVE R EPLAY (S HIN ET AL ., 2017)
CAB (H E & JAEGER , 2018)
EWC (K IRKPATRICK ET AL ., 2017)
SI (Z ENKE ET AL ., 2017)
HAT – M EDIUM
HAT – L ARGE
PARAMETERS
A≤10
0.1 M*
0.1 M
0.1 M
82.8%
86.0%
88.2%
89.7%
90.0%
91.6%
94.9%
95.2%
96.9%
97.1%
97.4%
98.6%
U NKNOWN *
0.1 M*
0.1 M
U NKNOWN *
0.7 M
5.8 M
5.8 M
0.7 M
5.8 M
parameters used in the literature, we consider a small, medium, and a large network based on a two-layer fully-connected
architecture as Zenke et al. (2017), with 100, 500, and 2000 hidden units, respectively. For the large network we set dropout
probabilities as Kirkpatrick et al. (2017). We use smax = 200 and c = 0.5 for the small network, and smax = 400 and
c = 0.5 for the medium and large networks. The results are available in Table 7.
C.3. Split MNIST
Another popular experiment is to split the MNIST data set into tasks and report the average accuracy after learning them
one after the other. We follow Lee et al. (2017) by splitting the data set using labels 0–4 and 5–9 as tasks and running the
experiment 10 times. We also match the base network architecture to the one used by Lee et al. (2017). We train HAT for
50 epochs with c = 0.1. Results are reported in Table 8. In preliminary experiments we observed that dropout could increase
accuracy by some percentage. However, to keep the same configuration as in the cited reference, we finally did not use it.
Table 8. Average accuracy on the split MNIST task, following the setup of Lee et al. (2017) using 10 runs (standard deviation into
parenthesis). Superscript (1) indicates results reported by Lee et al. (2017).
A PPROACH
SGD (G OODFELLOW ET AL ., 2014)1
L2-T RANSFER (E VGENIOU & P ONTIL , 2004)1
IMM-M EAN (L EE ET AL ., 2017)
IMM-M ODE (L EE ET AL ., 2017)
CAB (H E & JAEGER , 2018)
HAT
PARAMETERS
A≤2
1.9 M
1.9 M
1.9 M
1.9 M
1.9 M
1.9 M
71.3% (1.5)
85.8% (0.5)
94.0% (0.2)
94.1% (0.3)
94.9% (0.3)
99.0% (0.0)
Overcoming Catastrophic Forgetting with Hard Attention to the Task
D. Variations to the Proposed Approach
In this section, we want to mention a number of alternatives we experimented with during the development of HAT. The
purpose of the section is not the report a formal set of results, but to inform the reader about potential different choices when
implementing HAT, or variations of it, and to give an intuition on the outcome of some of such choices.
D.1. Embedding Learning
When we realized that the embedding weights etl were not changing much and that their gradients were small compared
to the rest of the network due to the introduced annealing of s, we initially tackled the issue by using a different learning
rate for the embeddings. With that, we empirically found that factors of 10–50 times the original learning rate were able to
tackle the issue, leading to performances that were almost as good as the final ones reported in the main paper. However,
the use of a different learning rate introduced an additional parameter that we could not conceptually relate to catastrophic
forgetting and that could have been tricky to tune for a generic setting.
We also studied the use of an adaptive optimizer such as Adagrad (Duchi et al., 2011) or Adam (Kingma & Ba, 2015) for the
embedding weights. The idea was that an adaptive optimizer would be able to automatically introduce an appropriate scaling
factor. We found that this option was effectively learning suitable values for etl . However, its performance was worse than
the constant-factor SGD boost explained above. Noticeably, introducing an adaptive optimizer also introduces a number of
new hyperparameters: type of optimizer, another learning rate, weight decays, etc.
D.2. Annealing
In our effort to further reduce the number of hyperparameters, we experimented for quite some time with the annealing
b−1
π
1+
s = tan
4
B−1
or using variants of
s = α + β tan
π b−1
2 B−1
.
The rationale for the first expression is that one starts with a sigmoid σ(sx) that is equivalent to a straight line of 45 degrees
for b = 1 and x ≈ 0. Then, with b increasing, it linearly increases the angle towards 90 degrees at x = 0. The second
expression is a parametric evolution of the first one.
These annealing schedules have the (sometimes desirable) feature that the maximum s is infinite, yielding a true step
function in inference time. Therefore, we obtain truly binary attention vectors atl and no forgetting. In addition, if we use the
first expression, we are able to remove the smax hyperparameter. Nonetheless, we found the first expression to perform
worse than the solution proposed in the main paper. The introduction of the second expression with α = 1 and β < 1
improved the situation, but results were still not as good as the ones in the main paper and the tuning of β was a bit tricky.
To conclude this subsection, note that if smax is large, for instance smax > 100, one can use
s = smax
b−1
,
B−1
which is a much simpler annealing formula that closely approximates the one in the main paper. However, one needs then to
be careful with the denominator of the embedding gradient compensation when s = 0.
D.3. Gate
We also studied the use of alternatives to the sigmoid gate. Apart from the rescaled tanh, an interesting alternative we
thought of was a clamped version of the linear function,
setl
1
t
al = max 0, min 1,
+
,
r
2
where r defines the ‘valid’ range for the input of the gate. This gate yields a much simpler formulation for the gradient
compensation described in the main paper. However, it implies that we need to set r, which could be considered a
Overcoming Catastrophic Forgetting with Hard Attention to the Task
further hyperparameter. It also implies that embedding values that are far away from 0, the step transition point, receive a
proportionally similar gradient to the ones that are close to it. That is, values of etl that yield atl that are very close to 0 or 1
(in the constant region of the pseudo-step function) are treated equal to the ones that are still undecided (in the transition
region of the pseudo-step function). We did not test this alternative gate quantitatively.
D.4. Cumulative Attention
In the most preliminary stages we used
a≤t
l =1−
h
1 − atl
1 − al≤t−1
i
for accumulating attention across tasks, but it was soon dismissed for the final max-based formula. The previous equation
could be interesting for online learning scenarios with limited model capacity, together with
≤t−1
t
a≤t
,
l = max al , κ al
where κ is a constant slightly lower than 1 (for instance κ = 0.9 or κ = 0.99).
D.5. Embedding Initialization
We ran a set of experiments using uniform initialization U(0, k1 ) for the embeddings etl instead of Gaussian N (0, 1). We
also experimented with N (k2 , 1). The idea behind these alternative initializations was that, for sufficiently large smax , all
or almost all atl start with a value of 1, which has the effect of distributing the attention over all units for more time at the
beginning of training. Using values of k1 ∈ [1, 6] and k2 ∈ [0.5, 2] yielded competitive results, yet worse than the ones
using N (0, 1). Our intuition is that a uniform initialization like U(0, 2) is better for a purely compressive approach, as used
in the last experiment of the main paper.
D.6. Attention Regularization
We initially experimented with a normalized L1 regularization
PL−1 PNl t
l=1
i=1 al,i
t
R A =
.
PL−1
l=1 Nl
Results were a small percentage lower than the ones with the attention-weighted regularization of the main paper. We also
exchanged the previous L1 regularization with the L2-based regularization
PL−1 PNl t 2
l=1
i=1 (al,i )
t
R A =
.
PL−1
l=1 Nl
With that, we observed similar accuracies as the L1 regularization, but under different values for the hyperparameter c.
D.7. Hard Attention to the Input
As mentioned in the main paper, no attention mask is used for the input (that is, there is no at0 ). We find this is a good
strategy for a general image classification problem and for first-layer convolutional filters in particular. However, if the input
consists of independent, isolated features, one may think of putting hard attention to the input as a kind of supervised feature
selection process. We performed a number of experiments using only fully-connected layers and the MNIST data as above,
and introduced additional hard attention vectors at0 that directly multiplied the input of the network. The results suggested
that it could potentially be a viable option for feature selection and data compression (Fig. 10).
E. A Note on Binary Masks
After writing a first version of the paper, we realized that the idea of a binary mask that affects a given unit could be
potentially traced back to the “inhibitory synapses” of McCulloch & Pitts (1943). This idea of inhibitory synapses is quite
unconventional and rarely seen today (Wang & Raj, 2017) and, to the best of our knowledge, no specific way for learning
Overcoming Catastrophic Forgetting with Hard Attention to the Task
Figure 10. Example of an input mask for MNIST data after training to convergence.
such inputs nor a specific function for them have been proposed. Weight-based binary masks are implicitly or explicitly
used by many catastrophic forgetting approaches, at least by Rusu et al. (2016); Fernando et al. (2017); Mallya & Lazebnik
(2017); Nguyen et al. (2017); Yoon et al. (2018). HAT is a bit different, as it learns unit-based hard attention vectors, with
possible (but not necessarily) binary values.
| 2cs.AI
|
Modeling and Control of Inverted Flight of a Variable-Pitch Quadrotor
Namrata Gupta, Mangal Kothari∗, and Abhishek†
Abstract
arXiv:1709.06407v1 [cs.SY] 19 Sep 2017
This paper carries out the mathematical modeling, simulation, and control law design for a quadrotor with
variable-pitch propellers. The use of variable-pitch propeller for thrust variation instead of RPM regulation facilitates generation of negative thrust, thereby augmenting the rate of change of thrust generation amenable for
aggressive maneuvering. Blade element theory along with momentum theory is used to estimate propeller thrust
and torque essential for formulating equation of motion of the vehicle. The proposed flight dynamics model is used
for non-linear control design using dynamic inversion technique, which is then used to stabilize, track reference
trajectory, and simulate flip maneuver. The rotor torque is an irrational function of the control input which makes
the control design challenging. To address this problem, the control design employs three loops. The outer loop
solves the translational dynamics to generate the thrust, pitch angle, and roll angle commands required to track
the prescribed trajectory. Using the command generated in the outer loop, the inner loop simplifies the rotational
dynamics to provide the desired rate of angular velocities. A control allocation loop is added to address the problem
of nonlinearity associated with rotor torque. This is done by introducing the derivative of thrust coefficient as a
virtual control input. These virtual inputs determine the derivatives of thrust and body moments, which in turn is
used to generate the required thrust and body moments. The concept is validated by showing attitude stabilization
in real flight for a variable pitch quadrotor. The performance of the proposed design is shown through simulated
results for attitude stabilization and trajectory following. Reverse thrust capability of variable-pitch quadrotor is
also shown by performing flip maneuver in which quadrotor roll angle changes from 0 to 180 degrees.
1
Introduction
Last decade saw the development of various configurations of Unmanned Aerial Vehicles (UAVs) capable of hovering
flight. UAV configurations ranging from flapping wing, rotary wing to cycloidal rotor concept have been developed
and studied in recent years (see Refs. [1, 2, 3, 4, 5, 6, 7]). Some of these designs have seen greater success than others.
Researchers around the world have been working on different rotary wing configurations such as the Micro Coaxial
∗ Department
† Department
of Aerospace Engineering, Indian Institute of Technology, Kanpur, Uttar Pradesh, India, email: [email protected]
of Aerospace Engineering, Indian Institute of Technology, Kanpur, Uttar Pradesh, India, email: [email protected]
1
Rotorcraft (MICOR) [3], muFly [4], RoboFly or the Samara Micro Air Vehicle (MAV) – a prop assisted mono blade [5],
Coanda UAV [6], and Cycloidal rotor MAV [7]. Among these the most successful configuration which caught the eyes
of researchers and amateurs alike in the early 2000s is the quadrotor configuration. Since then, the quadrotors have
been extensively studied and several papers have been authored studying their dynamics, stabilization, and control.
Some of the notable pioneering works include that by Bouabdallah et al. [8, 9] on ETH Zürich’s ‘OS4’ (a belt-driven
indoor quadrotor vehicle) and Castillo et al. [10]. The commercial success for this configuration can be gauged from
the fact that quadrotors with all up weights ranging from 50 grams to 15–20 kg can be bought off-the-shelf and can
be used for a variety of missions.
The conventional quadrotor with fixed pitch propellers is controlled by varying the RPM of the individual motors
and suffers from a few limitations: i) the rotational inertia of the motors limits the control bandwidth of the system [11];
and ii) the stabilization of larger quadrotors through RPM control alone becomes challenging as a point can be reached
where the torque required to change the RPM of the motor exceeds the capacity of the motor. Due to these limitations,
the current flight control strategy of quadrotors is not suitable for larger full scale vehicles meant for lifting heavy
payload.
These limitations can be overcome by employing a quadrotor design with variable-pitch control [12]. It would
appear that the use of variable-pitch propellers add complexity to a simple and robust quadrotor design. But, the
advantages of increased controller bandwidth due to the availability of reverse thrust from propellers and scalability to
full scale size justify the design. The idea of variable-pitch propeller based quadrotor is an old one. It has somehow not
attracted enough attention from researchers, until recently. In 1922, Georges de Bothezat and Ivan Jerome built and
flew the “Flying Octopus” a quadrotor with rotors located at each end of a truss structure of intersecting beams, placed
in the shape of a cross. Control of the machine was achieved by changing the pitch of each of the propellers [13]. The
Hoverbot, developed at University of Michigan by Johann Borenstein [14] is the first documented effort at designing
and flying a small scale quadrotor with variable pitch control. However, the Hoverbot never achieved flight beyond
tethered hovering.
In recent past, while, several hobbyists have demonstrated the construction and flight of remote controlled variablepitch quadrotors, a serious and organized effort of studying the flight mechanics and control of a variable-pitch
quadrotor was demonstrated by Cutler et al. [11, 15]. To keep the design simple, the four motor based design of fixed
pitch quadrotor was retained and a mechanism similar to tail rotor swashplate was used to change the blade pitch angle
for each of the propellers. The flight performance of variable-pitch quadrotor with fixed pitch ones was systematically
studied and the following conclusions were made: (i) the variable-pitch propeller quadrotor could generate significantly
large rate of change of thrust when compared to a fixed-pitch design, thereby improving the capability to perform
2
aggressive maneuvers; and (ii) the ability to generate negative thrust by variable-pitch propellers can be utilized to
perform aerobatics and allow for inverted flight. However, the dynamics of variable-pitch quadrotor UAV has not been
studied in detail and therefore model based control has not been applied for control and navigation of such quadrotors
until recently. It should be noted that the thrust for variable pitch propellers not only depend on RPM but also on
the blade pitch angle. Unlike RPM controlled quadrotors, designing a controller for these vehicles is challenging as
the function representing the relation between thrust and torque with blade pitch angle is not rational. Therefore,
the control design methods available for RPM controlled quadrotors cannot be applied as it is for variable-pitch
quadrotors. Previously, approaches like backstepping [16, 17], sliding mode [18], nonlinear dynamic inversion (NDI)
[19], [20], adaptive control [21, 22, 23, 24], have been applied to design controller for quadrotors. A comparison study
was carried out in [25] using visual feedback for stabilization and tracking.
Cutler et al. [26] carried out the analysis and control design for the variable-pitch quadrotor. However, the analytical
model developed for the variable-pitch quadrotor in this paper suffers from some serious errors. First, in the paper the
angle of attack used for computing the blade lift and drag forces has erroneously been replaced with geometric blade
pitch angle ignoring the contribution of the blade induced velocity (inflow) component which has significant effect on
angle of attack. The inflow angle is a function of thrust being generated by the rotor which in turn is a function of blade
pitch angle itself. Therefore, ignoring the induced inflow angle would result in significant overestimation of the lift
and drag forces. Further, this mistake simplifies the control design. Second, the paper assumes that the multiplication
of a constant “drag coefficient” with thrust force would result in yawing moment. This is only valid for a fixed-pitch
propeller which has constant thrust and torque coefficients. The torque responsible for yawing motion of a variable
pitch quadrotor is a function of blade pitch angle itself and therefore cannot be obtained by merely multiplying the
thrust with a constant factor for entire range of pitch angles. While the theoretical model was incorrect, authors
in [26] were able to demonstrate the flight of variable pitch quadrotor using PID control design. The flight dynamics
model of variable-pitch quadrotor aerial vehicle was first proposed in [27] in which the controller was used to perform
attitude stabilization and trajectory tracking. The detailed control design and ability to perform flip maneuver has
not been studied for a variable-pitch quadrotor using a model based controller and this is the focus of the present
paper.
The present paper focuses on development of flight dynamics model based on Blade Element Theory (BET) and
uniform inflow for variable pitch quadrotor is presented. Next, the control design based on dynamic inversion technique
is developed for aggressive maneuvering of variable-pitch quadrotors. The challenge associated with the control
allocation in variable-pitch quadrotors is addressed by use of an additional loop that dynamically allocates control to
generate the desired thrust and moments. Finally, the nonlinear controller is used to simulate the stabilization, flip
3
and upright/inverted trajectory following of the variable pitch quadrotor. The contributions of this paper are two
folds: (i) establish the detailed dynamics model of variable-pitch quadrotor aerial vehicle; and (ii) develop and apply
a nonlinear controller for stabilization, trajectory tracking and inverted flight. In addition to this, the new design is
validated by showing attitude stabilization in real flight.
2
Quadrotor Modeling
The strategy for controlling a variable-pitch quadrotor is significantly different from that of conventional fixed-pitch
propeller based quadrotor and is discussed in this section. After establishing the control strategy, the six degrees of
freedom (six-DOF) Newton-Euler equations representing the dynamics of variable-pitch quadrotor vehicle are derived.
2.1
Strategy for Control
Like the conventional quadrotor, the primary control of various motions (three translational, roll, and pitch motions)
for variable-pitch quadrotor is achieved by changing the thrust of different rotors in various combinations. However,
the mechanism of thrust variation is different. The change in thrust is achieved by simultaneously changing the pitch
angle (collective angle) of all the blades. The control of yawing motion and the mechanism involved is significantly
different as discussed below. It should be noted that any point of time all the rotors are operated at the same nominal
RPM which may be regulated about the specified value for setting the baseline value of thrust.
Figure 1: Translational flight and roll motion control of variable pitch quadrotor
The up/down motion is easily controlled by collectively increasing or decreasing the collective pitch angles for all the
4
rotors / propellers simultaneously. Side-wards flight can be achieved as explained in Fig. 1. For example, increasing
the collective input and thereby increasing the thrust of the two left rotors lifts the left side up and generates a net
thrust component to the right. Consequently, the quadrotor would move to the right. The change in torque/power of
the two left rotors is equal and opposite, therefore, the moment remains balanced and pure translational motion can
be achieved. By the same principle, increasing the collective of the two rear rotors would result in forward flight. The
yaw control is less intuitive.
Figure 2: Yaw motion control of variable pitch quadrotor
The method of generating yawing moment is identical to that used for coaxial and tandem helicopters and is
known as “differential collective”. In this, the collective pitch of the two diagonal rotors rotating in the same direction
is increased and the collective pitch of the other diagonal pair is reduced. The increased collective pitch results in
increasing the lift and drag forces experienced by both these rotors, while the other two rotors would experience an
identical reduction in lift. The rotors with increased lift and drag would experience an increase in profile and induced
torque components compared to the other two rotors which would experience a decrease in the total torque. This net
increase in the combined torque of all the rotors would result in yawing motion of the quadrotor as explained in Fig. 2.
It should be noted that this operation has no effect on translational motion, as the combined thrust of all four rotors
remain unchanged.
2.2
Kinematics
For describing the rigid body dynamics of the quadrotor two coordinate systems, as shown in Fig. 3, are employed: the
inertial and body fixed coordinates. All the physical quantities are transformed between the two coordinate systems
using the classical Euler angles (φ-roll, θ-pitch, ψ-yaw). For modeling using quaternions, refer [28]. The following
5
Figure 3: Coordinate systems used for development of equation of motion
expression relates the velocity of quadrotor in these two frames:
x
u
d
y = Ri v
b
dt
z
w
where
CθCψ
i
Rb =
CθSψ
−Sθ
SφSθCψ − CφSψ
SφSθSψ + CφCψ
SφCθ
(1)
CφSθCψ + SφSψ
CφSθSψ − SφCψ
,
CφCθ
Cβ , cos β and Sβ , sin β. [x,y,z] are the position in the inertial frame and [u,v,w] are the velocity components in
the body frame. Similarly, the following expression relates the body rates to Euler angle rates:
φ̇ 1
θ̇ = 0
0
ψ̇
SφSθ
Cθ
Cφ
Sφ
Cθ
CφSθ
Cθ
p
−Sφ q
Cφ
r
Cθ
where [p,q,r] are the angular velocity components (roll, pitch, and yaw) in the body frame.
6
(2)
2.3
Dynamics
The rigid body equation of motion of the quadrotor can be derived by applying the linear momentum and angular
momentum conservation laws. In present work, propulsive forces (thrust and torque from the motors) and the gravitational forces are assumed to be the dominant forces. The aerodynamic forces (such as lift and drag) acting on the
fuselage are neglected assuming them to be very small. Transforming the gravitational force to the body coordinate
axes, the translation dynamics of the quadrotor is given as follows:
u̇
v̇ =
ẇ
0
0
T ∗f lag
M
+ [Rbi ]T
0 rv − qw
0 + pw − ur
g
qu − pv
(3)
f lag = −sgn(cosφ)
where T is the total thrust from all the rotors, M is the mass of the quadrotor, and g represents gravitational
acceleration. Eq. (3) expresses quadrotor’s translational dynamics in the body fixed coordinate system. The variable
f lag decides the direction of thrust vector in the body coordinate system and is negative for roll angle (φ) less than
90◦ and becomes positive causing reversal of thrust direction for φ greater than 90◦ and less than or equal to 180◦.
The translational dynamics can also be be expressed in the inertial frame as:
ẍ
0
ÿ = Ri 0
b
T
z̈
−M
0
+ 0
g
(4)
Note that the translation dynamics is presented in both the body and inertial frames for its application in the control
design. It is safe to assume that the quadrotor is symmetric about x and y axes, which allows for the rotational
dynamics to be represented as:
ṗ
q̇ =
ṙ
Iyy −Izz
Ixx qr
+
Izz −Ixx
pr
Iyy
Ixx −Iyy
pq
Izz
l
Ixx
m
Iyy
n
Izz
(5)
where Ixx , Iyy , and Izz are moments of inertia about x-axis, y-axis, and, z-axis, respectively. By the virtue of symmetry,
the product of inertia terms are assumed to be zero. Here, l, m, and n are the components of the externally applied
moments known as rolling, pitching, and yawing moments, respectively. Eqs. (1)-(5) together represent the complete
equation representing the full six degrees of freedom for the quadrotor.
7
2.4
Rotor Dynamics
Unlike the conventional fixed-pitch quadrotors, the thrust from individual rotors, Ti , is varied by changing their
collective pitch input. The thrust and moment equilibrium equations for the “H” configuration (similar to “X”
configuration) of the quadrotor is derived about hover condition as shown below. Blade element theory along with
momentum theory [13] is used to calculate thrust and torque of each rotor as a function of thrust coefficient.
With the assumption of the blades being rigid, the aerodynamic forces and moments generated by each rotor can
be calculated using blade element theory in which each blade is divided in to a number of elements such that each
element is a 2D airfoil. In this, the contribution of each blade element to the total airload (lift, drag, and pitching
moment) is calculated and then integrated over the blade radius to calculate the net thrust and torque contribution
of each blade, which is then multiplied with number of blades of calculate the total thrust and torque from each rotor.
Using the approach given in [13], the non-dimensional thrust coefficient, CTi , and torque coefficient, CQi , for the
ith rotor are given by
CTi
=
CQi
=
λi
θ0i
1
σClα
−
2
3
2
Cdoi
1
λ2 Cl
λi Clα θ0i
σ
− i α +
2
3
2
4
(6)
(7)
where Clα is the lift curve slope, θ0i is the blade collective pitch angle of the ith rotor, λi is the induced inflow of the
ith rotor, Cdoi is the zero lift drag coefficient of the airfoil of the ith rotor, σ =
Nb c
πR .
Here, Nb is number of blades,
c is the chord length of the rotor, R is the rotor blade radius. These non-dimensional quantities can be converted to
corresponding dimensional parameters by using Ti = CTi ρAVtip 2 and Qi = CQi ρARVtip 2 , where ρ is the density of
air, A is the rotor disk area, Vtip = ΩR is the tip speed of rotor blade rotating with angular speed of Ω. The only
unknown parameter in Eqs. (6) and (7) is the inflow ratio λi which can be evaluated using momentum theory for the
hovering flight condition and is given by Eq. (8)
λi =
r
CTi
2
(8)
Substituting the value of λi in Eq. (6) gives
CTi
1
= σClα
2
θ0i
1
−
3
2
r
CTi
2
!
(9)
which upon rearrangement yields
θ0i
6CTi
3
=
+
σClα
2
8
r
CTi
2
(10)
Using the above pitch angle and inflow ratio in Eq. (7) gives
CQi
1
= σ
2
!
√
3
Cdoi
2CTi 2
+
σ
4
(11)
From the above definitions of thrust and torque it can be seen that
Ti = KCTi
(12)
Qi = KRCQi
(13)
where K = ρAVtip 2 . K is typically constant for the variable pitch quadrotor as the rotor speed is regulated about a
prescribed constant value. The total thrust generated by the vehicle is then given as:
T = f lag × (T1 + T2 + T3 + T4 )
(14)
T = f lag × K(CT1 + CT2 + CT3 + CT4 )
(15)
Rolling and pitching moments are obtained by cross multiplying thrust from each rotor with its respective moment
arm. Yawing moment is obtained from Eq. (11). Due to the relative sense of rotation, Rotors 1 and 3 produce torque
in positive z direction, while Rotors 2 and 4 produce torque in opposite direction. The contribution of blade drag to
total torque (shown in Eq. (11)) is independent of thrust and hence remains constant at all times and cancels out for
four rotors. The final expressions for total forces and moments acting on the quadrotor are shown in Eq. (16).
T = f lag × K(CT1 + CT2 + CT3 + CT4 )
l = d × K(CT1 − CT2 − CT3 + CT4 )
(16)
m = −f lag × Kd(CT1 + CT2 − CT3 − CT4 )
3
3
3
3
KR
n = −f lag × √ (|CT1 | 2 − |CT2 | 2 + |CT3 | 2 − |CT4 | 2 )
2
where d is the moment arm of rotors from the center of gravity.
3
Control Design
This section develops a controller for variable-pitch quadrotor for stabilization, tracking, and aggressive maneuvers
using nonlinear dynamic inversion approach [29], [20]. The variable-pitch quadrotor is an under actuated system like
9
the conventional RPM regulated quadrotor. However, the dynamics of variable-pitch quadrotor is relatively more
complex than the conventional quadrotors as the rotor thrust, roll, and pitch moment equations are linear functions
of control input whereas the roll moment is a nonlinear function of control input. Therefore, a closed form solution
to these equations is not possible and iterative online solution of the system of equations is tedious and impractical.
To address this problem, the control design incorporates three loops: outer loop, inner loop, and control allocation
loop. Note that the outer and inner loops are similar to the conventional design. This means that the outer loop is
responsible for trajectory tracking whereas the inner loop provides stability. An extra loop is added to dynamically
allocate control to determine blade pitch angles of individual rotors.
Let the state of variable-pitch quadrotor be X , [x y z φ θ ψ u v w p q r]T . For tracking and stabilization, the
output of quadrotor is chosen as Y , [x
y
z
φ θ
ψ]T . The control objective is to drive Y to some desired
output, Yd . In order to achieve this, the proposed design use a two loop structure by exploiting the time scale
separation principle. The outer loop operates on position yout = [x
roll angle, φd , and pitch angle, θd . The inner loop drives yin = [φ
Uin = [ld
md
y
θ
z]T and generates the desired thrust, Td ,
ψ]T to yind = [φd
θd
ψd ]T by generating
nd ]T . As the relation between torque and blade pitch angle is nonlinear, a control allocation loop
is included to solve the problem of nonlinearities. For this loop, the derivatives of thrust coefficients act as virtual
inputs, the value of which needs to be determined to generate the desired thrust and moments. The thrust coefficients
are computed by integrating the derivatives of thrust coefficients. The required blade pitch angle for individual rotors
is then calculated. The control allocation loop computes the required blade pitch angles to generate the desired thrust
and moments.
To differentiate whether the quadrotor is upright or inverted, while tracking the given trajectory, a command
variable σd is introduced, where σd = sgn(cosφd ). σd is negative when quadrotor is in the inverted flight, i.e, φd is
greater than 90o . Another variable f lip, which is set to zero to begin with, is used to check if the quadrotor has
achieved required φd . The variable f lip becomes 1, once desired φd is achieved.
3.0.1
Outer Loop Design
∆
The tracking error in position can be defined as e = yout − youtd , where yout and youtd are the current and desired
outputs of a quadrotor in the inertial frame. As the relative degree is two, we choose second order stable error dynamics
to synthesize the control as follows
ë + 2ζωn ė + ωn2 e = 0
10
(17)
From Eq. 4 and Eq. 17, we get
i
Rb
0
0
T
−M
+
0
0
g
=
ẍd
ÿ
d
z̈d
2ζout ωnout T
+
ẋd − ẋ
ẏ − ẏ
d
żd − ż
+
ωnout ωnout T
xd − x
y −y
d
zd − z
(18)
ζout and ωnout are 3 × 1 matrices. The required thrust and desired roll and pitch angles are given as:
Td = M f lag
p
ẍ2 + ÿ 2 + (g − z̈)2
φd = sin−1 (ux sin ψd − uy cos ψd )
θd = sin−1
(19)
ux cos ψd + uy sin ψd
cos φd
where
ux = M ẍ/Td
uy = M ÿ/Td
The quadrotor is commanded to perform flip maneuver by setting σd negative, which commands the quadrotor to
flip itself, before tracking the trajectory (this means f lip = 0). At this stage, controller only tracks the altitude and
attitude and hence first and second terms of ωnout are zero, as a result ẍ and ÿ are zero. The variables φd and θd
remain the same. After substituting ẍ and ÿ in Eq. 19, the desired thrust is given by Td = M ∗ f lag ∗ abs(z̈ − g). Once
the quadrotor is flipped (f lip = 1), it is commanded to follow the given trajectory and generate acceleration along x
and y directions, hence the first and second terms of ωnout become non-zero. The expressions for Td and θd remain
the same as in Eq. (19) whereas φd can be expressed as:
φd = π − sin−1 (ux sin ψd − uy cos ψd )
3.0.2
(20)
Inner Loop Design
For designing the inner loop, we again choose the second order stable error dynamics on attitude as:
φ̈ φ̈d
θ̈ = θ̈ + 2ζin ωnin T
d
ψ̈d
ψ̈
φ̇d − φ̇
θ̇ − θ̇
d
ψ̇d − ψ̇
11
+ ωnin ωnin T
φd − φ
θ −θ
d
ψd − ψ
(21)
The Eulerian angular rates in Eq. 21 is obtained by transforming the body rates to Eulerian rates using Eq. 2. The
onboard sensors measure the rate of rotation of a quadrotor in the body frame. The Eulerian angular acceleration
computed using the error dynamics is transformed to obtain desired body angular acceleration as follows
ṗ
q̇
ṙ
0
−Sθ
1
0 Cφ SφCθ
0 −Sφ CφCθ
=
φ̈
θ̈
ψ̈
0
0
0
+
0
Sφφ̇
Cφφ̇
φ̇
−SφSθθ̇ − CφCθφ̇ θ̇
ψ̇
−SθCφθ̇ + SφCθφ̇
Cθθ̇
(22)
This desired body angular acceleration relation in Eq. 22 is obtained by first inverting the Eq. 2 and then differentiating
it with respect to time. In order to generate the desired body rates in Eq. 22, the quadrotor needs to generate the
following moments:
ld
m
d
nd
Ixx ṗ + (Izz − Iyy )qr
= I q̇ + (I − I )pr
yy
xx
zz
Izz ṙ + (Iyy − Ixx )pq
(23)
Next, the task is to determine the required thrust coefficient (blade pitch angle) to generate the thrust and moments
calculated from the outer and inner loops.
3.0.3
Control Allocation Loop
For the given Td , ld , md , and nd , the task is to find CTi , ∀ i, i=1, 2, 3, 4 by solving Eq. 16. It can be seen that Eq. 16 is
not rational (the yawing moment equation), therefore it is difficult to explicitly obtain the values of CTi . To overcome
this challenge, an additional loop that computes the desired rate of change in blade pitch angles for the given Td , ld ,
md , and nd is used. The objective of control allocation location loop is to determine U = [ĊT1
that drives [T
l
m
n] to [Td
ld
md
ĊT2
ĊT3
ĊT4 ]T
nd ]. For achieving this, the following second order stable error dynamics
is chosen to synthesize the virtual control:
p̈ p̈d
q̈ = q̈ + 2ζCA ωnCA T
d
r̈d
r̈
ṗd − ṗ
q̇ − q̇
d
ṙd − ṙ
+ ωnCA ωnCA T
pd − p
q −q
d
rd − r
(24)
As the error dynamics are chosen on body rates, the desired body angular accelerations are obtained from Eq. 22.
The actual body angular acceleration is obtained from Eq. 5. By solving the above error dynamics, we get moment
12
rates as:
˙
l Ixx p̈ (Iyy − Ixx )(q ṙ + rq̇)
ṁ = I q̈ − (I − I )(pṙ + rṗ)
yy zz
xx
ṅ
Izz r̈
(Ixx − Iyy )(pq̇ + q ṗ)
(25)
Next, a first order error dynamics is applied on thrust to calculate its rate. The error dynamics is as follows
Ṫ = kp (Td − T )
(26)
where kp > 0 is some proportionality constant. Using Eqs. 25 and 26, ĊTi are computed as follows
Kf lag
ĊT1
Ċ
Kl
T2
=
Ċ
−f lagKl
T3
q
|cT 1 |
3KR
Ċt4
−f lag 2
2
Once virtual control input U = [ĊT1
ĊT2
Kf lag
Kf lag
Kf lag
−Kl
−Kl
Kl
−f lagKl
q
|ct2 |
3KR
f lag 2
2
f lagKl
q
|ct3 |
3KR
−f lag 2
2
f lagKl
q
|ct4 |
3KR
f lag 2
2
ĊT3
−1
Ṫ
l˙
ṁ
ṅ
(27)
ĊT4 ]T is obtained, it is integrated with the system dynamics to
obtain the thrust coefficients. For the given thrust coefficients, the desired pitch angle can be obtained by solving
Eq. 10. The control architecture block describing the control flow is shown in Fig. 4.
Figure 4: A control architecture block
13
4
Numerical Results
In this section, the performance of the controller for a variable-pitch quadrotor is demonstrated through four examples.
First, the stability of the inner loop is shown by stabilizing the perturbations given in the attitude from the hover
state. Next, the overall performance is demonstrated by tracking a given trajectory. Next, the controller performs a
flip maneuver on the quadrotor to enable inverted flight while maintaining altitude. Finally, the quadrotor tracks a
sinusoidal trajectory in inverted state. The parameters used for numerical simulation of the variable pitch quadrotor
are given in Table 1. Table 2 lists the control design parameters used for outer, inner and control allocation loops.
Table 1: Parameters for variable pitch quadrotor used for numerical results
Mass of quadrotor, M
1.34 kg
Radius of rotor blades, R
0.18 m
Chord of rotor blades, c
0.03 m
Distance of rotor axis from cg, d
0.3 m
5.23
Airfoil lift curve slope, Clα
0.01
Airfoil drag coefficient, Cd0
Number of blades, Nb
2
Rotational speed, Ω
282.7 rad/sec
Moment of Inertia, Ixx
1 × 10−3 kg-m/sec2
Moment of Inertia, Iyy
1 × 10−3 kg-m/sec2
Table 2: Parameters used for control design
ζout
[0.95 0.95 0.95]T
[4.7 4.7 4.7]T
ωnout
ζin
[0.92 0.92 0.92]T
[30.5 30.5 20.5]T
ωnin
ζCA
[0.91 0.91 0.91]T
[50 50 25]T
ωnCA
kp
10
4.1
Attitude Stabilization
The variable-pitch quadrotor is a fairly new concept and therefore it is necessary to validate that vehicle attitude can
be stabilized changing blade pitch angle before we give simulation results. Toward this, a proof-of-concept single power
plant electric powered variable pitch quadrotor UAV is designed. A PID controller based autopilot is developed and
implemented on open source Pixhawk autopilot board to demonstrate attitude stabilization. The attitude controller
designed on inner loop generates the desired roll, pitch, and yaw moments. The desired thrust is computed from the
altitude stabilization. For control allocation, we assume that thrust and the moments are linear functions of blade
pitch angle. This assumption makes CTi calculation simple and enables the computation of the desired blade pitch
angle using (9). Figure 4.1 shows the attitude tracking performance for roll (Fig. 5(a)), pitch (Fig. 5(b)) and yaw
(Fig. 5(c)) attitudes during closed-loop flight test of the proof-of-concept UAV. It can be observed that the controller
14
40
30
Roll attitude setpoint
Roll attitude achieved
30
Pitch attitude (deg)
Roll attitude (deg)
20
10
0
−10
−20
10
0
−10
−20
−30
−40
0
Pitch attitude setpoint
Pitch attitude achieved
20
2
4
6
Time (sec)
8
−30
0
10
2
(a) Roll attitude
4
6
Time (sec)
8
10
(b) Pitch attitude
90
Yaw attitude (deg)
60
Yaw attitude setpoint
Yaw attitude achieved
30
0
−30
−60
−90
0
2
4
6
Time (sec)
8
10
(c) Yaw attitude
is able to accurately track the commanded setpoints for each of the vehicle attitudes. The setpoints during the flight
test are being provided by human pilot through a joy-stick.
4.2
Position Stabilization
To demonstrate the ability of the controller to stabilize and maintain vehicle position in the event of disturbance,
the initial values of roll (φ), pitch (θ), and yaw (ψ) angles are perturbed by 45◦ , 30◦ and 10◦ , respectively. The
controller brings back the vehicle to hover attitude by reducing the given perturbations quickly. The time history
of attitude variation during this process is shown in Fig. 5(d). The act of stabilizing the quadrotor against large
disturbance results in rapid changes in individual rotor thrust resulting in change in position of the quadrotor. But,
once the attitude disturbance is controlled back to the desired state, the deviation in position is also reduced to zero.
The time history of variation in x, y, and z position coordinates during the stabilization is shown in Fig. 5(e). It
is observed that the attitude is stabilized in less than one sec and the position is restored in less than 1.5 sec. The
overall variation in position in three-dimensions is shown in Fig. 6 and is observed to be small. The time history of
variation of coefficient of thrust CT for individual rotors required for stabilizing the quadrotor is shown in Fig. 7. The
corresponding collective pitch input required to achieve this thrust coefficient is shown in Fig. 8. Even though the
vehicle is released from an attitude which is significantly disturbed from the desired hover attitude, the stabilization
15
Current Angle
Desired Angle
0
X (m)
φ (deg)
50
0
-50
0
0.5
1
Time (s)
1.5
-0.02
-0.04
-0.06
2
40
Y (m)
θ (deg)
0
0.5
1
Time (s)
1.5
2
0
0.5
1
Time (s)
1.5
2
0
0.5
1
Time (s)
1.5
2
0.1
20
0
0.05
-20
0
0.5
1
Time (s)
1.5
0
2
10
0.04
Z (m)
ψ (deg)
Current state
Desired trajectory
5
0
0
0.5
1
Time (s)
1.5
0.02
0
-0.02
2
(d) Attitude
(e) Position
Figure 5: Position and attitude variation during quadrotor stabilization
Initial position
Initial desired position
Current state
0.03
Z (m)
0.02
0.01
0
−0.01
0.1
0
0.05
Y (m)
−0.02
0 −0.06
−0.04
X (m)
Figure 6: Variation in position shown in three-dimensions during quadrotor stabilization
16
0.02
CT
2
CT
1
0.02
0.01
0.01
0
0
0
2
0
1
Time (s)
2
0
1
Time (s)
2
0.02
CT
CT
3
4
0.02
1
Time (s)
0.01
0.01
0
0
0
1
Time (s)
2
Figure 7: Time history of required thrust coefficients for stabilizing the quadrotor
of the quadrotor is achieved with moderate control actuation (less than 16◦ ). Next, the performance of the controller
in trajectory tracking is demonstrated.
4.3
Trajectory Tracking
The trajectory tracking capability of the controller is evaluated by commanding it to follow a sinusoidal path. The
initial position of the quadrotor is set as the origin (0,0,0), the attitude angles (φ, θ, ψ) are (0◦ , 0◦ , 0◦ ). The quadrotor
is commanded from this position to follow a sinusoidal input of sin( π2 t) meters in X, Y , and Z directions. The
controller is able to track the trajectory accurately as observed from the time history of variation of attitude and
position during the trajectory tracking shown in Figs. 9(a) and 9(b). In order to follow the given trajectory, the outer
loop generates required roll (φ) and pitch (θ) commands which are shown by dashed line in Fig. 9(a). The red line
shows the tracking of desired command.
The trajectory traced by the quadrotor in three-dimensions is shown in Fig. 10, which is a slanted circle. The given
command is shown by dashed line and the accurately tracked trajectory is shown using solid lines. Since, the x, y,
and z coordinates depicting the location of the vehicle are varying sinusoidally, it is expected that the controller input
would also vary sinusoidally as shown in Figs. 11 and 12. Again, it can observed from these figures that the controller
is able to regulate the blade pitch angles and thereby generate required thrust to track the prescribed trajectory.
17
20
(deg)
(deg)
20
10
θ
θ
0
0
1
2
10
0
0
0
2
0
1
Time (s)
2
0
1
Time (s)
2
20
θ 0 (deg)
(deg)
20
1
Time (s)
θ
0
4
10
3
10
0
0
0
1
Time (s)
2
Figure 8: Time history of required collective pitch inputs for stabilizing the quadrotor
X (m)
φ (deg)
10
0
0
2
4
6
Time (s)
8
10
50
2
0
1
Y (m)
θ (deg)
0
-1
-10
-50
0
2
4
6
Time (s)
8
10
0
2
4
6
Time (s)
8
10
0
2
4
6
Time (s)
8
10
0
-1
-100
0
2
4
6
Time (s)
8
10
2
2
0
1
Z (m)
ψ (deg)
Current state
Desired trajectory
1
Current Angle
Desired Angle
20
-2
0
-1
-4
0
2
4
6
Time (s)
8
10
(a) Attitude
(b) Position
Figure 9: Position and attitude variation during tracking of prescribed trajectory by quadrotor
18
Initial position
Current state
Initial desired position
Desired trajectory
1.5
Z (m)
1
0.5
0
-0.5
2
1
1
0
0
Y (m)
-1
-1
X (m)
Figure 10: Variation in position of quadrotor shown in three-dimensions during tracking of prescribed
trajectory
0.02
CT
CT
1
2
0.02
0.01
0
0
5
Time (s)
0
0
10
5
Time (s)
10
5
Time (s)
10
4
0.02
CT
3
0.02
CT
0.01
0.01
0
0
5
Time (s)
0.01
0
0
10
Figure 11: Time history of required thrust coefficients for tracking prescribed trajectory
19
(deg)
15
10
10
θ
0
2
1
θ0 (deg)
15
5
0
5
Time (s)
5
0
10
10
5
Time (s)
10
(deg)
15
10
4
10
θ
0
3
θ0 (deg)
15
5
Time (s)
5
0
5
Time (s)
5
0
10
Figure 12: Time history of required collective pitch inputs for tracking prescribed trajectory
0.01
100
X (m)
φ (deg)
200
Current angle
Desired angle
0.5
1
Time (s)
1.5
0
-0.01
0
0
2
0
0.5
1
Time (s)
1.5
2
0
0.5
1
Time (s)
1.5
2
0
0.5
1
Time (s)
1.5
2
1
Y (m)
0.01
θ (deg)
Current position
Initial position
0
0
-1
-2
-0.01
0
0.5
1
Time (s)
1.5
2
0.2
Z (m)
ψ (deg)
0.01
0
0.1
0
-0.1
-0.01
0
0.5
1
Time (s)
1.5
2
(a) Attitude
(b) Position
Figure 13: Position and attitude variation during flip maneuver performed by quadrotor
20
0.02
0
0
2
T
0.01
C
C
T
1
0.02
1
Time (s)
0
0
2
0.02
CT
3
T
1
Time (s)
2
1
Time (s)
2
4
0.02
C
0.01
0.01
0
0
1
Time (s)
0.01
0
0
2
Figure 14: Time history of required thrust coefficients for performing flip maneuver
4.4
Flip Maneuver
In this section, the capability of the variable pitch quadrotor and the developed controller is demonstrated by performing a complicated flip maneuver. In this maneuver, the quadrotor is simulated to fly upside down starting from
the stable hover position with roll, pitch and yaw attitude angles maintained at 0◦ . The controller then commands the
quadrotor to change the roll angle to 180◦ while maintaining the pitch and yaw angles. The time history of commanded
and achieved attitude angles of the vehicle is shown in Fig. 13(a). The time history of position of the quadrotor during
the flip maneuver is shown in Fig. 13(b). During this maneuver no attempt is made to control the position of the
quadrotor and only attitude is targeted. As a consequence, it can be observed that during the transition from 0◦ roll
attitude to 180◦ roll attitude, the quadrotor generates some acceleration which results in a small velocity along lateral
(Y ) direction which makes the Y coordinate position to increase with time.
After the execution of the flip maneuver, all the four rotors of the quadrotor produce thrust of equal magnitude
and has same value as that of thrust in hover mode (see Fig. 14). As expected, the rotors of the quadrotor, when in
inverted flight, operate at negative collective pitch angles to generate thrust in upward direction as shown in Fig. 15.
The trajectory of the quadrotor during the flip maneuver in Y − Z plane is shown in Fig. 16. The upright attitude
of the quadrotor at its original location is marked by number ‘1’ and is depicted using a square with dark shade in
the top half and light shade in the bottom half portion. The snapshots of the simulated flip maneuver are marked
by numbers ‘1’ through ‘7’. It is observed that the flipping of the quadrotor is completed by the time the quadrotor
reaches location ‘6’ as it attains upside down attitude marked by a square with bottom half in dark and top half in
21
(deg)
20
0
0
θ
0
2
1
θ0 (deg)
20
−20
0
1
Time (s)
−20
0
2
2
1
Time (s)
2
θ0 (deg)
20
0
0
4
3
θ0 (deg)
20
1
Time (s)
−20
0
1
Time (s)
−20
0
2
Figure 15: Time history of required collective pitch input for performing flip maneuver
light shade. The centre of mass of the quadrotor is observed to move by only 0.14 m in lateral direction and 0.07
m in vertical direction during the execution of the flip maneuver. The quadrotor maintains its altitude but drifts in
Y-direction due to the reason explained above.
4.5
Inverted Flight
For final demonstration of the performance of the controller, a sinusoidal trajectory is tracked by the quadrotor in
inverted orientation. Starting with hover, the quadrotor is commanded to perform a flip maneuver followed by tracking
of a sinusoidal trajectory of sin( π2 t) meters in X, Y , and Z directions.
Similar to trajectory following, the attitude required to track the trajectory is shown by dashed line in Fig. 17(a)
and actual attitude attained is shown using solid line. Fig. 17(a) shows that quadrotor flips within 1 sec attaining a
roll angle of 180◦, and then starts following the desired attitude to track the prescribed trajectory. The time history
of desired and tracked positions are shown in Fig. 17(b). After the initial deviation of X and Y location during the
flipping motion, the inverted quadrotor is able to track the desired trajectory with great precision. The corresponding
flight path in three-dimensions is shown in Fig. 10.
The rapid changes in the commanded thrust from the individual rotors is shown in Fig. 19. Figure 20 shows that
all the rotors operate at negative collective input after the quadrotor is inverted to produce thrust in upward direction
for tracking the trajectory in upside down attitude.
22
0.1
0.08
6
5
Z (m)
0.06
0.04
4
0.02
-0
7
3
1
2
-0.02
-0.6
-0.5
-0.4
-0.3
Y (m)
-0.2
-0.1
0
Figure 16: Trajectory in Y-Z plane during flipping
X (m)
φ (deg)
400
200
0
2
4
6
Time (s)
8
10
0
2
4
6
Time (s)
8
10
0
2
4
6
Time (s)
8
10
0
2
4
6
Time (s)
8
10
2
Y (m)
50
θ (deg)
0
-1
0
0
1
0
-1
-50
0
2
4
6
Time (s)
5
8
10
2
Current Angle
Desired Angle
Z (m)
ψ (deg)
Current state
Desired trajectory
1
0
1
0
-1
-5
0
2
4
6
Time (s)
8
10
(a) Attitude
(b) Position
Figure 17: Position and attitude variation during trajectory tracking with inverted quadrotor
23
Initial position
Current state
Initial desired position
Desired trajectory
1.5
Z (m)
1
0.5
0
−0.5
2
1
1
0
0
Y (m)
−1 −1
X (m)
Figure 18: Three Dimensional Variation in Position in tracking trajectory during inverted flight
T
0.01
0
0
C
CT
2
0.02
1
0.02
5
Time (s)
0
0
10
5
Time (s)
10
5
Time (s)
10
T
4
0.02
0.01
0
0
C
3
0.02
CT
0.01
5
Time (s)
0.01
0
0
10
Figure 19: Required values of thrust coefficients variation for tracking trajectory during inverted flight
5
Conclusions
This paper discusses the development of the flight dynamics model of a variable-pitch quadrotor which is suitable for
model based controller design. The thrust and moment for each rotor is calculated using Blade Element Theory and
momentum theory. Due to its ability to generate negative thrust, the variable-pitch quadrotor is known to offer higher
controller bandwidth, which is suitable for aggressive maneuvering and inverted flight. A novel nonlinear controller is
developed using dynamic inversion approach and demonstrated for stabilization, tracking, flipping and inverted flying
of the variable pitch quadrotor. The challenge associated with control allocation, due to non-rational relation between
blade pitch angle and rotor propulsive forces, is solved using an additional loop in the control design. The strategy of
controlling the quadrotor by changing the blade pitch angle is validated by showing attitude stabilization in real flight
24
20
(deg)
(deg)
20
0
θ
θ
0
0
2
1
0
−20
0
5
Time (s)
−20
0
10
10
5
Time (s)
10
20
(deg)
(deg)
20
5
Time (s)
0
θ
θ
0
0
3
4
0
−20
0
5
Time (s)
−20
0
10
Figure 20: Required values of Collective inputs variation for tracking trajectory during inverted flight
for a variable pitch quadrotor. The change in coordinate system due to flipping is taken care by introducing suitable
variable for booking keeping of the orientation. The performance of controller is demonstrated through numerical
simulations. As the controller is derived using six-DOF model, it is generic and can be employed for the whole flight
regime.
References
[1] Singh B, Chopra I. Insect-based hover-capable flapping wings for micro air vehicles: experiments and analysis.
AIAA journal. 2008;46(9):2115–2135.
[2] Seshadri P, Benedict M, Chopra I. Understanding micro air vehicle flapping-wing aerodynamics using force and
flowfield measurements. Journal of Aircraft. 2013;50(4):1070–1087.
[3] Bohorquez F, Samuel P, Sirohi J, Pines D, Rudd L, Perel R. Design, analysis and hover performance of a rotary
wing micro air vehicle. Journal of the American Helicopter Society. 2003;48(2):80–90.
[4] Schafroth D, Bouabdallah S, Bermes C, Siegwart R. From the test benches to the first prototype of the muFly
micro helicopter. In: Unmanned Aircraft Systems. Springer; 2008. p. 245–260.
[5] Ulrich ER, Humbert JS, Pines DJ. Pitch and heave control of robotic samara micro air vehicles. Journal of
Aircraft. 2010;47(4):1290–1299.
[6] Naudin JL. The GFS UAV Project. Retrieved on. 2006;25:2010.
25
[7] Benedict M, Jarugumilli T, Chopra I. Experimental Optimization of MAV-Scale Cycloidal Rotor Performance.
Journal of the American Helicopter Society. 2011;56(2):22005–22005.
[8] Bouabdallah S, Noth A, Siegwart R. PID vs LQ control techniques applied to an indoor micro quadrotor. In:
Intelligent Robots and Systems, 2004.(IROS 2004). Proceedings. 2004 IEEE/RSJ International Conference on.
vol. 3. IEEE; 2004. p. 2451–2456.
[9] Bouabdallah S, Murrieri P, Siegwart R. Design and control of an indoor micro quadrotor. In: Robotics and
Automation, 2004. Proceedings. ICRA’04. 2004 IEEE International Conference on. vol. 5. IEEE; 2004. p. 4393–
4398.
[10] Castillo P, Dzul A, Lozano R. Real-time stabilization and tracking of a four-rotor mini rotorcraft. IEEE Transactions on control systems technology. 2004;12(4):510–516.
[11] Cutler M, Ure NK, Michini B, How JP. Comparison of fixed and variable pitch actuators for agile quadrotors.
AIAA Paper. 2011;(2011-6406).
[12] Abhishek A, Gadekar R, Duhoon A, Kothari M, Kadukar S, Rane L, et al. Design, development, and closed-loop
flight testing of a single power plant variable pitch quadrotor unmanned air vehicle. In: the 73rd American
Helicopter Society Annual Forum. AHS; 2017. .
[13] Leishman JG. 2. In: Principles of Helicopter Aerodynamics. 2nd ed. Cambridge University Press; 2006. p.
115–170.
[14] Borenstein J. The hoverbot, an electrically powered flying robot. Ann Arbor. 1992;1001:48109–2110.
[15] Cutler M, How JP. Actuator constrained trajectory generation and control for variable-pitch quadrotors. In:
AIAA Guidance, Navigation, and Control Conference; 2012. p. 1–15.
[16] Madani T, Benallegue A. Backstepping control for a quadrotor helicopter. In: 2006 IEEE/RSJ International
Conference on Intelligent Robots and Systems. IEEE; 2006. p. 3255–3260.
[17] Nagaty A, Saeedi S, Thibault C, Seto M, Li H. Control and navigation framework for quadrotor helicopters.
Journal of intelligent & robotic systems. 2013;p. 1–12.
[18] Bouabdallah S, Siegwart R. Backstepping and sliding-mode techniques applied to an indoor micro quadrotor. In:
Proceedings of the 2005 IEEE international conference on robotics and automation. IEEE; 2005. p. 2247–2252.
26
[19] Das A, Subbarao K, Lewis F. Dynamic inversion with zero-dynamics stabilisation for quadrotor control. IET
control theory & applications. 2009;3(3):303–314.
[20] Prabhakaran B, Kothari M, Abhishek. Nonlinear control design for quadrotors. In: 2015 IEEE Workshop on
Computational Intelligence: Theories, Applications and Future Directions (WCI); 2015. p. 1–6.
[21] Dydek ZT, Annaswamy AM, Lavretsky E. Combined/composite adaptive control of a quadrotor UAV in the
presence of actuator uncertainty. In: AIAA Guidance, Navigation, and Control Conference; 2010. p. 2–5.
[22] Lee D, Kim HJ, Sastry S. Feedback linearization vs. adaptive sliding mode control for a quadrotor helicopter.
International Journal of control, Automation and systems. 2009;7(3):419–428.
[23] Mohammadi M, Shahri AM. Adaptive nonlinear stabilization control for a quadrotor UAV: theory, simulation
and experimentation. Journal of Intelligent & Robotic Systems. 2013;72(1):105–122.
[24] Shastry AK, Pattanaik A, Kothari M. Neuro-adaptive Augmented Dynamic Inversion Controller for Quadrotors.
IFAC-PapersOnLine. 2016;49(1):302–307.
[25] Carrillo LG, Dzul A, Lozano R. Hovering quad-rotor control: A comparison of nonlinear controllers using visual
feedback. IEEE Transactions on Aerospace and Electronic Systems. 2012;48(4):3159–3170.
[26] Cutler M, How JP. Analysis and control of a variable-pitch quadrotor for agile flight. Journal of Dynamic Systems,
Measurement, and Control. 2015;137(10):101002.
[27] Gupta N, Kothari M, Abhishek A. Flight dynamics and nonlinear control design for variable-pitch quadrotors.
In: 2016 American Control Conference (ACC). IEEE; 2016. p. 3150–3155.
[28] Parwana H, Kothari M. Quaternions and Attitude Representation. arXiv preprint arXiv:170808680. 2017;.
[29] Enns D, Bugajski D, Hendrick R, Stein G. Dynamic inversion: an evolving methodology for flight control design.
International Journal of control. 1994;59(1):71–91.
27
| 3cs.SY
|
arXiv:1711.05797v2 [math.GR] 27 Feb 2018
THE RESIDUAL FINITENESS OF (HYPERBOLIC)
AUTOMORPHISM-INDUCED HNN-EXTENSIONS
ALAN D. LOGAN
Abstract. We classify finitely generated, residually finite automorphisminduced HNN-extensions in terms of the residual separability of a
single associated subgroup. This classification provides a method
to construct automorphism-induced HNN-extensions which are not
residually finite. We prove that this method can never yield a
“new” counter-example to Gromov’s conjecture on the residual
finiteness of hyperbolic groups.
1. Introduction
A group H∗(K,φ) is called an automorphism-induced HNN-extension
if it has a relative presentation of the form
H∗(K,φ) = hH, t; tkt−1 = φ(k), k ∈ Ki
where φ ∈ Aut(H) and K H.
The main result of this note is a classification of finitely generated,
residually finite automorphism-induced HNN-extensions. A subgroup
K of H is residually separable in H if for all x ∈ H \ K there exists a
finite index, normal subgroup N of H, written N ⊳f H, such that x 6∈
KN (hence if ϕx : H → H/N is the natural map then ϕx (x) 6∈ ϕx (K)).
Theorem A. Suppose that H is finitely generated. Then G = H∗(K,φ)
is residually finite if and only if H is residually finite and K is residually separable in H.
We prove two corollaries of Theorem A. These corollaries can be easily applied to construct automorphism-induced HNN-extensions which
are not residually finite. Both corollaries relate to the subgroup-quotient
NH (K)/K. This subgroup-quotient plays a central role in a framework
for the construction of groups possessing certain properties and with
specified outer automorphism group [Log15] (see also [Log16] [Log17]).
Date: February 28, 2018.
2010 Mathematics Subject Classification. 20E06, 20E26, 20F67.
Key words and phrases. HNN-extensions, Residual finiteness, Hyperbolic groups.
1
2
ALAN D. LOGAN
Corollary 1.1. Suppose that H is finitely generated. If NH (K)/K is
not residually finite then G = H∗(K,φ) is not residually finite.
Corollary 1.2. Suppose that H is finitely generated and that NH (K)
has finite index in H. Then G = H∗(K,φ) is residually finite if and only
if both H and NH (K)/K are residually finite.
Hyperbolicity. It is a famous conjecture of Gromov that all hyperbolic groups are residually finite [nib93] [KW00] [Ol’00]. One might
hope to apply Corollary 1.1 to obtain a counter-example to this conjecture. However, Theorem B proves that Corollary 1.1 can produce
no “new” counter-examples to Gromov’s conjecture, in the sense that
if G = H∗(K,φ) is a counter-example where the subgroup-quotient
NH (K)/K is used to force G to be non-residually finite then the conditions of Theorem B hold, and so H is also a counter-example.
Theorem B. Suppose that G = H∗(K,φ) is hyperbolic and non-residually
finite, and that K
NH (K). Then K is finite, and H is hyperbolic
and non-residually finite.
Theorem B leaves the following question:
Question 1.3. Suppose that G = H∗(K,φ) is hyperbolic and non-residually
finite. Then is H is hyperbolic and non-residually finite?
We also have the following result:
Theorem C. Suppose that K NH (K) and that K contains an element of infinite order. Then Z × Z embeds into G = H∗(K,φ).
Automorphism-induced HNN-extensions can be thought of as “partial” mapping tori H ⋊φ Z. Theorem C proves that automorphisminduced HNN-extensions of free groups Fn ∗(K,φ) are not hyperbolic if
K NH (K), even if the “full” mapping torus Fn ⋊φ Z is hyperbolic.
Acknowledgments. I would like to thank the anonymous referee for
their extremely helpful comments.
2. Residual finiteness
We first prove Theorem A. Note that for G some group, if P ⊳f G
and H ≤ G then P ∩H ⊳f H. Also note that if H is a finitely generated
group, Q ⊳f H and φ ∈ Aut(H) then ∩i∈Z φi (Q) ⊳f H.
Proof of Theorem A. Suppose H is residually finite and K is residually
separable in H. Then H∗(K,φ) is residually finite [BT78, Lemma 4.4].
RESIDUAL FINITENESS OF CERTAIN HNN-EXTENSIONS
3
Suppose H∗(K,φ) is residually finite. Then H is residually finite, as
subgroups of residually finite groups are residually finite. Now, suppose
that K is not residually separable in H, and let x ∈ H \ K be such that
x ∈ KN for all finite index subgroups N of H. Let N ⊳f H∗(K,φ) be
arbitrary. It is sufficient to prove that txt−1 φ(x)−1 ∈ N . To see this
ini
clusion, first note that N ∩H ⊳f H. Consider L := ∩i∈Z φ N ∩ H , and
note that L ⊳f H. Then there exists k ∈ K such that xk −1 , φ(xk −1 ) ∈
L. Thus, xk −1 , φ(xk −1 ) ∈ N , and so xN = kN and φ(x)N = φ(k)N.
Then:
txt−1 φ(x)−1 N = tkt−1 φ(k)−1 N = N
Hence, txt−1 φ(x)−1 ∈ N as required.
We now prove Corollary 1.1.
Proof of Corollary 1.1. Suppose that NH (K)/K is not residually finite.
Then there exists some x ∈ NH (K) such that x ∈ NK for all N ⊳f
NH (K). Hence, for all N ⊳f H we have that x ∈ N ∩ NH (K) K,
and so x ∈ NK. Therefore, K is not residually separable in H, and so
H∗(K,φ) is not residually finite by Theorem A.
We now prove Corollary 1.2. We previously proved the analogous
result for the groups H∗(K,1), so where the inducing automorphism φ
is trivial [Log16, Proposition 2.2 ].
Proof of Corollary 1.2. By Theorem A and Corollary 1.1, it is sufficient to prove that if H and NH (K)/K are residually finite then K is
residually separable. So, suppose that H and NH (K)/K are residually
finite.
Additionally, suppose that x 6∈ NH (K). Clearly x 6∈ NH (K)K as
K ≤ NH (K). Then the subgroup N := ∩h∈H h−1 NH (K)h is a finite
index, normal subgroup of H such that x 6∈ NK, as required.
Suppose that x ∈ NH (K) \ K. Now, as NH (K)/K is residually
finite, there exists a map ϕx : NH (K)/K → Ax with Ax finite and
xK 6∈ ker(ϕx ). Therefore, there exists a map ϕ
fx : NH (K) → Ax
ϕx
which factors as NH (K) → NH (K)/K −→ Ax such that x 6∈ ker (f
ϕx ).
Then K ≤ ker (f
ϕx ) so x 6∈ ker (f
ϕx ) K. As ker (f
ϕx ) ⊳f NH (K) ⊳f H,
there exists N ⊳f H such that N ≤ ker (f
ϕx ). As x 6∈ ker (f
ϕx ) K and
NK ≤ ker (f
ϕx ) K we have that x 6∈ NK as required.
3. Hyperbolicity
We first prove Theorem C, as it is applied in the proof of Theorem B.
Recall that Theorem C gives a necessary condition for Z × Z to embed
into G = H∗(K,φ). As Z × Z does not embed into any hyperbolic
4
ALAN D. LOGAN
group, Theorem C gives a necessary condition for the hyperbolicity of
automorphism-induced HNN-extensions.
Proof of Theorem C. Consider an element k ∈ K of infinite order, and
consider a ∈ NH (K) \ K. Then the word W = a−1 t−1 φ(a)t has infinite
order in G, and indeed no power of W is contained in K. Now, as
aka−1 ∈ K we have that t−1 φ(aka−1 ) = aka−1 t−1 . Then W and k
commute as follows:
a−1 t−1 φ(a)t · k = a−1 t−1 φ(ak)t
= a−1 t−1 φ(aka−1 )φ(a)t
= k · a−1 t−1 φ(a)t
Therefore, hW, ki ∼
= Z × Z as required.
We now prove Theorem B.
Proof of Theorem B. By assumption, G = H∗(K,φ) is hyperbolic and
non-residually finite, and K
NH (K). Suppose that K is infinite.
Then K is an infinite torsion group by Theorem C. Now, as K ≤ G
with G hyperbolic, this is a contradiction [Gro87]. Hence, K is finite.
Suppose that H is residually finite. As K is finite we have that G
is residually finite [BT78, Theorem 3.1], a contradiction. Hence, H is
non-residually finite.
Finally, note that H is a quasi-convex subgroup of G as K and φ(K)
are finite. Hence, H is hyperbolic [BH99, Proposition III.Γ.3.7].
References
[ALP17] F. Ateş, A. D. Logan, and S.J. Pride, Automata and Zappa-Szèp products
of groups, in preparation (2017).
[Bau63] G. Baumslag, Automorphism groups of residually finite groups, J. London
Math. Soc. 38 (1963), 117–118. MR 0146271
[BH99] M. R. Bridson and A. Haefliger, Metric spaces of non-positive curvature,
Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 319, Springer-Verlag, Berlin, 1999.
MR 1744486
[BT78] B. Baumslag and M. Tretkoff, Residually finite HNN extensions, Comm.
Algebra 6 (1978), no. 2, 179–194. MR 484178
[Gro87] M. Gromov, Hyperbolic groups, Math. Sci. Res. Inst. Publ., vol. 8,
Springer, New York, 1987. MR 919829
[KM98] O. Kharlampovich and A. Myasnikov, Hyperbolic groups and free constructions, Trans. Amer. Math. Soc. 350 (1998), no. 2, 571–613. MR 1390041
[KW00] I. Kapovich and D.T. Wise, The equivalence of some residual properties of word-hyperbolic groups, J. Algebra 223 (2000), no. 2, 562–583.
MR 1735163
RESIDUAL FINITENESS OF CERTAIN HNN-EXTENSIONS
5
[Log15] A. D. Logan, The Bass-Jiang group for automorphism-induced HNNextensions, arXiv:1509.01847 (2015).
, On a question of Bumagin and Wise, New York J. Math. 22
[Log16]
(2016), 865–873. MR 3548127
[Log17]
, Every group is the outer automorphism group of an HNNextension of a fixed triangle group, arXiv:1709.06441 (2017).
[nib93] Geometric group theory 1991 problem list, Geometric group theory, Vol. 1
(Sussex, 1991) (Graham A. Niblo, ed.), London Math. Soc. Lecture Note
Ser., vol. 181, Cambridge Univ. Press, Cambridge, 1993, pp. 208–212.
MR 1238528
[Ol’00] A. Yu. Ol’shanskiı̆, On the Bass-Lubotzky question about quotients of hyperbolic groups, J. Algebra 226 (2000), no. 2, 807–817. MR 1752761
School of Mathematics and Statistics, University of Glasgow, Glasgow, G12 8QW, UK.
E-mail address: [email protected]
| 4math.GR
|
SEMIPROJECTIVITY AND SEMIINJECTIVITY IN DIFFERENT
CATEGORIES
arXiv:1802.05037v1 [math.CT] 14 Feb 2018
HANNES THIEL
Abstract. Projectivity and injectivity are fundamental notions in category
theory. We consider natural weakenings termed semiprojectivity and semiinjectivity, and study these concepts in different categories.
For example, in the category of metric spaces, (semi)injective objects are
precisely the absolute (neighborhood) retracts. We show that the trivial group
is the only semiinjective group, while every free product of a finitely presented
group and a free group is semiprojective.
To a compact, metric space X we associate the commutative C ∗ -algebra
C(X). This association is contravariant, whence semiinjectivity of X is related
to semiprojectivity of C(X). Together with Adam Sørensen, we showed that
C(X) is semiprojective in the category of all C ∗ -algebras if and only if X is
an absolute neighborhood retract with dim(X) ≤ 1.
1. Introduction
While being fundamental in category theory, the concepts of projectivity and
injectivity are often very restrictive. It is therefore natural to consider weaker
versions of these notions.
For example, injective objects in the category of metric spaces and continuous
maps are precisely the absolute retracts introduced by Borsuk in 1931. He also
defined a generalization, called absolute neighborhood retracts; see Definition 3.10.
Therefore, being an absolute neighborhood retract is a weak form of injectivity
in the category of metric spaces. While not many spaces are absolute retracts,
numerous naturally occurring spaces are absolute neighborhood retracts, including
topological manifolds, polyhedra and CW-complexes.
Given a compact, metric space X, we associate the algebra C(X) of continuous
complex-valued functions on X. This is a C ∗ -algebra with the supremum norm and
pointwise operations. The category CMetr of compact, metric spaces, and the category AbSC∗1 of abelian, unital, separable C ∗ -algebras are (contravariantly) equivalent. We therefore think of C ∗ -algebras as ‘noncommutative topological spaces’.
Using the contravariant correspondence between CMetr and AbSC∗1 , injectivity
of a compact, metric space X corresponds precisely to projectivity of C(X) in the
category AbSC∗1 . However, projectivity within the category C∗1 of unital C ∗ algebras and unital ∗ -homomorphisms is more restrictive - since there are more
lifting problems to be solved.
Date: February 15, 2018.
2010 Mathematics Subject Classification. Primary 18A05; Secondary 06B35, 06F05, 18A20,
20E05, 46L05, 54C55, 55M15.
Key words and phrases. projectivity, injectivity, semiprojectivity, semiinjectivity, free groups,
absolute retracts, absolute neighborhood retracts, C*-algebras.
The author was partially supported by the Deutsche Forschungsgemeinschaft (SFB 878).
1
2
HANNES THIEL
In 1985, Blackadar used the contravariant correspondence between CMetr and
AbSC∗1 to translate the concept of an absolute neighborhood retract to (noncommutative) C ∗ -algebras. He hence introduced a weak form of projectivity in the category of C ∗ -algebras, called semiprojectivity; see [Bla85a, Definition 2.1], [Bla85b,
Definition 2.10].
It is straightforward to generalize Blackadar’s definition of semiprojectivity to
general categories; see Definition 3.4. The dual notion is called semiinjectivity.
Semiinjective objects in the category of metric spaces are precisely the absolute
neighborhood retracts.
We characterize semiprojectivity and semiinjectivity in the category of groups:
A group is semiprojective if and only if it is a retract of a free product of a finitely
presented group and a free group; see Proposition 3.7. On the other hand, only the
trivial group is semiinjective; see Proposition 3.9.
One motivation to consider semiprojectivity and semiinjectivity is shape theory,
which is a machinery to study an object by approximating it by better-behaved
ones. Absolute neighborhood retracts are the building blocks of shape theory of
topological spaces. Analogously, semiprojective C ∗ -algebras are the building blocks
of noncommutative shape theory. In this context, it is of general interest to study
semiprojective C ∗ -algebras.
More specifically, semiprojectivity is a concept that is used at many different
places in the theory of C ∗ -algebras. For instance, it is often used that a C ∗ -algebra
that is ‘locally approximated’ by a certain class of semiprojective C ∗ -algebras is
already isomorphic to an inductive limit of such C ∗ -algebras; see [Thi11, Section 3].
Further, semiprojective C ∗ -algebras are used to study and classify C ∗ -algebras given
as inductive limits or as crossed products of dynamical systems. For example,
Elliott’s seminal classification of AF-algebras (inductive limits of finite-dimensional
C ∗ -algebras) by K-theory relies on the semiprojectivity of finite-dimensional C ∗ algebras. Semiprojectivity also plays a crucial role in the analysis of the structure
of crossed products by actions with the Rokhlin property in [OP12] and [Gar17].
Blackadar asked to determine, in terms of X, when C(X) is semiprojective among
all C ∗ -algebras. It is easy to see that X must be an absolute neighborhood retract,
but is that sufficient? Surprisingly, a dimensional restriction appears. Together
with Adam Sørensen we showed in [ST12] that C(X) is semiprojective if and only
if X is an absolute neighborhood retract with dim(X) ≤ 1; see Theorem 4.8.
This article is based on a talk presented at the conference ‘VI Coloquio Uruguayo
de Matemática’, held during December 20 to 22, 2017, in Montevideo, Uruguay.
Acknowledgements
I am thankful to Eusebio Gardella for his feedback on the first draft of this paper.
2. Concrete categories, Monomorphisms, Epimorphisms
Let C be a category. Given objects X and Y in C, we use HomC (X, Y ) to
denote the morphisms in C from X to Y .
Let ϕ : X → Y be a morphism in C. Then ϕ is an epimorphism (for short, ϕ is
epi), denoted ϕ : X ։ Y , if for every object Z and morphisms ψ1 , ψ2 : Y → Z with
ψ1 ◦ ϕ = ψ2 ◦ ϕ, we have ψ1 = ψ2 . Dually, ϕ is a monomorphism (for short, ϕ is
mono), denoted ϕ : X ֒→ Y , if for every object Z and morphisms ψ1 , ψ2 : Z → X
with ϕ ◦ ψ1 = ϕ ◦ ψ2 , we have ψ1 = ψ2 .
Example 2.1. Let Set denote the category of sets and (ordinary) mappings. A
morphism in Set is epi (mono) if and only if it is surjective (injective).
SEMIPROJECTIVITY AND SEMIINJECTIVITY IN DIFFERENT CATEGORIES
3
Recall that C is said to be locally small if HomC (X, Y ) is a set for any objects
X and Y . In this case, for each object X we obtain a covariant hom functor
HomC (X, ) : C → Set and a contravariant hom functor HomC ( , X) : C → Set.
A concrete category is a category C together with faithful functor U : C → Set.
In this case, we think of an object X in C as a set (namely U (X)) with additional
structure, and a morphism ϕ : X → Y is a mapping (namely U (ϕ) : U (X) → U (Y ))
that preserves the structure of the objects. In a concrete category, we usually
identify an object with its underlying set, and we identify a morphism with its
underlying set mapping.
A faithful functor reflects epimorphisms and monomorphisms. It follows that
in a concrete category, every surjective (injective) morphism is epi (mono). The
converse need not hold; see Examples 2.3 and 2.4.
Of particular interest is the case that C is locally small and that there exists
an object G in C (called a generator ) such that HomC (G, ) : C → Set is faithful;
see [Bor94, Corollary 4.5.9, p.155]. In that case, a morphism is mono if and only
if it is injective. As noted above, the backward implication holds in every concrete
category. To show the forward implication, let ϕ : X → Y be a monomorphism. To
show that ϕ is injective, let x and y be elements in the set underlying X such that
ϕ(x) = ϕ(y). Note that x is an element of HomC (G, X) and that ϕ(x) is just the
composition of morphisms ϕ ◦ x in C. Thus, the equality ϕ(x) = ϕ(y) really means
ϕ ◦ x = ϕ ◦ y. Now it follows directly from the definition of monomorphism that
x = y, as desired.
Dually, if a category has a cogenerator (an object K such that HomC ( , K) is
faithful), then a morphism is epi if and only if it is surjective.
Example 2.2. Let Gp denote the category of discrete groups and group homomorphisms, with the usual concretization that sends a group to its underlying set.
Then the group Z is a generator for Gp. Indeed, given a group G, there is a natural bijection between elements in G and group homomorphisms Z → G. See also
[Bor94, Example 4.5.17.c, p.160]. It follows that a morphism in Gp is mono if and
only if it is injective.
The category Gp has no cogenerator; see [Bor94, Proposition 4.7.3, p.169]. Nevertheless, a morphism in Gp is epi if and only if it is surjective. The forward
implication is not obvious; see [Lin70].
Example 2.3. Let Mon denote the category of monoids and monoid homomorphisms. The inclusion map ϕ : N → Z is a non-surjective epimorphism. To show
that ϕ is epi, let M be a monoid, and let ψ1 , ψ2 : Z → M be morphisms with
ψ1 ◦ ϕ = ψ2 ◦ ϕ. Then ψ1 (k) = ψ2 (k) for all k ≥ 0. We have 1M = ψ1 (0) =
ψ1 (−k)ψ1 (k) and 1M = ψ2 (k)ψ2 (−k) for all k ≥ 0. We deduce that
ψ1 (−k) = ψ1 (−k)ψ2 (k)ψ2 (−k) = ψ1 (−k)ψ1 (k)ψ2 (−k) = ψ2 (−k),
for all k ≥ 0. Thus, ψ1 = ψ2 , as desired.
Example 2.4. Consider the category of pointed, path connected spaces with
pointed, continuous maps. We let T = {z ∈ C : |z| = 1} be the circle with
base point 1. Let π : (R, 0) → (S 1 , 1) be given by π(t) := exp(2πit), for t ∈ R.
Then π is a non-injective monomorphism. To show that π is a monomorphism, let
(X, x0 ) be a path connected space, and let f1 , f2 : (X, x0 ) → (R, 0) be two pointed,
continuous maps satisfying π ◦ f1 = π ◦ f2 . Then f1 (x0 ) = 0 = f2 (x0 ). Given
x ∈ X, choose a path from x0 to x, that is, a continuous map p : [0, 1] → X with
p(0) = x0 and p(1) = x. Then f1 ◦ p and f2 ◦ p are two paths in R starting at 0.
We further have π ◦ f1 ◦ p = π ◦ f2 ◦ p. Since π is a covering, it has the unique path
4
HANNES THIEL
lifting property. It follows that f1 ◦ p = f2 ◦ p and hence
f1 (x) = (f1 ◦ p)(1) = (f2 ◦ p)(1) = f2 (x).
Thus, f1 = f2 , as desired.
Example 2.5. Let CMetr be the category of compact, metric spaces and continuous mappings, with the usual concretization sending a topological space to its
underlying set. The one-point space is a generator for CMetr. The interval [0, 1]
with its usual Hausdorff topology is a cogenerator for CMetr; see [Bor94, Proposition 4.7.8, p.173]. Thus, epimorphisms (monomorphisms) in CMetr are precisely
surjective (injective) continuous mappings.
3. Semiprojective and semiinjective objects
The following definition is standard in category theory.
Definition 3.1. Let C be a category, and let X be an object in C. Then X is said
to be projective if for every epimorphism π : Y → Z and every morphism ϕ : X → Z
there exists a morphism ϕ̃ : X → Y such that π ◦ ϕ̃ = ϕ. The morphism ϕ̃ is called
a lift of ϕ.
Dually, X is said to be injective if for every monomorphism ι : Z → Y and every
morphisms ϕ : Z → X there exists a morphism ϕ̃ : Y → X such that ϕ̃ ◦ ι = ϕ. The
morphism ϕ̃ is called an extension of ϕ.
Thus, X is projective (injective), if in the left (right) diagram below, for given
solid arrows, the dashed arrow exists making the diagram commutative:
>Y
ϕ̃
X
π
ϕ
YO
ϕ̃
/Z
ι
~
Xo
ϕ
?
Z.
Examples 3.2. (1) A group G is projective (in Gp) if and only if G is free. The
backward implication is easy to prove. To show the forward implication, choose a
free group F and a surjective group homomorphism π : F → G. Using that G is
projective, we obtain a morphism ϕ̃ : G → F that lifts the identity on G, that is,
such that π ◦ ϕ̃ = idG . This is shown in the following commutative diagram:
ϕ̃
G
idG
>F
π
/ G.
It follows that ϕ̃ is injective, and thus G is (isomorphic to) a subgroup of F . By the
Nielsen-Schreier theorem, every subgroup of a free group is again free. It follows
that G is free, as desired.
(2) Eilenberg and Moore showed that the trivial group is the only injective object
in Gp. We include a short proof, which is a variation of the proof in [Nog07].
Let G be an injective group, and let g ∈ G. Let F2 denote the free group of
rank two, with generators x and y. Let ϕ : F2 → G be the morphism satisfying
ϕ(x) = 1 and ϕ(y) = g. Let σ : F2 → F2 be the automorphism of F2 satisfying
σ(x) = y and σ(y) = x. We consider the semidirect product F2 ⋊σ Z2 . Let
ι : F2 → F2 ⋊σ Z2 denote the natural inclusion morphism. Use that G is injective to
SEMIPROJECTIVITY AND SEMIINJECTIVITY IN DIFFERENT CATEGORIES
5
obtain an extension ϕ̃ of ϕ. This is shown in the following commutative diagram.
GO d
ϕ
F2
ϕ̃
ι
/ F2 ⋊σ Z2 .
To simplify, we consider F2 as a subgroup of F2 ⋊σ Z2 . Let u ∈ F2 ⋊σ Z2 be the
element implementing σ. Then uxu−1 = y. We have ϕ̃(x) = ϕ(x) = 1 and hence
g = ϕ(y) = ϕ̃(uxu−1 ) = ϕ̃(u)ϕ̃(x)ϕ̃(u)−1 = ϕ̃(u)ϕ̃(u)−1 = 1.
Thus, G = {1}, as desired.
Recall that a partially ordered set I is said to be upward directed (downward
directed ) if for all i, j ∈ I there exists k ∈ I with i, j ≤ k (with k ≤ i, j). The
following definition is standard.
Definition 3.3. A direct system (an inverse system) in a category C is an upward
directed (downward directed) set I, together with objects Xi for i ∈ I and morphisms πi,j : Xi → Xj for i, j ∈ I with i ≤ j, satisfying πi,i = idXi for every i ∈ I
and satisfying πi,k = πj,k ◦ πi,j for all i, j, k ∈ I with i ≤ j ≤ k. The morphisms
πi,j are called the connecting morphisms of the system.
Given a direct system (I, Xi , πi,j ), a direct limit (also called inductive limit ) is
an object X together with a family π = (πi,∞ )i∈I of morphisms πi,∞ : Xi → X
satisfying πj,∞ ◦ πi,j = πi,∞ for all i, j ∈ I with i ≤ j, and such that (X, π) is
universal with these properties. Dually, given an an inverse system (I, Xi , πi,j ), an
inverse limit is an object X together with a family π = (π∞,i )i∈I of morphisms
π∞,i : X → Xi satisfying πi,j ◦ π∞,i = π∞,j for all i, j ∈ I with i ≤ j, and such that
(X, π) is universal with these properties.
The following definition of semiprojectivity was introduced by Blackadar in the
sequential setting for the category of C ∗ -algebras; see [Bla85a, Definition 2.1]. See
also [Bla85b, Definition 2.10]. The general (nonsequential) definition has also been
considered in [CLT18].
Definition 3.4. Let C be a category, and let X be an object in C. Then X is said
to be semiprojective if for every inductive system (I, Yi , πi,j ) in C with connecting
epimorphisms and for which the direct limit lim Yi exists, and for every morphism
−→
ϕ : X → lim Yi , there exist i ∈ I and a morphism ϕ̃ : X → Yi such that πi,∞ ◦ ϕ̃ = ϕ.
−→
The morphism ϕ̃ is called a partial lift of ϕ.
Dually, X is said to be semiinjective if for every inverse system (I, Yi , ιi,j ) in C
with connecting monomorphisms and for which the inverse limit lim Yi exists, and
←−
for every morphisms ϕ : lim Yi → X, there exist i ∈ I and a morphism ϕ̃ : Yi → X
←−
such that ϕ̃ ◦ ι∞,i = ϕ. The morphism ϕ̃ is called an partial extension of ϕ.
Given objects X and Y in a category, we say that X is a retract of Y if there
exist morphisms α : X → Y and β : Y → X with β ◦ α = idX . The proof of the
following result is straightforward.
Lemma 3.5. Let C be a category, and let X and Y be objects in C. Then:
(1) If X is a retract of Y , and if Y is (semi)projective, then so`is X.
(2) If X and Y are (semi)projective, then so is the coproduct X Y (assuming
it exists).
Remarks 3.6. (1) Under suitable countability assumptions, it is usually enough
to consider sequential direct limits in Definition 3.4. For example, a countable
group G is semiprojective if and only if for every sequential direct system (N, Dk ) of
6
HANNES THIEL
countable groups and with connecting epimorphisms, every morphism G → limk Dk
−→
has a partial lift.
To show the backward implication, let (I, Hi , πi,j ) be an arbitrary direct system
with connecting epimorphisms in Gp, and let ϕ : G → limi Hi be a morphism. Then
−→
there exist an increasing sequence of indices i(0) ≤ i(1) ≤ . . . in I, and countable
subgroups Dk ⊆ Hi(k) for all k ∈ N, such that the restriction of πi(k),i(k+1) to Dk
maps onto Dk+1 , and such that ϕ factors through limk Dk . This means that there
−→
exists a morphism ψ : G → limk Dk such that ϕ = γ◦ψ, where γ : limk Dk → limi Hi
−→
−→
−→
is the morphism obtained by the universal property of limk Dk applied for the
−→
morphisms (πi(k),∞ )|Dk : Dk → limi Hi . Given a partial lift ψ̃ : G → Dk for ψ, we
−→
obtain a partial lift for ϕ by composing ψ̃ with the inclusion Dk ⊆ Hi(k) .
The situation is shown in the following commutative diagram:
/ Hi(k)
: Dk
ψ̃
G
ψ
/ lim
D
−→k∈N k
γ
πi(k),∞
/ lim H .
3 −→i∈I i
ϕ
(2) Similarly, a separable C ∗ -algebra A is semiprojective if and only if every ∗ homomorphism from A to the direct limit of a sequential direct system of separable
C ∗ -algebras with surjective connecting maps has a partial lift.
(3) The concept of direct and inverse limits in a category can be generalized
to filtered (co)limits; see [Bor94, Section 2.13, p.75ff]. In some categories, it may
be appropriate to modify Definition 3.4 and consider filtered (co)limits instead of
direct and inverse limits.
Proposition 3.7. A group is semiprojective if and only if it is the retract of the
free product of a finitely presented group and a free group.
Proof. Let us show the backward implication. Using Lemma 3.5, it remains to
prove that every finitely presented group H is semiprojective. Choose a finitely
generated free group F and a finitely generated normal subgroup N ⊳ F such that
H is isomorphic to F/N . We identify H with F/N . Let r1 , . . . , rn ∈ N be a set of
elements that generate N as a normal subgroup of F .
To show that H is semiprojective, let (I, Hi , πi,j ) be a direct system in Gp with
connecting epimorphisms, and let ϕ : H → lim Hi be a morphism. Let γ : F → H
−→
be the quotient map. Since F is free, and hence projective, we can choose i0 ∈ I
and a lift ψ : F → Hi of ϕ ◦ γ. Given k ∈ {1, . . . , n}, we have
πi,∞ ψ(rk ) = (πi,∞ ◦ ψ)(rk ) = (ϕ ◦ γ)(rk ) = 1,
which allows us to choose ik ≥ i0 such that πi,ik ψ(rk ) = 1. Choose i′ ∈ I with
i1 , . . . , in ≤ i′ . Then πi,i′ maps ψ(r1 ), . . . , ψ(rn ) to 1. It follows that ψ(N ) ⊆
ker(πi,i′ ). Thus, πi,i′ ◦ γ factors through H, which provides the desired partial lift.
To show the forward implication, assume that G is a semiprojective group.
Choose a set X and a surjective group homomorphism γ : F (X) → G, where F (X)
denotes the free group on the set of generators X. Set N := ker(γ). Given a subset
A ⊆ X, we identify F (A) in the obvious way with a subgroup of F (X). Set
I := (A, B) : A ⊆ X finite, B ⊆ N finite, B ⊆ F (A) .
For (A, B) ∈ I set G(A,B) := F (A)/hBi, where hBi denotes the normal subgroup
of F (A) generated by B. Let γ(A,B) : F (A) → G(A,B) denote the quotient map.
SEMIPROJECTIVITY AND SEMIINJECTIVITY IN DIFFERENT CATEGORIES
7
We define a partial order on I by setting (A′ , B ′ ) ≤ (A, B) if A′ ⊆ A and B ′ ⊆ B.
Then I is an upward directed set. For (A, B) ∈ I we set
I(A,B) := (Ã, B̃) ∈ I : (Ã, B̃) ≥ (A, B) ,
and
H(A,B) := G(A,B) ⋆ F (X × I(A,B) ).
′
(A,B)
′
Given (A , B ) ≤ (A, B), let us define a surjective morphism π(A′ ,B ′ ) from
H(A′ ,B ′ ) to H(A,B) . The inclusion F (A′ ) → F (A) induces a morphism G(A′ ,B ′ ) →
G(A,B) . We let R be the subset of X × I(A′ ,B ′ ) such that X × I(A′ ,B ′ ) is the
disjoint union of A × {(A′ , B ′ )}, X × I(A,B) , and R. The first coordinate projection A × {(A′ , B ′ )} → A induces a morphism F (A × {(A′ , B ′ )}) → F (A) that
we postcompose with γ(A,B) to obtain a surjective morphism F (A × {(A′ , B ′ )}) →
(A,B)
G(A,B) . We define π(A′ ,B ′ ) as the free product of the morphisms G(A′ ,B ′ ) → G(A,B) ,
F (A × {(A′ , B ′ )}) → G(A,B) , the identity on F (X × I(A,B) ) and the trivial map
F (R) → {1}. This is shown in the following diagram:
H(A′ ,B ′ )
:=
(A,B)
π(A′ ,B ′ )
H(A,B)
:=
F (A × {(A′ , B ′ )}) ⋆ F (X × I(A,B) )
♣♣
♠♠
∼
=♠♠♠♠
♣♣♣
♣
♠
♣
♠
♣
♠
v♠♠♠
x♣x ♣♣
G(A,B) ⋆ F (X × I(A,B) )
G(A′ ,B ′ )
⋆
⋆
F (R)
(A,B)
It is straightforward to verify that the maps π(A′ ,B ′ ) are surjective and define an
inductive system (over the index set I). Moreover, there is a natural isomorphism
ϕ : G → lim(A,B)∈I G(A,B) .
−→
Using that G is semiprojective, we find (A, B) ∈ I and a partial lift ϕ̃ : G →
G(A,B) of ϕ. This shows that G is a retract of G(A,B) , which is the free product of
the finitely presented group H(A,B) and a free group.
Corollary 3.8. Every finitely presented group is semiprojective. Moreover, every
group is a direct limit of semiprojective groups (and one may also assume that the
connecting morphisms are surjective).
Proposition 3.9. The trivial group is the only semiinjective object of Gp.
Proof. Let G be a semiinjective group. We show that G is injective, whence it is
trivial as noted in Examples 3.2.
To show that G is injective, let H ⊆ K be an inclusion of groups, and let
ϕ : H → G be a morphism. We let ⋆n∈N K denote the free product of countably
many copies of K, and for each m ∈ N we let ιm : K → ⋆n∈N K be the natural
inclusion. The amalgamated free product ⋆n∈N,H K is defined as the quotient of
⋆n∈N K by the normal subgroup generated by ιn (h)ιm (h)−1 , for n, m ∈ N and
h ∈ H. For each m ∈ N we let ⋆n≥m,H K denote the subgroup of ⋆n∈N,H K generated
by all except the first m copies of K. This defines a decreasing sequence of subgroups
whose intersection is isomorphic to H.
Since G is semiinjective, there exist m and a partial extension ϕ̃ : ⋆n≥m,H K → G.
Composing with the morphism ιm : K → ⋆n≥m,H K, we obtain a morphism K → G
that extends ϕ, showing that G is injective.
The following definition is due to Borsuk. For more details we refer to the books
[Bor67] and [Hu65]. Recall that a retract from a topological space Y to a subspace
X is a continuous map r : Y → X that satisfies r(x) = x for all x ∈ X.
Definition 3.10. Let X be a metric space. Then:
8
HANNES THIEL
(1) X is called an absolute retract if whenever X is embedded as a closed subset
of another metric space Y , there exists a retract Y → X.
(2) X is called an absolute neighborhood retract if whenever X is embedded as
a closed subset of another metric space Y , there exist a neighborhood U of
X in Y and a retract U → X.
The equivalence between (1) and (2) in the following result is a standard fact
about absolute neighborhood retracts; see for example [Hu65, Theorem III.3.1,
III.3.2, p.83f]. The equivalence between (2) and (3) follows using that compact,
metric spaces are normal.
Proposition 3.11. Let X be a compact, metric space. Then the following are
equivalent:
(1) X is an absolute (neighborhood) retract.
(2) Given a compact, metric space Y and a closed subset Z ⊆ Y , and given a
continuous map ϕ : Y → X, there exists an extension of ϕ to a continuous
map ϕ̃ : Y → X (there exists a closed neighborhood C of Z in Y and an
extension ϕ̃ : C → X).
(3) X is a (semi)injective object in the category CMetr.
4. C ∗ -algebras
A C ∗ -algebra is a Banach algebra A with an involution such that ka∗ ak =
kak2 for all a ∈ A. A ∗ -homomorphism between C ∗ -algebras is a multiplicative,
∗
-preserving, linear map. We let C∗ denote the category of C ∗ -algebras and ∗ homomorphisms. The naive concretization of C∗1 associates to every C ∗ -algebra its
(usual) underlying set. However, this functor C∗ → Set is not representable.
Nevertheless, C∗ has a generator. Indeed, let G := C ∗ (x : kxk ≤ 1) be the
universal C ∗ -algebra generated by a contraction. Given a C ∗ -algebra A, there is a
natural bijection between HomC∗ (G, A) and the elements in the unit ball of A. To
see that G is a generator, we note that two ∗ -homomorphisms A → B are equal if
and only if they agree on the unit ball of A. Hence, a morphism in C∗ is mono if
and only if it is injective.
The category C∗ has no cogenerator. (The proof is analogous to that for Gp.)
Nevertheless, a morphism in C∗ is epi if and only if it is surjective. As for Gp, the
forward implication is not obvious; see [Rei70, Proposition 2] and [HN95].
It follows that isomorphisms in C∗ are exactly the bijective ∗ -homomorphism,
also called ∗ -isomorphisms, and such maps are automatically isometric.
For simplicity, we will restrict attention to the subcategory SC∗1 of unital, separable C ∗ -algebras and unital ∗ -homomorphisms. We let AbSC∗1 denote the full
subcategory of SC∗1 of abelian, unital, separable C ∗ -algebras.
Examples 4.1. (1) Given a Hilbert space H, the algebra B(H) of bounded linear
operators on H, equipped with the operator norm and the natural involution, is
a unital C ∗ -algebra. By the Gelfand-Naimark theorem, every C ∗ -algebra is ∗ isomorphic to norm-closed ∗ -subalgebra of B(H) for some Hilbert space H; see
[Bla06, Corollary II.6.4.10, p.109].
(2) For the Hilbert space H = ℓ2 ({1, 2, . . . , n}), we obtain that B(H) ∼
= Mn (C),
the algebra of complex n × n-matrices, has the structure of a C ∗ -algebra. By the
Artin-Weddenburn theorem, a C ∗ -algebra is finite-dimensional (as a complex vector
space) if and only if it is isomorphic to a finite direct sum of matrix algebras.
(3) Let X be a compact, metric space. Set
C(X) := f : X → C : f is continuous ,
SEMIPROJECTIVITY AND SEMIINJECTIVITY IN DIFFERENT CATEGORIES
9
equipped with pointwise addition, multiplication and involution, and with the norm
kf k := sup |f (x)| : x ∈ X ,
for f ∈ C(X). Then C(X) is a unital, commutative, separable C ∗ -algebra.
If Y is another compact, metric space, and if ϕ : X → Y is a continuous map,
then ϕ∗ : C(Y ) → C(X) given by ϕ∗ (f ) := f ◦ ϕ is a unital ∗ -homomorphism. This
defines a contravariant functor C( ) : CMetr → AbSC∗1 .
Proposition 4.2 (Gelfand; [Bla06, Theorem II.2.2.6, p.61]). Every abelian, unital, separable C ∗ -algebra is isomorphic to C(X) for some compact, metric space X.
Moreover, the functor C( ) : CMetr → AbSC∗1 defines a (contravariant) equivalence of categories.
Remark 4.3. Projectivity and semiprojectivity of C ∗ -algebras is defined with respect to the category C∗ of all C ∗ -algebras. When considering the category C∗1 , the
notion of projectivity changes. On the other hand, a unital C ∗ -algebra is semiprojective (in C∗ ) if and only if it is semiprojective in C∗1 .
Remark 4.4. Let X be a compact, metric space. If C(X) is (semi)projective in C∗1 ,
then it is also (semi)projective in AbSC∗1 - since there are fewer lifting problems
to solve. Using the (contravariant) equivalence between AbSC∗1 and CMetr from
Proposition 4.2, we deduce that C(X) is (semi)projective in AbSC∗1 if and only if
X is (semi)injective in CMetr. Thus, if C(X) is (semi)projective in C∗1 , then X
is an absolute (neighborhood) retract.
Examples 4.5. (1) C = C(pt) and C([0, 1]) are projective in C∗1 .
(2) Set S 1 := {z ∈ C : |z| = 1}, the unit circle. The C ∗ -algebra C(S 1 ) is
semiprojective, but not projective in C∗1 .
(2) Set D2 := {z ∈ C : |z| ≤ 1}, the two-disc. Then D2 is an absolute retract.
However, the C ∗ -algebra C(D2 ) is not even semiprojective.
(3) If X is a compact, metric space with S 1 ⊆ X, then C(X) is not projective
in C∗1 . Indeed, assume that ι : S 1 ֒→ X is an embedding. Let ϕ : S 1 → D2 be the
inclusion map. Since D2 is an absolute retract, there exists an extension of ϕ to
a continuous map ϕ̃ : X → D2 . This is shown in the left commutative diagram
below.
DO 2 b❉
ϕ
S
T
f˜
❉
π
❉
1
ι
❉
/X
/ C(S 1 )
C(D2 )
▲▲▲ ∗
s9
ι∗ sss
▲▲ϕ̃▲
s
▲▲
s
%
ss
C(X).
ϕ∗
We let T denote the Toeplitz algebra, that is, the sub-C ∗ -algebra of B(ℓ2 (N))
generated by the unilateral shift on ℓ2 (N). Sending this shift to the identity map
id ∈ C(S 1 ) induces a surjective ∗ -homomorphism π : T → C(S 1 ). The situation is
shown in the right commutative diagram above.
If C(X) were projective in C∗1 , then there would exist a lift ψ : C(X) → T for ι∗ .
Then ψ ◦ ϕ̃∗ is a lift for ϕ∗ . The image of the identity map id ∈ C(D2 ) under ψ ◦ ϕ̃∗
is a normal element in T that lifts the unitary id ∈ C(S 1 ). Using the Fredholm
index one can show that no such lift exists, showing that C(X) is not projective in
C∗1 .
(4) If X is a compact, metric space with D2 ⊆ X, then C(X) is not semiprojective in C∗1 . This is shown similarly as in (3); see [ST12, Remark 3.3].
10
HANNES THIEL
Let X be a compact, metric space such that C(X) is (semi)projective in C∗1 . As
observed in Remark 4.4, it follows that X is an absolute (neighborhood) retract.
The above examples show that the converse does not hold.
Conjecture 4.6 (Blackadar, [Bla06, II.8.3.8, p.163]). Let X be a compact, metric
space. Then C(X) is semiprojective if and only if X is an absolute neighborhood
retract with dim(X) ≤ 1.
The analog of this conjecture for projectivity was solved by Chigogidze and
Dranishnikov:
Theorem 4.7 ([CD10, Theorem 4.3]). Let X be a compact, metric space. Then
C(X) is projective in C∗1 if and only if X is an absolute retract with dim(X) ≤ 1.
Let us sketch the proof of the forward implication of Theorem 4.7. Assume that
C(X) is projective in C∗1 . Then X is an absolute retract; see Remark 4.4. To
show that dim(X) ≤ 1, assume that dim(X) ≥ 2. By [CD10, Proposition 3.2], it
follows that S 1 ⊆ X. As sketched in Examples 4.5(3), this implies that C(X) is
not projective.
We remark that the topological result that S 1 embeds into X uses both that
X is an absolute (neighborhood) retract and that dim(X) ≥ 2. Indeed, the space
[0, 1] is an example of an absolute retract that does not admit an embedding of the
circle. On the other hand, there exist compact metric spaces with dim(X) = ∞
such that every closed subset of X satisfies either dim(X) = 0 or dim(X) = ∞.
In particular, such a space does not admit an embedding of the circle. The point
is that such a behaviour is not possible for ‘well-behaved’ spaces such as absolute
(neighborhood) retracts.
Together with Adam Sørensen, we confirmed Blackadar’s conjecture.
Theorem 4.8 ([ST12, Theorem 1.2]). Let X be a compact, metric space. Then
C(X) is semiprojective if and only if X is an absolute neighborhood retract with
dim(X) ≤ 1.
Let us sketch the proof of the forward implication of Theorem 4.8. Assume
that C(X) is semiprojective in C∗1 . Then X is an absolute neighborhood retract;
see Remark 4.4. To show that dim(X) ≤ 1, assume that dim(X) ≥ 2. If we
could deduce that D2 embeds into X, then we would conclude that C(X) is not
semiprojective as mentioned in Examples 4.5(3).
The problem is that dim(X) ≥ 2 does not imply D2 ⊆ X. Indeed, Bing and
Borsuk constructed an absolute retract Y such that dim(Y ) = 3, but such that D2
does not embed into Y ; see [BB64]. Thus, we cannot assume that a disc embeds
into X. Instead, we use the following topological result:
Lemma 4.9 ([ST12, Remark 3.4]). Let X be a compact, metric space that is an
absolute neighborhood retract. Assume that dim(X) ≥ 2. Then X contains either
the space C1 of disjoint ‘smaller and smaller circles’, the Hawaiian earrings space
C2 , or the space C3 that is a variant of the Hawaiian earrings, each given as subsets
of R2 : (S(x, r) denotes the circle of radius r around x.)
[
[
C1 = (0, 0) ∪
S((1/2k , 0), 1/(4 · 2k )), C2 =
S((1/2k , 0), 1/2k ),
k≥1
k≥1
[
C3 = (x, x), (x, −x) : x ∈ [0, 1] ∪
{1/k} × [−1/k, 1/k].
k≥1
A boosted version of the argument in Examples 4.5(3) shows the following result,
which together with Lemma 4.9 shows the forward implication of Theorem 4.8.
SEMIPROJECTIVITY AND SEMIINJECTIVITY IN DIFFERENT CATEGORIES
(a) Space C1
(b) Space C2
11
(c) Space C3
Lemma 4.10 ([ST12]). Let X be a compact, metric space. If X contains any of
the spaces C1 , C2 , or C3 from Lemma 4.9, then C(X) is not semiprojective.
References
[BB64]
[Bla85a]
[Bla85b]
[Bla06]
[Bor94]
[Bor67]
[CD10]
[CLT18]
[Gar17]
[HN95]
[Hu65]
[Lin70]
[Nog07]
[OP12]
[Rei70]
[ST12]
[Thi11]
R. H. Bing and K. Borsuk, A 3-dimensional absolute retract which does not contain
any disk, Fund. Math. 54 (1964), 159–175. MR 0161312. Zbl 0118.18002.
B. Blackadar, Noncommutative shape theory, in Operator algebras and their connections with topology and ergodic theory (Buşteni, 1983), Lecture Notes in Math. 1132,
Springer, Berlin, 1985, pp. 38–45. MR 799562. Zbl 0618.46058.
B. Blackadar, Shape theory for C ∗ -algebras, Math. Scand. 56 (1985), 249–275.
MR 813640 (87b:46074). Zbl 0615.46066.
B. Blackadar, Operator algebras, Encyclopaedia of Mathematical Sciences 122,
Springer-Verlag, Berlin, 2006, Theory of C ∗ -algebras and von Neumann algebras,
Operator Algebras and Non-commutative Geometry, III. MR 2188261 (2006k:46082).
Zbl 1092.46003.
F. Borceux, Handbook of categorical algebra. 1, Encyclopedia of Mathematics and its
Applications 50, Cambridge University Press, Cambridge, 1994, Basic category theory.
MR 1291599 (96g:18001a).
K. Borsuk, Theory of retracts, Monografie Matematyczne, Tom 44, Państwowe
Wydawnictwo Naukowe, Warsaw, 1967. MR 0216473. Zbl 0153.52905.
A. Chigogidze and A. Dranishnikov, Which compacta are noncommutative ARs?,
Topology Appl. 157 (2010), 774–778. MR 2585410 (2011a:54017). Zbl 1187.54017.
A. Chigogidze, T. A. Loring, and H. Thiel, Nonseparable semiprojective C ∗ -algebras,
in preparation, 2018.
E. Gardella, Crossed products by compact group actions with the Rokhlin property,
J. Noncommut. Geom. 11 (2017), 1593–1626. MR 3743232.
K. H. Hofmann and K.-H. Neeb, Epimorphisms of C ∗ -algebras are surjective, Arch.
Math. (Basel) 65 (1995), 134–137. MR 1338245. Zbl 0829.46044.
S.-t. Hu, Theory of retracts, Wayne State University Press, Detroit, 1965. MR 0181977.
Zbl 0145.43003.
C. E. Linderholm, A group epimorphism is surjective, Amer. Math. Monthly 77 (1970),
176–177. MR 0255643. Zbl 0194.03501.
M. Nogin, A short proof of Eilenberg and Moore’s theorem, Cent. Eur. J. Math. 5
(2007), 201–204. MR 2287720. Zbl 1120.20054.
H. Osaka and N. C. Phillips, Crossed products by finite group actions with the Rokhlin
property, Math. Z. 270 (2012), 19–42. MR 2875821. Zbl 1244.46032.
G. A. Reid, Epimorphisms and surjectivity, Invent. Math. 9 (1969/1970), 295–307.
MR 0260829. Zbl 0191.13503.
A. P. W. Sørensen and H. Thiel, A characterization of semiprojectivity for commutative C ∗ -algebras, Proc. Lond. Math. Soc. (3) 105 (2012), 1021–1046. MR 2997045.
Zbl 1266.46043.
H. Thiel, Inductive limits of projective C ∗ -algebras, preprint (arXiv:1105.1979
[math.OA]), 2011.
12
HANNES THIEL
Hannes Thiel Mathematisches Institut, Fachbereich Mathematik und Informatik der
Universität Münster, Einsteinstrasse 62, 48149 Münster, Germany.
E-mail address: [email protected]
URL: www.math.uni-muenster.de/u/hannes.thiel/
| 4math.GR
|
Tensor Decompositions for Modeling
Inverse Dynamics
Stephan Baier ∗ Volker Tresp ∗∗
∗
arXiv:1711.04683v1 [cs.LG] 13 Nov 2017
Ludwig Maximilian University of Munich, Oettingenstr. 67, 80538
Munich (e-mail: [email protected]).
∗∗
Siemens AG and Ludwig Maximilian University of Munich,
Otto-Hahn-Ring 6, 81739 Munich (e-mail: [email protected])
Abstract:
Modeling inverse dynamics is crucial for accurate feedforward robot control. The model
computes the necessary joint torques, to perform a desired movement. The highly non-linear
inverse function of the dynamical system can be approximated using regression techniques. We
propose as regression method a tensor decomposition model that exploits the inherent threeway interaction of positions × velocities × accelerations. Most work in tensor factorization has
addressed the decomposition of dense tensors. In this paper, we build upon the decomposition
of sparse tensors, with only small amounts of nonzero entries. The decomposition of sparse
tensors has successfully been used in relational learning, e.g., the modeling of large knowledge
graphs. Recently, the approach has been extended to multi-class classification with discrete input
variables. Representing the data in high dimensional sparse tensors enables the approximation
of complex highly non-linear functions. In this paper we show how the decomposition of sparse
tensors can be applied to regression problems. Furthermore, we extend the method to continuous
inputs, by learning a mapping from the continuous inputs to the latent representations of the
tensor decomposition, using basis functions. We evaluate our proposed model on a dataset with
trajectories from a seven degrees of freedom SARCOS robot arm. Our experimental results
show superior performance of the proposed functional tensor model, compared to challenging
state-of-the art methods.
Keywords: Tensor modeling, tensor decomposition, inverse dynamics, robot dynamics,
supervised machine learning
1. INTRODUCTION
Within model-based robot control, an inverse dynamics
model is used to compute the necessary joint torques of the
robot’s motors for the execution of a desired movement.
The feedforward control command can be calculated using
the rigid-body formulation uF F = M (q)q̈ + F (q, q̇), with
q, q̇, q̈ being vectors of joint positions, joint velocities, and
joint accelerations. However, in practice many nonlinearities such as friction or actuator forces need to be taken
into account. Thus, methods modeling uF F = f (q, q̇, q̈)
using non-linear regression techniques have shown superior
performance in inferring the required joint torques for
feedforward robot control. The parameters of the function
f are estimated offline using collected trajectories of the
robot. Craig (2005); Nguyen-Tuong et al. (2008); Nakanishi et al. (2005).
Tensor models have been applied successfully in many
application areas, e.g., relational learning, multilinear time
invariant systems, factor analysis, and spatio-temporal
analysis, see Nickel et al. (2011); Pangalos et al. (2013b);
Mørup et al. (2006); Bahadori et al. (2014). Most literature on tensor modeling, however, is concerned with the
decomposition of dense tensors, i.e., most of the elements
in the tensor are nonzero. Models for sparse tensors have
mainly become popular for the application of modeling
large knowledge graphs, such as Yago, DBpedia, and Freebase, see Nickel et al. (2011); Suchanek et al. (2007);
Auer et al. (2007). In these models, the elements of the
tensor represent all possible triple combinations of entities
and relations in the knowledge graph. Only elements that
represent known facts from the knowledge graph are set
to one. This results in a very sparse tensor, where the
vast majority of elements are zero. Recently, the approach
has been extended to higher order tensors for the task of
classifying discrete sensor data, see Baier et al. (2016). The
tensor represents the space of all possible combinations
of sensor values. By learning a representation for each
possible value of all sensors, the decomposition allows for
approximating highly non-linear functions.
In this paper we build upon the approach of decomposing
sparse tensors, and apply it to inverse system identification. Our model exploits the inherent three-way interaction of positions × velocities × accelerations. We first
show how the method can be applied to regression tasks.
Furthermore, we extend the approach to continuous inputs, by including basis functions that map the continuous
inputs to the latent representations of the tensor decompositions. In this way, we retrieve a functional version
of tensor decompositions. The basis functions also imply
smoothness on the inputs, such that the model is able to
generalize well, in spite of the extreme sparsity. By using
multivariate basis functions we can group inputs, such
that the dimensionality of the tensor decomposition can
be reduced. In our inverse dynamics model we group the
joint positions, velocities, and accelerations of all degrees
of freedom of the robot, resulting in a tensor of order three.
This makes the powerful Tucker decomposition applicable
to the problem.
We evaluate our model on a dataset of a seven degrees of
freedom SARCOS robot arm that was introduced in Vijayakumar and Schaal (2000). An inverse dynamics model
is learned based on collected trajectories, and its performance is evaluated on a 10 percent test set. The results
show that our model outperforms a number of competitive
baseline methods, such as linear regression, radial basis
function networks (RBF-networks), and support vector
regression. Furthermore, the Tucker model shows superior
performance over a PARAFAC model.
The paper is structured as follows. The next section gives
an overview of related work. Section 2 shows how the
factorization of sparse tensors can be utilized for regression
problems, and how the tensor decompositions can be
extended to continuous inputs, using basis functions. In
Section 3 we describe a functional Tucker decomposition
for the task of modeling inverse dynamics. Related work is
discussed in Section 4. Section 5 presents the experimental
evaluation. Finally, we conclude our work in Section 6.
2. TENSOR DECOMPOSITIONS USING BASIS
FUNCTIONS
In this section we first show how the decomposition of
sparse tensors can be applied to regression problems with
discrete input variables. We then extend the model to
continuous inputs, by using basis functions, which map
the continuous input to the latent representations of the
tensor decompositions.
Y(v1 , v2 , ..., vS ) ≈
r̃
X
g(r)·A1 (v1 , r)·A2 (v2 , r)·. . .·AS (vS , r).
r=1
(2)
with g ∈ Rr̃ . As PARAFAC only models the diagonal of
the core tensor, its parameters scale linearly with the order
of the tensor; see Harshman (1970).
2.2 Discrete Input Regression
We consider a regression problem with S ∈ N discrete
input variables. Each of the input variables vi for i ∈
{1, ..., S} assumes one out of Fi ∈ N discrete values.
Furthermore, we consider a dependent variable y. We
model a regression function for a dataset of N training
examples {y j , (v1j , ..., vSj )}N
j=1 .
All training examples are mapped to a sparse tensor
Y ∈ RF1 ,...,FS . The tensor is filled with
Y(v1j , ..., vSj ) = y j ∀j ∈ {1, ..., N }.
(3)
The remaining entries of the tensor, which do not occur
in the training data, are left unknown. This results in Y
being a very sparse tensor.
The tensor Y is approximated using a low-rank tensor
decomposition, e.g., the PARAFAC decomposition see
equation 2. Using low ranks for r̃, the approximation
results in a dense tensor Φ. It describes the outcome
y for all combinations of the input variables (v1 , ..., vS ).
However, it would be impossible to compute and store the
whole approximated tensor Φ; thus, only the parameters
of the decomposition are stored. When predicting y for a
new set of input variables, the representations for that
tuple are indexed, and the approximation is computed
on demand. In principle any tensor decomposition can
be used for the approximation. However, in practice only
few decompositions, such as PARAFAC and Tensor Train
are scale-able to many dimensions, see Harshman (1970);
Oseledets (2011).
2.1 Tensor Decompositions
2.3 Continuous Inputs
Tensor decompositions are a generalization of low rank
matrix factorizations to higher order tensors. There are
multiple ways of decomposing a higher order tensor.
The proposed model so far only works for a discrete input
space. Furthermore, it does not imply any smoothness on
the values of the input variables. Although, this makes
it a powerful, highly non-linear model, it is prone to
overfitting. If the input values follow a natural ordering,
or if they are discretized from a continuous scale, the
model requires many more training examples to learn the
smoothness implicitly. To introduce smoothness explicitly,
and to extend the model to continuous inputs, we use
smooth basis functions for the latent parameters of the
decomposition. Instead of indexing the latent representation from a matrix, their values are computed using
basis functions. For example, all Ai in equation 2 can be
modeled using a radial basis function
Ai (vi , ri ) = exp (−γri kµri − vi k2 ).
(4)
This allows for continuous inputs vi ∈ R. The latent
representation is now modeled by the similarity of the
input to the center of the radial basis function. In this
way, similar inputs induce similar representations. The
parameters of the basis function are optimized during
training, to yield optimal regression results. Also a mixture
The full Tucker decomposition factorizes a tensor Y ∈
Rd1 ×···×dS into S matrices, including latent representations for all fibers in each mode. The tensor elements are
expressed by the interaction of the latent representations,
weighted by a core tensor G ∈ Rr̃×···×r̃ such that
Y(v1 , . . . , vS ) ≈
r̃
X
G(r1 , . . . , rS ) · A1 (v1 , r1 )·
r1 ,...,rS
(1)
A2 (v2 , r2 ) · . . . · AS (vS , rS )
di ×r̃
with Ai ∈ R
. The full Tucker decomposition does
not scale to high dimensions, as the core tensor G grows
exponentially with the dimensionality of the tensor; see
Tucker (1965).
A special case of the full Tucker decomposition is the
PARAFAC decomposition, where the core tensor G is
diagonal. All other interactions are left out, such that
positions
...
...
velocities
torques
accelerations
Fig. 1. Inverse dynamics model using a functional Tucker decomposition. The output tensors and the representation
matrices are replaced by functions (illustrated with dashed lines). The representations are computed given the
continuous inputs using Gaussian kernels.
of discrete and continuous inputs can easily be modeled, by
applying the basis functions only to the continuous inputs,
and learning representation matrices for the discrete input
variables.
It is also possible to group multiple inputs together into
one tensor mode, such that vi ∈ Rm , where m ∈ N
denotes the number of grouped inputs. In this way, the
representation of a tensor mode is calculated given a
vector of continous inputs. The grouping of input variables
reduces the dimensionality of the tensor decomposition,
and thus the number of free parameters.
3. APPLICATION TO INVERSE DYNAMICS
In the following we describe how the continuous tensor
decomposition proposed in section 2 can be applied to
inverse dynamics modeling.
3.1 Functional Tucker Decomposition
We describe a functional Tucker model for the approximation of the joint torques, necessary to perform a movement
of a robot arm. Figure 1 shows the model schematically.
We consider a robot with C ∈ N degrees of freedom (DoF).
In the following we denote the vectors p, ṗ, p̈, describing the
desired positions, velocities, and accelerations for each of
the c DoFs, as x1 , x2 , x3 ∈ Rc for syntactic reasons. The
vector y ∈ Rc describes the corresponding joint torques.
We model the function y = f (x1 , x2 , x3 ) using a functional
tensor decomposition model. Each input vector is modeled
by one dimension in the tensor decomposition, resulting in
third-order tensors Y, which describe the joint torques.
Each element of the vector y is modeled in a separate
model. The resulting three-dimensional tensors of the form
positions × velocities × accelerations, are then factorized
using the Tucker decomposition with limited rank, resulting in a tensor Φ ≈ Y, such that
Φ(x1 , x2 , x3 ) =
r̃
X
G(r1 , r2 , r3 ) · A1 (x1 , r1 ) · A2 (x2 , r2 )
r1 ,r2 ,r3
·A3 (x3 , r3 ).
(5)
A1 to A3 are functions, which map from the c-dimensional
input to the latent representations of the Tucker model.
We model the representations using multivariate Gaussian
kernels, such that
Ai (xi , ri ) = exp −(µri − xi )T Dri (µri − xi )
(6)
∀i ∈ {1, 2, 3},
with µri ∈ Rc representing the centers and Dri ∈ Rc×c
weighting the distance from the centers in the c dimensional input space. The closer a data point is to the center
of a basis function, the higher is its activation. Thus, the
centers of the basis functions can be seen as landmarks
in the input space. All three-way interactions, between
the representations of the three input dimensions, are
explicitly modeled and weighted by the elements of the
core tensor G.
3.2 Model Training
For training the model, we take a maximum likelihood
approach. We minimize the negative log-likelihood of the
collected dataset {y j , (xj1 , xj2 , xj3 )}N
j=1 as
l = −log
N
X
p(y j |xj1 , xj2 , xj3 , Θ),
(7)
j=1
where Θ includes the parameters of the decomposition and
the basis functions. Assuming a Gaussian distribution, we
get the squared error cost function
Table 1. Normalized mean squared error for all 7 degrees of freedom in percent. Mean and
standard deviation of ten random data splits.
Method
Linear Regression
RBF-Network Regression
Support Vector Regression
Functional-Tucker
Functional-PARAFAC
C=
N
X
DoF 1
6.80
2.64
0.88
0.59
1.64
DoF 2
11.62
1.79
0.67
0.28
1.14
(Y(xj1 , xj2 , xj3 ) − Φ(xj1 , xj2 , xj3 ))2 .
DoF 3
10.82
1.01
0.43
0.46
0.61
(8)
i=1
Note, that the cost function considers only nonzero elements of the tensor, i.e., the sparsity of the tensor is
exploited. We minimize equation 8 using gradient descent.
In experiments, we found the stochastic optimization algorithm Adam, see Kingma and Ba (2014), which adopts the
learning rate automatically for each parameter, to work
best for this task. The sampling of stochastic mini-batches
for each update has also shown advantageous, for speeding
up training.
We initialize the centers of the Gaussian kernel in a
preprocessing step, using three k-means clusterings, such
that
N
r̃ X
X
kxji − µri k2
(9)
Ji =
i
j=1
are minimized for i ∈ {1, . . . , 3}, see Lloyd (1982). All
matrices D are initialized with the identity matrix. The
elements of the core tensor G are initialized randomly
with a Gaussian distribution of mean zero and standard
deviation 0.05. While training all parameters are further
optimized. We implemented the model using the computational python library Theano, see Theano Development
Team (2016).
4. RELATED WORK
Multiway data analysis has found applications in a number
of different areas, such as signal processing, neuroscience,
and data mining, see Cichocki (2014); Mørup et al. (2006);
Harshman (1970); Kolda and Bader (2009). Recently,
tensor models also have found applications in control
engineering such as for modeling hybrid systems, see
Pangalos et al. (2013a) and multilinear dynamical systems,
see Rogers et al. (2013). Furthermore, tensor methods have
been applied to Boolean networks, see Cheng et al. (2010)
and pneumatic models, see Gróf et al. (2010).
The factorization of sparse matrices has become popular in
recommendation systems, especially due to its success in
the Netflix challenge, see Koren et al. (2009). Extensions
to higher order tensors can be found in the modeling
of large knowledge bases, such as Yago, DBpedia, or
Freebase, see Nickel et al. (2011, 2015). The multi-graphs
have been modeled using sparse three-dimensional tensors
and decompositions such as RESCAL, see Nickel et al.
(2011, 2015). The approach of factorizing sparse tensors
has further been exploited in Baier et al. (2016). Here, the
decomposition of sparse tensors is applied to multi-class
classification with discrete input features.
Tensor regression methods are concerned with the regression of high dimensional data, structured in a multidimen-
DoF 4
5.81
0.41
0.15
0.24
0.32
DoF 5
12.81
4.07
1.04
1.03
1.30
DoF 6
22.59
3.91
0.72
0.91
1.17
DoF 7
6.73
1.17
0.34
0.31
0.50
Mean ± std in %
11.03 ± 0.26
2.14 ± 0.19
0.60 ± 0.28
0.55 ± 0.24
0.96 ± 0.22
sional array. Tensor methods allow for efficient modeling
where traditional methods are often insufficient, due to
the complex structure of the data and the high input
dimensionality. Tensor regression learns a linear mapping
and deals with dense input tensors. Thus, their approach is
fundamentally different from ours; see Zhou et al. (2013);
Yu and Liu (2016).
Our approach shows some similarities to RBF-networks
which are able to approximate any non-linear function
by using radial basis functions. RBF-networks have been
successfully applied to a number of tasks including control
engineering, see Broomhead and Lowe (1988). The main
difference to our proposed functional Tucker model is
that RBF-networks learn one latent representation for
the complete input, and map it to the output; whereas,
the functional Tucker model learns a representation for
each tensor mode and jointes them using the tensor
decomposition model. In this way multi-way interactions
are modeled explicitly.
Inverse dynamics are traditionally modeled using the rigidbody formulation, see Craig (2005). However, general regression techniques such as locally weighted projection regression (LWPR), Gaussian Processes, and RBF-networks,
have shown advantageous for learning inverse dynamics,
see Vijayakumar and Schaal (2000); Rasmussen (2006).
The topic was subject to a number of studies, see Burdet
and Codourey (1998); Nguyen-Tuong et al. (2008). Support vector regression has shown superior performance for
this task.
5. EXPERIMENTS
In this section we evaluate our proposed method on an
inverse dynamics dataset including movements from a
seven degrees of freedom SARCOS robot arm. We compare
against various other state-of-the-art regression techniques
for this task.
5.1 Dataset
The dataset was introduced by Vijayakumar and Schaal
(2000). 1 It contains data from a SARCOS robot arm
with seven degrees of freedom. The data was collected
from the moving robot arm at 100Hz and corresponds to
7.5 minutes of movement. The dataset includes 21 input
dimensions, consisting of 7 joint torques, 7 joint positions,
7 joint velocities, and 7 joint accelerations. The whole
dataset consists of 42482 samples. We split the dataset
randomly into 90 percent training and 10 percent test
data. Additional 5 percent of the training set where used
as a validation set. The task is to learn a model on the
training data, which models the 7 joint torques, given the
1
http://www.gaussianprocess.org/gpml/data/
3
2
1.5
1
0.5
0
5.2 Baselines
functional Tucker
functional PARAFAC
2.5
nMSE in %
positions, velocities and accelerations. The offline learned
model can then be applied in the forward controller of the
robot. The dataset has been subject to some studies on
the topic, see Vijayakumar and Schaal (2000); Rasmussen
(2006). The regression task has been found to be highly
non-linear. Non-linear regression techniques outperformed
the rigid-body dynamics formulation by a large margin.
The performance of the regression techniques is evaluated
on the test set, which includes unseen movements. We
repeated the random split 10 times and report the average
results and the standard deviation of multiple trials.
5
10
15
20
25
30
35
40
45
50
Rank
We compare our model against various state-of-the art
regression techniques, modeling the function y = f (q, q̇, q̈).
The baseline models we consider are linear regression,
RBF-networks and support vector regression. In previous
studies support vector regression has shown the best
results on this task. For all baseline models a concatenated
vector x = [q, q̇, q̈] is built. The linear regression model
learns a linear mapping from the inputs to the outputs,
such that
y = W x + b.
(10)
RBF-networks model the regression problem as,
r̃
X
wi exp (−βi kx − ci k2 ).
(11)
y=
i=1
The parameters ci , βi and wi are learned using backpropagation. We initialized the parameters ci with the centroids
of a k-means clustering on the training data, where r̃ is
the number of centroids.
Support vector regression (see Smola and Vapnik (1997))
has shown state-of-the-art results in modeling inverse
dynamics. It predicts y as,
N
X
y=
(αj − αj? )k(xj , x) + b
(12)
j=1
0
with k(x, x ) being a kernel function. In the experiments
we use a Gaussian kernel. αj and αj? are Lagrange multipliers, which are determined during optimization. In our
experiments we use the libsvm library, see Chang and Lin
(2011).
Furthermore, we compare the functional Tucker model
proposed in Section 3 with a functional PARAFAC model.
For the functional PARAFAC model we replace the tensor
decomposition in equation 5 with a PARAFAC decomposition, as shown in equation 2.
5.3 Results
We report the normalized mean squared error (nMSE) for
the regression task, which is defined as the mean squared
error of all data points divided by the variance of the target
variable in the training data. Table 1 summarizes the mean
nMSE for all seven degrees of freedom in percent. In the
rightmost column the mean of all seven degrees of freedom
is shown. All results, as well as the standard deviation are
referring to the average result of 10 random data splits.
The performance of the regression techniques varies across
the DoFs. The linear model reaches an nMSE of 11.03%
Fig. 2. Normalized mean squared error of the functional
Tucker and functional PARAFAC model, in dependency of the embedding rank.
in average. The nonlinear RBF-networks performs much
better with an nMSE of 2.14% in average. The number of
of hidden neurons for the RBF-network was set to 1000.
With larger numbers the predictive performance did not
increase. The support vector regression model yields a
very good result of 0.60%. Here, we set the parameter C
to 600 and to 0.1. All hyperparameters were evaluated
on a separate validation set. Our proposed functional
Tucker model resulted in a slightly better nMSE of 0.55%.
Especially, for the first two DoFs the functional Tucker
model performs significantly better than support vector
regression. For the other DoFs the results of support
vector regression and functional Tucker decomposition
are very close to each other. The parameter efficient
functional PARAFAC model reaches an nMSE of 0.96%
in average. Figure 2 shows the performance of the two
functional tensor decomposition models in dependence of
the rank of the decompositions. For the Tucker model,
the performance converges at a rank of 30 and for the
PARAFAC model at a rank of 40. It is also notable that
both methods already perform relatively well with a very
small rank of 5. The nMSE of the Tucker model is 2.09%
with a rank of 5 and the nMSE of the PARAFAC model
is 2.43%. Both functional tensor models show clearly
better results than RBF-networks. This indicates that the
explicit modeling of the three-way interaction, yields a
significant improvement.
6. CONCLUSION
In this paper we apply a tensor model, that is based
on the Tucker decomposition, to inverse dynamics. Our
proposed model exploits the inherent three-way interaction of positions × velocities × accelerations. We show
how the decomposition of sparse tensors can be applied
to regression tasks. Furthermore, we propose to augment
the tensor decompositions with basis functions for allowing
continuous input variables. In this way, a functional version of a tensor decomposition can be derived. Representations for each tensor mode are induced through the basis
functions and fused by the tensor model. The parameters
of the basis functions are learned using backpropagation.
Experiments on an inverse dynamics dataset, derived from
a seven degrees of freedom robot arm, show promising results of our proposed model for the application of learning
inverse dynamics. The proposed functional Tucker model
outperforms RBF-networks, and even support vector regression, which has shown state-of-the-art performance on
this task. Our extension of tensor decomposition models
to continuous inputs enables a wide range of application
areas. Especially if an inherent multi-way structure exists
in the data, functional tensor models can be advantageous
over traditional techniques, by explicitly modeling the
multi-way interaction. Within automatic robot control the
approach might be further extended to learning also a
functional Tucker model for a feedback controller based
on the tracking errors.
REFERENCES
Auer, S., Bizer, C., Kobilarov, G., Lehmann, J., Cyganiak,
R., and Ives, Z. (2007). Dbpedia: A nucleus for a web
of open data. In The semantic web, 722–735. Springer.
Bahadori, M.T., Yu, Q.R., and Liu, Y. (2014). Fast multivariate spatio-temporal analysis via low rank tensor
learning. In Advances in neural information processing
systems, 3491–3499.
Baier, S., Krompass, D., and Tresp, V. (2016). Learning
representations for discrete sensor networks using tensor
decompositions. International Conference on Multisensor Fusion and Integration for Intelligent Systems.
Broomhead, D.S. and Lowe, D. (1988). Radial basis
functions, multi-variable functional interpolation and
adaptive networks. Technical report, DTIC Document.
Burdet, E. and Codourey, A. (1998). Evaluation of
parametric and nonparametric nonlinear adaptive controllers. Robotica, 16(01), 59–73.
Chang, C.C. and Lin, C.J. (2011). Libsvm: a library
for support vector machines. ACM Transactions on
Intelligent Systems and Technology (TIST), 2(3), 27.
Cheng, D., Qi, H., and Li, Z. (2010). Analysis and control
of Boolean networks: a semi-tensor product approach.
Springer Science & Business Media.
Cichocki, A. (2014). Tensor networks for big data analytics
and large-scale optimization problems. arXiv preprint
arXiv:1407.3124.
Craig, J.J. (2005). Introduction to robotics: mechanics and
control, volume 3. Pearson Prentice Hall Upper Saddle
River.
Gróf, P., Takarics, B., Petres, Z., and Korondi, P. (2010).
Tensor product model type polytopic decomposition of
a pneumatic system with friction phenomena taken into
account. In Applied Machine Intelligence and Informatics (SAMI), 2010 IEEE 8th International Symposium
on, 153–158. IEEE.
Harshman, R.A. (1970). Foundations of the parafac
procedure: Models and conditions for an” explanatory”
multi-modal factor analysis.
Kingma, D. and Ba, J. (2014).
Adam: A method
for stochastic optimization.
arXiv preprint
arXiv:1412.6980.
Kolda, T.G. and Bader, B.W. (2009). Tensor decompositions and applications. SIAM review, 51(3), 455–500.
Koren, Y., Bell, R., and Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer,
(8), 30–37.
Lloyd, S. (1982). Least squares quantization in pcm. IEEE
transactions on information theory, 28(2), 129–137.
Mørup, M., Hansen, L.K., Herrmann, C.S., Parnas, J., and
Arnfred, S.M. (2006). Parallel factor analysis as an
exploratory tool for wavelet transformed event-related
eeg. NeuroImage, 29(3), 938–947.
Nakanishi, J., Farrell, J.A., and Schaal, S. (2005). Composite adaptive control with locally weighted statistical
learning. Neural Networks, 18(1), 71–90.
Nguyen-Tuong, D., Peters, J., Seeger, M., and Schölkopf,
B. (2008). Learning inverse dynamics: a comparison.
In European Symposium on Artificial Neural Networks,
EPFL-CONF-175477.
Nickel, M., Murphy, K., Tresp, V., and Gabrilovich, E.
(2015). A review of relational machine learning for
knowledge graphs: From multi-relational link prediction
to automated knowledge graph construction. arXiv
preprint arXiv:1503.00759.
Nickel, M., Tresp, V., and Kriegel, H.P. (2011). A threeway model for collective learning on multi-relational
data. In Proceedings of the 28th international conference
on machine learning (ICML-11), 809–816.
Oseledets, I.V. (2011). Tensor-train decomposition. SIAM
Journal on Scientific Computing, 33(5), 2295–2317.
Pangalos, G., Eichler, A., and Lichtenberg, G. (2013a).
Tensor systems - multilinear modeling and applications.
In Proceedings of the 3rd International
Conference on Simulation and Modeling Methodologies, Technologies and Applications, 275–285. doi:
10.5220/0004475602750285.
Pangalos, G., Eichler, A., and Lichtenberg, G. (2013b).
Tensor systems-multilinear modeling and applications.
In SIMULTECH, 275–285.
Rasmussen, C.E. (2006). Gaussian processes for machine
learning.
Rogers, M., Li, L., and Russell, S.J. (2013). Multilinear
dynamical systems for tensor time series. In Advances
in Neural Information Processing Systems, 2634–2642.
Smola, A. and Vapnik, V. (1997). Support vector regression machines. Advances in neural information processing systems, 9, 155–161.
Suchanek, F.M., Kasneci, G., and Weikum, G. (2007).
Yago: a core of semantic knowledge. In Proceedings of
the 16th international conference on World Wide Web,
697–706. ACM.
Theano Development Team (2016). Theano: A Python
framework for fast computation of mathematical expressions.
arXiv e-prints, abs/1605.02688.
URL
http://arxiv.org/abs/1605.02688.
Tucker, L.R. (1965). Some Mathematical Notes on Threemode Factor Analysis [by]. Urbana, Department of
Psychology, University of Illinois.
Vijayakumar, S. and Schaal, S. (2000). Locally weighted
projection regression: An o (n) algorithm for incremental real time learning in high dimensional space. In International conference on machine learning, proceedings
of the sixteenth conference.
Yu, R. and Liu, Y. (2016). Learning from multiway data:
Simple and efficient tensor regression. In Proceedings of
the 33nd International Conference on Machine Learning
(ICML-16), 238–247.
Zhou, H., Li, L., and Zhu, H. (2013). Tensor regression
with applications in neuroimaging data analysis. Journal of the American Statistical Association, 108(502),
540–552.
| 3cs.SY
|
Polly’s Polyhedral Scheduling in the Presence of
Reductions
Johannes Doerfert, Kevin Streit, and
Sebastian Hack
Zino Benaissa
Qualcomm Innovation Center
San Diego
California, USA
Computer Science Department
Saarland University
Saarbrücken, Germany
[email protected]
arXiv:1505.07716v1 [cs.PL] 28 May 2015
<lastname>@cs.uni-saarland.de
ABSTRACT
to loop carried data dependences. An often used definition
for reductions describes them as an associative and commutative computation which reduces the dimensionality of a
set of input data [16]. A simple example is the array sum
depicted in Figure 1a. The input vector A is reduced to
the scalar variable sum using the associative and commutative operator +. In terms of data dependences, the loop has
to be computed sequentially because a read of the variable
sum in iteration i + 1 depends on the value written in iteration i. However, the associativity and commutativity of the
reduction operator can be exploited to reorder, parallelize
or vectorize such reductions.
While reordering the reduction iterations is always a valid
transformation, executing reductions in a parallel context requires additional “fix up”. Static transformations often use
privatization as fix up technique as it works well with both
small and large parallel tasks. The idea of privatization is
to duplicate the shared memory locations for each instance
running in parallel. Thus, each parallel instance works on
a private copy of a shared memory location. Using the privatization scheme we can vectorize the array sum example
as shown in Figure 1b. For the shared variable sum, a temporary array tmp_sum, with as many elements as there are
vector lanes, is introduced. Now the computation for each
vector lane uses one array element to accumulate intermediate results unaffected by the computations of the other
lanes. As the reduction computation is now done in the
temporary array instead of the original reduction location
we finally need to accumulate all intermediate results into
the original reduction location. This way, users of the variable sum will still see the overall sum of all array elements,
even though it was computed in partial sums first.
The polyhedral model provides a powerful mathematical abstraction to enable effective optimization of loop nests with
respect to a given optimization goal, e.g., exploiting parallelism. Unexploited reduction properties are a frequent
reason for polyhedral optimizers to assume parallelism prohibiting dependences. To our knowledge, no polyhedral loop
optimizer available in any production compiler provides support for reductions. In this paper, we show that leveraging
the parallelism of reductions can lead to a significant performance increase. We give a precise, dependence based,
definition of reductions and discuss ways to extend polyhedral optimization to exploit the associativity and commutativity of reduction computations. We have implemented
a reduction-enabled scheduling approach in the Polly polyhedral optimizer and evaluate it on the standard Polybench
3.2 benchmark suite. We were able to detect and model all
52 arithmetic reductions and achieve speedups up to 2.21×
on a quad core machine by exploiting the multidimensional
reduction in the BiCG benchmark.
Categories and Subject Descriptors
D 3.4 [Programming languages]: Processors—Compilers, Optimization
General Terms
Algorithms; Performance
Keywords
Compiler Optimization; Affine Scheduling; Reductions
1.
for (i = 0; i < 4 * N; i++)
sum += A[i];
INTRODUCTION
Over the last four decades various approaches [10, 11, 2,
17, 22, 6, 18, 21, 27, 31, 7] were proposed to tackle reductions: a computational idiom which prevents parallelism due
(a) Sequential array sum computation.
tmp_sum[4] = {0,0,0,0}
for (i = 0; i < 4 * N; i+=4)
tmp_sum[0:3] += A[i:i+3];
IMPACT 2015
Fifth International Workshop on Polyhedral Compilation Techniques
Jan 19, 2015, Amsterdam, The Netherlands
In conjunction with HiPEAC 2015.
sum += tmp_sum[0] + tmp_sum[1];
+ tmp_sum[2] + tmp_sum[3];
http://impact.gforge.inria.fr/impact2015
Figure 1: A canonical example of a single address reduction.
(b) Vectorized array sum computation.
1
Transformations as described above have been the main
interest of reduction handling approaches outside the polyhedral world. Associativity and commutativity properties
are used to extract and parallelize the reduction loop [10,
6] or to parallelize the reduction computation with regards
to an existing surrounding loop [17, 18, 21, 27, 31]. While
prior work on reductions in the polyhedral model [22, 23, 24,
9, 32] was focused on system of affine recurrences (SAREs),
we look at the problems a production compiler has to solve
when we allow polyhedral optimizations that exploit the reduction properties. To this end our work supplements the
polyhedral optimizer Polly [8], part of the LLVM [14] compiler framework with awareness for the associativity and
commutativity of reduction computations. While we are
still in the process of upstreaming, most parts are already
accessible in the public code repository.
The contributions of this paper include:
dependences between individual instances thereof. R has a
one-dimensional iteration space, as it is nested in the i-loop
only. Statements S and T have a two-dimensional iteration
space as they are nested in both the i-loop as well as in the
j -loop. The axes in the Figure correspond to the respective
loops. Single instances of each statement are depicted as
dots in the graph. Dependences between individual statement instances are depicted as arrows: dashed ones for regular data dependences and dotted ones for loop carried data
dependences.
Stmt R
j
NY-1
j
0
NY-1
In the polyhedral model the iteration space of a statement
Q is represented as a multidimensional Z-polytope IQ , defined by affine constraints on the iteration variables of loops
surrounding the statement, as well as global parameters.
The latter are basically loop invariant expressions like for example the upper bounds NX and NY of the loops in Figure 2.
As a consequence, the polyhedral model is only applicable
to well structured program parts with affine loop bounds
and memory access functions, so called Static Control Parts
(SCoP s) [8]. While, there are different over-approximations
to increase the applicability (e.g., by Benabderrahmane [1])
we will assume that all restrictions of SCoP s are fulfilled.
The dependences between two statements Q and T are
also represented as a multidimensional Z-polytope known
as the dependence polytope D<Q,T > . It contains a point <
iQ , iT > for every pair of instances < iQ >∈ IQ and < iT >∈
IT for which the latter depends on the former. To ease
reading we will however omit the index of the dependence
polytopes and only argue about the set of all dependences
D, defined as:
The remainder of this paper is organized as follows: We
give a short introduction into the polyhedral model in Section 2. Thereafter, in Section 3, our reduction detection is
described. Section 4 discusses the benefits and drawbacks
of different reduction parallelization schemes, including privatization. Afterwards, we present different approaches to
utilize the reduction properties in a polyhedral optimizer in
Section 5. In the end we evaluate our work (Section 6), compare it to existing approaches (Section 7) and conclude with
possible extensions in Section 8.
THE POLYHEDRAL MODEL
The main idea behind polyhedral loop nest optimizations
is to abstract from technical details of the target program.
Information relevant to the optimization goal is represented
in a very powerful mathematical model and the actual optimizations are well understood transformations on this representation. In the context of optimization for data locality
or parallelism, the relevant information is the iteration space
of each statement, as well as the data dependences between
individual statement instances.
0; i <
0;
= 0; j
= q[i]
= s[j]
0
Figure 3: Polyhedral representation of statements R, S and
T of the BiCG Sub Kernel of Figure 2.
• A dependence based approach to identify vectorization
and parallelization opportunities in the presence of reductions.
for (i =
q[i] =
for (j
S:
q[i]
T:
s[j]
}
}
0
Stmt T
i
0
• A sound model to relax memory dependences with regards to reductions and its use in reduction-enabled
polyhedral scheduling.
R:
i
NX-1
• A powerful algorithm to identify reduction dependences,
applicable whenever memory or value based dependence information is available.
2.
Stmt S
i
D := {< iQ , iT > | ∀Q, T ∈ SCoP :< iQ , iT >∈ D<Q,T > }
Later we will also distinguish all Write-After-Write (WAW
or output dependence) dependences of D by writing DWAW .
A loop transformation in the polyhedral model is represented as an affine function θQ for each statement Q. It is
often called scheduling or scattering function. This function translates a point in the original iteration space IQ of
statement Q into a new, transformed target space. One important legality criterion for such a transformation is that
data dependences need to be respected: The execution of
every instance of a source statement Q of a dependence has
to precede the execution of the corresponding target statement T in the transformed space. Formulated differently:
the target iteration vector of the value producing instance
of Q has to be lexicographically smaller1 than the target
NX; i++) {
< NY; j++) {
+ A[i][j] * p[j];
+ r[i] * A[i][j];
Figure 2: BiCG Sub Kernel of BiCGStab Linear Solver.
Figure 2 shows an example program containing three statements R, S and T in a loop nest of depth two. Figure 3
shows the polyhedral representation of the individual iteration spaces for all statements, as well as value-based data
1
To compare two vectors of different dimensionality, we simply fill up the shorter vector with zeros in the end to match
the dimensionality of the larger one.
2
iteration vector of the consuming instance of T :
< iQ , iT >∈ D ⇒ θQ (iQ ) θT (iT )
SCoP . The range of a memory instruction is defined as the
range of its affine access function:
(1)
ran(store x A[f(iS , p)]) := ran(A[f(iS , p)])
ran(load A[f(iS , p)]) := ran(A[f(iS , p)])
ran(A[f(iS , p)]) := A + ran(f(iS , p))
Multiple statements, or multiple instances of the same statement, that are mapped to the same point in the target space,
can be executed in parallel. However, implementations of
polyhedral schedulers [3, 29] usually generate scheduling
functions with full rank, thus rank (dom(θ)) = rank (img(θ)).
The parallelism is therefore not explicit in the scheduling
function but is exposed later when the polyhedral representation is converted to target code.
There are two things that make the described model particularly interesting for loop transformation: First, unlike
classical optimizers, a polyhedral optimizer does not only
consider individual statements, but instead individual dynamic instances of each statement. This granularity leads
to a far higher expressiveness. Second, the combination of
multiple classical loop transformations, like for instance loop
skewing, reversal, fusion, even tiling, typically used as atoms
in a sequence of transformations, can be performed in one
step by the scattering function. There is no need to come
up with and evaluate different, possibly equivalent or even
illegal combinations of transformations. Instead, linear optimization is used to optimize the scattering function for every
individual statement with respect to an optimization goal.
3.
Note the absence of any kind of control flow producing or
dependent instructions (φ instructions or branches). This
is a side effect of the limited scope of the reduction detection analysis. It is applied only to polyhedral statements, in
our setting basic blocks with exactly one store instruction.
Furthermore, we assume all loop carried values are communicated in memory. This setup is equivalent to C source
code statements without non-memory side effects.
3.1
DETECTING REDUCTIONS
Pattern based approaches on source statements are limited to find general reduction idioms [6, 21, 27]. The two
main restrictions are the amount of patterns in the compiler’s reduction pattern database and the sensitivity to the
input code quality or preprocessing steps. To become as independent as possible of source code quality and canonicalization passes we replace the pattern recognition by a simple,
data flow like analysis. This analysis will identify reductionlike computations within each polyhedral statement. Such a
computation is a potential candidate for a reduction, thus it
might be allowed to perform the computation in any order
or even in parallel. Afterwards, we utilize the polyhedral
dependence analysis [5] to precisely identify all reduction
dependences [20] in a SCoP , hence to identify the actual
reduction computations from the set of possible candidate
(reduction-like) computations.
l = load A[f(iS , p)]
x = y
Reduction-like Computations
Reduction-like computations are a generalization of the
reduction definition used e.g., by Jouvelot [11] or Rauchwerger [21]. Their main characteristic is an associative and
commutative computation which reduces a set of input values into reduction locations. Furthermore, the input values,
the control flow and any value that might escape into a nonreduction location needs to be independent of the intermediate results of the reduction-like computation. The difference
between reduction-like computations and reductions known
in the literature is the restriction on other appearances of
the reduction location in the loop nest. We do not restrict
syntactic appearances of the reduction location base pointer
as e.g., Rauchwerger [21] does, but only accesses to the actual reduction location in the same statement. This means
a reduction-like computation on A[i%2] is not invalidated
by any occurrence of A[i%2 + 1] in the same statement
or any occurrence of A in another statement.
It is crucial to stress that we define reduction-like computations for a single polyhedral statement containing only a
single store. Thus intermediate results of a reduction-like
computation can only escape if they are used in a different statement or outside the SCoP . As we focus on memory reductions in a single statement we will assume such
outside uses invalidate a candidate computation from being
reduction-like. To this end we define the function:
hasOutsideUses : Insts → bool
that returns true if an instruction is used outside its statement. In Section 3.3 we explain how the situation changes if
multiple statements are combined into compound statements
in order to save compile time.
Reconsider the array sum example in Figure 1. The reduction location is the variable sum, a scalar variable or
zero dimensional array. However, we do not limit reductionlike computations to zero dimensional reduction locations,
instead we allow multidimensional reduction locations, also
called histogram reductions [18], as well. The second example, Figure 2, shows two such multidimensional reductions.
The reduction locations are q[i] and s[j]. The first is
variant in the outer loop, the second in the inner loop.
To detect reduction-like computations we apply the detection function tS , shown in Figure 4, to the store in the
polyhedral statement S. The idea is to track the flow of
loaded values through computation up to the store. To
this end, tS (I) for any instruction I will assign each load
a symbol that describes how the value loaded by load used
up to and by I. We will use Rop to refer to the set of all
store x A[f(iS , p)]
z
Figure 5: SSA-based language subset.
The following discussion is restricted to the SSA-based
language subset (Insts) depicted in Figure 5. Our implementation however handles all LLVM-IR [14] instructions.
The binary operation is parametrized with and can be
instantiated with any arithmetic, bit-wise or logic binary operator. To distinguish associative and commutative binary
operators we use ⊕ instead. The load instruction is applied
to a memory location. It evaluates to the current value x
stored in the corresponding memory location. The store
instruction takes a value x and writes it to the given memory location. In both cases the memory location is described
as A[f(iS , p)], where A is a constant array base pointer
and f(iS , p) is an affine function with regards to outer loop
indices of the statement S (iS ) and parameters (p) of the
3
tS (l = load A[f(iS , p)])
:=
λl :
if (l 6= l) then ⊥ else
if (hasOutsideUses(l)) then > else ↑
tS (x = y
z)
:=
λl :
if ({ (tS (y))(l), (tS (z))(l) } = { ⊥, ⊥ }) then ⊥ else
if ¬(isCommutative( ) ∧ isAssociative( )) then > else
if (hasOutsideUses(x)) then > else
tS (store x A[f(iS , p)])
:=
λl :
if ({ (tS (y))(l), (tS (z))(l) } = { ↑, ⊥ }) then
else
if ({ (tS (y))(l), (tS (z))(l) } = { , ⊥ }) then
else >
if x ∈
/ Insts then ⊥ else
if (ran(l) ∩ ran(A[f(iS , p)]) = ∅) then > else
if (∃l0 : l 6= l0 ∧ ran(l0 ) ∩ ran(l) ∩ ran(A[f(iS , p)]) 6= ∅) then > else (tS (x))(l)
Figure 4: Detection function for reduction-like computations: tS : Insts → (loads(S) → Rop ).
four symbols. It includes the ⊥ indicating that the load
was not used by the instruction, the ↑ to express that it
was only loaded but not yet used in any computation, the >
stating that the loaded value may have been used in a nonassociative or non-commutative computation. Additionally,
the ⊕ is used when the loaded value was exactly one input of a chain of ⊕ operations. Note that only a load l
that flows with ⊕ into the store is a valid candidate for
a reduction-like computation and only if the load and the
store access (partially) the same memory. Furthermore, we
forbid all other load instructions in the statement to access
the same memory as both l and the store as that would
again make the computation potentially non-associative and
non-commutative.
If a valid load l was found, it is the unique load instruction inside the statement S that accesses (partially)
the same memory as the store s and (tS (s))(l) is an associative and commutative operation ⊕. We will refer to the
quadruple (S, l, ⊕, s) as the reduction-like computation Rc
of S and denote the set of all reduction-like computations in
a SCoP as Rc .
It is worth noting that we explicitly allow the access functions of the load and the store to be different as for example shown in Figure 6. In such cases a reduction can
manifest only for certain parameter valuations or, as shown,
for certain valuations of outer loop indices. Additionally, we
could easily extend the definition to allow non-affine but
Presburger accesses or even over-approximated non-affine
accesses if they are pure. It is also worth to note that our
definition does not restrict the shape of the induced reduction dependences.
computation, reduction dependences can be considered to be
“associative” and “commutative”. The latter allows a schedule to reorder the iterations participating in the reductionlike computation while it can still be considered valid, however all non-reduction dependences still need to be fulfilled.
We split the set of all dependences D into the set of reduction dependences Dρ ⊆ D and the set of non-reduction
dependences Dν := D \ Dρ . Now we can express the commutativity of a reduction dependence by extending the causality condition given in Constraint 1 as follows:
(2)
< iQ , iT >∈ Dρ =⇒ θQ (iQ ) 6= θT (iT )
(3)
Constraint 2 is the same as the original causality condition (Constraint 1), except that we restrict the domain to
non-reduction dependences Dν . For the remaining reduction dependences Dρ , Constraint 3 states that the schedule
θ can reorder two iterations freely, as long as they are not
mapped to the same time stamp. However, relaxing the
causality condition for reduction dependences is only valid
if D contains all transitive reduction dependences. This is
for example the case if D is computed by a memory-based
dependence analysis. In case only value-based dependence
analysis [5] was performed it is also sufficient to provide the
missing transitive reduction dependences e.g., by recomputing them using a memory-based dependence analysis.
Reconsider the BiCG kernel (Figure 2) and its non transitive (value-based) set of dependences D shown in Figure 3.
If we remove all reduction dependences Dρ from D, the only
constraints left involve statement R and the iterations of
statements S with j = 0. Consequently, there is no reason
not to schedule the other instances of statement S before
statement R.
To address the issue of only value-based dependences without recomputing memory-based ones we use the transitive
closure Dρ+ S of the reduction dependences for a statement S
(Equation 4). As the transitive closure of a Presburger relation is not always a Presburger relation we might have to use
an over-approximation to remain sound, however Pugh and
Wonnacot [19] describe how the transitive closure can also
be computed precisely for exact direction/distance vectors.
They also argue in later work [20] that the transitive closure
of value-based reduction dependences of real programs can
be computed in an easy and fast way.
If we now interpret Dρ+ S as a relation that maps instances
of a reduction statement S to all instances of S transi-
for (i = 0; i < N; i++)
for (j = 0; j < M; j++)
A[j] = A[j-i] + Mat[i][j];
Figure 6: Conditional reduction with different access functions.
3.2
< iQ , iT >∈ Dν =⇒ θQ (iQ ) θT (iT )
Reduction Dependences
While the data flow analysis performed on all polyhedral
statements only marks reduction-like computations, we are
actually interested in reduction dependences [20]. These loop
carried self dependences start and end in two instances of
the same reduction-like computation and they inherit some
properties of this computation. Similar to the reduction-like
4
tively dependent, we can define privatization dependences
Dτ (Equation 5). In simple terms, Dτ will ensure that
no non-reduction statement accessing the reduction location can be scheduled in-between the reduction statement
instances by extending the dependences ending or starting
from a reduction access to all reduction access instances.
This also means that in case no memory locations are reused
e.g., after renaming and array expansion [4] was applied, the
set of privatization dependences will be empty.
Dρ+ S := (Dρ ∩ < IS , IS >)+
non-reduction memory locations. This happens if and only
if intermediate values—and therefore the reduction load—
flow into multiple store instructions of the statement S.
Additionally, other store instructions are not allowed to
override intermediate values of the reduction computation.
Thus, (S, l, ⊕, s) can only be a reduction-like computations,
if for all other store instructions s’ in S:
(t(s’))(l) = ⊥ ∧ range(s’) ∩ range(s) ∩ range(l) = ∅
Furthermore, we cannot assume that all loop carried WAW
self dependences of a statement containing a reduction-like
computation are reduction dependences: other read and
write accesses contained in the statement could cause the
same kind of dependences. However, we are particularly
interested in dependences caused by the load and store
instruction of a reduction-like computation Rc . To track
these accesses separately we can pretend they are contained
in their own statement SRc that is executed at the same
time as S (in the original iteration space). This is only
sound as long as no other instruction in S accesses (partially)
the same memory as the load or the store, but this was
already a restriction on reduction-like computations. The
definition of reduction dependences (Equation 7) is finally
changed to:
(4)
Dτ := {< iT , Dρ+ S (iS ) > | < iT , iS >∈ D<T,S> }
∪ {< Dρ+ S (iS ), iT > | < iS , iT >∈ D<S,T > }
(5)
Privatization dependences overestimate the dependences
that manual privatization of the reduction locations would
cause. They are used to create alternative causality constraint for the reduction statements that enforce the initial
order between the reduction-like computation and any other
statement accessing the reduction locations. To make use of
them we replace Constraint 2 by Constraint 6.
< iQ , iT >∈ (Dν ∪ Dτ ) =⇒ θQ (iQ ) θT (iT )
(6)
If we now utilize the associativity of the reduction dependences we can compute intermediate results in any order before we combine them to the final result. As a consequence we can allow parallelization of the reduction-like
computation, thus omit Constraint 3; thereby eliminating
the reduction dependences Dρ from the causality condition
of a schedule completely. However, parallel execution of iterations connected by reduction dependences requires special “treatment” of the accesses during the code generation
as described in Section 4.
The restriction on polyhedral statements, especially that
it contains at most one store instruction, eases the identification of reduction dependences; they are equal to all loop
carried Write-After-Write self dependences over a statement
with a reduction-like computation2 . Thus, Dρ can be expressed as stated in Equation 7.
Dρ := DWAW ∩ { IS × IS | (S, l, ⊕, s) ∈ Rc }
3.3
Dρ := DWAW ∩ { ISRc × ISRc | Rc ∈ Rc }
(8)
It is important to note the increased complexity of the
dependence detection problem when we model reduction accesses separately. However, our experiments in Section 6
show that the effect is (in most cases) negligible. Furthermore, we want to stress that this kind of separation is not
equivalent to separating the reduction access at the statement level as we do not allow separate scheduling functions
for S and SRc . Similar to a fine-grained granularity at the
statement level, separation might be desirable in some cases,
however it suffers from the same drawbacks.
4.
PARALLEL EXECUTION
When executing accesses to a reduction location x , p times
in parallel, it needs to be made sure that the read-modifywrite cycle on x happens atomically. While doing exactly
that — performing atomic read-modify-write operations —
might be a viable solution in some contexts [31], it is generally too expensive. The overhead of an atomic operation
easily outweighs the actual work for smaller tasks [18]. Additionally, the benefit of vectorization is lost for the reduction, as atomic operations have to scalarize the computation
again. We will therefore focus our discussion and the evaluation on privatization as it is generally well-suited for the
task at hand [18].
(7)
General Polyhedral Statements
Practical polyhedral optimizer operate on different granularities of polyhedral statements; a crucial factor for both
compile time and quality of the optimization. While Clan 3
operates on C statements, Polly is based on basic blocks
in the SSA-based intermediate language of LLVM. The former eases not only reduction handling but also offers more
scheduling freedom. However, the latter can accumulate the
effects of multiple C statements in one basic block, thus it
can perform better with regards to compile time. Finding a
good granularity for a given program, e.g., when and where
to split a LLVM basic block in the Polly setting, is a research
topic on its own but we do not want to limit our approach to
one fixed granularity. Therefore, we will now assume a polyhedral statement can contain multiple store instructions,
thus we allow arbitrary basic blocks.
As a first consequence we have to check that intermediate
values of a reduction-like computation do not escape into
4.1
Privatizing Reductions
Privatization means that every parallel context c i , which
might be a thread or just a vector lane, depending on the
kind of parallelization, gets its own private location xi for
x . In front of the parallelized loop carrying a reduction dependence p, private locations x1 , · · · , xp of x are allocated
and initialized with the identity element of the corresponding reduction operation ⊕. Every parallel context c i now
non-atomically, and thus cheaply, modifies its very own location xi . After the loop, but before the first use of the x,
accumulation code needs to join all locations into x again,
thus: x := x ⊕ x1 ⊕ · · · ⊕ xp .
2
In this restricted environment we could also use the ReadAfter-Write (RAW) dependences instead of the WAW ones.
3
http://icps.u-strasbg.fr/˜bastoul/development/clan/
5
// (A) init
for (i = 0; i < NX; i++)
// (B) init
for (j = 0; j < NY; j++)
// (C) init
for (k = 0; k < NZ; k++)
P[j] += Q[i][j] * R[j][k];
// (C) aggregate
// (B) aggregate
// (A) aggregate
sulting code might be limited, which suggests that the scheduler should be aware of the implications of a chosen schedule
with respect to the efficiency of necessary privatization.
In Section 6.1 we discuss the effect of different placement
choices for the BiCG benchmark shown in Figure 2.
5.
MODELING REDUCTIONS
As mentioned earlier, the set D of all dependences is partitioned into the set Dρ of reduction induced dependences and
Dν of regular dependences. Reduction dependences inherit
properties similar to associativity and commutativity from
the reduction operator ⊕: the corresponding source and
target statement instances can be executed in any order—
provided ⊕ is a commutative operation—or in parallel—if
⊕ is at least associative. In order to exploit these properties the polyhedral optimizer needs to be aware of them. To
this end we propose different scheduling and code generation
schemes.
Figure 7: Possible privatization locations (A-C ) for the reduction over P[j].
Such a privatization transformation is legal due to the
properties of a reduction operation. Every possible user
of x sees the same result after the final accumulation has
been performed as it would have seen before the transformation. Nevertheless we gained parallelism which cannot be
exploited without the reduction properties. It might seem,
that the final accumulation of the locations needs to be performed sequentially, but note that the number of locations
does not necessarily grow with the problem size but instead
only with the maximal number of parallel contexts. Furthermore, accumulation can be done in logarithmic time by
parallelizing the accumulation correspondingly [6].
One positive aspect of using privatization to fix a broken
reduction dependence is that it is particularly well-suited for
both ways of parallelization usually performed in the polyhedral context: thread parallelism and vectorization. For
thread parallelism real private locations of the reduction
address are allocated; in case of vectorization, a vector of
suitable width is used.
As described, privatization creates “copies” of the reduction location, one for each instance possibly executed in parallel. While we can limit the number of private locations
(this corresponds to the maximal number of parallel contexts), we cannot generally bound the number of reduction
locations. Furthermore, the number of necessary locations,
as well as the number of times initialization and aggregation is needed, varies with the placement of the privatization
code.
Consider the example in Figure 7. Different possibilities
exist to exploit reduction parallelism: using placement C
for the privatization, the k -loop could be executed in parallel and only p private copies of the reduction location are
necessary. There is no benefit in choosing location B as we
then need p×NY privatization locations (we have NY different reduction locations modified by the j -loop and p parallel
contexts), but there is no gain in the amount of parallelism
(the j -loop is already parallel). Finally, choosing location
A for privatization might be worthwhile. We still only need
p × NY privatized values, but save aggregation overhead:
While for location C , p values are aggregated NX × NY
times and for location B , p × NY locations are aggregated
NX times, for location A, p × NY locations are aggregated
only once. Furthermore, the i-loop can now be parallelized.
In general, a trade-off has to be made between memory
consumption, aggregation time and exploitable parallelism.
Finding a good placement however is difficult and needs to
take the optimization goal, the hardware and the workload
size into account. Furthermore, depending on the scheduling, the choices for privatization code placement in the re-
Reduction-Enabled Code Generation
is a simple, non-invasive method to realize reductions
during the code generation, thus without modification
of the polyhedral representation of the SCoP .
Reduction-Enabled Scheduling
exploits the properties of reductions in the polyhedral
representation. All reduction dependences are basically ignored during scheduling, thereby increasing the
freedom of the scheduler.
Reduction-Aware Scheduling
is the representation of reductions and their realization via privatization in the polyhedral optimization.
The scheduler decides when and where to make use of
reduction parallelism. However, non-trivial modifications of the polyhedral representation and the current
state-of-the-art schedulers are necessary.
5.1
Reduction-Enabled Code Generation
The reduction-enabled code generation is a simple, noninvasive approach to exploit reduction parallelism. The only
changes needed to enable this technique are in the code generation, thus the polyhedral representation is not modified.
So far, dimensions or loops are marked parallel if they do
not carry any dependences. With regards to reduction dependences we can relax this condition, hence we can mark
non-parallel dimensions or loops as parallel, provided we add
privatization code, if they only carry reduction dependences.
To implement this technique we add one additional check to
the code generation that is executed for each non-parallel
loop of the resulting code that we want to parallelize. It
uses only non-reduction dependences Dν not D to determine if the loop exclusively carries reduction dependences.
If so, the reduction locations corresponding to the broken
dependences are privatized and the loop is parallelized.
Due to its simplicity, it is easily integrable into existing
optimizers while the compile time overhead is reasonably
low. However, additional heuristics are needed. First, to
decide if reductions should be realized e.g., if privatization
of a whole array is worth the gain in parallelism. And second, where the privatization statements should be placed
(cf. Section 4.1). Note that usually the code generator has
no, and in fact should not have any, knowledge of the optimization goal of the scheduler.
6
Parallel
Outer
Tile
Apart from the need for heuristics, reduction-aware code
generation also misses opportunities to realize reductions effectively. This might happen if the scheduler has no reason
to perform an enabling transformation or the applied transformation even disabled the reduction. Either way, it is hard
to predict the outcome of this approach.
5.2
212 × 212
2.31
0.75
0.32
1.54
214 × 214
3.91
0.72
0.10
1.60
215 × 215
2.19
0.96
0.16
2.21
Table 1: BiCG run-time results. The values are speedups
compared to the sequential Polly version, first for the 32core machine, then for the 4-core machine.
Reduction-Enabled Scheduling
execution environments or parameter values, there is more
work needed in order to (1) predict the effects of parallelization and privatization on the actual platform and to (2) express them as affine constraints in the scheduling objective
function.
In contrast to reduction-aware code generation, which is
basically a post-processing step, reduction-enabled scheduling actually influences the scheduling processes by eliminating reduction dependences beforehand. Therefore, the
scheduler is (1) unaware of the existence of reductions and
their dependences and (2) has more freedom to schedule
statements if they contain reduction instances. While this
technique allows to exploit reductions more aggressively, there
are still disadvantages. First of all, this approach relies
on reduction-aware code generation as a back-end, hence it
shares the same problems. Second, the scheduler’s unawareness of reduction dependences prevents it from associating
costs to reduction realization. Thus, privatization is implicitly assumed to come for free. Consequently, the scheduler
does not prefer existing, reduction-independent parallelism
over reduction parallelism and therefore may require unnecessary privatization code.
For the BiCG example (Figure 2) omitting the reduction
dependences might not result in the desired schedule if we
assume we are only interested in one level of outermost parallelism4 and furthermore that the statements S and T have
been split prior to the scheduling. In this situation we want
to interchange the outer two loops for the T statement in
order to utilize the inherent parallelism, not the reduction
parallelism. However, without the reduction dependences
the scheduler will not perform this transformation. In order
to decrease the severity of this problem, the reduction dependences can still be used in the proximity constraints of
the scheduler [29], thus the scheduler will try to minimize
the dependence distance between reduction iterations and
implicitly move them to inner dimensions. This solves the
problem for all Polybench benchmarks with regards to outermost parallelism, however it might negatively affect vectorization if e.g., the innermost parallel dimension is always
vectorized.
5.3
210 × 210
0.19
0.55
0.03
1.10
6.
EVALUATION
We implemented Reduction-Enabled Scheduling (c.f., Section 5.2) in the polyhedral optimizer Polly and evaluated the
effects on compile time and run-time on the Polybench 3.2.
We used an Intel(R) core i7-4800MQ quad core machine
and the standard input size of the benchmarks.
Our approach is capable of identifying and modeling all
reductions as described in Section 3: in total 52 arithmetic
reductions in 30 benchmarks 5 .
As described earlier, our detection virtually splits polyhedral statements to track the effects of the load and store
instructions that participate in reduction-like computations.
As this increases the complexity of the performed dependence analysis we timed this particular part of the compilation for each of the benchmarks and compared our ?? hybrid
dependence analysis to a completely ?? access-wise analysis
and the default ?? statement-wise one. We use the term hybrid because reduction accesses are tracked separately while
other accesses are accumulated on statement level.
As shown in Figure 8 (top) our approach takes up to 5×
as long (benchmark lu) than the default implementation but
in average only 85% more. Access-wise dependence computation however is up to 10× slower than the default and
takes in average twice as long as our hybrid approach. Note
that both approaches do not only compute the dependences
(partially) on the access level but also the reduction and
privatization dependences as explained in Section 3.2.
Figure 8 (bottom) shows the speedup of our approach
compared to the non-reduction Polly. The additional scheduling freedom causes speedups for the data-mining applications (correlation and covariance) but slowdowns especially
for the matrix multiplication kernels (2mm, 3mm and gemm).
This is due to the way Polly generates vector code. The
deepest dimension of the new schedule that is parallel (or
now reduction parallel) will be strip-mined and vectorized.
Hence the stride of the contained accesses, crucial to generate efficient vector code, is not considered. However, we do
not believe this to be a general shortcoming of our approach
as there are existing approaches to tackle the problem of
finding a good vector dimension [13] that would benefit from
the additional scheduling freedom as well as the knowledge
of reduction dependences.
Reduction-Aware Scheduling
Reduction-enabled scheduling results in generally good
schedules for our benchmark set, however resource constraints
as well as environment effects, both crucial to the overall performance, are not represented in the typical objective function used by polyhedral optimizers. In essence we believe,
the scheduler should be aware of reductions and the cost of
their privatization, in terms of memory overhead as well as
aggregation costs. This is especially true if the scheduler
is used to decide which dimensions should be executed in
parallel or if there are tight memory bounds (e.g., on mobile
devices).
In Section 6.1 we show that the execution environment as
well as the values of runtime parameters are crucial factors in
the actual performance of parallelized code, even more when
reductions are involved. While a reduction-aware scheduler
could propose different parallelization schemes for different
6.1
BiCG Case Study
Polybench is a collection of inherent parallel programs,
there is only one—the BiCG kernel— that depends on re5
This assumes the benchmarks are compiled with -ffastmath, otherwise reductions over floating point computations
are not detected.
4
A reasonable assumption for desktop computers or moderate servers with a low number of parallel compute resources.
7
duction parallelism. To study the effects of parallelization
combined with privatization of multidimensional reductions
in the BiCG kernel we compared two parallel versions to
the non-parallel code Polly would generate without reduction support. The first version “Outer ” has a parallel outermost loop and therefore needs to privatize the whole array
s. The second version “Tile” parallelizes the second outermost loop. Due to tiling, only “tile size” (here 32) locations of the q array need to be privatized. Table 1 shows
the speedup compared to the sequential version for both a
quad core machine and a 8 × 4-core server. As the input
grows larger the threading overhead as well as the interchip communication on the server will cause the speedup of
Tile to stagnate, however on a one chip architecture this
version generally performs best. Outer on the other hand
will perform well on the server but not on the 4-core machine. We therefore believe the environment is a key factor
in the performance of reduction-aware parallelization and a
reduction-aware scheduler is needed to decide under which
run-time conditions privatization becomes beneficial.
7.
dimensions. This is similar to the nested Recur operator introduced by Redon and Feautrier [22, 24]. Hence, reductions
are not only restricted to a single loop dimension, as in other
approaches [11, 6, 25], but can also be multidimensional as
shown in Figure 2.
7.2
RELATED WORK
Reduction aware loop parallelization has been a long lasting research topic. Different approaches to detect reduction,
to model them and finally to optimize them have been proposed. As our work has some intersection with all three
parts we will discuss them in separation.
7.1
Modeling
Modeling reductions was commonly done implicitly, e.g.,
by ignoring the reduction dependences during a post parallelization step [11, 17, 22, 18, 21, 30, 28]. This is comparable
to the reduction-enabled code generation described in Section 5.1. However, we believe the full potential of reductions
can only be exposed when the effects are properly modeled
on the dependence level.
The first to do so, namely to introduce reduction dependences, where Pugh and Wonnacot [20]. Similar to most
other approaches [22, 23, 27, 24, 7, 25, 32], the detection
and modeling of the reduction was performed only on Clike statements and utilizing a precise but costly access-wise
dependence analysis (see the upper part of Figure 8). In
their work they utilize both memory and value-based dependence information to identify statements with an iteration space that can be executed in parallel, possibly after
transformations like array expansion. They start with the
memory-based dependences and compute the value-based
dependences as well as the transitive self-dependence relation for a statement in case the statement might not be
inherently sequential.
Stock et al.[26] describe how reduction properties can be
exploited in the polyhedral model, however neither do they
describe the detection nor how omitting reduction dependences may affect other statements.
In the works of Redon and Feautrier [23] as well as the
extension to that by Gupta et al. [9] the reduction modeling is performed on SAREs on which array expansion [4]
and renaming was applied, thus all dependences caused by
memory reuse were eliminated. In this setting the possible
interference between reduction computation and other statements is simplified but it might not be practical for general
purpose compilers due to memory constraints. As an extension to these scheduling approaches on SAREs we introduced
privatization dependences. They model the dependences between a reduction and the surrounding statements without
the need for any special preprocessing of the input. However, we still allow polyhedral optimizations that will not
only affect the reduction statement but all statements in a
SCoP .
Detection
Reduction detection started with pattern based approaches
on source statements [11, 17, 22, 18, 21, 24] and evolved to
more elaborate techniques that use symbolic evaluation [6],
a data dependency graph [27] or even a program dependency
graph [17] to find candidates for reduction computations.
For functional programs Xu et al. [30] use a type system
to deduce parallel loops including pattern based reductions.
Their typing rules are similar to our detection function (Figure 4) we use to identify reduction-like computations.
Sato and Iwasaki [25] describe a pragmatic system to detect and parallelize reduction and scan operations based on
the ideas introduced by Matsuzaki et al. [15]: the representation of (part of) the loop as a matrix multiplication with
a state vector. They can handle mutually recursive scan
and reduction operations as well as maximum computations
implemented with conditionals, but they are restricted to
innermost loops and scalar accumulation variables. As an
extension Zou and Rajopadhye [32] combined the work with
the polyhedral model and the recurrence detection approach
of Redon and Feautrier [22, 24]. This combination overcomes many limitations, e.g., multidimensional reductions
(and scans) over arrays are handled. However, the applicability is still restricted to scans and reductions representable
in State Vector Update Form [12].
In our setting we identify actual reductions utilizing the
already present dependence analysis, an approach very similar to the what Suganuma et al. [27] proposed to do. However, we only perform the expensive, access-wise dependence
analysis for reduction candidates, and not for all accesses in
the SCoP . Nevertheless, both detections do not need the
reductions to be isolated in a separate loop as assumed by
Fisher and Ghuloum [6] or Pottenger and Eigenmann [18].
Furthermore, we allow the induced reduction dependences
to be of any form and carried by any subset of outer loop
7.3
Optimization
Optimization in the context of reductions is twofold. There
is the parallelization of the reduction as it is given in the input and the transformation as well as possible parallelization
of the input with awareness of the reduction properties. The
first idea is very similar to the reduction-enabled code generation as described in Section 5.1. In different variations,
innermost loops [25], loops containing only a reduction [6,
18] or recursive functions computing a reduction [30] were
parallelized or replaced by a call to a possibly parallel reduction implementation [28]. The major drawback of such
optimizations is that reductions have to be computed either in isolation or with the statements that are part of
the source loop that is parallelized. Thus, the reduction
statement instances are never reordered or interleaved with
8
other statement instances, even if it would be beneficial. In
order to allow powerful transformations in the context of reductions, their effect, hence the reduction dependences, as
well as their possible interactions with all other statement
instances must be known. The first polyhedral scheduling
approach by Redon and Feautrier [23] that optimally6 schedules reduction together with other statements assumed reductions to be computable in one time step. With such
atomic reduction computations there are no reduction statement instances that could be reordered or interleaved with
other statement instances. Gupta et al. [9] extended that
work and lifted the restriction on an atomic reduction computation. As they schedule the instances of the reduction
computation together with the instances of all other statements their work can be seen as a reduction-enabled scheduler that optimally minimizes the latency of the input.
To speed up parallel execution of reductions the runtime
overhead needs to be minimized. Pottenger [18] proposed to
privatize the reduction locations instead of locking them for
each access and Suganuma et al. [27] described how multiple
reductions on the same memory location can be coalesced.
If dynamic reduction detection [21] was performed, different
privatization schemes to minimize the memory and runtime
overhead were proposed by Yu et al. [31]. While the latter
is out of scope for a static polyhedral optimizer, the former
might be worth investigating once our approach is extended
to multiple reductions on the same location.
In contrast to polyhedral optimization or parallelization,
Gautam and Rajopadhye [7] exploited reduction properties
in the polyhedral model to decrease the complexity of a computation in the spirit of dynamic programming. Their work
on reusing shared intermediate results of reduction computations is completely orthogonal to ours.
While Array Expansion, as introduced by Feautrier [4], is
not a reduction optimization, it is still similar to the privatization step of any reduction handling approach. However,
the number of privatization copies the approach introduces,
the accumulation of these private copies as well as the kind
of dependences that are removed differ. While privatization
only introduces a new location for each processor or vector
lane, general array expansion introduces a new location for
each instance of the statement. In terms of dependences,
array expansion aims to remove false output and anti dependences that are introduced by the reuse of memory while
reduction handling approaches break output and flow dependences that are caused by a reduction computation. Because
of the flow dependences—the actual reuse of formerly computed values—the reduction handling approaches also need
to implement a more elaborate accumulation scheme that
combines all private copies again.
Instead we want to allow any transformation possible to our
scheduler with only one restriction: the integrity of the reduction computation needs to stay intact. In other words,
no access to the reduction location is scheduled between the
first and last instance of the reduction statement. This allows our scheduler not only to optimize the reduction statement in isolation, but also to consider other statements at
the same time without the need for any preprocessing to get
a SARE-like input.
To this end we presented a powerful reduction detection
based on computation properties and the polyhedral dependence analysis. Our design leverages the power of polyhedral loop transformations and exposes various optimization
possibilities including parallelism in the presence of reduction dependences. We showed how to model and leverage
associativity and commutativity to relax the causality constraints and proposed three approaches to make polyhedral
loop optimization reduction-aware. We believe our framework is the first step to handle various well-known idioms,
e.g., privatization or recurrences, not yet exploited in most
practical polyhedral optimizers.
Furthermore, we showed that problems and opportunities arising from reduction parallelism (see Section 6.1) have
to be incorporated into the scheduling process, thus the
scheduling in the polyhedral model needs to be done in a
more realistic way. The overhead of privatization and the
actual gain of parallelism are severely influenced by the execution environment (e.g., available resources, number of processors and cores, cache hierarchy), however these hardware
specific parameters are often not considered in a realistic
way during the scheduling process.
Extensions to this work include a working reduction-aware
scheduler and the modeling of multiple reduction-like computations as well as other parallelization preventing idioms.
In addition we believe that a survey about the applicability
of different reduction detection schemes as well as optimization approaches in a realistic environment is needed. In any
case this would help us to understand reductions not only
from the theoretical point of view but also from a practical
one.
8.
10.
9.
CONCLUSIONS AND FUTURE WORK
REFERENCES
[1] M.-W. Benabderrahmane, L.-N. Pouchet, A. Cohen,
and C. Bastoul. The polyhedral model is more widely
applicable than you think. In Proceedings of the 19th
Joint European Conference on Theory and Practice of
Software, International Conference on Compiler
Construction, CC’10/ETAPS’10, pages 283–303,
Berlin, Heidelberg, 2010. Springer-Verlag.
[2] G. E. Blelloch. Scans as primitive parallel operations.
IEEE Trans. Comput., 38(11):1526–1538, Nov. 1989.
[3] U. Bondhugula, M. M. Baskaran, S. Krishnamoorthy,
J. Ramanujam, A. Rountev, and P. Sadayappan.
Earlier work already utilized reduction dependences in different varieties, depending on how powerful the detection
was. Whenever reductions have been parallelized the reduction dependences have been implicitly ignored, in at least
two cases they have even been made explicit [20, 26]. However, to our knowledge, we are the first to add the concept
of privatization dependences in this context. The reason is
simple: we believe the parallel execution of a loop containing a reduction is not always the best possible optimization.
6
ACKNOWLEDGMENTS
We would like to thank Tobias Grosser, Sebastian Pop
and Sven Verdoolaege for the helpful discussions during the
development and implementation of this approach. Furthermore, we want to thank the reviewers who not only provided
extensive comments on how to further improve this work but
also pointed us to related work that was previously unknown
to us. Lastly, we would like to thank Tomofumi Yuki for giving us many helpful tips.
e.g., according to the latency
9
Figure 8: Evaluation results for Polybench 3.2. In the upper part the compile time for different grained dependence analyses
is shown, in the lower part the speedup of Polly with reduction support compared to Polly without reduction support.
[4]
[5]
[6]
[7]
[8]
[9]
Automatic transformations for
communication-minimized parallelization and locality
optimization in the polyhedral model. ETAPS ’08.
P. Feautrier. Array expansion. In Proceedings of the
2Nd International Conference on Supercomputing, ICS
’88, pages 429–441, New York, NY, USA, 1988. ACM.
P. Feautrier. Dataflow analysis of array and scalar
references. International Journal of Parallel
Programming, 20(1):23–53, 1991.
A. L. Fisher and A. M. Ghuloum. Parallelizing
complex scans and reductions. In Proceedings of the
ACM SIGPLAN 1994 Conference on Programming
Language Design and Implementation, PLDI ’94,
pages 135–146, New York, NY, USA, 1994. ACM.
Gautam and S. Rajopadhye. Simplifying reductions.
In Conference Record of the 33rd ACM
SIGPLAN-SIGACT Symposium on Principles of
Programming Languages, POPL ’06, pages 30–41, New
York, NY, USA, 2006. ACM.
T. Grosser, A. Größlinger, and C. Lengauer. Polly performing polyhedral optimizations on a low-level
intermediate representation. Parallel Processing
Letters, 22(4), 2012.
G. Gupta, S. Rajopadhye, and P. Quinton. Scheduling
reductions on realistic machines. In Proceedings of the
Fourteenth Annual ACM Symposium on Parallel
Algorithms and Architectures, SPAA ’02, pages
117–126, New York, NY, USA, 2002. ACM.
[10] P. Jouvelot. Parallelization by semantic detection of
reductions. In Proc. Of the European Symposium on
Programming on ESOP 86, pages 223–236, New York,
NY, USA, 1986. Springer-Verlag New York, Inc.
[11] P. Jouvelot and B. Dehbonei. A unified semantic
approach for the vectorization and parallelization of
generalized reductions. In Proceedings of the 3rd
International Conference on Supercomputing, ICS ’89,
pages 186–194, New York, NY, USA, 1989. ACM.
[12] P. M. Kogge and H. S. Stone. A parallel algorithm for
the efficient solution of a general class of recurrence
equations. IEEE Trans. Comput., 22(8):786–793, Aug.
1973.
[13] M. Kong, R. Veras, K. Stock, F. Franchetti, L.-N.
Pouchet, and P. Sadayappan. When polyhedral
transformations meet simd code generation. In
Proceedings of the 34th ACM SIGPLAN Conference
on Programming Language Design and
Implementation, PLDI ’13, pages 127–138, New York,
NY, USA, 2013. ACM.
[14] C. Lattner and V. Adve. Llvm: A compilation
framework for lifelong program analysis &
transformation. In Proceedings of the International
Symposium on Code Generation and Optimization:
Feedback-directed and Runtime Optimization, CGO
’04, pages 75–, Washington, DC, USA, 2004. IEEE
Computer Society.
[15] K. Matsuzaki, Z. Hu, and M. Takeichi. Towards
automatic parallelization of tree reductions in
10
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
dynamic programming. In Proceedings of the
Eighteenth Annual ACM Symposium on Parallelism in
Algorithms and Architectures, SPAA ’06, pages 39–48,
New York, NY, USA, 2006. ACM.
S. P. Midkiff. Automatic Parallelization: An Overview
of Fundamental Compiler Techniques. Synthesis
Lectures on Computer Architecture. 2012.
S. S. Pinter and R. Y. Pinter. Program optimization
and parallelization using idioms. In Proceedings of the
18th ACM SIGPLAN-SIGACT Symposium on
Principles of Programming Languages, POPL ’91,
pages 79–92, New York, NY, USA, 1991. ACM.
B. Pottenger and R. Eigenmann. Idiom recognition in
the polaris parallelizing compiler. In Proceedings of the
9th International Conference on Supercomputing, ICS
’95, pages 444–448, New York, NY, USA, 1995. ACM.
W. Pugh. Uniform techniques for loop optimization.
In Proceedings of the 5th International Conference on
Supercomputing, ICS ’91, pages 341–352, New York,
NY, USA, 1991. ACM.
W. Pugh and D. Wonnacott. Static analysis of upper
and lower bounds on dependences and parallelism.
ACM Trans. Program. Lang. Syst., 16(4):1248–1278,
July 1994.
L. Rauchwerger and D. Padua. The lrpd test:
Speculative run-time parallelization of loops with
privatization and reduction parallelization. In
Proceedings of the ACM SIGPLAN 1995 Conference
on Programming Language Design and
Implementation, PLDI ’95, pages 218–232, New York,
NY, USA, 1995. ACM.
X. Redon and P. Feautrier. Detection of recurrences in
sequential programs with loops. In Proceedings of the
5th International PARLE Conference on Parallel
Architectures and Languages Europe, PARLE ’93,
pages 132–145, London, UK, UK, 1993.
Springer-Verlag.
X. Redon and P. Feautrier. Scheduling reductions. In
Proceedings of the 8th International Conference on
Supercomputing, ICS ’94, pages 117–125, New York,
NY, USA, 1994. ACM.
X. Redon and P. Feautrier. Detection of scans in the
polytope model. Parallel Algorithms Appl.,
15(3-4):229–263, 2000.
S. Sato and H. Iwasaki. Automatic parallelization via
matrix multiplication. In Proceedings of the 32Nd
ACM SIGPLAN Conference on Programming
Language Design and Implementation, PLDI ’11,
pages 470–479, New York, NY, USA, 2011. ACM.
K. Stock, M. Kong, T. Grosser, L.-N. Pouchet,
F. Rastello, J. Ramanujam, and P. Sadayappan. A
framework for enhancing data reuse via associative
reordering. In Proceedings of the 35th ACM SIGPLAN
Conference on Programming Language Design and
Implementation, PLDI ’14, pages 65–76, New York,
NY, USA, 2014. ACM.
T. Suganuma, H. Komatsu, and T. Nakatani.
Detection and global optimization of reduction
operations for distributed parallel machines. In
Proceedings of the 10th International Conference on
Supercomputing, ICS ’96, pages 18–25, New York, NY,
USA, 1996. ACM.
[28] A. Venkat, M. Shantharam, M. Hall, and M. M.
Strout. Non-affine extensions to polyhedral code
generation. In Proceedings of Annual IEEE/ACM
International Symposium on Code Generation and
Optimization, CGO ’14, pages 185:185–185:194, New
York, NY, USA, 2014. ACM.
[29] S. Verdoolaege. Isl: An integer set library for the
polyhedral model. In Proceedings of the Third
International Congress Conference on Mathematical
Software, ICMS’10, pages 299–302, Berlin, Heidelberg,
2010. Springer-Verlag.
[30] D. N. Xu, S.-C. Khoo, and Z. Hu. Ptype system: A
featherweight parallelizability detector. In In
Proceedgins of 2nd Asian Symposium on Programming
Languages and Systems (APLAS 2004), LNCS 3302,
pages 197–212. Springer, LNCS, 2004.
[31] H. Yu, D. Zhang, and L. Rauchwerger. An adaptive
algorithm selection framework. In Proceedings of the
13th International Conference on Parallel
Architectures and Compilation Techniques, PACT ’04,
pages 278–289, Washington, DC, USA, 2004. IEEE
Computer Society.
[32] Y. Zou and S. Rajopadhye. Scan detection and
parallelization in ”inherently sequential” nested loop
programs. In Proceedings of the Tenth International
Symposium on Code Generation and Optimization,
CGO ’12, pages 74–83, New York, NY, USA, 2012.
ACM.
11
| 6cs.PL
|
arXiv:1208.3773v1 [cs.DC] 18 Aug 2012
Haskell#: Coordinating Functional Processes
Francisco Heron de Carvalho-Junior
Departamento de Computação
Universidade Federal do Ceará (UFC)
Fortaleza, Brazil
Rafael Dueire Lins
Centro de Informática
Universidade Federal de Pernambuco
Recife, Brazil
June 2004
Abstract
This paper presents Haskell# , a coordination language targeted
at the efficient implementation of parallel scientific applications on
loosely coupled parallel architectures, using the functional language
Haskell. Its programming environment encompasses an editor, a compiler into Petri nets, a Petri net animator and proof tool, and a skeleton library. Examples of applications, their implementation details
and performance figures are presented.
1
Introduction
The peak performance of parallel architectures is growing at a faster pace
than predicted by Moore’s law, that states that at each 18 months computer hardware becomes twice as fast and halves its sale price. However,
parallel programming tools have not being able to reconcile portability, scalability and a higher level of abstraction without imposing severe performance
penalties to applications [28].
1
The emerging technologies in the 1990s gave birth to new challenges in
high-performance computing. The advent of clusters [8], low cost supercomputers built on top of networks of workstations and personal computers,
disseminated supercomputing among academic institutions, industries and
companies [11, 4, 19, 20]. More recently, advances in wide area network interconnection technologies have made possible to use their infra-structure to
build distributed supercomputers of virtually infinite scale, the grids, which
are particularly suitable for addressing very coarse grained scientific computing applications. Great efforts to make these technologies viable are been
promoted, with promising results [34].
Clusters and grids sparkled a myriad of new applications in supercomputing for scientific computation. Most of them are not addressed adequately by
contemporary tools, yielding inefficient distribution of parallel programs [83].
In [10], some parallel programming approaches used in scientific computing
are compared in relation to scalability (efficiency), generality and abstraction
dimensions. MPI (Message Passing Interface) [63], the most widespread message passing library, provides scalability, generality, but is less abstract than
TCE (Tensor Contraction Engine) [7], PETSc [5], GA (Global Array) [66],
openMP [67], auto-parallelized C/Fortran90 and HPF (High Performance
Fortran) [32]. PETSc and TCE are specific purpose libraries for scientific
computing, providing a high level of abstraction and scalability. Implicit
approaches, such as C/Fortran90, present low scalability, high level of abstraction and high generality. These observations illustrate that, despite the
efforts conducted on the last decade, the need for new parallel programming
environments that reconcile a high-level of abstraction, modularity, and generality with scalability and peak performance is still a challenge [28, 77, 82].
This paper presents Haskell# , a process-oriented coordination language
[35] where Haskell [75], a language considered de facto a standard in lazy
functional programming, is used for programming at computation level. Haskell#
aims to provide high-level programming mechanisms without sacrificing performance significantly, by minimizing the overheads of the management of
parallelism. One of the most important concerns in Haskell# is to make
easier to prove correctness of programs. For that, a divide-and-conquer approach was adopted to increase the chances of formally analyzing programs:
the process network is completely orthogonal to the sequential blocks of code
(process functionality). Haskell allows sequential programs to be proved correct in a simpler fashion than their equivalent written in languages which
belong to other programming paradigms. The communication primitives
2
were designed in such a way as to allow their translation into Petri nets [72],
a well reputed formalism for the specification of concurrent systems, with
several analysis and verification tools [80, 12] available.
Haskell# emphasizes compositional programming and provides support
for skeletons [25]. Skeletons are used to expose topological information that
can guide the Haskell# compiler in the generation of more efficient code.
MPI (Message Passing Interface) [29] is used to manage parallelism without
claiming for any run-time support. Due to the recent development of interoperable [84] and grid enabled [49] versions of MPI Haskell# programs may
be executable on grids without any extra burden.
Examples of benchmark programs and their performance figures are provided, elucidating the most important aspects of programming in Haskell# .
This paper comprises five other sections. Section 2 gives background for
programming in Haskell# , focusing on programming abstractions. Section 3
presents motivating application examples of Haskell# programming. Section
4 presents details about current implementation of Haskell# for clusters. Section 5 presents performance figures about applications presented in Section 3
running on implementation described in Section 4. Section 6 concludes this
paper outlining the work in progress with Haskell# .
2
Programming in Haskell#
Haskell# programs are composed from a set of components, each one describing an application concern. Concerns may be functional or non-functional.
Examples of functional concerns are the calculation of an exact solution for
a system of linear equations and the calculation of a finite-difference approximation for a system of partial differential equations. An example of nonfunctional concern is the allocation of processes to processors. Components
may be reused among Haskell# programs.
In Haskell# programming, the process of composing components is inductive. Simple components, functional modules implemented in Haskell,
are basic building blocks. Given a collection of components, simple or composed ones, it is possible to define a new composed component by specifying
their composition through Haskell# Configuration Language (HCL). The result of this process is a hierarchy of components, where the main component,
describing the application functionality, is at the root. Components at the
leaves are simple components (always addressing functional concerns) and
3
intermediate nodes are composed components.
Under perspective of process-oriented coordination models [35], the collection of functional modules of a Haskell# program forms a computation
medium, while the collection of composed components forms a coordination
medium. The concerns on the parallel composition of Haskell functional computations are sufficiently and necessarily resolved at the coordination level.
The use of Haskell for programming the computation medium allows that
coordination and computation languages be really orthogonal. Lazy lists
allow the overlapping of communication and computation in process execution, without to need to embed coordination extensions in the code of the
functional modules.
The idea of hierarchical compositional languages implemented using configuration languages is not a recent idea [13, 1]. Haskell# difference from
its predecessor languages resides in its support for skeletons, by allowing to
partially parameterize the concern addressed by components, and its ability
for overlapping them, making possible to encapsulate cross-cutting concerns
[21]. The use of skeletons has gained attention of parallel programming community since last decade [25] and now it is supported by many languages
and paradigms [79]. The problem of modularizing cross-cutting concerns
have gained attention in software engineering research community, particularly for programming large scale object-oriented systems. An example
of cross-cutting concern is validation procedures executed by processes for
accessing computational resources in a grid environment. With respect to
this feature, Haskell# may implement the notions of AOP (Aspect Oriented
Programming) [52] and Hyperspaces [68] using an unified set of language
constructors. Skeletons may be overlapped, forming more complex ones.
Haskell# programs may be translated into Petri nets. This allows to
prove formal properties and to evaluate the performance of parallel programs
using automatic tools. Some previous work have addressed the problem of
translating Haskell# programs into Petri nets [56, 23]. The expressive power
of HCL for describing patterns of interaction among processes is equivalent
to descriptive power of labelled Petri nets [71].
Now, relevant details about how Haskell# programs are implemented
are presented. HCL abstractions for programming at coordination medium
are informally introduced and it is shown how simple components are programmed in Haskell. Motivating examples of Haskell# are presented in Section 3, illustrating the use of Haskell# programming abstractions. Appendix
B formalizes an algebra for describing semantics of Haskell# programming
4
tallies
tallying[1]
events
tracking[1]
tallies
tallying[2]
events
tracking[2]
tallies
tallying[3]
events
tracking[3]
events
tallies
tallying[4]
particles
user information
prob_def
tally entries
tally entries
tally entries
tally entries
tracking[4]
statistics[1]
statistics[2]
recip
average energy
tally entries
Figure 1: Process Network of MCP-Haskell#
abstractions. The informal description points at the corresponding Haskell#
algebraic constructions.
2.1
Programming Composed Components
Composed components, which form coordination medium of Haskell# programs, are programmed in HCL configurations. HCL programming corresponds to the inductive step in Haskell# programming task described in
last section. In what follows, the constructors used at coordination level
for programming Haskell# applications are informally introduced. Their for5
01.
02.
03.
04.
05.
06.
07.
08.
09.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
component MCP<n,m> with
iterator i range [1,n]
use Skeletons.{PipeLine, Workpool}
interface IProbDef () → (user info, particles, tally entries, recip, avg e, all tallies) where: IDispatcher () → particles
behavior: seq { recip!; avg e!; all tallies!; tally entries!; user info!;
repeat particles! until particles }
interface ITracking (user info,particles*) → (events*, totals) where: IPipeStage particles → events
behavior: seq { user info?; do process particles; totals! }
interface ITallying (tally entries, events*) → tallies* where: IPipeStage events → tallies
behavior: seq { tally entries?; do process events }
interface IStatistics (avg e, recip, totals, tallies*) → () where: ICollector tallies → ()
behavior: seq { avg e?; recip?; all tallies?; repeat tallies? until tallies; totals? }
unit pp; assign PipeLine<2> to pp
unit wp; assign WorkPool<n> to wp
unit
unit
unit
unit
prob def
track
tally
statistics
#
#
#
#
IProbDef wire tally entries all*2: distribute; assign ProbDef
ITacking
; assign Tracking
ITallying
; assign Tallying
IStatistics
; assign Statistics
to
to
to
to
prob def
track
tally
statistics
factorize wp.manager in → out to dispatcher # () → out, collector # in → ()
replace
replace
replace
replace
dispatcher
pp.stage[1]
pp.state[2]
collector
#
#
#
#
tallies →
particle→
events →
tallies →
particles
events
tallies,
(),
by
by
by
by
prob def
track
tally
statistics
#
#
#
#
(
(
(
(
,particles, , , , )
, particles) → (events, )
,events) → tallies
, , ,tallies) → () to manager
replicate pp into n; [/ replace wp.worker[i] by pp[i] /]
connect
connect
connect
connect
connect
connect
prob def→user info
to
prob def→tally entries[0] to
prob def→tally entries[1] to
to
prob def→recip
prob def→avg e
to
tracking→totals
to
tracking←user info,
tallies←tally entries,
statistics←tally entries,
statistics←recip,
statistics←avg e,
statistics←totals,
synchronous
synchronous
synchronous
buffered
buffered
buffered
replicate m statistics # (avg e, recip, totals, tallies,tally entries) → ()
adjust wire avg e: broadcast, recip: broadcast, totals: {# (map.sum.transpose) #}
tally entries<>: distribute, tallies<>: broadcast
Figure 2: HCL Code for MCP-Haskell#
mal syntax is presented in Appendix A. Appendix B brings their algebraic
semantics.
MCP-Haskell# A parallel version of MCP-Haskell [22] is used for exemplifying the syntax of HCL. MCP-Haskell [39] is a simplified sequential version
of MCNP, a code developed at Los Alamos during many years for simulating
the statistical behaviour of particles (photons, neutrons, electrons, etc.) while
they travel through objects of specified shapes and materials [15]. HCL code
of MCP-Haskell# is shown in Figure 2. The corresponding network topology is presented in Figure 1. The parallelism is obtained from three sources.
Firstly, tracking and tallying procedures must be executed concurrently using
6
Unit
cluster
Composed Component
Simple Component
Group of
Input Ports
process
Individual
Input Port
Channel
...
...
λ
...
Individual
Output port
Group of
Output Ports
Associating an argument
with an input port
Associating a return point
with an output port
Binding a port to an
argument/return point
Figure 3: Diagrammatic Notation for Haskell# Abstractions
a pipe-line. The main source of parallelism is the second. It comes from the
fact that particles may be tracked and tallied independently. To take advantage of this problem feature, a work pool pattern of interaction is employed,
where a manager process distributes jobs (particles) to worker processes, on
demand controlled by their availability, and collects the results from each
job. A third source of parallelism comes from the fact that the statistics of
different tallies may be computed in parallel. Thus, each statistical process
in the network is responsible for computing a specified set of tallies. In the
following sections, it is explained how a HCL configuration may implement
this network topology.
A HCL configuration starts with a header, declaring the name of the
composed component, its static formal parameters and its arguments and
return points. MCP-Haskell# has two static parameters, m and n, which
controls the number of parallel tasks, but no argument or return point is
defined. In general, arguments and return points are not defined for the main
component of an application. They are normally used in the configuration
of the encapsulated functional concerns.
7
2.1.1
The Basic Abstractions: Units and Channels
A Haskell# configuration is specified by a collection of units, which are abstractions for agents that execute a particular task. Units synchronize using
communication channels. The task performed by a unit is defined statically,
by assigning a component for it. Units may be viewed as a “glue” for composing components. Units have interfaces, comprising collections of input
and output ports. Interfaces are necessary for allowing units to be connected
through communication channels. An interface also describes a partial order
for the activation of ports during execution, characterizing the behavior of
a unit. A communication channel is defined by linking two ports from opposite directions through a communication mode: synchronous, buffered
and ready. Communication modes of Haskell# channels have direct correspondence to MPI primitives, ensuring their efficient implementation, and
have semantic equivalence with OCCAM [46] and CSP [43]. Ports linked
through a communication channel are said to be communication pairs.
In Figure 2, lines 20 to 26 have declarations of units, whose identifiers are
placed after the keyword unit. The assign declarations bind components
to units. The interface of a unit is declared after the clause “#”. In the
example, an interface class identifier is employed but it is possible to declare
an interface directly. This topic is discussed further in the next section.
The low level of abstraction provided by units, ports and channels is not
appropriate for programming large-scale and complex distributed parallel
programs. Next sections introduce additional abstractions intended to raise
the level of abstraction in HCL programming, simplifying the specification of
large-scale and complex process topologies. Essentially, they provide support
for partial topological skeletons.
2.1.2
Interface Classes
Haskell# incorporates the notion of interface class for representing interfaces of units that present equivalent behavior. Examples of interface class
declarations are shown in lines 07 to 18 of Figure 2. The identifier of an
interface class is configured after the interface keyword. The notation
(i1 , i2 , . . . , in ) → (o1 , o2 , . . . , om ) sets up n input ports and m output ports,
with the respective identifiers. In a where clause, an interface composition
operator (#) allows defining how an interface is obtained from the composition of existing ones. The semantics of the # operator is formalized in
8
Appendix B.
Units that declare the same interface name after “#” clause in unit declarations inherit the same behavior, specified in the corresponding interface
declaration.
A small language is embedded in behavior clause of interface declarations, intended to describe partial orders in the activation of ports. Its
combinators have semantic equivalence to operators of regular expressions
controlled by semaphores [47], which are regular expressions enriched with
an interleaving operator, represented in HCL by the combinator par, and
counter semaphores primitives, represented by the primitives wait and signal. This feature ensures that the HCL descriptive power is equivalent to the
power of terminal labelled Petri nets in describing the interaction patterns
between processes.
2.1.3
Wire Functions
In an assignment declaration, it is necessary to map input and output ports
of the unit to arguments and return points of the assigned component, respectively. The notation (i1 , i2 , . . . , in ) → (o1 , o2 , . . . , om ) may be used whenever the order of ports does not match the order of corresponding arguments/return points.
In fact, the association between the input and output ports and the arguments and return points of components in assign declarations defines how
Haskell# glues coordination and computation media. Whenever an argument is not bound to an input port, an explicit value must be provided to
it. Also, whenever a return point is not associated with an output port, it is
not evaluated.
In wire clauses of unit declarations, HCL allows to define a wire function
that maps a value received through an input port onto a value that is passed
to an argument. Analogously, it is possible to define a wire function that
receives a value produced at a return point yielding another value that is
sent through the associated output port. This increases the chances that a
component be reused without changing its internal implementation, in such
cases where there is some type incompatibility between the type or meaning
of arguments and the return points and the expected input and output ports
types and meaning at coordination level.
9
individual ports
wire
function
?
wire
function
!
All group of ports
?
?
wire
function
?
!
!
...
...
wire
function
!
Any group of ports
wire
function
wire
function
!
... ...
?
... ...
choosen port
choosen port
Figure 4: Wire Functions and Groups of Ports
2.1.4
Groups of Ports
Another useful feature of HCL is the replication of interface ports of a unit,
forming groups of ports where individual members are referenced using enumeration indexes. A group of ports is treated as an individual entity from
the local perspective of the unit. Thus, they are bound to a unique argument/return point and must be activated atomically. However, from a global
view, individual ports of the group are treated in separate, being possible to
connect them through different channels.
Groups of ports may be of two kinds: any or all. When a group of
input ports of kind all is activated, each port member must receive a value.
The array of values received is mapped to a unique value by using a wire
function. Then, the value is passed to the argument mapped onto the group
of ports. When an output group of ports is activated, the value yielded at
the return point mapped to it is transformed, using an wire function, into
an array of values that are sent through port members of the group. In
activation of groups of ports of kind any, one port belonging to the group is
chosen among ports whose communication pair is activated. Once the port
is chosen, communication occur like in individual ports. Notice that wire
functions are necessary for configuring groups of ports of kind all. Because
of that, groups of ports are configured in clause wire of unit declarations,
like exemplified in Figure 2 with tally entries group of ports. For configuring
a group of ports of kind any, use any keyword instead of all keyword, as
illustrated in the example. Figure 4 illustrate semantics of wire functions
10
and groups of ports.
2.1.5
Stream Ports
Stream ports allows to transmit sequences of values (streams) terminated
by an end marker. Haskell# streams may be nested (streams of streams) at
arbitrary nesting levels, which must be statically configured. Stream ports
of units for which it is assigned a simple component must be mapped to
argument and return points of lazy lists types in the functional module.
Nested streams are associated to nested lazy lists of at least the same nesting
level.
In interface declarations in lines 11 and 14, stream ports may be identified by the occurrence of sequences of symbols “*” after the identifier of
the port. The number of *’s indicates its nesting level. For instance, stream
ports particles and events of interface ITracking have nesting level equal to
one. In Figure 9, where Haskell code of the functional module Tracking is
presented, arguments and return points associated to particles and events
ports of track unit are lazy lists of nesting level greater than or equal to one.
Stream ports are essential for Haskell# expressiveness, once it allows overlapping communication operations and computations during the execution of
processes. The same approach is used by other parallel functional languages,
such as Eden [14].
2.1.6
Configuring Arguments and Return Points of Composed
Components
Arguments and return points of composed components are, respectively, input and output ports of units that are not connected through any communication channel. For speciying ports that must be connected to arguments
and return points, HCL supports bind declarations.
2.1.7
The Distinction Between Processes and Clusters
It is convenient to distinguish between units associated to simple and composed components. The former are called processes, while the latter are
called clusters. Processes are concrete entities and may be viewed as agents
that perform sequential computations programmed in Haskell. Clusters are
abstract entities and must be viewed as a parallel composition of processes.
The abstraction of clusters is essential for expressing hierarchical parallelism.
11
For example, in a constellation architecture 1 , a cluster must be associated
with a multiprocessing node, in such a way that its comprising processes are
allocated to processors for shared memory parallel execution. Instead of generating MPI code, the Haskell# compiler could generate openMP [67] code
for implementing communications among processes inside the cluster, more
appropriate for multiprocessors. The support for multiple hierarchies of parallelism is essential for grid computing architectures [48] and is recognized as
an important challenge for parallel programming languages designers [9].
In MCP-Haskell# specification, pp and wp are clusters, units respectively
associated to composed components PipeLine and Workpool, which represent skeletons. Units prob def, tally, track and statistics are declared as
processes. The components assigned to these units are functional modules,
written in Haskell.
2.1.8
Termination of Haskell# Programs
Units may be declared as repetitive or non-repetitive. Non-repetitive units
perform a task and go to their final state, while repetitive ones always go
back to their initial state, for executing its task once more. In HCL, a unit
is declared repetitive by placing a symbol “*” after the keyword unit in its
declaration. For declaring a cluster as repetitive, all units belonging to the
composed component assigned to it must be repetitive. Otherwise, an error
is detected and informed by HCL compiler.
A Haskell# program terminates whenever all non-repetitive units belonging to its main component terminates. If it has only repetitive units, it does
not terminate. Repetitive units may be used to model reactive applications.
A non-stream port of a repetitive unit may be connected to a stream port
of a non-repetitive unit. Each value produced in the stream is consumed in
an execution of the repetitive process.
2.1.9
Virtual Units and The Support for Skeletons
A skeletons was defined above as a composed component where its addressed
concern is partially defined or totally undefined. Now that the structure of
composed components was scrutinized, it is possible to define Haskell# skeletons in more precise terms. In fact, the concern addressed by a composed
1
Constellations have been defined as clusters of multiprocessor nodes with at least
sixteen processors per node [9, 28].
12
Interface or Interface Class
Assignmennt
Instantiation
interface declaration
Unit (non−virtual)
Virtual Unit
unit declaration
assign declaration
Figure 5: Configuration of Units
component is defined by the composition of concerns addressed by components assigned to its comprising units. If some unit of a component does not
have a component assigned to it, it is said that the component is partially
parameterized by its addressed concern. This kind of component is called
a partial topological skeleton. Units not assigned to a component are called
virtual units. In other skeleton-based languages, skeletons are usually total,
in the sense that all units are virtual. After instantiating a partial topological skeleton, or simply a skeleton, by assigning it to a unit comprising
a configuration of a component, it is possible to assign components to the
virtual units of the skeleton, configuring its addressed concern.
The components Farm and Workpool are examples of total skeletons.
They are used for structuring the topology of the MCP-Haskell# program.
They are instantiated by assigning them to units pp and wp, respectively.
The replaces declaration, exemplified in lines 30 to 33 of Figure 2, takes a
virtual unit from a skeleton and replaces it by another unit, such that there is
an homomorphism relation from interface of the original unit to the interface
of the new unit. This is formalized in Appendix B by the pair of relations
⊑ and ⊒ between interfaces. Indeed, replacing declarations are syntactic
sugaring of HCL. The same effect could be obtained by creating a new unit,
unifying it with the skeleton unit and assigning the appropriate component to
the resulting unit. For that reason, replacing declarations are not formalized
in Appendix B. This topic is revisited in the next section, where unification
is introduced.
2.1.10
Operations over Virtual Units and Overlapping of Skeletons
Two operations are useful for the specification of complex topologies through
the composition of skeletons: unification and factorization. Unification sub-
13
A
C
e
b
F
d
F
unification
A
CDE
B
D
e
a
a
e
b
G
B
d
b
G
c
factorization
E
c
Figure 6: An Illustrative Example of Unification/Factorization
stitutes a collection of virtual units by a new virtual unit, obeying the network connectivity and behavioral preserving restrictions formalized in the
Appendix B. In this process, ports, individual or groups, may be grouped.
To group groups of ports involves to merge their sets of ports. Factorization
performs inverse operation of unification. It takes a unit and splits it in
a collection of units, also respecting behavioral and networking connectivity
preserving restrictions. It may be needed to replicate communication pairs of
interface ports of a factorized unit for preserving network connectivity. Thus,
it is also necessary to configure wire functions whenever a new group of ports
is resulted from a factorization. For that, HCL provides clause adjust wire
in unification and factorization declarations.
In Figure 6, illustrative abstract examples of unification are presented,
illustrating duality between these operations. A more concrete example of
factorization is presented in line 28 of Figure 2, where manager unit from
Workpool skeleton is split up into units dispatcher and collector, dividing
tasks realized by the manager. Unification does not appear directly in example of Figure 2. But replacing declarations, like discussed in the last section,
is a syntactic sugaring of HCL that may be defined using unification. For
instance, consider replacing declaration in line 31. It can be rewritten using
the following equivalent code:
unit track’ # ITracking
..
.
unify pp.stage[1] # particle → events, track’ # (user info, particles) → (events, totals)
to track # ITracking (user info, particles) → (events, totals)
..
.
assign Tracking to track
14
A
A[1]
B
B[1]
A[2]
B[2]
replication
C
C
D
D
Figure 7: An Example of Replication
Unification, and consequently replacing declarations, allows for overlapping skeletons. In this sense, units from distinct skeletons may be unified
forming a new unit. Overlapping of skeletons is not supported by other
skeleton-based languages. In general, only nesting composition has been addressed and cost models have been defined incorporating this feature [38].
A further step is to work on defining new cost models that incorporate the
overlapping of skeletons.
2.1.11
Replicating Units
Another useful feature of HCL is to support replicate sub-networks from
the overall network of the units described by the configuration. For that,
a collection of units to be replicated and a natural number greater than
one are provided. Network preserving restrictions must be observed, making
necessary to replicate communication pairs of interface ports of a replicated
unit, like in factorization. Wire functions must be provided to resulted groups
of ports using the adjust wire clause.
Replication is exemplified in line 35 of Figure 2. The unit pp is replicated
into n units (pp[i], 0 ≤ i ≤ n − 1), which replace worker units of Workpool
skeleton. Figure 7 presents an illustrative example.
2.1.12
Indexed Notation
The # configuration language supports a special kind of syntactic sugaring
for allowing to declare briefly large collections of entities. The iterator
declaration employs one or more indexes and their ranging values. Syntactic
elements that appear enclosed in [/ and /] delimiters (variation scopes) are
unfolded, according to range of indexes that appears on its scope. The #
compiler incorporates a pre-processor for unfolding indexed notation.
In Figure 2, an iterator i is declared, varying from 1 to n. The replacing
15
a1
a2
an
main :: a1 −> a2 −> ... −> an −> (r1, r2, ..., rm)
r1
r2
rm
Figure 8: Simple Components in Haskell
declaration in line 35 is put in context of a variation scope. Thus, it may be
unfolded in the following code, assuming that n = 3:
replace wp.worker[1] by pp[1]; replace wp.worker[2] by pp[2]; replace wp.worker[3] by pp[3]
2.2
Programming Simple Components
Simple components are programmed using standard Haskell. No extensions
are necessary to Haskell for gluing functional modules in the coordination
medium. They are connected to units at the coordination medium by assignment declarations, where a mapping between ports of the unit interface
and argument/return points of the component is defined. Arguments of a
functional module are represented by the collection of arguments of its function named main, while return points are represented by the elements of the
returned tuple. The general signature of main is shown in Figure 8.
The main function may return values in the IO monad [90], but the I/O
concerns may be resolved at coordination level using a skeleton that implements an I/O aspect, an example of cross-cutting non-functional concern.
Figure 9 presents the Haskell code for the functional module Tracking
of MCP-Haskell# . Notice the correspondence of the arguments and return
points with the ports of the unit track. Functional modules are programmed
in pure Haskell. There is no reference in the computation code for any
16
module Tracking(main) where
import Track
import Tallies
import Mcp types
main :: User spec info → [(Particle,Seed)] → ([[Event]],[Int])
main user info particle list = let events’s = map f particle list in (events’s, tally bal event lists)
where
f (particle@( , , , e, ), sd) = (Create source e):(track user info particle [] sd)
Figure 9: A Functional Module from MCP-Haskell#
element declared at the coordination level of abstraction. Other examples
of functional modules, enforcing these characteristics, are provided in Figure
13.
2.3
Haskell# in the Parallel Functional Languages Context
Some authors have written papers on the evolution of parallel functional
languages [57, 41, 87]. It is convenient to analyze the evolution of parallel
functional programming by dividing it into two periods [57]. In first one, the
decades of 1970 and 1980, parallelism was viewed as possibility to make functional programs run faster. After that period, functional programming techniques have been viewed as a promising alternative to promote higher-level
parallel programming, mainly motivated by the use of skeletons implemented
using higher order functions [25].
The first attempts to embed the support for parallelism in functional languages suggested the technique of evaluating function arguments in parallel,
with the possibility of functions absorbing unevaluated arguments and perhaps also exploiting speculative evaluation [16]. However, the granularity
of the parallelism obtained from referential transparency in pure functional
languages is too fine, not yielding good performance on distributed architectures. Techniques for controlling granularity, either statically or dynamically,
produced little success in practice [44, 73, 50]. Implicit parallelizing compilers face difficulties to promote good load balancing amongst processors and
to keep the communication costs low. On the other hand, explicit parallelism
17
with annotations to control the demand of the evaluation of expressions, the
creation/termination of processes, the sequential and parallel composition of
tasks, and the mapping of these tasks onto processors yielded better results
[18, 51, 45, 76, 14, 86]. GpH adopts a semi-explicit approach, where programmers may annotate the code, but responsibility to decide when to evaluate
expressions in parallel is left to the compiler. Explicit approaches have the
disadvantage of cross-cutting the computation and the communication code,
not allowing to reason about these elements in isolation. Skeleton-based approaches have obtained a relative success in parallel functional programming
[26, 42, 64, 40].
The coordination paradigm [35] influenced the design of parallel functional languages in 1990s, being exploited from two perspectives. In the first
one, it is used for abstracting parallel concerns from specification of computations. Eden [14], Caliban [85], and Haskell# focuses on these ideas. In the
second one, a higher-order and non-strict style of functional programming
has been seen as a convenient way for specifying the coordination amongst
tasks. SCL [26] and Delirium [61] are examples of languages that employ
the functional paradigm at coordination level, describing computations using languages from other paradigms. Haskell# have other similarities with
Eden and Caliban besides adopting the coordination paradigm and Haskell
for describing computations. They all use constructors for explicit specification of network topologies where processes communicate through point-topoint and unidirectional channels. Like Eden, Haskell# employs lazy lists for
interleaving computation and computation and is strict in communication.
Higher order values can not be transmitted through channels. Eden includes
functionalities for specifying dynamic topologies, contraryse to Caliban and
Haskell# . Static parallelism is an important premise of Haskell# design, since
it is intended to analyze Haskell# programs by translating them into Petri
nets. Also, Haskell# is oriented for high performance computing, where static
parallelism is a reasonable assumption, and not for general concurrency. In
the next paragraphs, some important distinguishable features Haskell# are
discussed.
The Adoption of a configuration based approach for coordination.
Configuration languages [53], integrated to a lazy functional language like
Haskell, allows a complete separation between parallelism and computational
programming dimensions. No extensions are required to Haskell for program-
18
ming at computational level. Haskell and the HCL are orthogonal. Eden and
Caliban, examples of embedded coordination languages, extend Haskell syntax with primitives for “gluing” processes to the coordination medium. GpH
tries to separate parallel coordination code by using evaluation strategies
[88]. Evaluation strategies is an interesting idea, but after inspecting some
GpH programs that uses them, we noticed that a complete and transparent
separation of the parallelism and the computation is very difficult to obtain.
This is even worse when programmers want to reach peak performance of
applications at any cost. The experience with Haskell# , and other parallel
functional languages, has shown that a really transparent separation makes
easier to parallelize existing Haskell programs. This increases opportunities
for the reuse of code and allows independent specification and development
of functional modules and coordination code, reducing programming efforts
and costs. The ability of composing programs from parts using the configuration approach also makes Haskell# more suitable for programming large
scale high-performance applications than other parallel functional languages
[33, 27]. Programmers are forced to adopt a coarse grained view of parallelism that is convenient for clusters and grids.
The Modelling of parallel architectures. Developing general techniques
for freeing programmers from making decisions on the allocation of processes
to processing nodes of a parallel architecture is an old challenge to the parallel programming community. However, this problem is hard to be treated
in general. Existing mechanisms for this purpose, either dynamic or static,
apply efficiently to restricted instances of the general problem and some of
them are based on heuristics. With the advent of grids, cluster of heterogeneous nodes, constellations, etc, it is not expected that a unified approach,
covering all realistic cases, may appear. Because of that, Haskell# follows a
static and explicit approach for process allocation, as in Caliban. Eden and
GpH, on the other hand, let allocation decisions to the compiler. In Haskell# ,
it is possible to model both processes needs for optimal execution and architecture characteristics by using partial topological skeletons for treating
allocation as an aspect. Each skeleton may be implemented using specific
allocation policies convenient for different architectures.
The analysis of formal properties using Petri nets. The support for
proving and analysing of formal properties of parallel programs by using
19
Petri nets is one of the most important premisses that guided the design of
Haskell#. A compiler that translates HCL configurations into INA [80], a
Petri net analysis tool, was developed [55]. In [23], a new translation schema
incorporating some extensions to the original HCL was presented. Recently,
a new translation schema has appeared and we are working on a new compiler
for translating Haskell# programs into PNML [91], a format supported by
many Petri net analysis tools, and SPNL [36], for analysing the performance
of Haskell# programs by using stochastic Petri nets. TimeNET [92] will be
used for this purpose. Other parallel functional languages do not support
formal analysis of parallel programs.
Simple and portable implementations . Unlike other parallel functional languages, it was not necessary to modify or extend the run-time system of GHC for implementing Haskell# . Indeed, any Haskell compiler could
be used in alternative to GHC, with all optimizations enabled. Haskell# programs take advantage of the evolution of compilation techniques with little
efforts. Eden, for example, modifies GHC compiler and disables some of its
optimizations [69]. Modifications to the run-time system of the Haskell compiler makes difficult to adapt the parallel language extension to new versions
of the compiler. In Haskell# , internal changes to the GHC run-time system
do not require modifications to the code generated by the Haskell# compiler.
Only if the interface of some used library is changed, minor modifications
are necessary. GpH and Eden developers should also carefully analyze the
effects of modifications to their parallel run-time system.
Efficiency. Potentially, Haskell# compiler may generate efficient MPI code
without using advanced compilation techniques for parallel code. This is due
to the direct correspondence of HCL constructors to MPI primitives and the
use of skeletons to abstract specific interaction patterns. Languages that
use higher-level constructors, in the sense that parallelism is transparent or
implicit, have difficulties on promoting the generation of MPI code able to
take advantage of peak performance in cluster architectures and, mainly, in
grid computing environments.
20
Toroidal Solution
Farm Solution
farm_a
farm_b
+
11111111
00000000
00000000
11111111
00000000
11111111
distributor
1111
0000
0000
1111
0000
1111
peer[n]
0000
1111
0000
1111
0000
1111
peer[1]
11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
0000
1111
0000
1111
0000
1111
0000
1111
farm_a
+
farm_b
peer[0]
11111111
00000000
00000000
11111111
00000000
11111111
00000000
11111111
collector
=
peer[0]
11111
00000
00000
11111
00000
11111
00000
11111
distributor
00000
11111
00000
11111
00000
11111
111111111
000000000
000000000
111111111
000000000
111111111
000000000
111111111
1111
0000
0000
1111
0000
1111
0000
1111
peer[1]
collector
...
1111
0000
0000
1111
0000
1111
collector
0000
1111
0000
1111
0000
1111
11111111
00000000
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
11111111
000000000
111111111
000000000
111111111
000000000
111111111
000000000
111111111
...
...
11111
00000
00000
11111
00000
11111
distributor
distributor
00000
11111
00000
11111
00000
11111
1111
0000
0000
1111
0000
1111
0000
1111
peer[0]
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
peer[1]
0000
1111
0000
1111
0000
1111
distributor
peer[n]
peer[n]
+
torus
Workpool Solution
wp_a 111111
000000
000000
111111
111111
000000
manager
000000
111111
000000
111111
000000
111111
worker[0]
worker[1]
worker[n]
=
11111111
00000000
0000
0000 1111
1111
0000 1111
0000
0000
1111
0000
1111
... 1111
0000
1111
0000
1111
0000
1111
0000
0000 1111
1111
0000 1111
worker[0]
worker[1]
cell[0,0]
cell[0,0]
...
cell[0,0]
cell[0,0]
cell[0,0]
...
cell[0,0]
...
cell[0,0]
cell[0,0]
cell[0,0]
...
+
00000000
wp_a11111111
00000000
+ 11111111
manager
00000000
wp_b11111111
...
0000
1111
0000
1111
0000
1111
0000
1111
1111
0000
0000
1111
0000
1111
worker[n]
worker[1] ... 1111
0000
0000
1111
0000
1111
0000
1111
0000
1111
0000
1111
manager
...
0000
1111
1111
0000
0000
1111
worker[0]
0000
1111
0000
1111
0000
1111
11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
00000000
11111111
00000000
11111111
... 11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
00000000 11111111
11111111
00000000 11111111
wp_b
worker[n]
Figure 10: HCL Topologies For Matrix Multiplication Solutions
3
Motivating Examples
This section presents Haskell# implementations for three applications recently used for benchmarking the parallel functional languages Eden, GpH
and PMLS: Matrix Multiplication, LinSolv and Ray Tracer [59]. A Haskell#
implementation for a sub-set of NPB (NAS Parallel Benchmarks) [2] is also
presented. These applications will be used in Section 5 for performance evaluation of the current Haskell# implementation, presented in Section 4.
3.1
Matrix Multiplication
Given two square matrices A, B ∈ Z n×n , n ∈ N, a matrix C ∈ Z n×n is
calculated, such that Ci,j =
n
X
Ai,k ∗ Bk,j .
k=1
A trivial, fine-grained, parallel solution requires n × n processors. Each
processor computes an element Ci,j , from scalar product of row i of A and
column j of B. This solution is obviously impractical, since large matrices are
common in real applications, requiring a number of processors not supplied
by contemporary parallel architectures. Three approaches are commonly
used in order to aggregate computations for increasing granularity [59]:
• Row Clustering: each process computes a set of rows of C. For that,
the process needs the corresponding set of rows of A and all matrix B;
21
{- In FILE: ManagerSkelBC WP.hcl -}
configuration ManagerSkel a → (b,c) where
{- In FILE: MatMultBC Farm.hcl -}
configuration MatMult<N> where
use ReadMatrix,WriteMatrix -- functional modules
iterator i range [1,N]
unit rA # () → (a::VMatrix); assign ReadMatrix to rA
unit rB # () → (b::VMatrix); assign ReadMatrix to rB
unit wC # c::Matrix → ()
; assign WriteMatrix to wC
{- In FILE: MatMultBC WP.hcl -}
configuration MatMult<N> where
unit farm a ;
assign Farm<N, splitM, combineM > to farm a
unit farm b ;
assign Farm<N, splitM T, combineM > to farm b
iterator i range [1,N]
use Skeletons.Workpool
use MatrixMult, ManagerSkel -- MatrixMult is a func.
import MatrixMult WF(splitM,combineM)
use Skeletons.Farm
use ReadMatrix, MatrixMult, WriteMatrix -- functional modules
import MatrixMult WFs(splitM,splitM T,combineM)
module
unit wa; assign Workpool<N> to wa
unit wb; assign Workpool<N> to wb
/[ unify farm a.worker[i] # a → c,
farm b.worker[i] # b → c
to worker[i] # (a::VMatrix, b::VMatrix ) → c::Matrix
assign MatrixMult to worker[i]
/]
unify wa.manager # c → a, wb.manager # c → b
to manager # c → (a, b)
[/ unify wa.worker[i] # a → c, wb.worker[i] # b → c
to worker[i] # (a::VMatrix, b::VMatrix ) → c::Matrix
assign MatrixMult to worker[i] /]
assign ManagerSkel to manager
unify farm a.collector # c → (),
farm b.collector # c → ()
to collector # c::Matrix → ()
assign WriteMatrix to collector
assign ReadMatrix N → a to farm a.distributor # () → a
assign ReadMatrix N → b to farm b.distributor # () → b
Figure 11: Haskell# Configuration of Block Clustering using Workpool
and Farm
• Block Clustering: each process computes a block of the resulting
matrix C. For that, the corresponding rows of A and columns of B are
needed;
• Gentleman’s algorithm: the processes are organized in a torus (circular mesh) for performing a systolic computation [78]. Each process
computes a block in C. At initial state, the corresponding blocks in
A and B are arranged across processes. Then, they execute k steps,
where k is the number of rows and columns of processes. At each step,
a process sends the blocks from A and B that it contains to its left
and down neighbors, and receive new blocks from right and top neighbors. A local computation is performed and the resulting matrix is
accumulated.
The above solutions differ on the number and size of messages exchanged.
In Haskell# programs, composition of skeletons may be used to describe
topologies for the solutions. Firstly, consider implementations of row and
block clustering using Workpool skeleton, where a manager process distributes rows or blocks, respectively, as jobs to a collection of worker processes, on demand of their availability. Once a worker finishes a job, it sends
22
{- In FILE: MatMultTorus.hcl -}
configuration MatMult<N> where
iterator i range [1,N*N]
use Skeletons.{Torus, Farm}
useReadMatrix, MatrixMult, WriteMatrix
import MatrixMult WFs(splitM,combineM)
unit farm a; assign Farm<N*N,splitM,combineM> to farm a
unit farm b; assign Farm<N*N,splitM,combineM> to farm b
unit torus; assign Torus<N> to torus
[/ unify farm a.worker[i] # a → c,
farm b.worker[i] # b → c,
torus.cell[i/N][i%N] # (as l,bs t) → (as r,bs d)
to cell[i/N][i%N] # (a::Matrix, b::Matrix, as l:: [Matrix ], bs t:: [Matrix ])
→ (c::Matrix, as r :: [Matrix ], bs d:: [Matrix ]) /]
unify farm a.collector # c → (), farm b.collector # c → () to collector # c::Matrix → ()
assign ReadMatrix N → a to farm a.distributor # () → a
assign ReadMatrix N → b to farm b.distributor # () → b
[/ assign MatrixMult to cell[i/N][i%N] /]
assign WriteMatrix to collector
Figure 12: Systolic Matrix Multiplication using a Torus (HCL Code)
its result back to the manager and stay available for receiving another job.
This technique is suitable when the number of jobs exceed the number of
processors available. Load balancing is automatically achieved in architectures where processor workload or performance may vary. Because of that,
it has been widely used in grid computations [34]. The unit manager in
the Workpool skeleton in Haskell# has two groups of ports of kind any:
one for sending jobs to workers and another for receiving results from them.
Workers receive jobs from their input ports and send results through their
output ports.
Row and block clustering may also be implemented using Farm skeleton.
Now, a master process sends a job to each slave process. Ideally, jobs have
similar workload. After completing a job, slaves send the result to their master and finish. The master combines the solutions received from all slaves.
This approach may reduce significantly the number of messages exchanged
and minimizes the communication overheads by using underlying collective
communication primitives. In fact, the Farm skeleton is defined by overlapping of Gather and Scatter skeletons. Farm employs wire functions for
distributing and combining values sent to and received from slave processes.
For achieving better load balancing, processors must be homogeneous. This
is a reasonable assumption to be made in cluster architectures, but not in
grid ones.
23
module LS homSol(main) where
module MatMult Toroidal(main) where
import
import
import
import
import MatrixTypes
import List
main :: Int -¿ Int -¿ Matrix -¿ Matrix -¿
[Matrix] -¿ [Matrix] -¿ (Matrix, [Matrix], [Matrix])
main = mult’
mult’ nc nr sm1 sm2 sm1s sm2s = (result, toRight, toDown)
where toRight = take (nc-1) (sm1:sm1s)
toDown = take (nr-1) (sm2’:sm2s)
sm2’ = transpose sm2
sms = zipWith multMatricesTr
(sm1:sm1s) (sm2’:sm2s)
result = foldl1’ addMatrices sms
addMatrices :: Matrix
addMatrices m1 m2 =
where addVectors
addVectors
-¿ Matrix -¿ Matrix
zipWith addVectors m1 m2
:: Vector -¿ Vector -¿ Vector
v1 v2 = zipWith (+) v1 v2
multMatricesTr :: Matrix -¿ Matrix -¿ Matrix
multMatricesTr m1 m2 =
[[prodEscalar row col | col ¡- m2] | row ¡- m1]
Matrix
LUDecompMatrix (det, homsolv)
qualified Matrix (determinant)
ModArithm
main :: (SqMatrix Integer, Vector Integer)
-¿ [Integer] -¿ [[Integer]]
main (a,b) = gen xList
where
gen xList :: [Integer] -¿ [[Integer]]
gen xList ps = map get homSol ps
get homSol :: Integer -¿ [Integer]
get homSol p =
let
b0 = vecHom p b
a0 = matHom p a
modDet = modHom p (determinant p a0)
pmx = homsolv0 p a0 b0
((iLo,jLo),(iHi,jHi)) = matBounds a
in
(p : modDet : if modDet == 0
then [0]
else pmx)
slow determinant :: SqMatrix Integer -¿ Integer
slow determinant = Matrix.determinant
foldl1’ :: (a-¿a-¿a) -¿ [a] -¿ a
foldl1’ f (x:xs) = foldl’ f x xs
determinant :: Integer -¿ SqMatrix Integer -¿ Integer
determinant = det
foldl’ :: (a -¿ b -¿ a) -¿ a -¿ [b] -¿ a
foldl’ f a [] = a
foldl’ f a (x:xs) = foldl’ f (f a x) xs
homsolv0 :: Integer -¿ SqMatrix Integer -¿
Vector Integer -¿ [Integer]
homsolv0 p a0 b0 = vecCont v
where
v = homsolv p a0 b0
prodEscalar :: Vector -¿ Vector -¿ MyInteger
prodEscalar v1 v2 = sum (zipWith (*) v1 v2)
Figure 13: Functional Modules of Matrix Multiplication and LinSolv
Figure 11 presents the Haskell# configuration codes for block clustering
using Workpool and Farm skeletons. Two matrices are distributed, thus
it is necessary to overlap two instances of both skeletons, as illustrated in Figure 10. The units readA, readB and writeC are clustered to implement the
manager process. The implementation of row clustering makes use of identical topological description. Differences are on port types and implementation
of computations. This evidences the importance of reuse and composition in
Haskell# programming.
The Gentleman’s algorithm is implemented by overlapping two instances
of the Farm skeleton, one for each input matrix, with a Torus skeleton,
as in Figure 10. The Torus describes the interaction pattern among slave
processes from the overlapped Farms. The HCL code for this arrangement
is presented in Figure 12.
Haskell# components that implement the solutions above have the same
names and interfaces. Only internal details, concerning the parallelism strat24
manager
guessedNoOfPrimes
input_ab
cra
n
primes
pBound
ab
output
add
xList
primes
...
homSol[1]
wp
ation
1111
0000
0000
1111
0000
1111
0000
1111
worker[0]
0000
1111
0000
1111
0000
1111
1111
0000
0000
1111
0000
1111
0000
1111
worker[1]
0000
1111
0000
1111
0000
1111
...
v_manager
11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
+
00000000
11111111 00000000
00000000
11111111 ... 11111111
11111111
00000000
00000000
11111111
00000000
11111111
00000000
00000000 11111111
11111111
00000000 11111111
peer[1]
peer[2]
1111111111
0000000000
0000000000
1111111111
0000000000
1111111111
manager
root
1111
0000
0000
1111
0000
1111
0000
1111
worker[N]
0000
1111
0000
1111
0000
1111
homSol[N]
bcast_ab
unific
111111
000000
000000
111111
000000
111111
manager
000000
111111
000000
111111
000000
111111
homSol[2]
peer[N]
=
0000
1111 0000
0000
1111 ... 1111
1111
0000
0000
1111
0000
1111
0000
0000 1111
1111
0000 1111
worker[0]
worker[1]
worker[N]
Figure 14: Haskell# Topology For LinSolv
egy adopted, varies. Thus, they can be used interchangeably in an application
by nesting composition. The Haskell# visual programming environment allows several component versions to co-exist. The programmer may choose
the appropriate version, depending on the target parallel architecture. For
instance, implementing matrix multiplication using Farm may be more efficient in clusters. In grids, a Workpool may prove more suitable. In
supercomputers where processors are organized in a torus, the toroidal solution may be the best choice.
3.2
LinSolv
Given a matrix A ∈ Z n×n and a vector b ∈ Z n , n ∈ N, find an exact solution
to the linear system of equations of the form Ax = b.
The solution described here is exact and operates over arbitrary precision
integers. A multiple homomorphic image approach is adopted [54], consisting
of three stages [59]:
1. map the input data into several homomorphic images. The domain of
homomorphic images is Z modulo p (Zp ), where p is a prime number;
25
{- In FILE: ManagerLS WP.hcl -}
configuration LS Manager<N> # (ab,xList) → primes where
use LS Input, LS Primes, LS CRA, LS Output
unit
unit
unit
unit
input #
primes #
cra
#
output #
connect
connect
connect
connect
connect
() → (ab,pBound,n)
;
(unlucky primes*,pBound) → (guessedNoOfPrimes,primes*) ;
(n, guessedNoOfPrimes, xList*) → (c,unlucky primes*)
;
c → ()
;
primes → guessedNoOfPrimes
input ab → n
input ab → pBound
cra → unlucky primes
cra → c
to
to
to
to
to
assign
assign
assign
assign
LS
LS
LS
LS
Input to
Primes to
CRA
to
Output to
input
primes
cra
output
cra ← guessedNoOfPrimes
cra ← n
primes ← pBound
primes ← unlucky primes
output ← c
{- In FILE: LinSolv WP.hcl -}
configuration LinSolv<N> where
iterator i range [0,N-1]
use Skeletons.{Collective.BCast, Workpool}
use LS Manager, LS HomSol
unit bcast ab ; assign BCast<N> to bcast ab
unit wp
; assign Workpool<N> to wp
interface ILinSolv (ab, job) → (ab,job) where: ab@IBCast # job@IWorkpool
behavior: seq {do ab; do job}
unify wp.manager, bcast ab.root to ls manager # ILinSolv
assign LS Manager to ls manager
[/ unify wp.worker[i], bcast ab.peer[i] to ls worker[i] # ILinSolv
assign HomSol to ls worker[i] /]
Figure 15: LinSolv using a Workpool skeleton (HCL Code)
2. compute the solution in each of these images, using LU-decomposition
followed by forward and backward substitution;
3. combine the results of all images into a result in the original domain,
using a fold-based CRA (Chinese Remainder Algorithm) [58].
The parallel strategy implemented in Haskell# is based on Eden and
GpH versions [60]. A manager process distributes computations of homomorphic solutions as jobs to a collection of worker processes. The skeleton
Workpool was adopted to distribute prime numbers to workers and to
collect computed homomorphic solutions. The BCast collective communication skeleton is used for distributing working data (A and b) to the workers.
The Haskell# configuration code that implements this arrangement is presented in Figure 15. A composed component LS Manager is configured
for aggregating computations of functional modules LS Input (obtains input data A and b), LS Primes (computes the list of primes for calculating
homomorphic solutions), LS CRA (aggregates homomorphic solutions using Chinese Remainder Algorithm), and LS Output (outputs result x). In
26
manager
rt_parameters
output
bcast_world
11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
raytracer[0]
root
00000000
11111111
11111111
00000000
11111111
00000000
11111111
00000000
00000000
11111111
00000000
11111111
00000000
00000000 11111111
11111111
00000000 11111111
peer[1]
peer[2]
peer[N−1]
...
raytracer[1]
+
root
+
00000000
11111111 00000000
00000000
11111111 11111111
11111111
00000000
00000000
11111111
00000000
11111111
00000000
00000000 11111111
11111111
00000000 11111111
peer[1]
peer[2]
peer[N−1]
raytracer[N]
gather_res
scatter_xy
11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
00000000000
11111111111
raytracer[2]
v_manager
11111111111
00000000000
00000000000
11111111111
00000000000
11111111111
00000000
11111111
00000000
11111111
00000000
00000000
00000000 11111111
11111111
00000000 ... 11111111
11111111
00000000
11111111
00000000
11111111
00000000
11111111
00000000
00000000 11111111
11111111
00000000 11111111
peer[1]
peer[2]
peer[N−1]
1111111111
0000000000
0000000000
1111111111
0000000000
1111111111
0000
1111
0000
1111
0000
0000
0000 1111
1111
0000 ... 1111
1111
0000
1111
0000
1111
0000
1111
0000
0000 1111
1111
0000 1111
root
root
=
peer[1]
peer[2]
peer[N−1]
Figure 16: Haskell# Topology Composition For Ray Tracer
composed component LinSolv, the main component, a cluster is created
by assigning LS Manager to unit ls manager, which is configured in such
a way that it makes the role of root unit in BCast skeleton and manager
of Workpool skeleton. The functional module LS HomSol implements
computation of a homomorphic solution for a given prime number. It is assigned to units ls worker [i], for 0 ≤ i ≤ N − 1, obtained by unification of
worker units of Workpool and peer units of BCast. Notice that these
skeletons are overlapped. The cluster ls manager might be placed onto a
multiprocessor node, in such a way that processes input, primes, cra and
output could execute concurrently. Figure 14 illustrates topological specification of LinSolv. Figure 13 shows examples of functional modules of Matrix
Multiplication and LinSolv.
3.3
Ray Tracer
Given a collections of objects in the three dimensional space, calculate the
corresponding two dimensional image. All rays in a window (for each pixel in
the grid) are traced and their intersections with objects are computed. The
colour of an intersection point is computed based on the strength of the ray
27
{- In FILE: RT Manager.hcl -}
configuration RT Manager<N> # (rt raytracer←xy, world) → (rt parameters→xy[2], world) where
use RT Parameters, RT Result -- functional modules
unit rt parameters # () → (xy<2>, world) groups xy:broadcast assign RT Parameters to rt paramters
# (xy,res) → (); assign RT Result to rt output
unit rt output
unit rt raytracer
# (xy,world) → res
connect rt parameters → xy[1] to rt output ← xy
{- In FILE: RayTracer.hcl -}
configuration RayTracer<N> where
iterator i range [0,N-1]
use BCast, Scatter, Gather from Skeletons.Colletive
use RT RayTracer -- functional module
unit bcast world; assign BCast<N> to bcast world
unit scatter xy ; assign Scatter<N> to scatter xy
unit gather res ; assign Gather<N> to gather res
interface IRayTracer (xy.*, world.*, res.*) → (xy.*, world.*, res.*)
where: (xy@IBCast # world@IScatter # res@IGather )
behavior: seq {do world; do xy; do res}
[/ unify bcast ab.peer[i], scatter world.peer[i], gather res.peer[i] to rt worker[i] # IRayTracer
assign RT RayTracer to rt worker[i]
/]
unify bcast world.root, scatter xy.root, gather res.root to manager # IRayTracer
assign RT Manager to manager
unify manager.rt raytracer, rt worker[0]
Figure 17: Ray Tracer (HCL Code)
and texture of the object reached [59].
A data parallel solution is trivial, since rays can be traced independently
for each pixel. In Haskell# implementation, a direct mapping of the image
lines to N parallel processes, assuming one at each processor, is employed.
Each process receives the same number of lines to compute. This solution
yields load balancing in homogeneous clusters. The HCL for ray tracer is
presented in Figure 17 and its topology is described in Figure 16. It is implemented by overlapping three skeletons: BCast, Gatherv and Scatter.
The root units of these skeletons are unified to form the manager unit, responsible for distributing and collecting work among worker units, obtained
by overlapping their peer units. The manager also acts as a worker. Distribution and collection are specified by wire functions. The BCast skeleton
disseminates the world scene to workers. Scatter and Gatherv are used
to distribute jobs and collect the results from the workers.
28
3.4
NAS Parallel Benchmarks
This section presents the Haskell# implementations for a sub-set of NPB
(NAS Parallel Bechmarks) [2], a package comprising eight programs, specified
in NASA Research Center at Ames, USA, intended to benchmark the performance of parallel computing architectures for execution of the NAS (Numerical Aerodynamic Simulation) programs. NPB programs implemented in
Haskell# are:
• EP (Embarrassingly Parallel ) generates pairs of Gaussian deviates according to a specified scheme and tabulates the number of pairs in successive square anulli. It was developed to estimate the upper achievable
limit for floating point performance in a parallel architecture;
• IS (Integer Sorting) performs parallel sorting of N keys using bucket
sort algorithm. Keys are generated using a sequential algorithm described in [3] and must be uniformly distributed;
• CG (Conjugate Gradient) implements a solution to an unstructured
sparse linear system, based on conjugate gradient method. The inverse
power method is used to find an estimate of the largest eigenvalue of
a symmetric positive definite sparse matrix with a random pattern of
non zeros;
• LU (LU factorization) uses symmetric successive over-relaxation (SSOR)
procedure to solve a block lower triangular-block upper triangular system of equations resulting from an unfactored implicit finite-difference
discretization of the Navier-Stokes equations in three dimensions;
NPB programs exercise the expressiveness of HCL for describing SPMD
programs and for translating MPI programs into Haskell# . LU gave us an
important insight on how to facilitate programming of applications where
processes have a large number of input and output ports. CG and IS help
on evaluating the performance of collective communication skeletons.
3.4.1
The Embarassingly Parallel (EP) Kernel
The HCL code of EP is presented in Appendix C.1. It declares n units,
named ep unit[i], for 1 ≤ i ≤ n. The interface class that describes the behavior of these units, IEP, is formed by the composition of three instances
29
All Reduce
sy
All Reduce
sx
sy
sx
sx
...
EP
EP
sy
EP
ep[1]
ep[0]
ep[N−1]
q
q
q
All Reduce
Figure 18: EP Topology
of IAllReduce interface class, called sx, sy and q. The definition of channels
is specified by overlapping three instances of the AllReduce skeleton. For
that, clusters sx comm, sy comm, and q comm are associated with AllReduce component and their virtual units are unified. The HCL compiler uses
the topological information provided by AllReduce skeleton and generates
code that uses the MPI AllReduce primitive of MPI.
3.4.2
The Integer Sort (IS) Kernel
The HCL code of IS is shown in Appendix C.2. It declares a network of n
units, named is unit[i], for 1 ≤ i ≤ n. The interface class for describing
the behavior of IS units, called IIS, is a composition of interfaces IAllRe-
All Reduce
bst
0
bs
bst
IS
kb1
is[1]
kb2
bs
bst
...
IS
IS
is[0]
kb2
bst
bs
IS
is[N−1]
is[2]
kb2
kb1
kb1
All To All v
Figure 19: IS Topology
30
bs
kb2
kb1
u[1][2]
u[1][3]
u[1][4]
u[2][2]
u[2][3]
u[2][4]
broadcast
u[1][1]
broadcast
broadcast
u[2][2]
Factorize
u[2][1]
broadcast
u[1][2]
sum_arrays
sum_arrays
u[2][1]
sum_arrays
sum_arrays
u[1][1]
Figure 20: Transpose Skeleton Topology
duce, IAllToAllv and IRShif. A cyclic pattern of communication (repeat
combinator) now appears, due to presence of stream ports on specification
of IIS.
IS network topology is defined by overlapping skeletons AllReduce and
AllToAllv, for collective communication, and RShift, which performs a
data shift right amongst processes. Cluster units bs comm, kb comm and
k shift are assigned to them, respectively, and their virtual units are unified.
The interface components bs, kb, rshif t of IIS indicate which ports of IS
units participate in the skeleton instances, respectively.
3.4.3
The Conjugate Gradient (CG Kernel)
The original topology of CG, specified in FORTRAN/MPI, imposes that the
number of processes, organized in a rectangular mesh, is a power of two.
The version of CG in Haskell# is less restrictive. The programmer must
provide parameters dim (the number of mesh rows), and col f actor (the
number of mesh columns is obtained by multiplying it to dim). Any number
of units may be configured using this approach, but different configurations
may result in different performance. The programmer should adequate the
parameters values to the features of the execution environment. CG units
cg unit[i][j], for 1 ≤ i ≤ dim and 1 ≤ j ≤ dim ∗ col f actor. The HCL code
of CG is presented in Appendix C.3.
The interface class that describes the behavior of CG units, ICG, is a
composition of interface classes IAllReduce (rho, aux, rnorm, norm temp 1
and norm temp 2) and ITranspose (q and r). CG topology is defined by
overlapping AllReduce and Transpose skeletons. The former is used for
data exchange during parallel scalar products at mesh rows, and the latter
for data exchange in parallel matrix multiplications, whenever a transpose
operation is performed on data stored in processors. In MPI original code,
31
several calls to MPI Irecv primitive are needed to perform these operations,
making difficult to understand the structure of the topology without a careful
analysis of the parameters of the problem.
Five clusters are needed for each row of processes: rho comm[i], aux comm[i],
rnorm comm[i], norm temp 1[i], and norm temp 2[i], 1 ≤ i ≤ rows. The
AllReduce component is assigned to them. The Transpose component
is assigned to the other two clusters, q comm and r comm, encompassing all
processes in the network. Their units are unified producing the final Haskell#
topology of CG.
The Haskell# configuration code of Transpose is presented in Appendix
C.3.1. It organizes virtual units according to parameters dim and col f actor,
supplied by CG configuration. Firstly, a square mesh of units with dimension
dim is assembled. The ports are connected to transpose data amongst processors using appropriate wire functions applied on groups of ports. These
units are factorized in col f actor units, resulting in a square mesh with dim
rows and dim ∗ col f actor columns. The diagram in Figure 20 illustrates the
factorization process involved in Transpose specification. In order to make
it easier to understand, only channels connected to u[1][1] ports are shown.
They are replicated according to factorization rules.
3.4.4
The LU Factorization (LU Simulated Application)
The HCL code of LU is presented in Appendix C.4. LU organizes n process, where n is a power of two, in a grid. It employs the wavefront method
[6] in parallel computation. It differs from other NPB programs because
communication is performed by small messages of approximately 40 bytes.
Another particularity of LU is the great number of communication ports in
units (thirty input ports and thirty output ports). Skeletons Exchange 1b,
Exchange 3b, Exchange 4, Exchange 5, and Exchange 6 describes
communication topologies in several communication phases during execution,
using the wavefront method. The same nomenclature employed in the original LU versions are used here to make easier to compare the two approaches.
In these skeletons, there are several interfaces for virtual units that comprise
them. Their specification vary according to their position in the grid. Interface generalization is useful in such cases, avoiding classes of units to be
treated individually in the configuration.
32
# Compiler
#
s
on
t
ele
High−level
Topological
Information
sk
#
#
FRONT−END
flatten
lea
Flat
Representation
BACK−END
ve
s
λ
λ
Functional
Modules
# PROGRAM
Executable
Process
Wrapper
Modules
GHC
FFI
MPI
...
Processes onto
processors
mapping
Process
Allocator
11
00
Switching Hub
Figure 21: Simple Components in Haskell
4
Implementation
Haskell# may be implemented on top of a message passing library and a
sequential Haskell compiler, without any modifications or extensions to any
of them. MPI 1.1 and GHC (Glasgow Haskell Compiler) are currently used,
respectively. MPI is now considered the most efficient message passing library for clusters, providing standard bindings for C and Fortran. Recently,
MPI versions for grid computing have appeared [49]. GHC is now considered
state-of-the-art techniques for the compilation of lazy functional programs.
It supports FFI (Foreign Function Interface) [24] to make direct calls to MPI
routines from Haskell programs. The use of an efficient sequential Haskell
compiler has important impact on performance of Haskell# programs, since
Haskell# programs assumes medium and coarse grained parallelism, where
most of time is spent in sequential mode of execution. Haskell# implementations are easily portable to new MPI and GHC versions. Indeed, it is
possible to replace GHC with any Haskell compiler that supports FFI. All optimizations and extensions provided by the Haskell compiler may be enabled.
This is an important feature of Haskell# , since other parallel functional languages built on top of GHC need to modify its run-time system. The current
Haskell# implementation has already been tested on top of LAM-MPI 6.5.9
[17], MPICH 1.2.5.2 [37] and GHC versions 6.01 and 6.2 in clusters equipped
with RedHat Linux 8.0 and 9.0.
33
4.1
An Overview of the Haskell# Compilation Process
The Haskell# compiler has been entirely programmed in Haskell, using Alex
2.0 [30] and Happy 1.13 [62] for parsing. It is divided into two modules:
front-end and back-end. The compilation process is illustrated in Figure
21. The front-end module parses all components of a Haskell# program,
by traversing its tree of components, from application component to simple
components. A flat representation of the processes network is generated.
Relevant topological information, obtained from the use of skeletons, that
could guide the back-end for the generation of optimized code is stored. The
flat code is currently represented as an algebraic data type in Haskell, but it
is intended to implement it in XML (Extended Markup Language), allowing
to use it as an intermediate language for interfacing tools for the analysis
performance and formal properties in the programming environment under
development.
The back-end uses flat code and topological information for generating a
wrapper module for each process and for inferring the mapping of processes
onto processors of the target architecture. A wrapper module is a Haskell
program that controls the execution of a process. The wrapper modules and
the functional modules are compiled using GHC. The mapper is a program
that copies executable files onto the target machine where it will execute,
based on the mapping of processes onto processors inferred by the back-end.
4.2
Wrapper Modules
In Figure 22, the structure of a wrapper module is illustrated. A wrapper makes a call to the main function of the functional module associated
to with the process. The values produced at return points (ri , 1 ≤ i ≤
k) are copied concurrently to channel variables 2 (chan rj ), using functions
send stream and send atom, depending on the nature of the associated output port. The arguments provided to the main function (aj , 1 ≤ j ≤ n)
may also be obtained from channel variables (ON DEMAND chan rj ) or directly (FORCED actionj ), on demand of evaluation of return points. The
function perform actions controls the completion of the communication operations, according to a guide automaton that recognizes the behavior specified
in the process interface. Whenever an output port must be activated, perform actions evaluates perform communication, which reads a value from the
2
Type Chan t from Concurrent Haskell [74].
34
module Main(main) where
import
import
import
import
System(getArgs)
Concurrent(forkIO, Chan, newChan, newQSemN, waitQSemN, signalQSemN)
HHashSupport
qualified ¡Functional Module¿(main)
¡import declarations that appear in # code¿
main :: IO ()
main = do
argv ← getArgs
argc ← (return.length) argv -- MPI initialization
give args [] argv (mpi init BUFFER SIZE argc)
a1 chan :: Chan (Comm ¡Channel Type¿) ← newChan --Initializing channel variables for arguments
a2 chan :: Chan (Comm ¡Channel Type¿) ← newChan
...
an chan :: Chan (Comm ¡Channel Type¿) ← newChan
r1 chan :: Chan (Comm ¡Channel Type¿) ← newChan --Initializing channel variables for return points
r2 chan :: Chan (Comm ¡Channel Type¿) ← newChan
...
rk chan :: Chan (Comm ¡Channel Type¿) ← newChan
for each p, an individual port or group of ports involved in a collective operation:
p ← [mpi register port . . . | mpi register peer . . .]
let comm p = [SingleIPort | SingleOPort | GroupIPort | GroupOPort | Bcast | Gather | Scatter |
Scatterv | Allgather | Allgatherv | Allreduce | Alltoall | Alltoallv | Reduce Scatter | Scan] p · · ·
caut ← ¡code to setup guide automata¿
control automata init caut
sync ← newQSemN 0
forkIO (perform actions ¿¿ signalQSemN sem)
let a1 = [recv stream | recv atom] [ON DEMAND action1 | FORCED chan a1 ]
a2 = [recv stream | recv atom] [ON DEMAND action2 | FORCED chan a2 ]
···
an = [recv stream | recv atom] [ON DEMAND actionn | FORCED chan an ]
(r1 ,r2 ,. . .,rk ) = ¡Functional Module¿.main a1 a2 . . . an
forkIO ([send stream | send atom] r1 chan r1 ¿¿ signalQSemN sync)
forkIO ([send stream | send atom] r2 chan r2 ¿¿ signalQSemN sync)
...
forkIO ([send stream | send atom] rk chan rk ¿¿ signalQSemN sync)
waitQSemN (k+1) sem
mpi finalize
Figure 22: Wrapper Module
35
perform_actions !
calls
calls
perform_communication!
writeChan
recv_stream!
ON_DEMAND
FORCED
calls
(guide automaton)
readChan
λ
channel variable
perform_communication!
send_stream!
writeChan
readChan
channel variable
Figure 23: Schematic Representation of a Wrapper Module
corresponding return point and sends it through the active port. For input
ports, perform communication may be called inside recv stream or recv atom
functions, when an argument value is demanded. In this case, the operation
is validated by the guide automaton and a channel variable is not necessary.
However, in some collective communication operations, when a process sends
and then receives a value (the root process in a broadcast, for example), it is
needed to write and read, in a single call to perform communication, channel
variables associated to a return point and to an argument, respectively. This
is a situation where a channel variable is necessary for an argument. The
Haskell# compiler forces evaluation of the input ports inside perform actions
whenever it may infer that an input port must be strictly activated before
the activation of some output port. This is typical when the alt (choice)
constructor does not occur in process behavior specification. Figure 23 illustrates the use of channel variables.
Since processes spend some time with synchronization, concurrent evaluation of perform action and exit points, using send stream and send atom,
allow the overlapping of computations when a process is executing perform communication. In multiprocessors and super scalar processors, which
may execute instructions in parallel and speculate about their execution,
performance might be improved.
4.3
Guide automaton: Controlling Activation of Ports
A guide automaton is an abstract data type, implemented in C, used for controlling and validating the activation order of ports in execution of Haskell#
programs. It might be algebraically described by a tuple of the following
form:
C = (Π, Q, T, ϕ0 , ϕ1 , ρ, F, S, σ, π, γ, κ)
36
!
q 1r
!
q nr
π(q)=Choice
... .
..
...
γ (q)=True
...
!
q ml
Ql
π(q)=Join
kIO
for
γ (q)=False
transition
...
...
q 1l
π(q)=Fork
q
path
!
!
!
q 1l
Ql
q ml
q
current
state
forkIO
q
kIO
for
...
...
...
current
state
!
!
...
q
...
!
forward
states
...
π(q)=Forward
Qr
Figure 24: Guide Automaton
where:
• Π is a set of port identifiers that forms the alphabet of the guide automata;
• Q is a finite set of states;
• T is a finite set of transitions;
• ϕ0 : T → Q maps each transition to its origin state;
• ϕ1 : T → Q maps each transition to its target state;
• ρ : T → Π labels each transition with a port identifier;
• F ⊆ Q is a set of final states;
• S is a finite set of symbols, representing semaphores;
• σ : Q → 2S×N at associates states to semaphore updates. For instance,
consider a semaphore s ∈ S. If (s, n) ∈ σ(q) then the value of s must
be incremented by n when entering state q;
• π : Q → {f orward, choice, f ork, join} gives the kinds of the states;
• γ : Q → {T rue, F alse} maps choice states to an expression (termination condition of a repeat combinator) that evaluates to True or
False;
• κ : Q → 2Q × 2Q associates a state q, to a pair of set of states (Ql , Qr ),
whose meaning depends on π(q) (see the next paragraph).
37
States and transitions are represented as natural numbers. The initial
state is 0 (zero). Let q be the current state of a guide automaton. The function perform actions looks up κ(q) in order to choose the next communication
operation to be performed. For instance, consider κ(q) = (Ql , Qr ). There
must be a path from state q to each state in Ql ∪ Qr . If π(q) = f orward,
Qr = ∅ and Ql determines the forward states of q. Among them, the goal
states are chosen. For that, let us consider a set of transitions T r = {t |
ϕ1 (t) = q ′ ∧ q ′ ∈ Ql ∧ t is in a path from q to q ′ }. Port p is chosen from ports
{p | t ∈ T r ∧ ρ(t) = p}, among those whose communication pairs are active
at that instant (ready for communication). Forward states q, such that, for
some t ∈ T r , ϕ1 (t) = q and ρ(t) = p, are goal states. Choices appear only
in the implementation of occurrences of the alt constructor. The port p is
activated. If p is an output port (default case), it may cause the implicit activation of input ports, in recv stream or recv atom function calls, before completing communication. After any port activation in perform communication,
the advance automata function is called for updating the current automata
state, validating the operation, by raising an error whenever there is no transition from the current state labelled with the activated port, and updated
semaphores. After the activation of p, the guide automaton must be in one
of the goal states. Otherwise, the operation is invalid. If π(q) = choice, γ(q)
must be evaluated (termination condition of a repetition). If γ(q) is true,
the set of forward states of q is Ql , otherwise it is Qr . Choice states are
used in the implementation of occurrences of repeat and if combinators. If
π(q) = f ork, Qr = ∅ and ∀t : ϕ0 (t) = q : ϕ1 (t) ∈ Ql ∧ ρ(t) = ⊥. When a
fork state is reached, threads are forked for executing communication actions
starting from the states in Ql . All threads must reach the same join state,
where they finalize and resume execution from that state. If π(q) = join,
Ql = ∅ and Qr = ∅. Fork and join states are used to implement occurrences
of par combinator. If there is no forward state from current state and it is
a final state, perform actions finalizes.
Semaphores are updated in calls to advance automata. The function σ is
used to update their values according to the new current state. A semaphore
must have more than one value at a time. During execution, it must be guaranteed that all semaphores must be at least one positive value. Otherwise,
an error is informed. Negative values are discarded. Semaphores only exist
for validating non-regular patterns of communication that may be described
by labelled Petri nets [89]. However, in general, regular patterns of communication are sufficient to describe behavior of most of high-performance
38
Table 1: Meaning of parameters of mpi register pair and mpi register peers
Parameter
Direction
Source/Target rank
Channel tag
Collective Op. Type
Number of Processes
Processes in group
Buffer Size
Data Type
Reduce Operation
Is Probed Flag
Pair is Probed Flag
pair
⋆
⋆
⋆
⋆
⋆
⋆
peer
⋆
⋆
⋆
⋆
⋆
⋆
Description
Specifies if a port is for input or output
Rank of the process that owns its comm. pair
A number that identifies individually a channel
Kind of the collective communication operation
Number of processes in the collective operation
Ranks of processes in the collective operation
Buffer used for storing data to be transmitted
MPI data type (used in a reduce operations)
MPI operation (used in a reduce operations)
Flag indicating if a port belongs to a choice group
Flag indicating if the communication pair of a
port belongs to a choice group.
parallel programs [65, 70]. Thus, overhead due to semaphore updating might
be avoided for parallel programs where peak performance is critical.
4.4
Implementing Communication Operations
There are two kinds of communication operations in Haskell# : point-to-point
and collective. The former is implemented through simultaneous activation
of channel’s communication pairs. MPI tags, in message envelopes, represent communication channels in calls to point-to-point primitives. The later
is implemented using MPI support for dynamic configuration of communication groups and contexts and MPI collective communication primitives.
Groups of ports involved in a collective communication are called communication peers. Each communication pair is configured using the function
mpi register pair, while communication peers are configured in a single call
to mpi register peers. These functions are implemented in C, being called
from Haskell code through FFI. Their arguments, detailed in Table 1, set up
parameters for completion of communication operations over involved ports
during execution. A communication handle, an integer number, is returned
and bound to a variable for allowing to access pair/peers information whenever necessary.
The polymorphic and higher-order function perform communication has
one argument, a value from the algebraic data type PortInfo t u v, whose
constructors identifies the kind of communication operation to be performed:
SingleIPort, SingleOPort, GroupIPort, GroupOPort (point-to-point
39
communication), Bcast, Gather, Scatter, Scatterv, Allgather, Allgatherv, Allreduce, Alltoall, Alltoallv, Reduce Scatter, Scan (collective
communication. The PortInfo’s fields encapsulate necessary information for
completion of communication operations: communication handle, port type
(choice or combine), wire functions, and channel variables. The type variables t, u and v are used for generalization of channel variables and wire
functions types.
The MPI point-to-point communication primitive used for completion
of communication over an output individual port (SingleOPort) depends
on the communication mode of the channel where it is linked: buffered
(MPI Bsend), synchronous (MPI Ssend) or ready (MPI Rsend). For groups
of output ports of kind All, the corresponding asynchronous MPI sending
primitives (MPI Ibsend, MPI Issend and MPI Irsend) are used for initiating
the communication on each port belonging to the group. Then, a call to
MPI Waitall waits for the completion of all the returned request. Similarly, a
call to MPI Recv implements the communication on individual input ports,
while MPI Irecv (asynchronous) and MPI Waitall, implements groups of
input ports of kind All. Groups of ports of kind Any are implemented
using the channel probing protocol, which allows the verification of the status
of activation of communication pairs.
Transmitting streams and atom values. In Haskell# , a value of type t
is transmitted as a value of algebraic type Comm t, whose Haskell representation is depicted below:
data Comm t = Atom {data :: t} | Mid {data :: t} | End {depth::Int}
The Atom constructor encapsulates atomically transmitted values, while
streamed ones are encapsulated using Mid and End constructors. The
integer value in the End field represents the depth of a finalized stream.
For instance, consider a stream port p of type (Int,Int) and nesting factor 2 (p**::(Int,Int)). The lazy list associated to the port must be of type
[[(Int,Int)]]. Consider the lazy list [[[(1,2),(3,4)], [(5,6)]], [[(7,8),(9,0),(1,2)]],
[[], [(3,4)], [(5,6),(7,8)]]]. The list of values effectively transmitted through
the stream port p at each activation is [Mid (1,2),Mid (3,4), End 3, Mid
(5,6), End 3, End 2, Mid (7,8), Mid (9,0),Mid (1,2), End 3, End 2, End
3, Mid (3,4), End 3, Mid (5,6), Mid (7,8), End 3, End 2, End 1]. Whenever possible, stream communication is implemented using MPI persistent
40
communication objects, for minimizing communication overhead.
Marshalling Haskell Values to C Buffers. In order to transmit Haskell
values using MPI primitives, they must be marshalled onto C contiguous
buffers. For that, the Storable class, from FFI, is employed. Default Storable
instances are provided for basic data types. User defined data types should
be instantiated for this class. The Haskell# compiler traverses Haskell modules of the Haskell# program for finding user defined type values that must
be instantiated for the Storable class. Structured data types, such as lists,
arrays, tuples and algebraic data types must be packed and unpacked element by element. This could result in a considerable source of inefficiency
when number of elements is very large. The benchmarks presented in Section 5.1 evidence this fact. GHC provides unboxed arrays, whose values are
stored in contiguous memory areas and can be directly marshalled to MPI
buffers. Since most high performance computing applications operate over
arrays, and not using lists, unboxed arrays may be used in order to avoid
this source of inefficiency.
5
Performance Evaluation
This section presents some performance figures for Haskell# programs presented in Section 3. The architecture used is a Beowulf cluster comprising 16
dual Intel Xeon processors (clock: 2 GHz, RAM: 1GB), connected through
a Fast Ethernet (100MBs). Measures with 32 nodes were performed in dual
multiprocessing mode. MPICH 1.2.3 on top of TCP/IP was used for communication between processes.
5.1
Benchmarking Haskell# with NPB
The benchmark results of Haskell# versions of NPB kernels (EP, IS and CG)
are presented in Figure 25. The plots to left hand side present their respective
running times, while the plots at right hand side presents their corresponding
absolute speedups, comparing them to linear speedup, always represented
by a solid line.
Two problem instances were used for measuring performance of Haskell#
kernel versions (Table 2). In the second one, processes demand about twice
41
EP Running Time
EP Speedup
4722
15.8
EP-1
EP-2
15.8
Linear
EP-1
EP-2
Time
2373
589
296
150
75
37
595
298
1 2
4
8
Number of Processors
Speedup
EP
Running
EP
1189
7.8
7.9
3.9
4
2
1 2
16
4
8
Number of Processors
IS Running Time
16
2
1
IS Speedup
800
Linear
IS-1
IS-2
IS-1
IS-2
Time
448
445
9.5
8.5
208
101
64
36
112
83
1 2
4
8
Number of Processors
Speedup
IS
Running
IS
201
6.2
5.3
3.4
3.3
1.7
16
1.6
1 2
4
8
Number of Processors
16
CG Speedup
CG Running Time
4428
Linear
CG-1
CG-2
CG-1
CG-2
Time
2881
Running
CG
1427
1290
901
583
798
6
5.2
2161
2
4
Number of Processors
8
Speedup
CG
1
3.7
3.4
2.1
2.2
1.1
1.1
1
2
4
Number of Processors
Figure 25: Performance Figures of NPB kernels in Haskell#
42
8
Table 2: Instances of Problem Sizes Used to Run Each Kernel
Kernel 1st Problem Size
2nd Problem Size
EP
m = 25
m = 28
IS
total keys log2 = 20 total keys log2 = 21
max key log2 =16
max key log2 =17
CG
na = 14000
nonzer = 11
niter = 45
na = 18000
nonzer = 12
niter = 45
as much memory space as the first one, without exhausting physical memory resources of a single node of the cluster. The default problem classes
of NPB (S,W,A,B,C) were not used because they were tuned for use with
C/FORTRAN + MPI original versions. Due to laziness and the use of immutable arrays, sequential performance of Haskell# versions are about an
order of magnitude worse than the performance of the original versions of
NPB kernels, both considering time and space. Because of that, some default
problem sizes exhaust physical memory resources of cluster nodes, causing
virtual memory overheads that must be avoided in measures. The use of
mutable arrays could minimize this source of inefficiency, but they require
the encapsulation of computations inside the IO monad, preventing arrays
of being transmitted through lazy lists.
Also due to performance differences in sequential mode of execution, granularity of Haskell# processes is coarser than the granularity of processes
in original NPB versions. While Haskell computations execute slower than
C/FORTRAN computations, the amount of data transmitted is about the
same. The original speedup measures of NPB kernels serve only to establish the lower bounds of the performance of the cluster. One should not use
that to make assumptions and claims about relative efficiency of Haskell#
implementation.
Using GHC profiling tools [81], five main cost centres were identified in
CG and IS Haskell# implementations. Table 3 presents the impact of each of
them in parallel execution. The impact of cost centres in speedup is evaluated
on Table 4. By analyzing the data obtained, one may be conclude that:
1. If only time spent in computation is considered, the speedup is linear;
43
Table 3: Cost Centre Analysis of IS and CG (% of total execution time)
i
ii
iii
iv
v
i
ii
iii
iv
v
SEQ
45,9
54,1
90,2
9,8
2
35,4 3,0 7,4 4,6 45,3
79,1 1,5 2,1 4,7
7,7
4
IS-1 37,6 3,0 7,2 11,0 35,7 CG-1 70,9 1,8 3,2 11,6 5,8
8
36,0 2,7 7,2 20,8 28,7
57,5 3,5 5,6 24,0 2,7
16
34,0 2,5 7,1 27,8 24,4
50,5 4,1 7,3 32,7 1,9
SEQ
34,5
65,5
84,5
15,5
2
38,6 2,8 6,7 5,4 49,3
68,5 1,2 1,7 10,8 11,6
4
IS-2 35,3 2,8 6,8 11,7 38,9 CG-2 70,2 1,5 2,5 12,9 6,3
8
32,8 2,7 7,0 21,1 32,6
61,1 3,2 5,1 19,5 4,5
16
30,1 2,7 7,2 27,8 28,7
58,5 3,5 5,6 25,0 2,7
i: Raw computation time, ii: Evaluation of wire functions, iii: Marshalling,
iv: Communication and synchronization, v: Garbage Collection
2. The marshalling cost centre is the unique source of overhead inherent to
Haskell# implementation. The other ones are inherent to parallelism.
In some cases, marshalling overhead increases with the number of processors (CG-1 and CG-2). Marshalling could be avoided if GHC allows
to copy immutable arrays to contiguous buffers in constant time. But
this feature could not be provided yet;
3. The garbage collection overhead decreases by increasing the number
of processors used in parallel computation. This fact is attributed to
less use of heap when the problem size is split among more processors
and the enforcement of data locality. Cache behavior effects are also
being investigated. It is worthwhile to remember that garbage collector
parameters were tuned before execution. The results obtained here do
not guarantee that every Haskell program presents the same behavior;
4. In CG, whenever number of processors increases, the gains in performance due to the minimization of the garbage collection overhead appears to compensate losses due to the marshalling overhead. Thus, in
some cases, Haskell# overhead may be considered null. Indeed, assuming that arrays are copied directly and in constant time, the minimization of the garbage collection overhead could compensate their sources
of overhead that are inherent to parallelization;
44
Table 4: Influence of Cost Centres in Speedup
2
4
8
16
2
4
8
16
IS-1
IS-2
a
2,1
4,1
7,5
15,1
1,9
4,1
8,0
16,1
b
1,9
3,8
7,0
13,8
1,8
3,8
7,3
14,7
c
1,6
3,2
5,9
11,8
1,5
3,2
6,1
12,0
d
1,5
2,6
4,0
8,3
2,5
4,1
8,0
e
1,2
2,5
4,4
8,5
2,5
4,5
8,2
CG-1
CG-2
a
2,0
3,9
7,9
15,9
2,1
4,0
8,0
16,0
b
2,0
3,8
7,4
13,4
2,0
3,9
7,6
14,4
c
1,9
3,6
6,7
12,8
2,0
3,7
7,0
13,4
d
1,8
3,2
4,9
10,5
1,7
3,2
5,5
10,9
e
1,9
3,2
5,3
10,9
1,8
3,4
6,0
11,2
a: i, b: i/ii, c: i/ii/iii, d: i/ii/iii/iv, e: i/ii/iii/iv/v
The observations above are evidences that Haskell# programs are an efficient approach for parallelizing functional computations. The fact observed
that splitting of problems among processors may reduce the garbage collection overheads is another motivation for using Haskell# for parallelizing
scientific high-performance applications written in Haskell, in addition to the
gains in execution time of computations, since this kind of application normally processes large data structures stored in memory. The benchmarks
presented in the next section compare Haskell# to other parallel functional
languages.
5.2
Benchmarking Haskell# with Loild’s Benchmark
Suite
The benchmarking results of Haskell# implementations of Matrix Multiplication (MM), LinSolv (LS), and Ray Tracer (RT), based on Eden and GpH
versions presented in [59], are shown in Figure 26. The parameters are described on Table 5. Since the cluster used has nodes about three times as fast
as than nodes of the cluster used in Loild’s measures, the size of the problem
instance of MM and RT used in this paper are larger. This attempts to
approximate the sequential run-time of original measures and the increase of
granularity of computations. For LS, however, the same problem size is used
since its scalability is less sensitive to variations in problem size.
The speedup curves of LS and RT are nearly linear, while the speedup
45
MM Speedup
MM Run Time
265.3
Linear
Row
Block
Torus
Linear
Row
Block
Torus
220.3
11
137
6.9
73.3
64
48.7
45.7
5.8
5.4
4.1
3.6
63
32
20
12 4
8
16
Null
3.5
1.9
32
12 4
8
LS Running Time
16
Null
32
1
LS Speedup
1031
24.5
521
14.3
263
7.6
136
72
42
3.9
2
12
4
8
16
32
12
4
8
RT Running Time
16
32
RT Speedup
400
30.8
221
15.6
100
8
50
26
13
4
1.8
12
4
8
16
32
12
4
8
16
Figure 26: Performance Figures of MM, LS and RT in Haskell#
46
32
Table 5: Problem Instance Parameters for Loild’s Benchmark Suite
Matrix Multiplication
960 × 960 matrices of integers with maximum
value of 65536.
LinSolv
Dense 62 × 62 matrix of arbitrary precision
integers with maximum value of 216 − 1.
Ray Tracer
An 1000 × 1000 image (in pixels) with a scene
comprising 640 spheres.
curve of MM is negatively affected due to the overhead caused by marshalling
large nested lists of integers. For row and block clustering of MM, the times
measured in 16 processors were little worse than those obtained for 8 and
9 processors, respectively. The marshalling overhead could be minimized
by use of Haskell arrays instead of lists to represent matrices. The Haskell#
implementations for NPB kernels, where the amount of exchanged data is far
larger, evidence this hypothesis. The toroidal version of MM yields a better
performance scalability in comparison to row and block clustering, once the
amount of data transmitted is comparatively smaller.
It is important to observe that measures of LS and RT for 32 processors
were obtained on dual processing mode across 16 nodes of the cluster. Unexpected additional overhead was observed when executing MPI programs
using the dual mode processing capabilities. This effect was more easily observed when measuring the run time of Haskell# versions of NPB kernels
CG and IS, probably due to the large amount of data exchanged between
processors in collective communication. Because of that, the results for NPB
with 32 processors was not presented. Best speedup for LS and RT were
expected for 32 non-dual processors. For that reason, in following discussion, the measures with 16 processors is used as a reference for comparing
benchmarks of Haskell# to benchmarks of GpH, Eden, and PMLS.
Comparing Haskell# results to the best ones obtained for Eden, GpH,
and PNML described in [59], one may observe that Haskell# results are
slightly better in all cases. For example, MM using toroidal solution obtains
a speedup of 11.0 on 16 processors, while a speedup of approximately 5.0 was
the best obtained in Eden toroidal solution. For LS, the speedup obtained in
47
Haskell# is 14.3 on 16 processors, while the best speedup obtained in Eden
version was 14.0. For RT, a speedup of 15.6 was obtained by Haskell# on 16
processors, while 15.1 was the best speedup obtained in PMLS version.
The results presented herein are not yet sufficient to conclude that Haskell#
programs are always more efficient than their GpH, Eden and PMLS versions.
The two compared benchmarks were obtained for distinct architectures and
using different problem sizes. However, the results presented in this paper evidence that Haskell# implementation presents comparable behavior to wellknown and mature implementations of parallel functional languages, such
as GpH, Eden and PMLS. The results obtained are not surprising, since
Haskell# run-time system is very light in relation to the complex parallel
run-time systems of GpH, Eden, and PMLS, which try to hide some parallel management details from programmers at different degrees. Decades
of experience in parallel languages design have shown that as explicit as it
is a general parallel language, assuming that it has an efficient implementation, best scalability is obtained using a simpler run-time system. The
combination of the results obtained in this paper and in [59] only confirm
this hypothesis. In this sense, Haskell# is the most explicit of all, followed
by Eden, PMLS and GpH, respectively.
6
Conclusions and Lines for Further Work
This paper introduces Haskell# , a coordination language for describing parallel execution of functional computations in Haskell. Haskell# intends to
raise the level of abstraction in explicit message-passing parallel programming on distributed architectures, such as clusters, for the development of
large scale parallel scientific computing applications. Motivating examples,
implementation issues and performance figures of Haskell# benchmarks are
also presented.
After some years of design, implementation and evaluation, Haskell# has
reached some maturity. Several works unfoldings are on progress. Firstly, a
parallel programming environment based on Haskell# have been prototyped
in JAVA, including the support for programming with visual abstractions,
integration to Petri net tools for animation, proving of formal properties, and
performance evaluation of programs. It is also under development the use of
network simulators, such as Network Simulator (NS) [31] for simulating the
performance of parallel programs. Such tool will allow to study the effect of
48
modifications to network characteristics on performance of parallel programs.
This work has important impact on studying behavior of Haskell# programs
whenever executing on grids.
Since HCL is a coordination language orthogonal to Haskell, it is conceptually possible to use other languages, in alternative to Haskell, for programming functional modules. The parallel programming environment under
development assumes that Haskell is the ideal language for specifying, prototyping and evaluating the formal properties of parallel programs. Once parallel composition is proved be safe, programmers may implement the functional
modules using a language more appropriate for implementing the functionality of the simple components. For example, numerical intensive functional
modules could be implemented in Fortran, while sorting of large amount
of numbers in parallel may be implemented in C. JAVA can be used for
programming functional modules that make access to some database. This
kind of multi-lingual compositional approach is a further development. One
important design difficult with multilingual approach is to maintain the orthogonality between languages used at coordination and computations levels
in absence of lazy and higher-order functional programming. Imperative languages, for example, do not allow to hide the control flow. It is intended to
use techniques from aspect oriented programming (AOP) for addressing this
matter. In this direction, parallel composition could be treated as an aspect
of programming. The recent appearance of heterogeneous versions of MPI
[84] is important for making feasible a multi-lingual approach for Haskell# .
An even more relevant important topic to be addressed is to develop
cost models for Haskell# skeletons, incorporating the possibility of overlapping them, and to use it for allowing Haskell# compiler to make automatic
decisions, such as better allocation of processes to processors, use of special primitives, and special restrictions on communication modes, such as
the size of buffers. However, a recent idea is to design a meta-language for
programmers to teach explicitly Haskell# compiler on how to generate the
appropriate code for a given skeleton or a combination of skeletons. The
latter approach is more in tune with Haskell# design premisses. However, it
is not difficult to see that the two lines could be combined.
Further developments will address grid enabled implementations of Haskell# .
A grid enabled version of MPI, such as the recently proposed MPICH-G2 [49],
might be used.
49
References
[1] F. Arbab, P. Ciancarini, and C. Hankin. Coordination Languages for
Parallel Programming. Parallel Computing, 24(7):989–1004, 1998.
[2] D. H. Bailey and et al. The NAS Parallel Benchmarks. International
Journal of Supercomputing Applications, 5(3):63–73, 1991.
[3] D. H. Bailey, T. Harris, W. Shapir, R. van der Wijngaart, A. Woo,
and M. Yarrow. The NAS Parallel Benchmarks 2.0. Technical
Report NAS-95-020, NASA Ames Research Center, December 1995.
http://www.nas.nasa.org/NAS/NPB.
[4] M. Baker, R. Buyya, and D. Hyde. Cluster Computing: A High Performance Contender. IEEE Computer, 42(7):79–83, July 1999.
[5] S. Balay, K. Buschelman, W. Gropp, D. Kaushik, M. Knepley, L. C.
McInnes, B. Smith, and H. Zhang. PETSc Users Manual. Technical Report ANL-95/11 Revision 2.1.3, Argonne National Laboratory, Argonne,
Illinois, 1996. http://www.mcs.anl.gov/petsc.
[6] E. Barscz, R. Fatoohi, V. Venkatakrishnan, and S. Weeratunga. Solution of Regular, Sparse Triangular Systems on Vector and DistributedMemory Multiprocessors. Technical Report NAS RNR-93-007, NASA
Ames Research Center, April 1993.
[7] G. Baumgartner, Bernholdt D. E., D. Cociorva, R. Harrison, S. Hirata,
C. Lam, M. Nooijen, R. Pitzer, J. Ramanujam, and P. Sadayappan. A
High-Level Approach to Synthesis of High-Performance Codes for Quantum Chemistry. In 2002 ACM/IEEE conference on Supercomputing,
pages 1–10, 2002. Baltimore, Maryland, USA.
[8] D. J. Becker, T. Sterling, D. Savarese, J. E. Dorban, U. A. Ranawak, and
C. V. Packer. Bewoulf: A Parallel Workstation for Scientific Computation. In 1995 International Conference on Parallel Processing, 1995.
[9] G. Bell and J. Gray. What’s the Next in High Performance Computing.
Communications of the ACM, 45(2):91–95, 2002.
[10] D. E. Bernholdt, J. Nieplocha, and P. Sadayappan. Raising Level
of Programming Abstraction in Scalable Programming Models. In
50
IEEE International Conference on High Performance Computer Architecture (HPCA), Workshop on Productivity and Performance in HighEnd Computing (P-PHEC), pages 76–84. Madrid, Spain, IEEE Computer Society, 2004.
[11] M. Bertozzi, G. Chiola, G. Ciaccio, G. Conte, P. Marenzoni, A. Poggi,
and P. Rossi. DISCO Report on the State-of-the-Art of PC Cluster
Computing. Technical Report DISI-TR-98-09, DISI, Universitá de Genova, December 1998.
[12] E. Best, J. Esparza, B. Grahlmann, S. Melzer, S. Rmer, and F. Wallner.
The PEP Verification System. In Workshop on Formal Design of Safety
Critical Embedded Systems (FEmSys’97), 1997.
[13] J. M. Bishop. Languages for Configuration Programming: A Comparison. Technical Report 94-04, University of Pretoria, 1994.
[14] S. Breitinger, R. Lógen, Y. Ortega Mallén, and R. Peña. Eden - The
Paradise of Functional Concurrent Programming. Lecture Notes in Computer Science (EUROPAR’96), 1123:710–713, 1996.
[15] J. F. Briesmeister. MCNP - A General Monte Carlo N-Particle Transport Code. Technical Report LA-12625-M, Los Alamos National Laboratory, 1993.
[16] W. H. Burge. Recursive Programming Techniques. Addison-Wesley
Publishers Ltd., 1975.
[17] G. Burns, R. Daoud, and J. Vaigl. LAM: An Open Cluster Environment
for MPI. In Proceedings of Supercomputing Symposium, pages 379–386,
1994.
[18] F.W. Burton. Functional Programming for Concurrent and Distributed
Computing. Computer Journal, 30(5):437–450, 1987.
[19] R. Buyya (ed.). High Performance Cluster Computing: Architectures
and Systems. Prentice Hall, 1999.
[20] R. Buyya (ed.). High Performance Cluster Computing: Programming
and Applications. Prentice Hall, 1999.
51
[21] F. H. Carvalho Junior and R. D. Lins. Topological Skeletons in
Haskell# . In International Parallel and Distributed Processing Symposium (IPDPS). IEEE Press, April 2003. 8 pages.
[22] F. H. Carvalho Junior, R. D. Lins, and R. M. F. Lima. Parallelising
MCP-Haskell# for Evaluating Haskell# Parallel Programming Environment. In UnB, editor, 13th Brazilian Symposium on Computer Architecture and High-Performance Computing (SBAC-PAD 2001), September
2001.
[23] F. H. Carvalho Junior, R. D. Lins, and R. M. F. Lima. Translating
Haskell# Programs into Petri Nets. Lecture Notes in Computer Science
(VECPAR’2002), 2565:635–649, 2002.
[24] M. (editor) et all Chakravarty. The Haskell 98 Foreign Function Interface (FFI) 1.0 (An Addendum to Haskell 98 Report). Glasgow University, Departament of Computing, Functional Programming Research
Group, 2002.
[25] M. Cole. Algorithm Skeletons: Structured Management of Paralell Computation. Pitman, 1989.
[26] J. Darlington, Y. Guo, H.W. To, and J. Yang. Functional Skeletons for
Parallel Coordination. Lecture Notes in Computer Science, 966:55–68,
1995.
[27] F. DeRemer and H. H. Kron. Programming-in-the-Large versus
Programming-in-the-small. IEEE Transactions on Software Engineering, pages 80–86, June 1976.
[28] J. Dongarra, I. Foster, G. Fox, W. Gropp, K. Kennedy, L. Torczon,
and A. White. Sourcebook of Parallel Computing. Morgan Kauffman
Publishers, 2003.
[29] J. Dongarra, S. W. Otto, M. Snir, and D. Walker. An Introduction to the
MPI Standard. Technical Report CS-95-274, University of Tennessee,
January 1995. http://www.netlib.org/tennessee/ut-cs-95-274.ps.
[30] C. Dornan, I. Jones, and S. Marlow.
http://www.haskell.org/alex.
52
Alex User Guide, 2003.
[31] K. Fall and K. Varadhan. The NS Manual (formely NS Notes and
Documentation). Technical report, The VINT Project, A Collaboration
between researchers at UC Berkeley, LBL, USC/ISI, and Xerox PARC,
April 2002.
[32] High Performance Fortran Forum. High Performance Fortran, Language
Specification, Version 2.0, January 1997.
[33] I. Foster. Compositional Parallel Programming Languages. ACM Transactions on Programming Languages and Systems, 18(4):454–476, 1985.
[34] I. Foster and C. Kesselman. The Grid 2: Blueprint for a New Computing
Infrastructure. M. Kauffman, 2004.
[35] D. Gelernter and N. Carriero. Coordination Languages and Their Significance. Communications of the ACM, 35(2):97–107, February 1992.
[36] R. German. SPNL: Processes as Language-Oriented Building Blocks
of Stochastic Petri Nets. In 9th Conference on Computer Performance
Evaluation, Modelling Techniques and Tools, pages 123–134. Springer
Verlag, 1997.
[37] W. Gropp, E. Lusk, N. Doss, and A. Skjellum. A High-Performance,
Portable Implementation of the MPI Message Passing Interface Standard. Parallel Computing, 22(6):789–828, 1996.
[38] M. M. Hamdan. A Combinational Framework for Paralell Programming
Using Skeleton Functions. PhD thesis, Department of Computing and
Electrical Engineering, Hariot-Watt University, January 2000.
[39] J. Hammes, O. Lubeck, and W. Böhm. Comparing Id and Haskell in a
Monte Carlo Photon Transport Code. Journal of Functional Programming, pages 283–316, July 1995.
[40] K. Hammond, J. Berthold, and R. Loogen. Automatic Skeletons in
Template Haskell. Parallel Processing Letters, 13(3), 2003.
[41] K. Hammond and G. Michaelson. Research Directions in Parallel Functional Programming. Springer-Verlag, 1999.
[42] C. Herrman and C. Lengauer. A Higher-Order Language for Dividing
and Conquer. Paralell Processing Letters, 10(2-3):239–250, 2000.
53
[43] C. A. R. Hoare. Communicating Sequential Processes. Prentice-Hall
International Series in Computer Science, 1985.
[44] P. Hudak. Serial Combinators: “Optimal” Grains of Parallelism. In
FPCA’85, pages 382–399, September 1985.
[45] P. Hudak. Para-Functional Programming in Haskell. Parallel Functional
Languages and Compilers, B. K. Szymanski, Ed. ACM Press, New York,
pages 159–196, 1991.
[46] Inmos. Occam Programming Manual. Prentice-Hall, C.A.R. Hoare Series Editor, 1984.
[47] T. Ito and Y. Nishitani. On Universality of Concurrent Expressions with
Synchronization Primitives. Theoretical Computer Science, 19:105–115,
1982.
[48] N. Karonis, B. D. Supinski, I. Foster, W. Gropp, E. Lusk, and J. Bresnahan. Exploiting Hierarchy in Parallel Computer Networks to Optimize
Collective Operation Performance. In 14th International Parallel and
Distributed Processing Symposium, pages 377–384, 2000. Los Alamitos,
CA, USA.
[49] N. Karonis, B. Toonen, and I. Foster. MPICH-G2: A Grid-enabled
Implementation of the Messaging Passing Interface. Journal of Parallel
and Distributed Computing, 63(5):551–563, 2003.
[50] O. Kaser, C.R. Ramakrishnan, I. V. Ramakrishnan, and R. C. Sekar.
Equals - A Fast Parallel Implementation of a Lazy Language. Journal
of Functional Programming, 7(2):183–217, March 1997.
[51] P. Kelly. Functional Programming for Loosely-coupled Multiprocessors. Research Monographs in Parallel and Distributed Computing, MIT
Press, 1989.
[52] G. Kiczales, J. Lamping, Menhdhekar A., Maeda C., C. Lopes, J. Loingtier, and J. Irwin. Aspect-Oriented Programming. In Lecture Notes in
Computer Science (Object-Oriented Programming 11th European Conference – ECOOP ’97), volume 1241, pages 220–242. Springer-Verlag,
November 1997.
54
[53] J. Krammer. Distributed Software Engineering. In IEEE Computer Society Press, editor, Proc. 16th IEEE International Conference on Software, 1994.
[54] Lauer M. Computing by Homomorphic Images. In G. Buchberger, G. E.
Collins, R. Loos, and Albrecht R., editors, Computer Algebra - Symbolic
and Algebraic Computation, pages 139–168. Springer, 1982.
[55] R. M. F. Lima. Haskell# - Uma Linguagem Funcional Paralela - Ambiente de Programação, Implementação e Otimização. PhD thesis, Centro
de Informáica, UFPE, July 2000.
[56] R. M. F. Lima and R. D. Lins. Translating HCL Programs into Petri
Nets. In Proceedings of the 14th Brazilian Symposium on Software Engineering, 2000.
[57] R. D. Lins. Functional programming and parallel processing. In 2nd International Conference on Vector and Parallel Processing - VECPAR‘96
- LNCS 1215 Springer-Verlag, pages 429–457, September 1996.
[58] J. D. Lipson. Chinese Remainder and Interpolation Algorithms. In
SYMSAM’71 - Symposium on Symbolic and Algebraic Manipulation,
pages 372–391. Academic Press, 1971.
[59] H.-W. Loidl, S. Priebe, A. J. Rebon, P. W. Trinder, F. Rubio, N. Scaife,
K. Hammond, S. Horiguchi, L. Loogen, G. J. Michaelson, and R. Pe na.
Comparing Parallel Functional Languages: Programming and Performance. Higher-Order and Symbolic Logic, 16(3):203–251, September
2003.
[60] H. W. et al Loidl. LinSolv: A Case Study in Strategic Parallelism. In
Glasgow Workshop on Functional Programming, pages 15–17, 1997.
[61] S. Lucco and O. Sharp. Delirium: An Embedding Coordination Language. In Proceedings of the 1990 Conference on Supercomputing, pages
515–524. IEEE Computer Society Press, 1990.
[62] S. Marlow and A. Gill.
http://www.haskell.org/happy.
Happy
55
User
Guide,
2001.
[63] Message Passing Interface Forum. MPI: A Message-Passing Interface
Standard. International Journal of Supercomputer Applications and
High Performance Computing, 8(3-4):169–416, 1994.
[64] G. Michaelson, N. Scaife, and P. King. Nested Algorithmic Skeletons for
Higher Order Functions. Parallel Algorithms and Applications, 16:181–
206, 2001.
[65] E. E. Miller and R. H. Katz. Input/Output Behavior of Supercomputing
Applications. In Proceedings of Conference Supercomputing’91, pages
567–576, November 1991.
[66] J. Nieplocha, R. J. Harrison, and R. J. Littlefield. Global Arrays: A Non-Uniform-Memory-Access Programming Model for HighPerformance Computers. The Journal of Supercomputing, 10(2):169–
189, 1996.
[67] OpenMP Architecture Review Board. OpenMP: Simple, Portable, Scalable SMP Programming, 1997.
[68] H. Ossher and P. Tarr. Multi-Dimensional Separation of Concerns
and the Hyperspace Approach. In Proceedings of the Symposium on
Software Architectures and Component Technology: The State of the
Art in Software Development. Kluwer Academics, June 2000. University
of Twente, Enschede, The Netherlands.
[69] C. Pareja, R. Pe na, F. Rubio, and C. Segura. Optimizing Eden
by Transformation. In S. Gilmore, editor, Trends in Functional Programming (2nd Scottish Functional Programming Workshop), volume 1,
chapter 4, pages 39–52. Intellect Books, 2000.
[70] B. K. Pasquale and G. Plyzos. A Statis Analysis of I/O Characterization
of Scientific Applications in a Production Workload. In Proceedings of
Conference Supercomputing’93, pages 388–397, November 1993.
[71] J. L. Peterson. Petri Net Theory and Modeling of Systems. PrenticeHall, 1981.
[72] C. A. Petri. Kommunikation mit Automaten. Technical Report RADCTR-65-377, Griffiths Air Force Base, New York, 1(1), 1966.
56
[73] S. L. Peyton Jones, C. Clack, and J. Salkild. GRIP - A HighPerformance Architecture for Parallel Graph Reduction. FPCA’87: Conference on Functional Programming Languages and Computer Architecture - Springer-Verlag LNCS 274, pages 98–112, 1987.
[74] S. L. Peyton Jones, A. Gordon, and S. Finne. Concurrent Haskell. In
POPL’96 - Symposium on Principles of Programming Languages, ACM
Press, pages 295–308, January 1996.
[75] S. L. Peyton Jones and J. (editors) Hughes. Report on the Programming Language Haskell 98, A Non-strict, Purely Functional Language,
February 1999.
[76] M. J. Plasmeijer and M. van Eekelen. Functional Programming and
Parallel Graph Rewriting. Addison-Wesley Publishers Ltd., 1993.
[77] Post D. E. The Coming Crisis in Computational Science. In IEEE
International Conference on High Performance Computer Architecture
(HPCA), Workshop on Productivity and Performance in High-End
Computing (P-PHEC). Madrid, Spain, 2004.
[78] M. Quinn. Parallel Computing. McGraw-Hill, 1994.
[79] F. A. Rabhi and S. Gorlatch. Patterns and Skeletons for Parallel and
Distributed Computing. Springer, 2002.
[80] S. Roch and P. Starke. Manual: Integrated Net Analyzer Version 2.2,
1999.
[81] P. M. Sansom and S. L. Peyton Jones. Time and Space Profiling for
Non-Strict, Higher-Order Functional Languages. In Proceedings of the
22nd ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages, pages 355–366. ACM Press, 1995.
[82] V. Sarkar, C. Williams, and K. Ebcioğlu. Application Development Productivity Challenges for High-End Computing. In IEEE International
Conference on High Performance Computer Architecture (HPCA),
Workshop on Productivity and Performance in High-End Computing,
pages 14–18, 2004.
57
[83] D. B. Skillicorn and D. Talia. Models and Languages for Parallel Computation. ACM Computing Surveys, 30:123–169, June 1998.
[84] J. M. Squyres, A. Lumsdaine, W. L. George, J. G. Hagedorn, and J. E.
Devaney. The interoperable message passing interface (IMPI) extensions
to LAM/MPI. In Proceedings, MPIDC’2000, March 2000.
[85] F. Taylor. Parallel Functional Programming by Partitioning. PhD Thesis, Department of Computing, Imperial College of Science, Technology
and Medicine, University of London, January 1997.
[86] P. Trinder, K. Hammond, J. S. Mattson Junior, A. S. Partridge, and S.
P. L. Jones. GUM: A Portable Parallel Implementation of Haskell. In
PLDI’96 - Programming Languages Design and Implementation, pages
79–88, 1996.
[87] P. W. Trinder, H-W. Loidl, and R. F. Pointon. Parallel and Distributed
Haskells. Journal of Functional Programming, 12(4/5):469–510, July
2002.
[88] P.W. Trinder, K. Hammond, and H. W. Loidl. Algorithm + Strategy =
Parallelism. Journal of Functional Programming, 8(1):23–60, January
1998.
[89] R. Valk and G. Vidal Naquet. Petri Nets and Regular Languages. Journal of Computer and System Sciences, 23:299–325, 1981.
[90] P. Wadler.
Monads for Functional Programming.
Advanced
Functional Programming, LNCS 925, Springer-Verlag, 1995.
http://cm.bell-labs.com/cm/cs/who/wadler/papers/marktoberdorf/marktoberdorf.ps.gz.
[91] Weber M. and Kindler E. The Petri Net Markup Language. Lecture
Notes in Computer Science, 2002. Submitted for publication in april
2002.
[92] A. Zimmermann, J. Freiheit, R. German, and G. Hommel. Petri Net
Modelling and Performability Evaluation with TimeNET 3.0. In 11th
Int. Conf. on Modelling Techniques and Tools for Computer Performance
Evaluation (TOOLS’2000), pages 188–202. Lecture Notes in Computer
Sciente, 2000.
58
A
The Formal Syntax of HCL
In what follows, it is described a context-free grammar for HCL, the Haskell#
Configuration Language, whose syntax and programming abstractions were
informally presented in Section 2. Examples of HCL configurations and their
meanings were presented in Sections 2 and 3. The notation employed here
is similar to that used for describing syntax of Haskell 98 [75]. Indeed, some
non-terminals from that grammar are reused here, once some Haskell code
appears in HCL configurations. They are faced italic and bold. A minor
difference on notation resides on the use of (. . .)? , instead of [. . .], for describing optional terms. For simplicity, notation for indexed notation is ignored
from the description of formal syntax of HCL. It may be resolver by a preprocessor, before parsing.
A.1
Top-Level Definitions
configuration
→
header
→
static parameter list→
component interface→
declaration
→
|
|
|
A.2
header declaration 1 . . . declaration n (n ≥ 0)
component ID static parameter list ? component interface ?
< ID1 . . . IDn > (n ≥ 0)
ports naming
import decl | use decl
| iterator decl | interface decl
unit decl
| assign decl | replace decl | channel decl
unify decl | factorize decl | replicate decl | bind decl
haskell code
Use Declaration
use decl → use use spec
use spec → id | id.use spec | id.{ use spec 1 , . . . , use spec n }
A.3
(n ≥ 1)
Import Declaration
import decl → impdecl
A.4
Iterator Declaration
iterator decl → iterator id1 , . . ., idn range [ numeric exp , numeric exp ] (n ≥ 1)
59
A.5
Interface Declaration
interface decl → interface (context =>)? ID tyvar 1 . . . tyvar k interface spec
interface spec → interface ports spec
(where : interface inheritance)? (behavior : behavior expression)?
A.5.1
Interface Ports Description
interface ports spec → port spec list -> port spec list
port spec list
→ port spec | ( port spec 1 , . . . , port spec n ) (n ≥ 2)
→ id (*)? (:: atype)? | id
port spec
A.5.2
Interface Composition
interface inheritance
→
→
interface slice
ports naming composition →
|
ports naming
→
port naming list
→
A.5.3
Interface Behavior
behavior expression →
action
→
|
|
|
condition
→
disjunction
→
sync conjunction
→
simple conjunction →
A.6
interface slice 1 # . . . # interface slice k (k ≥ 1)
id @ ID | ID ports naming composition
ports naming
( ports naming 1 # . . . # ports naming n ) (n ≥ 1)
port naming list -> port naming list
id | ( id1 , . . . , idn ) (n ≥ 1)
(sem id1 , . . . , idn )? : action (n ≥ 1)
par { action 1 ; . . . ; action n } | seq { action 1 ; . . . ; action n }
alt { action 1 ; . . . ; action n } | repeat action condition ?
if condition then action else action
id ! | id ? | signal id | wait id (n ≥ 2)
until disjunction | counter numeric exp
conjunction 1 ‘|’ . . . ‘|’ conjunction n (n ≥ 1)
h simple conjunction i | simple conjunction
id | ( id1 & . . . & idn ) (n ≥ 1)
Unit Declaration
→ unit unit spec
unit decl
unit spec
→ (*)? id (# unit interface)? (wire wf setup 1 , . . . , wf setup n )?
unit interface→ ID ports naming composition ? | interface spec
60
wf setup
→
group spec →
group type →
wire function →
A.7
id (group type group spec)? (: wire function)?
{ id1 , . . ., idn } | * numeric exp
any | all
? | exp
Assignment Declaration
assign decl
→
assigned component →
actual parameter list→
→
assigned unit
A.8
assign assigned component to assigned unit
ID actual parameter list ? ports naming composition ?
< numeric exp 1 , . . . , numeric exp n > (n ≥ 1)
qid ports naming composition ?
Replace Declaration
replace decl → replace qid ports naming composition ? by operand unit
A.9
Channel Declaration
channel decl → connect qid -> qid to qid <- qid , comm mode
comm mode → synchronous | buffered numeric exp | ready
A.10
Unification Declaration
→ unify operand unit 1 , . . . , operand unit n to unit spec
adjust wire wf setup 1 , . . . , wf setup k (n ≥ 2, k ≥ 1)
operand unit
→ qid # interface pattern 1 . . . # interface pattern n (n ≥ 1)
interface pattern → port pattern list -> port pattern list | id
port pattern list → pattern | ( pattern 1 , . . . , pattern n )
pattern
→ id | @ qid | |
unify decl
A.11
Factorization Declaration
factorize decl → factorize operand unit to unit spec 1 . . . unit spec n
adjust wire wf setup 1 , . . . , wf setup k (n ≥ 2, k ≥ 1)
61
A.12
Replication Declaration
replicate decl → replicate operand unit 1 , . . . , operand unit n into numeric exp
adjust wire wf setup 1 , . . . , wf setup k (n ≥ 2, k ≥ 1)
A.13
Bind Declaration
bind declaration → bind qid -> qid to -> id | bind qid <- qid to <- id
A.14
Miscelaneous
haskell code → topdecls
qid → id1 ‘.’ . . . ‘.’ idn (n ≤ 2)
qID → ID1 ‘.’ . . . ‘.’ IDn (n ≤ 2)
B
An Algebraic Semantics for Haskell# Components
This appendix presents an algebra intending to formalize semantics of Haskell#
programming abstractions at coordination level. A Haskell# component H
may be defined by an algebra with the following elements:
H =< G, R, C >
where G is a set of generators, R is a set of relations on generators, and
C is a set of restrictions on relations, defined as following:
G=
C,
composed components
S,
simple components
U,
units
G,
ports groupings
P,
individual ports
R,
kinds of processes: repetitive or non-repetitive
D,
port directions: input or output
T,
port type: any or all
M communication modes: synchronous, buffered or ready
62
R=
ω : {⋆} → C ∪ S
δ : C → 2U ,
ψ : P → P,
γ : U → C ∪ S,
π : G → U × D,
β : U → G∗ ,
τ : Q → 2P × T,
ρ : U → R,
ν :P ×P ×M
λ : G → N at
∗
ι : U → 2G × 2G
main component
units that comprise a component
association of ports to argument/return points
component associated to a unit
unit of a port grouping
behavior of a unit
grouping of ports
type of process
communication channels
nesting factor of a stream port
interface of a unit
C = {R1, R2 , R3 , R4 , R5 , R6 , R7 , R8 , R9 , R10 , R11, R12 }
A Haskell# program is a component that may execute. Essentially, it does
not have virtual units in its composition (it is not a partial skeleton). All
units are assigned to a component (∀u : u ∈ U : (∃c : c ∈ C ∪ S : γ(u) = c)).
In what follows, the restrictions from R1 to R12 are described. They are
formulas in predicate logic (predicate) of the following form (∀b1 , b2 , . . . , bn :
R : P ) or (∃b1 , b2 , . . . , bn : R : P ), where ∀ and ∃ are the usual existential
quantifier, bi , 1 ≤ i ≤ n, are bound variables, R is a formula that specifies
the set of values of bound variables, and P is a logical predicate.
The restriction R1 states that component ω (main component) is the only
component that is not assigned to any unit:
R1 ⊢ ∀u : u ∈ U : γ(u) 6= ω
(1)
R2 states that cyclic dependencies may not occur in component hierarchy:
R2 ⊢ ∀u : u ∈ U ∧ γ(u) 6= ⊥ : u ∈
/ (δ ◦ γ)(u)
s∈S
where : δ(s) = ∅
S
δ(c) = u∈δ(c) (δ ◦ γ)(u) c ∈ C
(2)
R3 states that a cluster is repetitive whenever all units belonging to its
assigned component are repetitive:
R3 ⊢ ∀u : u ∈ U ∧ (∃c : c ∈ C : γ(u) = c ∧ δ(c) 6= ∅) :
ρ(u) = Repetitive ⇔ (∀u : u ⊆ (δ ◦ γ)(u)) : ρ(u) = Repetitive)
63
(3)
R4 state that groups of ports are disjoint, R5 states that all individual
ports belong to a group of ports, and R6 states that groups of ports must
not be empty:
R4 ∀g, g ′ : g, g ′ ∈ G : π(g) ∩ π(g ′ ) = ∅
R5 ∀p ∃g : p ∈ P ∧ g ∈ G : p ∈ τ (g)
(4)
R6 ∀g : g ∈ G : τ (g) 6= ∅
In the algebra, all ports are treated as non-empty groups. Thus, an
individual port in a Haskell# program is represented as a group containing
an unique port. The restrictions above makes possible to define a “inverse”
relation τ 3 , such that τ (p) returns the group g that p belongs. It is useful
for simplifying next formulations.
Restrictions R7 , R8 , and R9 specifies rules for formation of channels.
Respectively, they say that channels are point-to-point, unidirectional and
have the same nesting factors:
R7 ⊢ (po1 , pi1 , m) ∈ ν ∧ (po2 , pi2 , m) ∈ ν ⇒ po1 = po2 ⇔ pi1 = pi2
R8 ⊢ (po , pi , m) ∈ ν ⇒ (∃u, u′ : u, u′ ∈ U : (π ◦ τ )(po ) = (u, Output) ∧ (π ◦ τ )(pi ) = (u′ , Input))
R9 ⊢ (po , pi , m) ∈ ν ⇒ (λ ◦ τ )(po ) = (λ ◦ τ )(pi ))
(5)
Let u be a cluster (γ(u) ⊆ C) and p be an individual port belonging to
group g, such that π(g) = (u, d), for d ∈ D (p belongs to interface of u). The
restriction R10 ensures that ψ(p) (argument or exit point of γ(u)) is a port
with the same direction of p belonging to interface of unit u′ , such that u′ is
a unit belonging to the component γ(u) (u′ ∈ (δ ◦ σ)(u)):
R10 ⊢ ∀p : (π ◦ τ )(p) = (u, d) ∧ γ(u) ⊆ C : ((π ◦ ψ)(p) = (u′ , d) ∧ u′ ∈ (δ ◦ γ)(u))
(6)
The restriction R11 defines the relation ι, which describes the interface
of a unit:
R11 ⊢ ι(u) =< {g | π(g) = (u, d)}, β(u) >
3
This is not the strict mathematical notion of inverse function, from set theory.
64
(7)
R12 says that ports belonging to the same group whose communication
pairs also belongs to the same groups are essentially the same port.
R12 ⊢ (po1 , pi1 , m1 ) ∈ ν ∧ (po2 , pi2 , m2 ) ∧ τ (po1 ) = τ (po2 ) ∧ τ (pi1 ) = τ (pi2 ) ⇒ po1 = po2 ∧ pi1 = pi2
(8)
B.1
Formalizing Interfaces
This section formalizes homomorphism relations between interfaces, which
are essential for formalizing unification and factorization operations in the
next section.
B.1.1
The # Operator
The # operator allows for combining to interfaces, generating a new interface
that inherits characteristics from original ones. It is defined as following:
ˆ B2 >, where I1 =< Q1 , B1 > and I2 =< Q2 , B2 >
I1 # I2 =< Q1 ∪ Q2 , B1 ∪
(9)
ˆ
The sets of ports from operand interfaces may overlap. The operator ∪
generates a new formal language describing a behavior for interface I1 #I2 ,
which is compatible with original behavior of I1 and I2 , in separate. Given
an interleaving operator ⊙, from concurrent expressions [47] and ℓ a function that returns the language generated by a concurrent expression, formal
ˆ is:
definition of ∪
ˆ B2 = ℓ [(w1 ⊙ u1 ) s (w2 ⊙ u2 ) s . . . s (wn ⊙ un )] , n ≥ 1
B1 ∪
where
s ∈ Q1 ∩ Q2
w1 s w2 s . . . s wn ∈ B1
u 1 s u 2 s . . . s u n ∈ B2
w1 w2 . . . wn ∈ (Q1 − {s})∗
u1 u2 . . . un ∈ (Q2 − {s})∗
65
(10)
ˆ B2 corresponds to interIf operand interfaces do not overlap ports, B1 ∪
leaving of their original behaviors (B1 ⊙ B2 ). Overlapping ports may be
interpreted as synchronization points when combining formal languages B1
and B2 .
B.1.2
Homomorphisms Between Interfaces
Let I1 =< Q1 , B1 > and I2 =< Q2 , B2 > be interface classes. Let H be a
pair < h : Q1 → Q2 , h : B1 → Q2 ∗ >, where h is defined as following:
h(ǫ) = ǫ
h(aw) = h(a)h(w)
(11)
With respect to H =< h, h >, the following interface relations are defined:
H
I1 ⊑ I2 ⇔ Im(h) ⊆ B2
H
I1 ⊒ I2 ⇔ Im(h) ⊇ B2
(12)
H
I1 ≡ I2 ⇔ Im(h) = B2
Relations ⊑ and ⊒ characterize homomorphisms between interfaces, while
≡ characterize isomorphisms between them.
B.2
An Algebra for Haskell# Programming
Now, it is defined an algebra to formalize Haskell# programming task. Operations over units are defined here: unification, factorization, replication
and assignment. They may be used to overlap and nest components that
comprise a Haskell# component. An algebra for Haskell# programming is
defined as:
< {H}, {u : H × H, f : H × H, a : H × H, r : H × H, i : H × H}, ∅ >
where generator H contains all well-formed Haskell# components. The
relations u, f , a and r represents sets of pairs (h1 , h2 ), h1 ∈ H and h2 ∈ H,
where h2 is a Haskell# component obtained from Haskell# component h1
from an application of unification, factorization, assignment or replication
operations, respectively, defined further. The relation i is a identity relation
containing pairs (h, h), ∀h ∈ H.
66
In what follows, assignment, unification, factorization, and replication operations, homomorphisms between Haskell# components, are defined. Since
all Haskell# components may be described using HCL configurations that
should be generated using a context-free grammar, the set H is recursively
enumerable. Thus, in what follows, the i-ith Haskell# program, i ≥ 0, is
denoted by #i = (Gi , Ri , Ci ), where Gi = {Ci , Si , Ui , Gi , Pi , Ri , Di , Ti , Mi },
Gi = {ωi , δi , ψi , γi , πi , βi , τi , ρi , νi , λi }.
B.2.1
Unification and Factorization
Unification and factorization, informally introduced in Section 2.1.10, are
formalized here as mutually reversible relations in the algebra of Haskell#
programming. For instance, consider two Haskell# components and their
algebraic description, denoted by #k and #j , for some j, k ≥ 0. Consider
V̂ = hv1 , v2 , . . . , vn i an ordered sub-set of virtual units in Uk , and their
respective interfaces Î = hI1 , I2 , . . . , In i, such that Ii = ι(vi ) = (Qi , Bi )
, for 1 ≤ i ≤ n. Also, consider a virtual unit v ∈ Uj and its interface
I = ι(v) = (Q, B). A set of interface mappings Ĥ = hH1 , H2 , . . . , Hn i,
where Hi = (hi , hi ) maps interface Ii to interface I is defined. Suppose that
#j is obtained from #k by unification of virtual units in V̂ to a unique virtual
unit v. It is also supposed correct to say that #k is obtained from #j by
factorization of the virtual unit v onto the set of virtual units V̂.
Two restrictions may be ensured in a correct application of unification
and factorization operations. The first one imposes behavior preserving restrictions for units, stating that v is a proper unification of virtual units in
Hi
set V̂ if Ii ⊒ I, for 1 ≤ i ≤ n. Analogously, units in V̂ constitute a proper
Hi
factorization of v if Ii ⊑ I, for 1 ≤ i ≤ n. The second one establishes restrictions for preservation of network connectivity. But before to talk about
them, it is necessary to define relation τ̂ : Q → 2Pj . It makes possible to
formalize partitioning of groups of ports, which must be configured explicitly
in factorizations. In unifications, it is not necessary to configure τ̂ explicitly
using HCL, since the inverse of partitioning of groups of ports is the union of
them, which is resolved by merging the groups. Ports b and e in Figure 6 are
examples of partitioning (right to left) and union (left to right) of ports. The
relation τ̂ must satisfy the restriction defined in Equation 13, which relates
it with interface mapping Ĥ.
67
∀q ′ : q ′ ∈ Gj ∧ (∃q : q ∈ Q : Ĥ(q) = q ′ ) :
[
q∈R
τ̂ (q) = τ (q ′ ), R = {q | q ∈ Q ∧ Ĥ(q) = q ′ }
(13)
In this paragraph, restrictions for ensuring preservation of network connectivity with respect to unification/factorization are discussed. In the trivial case, where overlapping of ports does not occur (Ĥ(q1 ) = Ĥ(q2 ) ⇒
τ̂ (q1 )∩τ̂ (q2 ) = ∅), all ports and channels are preserved (Pj = Pk and νj = νk )
after applying unification/factorization. Essentially, only the sets of units
(Uj − Uk = {v} ∧ Uk − Uj = V̂), ownership of ports (relations πk and πj ), and
grouping of ports (relations τk and τj ) differs between #j and #k . Ownership
and grouping of ports is affected by interface mappings Ĥ. If overlapping
of port occurs, some adjustment of ports and channels may be necessary in
order to ensure obedience to restrictions for channel formation. For instance,
consider a port p, such that ∃Q : Q ⊆ Q : (∀q : q ∈ Q : p ∈ τ̂ (q)) ∧ |Q| ≥ 2.
From the perspective of factorization, p is interpreted as a port of unit v,
in component #j , that have more than one port in Pk associated to it, possibly all belonging to distinct units in the set V̂ of component #k . For
ensuring point-to-point nature of channels (R7 ), the communication pair of
p, p ((p, p, m) ∈ νj ∨ (p, p, m) ∈ νj ), must be replicated in |Q| copies as
consequence of factorization. They are connected to the ports belonging
to groups in Q that have association to p. From the perspective of unification, p is a port of #j that comes from unification of a set of ports
Q = {p′ | p′ ∈ Pk ∧ p ∈ τ̂ (p′ )} of #k . The communication pairs of ports in
Q, Q, are members of the same group of ports (∃g : g ∈ Gk : Q ⊆ τk (g)).
In such case, in order to satisfy restriction R12 , ports in P are unified in a
single port p in #j , the communication pair of p.
B.2.2
Assignment
In an executable Haskell# program, application component must not contain
virtual units. Thus, it is necessary to define an operation for associating components to virtual units (nesting composition). Let #k and #i be Haskell#
programs, v ∈ Vk be a virtual unit in program #k , and ψ a mapping from
ports of interface of v to arguments and exit points of ωi (main component
of #i ).
68
Assignment of main component of #i (ωi ) to virtual unit v of #k , produces a new program #k , the union of generators and relations from two
programs, where v is associated to ωi through γk . Arguments and exit points
of ωi are associated to v ports through ψk , using ψ.
B.2.3
Replication
Let #k be a Haskell# program. Given a positive integer r > 1 and a collection
of units U ⊆ Uk , U = {u1 , u2 , . . . , un }, it is possible to replicate the subnetwork induced by units in U in r copies, forming a new program #j . In
order to maintain network connectivity and attendance to Haskell# algebra
restrictions, when defining #j from #k , it is necessary to replicate ports from
units that are not in U but are connected to any port of some unit in U. HCL
allows for specifying wire functions for new groups. Channels connecting unit
ports between units in U are also replicated in n copies, one connecting each
pair of ports from the n units copies.
C
C.1
HCL Code for NPB Benchmarks EP, IS,
CG, and LU
EP
component EP<no nodes,mk, mm, nn, nk, nq, epsilon, a, s> with
#define PARAMETERS (EP Params i no nodes mk mm nn nk nq epsilon a s)
iterator i range [1..no nodes]
use Skeletons.Collective.AllReduce
use EP FM −− EP Functional Module
interface IEP (sx, sy, q) → (sx,sy,q) where: sx@IAllReduce Double # sy@IAllReduce Double # q@IAllReduce UDVector
behaviour: seq {do sx; do sy; do q}
unit sx comm; assign AllReduce<no nodes, MPI SUM, MPI DOUBLE> to sx comm
unit sy comm; assign AllReduce<no nodes, MPI SUM, MPI DOUBLE> to sy comm
unit q comm ; assign AllReduce<no nodes, MPI SUM, MPI DOUBLE> to q comm
[/ unify sx comm.p[i] # sx, sy comm.p[i] # sy, q comm.p[i] # q to ep unit[i] # IEP
assign EP FM (PARAMETERS, sx, sy, q) → (sx, sy, q) to ep unit[i] # sx # sy # q /]
C.2
IS
component IS<problem class, num procs, max key log2, num buckets log2,
total keys log2, max iterations, max procs, test array size> with
#define PARAMETERS (IS Params problem class num procs max key log2 num buckets log2
total keys log2 max iterations max procs test array size)
iterator i range [1, num procs]
69
use Skeletons.{Misc.RShift, Collective.{AllReduce, AllToAllv}}
use IS FM −− IS Functional Module
interface IIS (bs*, kb*, k) → (bs*, kb*, k) where: bs@IAllReduce (UArray Int Int) # kb@IAllToAllv (Int, Ptr Int) # k@RShift Int
behaviour: seq {repeat seq {do bs; do kb} until <bs & kb>; do k}
unit bs comm ; assign AllReduce<num procs, MPI SUM, MPI INTEGER> to bs comm
to kb comm
unit kb comm ; assign AllToAllv<num procs>
; assign RShift<num procs> 0 →
to k shift
unit k shift
[/ unify bs comm.p[i] # bs, kb comm.p[i] # kb, k comm.p[i] # k to is unit[i] # IIS
assign IS FM (PARAMETERS, bs, kb, k) → (bs, kb, k) to is unit[i] # bs # kb # k /]
C.3
C.3.1
O kernel CG
Esqueleto Transpose
component Transpose<dim, col factor>
iterator i, j range [1..dim]
iterator k range [1..col factor]
interface ITranspose (x::UDVector) → (w::UDVector) behaviour: seq { w!; x? }
[/ unit trans[i][j] # ITranspose wire x all*dim:?, w all*dim:? /]
[/ connect trans[i][j] → w[k] to trans[k][i] ← x[j] /]
[/ factorize trans[i][j] # w → x to [/ u[(.i-1)*col factor+k][.j] # w → x /]
adjust wires w: sum arrays, x: split and scatter /]
C.3.2
Componente CG
component CG<dim, col fator, na, nonzer, shift, niter, rcond zvv> # () → (zeta, x) with
#define PARAMETERS (CG Params dim (dim*col factor) na nonzer shift niter rcond zvv)
use Skeletons.MPI.Collective.AllReduce
use Transpose
use CG FM −− CG Functional Module
index i range [1..dim]
index j range [1..col factor]
interface ICG (r*,q**,rho**,aux**,rnorm*,norm temp 1*,norm temp 2*)
→ (r*,q**,rho**,aux**,rnorm*,norm temp 1*,norm temp 2*, x::Array Int Double, zeta::Double)
where: q@ITranspose # rho@IAllReduce Double # aux@IAllReduce Double # rnorm@IAllReduce Double #
r@ITranspose # norm temp 1@IAllReduce Double # norm temp 2@IAllReduce Double #
behaviour: repeat seq {do rho; repeat seq {do q; do aux; if rho then do rho else skip }
until <q & aux & rho>;
do r; do rnorm; do norm temp 1; do norm temp 2;}
until <r & rnorm & q & aux & rho & norm temp 1 & norm temp 2>
assign
unit q comm;
unit r comm;
assign
assign
[/ unit rho comm[i];
unit aux comm[i];
assign
assign
unit rnorm comm[i];
unit norm temp 1 comm[i];assign
unit norm temp 2 comm[i];assign
Transpose<dim,
Transpose<dim,
AllReduce<dim
AllReduce<dim
AllReduce<dim
AllReduce<dim
AllReduce<dim
dim *
dim *
* col
* col
* col
* col
* col
dim * col factor>
dim * col factor>
factor, MPI SUM, MPI
factor, MPI SUM, MPI
factor, MPI SUM, MPI
factor, MPI SUM, MPI
factor, MPI SUM, MPI
to
to
DOUBLE>to
DOUBLE>to
DOUBLE>to
DOUBLE>to
DOUBLE>to
q comm
r comm
rho comm[i]
aux comm[i]
rnorm comm[i]
norm temp 1 comm[i]
norm temp 2 comm[i] /]
[/ unify q comm.u[i][j] # q, rho comm[i].p[j] # rho, aux comm[i].p[j] # aux,
r comm.u[i][j] # r, rnorm comm[i].p[j] # rnorm,
norm temp 1 comm[i].p[j] # norm temp 1, norm temp 2 comm[i].p[j] # norm temp 2
to cg[i][j] # ICG
assign CG FM (PARAMETERS, q, rho, r, aux, rnorm, norm temp 1, norm temp 1)
→ (q, rho, aux, r, rnorm, norm temp 1, norm temp 2)
to cg[i][j] # q # rho # aux # r # rnorm # norm temp 1 # norm temp 2
/]
70
C.4
A Aplicação Simulada LU
C.4.1
Esqueleto Exchange 1b
component Exchange 1b < xdiv , ydiv ,itmax> with
iterator m range [0..( ydiv -1)]
iterator n range [0..( xdiv -1)]
interface Exchange 1b # (from north**,from west**, from south**, from east** :: UArray (Int,Int) Double)
→ (to south**, to east**, to north**,tp west** :: UArray (Int,Int) Double)
behaviour: repeat { seq {repeat seq{from north?;
from west?;
to south!;
to east!} until <from north & from west &
to south & to east>;
repeat seq{from south?;
from east?;
to north!;
to west!} until <from south & from east &
to north & to west>
} until itmax
[/unit bigLoop[n][m] # Exchange 1b /]
[/ connect
connect
connect
connect
C.4.2
bigLoop[n][m]
bigLoop[n][m]
bigLoop[n][m]
bigLoop[n][m]
→
→
→
→
to
to
to
to
south to bigLoop[(n+1) mod xdiv ][m] ← from north
east to bigLoop[n][(m+1) mod ydiv ] ← from west
north to bigLoop[(n+ xdiv -1) mod xdiv ][m] ← from south
west to bigLoop[n][(m+ ydiv -1) mod ydiv ] ← from east /]
Esqueleto Exchange 3b
component Exchange 3b < xdiv , ydiv > with
iterator m range [0..(ydiv-1)]
iterator n range [0..(xdiv-1)]
interface IExchange 3b # (from north*,from south*, from east*, from west*::UArray Int Double)
→(to north*, to south*, to east*, to west*:: UArray Int Double)
behaviour: repeat seq {to south!;from noth?; to north!;from south?;
to east!; from west?; to west!; from east?} until to south
[/ unit g1[n][m] # IExchange 3b /]
[/ connect
connect
connect
connect
C.4.3
g1[n][m]
g1[n][m]
g1[n][m]
g1[n][m]
→
→
→
→
g1
g1
g1
g1
ts to g1 [(n+1) mod xdiv ][m]← g1 fn
tn to g1 [(n+ xdiv -1) mod xdiv ][m]← g1 fs
te to g1 [n][(m+1) mod ydiv ]← g1 fw
tw to g1 [n][(m+ ydiv -1) mod ydiv ]← g1 fe /]
Esqueleto Exchange 4
component Exchange 4 < xdiv , ydiv > with
iterator
iterator
iterator
iterator
n range [0..(ydiv-2)]
s range [1..(ydiv-1)]
l range [0..(xdiv-2)]
r range [1..(xdiv-1)]
iterator i range [1..(ydiv-2)]
iterator j range [1..(xdiv-2)]
interface IExchange 4
interface IExchange 4 Null specializes IExchange 4
interface IExchange 4 Border # (in::UArray Int Double) → (out::UArray Int Double)
behaviour: seq {out!;in?} specializes IExchange 4
interface IExchange 4 Corner NW # (in1, in2::UArray Int Double) → ()
behaviour: seq {in1?;in2?} specializes IExchange 4
71
interface IExchange 4 Corner SE # ()→(out1, out2::UArray Int Double)
behaviour: seq {out1!;out2!} specializes IExchange 4
[/ unit h0[i][j] # IExchange 4 Null /]
unit h0[0][0] # IExchange 4 Corner NW
unit h0[ xdiv -1][ ydiv -1] # IExchange 4 Corner SE
[/ unit h0[n][0] # IExchange 4 Border /]
[/ unit h0[s][ xdiv -1] # IExchange 4 Border /]
[/ unit h0[l][0] # IExchange 4 Border /]
[/ unit h0[r][ ydiv -1] # IExchange 4 Border /]
[/ connect h0[0][s] → out to h0[0][s-1] ← in /]
[/ connect h0[ xdiv -1][s] → out to h0[ xdiv -1][s-1] ← in /]
[/ connect h0[r][0] → out to h0[r-1][0] ← in /]
[/ connect h0[r][ ydiv -1] → out to h0[r-1][ ydiv -1] ← in /]
C.4.4
Esqueleto Exchange 5
component Exchange 5 < xdiv , ydiv > with
iterator
iterator
iterator
iterator
m range [0..(ydiv-1)]
n range [0..(xdiv-1)]
i range [1..(ydiv-2)]
j range [1..(xdiv-2)]
interface generalization IExchange 5
interface
interface
interface
interface
IExchange 5
IExchange 5
IExchange 5
IExchange 5
behaviour:
Null specializes IExchange 5
Top # (in::UArray Int Double)→() behaviour: in? specializes IExchange 5
Bottom # ()→(out::UArray Int Double) behaviour: out! specializes IExchange 5
Side # (in::UArray Int Double)→(out::UArray Int Double)
seq {out!;in?} specializes IExchange 5
[/ unit h1[i][m] # IExchange 5 Null /]
unit h1[0][0] # IExchange 5 Top
unit h1[0][ ydiv -1] # IExchange 5 Top
unit h1[ xdiv -1][0] # IExchange 5 Bottom
unit h1[ xdiv -1][ ydiv -1] # IExchange 5 Bottom
[/ unit h1[j][0] # IExchange 5 Side
unit h1[j][ ydiv -1] # IExchange 5 Side /]
[/ connect h1[l][0]→out to h1[l-1][0]←in /] [/ connect h1[l][ ydiv -1]→out to h1[l-1][ ydiv -1]←in /]
C.4.5
Esqueleto Exchange 6
component Exchange 6 < xdiv , ydiv > with
iterator m range [0..(ydiv-1)]
iterator n range [0..(xdiv-1)]
iterator i range [1..(ydiv-2)]
iterator j range [1..(xdiv-2)]
interface generalization IExchange 6
interface
interface
interface
interface
IExchange 6
IExchange 6
IExchange 6
IExchange 6
behaviour:
Null specializes IExchange 6
Left # (in::UArray Int Double) → () behaviour: in? specializes IExchange 6
Right # ()→(out::UArray Int Double) behaviour: out! specializes IExchange 6
Side # (in::UArray Int Double) → (out::UArray Int Double)
seq {out!; in?} specializes IExchange 6
[/ unit h1[i][m] # IExchange 6 Null /]
unit h1[0][0] # IExchange 6 Left
72
unit h1[0][ ydiv -1] # IExchange 6 Left
unit h1[ xdiv -1][0] # IExchange 6 Right
unit h1[ xdiv -1][ ydiv -1] # IExchange 6 Right
[/ unit h1[j][0] # IExchange 6 Side
unit h1[j][ ydiv -1] # IExchange 6 Side /]
[/ connect h1[0][l] → out to h1[0][l-1] ← in /] [/ connect h1[ xdiv -1][l] → out to h1[ xdiv -1][l-1 ← in /]
C.4.6
Componente LU (Esqueleto de Aplicação)
component LU <nprocs,problem size,dt default,itmax> with
#define d ilog2(nprocs)/2
#define xdiv (ipow2(if (d*2 == ilog2(nprocs), d, d + 1)))
#define ydiv (ipow2(d))
#define PARAMETERS (LU Params nprocs problem size dt default itmax)
use Skeletons.MPI.{AllReduce,BCast}
use Exchange 1b, Exchange 3b, Exchange 4, Exchange 5, Exchange 6
use LU FM −− LU Functional Module
iterator m range [1,ydiv]
iterator n range [1,xdiv]
interface ILU (ipr,inorm,itmax,nx0,ny0,nz0,dt,omega,tolrsd,rsdnm*,errnm,frc1,frc2,frc3,rsd1,rsd0,u1,phis,phiver,phivor)
→ (ipr,inorm,itmax,nx0,ny0,nz0,dt,omega,tolrsd,rsdnm*,errnm,frc1,frc2,frc3,rsd1,rsd0,u1,phis,phiver,phivor)
where: ipr, inorm, itmax, nx0, ny0, nz0 @IBCast Int
#
dt, omega
@IBCast Double
#
tolrsd
@IBCast MyArray1d,
#
rsdnm, errnm
@IAllReduce MyArray1d #
frc1, frc2, frc3
@IAllReduce Double
#
rsd1
@IExchange 1b
#
#
rsd0, u1
@IExchange 3b
#
phis
@IExchange 4
phiver
@IExchange 5
#
#
phihor
@IExchange 6
behaviour: seq { do ipr; do inorm; do itmax; do nx0; do ny0; do nz0; do dt; do omega; do tolrsd; do rsd0;
do u1; do rsdnm; do rsd1; do u1; do rsdnm; do errnm; do phis; do frc1; do phiver; do frc2;
do phihor; do frc3 }
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
unit
ipr comm
inorm comm
itmax comm
nx0 comm
ny0 comm
nz0 comm
dt comm
omega comm
tolrsd comm
rsd0 comm
u1 comm
rsdnm comm
ssor comm
errnm comm
phis comm
frc1 comm
phiver comm
frc2 comm
phihor comm
frc3 comm
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
;
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
assign
BCast
BCast
BCast
BCast
BCast
BCast
BCast
BCast
BCast
Exchange 3b
Exchange 3b
AllReduce
Exchange 1b
AllReduce
Exchange 4
AllReduce
Exchange 5
AllReduce
Exchange 6
AllReduce
[/ unify ipr comm.p[n][m]
inorm comm.p[n][m]
itmax comm.p[n][m]
dt comm.p[n][m]
omega comm.p[n][m]
tolrsd comm.p[n][m]
nx0 comm.p[n][m]
ny0 comm.p[n][m]
nz0 comm.p[n][m]
rsd0 comm.g1[n][m]
u1 comm.g1[n][m]
#
#
#
#
#
#
#
#
#
#
#
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
<
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
xdiv
* ydiv >
* ydiv >
* ydiv >
* ydiv >
* ydiv >
* ydiv >
* ydiv >
* ydiv >
* ydiv >
, ydiv >
, ydiv >
* ydiv , mpi double,
, ydiv , itmax, nz>
* ydiv , mpi double,
, ydiv >
* ydiv , mpi double,
, ydiv >
* ydiv , mpi double,
, ydiv >
* ydiv , mpi double,
ipr ,
inorm ,
itmax ,
dt ,
omega ,
tolrsd,
nx0 ,
ny0 ,
nz0 ,
rsd0 ,
u1 ,
73
mpi
mpi
mpi
mpi
mpi
to
to
to
to
to
to
to
to
to
to
to
sum>
to
to
sum>
to
to
sum>
to
to
sum>
to
to
sum>
to
ipr comm
inorm comm
itmax comm
nx0 comm
ny0 comm
nz0 comm
dt comm
omega comm
tolrsd comm
rsd0 comm
u1 comm
rsdnm comm
ssor comm
errnm comm
phis comm
frc1 comm
phiver comm
frc2 comm
phihor comm
frc3 comm
rsdnm comm.p[n][m]
ssor comm.bigLoop[n][m]
errnm comm.p[n][m]
phis comm.h0[n][m]
frc1 comm.p[n][m]
phiver comm.h1[n][m]
frc2 comm.p[n][m]
phihor comm.h2[n][m]
frc3 comm.p[n][m]
#
#
#
#
#
#
#
#
#
rsdnm ,
rsd1 ,
errnm ,
phis ,
frc1 ,
phiver,
frc2 ,
phihor,
frc3 to lu[n][m] # ILU
assign LU FM(PARAMETERS,ipr,inorm,itmax,nx0,ny0,nz0,dt,omega,tolrsd,rsdnm,errnm,
frc1,frc2,frc3,rsd1,rsd0,u1,phis,phiver,phivor)
→ (ipr,inorm,itmax,nx0,ny0,nz0,dt,omega,tolrsd,rsdnm,errnm,
frc1,frc2,frc3,rsd1,rsd0,u1,phis,phiver,phivor)
to lu[n][m] # ipr # inorm # itmax # nx0 # ny0 # nz0 # dt # omega # tolrsd # rsdnm # errnm
# frc1 # frc2 # frc3 # rsd1 # rsd0 # u1 # phis # phiver # phivor /]
74
| 6cs.PL
|
arXiv:1703.10088v1 [math.GR] 29 Mar 2017
Profinite semigroups
Revekka Kyriakoglou, Dominique Perrin
Université Paris-Est, LIGM
February 12, 2018
Abstract
We present a survey of results on profinite semigroups and their link
with symbolic dynamics. We develop a series of results, mostly due to
Almeida and Costa and we also include some original results on the
Schützenberger groups associated to a uniformly recurrent set.
Contents
1 Introduction
2
2 p-adic numbers
4
3 The Fibonacci morphism
5
4 Topological spaces, groups and
4.1 Topological spaces . . . . . .
4.2 Topological semigroups . . . .
4.3 Topological groups . . . . . .
semigroups
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . .
5 Profinite semigroups
5.1 Projective limits . . . . . . . . . . . . .
5.2 Profinite semigroups . . . . . . . . . . .
5.3 Profinite groups . . . . . . . . . . . . . .
5.4 Endomorphisms of profinite semigroups
5.5 The ω operator . . . . . . . . . . . . . .
5.6 The free profinite monoid . . . . . . . .
5.7 The free profinite group . . . . . . . . .
5.8 Recognizable sets . . . . . . . . . . . . .
5.9 The natural metric . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
6
8
8
9
9
11
12
13
13
14
15
16
17
6 Profinite codes
18
6.1 Finite codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
6.2 Codes and submonoids . . . . . . . . . . . . . . . . . . . . . . . . 19
1
7 Uniformly recurrent sets
20
7.1 Uniformly recurrent pseudowords . . . . . . . . . . . . . . . . . . 20
7.2 The J -class J(F ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
7.3 Fixed points of substitutions . . . . . . . . . . . . . . . . . . . . 24
8 Sturmian sets and tree sets
26
8.1 Sturmian sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
8.2 Tree sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
9 Return words
29
9.1 Left and right return words . . . . . . . . . . . . . . . . . . . . . 29
9.2 Limit return sets . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
10 Schützenberger groups
33
10.1 Groups of tree sets . . . . . . . . . . . . . . . . . . . . . . . . . . 34
10.2 Groups of fixed points of morphisms . . . . . . . . . . . . . . . . 35
10.3 Groups of bifix codes . . . . . . . . . . . . . . . . . . . . . . . . . 37
1
Introduction
The theory of profinite groups originates in the theory of infinite Galois groups
and of p-adic analysis (see [33]). The corresponding theory for semigroups was
considerably developed by Jorge Almeida (see [2] for an introduction). The
initial motivation has been the theory of varieties of languages and semigroups,
put in correspondance by Eilenberg’s theorem. Later, Almeida has initiated the
study of the connexion between free profinite semigroups and symbolic dynamics
(see [1]). He has shown in [7] that minimal subshifts correspond to maximal
J -classes of the free profinite monoid. Moreover, the Schützenberger group of
such a J -class is a dynamical invariant of the subshift [4]. Finally, it is shown
in [5] that if a minimal system satisfies the tree condition (as defined in [11]),
the corresponding group is a free profinite group.
In these notes, we give a gentle introduction to the notions used in profinite
algebra and develop the link with minimal sets.
Our motivation to explore the profinite world is the following. We are interested in the situation where we fix a uniformly recurrent set F on an alphabet
A (in general a non rational set, like the set of factors of the Fibonacci word).
We want to study sets of the form F ∩ L where L is a rational set on A, that
is the inverse image through a morphism ϕ : A∗ → M of a subset of a finite
monoid M . The aim is thus to develop a theory of automata observing through
the filter of a non rational set.
We are particularly interested in sets L of the form
1. L = wA∗ w for some word w ∈ F (linked to the complete return words to
w)
2. L = h−1 (H) where h : A∗ → G is a morphism onto a finite group G and
H is a subgroup of G.
2
This question has been studied successfully in a number of cases starting with
a Sturmian set F in [9] and progessively generalizing to sets called tree sets
in [11].
The framework of profinite semigroups allows one to work simultaneously
with all rational sets L. This is handled, as we shall see, both through the
definition of an inverse limit and through a topology on words. This topology
is defined by introducing a distance on words: two words are close if one needs
a morphism ϕ on a large monoid M to distiguish them. For example, for any
word x, the powers xn! of x will not be distinguished by a monoid with less than
n elements. Thus one can consider a limit, called a pseudoword, and denoted
xω which has the same image by all morphisms ϕ onto a finite monoid M .
The moment of inspiration was the derivation by Ameida and Costa [5] of a
result on profinite groups (the Schützenberger group of a tree set is free) which
uses the main result of [11]. This result gives a global version of properties like
the Finite Index Basis Property, as defined in [12]. One of the goals of this
paper is to develop this connexion.
We begin with two motivating examples: the first one concerns p-adic and
profinite numbers (Section 2) and the second one the Fibonacci morphism (Section 3).
We give in Section 5 an introduction to the basic notions concerning profinite
semigroups. We have chosen a simplified presentation which uses the class of all
semigroups instead of working inside a pseudovariety of semigroups. It simplifies
the statements but the proofs work essentially in the same way.
In Section 6 we describe results concerning codes in profinite semigroups.
The main result, from [29], is that any finite code is a profinite code (Theorem 6.1).
In Section 7 we introduce uniformly recurrent pseudowords. We prove a
result of Almeida (Theorem 7.2) showing that uniformly recurrent pseudowords
can be characterized in algebraic terms, as the J -maximal elements of the free
profinite monoid.
In Section 8, we recall some basic properties of Sturmian sets and their
generalization, the tree sets, introduced in [11].
In Section 9, we prove several results due to Costa and Almeida concening
the presentation of the Schützenberger group of a uniformly recurrent set.
In Section 10, we prove a new result (Thorem 10.18) concerning the Schützenberger
groups of uniformly recurrent sets.
Acknowledgements We would like to thank Jorge Almeida and Alfredo
Costa for their help in the preparation of this manuscript. It was written in
connexion with a workshop held in Marne la Vallée on january 20-22 2016
and gathering around them Marie-Pierre Béal, Valérie Berthé, Francesco Dolce,
Pavel Heller, Julien Leroy, Jean-Eric Pin and the autors.
3
2
p-adic numbers
We begin with a motivating example (see [24] for an introduction to this subject). Let p be a prime number and let Zp denote the ring of p-adic integers,
namely the completion of Z under the p-adic metric. This metric is defined by
the norm
(
p− ordp (x) if x 6= 0
|x|p =
0
otherwise
where ord(x)p (x) is the largest n such that pn divides x. For the topology
defined by this metric, Zp is compact.
Any element γ ∈ Zp has a unique p-adic expansion
γ = c0 + c1 p + c2 p2 + . . . = (. . . c3 c2 c1 c0 )p
Note that infinite expansions may represent ordinary integers. For example
(. . . 111)2 = −1.
We can express the expansion of the elements in Zp as
Zp
=
lim Z/pn Z
=
{(an )n≥1 ∈
Y
Z/pn Z | for all n, an+1 ≡ an mod pn }
n≥1
This expresses the ring Zp as a projective limit of the rings Z/pn Z. The direct
product of all rings Zp
Y
Ẑ =
Zp
p
over all prime numbers p is the ring of profinite integers. Its topolology is that
induced by the product. As a product of compact spaces, Ẑ is itself compact.
It does not depend on a particular number p and shares with all rings Zp the
property of being a compact topological space. Thus it is a compact covering
of all rings Zp .
One may define it equivalently as the projective limit of all cyclic groups
Y
Ẑ =
Z/nZ
n≥1
=
{(an )n≥1 ∈
Y
Z/nZ | for all n|m, am ≡ an mod n}
n≥1
or as the projective limit of the cyclic groups Z/n!Z
Y
Ẑ =
Z/n!Z
n≥1
=
{(an )n≥1 ∈
Y
Z/n!Z | for all n, an+1 ≡ an mod n!}
n≥1
4
The last representation corresponds to an expansion of the form
γ = c1 + c2 2! + c3 3! + . . . = (. . . c3 c2 c1 )!
with digits 0 ≤ ci ≤ i. This expansion forms the factorial number system
(see [23]).
Note that this time
−1 = (. . . 321)!
which holds because 1 + 2.2! + . . . + n.n! = (n + 1)! − 1, as one may verify by
induction on n. The profinite topology on Ẑ can also be defined directly by
the norm |x|! = 2−r(x) where r(x) is the largest n such that n! divides x. A
sequence converges with respect to this topology if the expansions converge in
the usual sense, that is the number of equal digits, starting from the right, tends
to infinity.
It is possible to define profinite Fibonacci numbers (see [26]). Indeed, Fibonacci numbers are defined by F0 = 0, F1 = 1 and Fn = Fn−1 + Fn−2 . The
definition can be extended to negative n by Fn = (−1)n−1 F−n . Then one may
extend the function n 7→ Fn to a continuous function from Ẑ into itself. We
note that since n! tends to 0 in Ẑ, the sequence Fn! tends to F0 = 0. Similarly
Fn!+2 and Fn!+1 tend to 1. We come back to this in the next section.
3
The Fibonacci morphism
In the last example, we have seen how a linear recurrent sequence (the Fibonacci
numbers) can interestingly be extended to a topological limit. As well known,
Fibonacci numbers are the lengths of the Fibonacci words defined inductively
by w0 = b, w1 = a and satisfying the recurrence relation wn+1 = wn + wn−1
for n ≥ 1. Then |wn | = Fn+1 . In the same way as one may embed the ring of
ordinary integers into the compact ring of profinite integers, we will see that one
may extend the Fibonacci sequence to a converging sequence of pseudowords
whose lengths are profinite integers.
The Fibonacci morphism is the morphism ϕ : A∗ → A∗ with A = {a, b}
defined by ϕ(a) = ab and ϕ(b) = a. The sequence of ϕn (a)
ϕ(a)
=
ab
ϕ2 (a)
ϕ3 (a)
=
=
aba
abaab
ϕ4 (a)
= abaababa
···
is the Fibonacci sequence of words of length equal to the Fibonacci numbers.
One has Fn = |ϕn−2 (a)| for n ≥ 2 (and for n ∈ Z with appropriate extensions).
The Fibonacci sequence of words converges in the space AN to the Fibonacci
infinite word
x = abaababa · · ·
5
It is a fixed-point of ϕ in the sense that ϕ(x) = x. From another point of
view, this sequence is not convergent. Indeed, the terms of the sequence end
alternately with a or b and thus can be distinguished by a morphism from A∗
into a monoid with 3 elements. A sequence converging in this stronger sense is
ϕn! (a)
ϕ(a)
=
ab
2
ϕ (a)
ϕ6 (a)
=
=
aba
abaababaabaababaababa
ϕ24 (a)
=
abaababa · · · abaababa
···
The limit in the sense of profinite topology, to be defined below, is a pseudoword
denoted ϕω (a) which begins by the Fibonacci infinite word (ϕn (a))n≥0 and ends
with the left infinite word (ϕ2n (a))n≥0 .
Its length in Ẑ is the limit of the Fibonacci numbers Fn!+2 which is F2 = 1.
We shall come back to this interesting property in Example 7.15.
4
Topological spaces, groups and semigroups
In this section, we recall the basic definitions of topology and the notion of
topological semigroup.
4.1
Topological spaces
We begin with an introduction to the basic notions of topology (see [34] for
example). A topological space is a set S with a family F of subsets such that
(i) it contains ∅ and S,
(ii) it is closed under union,
(iii) it is closed under finite intersection
The elements of F are called open sets. The complement of an open set is called
a closed set. A clopen set is both open and closed.
For any set S, the discrete topology is the topology for which all subsets are
open.
For any subset X of a topological set S, the topology induced on X by the
topology of S corresponds to the family of open sets X ∩ Y for Y open in S.
A limit point of a subset X of a topological space S is a point x such that any
open set containing x contains a point of X distinct of x. As a particlular case,
it is said to be an accumulation point if every open set containing x contains an
infinite number of points of X.
The closure of a subset X of S, denoted X̄ is the union of X and all its limit
points. It is also the smallest closed set containing X. A set is closed if and
only if X = X̄. The set X is dense in S is X̄ = S.
6
The notion of limit point can also be defined for a sequence. A point x
is said to be a limit point of the sequence (xn )n≥0 if any open set containing
x contains all but a finite number of terms of the sequence. It is said to be
an accumulation point of the sequence if every open set containing x contains
an infinite number of elements of the sequence. A limit point of the sequence
(xn )n≥0 is a limit point of the set {xn | n ≥ 0}.
A map ϕ : X → Y between topological spaces X, Y is continuous if for any
open set U ⊂ Y , the set ϕ−1 (U ) is open in X.
A basis of the family of open sets is a family B of sets such that any open
set is a union of elements of B.
Given a family of topological spaces
Xi indexed by a set I, the product
Q
topology on the direct product X = i∈I Xi is defined as the coarsest topology
such that the projections πi : X → Xi are continuous. A basis
Q of the family of
open sets is the family of open boxes, that is sets of the form i∈I Ui where the
Ui are open sets such that Ui 6= Xi only for a finite number of indices i.
Example 4.1 The set AN of infinite words over A is a topological space for the
discrete topology on A. It is not a semigroup (a big motivation for introducing
pseudowords!). However, there is a well defined product of finite words and
infinite words. The open sets are the sets of the form XAN for X ⊂ A∗ .
Metric spaces form a vast family of topological spaces. A metric space is
a space S with a function d : S × S → R, called a distance, such that for all
x, y, z ∈ S,
(i) d(x, y) = 0 if and only if x = y,
(ii) d(x, y) = d(y, x)
(iii) d(x, z) ≤ d(x, y) + d(y, z) (triangle inequality).
Any metric space can be considered as a topological space, considering as open
sets the unions of open balls Bε (x) = {y ∈ S | d(x, y) < ε} for x ∈ S and ε ≥ 0.
For example, the set Rn is a metric space for the Euclidean distance.
A topological space is a Hausdorff space if any two distinct points belong to
disjoint open sets. A metric space is a Hausdorff space. In a Hausdorff space,
every limit point of a set is an accumulation point.
A topological space is compact if it is a Hausdorff space and if from any
family of open sets whose union is S, one may extract a finite subfamily with
the same property.
A closed subset of a compact space is compact and, by Tychononoff’s theorem, any product of compact spaces is compact.
One may verify that in a compact space every infinite set has an accumulation point (the converse is true in metric spaces). Indeed, if X is an infinite
subset of a compact space S without accumulation point, there is for every
x ∈ S an open set containing x which contains only a finite number of elements
of X. For a finite set F ⊂ X, denote by OF the union of the Ox such that
Ox ∩ X = F . Then the sets OF form a family of open sets whose union is S
7
every finite subfamily of which intersects at most finitely many points in X.
Thus no finite subfamily may cover S, a contradiction.
The clopen sets in a product of compact spaces Si are the finite unions of
clopen boxes, that is the sets of the form Πi∈I Ki where each Ki is clopen in Si
and Ki = Si for all but a finite number of indices i.
4.2
Topological semigroups
A semigroup is a set with an associative operation. A monoid is a semigroup
with a neutral element.
A topological semigroup is a semigroup S endowed with a topology such that
the semigroup operation S × S → S is continuous. A topological monoid is a
topological semigroup with identity.
Example 4.2 A finite semigroup can always be viewed as a topological semigroup under the discrete topology.
Example 4.3 As a less trivial example, the set R of nonnegative real numbers
is a topological semigroup for the addition and the interval [0, 1] is a topological
semigroup for the multiplication.
A compact monoid is a topological monoid which is compact (as a topogical
space). Note that we assume a compact space to satisfy Hausdorff separation
axiom (any two distinct points belong to disjoint open sets). Note also the
following elementary property of compact monoids. Recall that, is a monoid
M , u ∈ M is a factor of v ∈ M if v ∈ M uM .
4.3
Topological groups
A topological group is a group with a topology such that the multiplication and
taking the inverse are continuous operations. It is in particular a topological
semigroup.
Example 4.4 The set R of real numbers with the usual topology is a topological
group under addition.
Any closed subgroup is a topological group for the induced topology. Moreover, since multiplication is continuous, the cosets Hg of an open (resp. closed)
subgroup are open (resp. closed).
Every open subgroup H of a topological group G is also closed since its
complement is the union of all cosets Hg for g ∈ G \ H which are open.
The following is [33, Lemma 2.1.2].
Proposition 4.5 In a compact group, a subgroup is open if and only if it is
closed and of finite index.
8
Proof. Assume that H is an open subgroup of G. We have already seen that
H is also closed. The union of the cosets of H form a covering by open sets.
Since G is compact, there is a finite subfamily covering G and thus H has finite
index.
Conversely, if H is a closed subgroup of finite index, then the complement
of H is the union of the cosets Hg for g ∈
/ H and thus H is open.
5
Profinite semigroups
In this section, we introduce the notions of profinite semigroup and of profinite
group. We begin with the notion of projective limit.
5.1
Projective limits
We want to define profinite semigroups as some kind of limit of finite semigroups
in such a way that properties true in all finite semigroups will remain true in
profinite semigroups. For this we need the notion of projective limit.
A projective system (or inverse system) of semigroups is given by
(i) a directed set I, that is a poset in which any two elements have a common
upper bound,
(ii) for each i ∈ I, a topological semigroup Si ,
(iii) for each pair i, j ∈ I with i ≥ j, a connecting morphism ψi,j : Si → Sj
such that ψi,i is the identity on Si and for i ≥ j ≥ k, ψi,k = ψi,j ◦ ψj,k .
Example 5.1 Let I be the set of natural integers ordered by divisibility: n ≥ m
if m|n. The family of cyclic groups (Z/nZ)n∈I forms a projective system for the
morphisms ψn,m defined by ψn,m (x) = x mod m.
In the same way, the family of cyclic groups (Z/n!Z)n≥0 indexed by the set
I of natural integers with the natural order is a projective system.
The projective limit (or inverse limit ) of this projective system is a topological semigroup S together with morphisms Φi : S → Si such that for all i, j ∈ I
with i ≥ j, ψi,j ◦ Φi = Φj , and for any topological semigroup T and morphisms
Ψi : T → Si such that for all i, j ∈ I with i ≥ j, ψi,j ◦ Ψi = Ψj , there exists a
morphism θ : T → S such that Φi ◦ θ = Ψi for all i ∈ I.
The uniqueness of the projective limit can be verified (“as a standard diagram chasing exercise” [2]). The
Q existence can be proved by considering the
subsemigroup S of the product i∈I Si consisting of all (si )i∈I such that, for
all i, j ∈ I with i ≥ j,
ψi,j (si ) = sj
endowed with the product topology. The maps Φi : S → Si are the projections,
that is, if s = (si )i∈I , then Φi (s) = si .
9
T
θ
Ψj
Ψi
S
Φj
Φi
ψi,j
Si
Sj
Figure 1: The projective limit.
One defines in the same way a projective system of monoids or groups and
a projective limit of monoids or groups. For a projective system of monoids,
one has to take all morphisms as monoid morphisms and similarly for groups
(actually a monoid morphism between groups is already a group morphism).
Example 5.2 The projective limit of the family of cyclic groups (Example 5.1)
is the group of profinite integers.
A variant of this construction allows to specify a fixed generating set for all
semigroups. An A-generated topological semigroup is a topological semigroup
S together with a mapping ϕ : A → S whose image generates a subsemigroup
dense in S. A morphism between A-generated topological semigroups ϕ : A → S
and ψ : A → T is a continuous morphism θ : S → T such that θ ◦ ϕ = ψ. We
denote θ : ϕ → ψ such a morphism.
A
ϕ
S
ψ
θ
T
Figure 2: A morphism of A-generated semigroups
A projective system in this category of objects is given by a directed set I
and for each i ∈ I, an A-generated topological semigroup ϕi : A → Si , and for
each pair i, j ∈ I with i ≥ j, a connecting morphism ψi,j : ϕi → ϕj such that
ψi,i is the identity on Si and for i ≥ j ≥ k, ψi,k = ψi,j ◦ ψj,k .
The projective limit of this projective system is an A-generated topological
semigroup Φ : A → S together with morphisms Φi : Φ → ϕi such that for
all i, j ∈ I with i ≥ j, ψi,j ◦ Φi = Φj , and for any A-generated topological
semigroup Ψ : A → T and morphisms Ψi : Ψ → ϕi such that for all i, j ∈ I
with i ≥ j, ψi,j ◦ Ψi = Ψj , there is a morphism θ : Ψ → Φ such that Φi ◦ θ = Ψi
for all i ∈ I.
10
A
Ψ
Ψ
T
ϕi
θ
θ
Ψj
Ψi
ϕi
Ψj
Ψi
S
Φ
Φj
Φi
ϕj
ψi,j
Φj
Φi
ϕj
Si
ψi,j
Sj
Figure 3: The projective limit of a family of A-generated semigroups.
5.2
Profinite semigroups
A profinite semigroup is a projective limit of a projective system of finite semigroups.
Example 5.3 The ring of profinite integers is a profinite group.
A profinite semigroup is compact. Indeed, let S be the projective limit of a
family Si of finite semigroups. Then S is a closed submonoid of a direct product
of the finite and thus compact semigroups Si and thus is compact.
A topological space is connected if it is not the union of two disjoint open sets.
A subset of a topological space is connected if it is connected as a subspace, that
is it cannot be covered by the union of two disjoint open sets. Every topological
space decomposes as a union of disjoint connected subsets, called its conneted
components.
A topological space is
(i) totally disconnected if its connected components are singletons,
(iii) zero-dimensional if it admits a basis consisting of clopen sets.
The term zero-dimensional is by reference to a notion of dimension in topological
spaces (the Lebesgue dimension). The following result, from [2] gives a possible
direct definition of profinite semigroup without using projective limits. A topological semigroup is residually finite if for any u, v ∈ S there exists a continuous
morphism ϕ : S → M into a finite semigroup M such that ϕ(u) 6= ϕ(v).
Theorem 5.4 The following conditions are equivalent for a compact semigroup
S.
(i) S is profinite,
(ii) S is residually finite as a topological semigroup,
(iii) S is a closed subsemigroup of a direct product of finite semigroups,
11
(iv) S is totally disconnected,
(v) S is zero-dimensional.
The explicit construction of the projective limit shows that (i)⇒ (ii) and (ii)⇒
(iii) results from the definitions. For (iii)⇒ (i), see [2]. Since a product of
totally disconnected spaces is totally disconnected, we have (iii)⇒ (iv). The
equivalence (iv)⇔ (v) holds for any compact space. Finally, the implication
(v)⇒ (ii) results from Hunter’s Lemma (see [2]).
Corollary 5.5 The class of profinite semigroups is closed under taking closed
subsemigroups and direct product.
The following is [2, Proposition 3.5]. A subset K of a semigroup S is recognized by a morphism ϕ : S → M if K = ϕ−1 ϕ(K).
Proposition 5.6 Let S be a profinite semigroup. A subset K ⊂ S is clopen if
and only if it is recognized by a continuous morphism ϕ : S → T into a finite
semigroup T .
Proof. The condition is sufficient since the set K is the inverse image under a
continuous function of a clopen set. Conversely assume that K is clopen. Since
S
Q is profinite, it is by Theorem5.4 a closed subsemigroup of a direct product
i∈I Si of finite semigroups Si . Since K is clopen, it is a finite union of clopen
boxes, that is a finite union of sets Πi∈I Ki with Ki = Si for all but a finite
number of indices i. Thus there is a finite set F ⊂ I such that being in K only
depends on the coordinates in F . Then the projection ϕ : S → Πi∈F Si is a
continuous morphism into a finite semigroup which recognizes K. Indeed, for
s ∈ ϕ−1 ϕ(K), there is t ∈ K such that ϕ(s) = ϕ(t). Since s and t agree on their
F -coordinates, we have s ∈ K.
5.3
Profinite groups
A profinite group is a projective limit of a projective system of finite groups.
A topological group which is a profinite semigroup is actually a profinite
group. Indeed, let S be the projective limit of a family Si of finite semigroups.
We may assume that the morphisms Φi : S → Si are surjective. If S is a group,
since the image of a group by a semigroup morphism is a group, each Si is a
finite group and thus S is a profinite group.
Any profinite group is a compact group. Indeed, it is a closed subgroup of a
direct product of finite and thus compact groups.
A simple example of a compact group which is not profinite is the multiplicative group [0, 1], which has no nontrivial image which is a finite group.
A closed subgroup of a profinite group is profinite. Indeed, let H be a closed
subgroup of a profinite group G. Then H is a closed subsemigroup of G and
thus it is a profinite semigroup by Corollary 5.5, whence a profinite group.
12
A group G is Hopfian if every endomorphism of G which is onto is an isomorphism. The following is [33, Proposition 2.5.2]. It is an analogue property
for profinite groups.
Proposition 5.7 Let G be a finitely generated profinite group. and let ϕ : G →
G be a continuous surjective morphism. Then ϕ is an isomorphism.
Proof. We show that Ker(ϕ) is contained in any subgroup of finite index of G.
Since G is profinite, it will imply that Ker(ϕ) = {1}.
For each finite group F , since G is finitely generated, there is a finite number
of morphisms from G into F . Thus there is only a finite number of morphisms
from G into a finite group of order n and thus also a finite number of normal
subgroups of index n of G.
Let Un be the family of normal subgroups of G of index n. Let Φ : Un → Un
be defined by Φ(U ) = ϕ−1 (U ). Then Φ is an injective map from a finite set into
itself and thus it is a bijection.
Let U be a normal subgroup of index n of G. Then, since Φ is surjective, we
have U = ϕ−1 (V ) for som V ∈ Un . This implies that Ker(ϕ) ⊂ U , which was
to be proved.
5.4
Endomorphisms of profinite semigroups
For a topological semigroup, we denote by End(S) the monoid of all its continuous endomorphisms. We consider End(S) as a topological monoid for the
pointwise convergence.
The following result is [2, Theorem 4.14].
Theorem 5.8 Let S be a finitely generated profinite semigroup. Then End(S)
is a profinite monoid.
5.5
The ω operator
Recall that an idempotent in a semigroup S is an element e ∈ S such that
e2 = e.
In a finite semigroup S, the semigroup generated by s ∈ S can be represented
as in Figure 4 (the frying pan) with the index i and the period p such that
si+p = si . It contains a unique idempotent, which is of the form snp with n
such that np ≥ i. Thus, Z/pZ is a maximal subroup of the semigroup generated
by s, which coincides with its minimal ideal.
In a compact semigroup S, just as in a finite semigroup, the closure of the
semigroup generated by an element s ∈ S contains a unique idempotent, denoted
sω . If S is profinite, it is the limit of the sequence sn! .
We note that in any profinite group, one has xω = 1 since the neutral element
is the only idempotent of G.
13
s
si+1
s
s
s
s2
s
s
si
···
s
si+p−1
s
Figure 4: The semigroup generated by s.
In general, the index can be a finite integer i, in which case we have sω+i =
s . It can also be infinite and equal to ω, as in the semigroup N̂. Independently,
the period can also be finite or infinite.
We also note that for any endomorphism ϕ of a profinite monoid S, the
endomorphism ϕω is a well defined endomorphism of S.
ω
5.6
The free profinite monoid
Consider the projective system formed by representatives of isomorphism classes
of all A-generated finite monoids (the finite monoids are considered as topological monoids for the discrete topology). For ϕ : A → M and ψ : A → N , one
has ϕ ≥ ψ if there is a morphism µ : M → N such that µ ◦ ϕ = ψ. Note that
ϕ, ψ, µ have to be surjective.
c∗ is the projective
The free profinite monoid on a finite alphabet A, denoted A
limit of this family. It has the following universal property (see Figure 5).
c∗ is such that for any map
Proposition 5.9 The natural mapping ι : A → A
ϕ : A → M into a profinite monoid there exists a unique continuous morphism
c∗ → M such that ϕ̂ ◦ ι = ϕ.
ϕ̂ : A
A
ϕ
ι
c∗
A
ϕ̂
M
c∗ .
Figure 5: The universal property of A
c∗ are called pseudowords and the elements of A
c∗ \ A∗ are
The elements of A
called infinite pseudowords.
The free profinite monoid on one generator is commutative. Its image by
the map an 7→ n is the monoid of profinite natural numbers , denoted N̂.
14
c∗ is discrete. Indeed,
The topology induced A∗ by the profinite topology on A
∗
∗
if u ∈ A is a word of length n, the quotient of A by the ideal formed by the
words of length greater than n is a finite monoid. The congruence class of u for
the corresponding quotient is reduced to u.
c∗ is a profinite natural number. The
The length |x| of a pseudoword x ∈ A
c∗ → |x| ∈ N̂ is the continuous morphism λ such that λ(a) = 1 for
map x ∈ A
every a ∈ A.
Example 5.10 The length of xy ω is |x| + ω.
5.7
The free profinite group
Likewise, the free profinite group , denoted F\
G(A) is the projective limit of
the projective system formed by the isomorphism classes of A-generated finite
groups.
The topology on the free group F G(A) induced by the topology of F\
G(A) is
not discrete. Indeed, for any x in F G(A), the sequence xn! tends to 1. Thus A∗
c∗ onto F\
is dense in F\
G(A) and there is an onto homomorphism from A
G(A).
The topology induced on F G(A) by the topology of F\
G(A) is also called the
Hall topology. It has been indeed introduced by M. Hall in [20]. Note that, since
A∗ is embedded in F G(A), we actually have two topologies on A∗ respectively
c∗ and F\
induced by the topologies of A
G(A). To distinguish them, the first one
is called the pro-M topology and the second one the pro-G topology. The first
one is strictly stronger than the second one. The pro-G topology on A∗ was
introduced by Reutenauer in [32].
The image of the free profinite group on one generator a by the map an 7→ n
is the group Ẑ of profinite integers (see Section 2). The length |x| of an element
x of F\
G(A) is a profinite integer. The map x ∈ F\
G(A) → |x| ∈ Ẑ is the
unique continuous morphism such that |a| = 1 for every a ∈ A. In particular
|a−1 | = −1.
The following result is from [20, p. 131]. The property is not true for non
finitely generated subgroups. The classical example is the commutator subgroup
F ′ of a finitely generated free group F , which is known to be a free group of
countable infinite rank. In the topology induced on F ′ by the profinite topology
on F , there are only countably many open subgroups while the number of open
subgroups in the profinite topology of F ′ is uncountable.
Proposition 5.11 Any finitely generated subgroup H of F G(A) is closed for
the topology induced on F G(A) by the pro-G topology.
The proof relies on the following result (see [19, Theorem 5.1] or [28, Proposition
3.10]). Recall that a free factor of a group G is a subgroup H such that G is
a free product of H and a subgroup K of G. When G is a free group, this is
equivalent to the following property: for some basis X of H there is a subset Y
of G such that X ∪ Y is a basis of G.
15
Theorem 5.12 (Hall) For any finitely generated subgroup H of F G(A) and
any x ∈ F G(A) \ H, there is a subgroup of finite index K such that H is a free
factor of K and x ∈
/K .
To deduce Proposition 5.11 from Hall’s Theorem, consider a sequence (xn ) of
elements of H converging to some x ∈ F G(A). Suppose that x ∈
/ H. By
Theorem 5.12, there is a subgroup K of finite index in F G(A) containing H
such that x ∈
/ K. Thus (xn ) cannot converge to x.
It follows from Theorem 5.12 that one has the following result [14, Corollary
2.2]
Corollary 5.13 Any injective morphism ϕ : F G(B) → F G(A) between finitely
generated free groups extends to an injective continuous morphism ϕ̂ : F\
G(B) →
F\
G(A)
Proof. Let H = ϕ(F G(B)). Then ϕ is an isomorphism between F G(B) and H
b the completion of H
which extends to an isomorphism between F\
G(B) and H,
with respect to the profinite metric. But, by Proposition 5.11, the subgroup H
b is the same as the closure H̄ in F\
is closed in F G(A) and thus H
G(A), which
\
\
shows that ϕ
b is an injective morphism from F G(B) into F G(A).
5.8
Recognizable sets
A subset X of a monoid M is recognizable if it is recognized by a morphism into
a finite monoid, that is, if there is a morphism ϕ : M → N into a finite monoid
N which recognizes X. We also say that X is recognized by ϕ.
Proposition 5.14 The following conditions are equivalent for a set X ⊂ A∗ .
(i) X is recognizable.
c∗ is open and X = X̄ ∩ A∗ .
(ii) the closure X̄ of X in A
c∗ .
(ii) X = K ∩ A∗ for some clopen set K ⊂ A
Proof. Assume that X is recognized by a morphism ϕ : A∗ → S from A∗ into a
c∗ , there is a unique continuous
finite monoid S. By the universal property of A
−1
morphism ϕ̂ extending ϕ. Then X = ϕ̂ ϕ(X) is open and satisfies X = X̄ ∩A∗ .
Thus (i)⇒ (ii). The implication (ii)⇒ (iii) is trivial. Finally, assume that (iii)
c∗ → S into
holds. By Proposition 5.6 there exists a continuous morphism ψ : A
a finite monoid S which recognizes K. Let ϕ be the restriction of ψ to A∗ . Then
X = A∗ ∩ K = A∗ ∩ ψ −1 ψ(K) and so X is recognizable.
c∗ wA
c∗ = A∗ wA∗ are clopen sets and so
As an example, the sets of the form A
c∗ = wA∗ . This shows that a pseudoword has a well-defined set
are the sets wA
of finite factors and a well-defined prefix of every finite length.
16
The analogue of Proposition 5.6 for the pro-G topology is also true (it actually holds in the pro-V topology for any pseudovariety V ). Thus a set X is
recognizable by a morphim on a finite group if and only if X = K ∩ A∗ for some
clopen set K ⊂ F\
G(A). In condition (ii), one has to add that X̄ ∩ A∗ = X, a
condition always satisfied for the closure with respect to the pro-M topology.
5.9
The natural metric
The natural metric on a profinite monoid M is defined by
(
2−r(u,v) if u 6= v
d(u, v) =
0
otherwise
where r(u, v) is the minimal cardinality of a monoid N for which there is a
continuous morphism ϕ : M → N such that ϕ(u) 6= ϕ(v).
It is actually an ultrametric since it satisfies the condition
d(u, w) ≤ min(d(u, v), d(v, w))
stronger than the triangle inequality.
Proposition 5.15 For a finitely generated profinite semigroup S, the topology
is induced by the natural metric.
Proof. Denote Bε (u) = {v ∈ S | d(u, v) < ε}. Let K be a clopen set in S.
By Proposition 5.6 there is a continuous morphism ϕ : S → T into a finite
semigroup T which recognizes K. Let n = Card(T ) and set ε = 2−n . For
s ∈ S and t ∈ Bε (s), we have d(s, t) < ε and thus r(s, t) > n. It implies that
ϕ(s) = ϕ(t). Since ϕ recognizes K, we conclude that t ∈ K. Thus the ball
Bε (s) is contained in ϕ−1 (t) and thus K is a union of open balls. Thus, since
the clopen sets form a basis of the topology, any open set is a union of open
balls.
Conversely, consider the open ball B = B2−n (u). Since there is a finite
number of isomorphism types of semigroups with at most n elements, and since
S is finitely generated, there are finitely many kernels of continuous morphisms
into such semigroups and so their intersection is a clopen congruence on S.
It follows that there exists a continuous morphism ϕ : S → T into a finite
semigroup such that ϕ(u) = ϕ(v) if and only if r(u, v) > n. Hence B = ϕ−1 ϕ(B)
so that B is open.
This leads to an alternative definition of the free profinite monoid.
Theorem 5.16 For a finite set A, the completion of A∗ for the natural metric
c∗ .
is the free profinite monoid A
17
Presentations of profinite semigroups A congruence of a profinite semigroup is called admissible if its classes are closed and the quotient is profinite.
In other terms, admissible congruences are the kernels of continuous homomorphisms into profinite monoids.
c∗ , the profinite
Given a set X and a binary relation R on the monoid X
∗
c
semigroup hX | Ri is the quotient of X by the admissible congruence generated
by R. It is also said to have the presentation hX | Ri.
The same notion holds for groups instead of semigroups and use the notation
hX | RiS or hX | RiG to specify if the presentation is as a profinite semigroup
or as a profinite group.
The following is [4, Lemma 2.2].
Proposition 5.17 Let S be a profinite semigroup and let ϕ be an automorphism
c∗ onto S and let Φ be a
of S. Let π be a continuous homomorphism from A
∗
c
continuous endomorphism of A such that the diagram below commutes. Then
Φ
c∗
c∗
A
A
π
S
π
ϕ
S
S has the presentation hA | Ri with R = {(Φω (a), a) | a ∈ A}.
6
Profinite codes
We will expore the notion of a code in the free profinite monoid.
6.1
Finite codes
Let A, B be finite alphabets. Any morphism β : B ∗ → A∗ extends uniquely by
c∗ → A
c∗ . A finite set X ⊂ A
c∗ is called
continuity to a continuous morphism β̂ : B
a profinite code if the continuous extension β̂ of any morphism β : B ∗ → A∗
inducing a bijection from B onto X is injective.
The following statement is from [29].
Theorem 6.1 Any finite code X ⊂ A+ is a profinite code.
Proof. Let β : B ∗ → A∗ be a coding morphism for X. We have to show
c∗ of distinct elements, we have β̂(u) 6= β̂(v), that is,
that for any pair u, v ∈ B
c∗ → M into a finite monoid M such that
there is a continuous morphism α̂ : A
c∗ → N be a continuous morphism into a
α̂β̂(u) 6= α̂β̂(v). For this, let ψ : B
finite monoid N such that ψ(u) 6= ψ(v). Let P be the set of proper prefixes of
X and let T be the prefix transducer associated to β (see [10]). Let α be the
morphism from A∗ into the monoid of P × P -matrices with elements in N ∪ 0
18
defined as follows. For x ∈ A∗ and p, q ∈ P , we have
(
x|y
ψ(y) if there is a path p → q
α(x)p,q =
0
otherwise.
Then M = α(A∗ ) is a finite monoid and α extends to a continuous morphism
c∗ → M . Since, by [10, Proposition 4.3.2], the transducer T realizes the
α̂ : A
decoding function of X, we have αβ(y)1,1 = ψ(y) for any y ∈ B ∗ . By continuity,
c∗ . Then α̂ is such that α̂β̂(u) 6= α̂β̂(v).
we have α̂β̂(y)1,1 = ψ(y) for any y ∈ B
Indeed α̂β̂(u)1,1 = ψ(u) 6= ψ(v) = α̂β̂(v)1,1 .
Example 6.2 Let A = {a, b} and X = {a, ab, bb}. The set X is a suffix code.
It has infinite deciphering delay since abb · · · = a(bb)(bb) · · · = (ab)(bb)(bb) · · · .
Nonetheless, X is a profinite code in agreement with Theorem 6.1. Note that,
c∗ , the pseudowords a(bb)ω and ab(bb)ω are distinct (the first one is a limit
in A
of words of odd length and the second one of words of even length).
Theorem 6.1 shows that the closure of the submonoid generated by a finite code
is a free profinite monoid. This has been extended to rational codes in [6].
Actually, for any rational code X, the profinite submonoid generated by X is
free with basis the closure X̄ of X [6, Corollary 5.7]. Note that we have here a
free profinite monoid with infinite basis (see [6] for an extension of the notion
of free profinite monoid to infinite alphabets).
c∗ is
Example 6.3 Let A = {a, b} and X = a∗ b. Then the profinite monoid X
v
free with basis the uncoutable set X̄ = {a b | v ∈ N̂}.
6.2
Codes and submonoids
A submonoid N of a monoid M is called stable if for any u, v, w ∈ M , whenever
u, vw, uv, w ∈ N then v ∈ N . It is called right unitary if for every u, v ∈ A∗ ,
u, uv ∈ N implies v ∈ N . It is well known that a submonoid of A∗ is stable if
and only if it is generated by a code and it is unitary if and only if it is generated
by a prefix code.
The following statement extends these notions to pseudowords.
Proposition 6.4 Let N be a recognizable submonoid of A∗ . If N is stable (resp.
c∗ is a stable (resp. unitary) submonoid of A
c∗ .
unitary) its closure in A
c∗ . Assume that N is stable.
Proof. The set N̄ is clearly a submonoid of A
c∗ be such that uv, w, u, vw ∈ N̄ . Let (un ), (vn ) and (wn ) be
Let u, v, w ∈ A
sequences of words converging to u, v and w respectively with un , wn ∈ N .
c∗ and
By Proposition 5.14, the closure N̄ of the submonoid N is open in A
N = N̄ ∩ A∗ . Since N̄ is open, we have un vn , vn wn ∈ N̄ for large enough n
and thus also un vn , vn wn ∈ N . Since N is stable, this implies vn ∈ N for large
19
enough n, which implies v ∈ N̄ . Thus N̄ is stable. The proof when N is unitary
is similar.
When X is a prefix code, every word w can be written in a unique way w = xp
with x ∈ X ∗ and p ∈ A∗ \ XA∗ , a word without any prefix in X. When X is
a maximal prefix code, a word which has no prefix in X is a proper prefix of a
word in X and thus every word w can be written in a unique way w = xp with
x ∈ X ∗ and p a proper prefix of X. This extends to pseudowords as follows.
Proposition 6.5 Let X ⊂ A+ be a finite maximal prefix code and let P be
c∗ has a unique factorization
its set of proper prefixes. Any pseudoword w ∈ A
∗
w = xp with x ∈ X and p ∈ P .
Proof. Let (wn ) be a sequence of words converging to w. For each n, we have
wn = xn pn with xn ∈ X ∗ and pn ∈ P . Taking a subsequence, we may assume
that the sequences (xn ) and (pn ) converge to x ∈ X ∗ and p ∈ P̄ . Since X is
finite, P is finite and thus P̄ = P . This proves the existence of a factorization.
c∗ and p, p′ ∈ P .
To prove the uniqueness, consider xp = x′ p′ with x, x′ ∈ X
′
′
Then, assuming that p is longer than p , we have p = up and xu = x′ for some
u ∈ A∗ . Since X ∗ is unitary, we have u ∈ X ∗ and thus u = ε.
7
Uniformly recurrent sets
In this section, we study the closure in the free profinite monoid of a uniformly
recurrent set.
7.1
Uniformly recurrent pseudowords
Let A be a finite alphabet. A set of finite words on A is factorial if it contains
the alphabet A and all the factors of its elements.
A factorial set F is recurrent if for any x, z ∈ F there is some y ∈ F such
that xyz ∈ F .
Recall that an infinite factorial set F of finite words is said to be uniformly
recurrent or minimal if for any x ∈ F there is an integer n ≥ 1 such that x is
a factor of every word in F of length n.
A uniformly recurrent set is obviously recurrent.
Note that a uniformly recurrent set is actually minimal for inclusion among
the infinite factorial sets. Indeed, assume first that F is uniformly recurrent.
Let F ′ ⊂ F be an infinite factorial set. Let x ∈ F and let n be such that x is
factor of any word of F of length n. Since F ′ is factorial infinite, it contains a
word y of length n. Since x is a factor of y, it is in F ′ . Thus F is mimimal.
Conversely consider x ∈ F and let T be the set of words in F which do not
contain x as a factor. Then T is factorial. Since F is minimal among infinite
factorial sets, T is finite. Thus there is an n such that x is a factor of all words
of F of length n.
20
A factorial set F is periodic if it is the set of factors of a finite word w. One
may always assume w to be primitive, that is not a power of another word. In
this case, the length of w is called the period of F . A periodic set is obviously
uniformly recurrent.
For an infinite pseudoword w, we denote by F (w) the set of finite factors of
w. It is an infinite factorial set.
An infinite pseudoword w is uniformly recurrent if F (w) is uniformly recurrent. This is the same definition as the definition commonly used for infinite
words.
The following is [7, Lemma 2.2].
Proposition 7.1 An infinite pseudoword is uniformly recurrent if and only if
all its infinite factors have the same finite factors.
Proof. Let w be an infinite pseudoword. Assume first that w is uniformly
recurrent. Let u be an infinite factor of w. Let us show that F (u) = F (w). The
inclusion F (u) ⊂ F (w) is clear. Conversely, let x ∈ F (w). Since w is uniformly
recurrent, x is a factor of every long enough finite factor of w. In particular, x
is a factor of every long enough finite prefix of u. Thus x ∈ F (u).
Conversely, assume that F (w) = F (u) for all infinite factors u of w. Let v be
a finite factor of w. Arguing by contradiction, assume that there are arbitrary
long factors of w which do not have v as a factor. This infinite set contains, by
Proposition ??, a subsequence converging to some infinite pseudoword u which
is also a factor of w, and such that v 6∈ F (u), a contradiction.
Recall that the J -order in a monoid M is defined by x ≤J y if x is a factor
of y. Two elements x, y are J -equivalent if each one is a factor of the other
(this is one of the Green’s relations , see [10]).
Replacing the notion of factor by prefix (resp. suffix), one obtains the Rorder (resp. L-order). Thus, two elements x, y of a monoid M are R-equivalent
(resp. L-equivalent) if xM = yM (resp. M x = M y). The H-equivalence is the
intersection of R and L. In any monoid, one has RL = LR and one denotes D
the equivalence RL = LR which is the supremum of R and L. In a compact
monoid, one has D = J . The proof is the same as in a finite monoid. It uses
the fact that a compact monoid satisfies the stability condition: if x ≤R y and
xJ y, then xRy and dually for L.
The following result is [7, Theorem 2.6]. It gives an algebraic characterization
of uniform recurrence in the free profinite monoid.
Theorem 7.2 An infinite pseudoword is uniformly recurrent if and only if it
is J -maximal.
The proof uses three lemmas. An element s of a semigroup S is regular if there
is some x ∈ S such that sxs = s. In a compact semigroup, a J -class contains
a regular element if and only if all its elements are regular, if and only if it
contains an idempotent.
21
For a pseudoword w, we denote X(w) the set of all infinite pseudowords
which are limits of sequences of finite factors of w.
Lemma 7.3 Let w be uniformly recurrent pseudoword over a finite alphabet A.
1. Every element of X(w) is a factor of w.
c∗ .
2. All elements of X(w) lie in the same J -class of A
3. Every element of X(w) is regular.
Proof. Assertion 1 results from the fact that F (w) is closed.
Suppose that u, v ∈ X(w). By Proposition 7.1, they have the same set of
finite factors. Thus, by Assertion 1, they are J -equivalent.
Assume that u is the limit of a sequence (un )n≥0 of finite factors of w. Since
w is recurrent, there are finite words vn such that un vn un is a factor of w. If
v is an accumulation point of the sequence vn , then uvu is a factor of w which
belongs to X(w). By Assertion 2, it is J -equivalent to u. In a compact monoid,
by the stability condition, this implies that u and uvu are H-equivalent and
thus that u is regular (indeed, uHuvu implies uHu(vu)ω and thus (vu)ω is an
idempotent in J(u)).
Lemma 7.4 Let w be a uniformly recurrent pseudoword. Each H-class contained in the J -class of w contains some element of X(w).
Proof. Let u ∈ J(w). Denote by xn and yn the prefix and the suffix of u of
length n. Since w is uniformly recurrent by Proposition 7.1, there is a factor
tn of w of length at least 2n having xn as a prefix and yn as a suffix. Taking
a subsequence, we may assume that the sequences (xn ), (yn ) and (tn ) converge
to x, y, t. Then x, y, t ∈ J(w) by Lemma 7.3(2). Since x ≥R t and t ≤L y, by
stability, we obtain uHt.
Lemma 7.5 Let u be a uniformly recurrent pseudoword and suppose v is a
pseudoword such that uv is still uniformly recurrent. Then u and uv are Requivalent.
Proof. Suppose first that v is finite. Let un be the suffix of u of length n. Since
u is an infinite factor of uv which is uniformly recurrent, they have the same
finite factors by Lemma 7.1. Hence for every n there is some m(n) such that
um(n) = xn un vyn . By compactness, we may assume by taking subsequences that
the sequences xn , yn , un converge to x, y, u′ respectively. Then by continuity of
the multiplication, the sequence um(n) converges to xu′ vy. Since the limits of
two convergent sequences of suffixes of the same pseudoword are L-equivalent,
we obtain that xu′ vyLu and thus uRu′ v by stability. Since u′ is the limit of a
sequence of suffixes of u, there is some factorization of the form u = zu′ . Since
the R-equivalence is a left congruence, we finally obtain uv = zu′ vRzu′ = u.
Assume next that v is infinite. We assume by contradiction that u >R uv.
Let vn be a sequence of finite words converging to v. Taking a subsequence, we
22
may assume that uvn >R u for all n. Thus for each n, we have a factorization
vn = xn an yn with an ∈ A such that uRuxn >R uxn an ≥R uv. Since the
c∗ is compact we may, up to taking a subsequence, assume
alphabet is finite and A
that the letter sequence an is constant and the sequences xn , yn converge to x
and y respectively. Thus we have p = xay with a ∈ A and uRux >R uxa ≥R uv.
On the other hand, since ux and uxa are infinite factors of uv, they are both
uniformly recurrent by Proposition 7.1. By the first part, we have uxRuxa, a
contradiction.
A dual result holds for the L-order.
Proof of Theorem 7.2. Suppose first that w is J -maximal as an infinite pseudoword. If v is an infinite factor of w, it is J -equivalent to w. Hence v, w have
the same factors and, in particular, the same finite factors. By Proposition 7.1,
w is uniformly recurrent.
c∗ \ A∗ are such that u ≥J w with w
Suppose conversely that u, w ∈ A
c∗ . By the dual of Lemma 7.5,
uniformly recurrent. Set w = puq with p, q ∈ A
we have puLu. And by Lemma 7.5, we have puRpuq. Thus u and w are J equivalent.
c∗ is made of one H-class. The J -class of
Example 7.6 The J -class of aω in A
ω
(ab) has four H-classes. It is represented in Figure 6.
(ab)ω (ab)ω a
b(ab)ω (ba)ω
Figure 6: The J -class of (ab)ω .
7.2
The J -class J(F )
Let F be a uniformly recurrent set of finite words on the alphabet A. The
c∗ is also factorial (see [3, Proposition 2.4]). The proof relies
closure F̄ of F in A
on the following useful lemma from [3, Lemma 2.5].
c∗ and every sequence (wn ) converging to uv,
Lemma 7.7 For every u, v ∈ A
there are sequences (un ), (vn ) such that lim un = u, lim vn = v and (un vn ) is a
subsequence of (wn ).
The set of two-sided infinite words with all their factors in F is denoted
X(F ). It is closed for the product topology of AZ . It is also invariant by the
shift σ : AZ → AZ defined by y = σ(x) if yn = xn+1 for any n ∈ Z. Such a
23
closed and shift invariant set is called a subshift . It is classical that a subshift
is of the form X(F ) for some uniformly recurrent set if and only if it is minimal
(see [27] for example).
By the results of Section 7, all the infinite pseudowords in the closure F̄ of
F are J -equivalent. We denote by J(F ) their J -class.
Example 7.8 The Fibonacci morphism ϕ is primitive. The set F of factors of
the words ϕn (a) for n ≥ 1 is called the Fibonacci set . It contains the infinitely
recurrent pseudowords ϕω (a) and ϕω (b).
For a uniformly recurrent set F , the J -class J(F ) can be described as follows.
c∗ , denote by →
x the right infinite word whose finite prefixes are those
For x ∈ A
←
of x. Symmetrically, x is the left inifnite word whose finite suffixes are those of
x. The following is [3, Lemma 6.6].
Proposition 7.9 For a uniformly recurrent set F , two words u, v ∈ J(F ) are
← ←
→ →
R-equivalent if and only if u = v and L-equivalent if and only if u = v
It follows from this that w ∈ J(F ) belongs to a subgroup if and only if the
← →
two-sided infinite word w · w has all its factors in F . Indeed, the finite factors
of w2 are those of w plus the the products uv where u is a finite suffix of w and
v is a finite prefix of w [3, Lemma 8.2].
Thus the maximal subgroups of J(F ) are in bijection with the elements of the
set X(F ) of two-sided infinite words with all their factors in F . For x ∈ X(F ),
we denote by Hx the maximal subgroup corresponding to x.
Example 7.10 Let F be the Fibonacci set and let w = ϕω (a). The right
←
→
infinite word w is the Fibonacci word. The left infinite word w is the word with
→
←
suffixes ϕ2n (a) (see Section 3). The two sided infinite word w · w is a fixed
point of ϕ2 .
Example 7.11 let A = {a, b} and let ϕ : A∗ → A∗ be defined by ϕ(a) = ab
and ϕ(b) = a3 b. Let F be the set of factors of ϕω (a). Since ϕ is primitive, F
is uniformly recurrent. The pseudowords ϕω (a) and ϕω (b) belong to the same
→
→
←
←
H-class of J(F ). Indeed, we have ϕω (a)=ϕω (b) and ϕω (a)=ϕω (b).
7.3
Fixed points of substitutions
Let A be a finite alphabet. Let ϕ : A∗ → A∗ be a morphism, also called a
substitution over A. Then ϕ extends uniquely by continuity to a morphism still
c∗ → A
c∗ . The monoid End(A
c∗ ) is profinite by Theorem 5.8. Thus
denoted ϕ : A
ω
the morphism ϕ is well defined as the unique idempotent in the closure of the
semigroup generated by ϕ.
A substitution ϕ : A∗ → A∗ is said to be primitive if there is an integer n
such that all letters appear in every ϕn (a) for a ∈ A.
24
A fixed point of a substitution ϕ is an infinite word x ∈ AN such that
ϕ(x) = x. As well known, a fixed point of a primitive substitution is uniformly
recurrent (see [17, Proposition 1.3.2] for example).
The following is a particular case of [7, Theorem 3.7] (in which the notion
weakly primitive substitution is introduced).
Theorem 7.12 Let ϕ be a primitive substitution over a finite alphabet A. Then
the pseudowords ϕω (a) with a ∈ A are uniformly recurrent and are all J equivalent.
Proof. Let u be a finite factor of ϕω (a) for some a ∈ A. Then there is an integer
N such that u is a factor of any factor of length N of ϕn (b) for all b ∈ A. Thus
u is a factor of any finite factor of length N of any ϕω (b). This proves both
claims.
Example 7.13 The Fibonacci morphism ϕ : a 7→ ab, b 7→ a is primitive. Thus
the pseudowords ϕω (a) and ϕω (b) are uniformly recurrent and J -equivalent.
Example 7.14 The Thue-Morse substitution is the morphism τ : a 7→ ab, b 7→
ba. It is primitive. The unique fixed point x = abbabaab · · · of τ beginning with
a is called the Thue-Morse infinite word. The set of its factors is called the
Thue-Morse set.
Given a substitution ϕ over A, we denote by ϕG the endomorphism of F\
G(A)
c∗ → F\
such that the following diagram commutes where π : A
G(A) denotes the
ϕ
c∗
c∗
A
A
π
F\
G(A)
π
ϕG
F\
G(A)
canonical projection.
For a substitution ϕ over A, the endomorphism ϕω
G is the identity if and
only if ϕ is invertible (as a map from F G(A) into itself).
Indeed, if ϕ is an automorphism of F G(A), then its extension to F\
G(A) is
also an automorphism. But then ϕω
G is the identity since in a group one has
xω = 1 for any element x.
Conversely, if ϕω
G is the identity, then ϕ is a bijection from F G(A) onto itself
and thus it is an automorphism of F G(A).
Example 7.15 Let A = {a, b} and let ϕ : a 7→ ab, b 7→ a be the Fibonacci
morphism. Then ϕ is an automorphism of F G(A) since ϕ−1 : a 7→ b, b 7→ b−1 a.
ω
Accordingly ϕω
G is the identity. In particular, one has ϕG (a) = a. This explains
in a simple way that the length of ϕω
(a)
is
equal
to
1
(see
Section 3).
G
25
8
Sturmian sets and tree sets
Let F be a factorial set on the alphabet A. For w ∈ F , we denote
LF (w)
= {a ∈ A | aw ∈ F },
RF (w)
EF (w)
= {a ∈ A | wa ∈ F },
= {(a, b) ∈ A × A | awb ∈ F }
and further
ℓF (w) = Card(LF (w)),
rF (w) = Card(RF (w)),
eF (w) = Card(EF (w)).
For w ∈ F , we denote
mF (w) = eF (w) − ℓF (w) − rF (w) + 1.
A word w is called neutral if mF (w) = 0. A factorial set F is neutral if every
word in F is neutral.
Example 8.1 The Fibonacci set is neutral as any Sturmian set.
Example 8.2 The Thue-Morse set T is not neutral. Indeed, since A2 ⊂ T , one
has mT (ε) = 1.
Example 8.3 Let ϕ : a 7→ ab, b 7→ a3 b be as in Example 7.11. Let F be the set
of factors of ϕω (a). It is not neutral since m(a) = 1 and m(aa) = −1.
A neutral set has complexity kn + 1 where k = Card(A) − 1 (see [11]).
8.1
Sturmian sets
We recall here some notions concerning episturmian words (see [9] for more
details and references).
A word w is right-special (resp. left-special ) if ℓF (w) ≥ 2 (resp. rF (w) ≥ 2).
A right-special (resp. left-special) word w is strict if ℓF (w) = Card(A) (resp.
rF (w) = Card(A)). In the case of a 2-letter alphabet, all special words are
strict.
By definition, an infinite word x is episturmian if F (x) is closed under reversal and if F (x) contains, for each n ≥ 1, at most one word of length n which
is right-special.
Since F (x) is closed under reversal, the reversal of a right-special factor of
length n is left-special, and it is the only left-special factor of length n of x. A
suffix of a right-special factor is again right-special. Symmetrically, a prefix of
a left-special factor is again left-special.
As a particular case, a strict episturmian word is an episturmian word x
with the two following properties: x has exactly one right-special factor of each
length and moreover each right-special factor u of x is strict, that is satisfies the
inclusion uA ⊂ F (x) (see [15]).
26
For a ∈ A, denote by ψa the morphism of A∗ into itself, called elementary
morphism, defined by
(
ab if b 6= a
ψa (b) =
a otherwise
Let ψ : A∗ → End(A∗ ) be the morphism from A∗ into the monoid of endomorphisms of A∗ which maps each a ∈ A to ψa . For u ∈ A∗ , we denote by
ψu the image of u by the morphism ψ. Thus, for three words u, v, w, we have
ψuv (w) = ψu (ψv (w)). A palindrome is a word w which is equal to its reversal.
Given a word w, we denote by w(+) the palindromic closure of w. It is, by
definition, the shortest palindrome which has w as a prefix.
The iterated palindromic closure of a word w is the word Pal(w) defined
recursively as follows. One has Pal(1) = 1 and for u ∈ A∗ and a ∈ A, one has
Pal(ua) = (Pal(u)a)(+) . Since Pal(u) is a proper prefix of Pal(ua), it makes sense
to define the iterated palindromic closure of an infinite word x as the infinite
word which is the limit of the iterated palindromic closure of the prefixes of x.
Justin’s Formula is the following. For every words u and v, one has
Pal(uv) = ψu (Pal(v)) Pal(u) .
This formula extends to infinite words: if u is a word and v is an infinite word,
then
Pal(uv) = ψu (Pal(v)) .
(1)
There is a precise combinatorial description of standard episturmian words (see
e.g. [21, 18]).
Theorem 8.4 An infinite word s is a standard episturmian word if and only if
there exists an infinite word ∆ = a0 a1 · · · , where the an are letters, such that
s = lim un ,
n→∞
where the sequence (un )n≥0 is defined by un = Pal(a0 a1 · · · an−1 ). Moreover, the
word s is episturmian strict if and only if every letter appears infinitely often
in ∆.
The infinite word ∆ is called the directive word of the standard word s. The
description of the infinite word s can be rephrased by the equation
s = Pal(∆) .
As a particular case of Justin’s Formula, one has
un+1 = ψa0 ···an−1 (an )un .
The words un are the only prefixes of s which are palindromes.
27
(2)
Example 8.5 The Fibonacci word is a standard episturmian word with directive word ∆ = ababa · · · . Indeed, by Formula (2) one has Pal(∆) = ψab (Pal(∆)).
Since ψab = ϕ2 where ϕ is the Fibonacci morphism, we have Pal(∆) = ϕ2 (Pal(∆)).
This shows that Pal(∆) is the Fibonacci word.
We note that for n ≥ 1, one has
|un+1 | ≤ 2|un |.
(3)
Indeed, set un = u′n ba. If an = a, the word u′n baabũ′n is palindrome of length at
most 2|un |. If an = b, then u′n babu˜′n is a palindrome of length strictly less than
2|un |.
Example 8.6 As a consequence of Equation (2), when s is the Fibonacci word
and ϕ the Fibonacci morphism, we have for every n ≥ 0
un+1 = ϕn (a)un .
(4)
In view of Equation (2), we need to show that ϕn (a) = ψa0 ···an−1 (an ). By Example 8.5, the directive word of s is ababa · · · . If n is even, then ψa0 ···an−1 (an ) =
ψ(ab)n/2 (a) = ϕn (a) since ψab = ϕ2 . If n is odd, then ψa0 ···an−1 (an ) = ψ(ab)(n−1)/2 a (b) =
ϕn−1 (ab) = ϕn (a) and the property is true also.
As a consequence, we have in the prefix ordering for every n ≥ 0,
un < ϕn+1 (a).
(5)
Indeed, both words are prefixes of the Fibonacci word and it is enough to compare their lengths. For n = 0, we have |u0 | = 0 and |ϕ(a)| = 2. Next, for
n ≥ 1, we have by (4), un = ϕn−1 (a)un−1 . Arguing by induction, we obtain
|un | < |ϕn−1 (a)| + |ϕn (a)| = |ϕn+1 (a)|.
8.2
Tree sets
Let F be a factorial set of words. For w ∈ F , we consider the set EF (w) as an
undirected graph on the set of vertices which is the disjoint union of LF (w) and
RF (w) with edges the pairs (a, b) ∈ EF (w). This graph is called the extension
graph of w.
A factorial set F is called biextendable if every w ∈ F can be extended on
the left and on the right, that is such that ℓF (w) > 0 and rF (w) > 0.
A biextendable set is a tree set if for every w ∈ F , the graph EF (w) is a
tree. A tree set is neutral.
More generally one also defines a connected (resp. acyclic), as a biextendable
set F such that for every w ∈ F , the graph EF (w) is connected (resp. acyclic).
Thus a biextendable set is a tree set if and only if it is both connected and
acyclic.
Example 8.7 The Fibonacci set is a tree set. This follows from the fact that
it is a Sturmian set (see [11]) and that every Sturmian set is a tree set.
28
Example 8.8 The Tribonacci set is the set of factors of the fixed point of the
morphism ϕ : a 7→ ab, b 7→ ac, c 7→ a. It is a also a tree set (see [11]). The graph
E(ε) is represented in Figure 7.
a
c
b
b
c
a
Figure 7: The extension graph of ε in the Tribonacci set.
Example 8.9 Let ϕ : a 7→ ab, b 7→ a3 b be as in Example 7.11. Let F be the
set of factors of ϕω (a). The graphs E(a) and E(aa) are shown in Figure 8.
The first graph has a cycle of length 4 and the second one has two connected
a
a
a
a
b
b
b
b
Figure 8: The extension graphs E(a) and E(aa).
components. Thus F is not a tree set (it is not either an acyclic or a connected
set, as defined in [11]).
Example 8.10 Let T be the Thue-Morse set (see example 7.14). The set F is
not a tree set since ET (ε) is the complete bipartite graph K2,2 on two sets with
2 elements.
9
Return words
In this section, we introduce return words. We begin with the classical notion
of left and right return words in factorial sets. We then develop a notion of limit
return sets for pseudowords.
9.1
Left and right return words
Let F be a factorial set. A return word to x ∈ F is a nonempty word w ∈ F
such that xw begins and ends by x but has no internal factor equal to x. We
denote by RF (x) the set of return words to x.
For x ∈ F , we denote
ΓF (x) = {w ∈ F | xw ∈ F ∩ A∗ x}.
29
Thus RF (x) is the set of nonempty words in ΓF (x) without any proper prefix
in ΓF (x). Note that ΓF (x) is a right unitary submonoid of A∗ .
One also defines a left return word to x ∈ F as a nonempty word such that
wx begins and ends with x but has no internal factor equal to x. We denote by
R′F (x) the set of left return words to x. One has obviously R′F (x) = xRF (x)x−1 .
Example 9.1 Let F be the Fibonacci set. The sets of right and left return
words to a are RF (a) = {a, ba} and R′F (a) = {a, ab}. Similarly, RF (b) =
{ab, aab} and R′F (b) = {ba, baa}.
Example 9.2 Let F be a periodic set. Let w be a primitive word of length n
such that F = F (w∗ ). Then, for any word x ∈ F of length at least n, the set
RF (w) is reduced to one word of length n.
The following is [13, Equation (4.2)].
Proposition 9.3 Let F be a factorial set. For any x ∈ F , one has ΓF (x) =
RF (x)∗ ∩ x−1 F .
Proof. If a nonempty word w is in ΓF (x) and is not in RF (x), then w = uv with
u ∈ ΓF (x) and v nonempty. Since γF (x) is right unitary, we have v ∈ ΓF (x),
whence the conclusion w ∈ RF (x)∗ by induction on the length of w. Moreover,
one has xw ∈ F and thus w ∈ x−1 F .
Conversely, assume that w is a nonempty word in RF (x)∗ ∩ w−1 F . Set
w = uv with u ∈ RF (x) and v ∈ RF (x)∗ . Then xw = xuv ∈ A∗ xv ⊂ A∗ x and
xw ∈ F . Thus w ∈ ΓF (x).
Note that, as a consequence, for x, y ∈ F such that xy ∈ F , we have
RF (xy) ⊂ RF (y)∗ .
(6)
Indeed, if w ∈ RF (xy), then xyw ∈ F ∩ A∗ xy implies yw ∈ F ∩ Ay and thus
the result follows since ΓF (y) ⊂ RF (y)∗ by Proposition 9.3.
The dual of Proposition 9.3 and of Equation (6) hold for left return words.
By a result of [8], in a uniformly recurrent neutral set, one has
Card(RF (x)) = Card(A)
(7)
for every x ∈ F .
Example 9.4 Let ϕ : a 7→ ab, b 7→ a3 b be as in Example 7.11. Let F be the set
of factors of ϕω (a). One has RF (a) = {a, ba} but RF (aa) = {a, babaa, babababaa}.
Thus the number of return words is not constant in a uniformly recurrent set
which is not neutral.
There is an explicit form for the return words in a Sturmian set (see [21,
Theorem 4.4, Corollaries 4.1 and 4.5]).
30
Proposition 9.5 Let s be a standard strict episturmian word over A, let ∆ =
a0 a1 · · · be its directive word, and let (un )n≥0 be its sequence of palindrome
prefixes.
(i) The left return words to un are the words ψa0 ···an−1 (a) for a ∈ A.
(ii) For each factor u of s, let n be the minimal integer such that u is a factor
of un . There there is a unique word z such that zu is a prefix of un and
the left return words to u are the words z −1 yz, where y ranges over the
left return words to un .
Example 9.6 Let ϕ be the Fibonacci morphism and let F be the Fibonacci
set. We have for n ≥ 1,
R′F (ϕ2n (aa)) = {ϕ2n (a), ϕ2n+1 (a)}.
(8)
For example, ϕ2 (aa) = abaaba and R′F (ϕ2 (aa)) = {aba, abaab}. Note that
Equation (8) does not hold for n = 0. Indeed, R′F (aa) = {aab, aabab} =
6 {a, ab}.
To show (8), we first observe that in the prefix order for n ≥ 1,
u2n < ϕ2n (aa) ≤ u2n+1 .
(9)
u2n+1 = ϕ2n (a)u2n ,
(10)
Indeed, one has by (4)
and, since n ≥ 1,
u2n = ϕ2n−1 (a)u2n−1 = ϕ2n−1 (a)ϕ2n−2 (a)u2n−2 = ϕ2n (a)u2n−2 .
(11)
Thus u2n+1 = ϕ2n (a)u2n = ϕ2n (aa)u2n−2 . This proves the second inequality
in (9). Since |u2n | ≤ 2|u2n−1 | by (3), and |u2n−1 | < |ϕ2n (a)| by (5), we obtain
|u2n | < 2|ϕ2n (a)| and this proves the first inequality .
By (9), the minimal integer m such that ϕ2n (aa) is a factor of um is m =
2n + 1. Thus, by Proposition 9.5 (ii), one has R′F (ϕ2n )(aa) = R′F (u2n+1 ). On
the other hand, by Proposition 9.5 (i), we have R′F (u2n+1 ) = {ψ(ab)n a (a), ψ(ab)n a (b)}.
Since ψ(ab)n a (a) = ϕ2n (a) and ψ(ab)n a (b) = ϕ2n (ab) = ϕn+1 (a), this proves (8).
The following is [11, Theorem 4.5]. It shows that in a tree set, a property
much stronger than Equation (7) holds.
Theorem 9.7 Let F be a uniformly recurrent tree set. For any x ∈ F , the set
RF (x) is a basis of F G(A).
The proof uses Equation (7) and the following result [11, Theorem 4.7].
Theorem 9.8 Let F be a uniformly recurrent connected set. For any w ∈ F ,
the set RF (w) generates the free group F G(A).
Example 9.9 Let F be the Tribonacci set on A = {a, b, c}. Then RF (a) =
{a, ba, ca}, which is easily seen to be a basis of F G(A).
31
9.2
Limit return sets
Let k ≥ 1 be an integer. A recurrent set F is k-bounded if every x ∈ F has at
most k return words. The set F is bounded if it is k-bounded for some k.
Thus, by Equation (7), a neutral set is k-bounded with k = Card(A).
Clearly, any bounded set is uniformly recurrent. There exist uniformly recurrent sets which are not bounded (see [16, Example 3.17]).
Let F be a k-bounded set. Consider an element x which belongs to a group
in J(F ). Let rn be a sequence of finite prefixes of x strictly increasing for
the prefix order. Similarly, let ℓn be a sequence of finite suffixes of x stricly
increasing for the suffix order. Since x2 ∈ J(F ), we have ℓn rn ∈ F . Let
Rn = rn RF (ℓn rn )rn−1 .
Up to taking a subsequence, we may assume that the set Rn has a fixed number
ℓ ≤ k of elements rn,1 , . . . rn,ℓ and that the sequence (rn,1 , . . . , rn,ℓ ) is convergent
ℓ
c∗ . Its limit R is called a limit return set to x. The sequence (ℓn , rn , Rn )
in A
is called an approximating sequence for R.
We note that
Rn+1 ⊂ R∗n .
(12)
Indeed, set rn+1 = rn sn . Since ℓn+1 rn is a prefix of ℓn+1 rn+1 , we have by the
dual of (6) the inclusion R′F (ℓn+1 rn+1 ) ⊂ R′F (ℓn+1 rn )∗ . Thus
RF (ℓn+1 rn+1 )
−1 −1
= rn+1
ℓn+1 R′F (ℓn+1 rn+1 )ℓn+1 rn+1 ,
−1 −1
⊂ rn+1
ℓn+1 R′F (ℓn+1 rn )∗ ℓn+1 rn+1
⊂ sn−1 RF (ℓn+1 rn )∗ sn .
Since ℓn rn is a suffix of ℓn+1 rn , we have by (6) the inclusion RF (ℓn+1 rn ) ⊂
RF (ℓn rn )∗ . Thus we obtain
Rn+1
−1
= rn+1 RF (ℓn+1 rn+1 )rn+1
,
−1
⊂ rn+1 sn−1 RF (ℓn+1 rn )∗ sn rn+1
,
⊂ rn RF (ℓn rn )∗ rn−1 = R∗n .
Example 9.10 Let ϕ be the Fibonacci morphism and F be the Fibonacci set.
We will show that {ϕω (a), ϕω (ba)} is a limit return set to x = ϕω (a).
For this, consider the sequence ℓn = rn = ϕ2n (a). The sequence is increasing
both for the prefix and the suffix order and its terms are both prefixes and
′
suffixes of x. The set Rn = rn RF (ℓn rn )rn−1 = ℓ−1
n RF (ℓn rn )ℓn is by (8)
Rn
=
ϕ2n (a)−1 {ϕ2n (aa), ϕ2n+1 (a)}ϕ2n (a) = {ϕ2n (a), ϕ2n (b)ϕ2n (a)}
=
{ϕ2n (a), ϕ2n (ba)}.
The sequence (ϕn! (a), ϕn! (ba)} converges to {ϕω (a), ϕω (ba)}, which proves the
claim.
32
The following example shows that in degenerated cases, a limit return set
may contain finite words.
Example 9.11 Let F be a periodic set of period n. By Example 9.2 a return
set to any word of length larger than n is formed of one word of length n. Thus
any return set to a pseudoword x ∈ J(F ) is formed of one word of length n.
10
Schützenberger groups
Let M be a topological monoid. For an element x ∈ M , we denote by H(x) the
H-class of x.
Let H be an H-class of M . Set T (H) = {x ∈ M | Hx = H}. Each
x ∈ T (H) defines a map ρx : H → H defined by ρx (h) = hx. The set of the
translations ρx for x ∈ T (H) is a topological group acting by permutations on
H, denoted Γ(H). The groups corresponding to different H-classes contained
in the same J -class J are continuously isomorphic and the equivalence called
the Schützenberger group of J, denoted G(J).
If J is a regular J -class, any H-class of J which is a group is isomorphic to G(J). Indeed, H ⊂ T (H) and the restriction to H of the mapping
ρ : x ∈ T (H) → ρx ∈ Γ(H) is an isomorphism (see [25] for a more detailed
presentation).
The following is [4, Proposition 5.2].
Theorem 10.1 Let F be a non-periodic bounded set. Let x ∈ J (F ) be such
that H(x) is a group and let R be a limit return set to x. Then H(x) is the
closure of the semigroup generated by R.
Note that Theorem 10.1 does not hold without the hypothesis that F is nonperiodic (see Example 9.11).
This fundamental result is the key to understand the role played by the
group generated by return words. Actually, let (ℓn , rn , Rn ) be an approximating
sequence for R. Then, since Rn+1 ⊂ R∗n by (12), we have R∗ = ∩n≥0 R∗n . Thus
the group H(x) is the intersection of the submonoids R∗n and each of them is
the closure of the submonoid generated by Rn .
Example 10.2 Let ϕ be the Fibonacci morphism and let F be the Fibonacci
set. The H-class of the pseudoword x = ϕω (a) is a group. Indeed, H(x)
contains the idempotent ϕω (aω ). The group H(x) is the closure of the semigroup
generated by x and y = ϕω (ba), that is, isomorphic to F[
(A).
The following statement is a generalization of [5, Theorem 6.5]. We denote
by G(F ) the Schützenberger group of J(F ).
c∗ → G be a
Theorem 10.3 Let F be a non-periodic bounded set and let f : A
∗
c
continuous morphism from A onto a profinite group G. The following conditions are equivalent.
33
(i) The restriction of f to G(F ) is surjective.
(ii) For every w ∈ F , the submonoid f (RF (w)∗ ) is dense in G.
Proof. Let x ∈ J(F ) be such that H(x) is a group. Let (ℓn , rn ) be a sequence
of pairs of suffixes and prefixes of x of strictly increasing length. Taking a
subsequence, we may assume that Rn = rn RF (ℓn rn )rn−1 converges to the limit
return set R. Since, by (12) we have Rn+1 ⊂ R∗n , the semigroup generated by
R is ∩n≥0 R∗n .
(i) implies (ii). Assume by contradiction that f (RF (w)∗ ) is not dense in G.
Since w is a factor of x, we may assume that r0 ends with w. Then RF (ℓ0 r0 ) ⊂
RF (w)∗ . This implies that f (RF (ℓ0 r0 )∗ ) is not dense in G. Since f (R∗0 ) is
conjugate to f (RF (ℓ0 r0 )∗ ), the same holds for f (R∗0 ). Thus f (R∗ ) is not dense
in G. But by Theorem 10.1, H(x) is the closure of the semigroup generated by
R. We conclude that f (H(x)) is not dense in G.
(ii) implies (i). Since H(x) = R∗ = ∩n≥0 R∗n , we have f (H(x)) = ∩n≥0 f (R∗n ).
But Rn is conjugate to R(ℓn rn ) and thus by (ii), each f (R∗n ) is dense in G. Thus
f (H(x)) = G.
Corollary 10.4 Let F be a non-periodic bounded set on the alphabet A. The
following conditions are equivalent.
(i) The restriction to any maximal subgroup of J(F ) of the natural projection
c∗ → F\
pG : A
G(A) is surjective.
(ii) For each w ∈ F the set RF (w) generates the free group F G(A).
Proof. We apply Theorem 10.3 with f being the identity. Let x ∈ J(F ) be such
that H(x) is a group.
(i) implies (ii). Let w ∈ F . There is a maximal subgroup of J(F ) contained in the topological closure of RF (w)∗ in the free profinite monoid (indeed,
RRF (ln rn )∗ is a subset of RF (w)∗ , for some suitable sequence of words ln rn as
defined some pages before, for infinitely many n).
It follows that the topological closure of RF (w)∗ in the free profinite group
generated by A is the whole free profinite group. This implies that RF (w)
generates F G(A).
(ii) implies (i). By Theorem 10.3, the restriction to H(x) of the projection
c∗ → F\
pG : A
G(A) is surjective.
10.1
Groups of tree sets
We now consider uniformly recurrent tree sets. Note that a uniformly recurrent
tree set F is non-periodic. Indeed, if F is the set of factors of w∗ with w
primitive, then RF (w) = {w} since w does not overlap nontrivially w2 . Thus
F is not a tree set by Theorem 9.7.
Theorem 10.5 Let F be a uniformly recurrent tree set. Then the following
assertions hold.
34
1. The group G(F ) is the free profinite group on A. More precisely, the
restriction to any maximal subgroup of J(F ) of the natural projection pG :
c∗ → F\
A
G(A) is an isomorphism.
2. Let H be a subgroup of finite index n in F G(A). For any maximal group
G in J(F ), G ∩ p−1
G (H̄) is a subgroup of index n of G.
Proof. 1. This results directly from Corollary 10.4 since by Theorem 9.7, RF (w)
is a basis of F G(A) for every w ∈ F when F is a uniformly recurrent tree set.
Since H(x) is the closure of a semigroup generated by Card(A) elements, there
is a continuous morphism ψ from F\
G(A) onto H(x). Thus pG ◦ ψ is continuous
\
surjective morphism from F G(A) onto itself. By Proposition 5.7, it implies that
it is an isomorphism. This proves the first assertion.
2. This results from Corollary 10.4 since the restriction α of pG to G is an
−1
(H̄).
isomorphism from G onto F\
G(A) and G ∩ p−1
G (H̄) = α
Example 10.6 Let F be the Fibonacci set. We have seen that G(F ) is the free
profinite group on A (Example 9.10).
10.2
Groups of fixed points of morphisms
Let ϕ : A∗ → A∗ be a primitive substitution and let F (ϕ) be the set of factors
of a fixed point of ϕ. We denote by J(ϕ) the closure of F (ϕ) and by G(ϕ) the
Schützenberger group of J(ϕ).
A connexion for ϕ is a word ba with b, a ∈ A such that ba ∈ F (ϕ), the first
letter of ϕω (a) is a and the last letter of ϕω (b) is b. Every primitive substitution
has a connexion [7, Lemma 4.1]. A connective power of ϕ is a finite power ϕ̃ of
ϕ such that the first letter of ϕ̃(a) is a, the last letter of ϕ̃(b) is b. We denote
Xϕ (a, b) = aRF (ba)a−1 . The set Xϕ (a, b) is a code.
Example 10.7 Let τ : a 7→ ab, b 7→ ba be the Thue-Morse morphism. The
word aa is a connection for τ and τ̃ = τ 2 is a connecting power of τ . The
set X = Xτ (a, a) has four elements x = abba, y = ababba, z = abbaba and
t = ababbaba.
The following is [4, Theorem 5.6].
Theorem 10.8 Let ϕ be a non periodic primitive substitution. Consider a
connexion ba for ϕ and a connective power ϕ̃. The intersection Hba of the
R-class of ϕω (a) with the L-class of ϕω (b) is a group and Hba = ϕ̃(Hba ) =
ϕω (Hba ).
The proof uses the notion of recognizablity of a substitution. We give the
definition in the following form (see [22] for the equivalence with equivalent
forms). Given a morphism ϕ : A∗ → A∗ , a pair (q, r) of words in A∗ is synchronizing if for any p, s, t ∈ A∗ such that ϕ(t) = pqrs, one has t = uv with
35
u
v
ϕ
p
ϕ
q
r
s
Figure 9: A synchronizing pair.
ϕ(u) = pq and ϕ(v) = rs (see Figure 9). Let F be the set of factors of a fixed
point of ϕ. The morphism ϕ is recognizable if there is an integer n ≥ 1 such that
for any x, y ∈ F ∩ An such that xy ∈ F , the pair (ϕ(x), ϕ(y)) is synchronizing.
By a result of Mossé [30], any non-periodic primitive substitution is recognizable
(see [22] for a new version of the proof).
The following is the main result if [4] (Theorem 6.2).
Theorem 10.9 Let ϕ be a non-periodic primitive substitution over the alphabet
A. Let ba be a connexion of ϕ and let Xϕ = Xϕ (a, b). Then G(ϕ) admits the
presentation
hX | ϕ̃ω
X,G (x) = x, x ∈ XiG .
Example 10.10 Let τ : a 7→ ab, b 7→ ba be the Thue-Morse morphism. We
have seen in Example 10.7 that the word aa is a connection for τ and τ̃ = τ 2
is a connecting power of τ . The set X = Xτ (a, a) has four elements x = abba,
y = ababba, z = abbaba and t = ababbaba. By Theorem 10.9, the group G(τ ) is
ω
the group generated by X with the relations τX,G
(u) = u for u ∈ X. Actually,
ω
ω−1
ω
ω
since τ (y)τ
(x)τ (z) = τ (t), the relation xy −1 z = t is a consequence of
the relations above and thus G(ϕ) is generated by x, y, z.
Let f : A∗ → G be a morphism from A∗ into a finite group G and let
ϕ : A∗ → A∗ be a morphism. We denote by ϕG the map from GA into itself
defined as follows. Consider h ∈ GA . We may naturally extend h to a map from
A∗ into G. For a ∈ A, we define the image of a by ϕG (h) as ϕG (h)(a) = h(ϕ(a)).
We say that ϕ has finite f -order if there is an integer n ≥ 1 such that ϕnG (f ) = f .
The least such integer is called the f -order of ϕ.
Any substitution ϕ which is invertible in F G(A) is of finite h-order for any
morphism h into a finite group. Indeed, since G is finite, there are integers n, m
with n < m such that ϕn+m
= ϕnG . Since ϕ is invertible, ϕm
G is the identity.
G
Example 10.11 Let ϕ : a 7→ ab, b 7→ a be the Fibonacci substitution and
let h : A∗ → Z/2Z be the parity of the length, that is the morphism into the
additive goup of integers modulo 2 sending each letter to 1. Then ϕ is of h-order
3.
The following is a consequence of Theorem 10.9, using [4, Proposition 3.2].
Corollary 10.12 Let ϕ be a non-periodic primitive substitution over A and let
c∗ → G
h : A∗ → G be a morphism onto a finite group. The restriction of ĥ : A
to any maximal subgroup of J(ϕ) is surjective if and only if ϕ has finite h-order.
36
Example 10.13 Let ϕ be as in Example 8.9, let G = Z/2Z and let h : A∗ → G
be the parity of the length. Then ϕG (h) = (0, 0) and ϕG (0, 0) = (0, 0). Thus
ϕ does not have finite h-order. Actually, any pseudoword in J(ϕ) which is in
the image of ϕ̂ has even length and thus is mapped by h to 0. Thus, by Theorem 10.8, there is a maximal group G in J(F ) which contains only pseudowords
of even length and therefore ĥ(G) = {0}, showing that the restriction of ĥ to G
is not surjective.
The following example is from [4, Section 7.2].
Example 10.14 Let ϕ : a 7→ ab, b 7→ a3 b be as in the previous example and let
h : A∗ → A5 be the morphism from A∗ onto the alternating group A5 defined
by h : a 7→ (123), b 7→ (345). One may verify that ϕ has h-order 12. Thus A5
is a quotient of G(ϕ). It is not known whether any finite group is a quotient of
G(ϕ).
Proper substitutions A substitution ϕ over A is proper if there are letters
a, b ∈ A such that for every d ∈ A, ϕ(d) starts with a and ends with b. Theorem 10.9 takes a simpler form for proper substitutions. The following is [4,
Theorem 6.4].
Theorem 10.15 Let ϕ be a non-periodic proper primitive substitution over a
finite alphabet A. Then G(ϕ) admits the presentation
hA | ϕω
G (a) = a, a ∈ AiG .
The proof uses Proposition 5.17 applied with the diagram of Figure 10.
ϕω+1
c∗
c∗
A
A
ϕω
H
ϕω
ϕ
H
Figure 10: A commutative diagram
Example 10.16 Let A = {a, b} and let ϕ : a 7→ ab, b 7→ a3 b. The morphism
ϕ is proper. Thus, by Theorem 10.15, the Schützenberger group of J(ϕ) has
ω
the presentation ha, b | ϕω
G (a) = a, ϕG (b) = bi. Since the image of F G(A) by
ϕ is included in the subgroup generated by words of length 2, the relations
ω
ϕω
G (a) = a and ϕG (b) = b are nontrivial and thus G(F ) is not a free profinite
group of rank two (it is actually not a free profinite group, see [7, Example 7.2]).
10.3
Groups of bifix codes
For an automaton A, we denote by ϕA the natural morphism from A∗ onto the
transition monoid of A.
37
Let F be a recurrent set. For any finite automaton A = (Q, i, T ), we denote
by rankA (F ) the minimum of the ranks of the maps ϕA (w) for w ∈ F . By [31,
Proposition 3.2], the set of elements of ϕA (F ) of rank rankA (F ) is included
in a regular J -class, called the F -minimal J -class of the monoid ϕA (A∗ ) and
denoted JA (F ). The structure group of this J -class is denoted GA (F ).
Let X ⊂ A+ be a code. A parse of a word w ∈ A∗ with respect to X is a
triple (p, x, q) with w = pxq such that p has no suffix in X, x ∈ X ∗ and q has
no prefix in X.
c∗ with respect to X is a triple (p, x, q) with
A parse of a profinite word w ∈ A
c∗ and q has no prefix in X.
w = pxq such that p has no suffix in X, x ∈ X
c∗ with respect to a finite maximal prefix
The number of parses of w ∈ A
code X is equal to the number of its prefixes which have no suffix in X. Indeed
the map (p, x, q) 7→ p assigning to each parse its first component is bijective by
Proposition 6.5.
Let F be a recurrent set. A bifix code X ⊂ F is F -maximal if it is not
properly contained in any bifix code Y ⊂ F .
A bifix code X ⊂ F is F -thin if there is a word w ∈ F which is not a factor
of X. When F is uniformly recurrent, a set X ⊂ F is F -thin if and only if it is
finite.
Let F be a recurrent set. The F -degree of a bifix code X, denoted dX (F ),
is the maximal number of parses of a word in F . A bifix code X is F -maximal
and F -thin if and only if its F -degree is finite. In this case a word of F has
dX (F ) parses if and only if it is not an internal factor of a word of X (see [9,
Theorem 4.2.8]).
Proposition 10.17 Let F be a uniformly recurrent set and let X be a finite
F -maximal bifix code. The number of parses of any element of F̄ is equal to
dX (F ).
Proof. Let w ∈ F̄ and let (un ) be a sequence of elements of F converging to
w. Since each long enough un has dX (F ) parses, we may assume that all un
have dX (F ) of parses. We may then number the parses of un as (pn,i , xn,i , qn,i )
in such a way that for fixed i, each sequence converges to (pi , xi , qi ). Since X
is finite, the sequences (pn,i ) and qn,i are ultimately constant and the sequence
c∗ . Thus w has dX (F ) parses. There cannot exist
(xn,i ) converges to some xi ∈ A
more than dX (F ) parses of a word in F̄ since the number of parses is equal to
the number of suffixes which are prefixes of X.
Let X be an F -thin and F -maximal bifix code. The F -degree of X is equal to
the F -minimal rank of the minimal automaton A of X ∗ . We denote by ϕX the
morphism ϕA , by JX (F ) the J -class JA (F ) and by GX (F ) the group GA (F ),
called the F -group of X. It is a permutation group of degree dX (F ).
We recall that for any uniformly recurrent tree set F a finite bifix code
X ⊂ F is F -maximal of F -degree d if and only if it is a basis of a subgroup of
index d (Finite Index Basis Theorem, see [12, Theorem 4.4]).
38
We prove the following result. It has the interesting feature that the hypothesis made on profinite objects has a consequence on finite words.
Theorem 10.18 Let F be a uniformly recurrent set, let Z be a goup code of
degree d and let X = Z ∩ F . Let h : A∗ → G be the morphism from A∗ onto the
syntactic monoid of Z ∗ . The restriction of ĥ to a maximal subgroup of J(F ) is
surjective if and only if the following properties hold.
(i) X is an F -maximal bifix code of F -degree d.
(ii) GX (F ) is isomorphic to G.
(iii) The morphism ϕˆX maps each H-class of J(F ) which is a group onto
GX (F ).
Proof. Set ϕ = ϕX and M = ϕ(A∗ ). Let (Q, i, i) be the minimal automaton of
Z ∗ . Thus G is a transitive permutation group on the set Q which has d elements
and h(Z ∗ ) is the stabilizer of i ∈ Q.
By [9, Theorem 4.2.11], since F is recurrent, the set X is an F -thin F maximal bifix code of F -degree at most d. Since F is uniformly recurrent, X is
finite.
Let x ∈ J(F ) be such that H = H(x) is a group such that the restriction of
ĥ to H(x) is surjective.
Since ĥ maps H(x) onto G, the pseudoword x has d parses with respect to
Z and thus with respect to X. Indeed, for any p ∈ Q, ĥ(x) sends p on some
q ∈ Q. Then x has the interpretation (u, v, w) with ph(u) = i, iĥ(v) = i and
ih(w) = q. This implies that dX (F ) = d and proves (i).
It is clear that ϕ̂(J(F )) is contained in JX (F ) since JX (F ) contains the
image by ϕ of every long enough word of F . Thus ϕ̂(x) is in JX (F ) ∩ ϕ(F ) and
its H-class K is a group.
ĥ
H(x)
G
ϕ̂
α
K
Figure 11: The reduction onto G.
Let w be a word in F and not a factor of X such that ϕ(w) ∈ K. Let P be
the set of suffixes of w which are proper prefixes of X. Since Card(P ) = dX (F ),
K is a permutation group on the set {i · p | p ∈ P } which is identified by an
isomorphism α with a subgroup of G.
If the map ĥ is surjective from H(x) onto G, the commutativity of the
diagram in Figure 11 forces α to be surjective. Moreover ϕ̂ maps H(x) onto K.
The converse is also true.
39
In the case where F is a tree set, the hypothesis of Theorem 10.18 is satisfied
by assertion 1 in Theorem 10.5. The conclusion of Theorem 10.18 is implied by
assertion 2.
Example 10.19 Let A = {a, b} and let Z = A2 . Let F be the Fibonacci
set. Then X = {aa, ab, ba}. The group GX (F ) is the cyclic group of order 2,
in agreement with the fact that X generates the kernel of the morphism from
F G(A) onto Z/2Z sending a, b to 1. The minimal automaton of X ∗ is shown
in Figure 12 on the left and the 0 minimal ideal of its transition monoid M is
represented on the right.
a, b
2
1, 2 1, 3
a
1
b
a
1/2, 3 * a
* ab
1/2 * ba
b
3
Figure 12: The minimal automaton of X ∗ and the F -minimal D-class.
Note that, in the above example, the F -minimal D-class D of M is the image of
the J -class J(F ). The following example shows that this may be true although
the image of J(F ) in the monoid M is strictly included in D.
Example 10.20 Let F be the set of factors of the fixed point of the morphism
ϕ : a 7→ ab, b 7→ a3 b (as in Example 7.11). The set F ∩ A2 is the same as in
Example 10.19 and the F -minimal D-class is also the same. However, J(F )
contains maximal groups formed of words of even length and thus its image in
M is aperiodic, that is has trivial subgroups.
We now deduce from Corollary 10.12 the following statement which gives
information on the groups GX (F ) when X is an F -maximal bifix code in a set
F which is not a tree set. It would be interesting to have a direct proof of this
statement which does not use profinite semigroups.
Theorem 10.21 Let ϕ be a primitive non-periodic substitution over the alphabet A and let F be the set of factors of a fixed point of ϕ. Let Z be a group code
of degree d on A and let h be the morphism from A∗ onto the syntactic monoid
of Z ∗ . Set X = Z ∩ F . If ϕ has finite h-order, then X is an F -maximal bifix
code of F -degree d and GX (F ) is isomorphic to G.
Proof. By Corollary 10.12, the hypothesis of Theorem 10.18 is satisfied and thus
the conclusion using conditions (i) and (ii).
Example 10.22 Let F and ϕ be as in Example 8.3. We consider, as in [4], the
morphim h : A∗ → A5 from A∗ onto the alternating group of degree 5 defined
by h : a 7→ (123), b 7→ (345). We have seen in Example 10.14 that ϕ has h-order
40
12 and thus, by Corollary 10.12, ĥ induces a surjective map from any maximal
subgroup of J(ϕ) onto A5 .
Let Z be the bifix code generating the submonoid stabilizing 1 and let X =
Z ∩ F . The F -maximal bifix code X has 8 elements. It is represented in
Figure 13 with the states of the minimal automaton indicated on its prefixes.
In agreement with Theorem 10.21, the F -degree of X is 5. The F -minimal
a 1
a
a 14
a 3 b
b
a
a
5
7
9
11 b
a
a 2 b
15
a
1
a
a
b
1 b
4
6 b
16
a 12
a
1
8
10 b
a
a
9
11 b
17
b
15
a
1
1
a
b
a
14
17
15
1
a
b
a
14
17
15
1
15
a
1
Figure 13: The bifix code X.
D-class is represented in Figure 14. The word a3 has rank 5 and RF (a3 ) =
1, 2, 3, 16, 17
1, 4, 5, 14, 15
1, 2, 6, 7, 17
a3
a3 b
a3 ba
1/2, 4/3, 6, 15/
1, 4, 8, 9, 15
* a3 bab
1, 2, 6, 10, 11
1, 2, 3, 12, 14
8/9
1/2/11, 17/6
*
ba3
7, 9, 10
1/3, 6, 15/9, 14
aba3
*
5, 8/4
1/11, 17/7, 16 *
baba3
bab
3, 6/2
1/2, 4/9, 14
*
*
3, 6, 15/5, 12
*
1/2, 4/3, 6, 15
10/11
Figure 14: The F -minimal D-class.
{babaaa, babababaaa}. The corresponding permutations defined on the image
41
{1, 2, 3, 16, 17} of a3 are respectively
(1, 2, 16, 3, 17) (1, 17, 16, 2, 3)
which generate A5 .
The next example is from [31].
Example 10.23 Let F be the Thue-Morse set and let A be the automaton
represented in Figure 15 on the left. The word aa has rank 3 and image I =
a
2 b
a
1
b
3
a
b
4
b
6
b
10
b
1
a
b
1
1
1
5
8
a
a
1
a
7
0
b
9
a
a
11
a
1
b
b
1
b
a
2
Figure 15: An automaton of F -degree 3 with trivial F -group
{1, 2, 4}. The action on the images accessible from I is given in Figure 16. All
b
1, 2, 8
1, 3, 10
a
b
b
b
a
b
a
1, 2, 4
1, 3, 6
1, 3, 5
1, 2, 7
1, 3, 9
1, 2, 11
a
a
Figure 16: The action on the minimal images
words with image {1, 2, 4} end with aa. The paths returning for the first time to
{1, 2, 4} are labeled by the set RF (aa) = {b2 a2 , bab2 aba2 , bab2 a2 , b2 aba2 }. Thus
rankA (F ) = 3 by [31, Theorem 3.1]. Moreover each of the words of RF (a2 )
defines the trivial permutation on the set {1, 2, 4}. Thus GA (F ) is trivial.
The fact that dA (F ) = 3 and that GA (F ) is trivial can be seen directly as
follows. Consider the group automaton B represented in Figure 15 on the right
and corresponding to the map sending each word to the difference modulo 3 of
the number of occurrences of a and b. There is a reduction ρ from A onto B
such that 1 7→ 0, 2 7→ 1, and 4 7→ 2. This accounts for the fact that dA (S) = 3.
Moreover, one may verify that any return word x to a2 has equal number of a
and b (if x = uaa then aauaa is in F , which implies that aua and thus uaa have
the same number of a and b). This implies that the permutation ϕB (x) is the
identity, and therefore also the restriction of ϕA (x) to I.
42
Example 10.24 Consider again the Thue-Morse substitution τ and the ThueMorse set F as in Example 10.10. Let h be the morphism h : a 7→ (123), b 7→
(345) from A∗ onto the alternating group A5 (already used in Example 9.4).
One may verify that τ has h-order 6 and thus, by Corollary 10.12, h extends to
a surjective continuous morphim from any maximal subgroup of J(ϕ) onto A5 .
Let Z be the group code generating the submonoid stabilizing 1 and let
X = Z ∩ F . The F -maximal bifix code X is represented in Figure 17. We
represent in Figure 17 only the nodes corresponding to right special words, that
is, vertices with two sons.
abba
3
ab
2
a
1
b
4
b
ba
a
ba
1
5
6
7
1
aba
ba
a
b2 a
a
τ 2 (b)
1
1
1
τ 2 (a)
1
τ 2 (a)
1
τ 3 (a) 10
2
12 τ (b)a
8
ba
1τ 2 (a) 1
1
τ 2 (a) 1
1
2
τ 2 (a) 11 τ (bba) 12
τ 2 (b)a
9
ba
1
1
τ 3 (b)
Figure 17: The bifix code X.
The image of τ 4 (b) is {1, 3, 4, 9, 10} and thus it is minimal. The action on
its image is shown in Figure 18. The return words to τ 4 (b) are τ 4 (b), τ 3 (a)
and τ 5 (ab). The permutations on the image of τ 4 (b) are the 3 cycles of length 5
indicated in Figure 18. Since they generate the group A5 , we have GX (F ) = A5 .
τ 3 (a) | (1, 10, 9, 3, 4)
τ 4 (b) | (1, 9, 10, 3, 4)
{1, 3, 4, 9, 10}
τ 4 (a)
{1, 2, 7, 8, 12}
τ 4 (b) | (1, 10, 9, 4, 3)
Figure 18: The action on the minimal images.
References
[1] Jorge Almeida. Dynamics of implicit operations and tameness of pseudovarieties of groups. Trans. Amer. Math. Soc., 354(1):387–411, 2002.
[2] Jorge Almeida. Profinite semigroups and applications. In Structural theory
of automata, semigroups, and universal algebra, volume 207 of NATO Sci.
43
Ser. II Math. Phys. Chem., pages 1–45. Springer, Dordrecht, 2005. Notes
taken by Alfredo Costa.
[3] Jorge Almeida and Alfredo Costa. Infinite-vertex free profinite semigroupoids and symbolic dynamics. J. Pure Appl. Algebra, 213(5):605–631,
2009.
[4] Jorge Almeida and Alfredo Costa. Presentations of Schützenberger groups
of minimal subshifts. Israel J. Math., 196(1):1–31, 2013.
[5] Jorge Almeida and Alfredo Costa. A geometric interpretation of the
Schützenberger group of a minimal subshift. Ark. Mat., 54(2):243–275,
2016.
[6] Jorge Almeida and Benjamin Steinberg. Rational codes and free profinite
monoids. J. Lond. Math. Soc. (2), 79(2):465–477, 2009.
[7] Zh. Almeı̆da. Profinite groups associated with weakly primitive substitutions. Fundam. Prikl. Mat., 11(3):13–48, 2005.
’
[8] Lubomı́ra
Balková, Edita Pelantová, and Wolfgang Steiner. Sequences with
constant number of return words. Monatsh. Math., 155(3-4):251–263, 2008.
[9] Jean Berstel, Clelia De Felice, Dominique Perrin, Christophe Reutenauer,
and Giuseppina Rindone. Bifix codes and Sturmian words. J. Algebra,
369:146–202, 2012.
[10] Jean Berstel, Dominique Perrin, and Christophe Reutenauer. Codes and
Automata. Cambridge University Press, 2009.
[11] Valérie Berthé, Clelia De Felice, Francesco Dolce, Julien Leroy, Dominique
Perrin, Christophe Reutenauer, and Giuseppina Rindone. Acyclic, connected and tree sets. Monatsh. Math., 176:521–550, 2015.
[12] Valérie Berthé, Clelia De Felice, Francesco Dolce, Julien Leroy, Dominique
Perrin, Christophe Reutenauer, and Giuseppina Rindone. The finite index
basis property. J. Pure Appl. Algebra, 219(7):2521–2537, 2015.
[13] Valérie Berthé, Clelia De Felice, Francesco Dolce, Julien Leroy, Dominique
Perrin, Christophe Reutenauer, and Giuseppina Rindone. Maximal bifix
decoding. Dicrete Math., 338:725–742, 2015.
[14] Thierry Coulbois, Mark Sapir, and Pascal Weil. A note on the continuous
extensions of injective morphisms between free groups to relatively free
profinite groups. Publ. Mat., 47(2):477–487, 2003.
[15] Xavier Droubay, Jacques Justin, and Giuseppe Pirillo. Episturmian words
and some constructions of de Luca and Rauzy. Theoret. Comput. Sci.,
255(1-2):539–553, 2001.
44
[16] Fabien Durand, Julien Leroy, and Gwenaël Richomme. Do the properties
of an S-adic representation determine factor complexity? J. Integer Seq.,
16(2):Article 13.2.6, 30, 2013.
[17] N. Pytheas Fogg. Substitutions in dynamics, arithmetics and combinatorics, volume 1794 of Lecture Notes in Mathematics. Springer-Verlag,
Berlin, 2002. Edited by V. Berthé, S. Ferenczi, C. Mauduit and A. Siegel.
[18] Amy Glen and Jacques Justin. Episturmian words: a survey. Theor. Inform. Appl., 43:403–442, 2009.
[19] Marshall Hall, Jr. Coset representations in free groups. Trans. Amer. Math.
Soc., 67:421–432, 1949.
[20] Marshall Hall, Jr. A topology for free groups and related groups. Ann. of
Math. (2), 52:127–139, 1950.
[21] Jacques Justin and Laurent Vuillon. Return words in Sturmian and episturmian words. Theor. Inform. Appl., 34(5):343–356, 2000.
[22] Karel Klouda and Stepan Starosta. Characterization of circular dol systems. 2014. http:\arxiv.org/abs/1401.0038.
[23] Donald E. Knuth. The art of computer programming. Vol. 2. AddisonWesley, Reading, MA, 1998. Seminumerical algorithms, Third edition [of
MR0286318].
[24] Neal Koblitz. p-adic numbers, p-adic analysis, and zeta-functions, volume 58 of Graduate Texts in Mathematics. Springer-Verlag, New York,
second edition, 1984.
[25] Gérard Lallement. Semigroups and combinatorial applications. John Wiley
& Sons, New York-Chichester-Brisbane, 1979. Pure and Applied Mathematics, A Wiley-Interscience Publication.
[26] Hendrick Lenstra. Profinite Fibonacci numbers.
Wiskunde, 6:297–300, 2005.
Nieuw Archief voor
[27] Douglas Lind and Brian Marcus. An introduction to symbolic dynamics
and coding. Cambridge University Press, Cambridge, 1995.
[28] Roger C. Lyndon and Paul E. Schupp. Combinatorial group theory. Classics
in Mathematics. Springer-Verlag, Berlin, 2001. Reprint of the 1977 edition.
[29] S. Margolis, M. Sapir, and P. Weil. Irreducibility of certain pseudovarieties.
Comm. Algebra, 26(3):779–792, 1998.
[30] Brigitte Mossé. Puissances de mots et reconnaissabilité des points fixes
d’une substitution. Theoret. Comput. Sci., 99(2):327–334, 1992.
45
[31] Dominique Perrin. Codes and automata in minimal sets. In Combinatorics
on Words - 10th International Conference, WORDS 2015, Kiel, Germany,
September 14-17, 2015, Proceedings, pages 35–46, 2015.
[32] Christophe Reutenauer. Une topologie du monoı̈de libre. Semigroup Forum,
18(1):33–49, 1979.
[33] Luis Ribes and Pavel Zalesskii. Profinite groups, volume 40 of Ergebnisse
der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer-Verlag, Berlin,
second edition, 2010.
[34] Stephen Willard. General topology. Dover Publications, Inc., Mineola,
NY, 2004. Reprint of the 1970 original [Addison-Wesley, Reading, MA;
MR0264581].
46
| 4math.GR
|
arXiv:1604.04146v1 [cs.NE] 14 Apr 2016
A Discrete Firefly Algorithm to Solve a Rich Vehicle Routing
Problem Modelling a Newspaper Distribution System with
Recycling Policy
Eneko Osaba1,2 , Xin-She Yang1 , F. Diaz2 , E. Onieva2 , A. D. Masegosa2 , A. Perallos2
1) E. Osaba, X.S. Yang
School of Science and Technology, Middlesex University
Hendon Campus, London, NW4 4BT, United Kingdom
2) E. Osaba, F. Diaz, E. Onieva, A. Masegosa, A. Perallos
Deusto Institute of Technology (DeustoTech), University of Deusto
Av. Universidades 24, Bilbao 48007, Spain
Citation Details:
E. Osaba, Xin-She Yang, F. Diaz, E. Onieva, A. D. Masegosa, A. Perallos,
A Discrete Firefly Algorithm to Solve a Rich Vehicle Routing Problem Modelling a
Newspaper Distribution System with Recycling Policy, Soft Computing, Published
First Online 18 March 2016. DOI 10.1007/s00500-016-2114-1
Abstract
A real-world newspaper distribution problem with recycling policy is tackled in this
work. In order to meet all the complex restrictions contained in such a problem, it
has been modeled as a rich vehicle routing problem, which can be more specifically
considered as an asymmetric and clustered vehicle routing problem with simultaneous
pickup and deliveries, variable costs and forbidden paths (AC-VRP-SPDVCFP). This
is the first study of such a problem in the literature. For this reason, a benchmark
composed by 15 instances has been also proposed. In the design of this benchmark, real
geographical positions have been used, located in the province of Bizkaia, Spain. For
the proper treatment of this AC-VRP-SPDVCFP, a discrete firefly algorithm (DFA)
has been developed. This application is the first application of the firefly algorithm
to any rich vehicle routing problem. To prove that the proposed DFA is a promising
technique, its performance has been compared with two other well-known techniques:
an evolutionary algorithm and an evolutionary simulated annealing. Our results have
shown that the DFA has outperformed these two classic meta-heuristics.
1
Introduction
Transportation is an important factor for today’s smart society. Public transportation, for
example, is used by almost the whole population, and it directly affects the quality of life.
However, modelling and planning such complex transport system is a very challenging task.
Here, this paper focuses on another sort of transport: the transportation in the business
world. Due to the rapid advancement of technologies, logistics systems have become very
important for media companies. The fact that anyone in the world can be well connected has
1
led to complex transport networks that are very demanding and are becoming increasingly
important. Therefore, an efficient logistic network can make a huge difference for companies
and relevant business operations.
To cite one fact that highlights the importance of logistics in this sector, in some
businesses, such as groceries delivery, the distribution costs can lead to an increase in the
product price up to 70% [1]. Thanks to cases like this, it is obvious to show the importance
of this sector.
Therefore, this paper focuses on a real-world logistic problem, and its effective treatment.
The real-world situation dealt in this work is related to the daily delivery of newspapers.
Specifically, the object of this study is a medium-sized newspaper distribution company,
which offers its services at the regional level. This company has certain mandatory
principles, which must be taken into account when performing the daily planning tasks
of deliveries. One of these principles is a strict recycling system. Another one is the
treatment of the different towns, or cities, as separate units. Besides that, the company
considers certain factors for the scheduling process, as variable travel times (depending on
the hour they are carried out), or the transit prohibition in certain streets. Although this
paper is focused on a specific company located in Bizkaia (Spain), it is noteworthy that
the goal of this work is to develop a model that is applicable to every company of similar
characteristics.
Therefore, the aim of this paper is to address efficiently this Newspaper Distribution
System with Recycling Policy (NDSRP). For this purpose, the NDSRP has been modeled
as a Multi-Attribute, or Rich-Vehicle Routing Problem (R-VRP). Nowadays, as indicated
in [2] or [3], R-VRPs form a hot topic in the scientific community. These kinds of problems
are special cases of the well-known conventional vehicle routing problem (VRP) [4], with the
distinction of having multiple constraints and complex formulations. This sort of problems
can have a great scientific interest, since such NP-Hard problems present a tough challenge
to solve. Furthermore, their social interest is also high, as their applicability to real-world
situations is greater than the conventional versions of routing problems.
To be more specific in the present paper, an Asymmetric and Clustered Vehicle Routing
Problem with Simultaneous Pickup and Deliveries, Variable Costs and Forbidden Paths
(AC-VRP-SPDVCFP) was proposed to address the presented NDSRP. Some examples of
recently developed R-VRPs can be [5] or [6]. In the former work, the authors propose
an R-VRP with hard and soft time windows, heterogeneous fleet, customers priorities and
vehicle-customer constraints. The solution proposed by the authors of this paper has already
been integrated into the optimization tool of a fleet management system used in the Canary
Islands. In the latter work, an R-VRP was proposed to deal with the perishable food
management. In this work, a heterogeneous fleet site-dependent VRP with multiple time
windows was presented. Another example of recently developed R-VRP can be the proposed
in [7]. In their paper, an R-VRP was developed for the olive oil collection in Tunisia. The
R-VRP designed in this work was a multi-product, multi-period and multi-compartment
VRP. These were some of the examples that justified the increasing interest of R-VRP in the
scientific community. For further information on R-VRP, readers can refer to the surveys
[2] and [3].
Though some appropriate methods can be found in the literature to address such
complex problems, arguably the most successful techniques to solve these R-VRPs are the
heuristics and metaheuristics [8]. In this study, the attention has been focused in the last
ones. In fact, to solve the AC-VRP-SPDVCFP proposed in this paper one metaheuristic
has been developed. Some classical examples of metaheuristics are the Tabu search [9]
2
and simulated annealing (SA) [10] as local search methods, and genetic algorithm (GA)
[11, 12], ant colony optimization [13] and particle swarm optimization [14] as population
ones. Though such methods were proposed some decades ago, they still attract the attention
in the scientific community [15, 16, 17] and metaheuristics, especially new metaheuristic
algorithms and their proper implementations still form a hot topic in the field. In fact,
many different metaheuristics have been proposed in the last two decades, which have been
successfully applied to various problems and fields. Some examples of these techniques are
the imperialist competitive algorithm, presented in 2007 by Gargari and Lucas [18], the
bat algorithm, proposed by Yang in 2010 [19], or the harmony search, presented in 2001 by
Geem et al. [20].
Another type of technique that shows a good performance applied to routing problems
are the memetic algorithms [21, 22]. In [23], for example, a hybrid genetic algorithm is
presented for solving a large class of vehicle routing problems with time-windows. On the
other hand, in [24] a hybrid genetic algorithm is presented to solve an R-VRP composed
by multidepot and periodicity features. In [25] a decomposition based memetic algorithm
is proposed for a multi-objective VRP with time windows. Another kind of R-VRP is
solved by a hybrid techniques in [26]. In this case, the characteristics of the problem are
clustered backhauls and 3D loading constraints. Finally, a multiperior VRP with profit is
addressed by a memetic algorithm in [27]. Works cited in this paragraph are some recent
and interesting examples of the whole literature, some other recent examples can be found
in, for example, [28, 29, 30].
Therefore, in this paper one metaheuristic proposed a few years ago is used to solve the
presented problem. This technique is the firefly algorithm (FA), proposed by Yang in 2008
[31]. This nature-inspired algorithm is based on the flashing behaviour of fireflies, which
acts as a signal system to attract other fireflies. As can be seen in several surveys [32, 33],
the FA has been successfully applied to many different optimization fields and problems
since its proposal. In addition, it still attracts a lot of interests in the current scientific
community [34, 35, 36]. Nevertheless, the FA has never been applied to any R-VRP. This
lack of works, along with the growing scientific interest in bio-inspired algorithms, and the
good performance shown by the FA since its proposal in 2008, has motivated its use in this
study.
In this way, the main novelties and contributions of this paper can be listed in the
following way. It is noteworthy that these originalities and contributions have motivated
the realization of this work:
• In this paper, a DPRP is dealt with using an R-VRP. As will be mentioned in the
following section, the problem of newspaper distribution has been previously addressed
in the literature, but never using an R-VRP as complex as the one presented in this
paper. Furthermore, it is the first time that an R-VRP of this characteristics is
proposed in the literature. For this reason, a benchmark composed by 15 instances
has been developed, based on real geographic locations.
• In this work, a discrete version of the FA (DFA) is presented for solving the proposed
R-VRP. Until the time of writing, the FA has never been applied to any R-VRP.
Anyway, the main novelty of the presented DFA is not only its application field. The
technique developed in this work uses the Hamming Distance function to measure
the distance between two fireflies of the swarm. This approach has been rarely
used previously, as well as the move strategy used by the fireflies, which is based
3
on evolution strategies. All these characteristics are described in following sections.
Additionally, in order to prove that the DFA is a promising technique to solve the
designed AC-VRP-SPDVCFP, the results obtained by the DFA are compared with
the ones obtained by the evolutionary algorithm (EA) based on mutations and the
evolutionary simulated annealing (ESA) [37].
The rest of the paper is organized as follows. In Section 2, the real-world problem
addressed in this work is described. In Section 3, the R-VRP proposed to deal with the real
problem is described in detail. Furthermore, in Section 4, the developed DFA is described
in detail. After that, the experimentation carried out is detailed in Section 5, as well as the
benchmark designed for the presented R-VRP. This paper then concludes with some brief
conclusions and further research topics in Section 6.
2
Newspapers Distribution with Recycling Policy
As has been pointed in the previous section, the real-world situation faced in this work is
related to the newspaper distribution. More specifically, the object of study is a mediumsized newspaper distribution company. The area of the coverage of this company is at a
provincial level. This means that the company has to serve a set of customers geographically
distributed in separate towns and cities. The company in question has some principles,
which are the base of their logistics planning. The first principle is to treat towns, or cities,
as separate units. In this way, if one vehicle enters a city, or a town, it is forced to serve each
and every customer located therein. Therefore, one vehicle is banned for entering one city
or town whether it does not have sufficient capacity to meet the demand of all customers
deployed there.
On the other hand, due to the current environmental requirements, the company has
a simple but robust paper recycling policy. In this case, the objects to recycle are the
newspapers that were not sold the previous day. Thus, as can be deduced, vehicles not only
have to meet the delivery demands of the customers. Besides that, they have to collect at
each point those newspapers that have not been sold the day before.
In addition, the company has to take into account certain factors in the routes planning
process. The first one is related to the hours at which the deliveries and collections are
done. The service is performed daily during morning from 6:00am to 15:00. Within this
time window exists one range, established between 8:00am and 10:00am, considered as
“peak hours”. In this way, traveling costs from one point to another are greater if they are
performed at “peak hours”. Besides this, with the aim of respecting all the traffic rules,
vehicles cannot go through prohibited roads.
Throughout the past few decades, the problem of the newspaper delivery has been
addressed many times in the literature. In 1996, Ree and Yoon presented a two-stage
heuristic for a newspaper distribution problem, taking as a case study of a major press
company of Korea [38]. The problem used in this study is a multi-VRP with time windows,
which is faced in two different stages by the proposed heuristic. On the other hand, in [39] a
free newspaper delivery problem is addressed. In their study, the problems was decomposed
in two phases. In the first stage, the delivery plan was created, based on concept taken from
the Inventory Routing Problem [40]. Once the delivery plan was fixed, the resulting problem
was solved in the second stage as a variant of the conventional VRP with time windows
(VRPTW) [41]. They used a heuristic for solving this second stage problem. The same
VRPTW was used in [42] to face a case study of the newspaper delivery problem focused
4
in the city of Bangkok and a modified sweep algorithm was presented to solve the proposed
VRPTW. These papers are some of the examples that can be found in the literature. Many
others can be found, considering, for example, the problem of delivering newspapers to
private subscribers [43, 44].
Regarding the problem addressed in this paper, some originalities and novelties are
proposed: the features of the variable travel times and forbidden paths have never been
used for the newspaper distribution system, as well as the recycling policy applied in this
work. In addition, as we will see in the rest of the paper, the R-VRP proposed in this work
has a great number of constraints, making easier its application to the real world. Finally,
the technique developed in this work solves the proposed AC-VRP-SPDVCFP in only one
phase, in contrast with the approaches presented in [38, 39]. This is an advantage because
solving the problem in only one phase, the runtimes and the computational effort can be
considerably reduced.
3
Asymmetric and Clustered Vehicle Routing Problem with
Simultaneous Pickups and Deliveries, Variable Costs and
Forbidden Paths
As has been stated earlier in this paper, the real-world situation of the newspaper
distribution has been modeled as an R-VRP. In this section, a detailed description of the
presented problem is provided. First of all, in Section 3.1, the basic features of the problem
are detailed, followed by the mathematical formulation of the proposed R-VRP is depicted
in Section 3.2.
3.1
General description of the problem
The proposed R-VRP has the following general characteristics. It is noteworthy that all
these features proposed here are to take into account the conditions stated in Section
2. Besides those, some additional restrictions have been considered with the intention
of developing a model closer to real conditions.
1. Asymmetry: The traveling costs in the proposed R-VRP are asymmetric. This means
that the traveling cost from any i node to another j node is different from the reverse
trip cost. It is important to highlight that, in the problem proposed in this paper,
this asymmetry appears in every node-to-node travel. This feature is not common in
most routing problems that can be found in the literature, and it brings realism to the
problem. Anyway, asymmetric costs have been applied previously in the literature in
a different context [45, 46, 47]. It is noteworthy that this feature is very valuable in
real-world applications.
2. Clusterized : With this attribute, the different clients that make up the system are
grouped in different sets or clusters. In this case, each cluster represents a city or
town. The condition that vehicles must meet is the following: if one vehicle meets the
demand of any customer belonging to any cluster, this vehicle must serve each and
every customer of this cluster. In this way, one vehicle cannot enter a cluster if it does
not have enough capacity to serve all customers in this city or town. This feature has
been used previously in many studies in a different context. [48, 49, 50].
5
3. Simultaneous Pickup and Delivery: This property is an adaptation of the often used
pickup and delivery system of some routing problems [51, 52]. Basically, this system
consists in the existence of two types of nodes: the delivery nodes and the collect
nodes. The former ones are those points in where newspapers are delivered. On the
other hand, in collect nodes, newspapers are collected, with the intention of bringing
them back to the center for their subsequent recycling.
In addition, it is important to highlight that, due to the simultaneous nature of
the real-world situation, one customer can ask for both delivery and collection of
newspapers. In this way, both delivery nodes and delivery-collect nodes can be found
in the system. On the other hand, it is assumed that all customers request the delivery
of newspapers. For these reason, customers which demand only the collection of
newspapers have not been taken into account.
4. Variable Travel Times: In real transportation situations, the travel between two points
does not always take the same cost, either temporarily or economically. In many cases,
this cost is subject to some external variables, such as the hour, the traffic or the
weather. With the intention of adding more realism to the problem, this situation has
been represented in the R-VRP proposed in this work. To this end, it has established a
schedule between 6:00am and 15:00. Within this schedule, two different time-periods
have been established, called “peak hours” and “off-peak hours”. Peak hours are
composed by two hours, between 8:00am and 10:00am. All trips performed in this
time window imply a higher cost, comparing with the costs of conducting the same
trip in “off-peak” period. A similar characteristic has been previously used a few
times in the literature [53].
5. Forbidden Paths: In the real world, it is common to find one-way roads, where the
traffic in a particular direction is prohibited. There are also pedestrian streets, where
vehicles cannot go through. With the aim of recreating this common situation, the
problem has certain arcs (i, j) which cannot be used in the final solution. A similar
philosophy has been used previously in some studies in the literature [54].
With the above assumptions and simplifications, the proposed AC-VRP-SPDVCFP is
an R-VRP, whose objective is to find a set of r routes, trying to minimize the total cost of
the complete solution, and taking into account the two different kinds of nodes, respecting
the restrictions of the clusters and vehicles capacities (Q) and not traveling through any
forbidden path. As can be seen, being an R-VRP problem, the AC-VRP-SPDVCFP has
multiple constraints. It is important to point out that all these restrictions reduce the size
of the search space which encompasses the feasible solutions. Anyway, all these constraints
make the process of generating feasible solutions and successors to be a very complex task.
In Figure 1, a possible 14-noded instance of the proposed AC-VRP-SPDVCFP is
represented. A feasible solution for this instance is also shown in Figure 1.
3.2
Mathematical description of the problem
The presented AC-VRP-SPDVCFP can be defined on a complete graph G = (V, A) where
V = {v0 , v1 , . . . , vn } is the set of vertices which represent the nodes of the system. On
the other hand, A = {(vi , vj ) : vi , vj ∈ V, i 6= j} is the set of arcs which represent
the interconnections between nodes. Each arc has an associated cost dij . Due to the
asymmetry feature, dij 6= dji . It is noteworthy that, in this case, the cost of traveling
6
Figure 1: Possible AC-VRP-SPDVCFP instance composed by 14 nodes, and one feasible
solution. Gray paths represent forbidden arcs.
from i to j is always different from the cost of traveling from j to i. The cost of traveling
via a forbidden arc is infinite. In this way, it is ensured that they will not appear in
the final solution. Furthermore, the vertex v0 represents the depot, and the rest are the
visiting points. Additionally, V is divided into c + 1 mutually exclusive nonempty subsets,
C = {V0 , V1 , ..., Vc }, each one for each cluster. These subsets hold the following conditions:
V = V0 ∪ V1 ∪ ... ∪ Vc
Va ∩ Vb = ∅,
a, b ∈ 0, 1, ..., c, a 6= b
(1)
(2)
It is noteworthy that V0 only contains v0 , which depicts the depot. The remaining
n nodes are divided into c clusters. Besides that, each node i has assigned two kinds of
demands: one associated with the delivery di > 0 and the other with the pick-up pi ≥ 0.
The proposed AC-VRP-SPDVCFP can be mathematically formulated in the following
way. It is important to highlight that yij depicts the demand picked-up in clients routed up
to node i (including node i) and transported in arc (i, j). On the other hand, zij represents
the demand to be delivered to clients routed after node i and transported in arc (i, j) [55].
Furthermore, wsr , is a binary variable, which takes 1 value whether vehicle r enters the
cluster s, and 0 in other case. Finally, the binary variable xrij is 1 if the vehicle r uses the
arc (i, j), and 0 otherwise.
The main problem is now to minimize:
n X
n X
k
X
dij xrij
(3)
i=0 j=0 r=1
where
xrij ∈ {0, 1},
i, j = 0, . . . , n; i 6= j; r = 1 . . . k,
wsr ∈ {0, 1},
yij ≥ 0,
r = 1, . . . , k; s = 1, ..., c,
i, j = 0, . . . , n,
7
(4)
(5)
(6)
zij ≥ 0,
i, j = 0, . . . , n.
(7)
This is subject to the following constraints:
n X
k
X
xrij = 1,
j = 0, . . . , n; i 6= j,
(8)
xrij = 1,
i = 0, . . . , n; j 6= i,
(9)
i=0 r=1
n X
k
X
j=0 r=1
n X
k
X
xr0j = k,
(10)
xri0 = k,
(11)
j=0 r=1
n X
k
X
i=0 r=1
n
X
xrij −
i=0
n
X
n
X
xrjl = 0, j = 0, . . . , n; r = 1 . . . k,
(12)
xrli = 0, i = 0, . . . , n; r = 1 . . . k,
(13)
l=0
xrij −
j=0
n
X
l=0
n
X
yji −
n
X
yij = pj ,
j = 0, . . . , n,
(14)
zij = dj ,
j = 0, . . . , n,
(15)
i, j = 0, ..., n,
(16)
dij xrij < ∞,
j = 0, . . . , n; i 6= j,
(17)
dij xrij < ∞,
i = 0, . . . , n; j 6= i,
(18)
i=0
i=0
n
X
n
X
zji −
i=0
i=0
yij + zij ≤ Q
k
X
xrij ,
r=1
n X
k
X
i=0 r=1
n X
k
X
j=0 r=1
k
X
wsk = 1 s = 1, ..., c.
(19)
r=1
The first clause represents the objective function, which is the sum of the costs of all
routes of the solution, and it must be minimized. The formulas (4), (5), (6) and (7) depict
the nature of the variables xrij , wsr , yij and zij . Equations (8) and (9) assure that all the
nodes are visited exactly once. On the other hand, constraints (10) and (11) ensure that
the total amount of vehicles leaving the depot, and the number of vehicles that return to it
is the same. Besides, the correct flow of each route is ensured by functions (12) and (13).
Additionally, restrictions (14) and (15) guarantee that the flows for the collection and
delivery demands, respectively, are properly conducted. These formulas ensure that both
8
demands are satisfied for every customer. Furthermore, constraint (16) assures that the
total capacity of any vehicle is never exceeded; it also establishes that pick-up and delivery
demands will only be transported using arcs included in the solution [55].
On the other hand, functions (17) and (18) ensure that every trip between one node
and another has not an infinite cost. Thus, it is guaranteed that forbidden paths will not
form part of the final solution. Finally, restriction (19) assures that only one vehicle enters
every cluster. This function, together with the above mentioned (8) and (9), ensures that
all the customers belonging the same cluster are visited by the same vehicle.
4
Firefly algorithm
As has been stated in the introduction, a discrete firefly algorithm (DFA) is proposed in
this work to address the designed AC-VRP-SPDVCFP. In this section, the description of
the classic FA is shown first (Section 4.1). Then, the proposed DFA is described in detail
in Section 4.2.
4.1
Classic Firefly Algorithm
The basic FA was first developed by Xin-She Yang in 2008 [31, 56], and it was based on the
idealized behaviour of the flashing characteristics of fireflies. To properly understand the
algorithm, it is important to highlight the following three idealized rules [31]:
• All the fireflies of the swarm are unisex, and one firefly will be attracted to other ones
regardless of their sex.
• Attractiveness is proportional to the brightness, which means that, for any two fireflies,
the brighter one will attract the less bright one. The attractiveness decreases as the
distance between the fireflies increases. Furthermore, if one firefly is the brightest one
of the swarm, it moves randomly.
• The brightness of a firefly is directly determined by the objective function of the
problem under consideration. In this manner, for a maximization problem, the
brightness can be proportional to the objective function value. On the other hand,
for a minimization problem, it can be the reciprocal of the objective function value.
In Algorithm 1, the pseudocode of the basic FA is shown, which was proposed by Yang
in [31]. In line with this, there are three important factors to consider in the FA: the
attractiveness, the distance and the movement. In the basic version of the FA these factors
are tackled in the following way. First of all, the attractiveness of a firefly is determined by
its light intensity, and it can be calculated using this formula:
β(r) = β0 e−γr
2
(20)
On the other hand, in the basic FA the distance rij between any two fireflies i and j is
calculated using the Cartesian distance, and it is computed by the following equation:
v
u d
uX
rij = ||Xi − Xj || = t (Xi,k − Xj,k )2
(21)
k=1
9
Algorithm 1: Pseudocode of the basic FA.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Define the objective function f (x);
Initialize the firefly population X = x1 , x2 , ..., xn ;
Define the light absorption coefficient γ;
for each firefly xi in the population do
Initialize light intensity Ii ;
end
repeat
for each firefly xi in the swarm do
for each other firefly xj in the swarm do
if Ij > Ii then
Move firelfy xi toward xj ;
end
Attractiveness varies with distance r via exp(-γr);
Evaluate new solutions and update light intensity;
end
end
Rank the fireflies and find the current best;
until termination criterion reached ;
Rank the fireflies and return the best one;
where Xi,k is the kth component of the spatial coordinate Xi of the ith firefly. Finally, the
movement of a firefly i toward any other brighter firefly j is determined by this formula:
2
Xi = Xi + β0 e−γrij (Xj − Xi ) + α(rand − 0.5)
(22)
where α is the randomization parameter and rand is a random number uniformly distributed
in [0,1]. On the other hand, the second term of the equation stems from the attraction
assumption.
4.2
The proposed Discrete Firefly Algorithm
It is noteworthy that the original FA was developed primarily for solving continuous
optimization problems. For this reason, the classic FA cannot be applied directly to solve
the proposed AC-VRP-SPDVCFP. Therefore, some modifications of the original FA are
needed in order to prepare it for addressing such AC-VRP-SPDVCFP problem.
First of all, in the proposed DFA, each firefly in the swarm represents a possible and
feasible solution for the AC-VRP-SPDVCFP. All the fireflies are initialized randomly.
Additionally, as has been detailed in Section 3, the sum of the costs of all routes of the
solution has been used as the objective function. Therefore, the AC-VRP-SPDVCFP is a
minimization problem, in which the fireflies with a lower objective function value are the
most attractive ones. In addition, the concept of light absorption is also represented in this
version of the FA. In this case, γ = 0.95, and this parameter is used in the way as can be
seen in Equ. (22). This parameter has been set following the guidelines proposed in several
studies of the literature [56, 31].
Besides that, the distance between two different fireflies is represented by the Hamming
Distance. The Hamming distance between two fireflies is the number of non-corresponding
10
elements in the sequence. In the proposed AC-VRP-SPDVCFP, the comparison is made
cluster by cluster. For example, taking into account two random fireflies, and one random
cluster c composed by 8 nodes:
x1 (cluster-c) : {0, 1, 2, 3, 4, 5, 6, 7},
x2 (cluster-c) : {0, 1, 3, 2, 5, 4, 6, 7},
the Hamming Distance between x1 and x2 for the cluster k would be 4. This same
comparison is made for every cluster. In this way, the total distance between fireflies i
and j is the sum of all the distances for every cluster.
Finally, the movement of a firefly i attracted to another brighter firefly j is determined
by
n = Random(2, rij · γ g )
(23)
xi = InsertionFunction(xi , n)
(24)
where rij is the Hamming Distance between firefly i and firefly j, and g is the iteration
number. In this case, the length of the movement of a firefly will be a random number
between 2 and rij · γ g . As for the movement function, the Insertion Function has been
used. This function selects and extracts one random node from a random route. After
that, this node is re-inserted in a random position inside the cluster of the selected node.
This function takes into account the capacity constraint, in order not to create infeasible
solutions.
Following the same philosophy as other previously developed and published FA for the
Traveling Salesman Problem [57, 58], fireflies in the proposed DFA do not have directions to
move. Instead, fireflies move using evolution strategies. In this way, each firefly moves using
n times the Insertion Function, generating n potential successors. After these n movements,
the best one is performed, generating the new firefly.
The pseudocode of the presented DFA is shown in Algorithm 5. In lines 1-3 the
initialization phase of the algorithm is carried out, in which fireflies are initialized and
evaluated. Besides, the γ parameter is initialized as previously described. In addition,
in lines 10-12 the movement process is performed. In line 10 the distance between the
selected xi and xj is calculated via Hamming Distance. Once the distance is obtained, the
n parameter is calculated, which is a random number between 2 and rij · γ g . Finally, the
movement is performed in line 12 using the Insertion Function as explained before. After
this movement process, fireflies are evaluated in line 14 and ranked in line 17. This iterative
procedure is repeated until termination criterion is reached.
5
Experimentation
The computational experiments carried out in this study are described in this section.
First of all, the details of the proposed benchmark for the AC-VRP-SPDVCFP are detailed
(Section 5.1). After that, the results obtained by the developed DFA for this benchmark
are presented (Section 5.2). In order to prove that the DFA is a promising metaheuristic
for solving routing problems, the results obtained by the DFA have been compared with
the ones obtained by the EA and the ESA. In addition, two different statistical tests have
been conducted, with the aim of obtaining rigorous and fair conclusions (Section 5.3).
11
Algorithm 2: Pseudocode of the proposed DFA.
Initialize the firefly population X = x1 , x2 , ..., xn ;
γ = 0.95;
for each firefly xi in the population do
Ii = f itness(xi );
end
repeat
for each firefly xi in the swarm do
for each other firefly xj in the swarm do
if Ij < Ii then
rij = HammingDistance(xi ,xj );
n = Random(2,rij · γ g );
xi = InsertionFunction (xi ,n);
end
Evaluate new solutions and update light intensity;
end
end
Rank the fireflies and find the current best;
until termination criterion reached ;
Rank the fireflies and return the best one;
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
5.1
The proposed benchmark for the VRP
The type of R-VRP proposed in this paper has never been treated before in the literature.
It is for this reason that there is no benchmark available in the literature for the AC-VRPSPDVCFP. In this work, a benchmark composed by 15 instances is proposed. At the same
time, these instances are composed of 50 to 100 nodes. Every node represents a customer,
and all nodes are placed in real geographical locations, which are located in the province of
Bizkaia, Spain. In addition, the maximum number of clusters has been established as 10,
existing also instances with 5, and 8 of them. In the Figure 2, a map with the geographical
locations of the depot, the customers and the clusters are shown. This map has been made
using Open Streep Maps technology, via uMap tool1 .
The clusters have been organized in order of appearance. That is, the nodes 1-10
compose cluster 1, nodes 11-20 for cluster 2, and so on. It is important to highlight that
every cluster has the same number of customers in all instances. Besides that, as has been
pointed out in Section 3.1, each customer has assigned two kinds of demands: one associated
with the delivery di and the other with the pick-up pi . The assignments of these demands
have been performed following this procedure:
1
di = 10, pi = 5,
∀i ∈ {1, 5, 9, . . . , 97},
(25)
di = 10, pi = 0,
∀i ∈ {2, 6, 10, . . . , 98},
(26)
di = 5, pi = 3,
∀i ∈ {3, 7, 11, . . . , 99},
(27)
di = 5, pi = 0,
∀i ∈ {4, 8, 12, . . . , 100},
(28)
http://umap.openstreetmap.fr
12
Figure 2: Geographical locations of the depot, customers and clusters around the province
of Bizkaia. Source: Open Street Maps, via uMap, accessed Sept 2015.
Algorithm 3: Procedure of travel costs assignment for “off-peak” period.
1
2
3
4
5
6
7
8
9
10
for ∀i ∈ {1, 2, . . . , 99} do
for ∀j ∈ {i + 1, . . . , 100} do
dij = EuclideanDistance(i,j);
if j is an odd number then
dji = EuclideanDistance(j,i) · 1.2 ;
else
dji = EuclideanDistance(j,i) · 0.8 ;
end
end
end
where d0 =0 and p0 =0, taking into account that v0 is considered the depot. In addition, the
costs of traveling from any customer i to other customer j have been established following
the procedure depicted in Algorithm 3. By this method, the asymmetry characteristic is
met. It is important to highlight that these costs are assigned for the “off-peak” period.
These costs are incremented when they are conducted on the “peak” period, following the
procedure shown in Algorithm 4. In is noteworthy that, with the intention of simplifying
the complexity of the problem, the traveling time between any node i and j is the same as
its traveling cost (in seconds).
Finally, depending on the instance, some paths are chosen in each cluster to be
forbidden. In Table 1, the characteristics of all the instances developed for the benchmark
are summarized. In order to understand the content of this table correctly, the following
clarifications should be made: Osaba 50 1 1 and Osaba 50 1 2 are comprised by 5 clusters,
which are the clusters {1, 3, 5, 7, 9}. On the other hand, Osaba 50 2 1 and Osaba 50 2 2 are
made up by clusters {2, 4, 6, 8, 10}. Moreover, clusters in Osaba 50 1 3 and Osaba 50 1 4
are composed by 5 nodes. In this case, this customers are the first 5 of every cluster. This
is in contrast with what happens in Osaba 50 2 3 and Osaba 50 2 4, where the 10 clusters
are comprised of the last 5 clients of every of them. Lastly, to complete all Osaba 80 X
13
Algorithm 4: Procedure of travel costs assignment for “peak” period.
1
2
3
4
5
6
7
8
9
10
for ∀i ∈ {1, 2, . . . , 99} do
for ∀j ∈ {i + 1, . . . , 100} do
dij = EuclideanDistance · 1.3;
if j is an odd number then
dji = (EuclideanDistance(j,i) · 1.2) · 1.2 ;
else
dji = (EuclideanDistance(j,i) · 0.8) · 1.4 ;
end
end
end
Table 1: Summary of the benchmark proposed for the AC-VRP-SPDVCFP. Forbidden paths
depicts the number of forbidden arcs in each cluster.
Instance
Nodes Clusters Vehic. capacity Forbidden paths
Osaba 50 1 1
50
5
240
5
Osaba 50 1 2
50
5
160
10
Osaba 50 1 3
50
10
240
5
Osaba 50 1 4
50
10
160
10
Osaba 50 2 1
50
5
240
5
Osaba 50 2 2
50
5
160
10
Osaba 50 2 3
50
10
240
5
Osaba 50 2 4
50
10
160
10
Osaba 80 1
80
8
240
5
Osaba 80 2
80
8
160
10
Osaba 80 3
80
10
240
5
Osaba 80 4
80
10
160
10
Osaba 100 1
100
10
140
5
Osaba 100 2
100
10
260
10
Osaba 100 3
100
10
320
10
instances, the first 8 clusters, or nodes (depending on the case) have been chosen.With the
aim of allowing the replication of this experimentation, it is noteworthy that the benchmark
developed is available under request to the corresponding author of this paper, or via Web2
5.2
Results
All the tests conducted in this work have been performed on an Intel Core i5 2410 laptop,
with 2.30 GHz and a RAM of 4 GB. Java has been used as the programming language. All
the 15 instances described in the previous section have been used in this experiment. Every
instance has been run 20 times. As has been said before, the results obtained by the DFA
are compared with the ones obtained by the EA and the ESA. The reason why these two
techniques have been used for this experimentation can be summarized as follows: First of
all, both meta-heuristics are well-known, and they have been frequently used to solve routing
problems. Proving that the DFA outperforms these two techniques can be concluded that
it is a promising technique to solve R-VRPs. On the other hand, all these three techniques
have two similarities: all of them base the movement of their individuals on a short-step
2
http://research.mobility.deustotech.eu/media/publication
SPDVCFP.rar.
14
resources/Instances Osaba AC-VRP-
Table 2: Parametrization of the EA and ESA for the proposed AC-VRP-SPDVCFP, where
−sup∆f is the difference in the objective function of the best and the worse individuals of
the initial population, and p=0.95.
Parameter
Population size
Mutation functions
Mutation prob.
Survivor func.
EA
Value
100
Insertion Function
1.0
70% Elitist - 30% Random
ESA
Parameter
Value
Population size
100
Successor Function Insertion Function
Temperature
−sup∆f /ln(p)
Cooling constant 0.95
operator, and they are easy to implement and can be adapted to solve new problems.
It is worth highlighting that, as far as possible, the same operators and similar
parameters have been used for all the algorithms implemented for the experimentation.
In this way, the intention is to conclude which algorithm obtains better results using similar
operators and a similar number of times. Furthermore, with the intention of facilitating the
replicability of this study, the parameters used for the EA and ESA are shown in Table 2. It
is worth pointing out that all the individuals are randomly generated. Additionally,Pas for
the termination criterion, every algorithm finishes its execution when there are n + nk=1 k
function evaluations without improvements in the best solution, where n is the size of the
problem. In addition, the parameters used for DFA are the described in Section 4.2. In
case of DFA, a population of 100 fireflies has also been used.
Finally, in this study, the permutation codification has been used for the representations
of the solutions. Thus, each solution X is encoded by an unique permutation of numbers,
which represents the different routes that compose that solution. Besides that, with the
aim of distinguishing the routes in one solution, they are separated by zeros.
Before starting the tests, a small study about the parametrization of the DFA is
shown. It should be taken into account that performing a comprehensive study on the
parameterization of the algorithm would be very extensive. That study has been planned
as future work, since the objective of this paper is to present the NDSRP model, and to
demonstrate that the DFA shows an adequate performance applied to an R-VRP. For this
reason, in this paper a small portion of that study is shown, in order to justify the population
size used. For that, we compare 4 different versions of the DFA, each one with a different
population size. Each version is called DF Ax , where x is the popualtion size. For this small
test, 4 different instances have been used. In Table 3 the results of this experimentation are
depicted. In this table, the results average (avg.), and average runtime (Time, in seconds)
are shown. In addition, the last row of the table represents the average ranking for each
alternative.
Table 3: Small study about the population size on the DFA
Instance
DF A25
Name
Avg. Time
Osaba 50 1 1
51945.7 14.8
Osaba 50 1 2
57398.7 15.8
Osaba 80 3
92990.39 33.7
Osaba 100 1 110206.94 51.4
Ranking
4
DF A50
Avg. Time
51561.3 25.8
56721.8 26.9
91663.8 47.3
108241.6 81.8
3
DF A100
DF A150
Avg. Time
Avg. Time
50989.5 37.9 50934.3 68.9
56203.8 35.1 56213.7 71.8
89512.0 75.3 89531.0 112.3
107799.5 112.8 107745.7 176.4
1.5
1.5
Some conclusions can be drawn if the data presented in Table 3 are analyzed. It can be
seen a slight trend of improvement in the results when the population size increases. Even
so, this fact involves a significant increase of runtime, which is not directly proportional to
15
Table 4: Results of DFA, ESA and EA for the proposed AC-VRP-SPDVCFP.
Instance
DFA
Name
Avg. S. dev. Time C. T.
Osaba 50 1 1 50989.5 234.9 37.9 26.2
Osaba 50 1 2 56203.8 203.4 35.1 25.4
Osaba 50 1 3 71730.0 1486.2 37.6 22.3
Osaba 50 1 4 78883.8 1193.4 35.7 23.9
Osaba 50 2 1 49276.0 392.4 38.6 22.3
Osaba 50 2 2 54589.6 628.1 37.8 28.5
Osaba 50 2 3 69631.4 2400.1 38.4 25.9
Osaba 50 2 4 80543.7 1512.3 36.1 22.6
Osaba 80 1
82307.8 1043.4 72.4 57.6
Osaba 80 2
89324.9 698.0 74.0 60.7
Osaba 80 3
89512.0 1414.0 75.3 60.4
Osaba 80 4 104601.9 1299.2 75.0 64.8
Osaba 100 1 107799.5 1501.8 136.1 112.8
Osaba 100 2 100522.9 1683.0 138.0 109.2
Osaba 100 3
95553.7 1470.6 139.7 105.3
Avg.
51632.6
56929.4
72298.3
79168.0
49751.3
54871.6
71352.6
81122.0
80834.2
89989.6
89431.2
105141.7
109183.6
101708.0
95641.3
ESA
S. dev. Time
479.0 31.0
459.1 31.7
1484.4 32.1
1788.8 29.4
538.3 33.0
621.0 31.8
2883.4 33.6
1675.3 32.4
1801.4 67.0
1134.4 67.6
2680.1 68.5
1963.3 69.4
1540.4 129.3
1644.1 130.4
2687.3 128.0
C. T.
Avg.
22.7 51569.3
21.0 57008.8
17.4 72490.7
20.0 79207.9
19.8 49416.8
23.7 55007.3
21.6 71286.3
19.1 81297.3
53.4 81779.6
54.3 90090.6
55.2 89883.1
61.3 106689.3
108.1 109614.4
102.0 101908.1
98.4 95893.7
EA
S. dev.
649.7
540.9
1413.4
1612.4
760.6
837.0
3037.6
1894.3
2018.4
1032.5
2949.6
1832.9
1700.7
1699.0
2472.0
Time
38.4
34.0
36.1
34.8
36.9
38.1
37.5
36.0
71.8
73.4
75.6
74.8
140.7
141.0
140.3
C. T.
26.8
26.4
21.8
24.5
20.4
27.9
25.1
23.1
58.0
59.4
59.8
65.3
116.7
111.9
108.4
the improvement in results. In this paper, to achieve the proposed, the option which best
balances the runtime and results quality has been selected. This option is DF A100 . This
version has acceptable execution times, and is the alternative that more improvement offers
regarding its previous version. DF A150 , on the other hand, needs very high execution times
without offering significant improvements in the results quality.
Once carried out this small study, the results obtained by the three techniques for the
proposed benchmark are summarized in Table 4. In this table, the results average (avg.),
standard deviation (S. dev), average runtime (Time, in seconds) and average convergence
time (C. T, in seconds) are shown. Additionally, the best results averages have been
represented bolded. Besides that, the best results found for each instance are represented
in Table 5. This table also shows the number of vehicles needed to perform every solution,
and the technique which reached each of these results. Since this is the first appearance
of the AC-VRP-SPDVCFP in the literature, these solutions are considered the best ones
found until the publication of this paper.
5.3
Analysis and Discussion
Analyzing the results in Table 4, the first conclusion that can be drawn is the following
one: DFA outperforms clearly the other algorithms in terms of results. Specifically, the
DFA performs better than ESA in 86.66% of the instance (13 out of 15), and in 93.33% of
the cases comparing with the EA (14 out of 15). Additionally, the supremacy of DFA can
also be seen in Table 5, having obtained the best solution in 80% of the instances (12 out
of 15). In relation with the data shown in Table 5, in Figure 3 and Figure 4, the partial
representation of the best solutions found by the DFA for the instances Osaba 80 2 and
Osaba 100 1 are shown. It is noteworthy that, due to the complex nature of the problem,
the complete representation of these solutions would lead to maps with some overlapping
lines, complicating the visibility of the whole solution. Therefore, these solutions are shown
at cluster-level, showing also several clusters in detail.
Another important factor that is worth mentioning is the robustness of the DFA in
16
Table 5: Best solutions found for each instance of the proposed benchmark.
Name
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
Osaba
50 1 1
50 1 2
50 1 3
50 1 4
50 2 1
50 2 2
50 2 3
50 2 4
80 1
80 2
80 3
80 4
100 1
100 2
100 3
Best Result Vehicles Technique
50641.46
2
DFA
55923.48
3
DFA
68535.19
2
DFA
76276.83
3
DFA
48819.71
2
DFA
53876.86
3
DFA
65059.03
2
DFA
77497.06
3
DFA
77660.84
3
ESA
87835.50
4
DFA
83713.72
3
ESA
101497.11
5
DFA
105165.35
5
DFA
97924.51
4
DFA
88966.48
3
ESA
Figure 3: Partial representation of the best solution found by the DFA for the instance
Osaba 80 2. Source: Open Street Maps, via uMap, accessed Sept 2015.
relation to the other techniques. It should be clarified that the robustness is the capacity
of a technique to obtain similar solutions in every run. As can be seen in Table 4, the
standard deviation of the results obtained by the DFA is lower than the ones presented
by the other metaheuristics in most instances (12 out of 15). This means that the quality
of the solutions provided by the DFA move in a narrow range. This characteristic gives
robustness and reliability to the algorithm, something crucial if the technique is applied a
real environment.
Regarding the runtimes, in Table 4 it can be seen how DFA and EA have similar
execution times, while the ESA takes less time than its competitors. These same conclusions
can be drawn looking at the convergence times, where ESA shows that it needs less time
to reach the final solution. This fact can be analyzed in the following way: The DFA needs
more time to reach its final solution, showing a better exploration capacity than the ESA.
On the other hand, it shows a better exploitation capacity than the EA, since it obtains
better results needing similar execution and convergence times.
Besides that, two different statistical tests have been conducted in order to obtain
rigorous and fair conclusions. It is important to clarify that the result averages obtained
by each technique have been use to perform these tests. The guidelines given by Derrac
et al. in [59] have been followed to perform this statistical analysis. First of all, the
17
Figure 4: Partial representation of the best solution found by the DFA for the instance
Osaba 100 1. Source: Open Street Maps, via uMap, accessed Sept 2015.
Friedman’s non-parametric test for multiple comparisons has been used to check if there
are any significant differences among all the techniques. The resulting Friedman statistic
has been 17.73. Taking into account that the confidence interval has been stated at the
99% confidence level, the critical point in a χ2 distribution with 2 degrees of freedom is
9.21. Since 17.73>9.21, it can be concluded that there are significant differences among
the results reported by the three compared algorithms, being DFA the one with the lowest
rank. Finally, regarding this Friedman’s test, the computed p-value has been 0.000141.
To evaluate the statistical significance of the better performance of DFA, the Holm’s
post-hoc test has been conducted using DFA as control algorithm. The unadjusted and
adjusted p-values obtained through the application of Holm’s post-hoc procedure can be
seen in Table 7. Analyzing this data, and taking into account that all the p-values are lower
than 0.05, it can be concluded that DFA is significantly better than ESA and EA at a 95%
confidence level.
Table 6: Average rankings returned by the Friedman’s non-parametric test for DFA, ESA
and EA.
Algorithm
DFA
ESA
EA
Average Ranking
1.2
2.0667
2.7333
Table 7: Unadjusted and adjusted p-values obtained through the application of Holm’s
post-hoc procedure using DFA as control algorithm.
Algorithm
ESA
EA
6
Unadjusted p
0.017622
0.000027
Adjusted p
0.017622
0.000054
Conclusions and Further Work
In this work, a new version of the newspaper delivery problem with recycling policy has been
tackled. This problem has been modelled as a rich vehicle routing problem, specifically,
as an asymmetric and clustered vehicle routing problem with simultaneous pickup and
deliveries, variable costs and forbidden paths. It is the first time that a problem like
18
this is presented in the literature, for this reason, a benchmark composed by 15 instances
has been also developed, using real-world geographical locations. To deal with such a
complex problem, a discrete firefly algorithm has been developed. This application can
be considered as the first application of a firefly algorithm to any rich vehicle routing
problem. To prove that the proposed DFA is a promising technique, its performance has
been compared with those obtained by two other well-known techniques: the evolutionary
algorithm, and the evolutionary simulated annealing. The DFA has outperformed these two
classic metaheuristics.
As for future work, it is intended to extend the application of the DFA to other complex
real-world situations, related to transportation and logistics. In addition, comparison
with more algorithms and techniques will be carried out. Various improvements will be
investigate so as to see if the results shown in this work for the AC-VRP-SPDVCFP can be
improved. Besides this, it is intended to conduct a thorough study on the parameterization
of the DFA (analyzing, for example, the time complexity). This study has not been done in
this work because the main objective is to present the NDSRP model, and to demonstrate
that the DFA shows an adequate performance applied to an R-VRP. Finally, it would be
useful to perform AC-VRP-SPDVCFP instances using random changes in the asymmetric
travel costs.
Acknowledgement
This project was supported by the European Unions Horizon 2020 research and innovation
programme through the TIMON: Enhanced real time services for optimized multimodal
mobility relying on cooperative networks and open data project (636220). As well
as by the projects TEC2013-45585-C2-2-R from the Spanish Ministry of Economy and
Competitiveness, and PC2013-71A from the Basque Government.
References
[1] Golden, B.L., Wasil, E.A.: Or practicecomputerized vehicle routing in the soft drink
industry. Operations research 35(1) (1987) 6–17
[2] Vidal, T., Crainic, T.G., Gendreau, M., Prins, C.: Heuristics for multi-attribute vehicle
routing problems: A survey and synthesis. European Journal of Operational Research
231(1) (2013) 1–21
[3] Lahyani, R., Khemakhem, M., Semet, F.: Rich vehicle routing problems: From a
taxonomy to a definition. European Journal of Operational Research 241(1) (2015)
1–14
[4] Toth, P., Vigo, D.: The vehicle routing problem. Society for Industrial and Applied
Mathematics (2001)
[5] de Armas, J., Melián-Batista, B., Moreno-Pérez, J.A., Brito, J.: Gvns for a real-world
rich vehicle routing problem with time windows. Engineering Applications of Artificial
Intelligence 42 (2015) 45–56
[6] Amorim, P., Parragh, S.N., Sperandio, F., Almada-Lobo, B.: A rich vehicle routing
problem dealing with perishable food: a case study. Top 22(2) (2014) 489–508
19
[7] Lahyani, R., Coelho, L.C., Khemakhem, M., Laporte, G., Semet, F.: A multicompartment vehicle routing problem arising in the collection of olive oil in tunisia.
Omega 51 (2015) 1–10
[8] Caceres-Cruz, J., Arias, P., Guimarans, D., Riera, D., Juan, A.A.: Rich vehicle routing
problem: Survey. ACM Computing Surveys (CSUR) 47(2) (2014) 32
[9] Glover, F.: Tabu search, part i. ORSA Journal on computing 1(3) (1989) 190–206
[10] Kirkpatrick, S., Gellat, C., Vecchi, M.: Optimization by simmulated annealing. science
220(4598) (1983) 671–680
[11] Goldberg, D.: Genetic algorithms in search, optimization, and machine learning.
Addison-Wesley Professional (1989)
[12] De Jong, K.: Analysis of the behavior of a class of genetic adaptive systems. PhD
thesis, University of Michigan, Michigan, USA (1975)
[13] Dorigo, M., Blum, C.: Ant colony optimization theory: A survey. Theoretical computer
science 344(2) (2005) 243–278
[14] Kennedy, J., Eberhart, R., et al.: Particle swarm optimization. In: Proceedings of
IEEE international conference on neural networks. Volume 4., Perth, Australia (1995)
1942–1948
[15] Rodriguez, A., Gutierrez, A., Rivera, L., Ramirez, L.: Rwa: Comparison of genetic
algorithms and simulated annealing in dynamic traffic. In: Advanced Computer and
Communication Engineering Technology. Springer (2015) 3–14
[16] Cao, B., Glover, F., Rego, C.: A tabu search algorithm for cohesive clustering problems.
Journal of Heuristics (2015) 1–21
[17] İnkaya, T., Kayalıgil, S., Özdemirel, N.E.: Ant colony optimization based clustering
methodology. Applied Soft Computing 28 (2015) 301–311
[18] Atashpaz-Gargari, E., Lucas, C.: Imperialist competitive algorithm: an algorithm for
optimization inspired by imperialistic competition. In: IEEE Congress on Evolutionary
Computation. (2007) 4661–4667
[19] Yang, X.S.:
A new metaheuristic bat-inspired algorithm.
cooperative strategies for optimization. Springer (2010) 65–74
In: Nature inspired
[20] Geem, Z.W., Kim, J.H., Loganathan, G.: A new heuristic optimization algorithm:
harmony search. Simulation 76(2) (2001) 60–68
[21] Moscato, P., Cotta, C.: A gentle introduction to memetic algorithms. In: Handbook
of metaheuristics. Springer (2003) 105–144
[22] Nalepa, J., Blocho, M.: Co-operation in the parallel memetic algorithm. International
Journal of Parallel Programming (2014) 1–28
[23] Vidal, T., Crainic, T.G., Gendreau, M., Prins, C.: A hybrid genetic algorithm with
adaptive diversity management for a large class of vehicle routing problems with timewindows. Computers & Operations Research 40(1) (2013) 475–489
20
[24] Vidal, T., Crainic, T.G., Gendreau, M., Lahrichi, N., Rei, W.: A hybrid genetic
algorithm for multidepot and periodic vehicle routing problems. Operations Research
60(3) (2012) 611–624
[25] Qi, Y., Hou, Z., Li, H., Huang, J., Li, X.: A decomposition based memetic algorithm for
multi-objective vehicle routing problem with time windows. Computers & Operations
Research 62 (2015) 61–77
[26] Bortfeldt, A., Hahn, T., Männel, D., Mönch, L.: Hybrid algorithms for the vehicle
routing problem with clustered backhauls and 3d loading constraints. European Journal
of Operational Research 243(1) (2015) 82–96
[27] Zhang, Z., Che, O., Cheang, B., Lim, A., Qin, H.: A memetic algorithm for the
multiperiod vehicle routing problem with profit. European Journal of Operational
Research 229(3) (2013) 573–584
[28] Nagata, Y., Bräysy, O., Dullaert, W.: A penalty-based edge assembly memetic
algorithm for the vehicle routing problem with time windows. Computers & Operations
Research 37(4) (2010) 724–737
[29] Nalepa, J., Blocho, M.: Adaptive memetic algorithm for minimizing distance in the
vehicle routing problem with time windows. Soft Computing (2015) 1–19
[30] Marinakis, Y., Marinaki, M., Spanou, P.: A memetic differential evolution algorithm
for the vehicle routing problem with stochastic demands. In: Adaptation and
Hybridization in Computational Intelligence. Springer (2015) 185–204
[31] Yang, X.S.: Nature-inspired metaheuristic algorithms. Luniver press, UK (2008)
[32] Fister, I., Yang, X.S., Fister, D., Fister Jr, I.: Firefly algorithm: A brief review of
the expanding literature. In: Cuckoo Search and Firefly Algorithm. Springer (2014)
347–360
[33] Fister, I., Fister Jr, I., Yang, X.S., Brest, J.: A comprehensive review of firefly
algorithms. Swarm and Evolutionary Computation (2013)
[34] Ma, Y., Zhao, Y., Wu, L., He, Y., Yang, X.S.: Navigability analysis of magnetic
map with projecting pursuit-based selection method by using firefly algorithm.
Neurocomputing (2015)
[35] Liang, R.H., Wang, J.C., Chen, Y.T., Tseng, W.T.:
An enhanced firefly
algorithm to multi-objective optimal active/reactive power dispatch with uncertainties
consideration. International Journal of Electrical Power & Energy Systems 64 (2015)
1088–1097
[36] Zouache, D., Nouioua, F., Moussaoui, A.: Quantum-inspired firefly algorithm with
particle swarm optimization for discrete optimization problems. Soft Computing (2015)
1–19
[37] Yip, P.P., Pao, Y.H.: Combinatorial optimization with use of guided evolutionary
simulated annealing. IEEE Transactions on Neural Networks 6(2) (1995) 290–295
21
[38] Ree, S., Yoon, B.S.: A two-stage heuristic approach for the newspaper delivery
problem. Computers & industrial engineering 30(3) (1996) 501–509
[39] Archetti, C., Doerner, K.F., Tricoire, F.: A heuristic algorithm for the free newspaper
delivery problem. European Journal of Operational Research 230(2) (2013) 245–257
[40] Campbell, A., Clarke, L., Kleywegt, A., Savelsbergh, M.: The inventory routing
problem. In: Fleet management and logistics. Springer (1998) 95–113
[41] Kallehauge, B., Larsen, J., Madsen, O.B., Solomon, M.M.: Vehicle routing problem
with time windows. Springer (2005)
[42] Boonkleaw, A., Suthikarnnarunai, N., Srinon, R.: Strategic planning and vehicle
routing algorithm for newspaper delivery problem: Case study of morning newspaper,
bangkok, thailand. In: Proceedings of the World Congress on Engineering and
Computer Science. Volume 2. (2009) 1067–1071
[43] Hurter, A.P., Van Buer, M.G.: The newspaper production/distribution problem.
Journal of Business Logistics 17 (1996) 85–107
[44] Van Buer, M.G., Woodruff, D.L., Olson, R.T.: Solving the medium newspaper
production/distribution problem. European Journal of Operational Research 115(2)
(1999) 237–253
[45] Laporte, G., Mercure, H., Nobert, Y.: An exact algorithm for the asymmetrical
capacitated vehicle routing problem. Networks 16(1) (1986) 33–46
[46] Toth, P., Vigo, D.: A heuristic algorithm for the symmetric and asymmetric vehicle
routing problems with backhauls. European Journal of Operational Research 113(3)
(1999) 528–543
[47] Herrero, R., Rodrı́guez, A., Cáceres-Cruz, J., Juan, A.A.: Solving vehicle routing
problems with asymmetric costs and heterogeneous fleets. International Journal of
Advanced Operations Management 6(1) (2014) 58–80
[48] Chisman, J.A.: The clustered traveling salesman problem. Computers & Operations
Research 2(2) (1975) 115–119
[49] Battarra, M., Erdogan, G., Vigo, D.: Exact algorithms for the clustered vehicle routing
problem. Operations Research 62(1) (2014) 58–71
[50] Vidal, T., Battarra, M., Subramanian, A., Erdogan, G.: Hybrid metaheuristics for the
clustered vehicle routing problem. Computers & Operations Research 58(1) (2014)
87–99
[51] Wang, C., Mu, D., Zhao, F., Sutherland, J.W.: A parallel simulated annealing method
for the vehicle routing problem with simultaneous pickup–delivery and time windows.
Computers & Industrial Engineering 83 (2015) 111–122
[52] Li, J., Pardalos, P.M., Sun, H., Pei, J., Zhang, Y.: Iterated local search embedded
adaptive neighborhood selection approach for the multi-depot vehicle routing problem
with simultaneous deliveries and pickups. Expert Systems with Applications 42(7)
(2015) 3551–3561
22
[53] Haghani, A., Jung, S.: A dynamic vehicle routing problem with time-dependent travel
times. Computers & operations research 32(11) (2005) 2959–2986
[54] Villeneuve, D., Desaulniers, G.: The shortest path problem with forbidden paths.
European Journal of Operational Research 165(1) (2005) 97–107
[55] Montané, F.A.T., Galvao, R.D.: A tabu search algorithm for the vehicle routing
problem with simultaneous pick-up and delivery service. Computers & Operations
Research 33(3) (2006) 595–619
[56] Yang, X.S.: Firefly algorithms for multimodal optimization. In: Stochastic algorithms:
foundations and applications. Springer (2009) 169–178
[57] Jati, G.K., et al.: Evolutionary discrete firefly algorithm for travelling salesman
problem. Volume 6943. Springer (2011)
[58] Zhou, L., Ding, L., Qiang, X.: A multi-population discrete firefly algorithm to solve
tsp. In: Bio-Inspired Computing-Theories and Applications. Springer (2014) 648–653
[59] Derrac, J., Garcı́a, S., Molina, D., Herrera, F.: A practical tutorial on the use of
nonparametric statistical tests as a methodology for comparing evolutionary and swarm
intelligence algorithms. Swarm and Evolutionary Computation 1(1) (2011) 3–18
23
| 9cs.NE
|
Improved EEG Event Classification Using Differential Energy
A. Harati, M. Golmohammadi, S. Lopez, I. Obeid and J. Picone
Neural Engineering Data Consortium, Temple University
Philadelphia, Pennsylvania, USA
{amir.harati, meysam, silvia.lopez, obeid, picone}@temple.edu
Abstract— Feature extraction for automatic classification of EEG
signals typically relies on time frequency representations of the
signal. Techniques such as cepstral-based filter banks or wavelets
are popular analysis techniques in many signal processing
applications including EEG classification. In this paper, we
present a comparison of a variety of approaches to estimating and
postprocessing features. To further aid in discrimination of
periodic signals from aperiodic signals, we add a differential
energy term. We evaluate our approaches on the TUH EEG
Corpus, which is the largest publicly available EEG corpus and an
exceedingly challenging task due to the clinical nature of the data.
We demonstrate that a variant of a standard filter bank-based
approach, coupled with first and second derivatives, provides a
substantial reduction in the overall error rate. The combination of
differential energy and derivatives produces a 24% absolute
reduction in the error rate and improves our ability to
discriminate between signal events and background noise. This
relatively simple approach proves to be comparable to other
popular feature extraction approaches such as wavelets, but is
much more computationally efficient.
I. INTRODUCTION
Electroencephalograms (EEGs) are used in a wide range of
clinical settings to record electrical activity along the scalp.
EEGs are the primary means by which neurologists diagnose
brain-related illnesses such as epilepsy and seizures [1]. We
have developed a system, known as AutoEEGTM, that
automatically interprets EEGs, and delivers high performance
on clinical data [2]. An overview of the system is shown in
Figure 1. It incorporates a traditional hidden Markov model
(HMM) based system and uses two stages of postprocessing to
produce epoch labels. An N-channel EEG is transformed into N
independent feature streams using a standard sliding window
based approach. These features are then transformed into EEG
signal event hypotheses using a standard HMM recognition
system [3]. These hypotheses are postprocessed by examining
temporal and spatial context to produce epoch labels.
discharges that can be focal or lateralized over one hemisphere.
These signals display quasi-periodic behavior. GPED events are
similar to PLEDs, and manifest themselves as periodic shortinterval diffuse discharges, periodic long-interval diffuse
discharges and suppression-burst patterns according to the
interval between the discharges. Triphasic waves, which
manifest themselves as diffuse and bilaterally synchronous
spikes with bifrontal predominance, typically at a rate of 1-2 Hz,
are also included in this class.
The system also detects three events used to model background
noise: (1) artifacts (ARTF) are recorded electrical activity that is
not of cerebral origin, such as those due to the equipment, patient
behavior or the environment; (2) eye movement (EYEM) are
common events that can often be confused with a spike; (3)
background (BCKG) is used for all other signals.
These six classes were arrived at through several iterations of a
study conducted with Temple University Hospital neurologists.
Automatic labeling of these events allows a neurologist to
rapidly search long-term EEG recordings for anomalous
behavior. Performance requirements for this application are
extremely aggressive. For the system to be clinically useful,
detection rates for the three signal classes must be at least 95%
with a false alarm rate below 5%. This is a challenge for clinical
data because the recordings contain many artifacts that can
easily be interpreted as spikes. Therefore, neurologists still rely
on manual review of data in clinical applications.
Hence, a unique aspect of the work reported here is that we have
used the TUH EEG Corpus [2] for evaluation. TUH EEG is the
world’s largest publicly available database of clinical EEG data,
comprising more than 28,000 EEG records and over 15,000
patients. It represents the collective output from Temple
Epochs are typically 1 sec in duration, while features are
computed every 0.1 secs using 0.2 sec analysis windows. These
parameters were optimized experimentally [2] in a previous
study. Neurologists review EEGs in 10 sec windows, and it is
common that pattern recognition systems classify 1 sec epochs.
We further divide these 1 sec epochs into 10 frames of 0.1 secs
each so that we can model an epoch with an HMM.
The system detects three events of clinical interest [4]: (1) spike
and/or sharp waves (SPSW), (2) periodic lateralized
epileptiform discharges (PLED), and (3) generalized periodic
epileptiform discharges (GPED). SPSW events are epileptiform
transients that are typically observed in patients with epilepsy.
PLED events are indicative of EEG abnormalities and often
manifest themselves with repetitive spike or sharp wave
Figure 1. A two-level architecture for automatic interpretation of
EEGs that integrates hidden Markov models for sequential decoding
of EEG events with deep learning for decision-making based on
temporal and spatial context.
University Hospital’s Department of Neurology since 2002 and
is an ongoing data collection project. EEG signals were recorded
using several generations of Natus Medical Incorporated’s
NicoletTM EEG recording technology. The raw signals obtained
from the studies consist of multichannel recordings that vary
between 20 and 128 channels sampled at a minimum of 250 Hz
minimum using a 16-bit A/D converter. The data is stored in a
proprietary format that has been exported to EDF with the use of
NicVue v5.71.4.2530. In our study, we have resampled all the
data to a common sample frequency of 250 Hz.
II. EEG FEATURES
Our system uses a fairly standard cepstral coefficient-based
feature extraction approach similar to the Mel Frequency
Cepstral Coefficients (MFCCs) used in speech recognition
[3],[5],[6]. Though popular alternatives to MFCCs in EEG
processing include wavelets, which are used by many
commercial systems, our experiments with such features have
shown very little advantage over MFCCs [7] on the TUH EEG
Corpus. Therefore, in this study we have focused on filter bank
approaches. Further, unlike speech recognition which uses a mel
scale for reasons related to speech perception, we use a linear
frequency scale for EEGs, since there is no physiological
evidence that a log scale is meaningful [4].
The focus of this paper is an exploration of some traditional
tuning parameters associated with cepstral coefficient
approaches. In this study, we limit our explorations to the
tradeoffs in computing energy and differential features, since
these have the greatest impact on performance.
It is common in the MFCC approach to compute cepstral
coefficients by computing a high resolution fast Fourier
Transform, downsampling this representation using an
oversampling approach based on a set of overlapping bandpass
filters, and transforming the output into the cepstral domain
using a discrete cosine transform [8],[9]. The zeroth-order
cepstral term is typically discarded and replaced with an energy
term as described below.
There are two types of energy terms that are often used: time
domain and frequency domain. Time domain energy is a
straightforward computation using the log of the sum of the
squares of the windowed signal:
!" = log
'
(
(.'
/01
)(+)
-
have introduced a differential energy term that attempts to model
the long-term change in energy. This term examines energy over
a range of M frames centered about the current frame, and
computes the difference between the maximum and minimum
over this interval:
!; = <=) !2 <
>
− <@+ !2 <
>
(2)
We typically use a 0.9 sec window for this calculation. This
simple feature has proven to be surprisingly effective.
The final step to note in our feature extraction process is the
familiar method for computing derivatives of features using a
regression approach [5],[8],[9]:
B" =
H /(C
DEF .CDGF )
FIJ
K
- H
FIJ /
(3)
where B" is a delta coefficient, from frame L computed in terms
of the static coefficients M"N/ to M"./ . A typical value for N is 9
(corresponding to 0.9 secs) for the first derivative in EEG
processing, and 3 for the second derivative. These features,
which are often called deltas because they measure the change
in the features over times, are one of the most well-known
features in speech recognition [8]. We typically use this
approach to compute the derivatives of the features and then
apply this approach again to those derivatives to obtain an
estimate of the second derivatives of the features, generating
what are often called delta-deltas. This triples the size of the
feature vector (adding deltas and delta-deltas), but is well-known
to deliver improved performance. This approach has not been
extensively evaluated in EEG processing.
Dimensionality is something we must always pay attention to in
classification systems since our ability to model features is
directly related to the amount of training data available. The use
of differential features raises the dimension of a typical feature
vector from 9 (e.g., 7 cepstral coefficients, frequency domain
energy and differential energy) to 27. There must be sufficient
training data to support this increase in dimensionality or any
improvements in the feature extraction process will be masked
by poor estimates of the model parameters (e.g., Gaussian means
and covariances). As we will show in the next section, the TUH
EEG Corpus is large enough to support such studies.
(Error! No sequence specified.)
We use an overlapping analysis window (a 50% overlap was
used here) to ensure a smooth trajectory of this features.
The energy of the signal can also be computed in the frequency
domain by computing the sum of squares of the oversampled
filter bank outputs after they are downsampled:
!2 = 345
(.'
801
6 7
-
(1)
This form of energy is commonly used in speech recognition
systems because it provides a smoother, more stable estimate of
the energy that leverages the cepstral representation of the
signal. However, the virtue of this approach has not been
extensively studied for EEG processing.
In order to improve differentiation between transient pulse-like
events (e.g., SPSW events) and stationary background noise, we
Figure 2. An illustration of how the differential energy term
accentuates the differences between spike-like behavior and noiselike behavior. Detection of SPSW events is critical to the success of
the overall system.
III. EXPERIMENTATION
No.
We have used a subset of TUH EEG that has been manually
labeled for the six types of events described in Section I. The
training set contains segments from 359 sessions while the
evaluation set was drawn from 159 sessions. No patient appears
more than once in the entire subset, which we refer to as the TUH
EEG Short Set. A distribution of the frequency of occurrence of
the 6 types of events in the training and evaluation set is shown
in Table 1. The training set was designed to provide a sufficient
number of examples to train statistical models such as HMMs.
Note that some classes, such as SPSW, occur much less
frequently in the actual corpus than common events such as
BCKG. In fact, 99% of the data is assigned to the class BCKG,
so special care must be taken to build robust classifiers for the
non-background classes. High performance detection of EEG
events requires dealing with infrequently occurring events since
the majority of the data is normal (uninformative). Hence, the
evaluation set was designed to contain a reasonable
representation of all classes.
We refer to the 6 classes shown in Table 1 as the 6-way
classification problem. This is not necessarily the most
informative performance metric. It makes more sense to collapse
the 3 background classes into one category. We refer to this
second evaluation paradigm as a 4-way classification task:
SPSW, GPED, PLED and BACKG. The latter class contains an
enumeration of the 3 background classes. Finally, in order that
we can produce a DET curve [10], we also report a 2-way
classification task in which we collapse the data into a target
class (TARG) and a background class (BCKG).
DET curves are generated by varying a threshold typically
applied to likelihoods to evaluate the tradeoff between detection
rates and false alarms. However, it is also instructive to look at
specific numbers in table form. Therefore, all experiments
reported in the tables use a scoring penalty of 0, which
essentially means we are evaluating the raw likelihoods returned
from the classification system. In virtually all cases, the trends
shown in these tables hold up for the full range of the DET curve.
A. Absolute Features
The first series of experiments was run on a simple combination
of features. A summary of these experiments is shown in
Table 2. Cepstral-only features were compared with several
energy estimation algorithms. It is clear that the combination of
frequency domain energy and differential energy provides a
substantial reduction in performance. However, note that
differential energy by itself (system no. 4) produces a noticeable
Train
Event
SPSW
GPED
PLED
EYEM
ARTF
BCKG
Total:
No.
645
6184
11,254
1,170
11,053
53,726
84,032
Eval
% (CDF)
No.
0.8% ( 1%)
7.4% ( 8%)
13.4% ( 22%)
1.4% ( 23%)
13.2% ( 36%)
63.9% (100%)
100.0% (100%)
567
1,998
4,677
329
2,204
19,646
29,421
% (CDF)
1.9% ( 2%)
6.8% ( 9%)
15.9% ( 25%)
1.1% ( 26%)
7.5% ( 33%)
66.8% (100%)
100.0% (100%)
Table 1. An overview of the distribution of events in the subset of the
TUH EEG Corpus used in our experiments.
System Description
Dims.
6-Way
4-Way
2-Way
1
Cepstral
7
59.3%
33.6%
24.6%
2
Cepstral + Ef
8
45.9%
33.0%
24.0%
3
Cepstral + Et
8
44.9%
33.7%
24.8%
4
Cepstral + Ed
8
55.2%
32.8%
24.3%
5
Cepstral + Ef +Ed
9
39.2%
30.0%
20.4%
Table 2. Performance on the TUH EEG Short Set of the base cepstral
features augmented with an energy feature. System no. 5 uses both
frequency domain and differential energy features. Note that the
results are consistent across all classification schemes.
degradation in performance. Frequency domain energy clearly
provides information that complements differential energy.
The improvements produced by system no. 5 hold for all three
classification tasks. Though this approach increases the
dimensionality of the feature vector by one element, the value of
that additional element is significant and not replicated by
simply adding other types of signal features [11].
B. Differential Features
A second set of experiments were run to evaluate the benefit of
using differential features. These experiments are summarized
in Table 3. The addition of the first derivative adds about 7%
absolute in performance (e.g., system no. 6 vs. system no. 1).
However, when differential energy is introduced, the
improvement in performance drops to only 4% absolute.
The story is somewhat mixed for the use of second derivatives.
On the base cepstral feature vector, second derivatives reduce
the error rate on the 6-way task by 4% absolute (systems no. 1,
6 and 11). However, the improvement for a system using
differential energy is much less pronounced (systems no. 5, 10
and 15). In fact, it appears that differential energy and
derivatives do something very similar. Therefore, we evaluated
a system that eliminates the second derivative for differential
energy. This system is labeled no. 16 in Table 3. We obtained a
small but significant improvement in performance over system
no. 10. The improvement on 4-way classification was larger,
which indicates more of an impact on differentiating between
No.
System Description
Dims. 6-Way 4-Way 2-Way
6
Cepstral + Δ
14
56.6% 32.6% 23.8%
7
Cepstral + Ef + Δ
16
43.7% 30.1% 21.2%
8
Cepstral + Et + Δ
16
42.8% 31.6% 22.4%
9
Cepstral + Ed + Δ
16
51.6% 30.4% 22.0%
10
Cepstral + Ef +Ed + Δ
18
35.4% 25.8% 16.8%
11
Cepstral + Δ + ΔΔ
21
53.1% 30.4% 21.8%
12
Cepstral + Ef + Δ + ΔΔ
24
39.6% 27.4% 19.2%
13
Cepstral + Et + Δ + ΔΔ
24
39.8% 29.6% 21.1%
14
Cepstral + Ed + Δ + ΔΔ
24
52.5% 30.1% 22.6%
15
Cepstral + Ef +Ed + Δ + ΔΔ
27
35.5% 25.9% 17.2%
16
(15) but no ΔΔ for Ed
26
35.0% 25.0% 16.6%
Table 3. The impact of differential features on performance is shown.
For the overall best systems (nos. 10 and 15), second derivatives do
not help significantly. Differential energy and derivatives appear to
capture similar information.
PLEDs, GPEDs and SPSW vs. background. This is satisfying
since this this feature was designed to address this problem.
The results shown in Tables 1-3 hold up under DET curve
analysis as well. DET curves for systems nos. 1, 5, 10, and 15
are shown in Figure 3. We can see that the relative ranking of
the systems is comparable over the range of the DET curves.
First derivatives deliver a measurable improvement over
absolute features (system no. 10 vs. no. 5). Second derivatives
do not provide as significant an improvement (system no. 15 vs.
no. 10). Differential energy provides a substantial improvement
over the base cepstral features.
It should be noted that user requirements for this type of
technology includes an extremely low false alarm rate.
Neurologists have expressed a need for a false alarm rate on the
order of no more than one or two per day per bed while
maintaining a detection rate of 95%. In related work we are able
to approach these levels of performance using postprocessing
steps alluded to in Figure 1. At these levels of performance, the
differences between systems becomes more significant, and the
use of second derivatives can potentially be more significant.
IV. SUMMARY
In this paper, we have essentially calibrated some important
algorithms used in feature extraction for EEG processing. We
have shown that traditional feature extraction methods used in
other fields such as speech recognition are relevant to EEGs. The
use of a novel differential energy feature improved performance
for absolute features (system nos. 1-5), but that benefit
diminishes as first and second order derivatives are included. We
have shown there is benefit to using derivatives and there is a
small advantage to using frequency domain energy.
In related research [7],[11] we are evaluating approaches based
on wavelets and other time-frequency representations.
Preliminary results seem to indicate there are no significant
benefits to these representations. Hence, in this work we have
focused on optimization of our standard approach. Future work
will focus on new feature extraction methods based on principles
of deep learning [12], discriminative training [13] and
nonparametric Bayesian models [14].
ACKNOWLEDGEMENTS
The primary funder of this research was the QED Proof of
Concept program of the University City Science Center (Grant
No. S1313). Research reported in this publication was also
supported by the National Human Genome Research Institute
of the National Institutes of Health under Award Number
U01HG008468 and the National Science Foundation through
Major Research Instrumentation Grant No. CNS-09-58854. The
TUH EEG database work was funded by (1) the Defense
Advanced Research Projects Agency (DARPA) MTO under the
auspices of Dr. Doug Weber through the Contract No.
D13AP00065, (2) Temple University’s College of Engineering
and (3) Temple University’s Office of the Senior Vice-Provost
for Research. Finally, we are also grateful to Dr. Mercedes
Jacobson, Dr. Steven Tobochnik and David Jungries of the
Temple University School of Medicine for their assistance in
developing the classification paradigm used in this study and
for preparing the manually annotated data.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
Figure 3. A DET curve analysis of feature extraction systems that
compares absolute and differential features. The addition of first
derivatives provides a measurable improvment in performance while
second derivatives are less beneficial.
[14]
T. Yamada and E. Meng, Practical Guide for Clinical Neurophysiologic
Testing: EEG. Philadelphia, Pennsylvania, USA: Lippincott Williams &
Wilkins, 2009.
A. Harati, S. Lopez, I. Obeid, M. Jacobson, S. Tobochnik, and J. Picone,
“THE TUH EEG CORPUS: A Big Data Resource for Automated EEG
Interpretation,” in Proceedings of the IEEE SPMB, 2014, pp. 1–5.
J. Picone, “Continuous Speech Recognition Using Hidden Markov
Models,” IEEE ASSP Mag., vol. 7, no. 3, pp. 26–41, Jul. 1990.
S. Sanei and J. A. Chambers, EEG signal processing. Hoboken, New
Jersey, USA: Wiley-Interscience, 2008.
J. Lyons, “Mel Frequency Cepstral Coefficient (MFCC) tutorial,”
Practical Cryptography, 2015 (available: http://practicalcryptography.
com/miscellaneous/machine-learning/guide-mel-frequency-cepstralcoefficients-mfccs/.
S. Davis and P. Mermelstein, “Comparison of Parametric Representations
for Monosyllabic Word Recognition in Continuously Spoken Sentences,”
IEEE Trans. ASSP, vol. 28, no. 4, pp. 357–366, 1980.
P. Garrit, et al., “Wavelet Analysis for Feature Extraction on EEG
Signals,” presented at Temple University CoE Res. Exp. for UG Conf.,
2015 (available at http://www.isip.piconepress.com/publications/
unpublished/conferences/2015/summer_of_code/wavelets/).
X. Huang, A. Acero, and H.-W. Hon, Spoken Language Processing: A
Guide to Theory, Algorithm and System Development. Upper Saddle
River, New Jersey, USA: Prentice Hall, 2001.
J. Picone, “Signal modeling techniques in speech recognition,” Proc.
IEEE, vol. 81, no. 9, pp. 1215–1247, 1993.
A. Martin, G. Doddington, T. Kamm, M. Ordowski, and M. Przybocki,
“The DET curve in assessment of detection task performance,” in
Proceedings of Eurospeech, 1997, pp. 1895–1898.
A. Moura, I. Obeid, and J. Picone, “Feature Extraction Methods for EEG
Event Detection,” in Temple University College of Eng. Res. Exp. for
Undergrad. Conf., 2015 (available at http://www.isip.piconepress.com/
publications/unpublished/conferences/2015/summer_of_code/features/).
J. Snoek, R. P. Adams, and H. Larochelle, “Nonparametric guidance of
autoencoder representations using label information,” J. Mach. Learn.
Res., vol. 13, no. 1, pp. 2567–2588, 2012.
D. Povey, et al., “Boosted MMI for model and feature-space
discriminative training,” Proceedings of the IEEE Int. Conf. on ASSP, Las
Vegas, Nevada, USA, pp. 4057–4060, 2008.
A. Harati and J. Picone, “A Doubly Hierarchical Dirichlet Process Hidden
Markov Model with a Non-Ergodic Structure,” submitted to the
IEEE/ACM Trans. Audio, Speech, Lang. Process., 2015.
| 1cs.CV
|
Grader variability and the importance of reference
standards for evaluating machine learning models for
diabetic retinopathy
arXiv:1710.01711v1 [cs.CV] 4 Oct 2017
Jonathan Krause1 , Varun Gulshan1 , Ehsan Rahimy2 , Peter Karth3 , Kasumi Widner1 ,
Greg S. Corrado1 , Lily Peng1? , Dale R. Webster1?
1
Google Research
2
Palo Alto Medical Foundation
3
Oregon Eye Consultants
Abstract. Diabetic retinopathy (DR) and diabetic macular edema are common
complications of diabetes which can lead to vision loss. The grading of DR is a
fairly complex process that requires the detection of fine features such as microaneurysms, intraretinal hemorrhages, and intraretinal microvascular abnormalities.
Because of this, there can be a fair amount of grader variability. There are different methods of obtaining the reference standard and resolving disagreements between graders, and while it is usually accepted that adjudication until full consensus will yield the best reference standard, the difference between various methods
of resolving disagreements has not been examined extensively. In this study, we
examine the variability in different methods of grading, definitions of reference
standards, and their effects on building deep learning models for the detection of
diabetic eye disease. We find that a small set of adjudicated DR grades allows
substantial improvements in algorithm performance. The resulting algorithm’s
performance was on par with that of individual U.S. board-certified ophthalmologists and retinal specialists.
1
Introduction
Diabetic retinopathy (DR) and diabetic macular edema (DME) are common complications of diabetes which can lead to vision loss. There are several methods for assessing
the severity of diabetic eye disease, such as the Early Treatment Diabetic Retinopathy Study (ETDRS) Grading System [1], the Scottish Diabetic Retinopathy Grading
System [2], and the International Clinical Diabetic Retinopathy (ICDR) disease severity scale [3]. The ICDR scale is one of the more commonly used grading rubrics and
consists of a 5 point grade for DR – no, mild, moderate, severe, and proliferative.
The grading of DR is a fairly complex process that requires the detection of fine
features such as microaneurysms (MA), intraretinal hemorrhages, and intraretinal microvascular abnormalities (IRMA). Because of this, there can be a fair amount of grader
variability [1,4,5]. This is not surprising as grader variability is a well-known issue with
human interpretation of imaging in other medical fields such as radiology [6] or pathology [7].
There are different methods of obtaining the reference standard and resolving disagreements between graders. One method consists of taking the majority decision from
?
Equal contribution
a group of three or more independent graders. Another method consists of having a
group of two or more graders work independently and then having a third generally
more senior grader arbitrate disagreements, with the senior graders decision serving
as the reference standard. Lastly, it is also common to have a group of three or more
graders grade independently and then discuss disagreements until there is full consensus
on the final grade. It is usually accepted that adjudication until full consensus will yield
the best reference standard, but the difference between various methods of resolving
disagreements has not been examined extensively.
Deep learning [8] is a family of machine learning techniques that allows computers to learn the most predictive features directly from images, given a large dataset of
labeled examples, without specifying rules or features explicitly. It has also been applied recently in medical imaging, producing highly accurate algorithms for a variety
of classification tasks, including melanoma [9] and diabetic retinopathy (DR) [10, 11].
Because the network is trained to predict labels that have been paired with the images, it
is imperative that the labels accurately represent the state of disease found in the image,
especially for the evaluation sets (e.g. tuning and external validation/test sets).
In this study, we examine the variability in different methods of grading, definitions of reference standards, and their effects on building deep learning models for the
detection of diabetic eye disease.
2
2.1
Methods
Datasets
In this work, we build upon the datasets used by Gulshan et al. [10] for algorithm development and validation. The development dataset used consists of images obtained from
the EyePACS clinics and three eye hospitals in India (Aravind eye hospital, Sankara
Nethralaya, and Narayana Nethralaya) among patients presenting themselves for diabetic retinopathy screening. While Gulshan et al. [10] used only macula-centered images for development, in this work we also use temporal- and disc-centered field of view
images, and in addition use more patient data acquired from the EyePACS clinics (see
Table 1 for distributions). The clinical validation datasets used in Gulshan et al. [10]
are also used as part of the development set. These consist of images from the EyePACS clinics and the publicly available Messidor 2 dataset [12, 13]. All images were
deidentified according to Health Insurance Portability and Accountability Act Safe Harbor prior to transfer to study investigators. Ethics review and institutional review board
exemption was obtained using Quorum Review IRB.
For clinical validation, further macula-centered images were acquired from the EyePACS clinics (imaged between May-October 2015). A variety of cameras were used,
including Centervue DRS, Optovue iCam, Canon CR1/DGi/CR2, and Topcon NW using 45 fields of view. EyePACS images were acquired as a part of routine clinical care
for diabetic retinopathy screening. These did not overlap with any images used in the
development dataset.
2
2.2
Grading and Adjudication
There were three kinds of grades used in the development dataset: grading by ophthalmologists, grading by the EyePACS graders, and adjudicated grading by multiple
retina specialists. The first grading source was the same as in Gulshan et al. [10], where
each image was independently graded by ophthalmologists (post-residency or trainees
in their 4th year). The second source of grades was provided by the EyePACS clinics,
graded according to the EyePACS protocol [14]. In the EyePACS protocol, all three
images of an eye are graded together and assigned a single grade. For development, we
use the grade provided for the entire eye as a grade for each image – while this yielded
a noisy label, we still found it helpful in training the algorithm (see Results). These
grades were available for all images in the development set obtained from EyePACS.
The third type of grades available were obtained on a small subset of images (referred
to as the tuning set), using an adjudication protocol. The tuning set consisted of 1,989
images from the EyePACS clinics and 1,748 images from the Messidor-2 set. The images were first independently graded by three fellowship-trained retina specialists and
then all disagreements resolved by face-to-face discussions. These adjudicated grades
were only used for tuning the algorithm hyperparameters (e.g. image resolution, learning rate) and making model choices (e.g. network architectures), but not for training the
model parameters.
The test set was graded using the same adjudication protocol the tuning set (faceto-face adjudication by three retina specialists). In addition, the test set was graded by
three ophthalmologists (distinct from the retina specialists who adjudicated the test set),
in order to measure the performance of the ophthalmologists relative to the algorithm in
an unbiased setting. After grading, all disagreements between the adjudicated consensus of retinal specialists and majority decision of the ophthalmologists were reviewed
manually by a retinal specialist, who also assigned a likely reason for the discrepancy.
2.3
Algorithm
Our deep learning algorithm for predicting diabetic retinopathy and diabetic macular
edema builds upon the architecture used in Gulshan et al. [10], which we first provide
a summary of for sake of completion. Gulshan et al. trained a convolutional neural network [15] for predicting multiple binary predictions, such as moderate or worse diabetic
retinopathy, referable diabetic macular edema, or gradability of image. The input to the
neural network is an image of a fundus, and through the use of many stages of computation, parameterized by millions of numbers, the network outputs a real-valued number
between 0 and 1 for each of those binary predictions, indicating how confident it is that
the image falls in the referable category.
The parameters of a neural network are determined by training it on a dataset of
fundus images. Repeatedly, a model is given an image with a known severity rating
for diabetic retinopathy, and the model predicts its confidence in diabetic retinopathy, slowly adjusting its parameters over the course of the training process in order
to become more accurate. The network used in Gulshan et al. was an Inception-v3
model [16], pre-trained on the ImageNet [17] dataset to speed up training. The model
was trained via distributed stochastic gradient descent and evaluated periodically on
3
a tuning dataset throughout the training process. The tuning dataset was used to determine model hyperparameters (parameters of the model architecture that cannot be
trained with gradient descent). Finally, Gulshan et al. created an ensemble of models,
training 10 such Inception-v3 models and combining their predictions, which improves
performance and robustness.
2.4
Algorithmic Improvements
Compared to Gulshan et al., we make a number of improvements to the core neural
network. First, we trained our model on a much larger set of images, obtained from the
teleophthalmology provider EyePACS [18]. Since it would be prohibitively expensive
to re-label all of these images with US-licensed ophthalmologists (as done in Gulshan
et al.), we also use labels determined by the EyePACS grading centers to train our
diabetic retinopathy classifier. In order to use both our ophthalmologist grading and
the EyePACS grading effectively, we treat the labels obtained from these two independent sources as separate prediction targets, i.e. we train our model to predict both an
EyePACS grade for diabetic retinopathy and a grade determined by our own labeling
procedure, if present.
Using this larger training set, we performed a more extensive search for wellperforming hyperparameters using a gaussian process bandit algorithm [19]. One significant change in model hyperparameters that resulted is an increase in the resolution
of input images: the model used in this work has an input resolution of 779 x 779 pixels,
a large increase over the 299 x 299 pixels used in Gulshan et al. This has the effect of
increasing the effective resolution as which lesions are seen. For example, a microaneurysm or small hard exudate might occupy only one pixel or less within a 299 x 299
image [20, 21], but at this higher resolution, it would occupy roughly 3.4 x 3.4 pixels –
still small, but more than 10 times the area.
Other significant changes include upgrading our model architecture from Inceptionv316 to Inception-v4 and predicting a 5-class rating for diabetic retinopathy (Gulshan
et al. only predicted binary referables).
2.5
Evaluation
One advantage of using a machine learning-based system to predict diabetic retinopathy
is that it explicitly outputs a confidence in the range [0,1] for each image. By applying
a threshold to this confidence to determine referability, we can tune the algorithms sensitivity and specificity based on a particular use case – as one lowers the threshold,
the algorithm becomes more sensitive, but less specific. Performance across many different thresholds is depicted in a receiver operating curve (ROC), and is summarized
with the area under the curve (AUC). The primary evaluation of our model is thusly
AUC for different levels of diabetic retinopathy (e.g. AUC for predicting moderate and
above diabetic retinopathy). In addition, for some experiments on 5-class prediction,
we evaluate using the quadratic-weighted Cohens kappa [22]. The decision criterion
for predicting 5-class DR severity is based on a series of thresholds: if the predicted
probability of proliferative DR is above a threshold (determined to reach a target sensitivity of on the tuning set), the algorithms prediction is proliferative DR. If not, then
4
a separate threshold is applied to the algorithms prediction for severe DR or worse to
determine whether the algorithm predicts severe DR, and so on down to mild DR. If
none of predicted probabilities pass their corresponding threshold, no DR is predicted.
3
3.1
Results
Grading and Adjudication
The baseline characteristics of the development and clinical validation datasets are described in Table 1. The training portion of the development set consisted of over 1.6M
fundus images from 238,610 unique individuals. The tune portion of the development
set consisted of adjudicated grades for 3,737 images from 2,643 unique individuals. The
clinical validation set consisted of 1,958 photos from 998 unique individuals. Compared
to the clinical validation set, the development set had a slightly higher proportion of abnormal images.
A comparison of the grades generated by the majority decision of the retinal specialists in the panel before adjudication against the adjudicated reference standard is
outlined in Table 2. Most grades remained within 1 step of each level of severity. For
DR, there were 27 instances (1.5% of images) where the difference was two steps, and
none that were three or more. Adjudication more often yielded new grades with higher
levels of severity. The weighted-kappa score for the grade determined by the majority
decision of retinal specialists and by adjudication was 0.91 (Table 3).
To better characterize the difference in each retinal specialists pre-adjudication grade
and the adjudicated consensus for detecting moderate and above diabetic retinopathy,
we measured the sensitivity and specificity of the each specialist. While all three specialists were very specific ( 99%), sensitivity ranged from 74.4-82.1%, corresponding to
weighted-kappa scores of 0.82-0.91. The majority decision of the panel of three retina
specialists showed higher sensitivity (88.1%) than grades from individual specialists
(Table 3), leaving 11.9% of further cases with moderate or above DR that required the
adjudicated consensus to detect.
In addition to having retinal specialists grade and adjudicate the clinical validation
datasets, we also had three U.S.-board certified ophthalmologists grade the same set
(Table 4). Quadratic-weighted kappa values were generally good (0.80-0.84), but somewhat lower for ophthalmologists than for the retina specialists (Table 5). The majority
decision of the ophthalmologists yielded higher agreement (weighted kappa: 0.87) than
individual ophthalmologists alone. Disagreements were more common in cases where
the adjudicated consensus yielded referable disease (Figure 1).
In-person face-to-face adjudication sessions of the three retina specialists yielded
considerable insights in the grading process. First, image artifacts, especially those resembling typical pathology such as microaneurysms, were a common source of disagreement. This was carried over many images, as artifacts are often caused by consistent lens defects. The adjudication process was critical in correcting these errors. Additionally, there was initial subjective variance in exact definition of grades and de facto
boundaries between grades at the outset of the adjudication process. While traditional
grading definitions are fixed, it became very clear that there are gray areas and significant variance within these definitions, which caused initial disagreement. Once all
5
three retina specialists converged on a more precise set of definitions, staying within but
also refining the traditional definitions, initial disagreements were reduced over time.
Overall, each of the retina specialists voiced that the precision used in the adjudication
process was above that typically used in everyday clinical practice. A summary of the
causes of errors in the non-adjudicated grades generated from the majority decision of
the ophthalmologists is presented in Table 6. The most common causes were missed
microaneurysms (36%), artifacts (20%), and disagreement regarding whether a lesion
was an MA or a hemorrhage (16%). 36 out of 193 disagreements differed by at least two
grades (e.g. no DR vs. moderate or moderate vs. proliferative). Overgrading, i.e. cases
where the majority decision of the ophthalmologists was more severe than the final
adjudicated grade, accounted for 14 (7.3%) of the disagreements, while undergrading
accounted for 22 (11.4%) of the cases.
3.2
Model Results
We show our main results in Table 7, comparing performance of our model and its improvements using either the majority decision and adjudicated grades as the reference
standard. When using majority decision as the reference standard, increasing image resolution shows no clear effect on predicting mild or worse DR (Figure 2). However, when
using adjudicated grades, increasing resolution improves the models AUC from 0.930
to 0.986 for mild or worse DR. However, increasing resolution beyond roughly 500 x
500 pixel input seemed to have marginal performance gains. Another benefit of evaluating with adjudicated grades is in reduced metric variability, where one can observe
significantly reduced metric variance (confidence intervals) across all input resolutions.
We illustrate the performance of our final model using the adjudicated consensus
as the reference standard in Table 8. On 5-class DR prediction, our model achieved a
quadratic-weighted kappa of 0.84, which is on par with the performance of individual
ophthalmologists and retinal specialists. There were 257 disagreements between the
adjudicated consensus and the model. Out of the 56 (21.8%) disagreements that were
two or more steps, 53 (20.6%) were the result of overgrading, i.e. the model predicted a
higher grade than the adjudicated consensus, and 3 (1.2%) resulted from undergrading.
In addition to 5-point grading analysis for DR, we also demonstrate the models performance in binary classification tasks. Figure 3 summarizes the models performance
as well as that of the ophthalmologists and retinal specialists in detecting mild or worse
DR, moderate or worse DR, severe or worse DR, and proliferative DR. For all of these
classification tasks, the algorithms performance was roughly on par with that of individual ophthalmologists and retinal specialists. For mild or worse DR the algorithm had
a sensitivity of 0.970, specificity of 0.917 and AUC of 0.986.
4
Discussion
Deep learning has garnered much attention recently due to it ability to create highly accurate algorithms from large datasets of labeled data without feature engineering. However, care must be taken to evaluate these algorithms against as high-quality ground
truth labels as possible. In previous work [10], we used majority decision as the way
6
of generate the ground truth. This works suggests that adjudication provides a more
rigorous reference standard, especially for the identification of artifacts and missed microaneurysms. However, adjudication is costly. We demonstrate that by adjudicating a
small subset (0.22%) of the training image grades for the tune set, we are able to significantly improve model performance without adjudicating the entire training corpus of
images. The resulting models performance was roughly on par with that of individual
ophthalmologists and retinal specialists. Using adjudication as the reference standard
also improves algorithm development, allowing us to see an improvement in AUC when
using higher-resolution images as input and reduced metric variability during training.
5
Acknowledgements
From Google Research: Yun Liu, Derek Wu, Katy Blumer, Philip Nelson
From EyePACS: Jorge Cuadros
References
1. Early Treatment Diabetic Retinopathy Study Research Group et al. Grading diabetic
retinopathy from stereoscopic color fundus photographsan extension of the modified airlie
house classification: Etdrs report number 10. Ophthalmology, 98(5):786–806, 1991. 1
2. Diabetic retinopathy screening services in scotland: A training handbook July 2003: page
17. http://www.ndrs-wp.scot.nhs.uk/?page_id=1609. Accessed: June 21,
2017. 1
3. American academy of ophthalmology. international clinical diabetic retinopathy disease
severity scale, detailed table. http://www.icoph.org/dynamic/attachments/
resources/diabetic-retinopathy-detail.pdf. Accessed: Oct 14, 2016. 1
4. Ingrid U Scott, Neil M Bressler, Susan B Bressler, David J Browning, Clement K Chan,
Ronald P Danis, Matthew D Davis, Craig Kollman, Haijing Qin, Diabetic Retinopathy Clinical Research Network Study Group, et al. Agreement between clinician and reading center
gradings of diabetic retinopathy severity level at baseline in a phase 2 study of intravitreal
bevacizumab for diabetic macular edema. Retina (Philadelphia, Pa.), 28(1):36, 2008. 1
5. Helen K Li, Larry D Hubbard, Ronald P Danis, Adol Esquivel, Jose F Florez-Arango,
Nicola J Ferrier, and Elizabeth A Krupinski. Digital versus film fundus photography for
research grading of diabetic retinopathy severity. Invest Ophthalmol Vis Sci, 51(11):5846–
5852, 2010. 1
6. Joann G Elmore, Carolyn K Wells, Carol H Lee, Debra H Howard, and Alvan R Feinstein. Variability in radiologists’ interpretations of mammograms. New England Journal
of Medicine, 331(22):1493–1499, 1994. 1
7. Joann G Elmore, Gary M Longton, Patricia A Carney, Berta M Geller, Tracy Onega,
Anna NA Tosteson, Heidi D Nelson, Margaret S Pepe, Kimberly H Allison, Stuart J Schnitt,
et al. Diagnostic concordance among pathologists interpreting breast biopsy specimens.
Jama, 313(11):1122–1132, 2015. 1
8. Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–
444, 2015. 2
9. Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau,
and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural
networks. Nature, 542(7639):115–118, 2017. 2
7
10. Varun Gulshan, Lily Peng, Marc Coram, Martin C Stumpe, Derek Wu, Arunachalam
Narayanaswamy, Subhashini Venugopalan, Kasumi Widner, Tom Madams, Jorge Cuadros,
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. 2, 3, 7
11. Rishab Gargeya and Theodore Leng. Automated identification of diabetic retinopathy using
deep learning. Ophthalmology, 2017. 2
12. Etienne Decencière, Xiwei Zhang, Guy Cazuguel, Bruno Laÿ, Béatrice Cochener, Caroline
Trone, Philippe Gain, Richard Ordonez, Pascale Massin, Ali Erginay, et al. Feedback on a
publicly distributed image database: the messidor database. Image Analysis & Stereology,
33(3):231–234, 2014. 2
13. Gwénolé Quellec, Mathieu Lamard, Pierre Marie Josselin, Guy Cazuguel, Béatrice Cochener, and Christian Roux. Optimal wavelet transform for the detection of microaneurysms
in retina photographs. IEEE Transactions on Medical Imaging, 27(9):1230–1241, 2008. 2
14. Eyepacs digital retinal image grading protocol narrative.
https:
//www.eyepacs.org/consultant/Clinical/grading/
EyePACS-DIGITAL-RETINAL-IMAGE-GRADING.pdf. Accessed: June 21, 2017. 3
15. Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning
applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. 3
16. Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2818–2826, 2016. 3
17. 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. 3
18. Jorge Cuadros and George Bresnick. Eyepacs: an adaptable telemedicine system for diabetic
retinopathy screening. Journal of diabetes science and technology, 3(3):509–516, 2009. 4
19. Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and
D Sculley. Google vizier: A service for black-box optimization. In Proceedings of the
23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,
pages 1487–1495. ACM, 2017. 4
20. H Kolb, E Fernandez, and R Nelson. Facts and figures concerning the human retina–
webvision: The organization of the retina and visual system. 1995. 4
21. Peter H Scanlon, Ahmed Sallam, and Peter van Wijngaarden. A practical manual of diabetic
retinopathy management. John Wiley & Sons, 2017. 4
22. Jacob Cohen. A coefficient of agreement for nominal scales. Educational and psychological
measurement, 20(1):37–46, 1960. 4
8
Tables and Figures
Development
Train
1,665,151
Tune
3,737
Clinical Validation
EyePACS-2
1,958
Images (#)
Patient Demographics
Unique Individuals (#)
238,610
2,643
999
Age (Average ± Stdev)
53.5 ± 11.6
54.3 ± 11.1
54.9 ± 10.9
Female / Total patients
140,183/230,556 (60.8%) 1,101/1,767 (62.3%) 607/999 (60.7%)
where gender was known
Image Quality Distribution
Fully gradable / Total images
where image quality was
1,343,726/1,529,771 (87.8%) 3,547/3,737 (94.9%) 1,813/1,958 (92.6%)
assessed
Development
Train
Tune
Disease Severity
Distribution
Total images where DR
was assessed
No diabetic retinopathy
Mild
Moderate
Severe
Proliferative
Total images where DME
was assessed
Referable diabetic macular
edema
Validation
EyePACS-2
#
%
#
%
#
%
1,662,646
100.0%
3,547
100.0%
1,813
100.0%
1,164,368
152,938
249,138
50,640
45,562
70.0%
9.2%
15.0%
3.0%
2.7%
2,417
458
497
106
69
68.1%
12.9%
14.0%
3.0%
1.9%
1,478
125
144
37
20
81.5%
6.9%
7.9%
2.0%
1.1%
252,544
100.0%
3,547
100.0%
1,813
100.0%
28,807
11.4%
699
19.7%
228
12.6%
Table 1. Baseline characteristics. Summary of image characteristics and available demographic
information in the development and clinical validation datasets. The adjudicated reference standard was used for computing the DR and DME distributions on the Tune and EyePACS-2 datasets,
and the majority reference standard was used for the Train dataset. For DR, the majority was
taken over both the EyePACS partner grade, and the grading done by our ophthalmologists (using whichever grades are available). For DME, the majority is taken over grades from our ophthalmologists. For Image quality distributions, the EyePACS partner grade was used for the Train
dataset and an adjudicated image quality standard was used for the Tune and EyePACS-2 datasets.
9
Majority of retinal specialist grading
before adjudication
No Mild Moderate Severe Proliferative
No
1,469
4
5
0
0
Mild
58
62
5
0
0
Adjudicated
Moderate
22
3
118
1
0
Consensus
Severe
0
0
13
36
1
Proliferative
0
0
0
1
15
Majority of retinal specialist
grading before adjudication
Not Referable Referable DME
Adjudicated Not Referable
1,693
4
16
100
Consensus Referable DME
Table 2. Comparison of retinal specialist grades before and after adjudication on the Validation
dataset. Confusion matrix for DR and DME between the grade determined by majority decision
and adjudicated consensus.
DR
Retina Specialist A
Retina Specialist B
Retina Specialist C
Majority Decision
(Retinal Specialists)
DME
Sensitivity
Specificity
74.6%
74.4%
82.1%
99.1%
99.3%
99.3%
Quadraticweighted kappa
0.82
0.80
0.91
88.1%
99.4%
0.91
Sensitivity
Specificity
kappa
70.1%
92.7%
86.2%
99.5%
99.3%
99.8%
0.78
0.91
0.90
86.2%
99.8%
0.90
Table 3. Agreement between each retina specialist and the adjudicated reference standard on the
Validation dataset. Retina specialists correspond to those who contributed to the final adjudicated
reference standard. Sensitivity and specificity metrics reported are for moderate or worse DR
and Referable DME. Agreement between the pre-adjudication 5-point DR grade and the final
adjudicated grade is also measured by the quadratic-weighted kappa. Since DME classification is
binary, the quadratic-weighted kappa and kappa yield the same number.
10
Majority of ophthalmologist grading
No Mild Moderate Severe Proliferative
No
1,431
33
10
0
1
Mild
67
39
18
1
0
Adjudicated
Moderate
20
14
94
14
2
Consensus
Severe
0
0
9
41
0
Proliferative
0
0
1
2
13
Majority of ophthalmologist
grading
Not Referable Referable DME
Adjudicated Not Referable
1,683
7
Consensus Referable DME
40
75
Table 4. Comparison of ophthalmologist grades versus adjudicated grades from retina specialists
on the Validation dataset. Confusion matrix for DR and DME between the grade determined by
majority decision of the ophthalmologists and the adjudicated consensus of retinal specialists.
DR
Ophthalmologist A
Ophthalmologist B
Ophthalmologist C
Majority Decision
(Ophthalmologists)
DME
Sensitivity
Specificity
67.4%
72.0%
81.1%
98.0%
90.1%
91.2%
Quadraticweighted kappa
0.84
0.80
0.81
83.8%
98.1%
0.87
Sensitivity
Specificity
kappa
69.1%
49.5%
68.8%
99.7%
99.0%
99.5%
0.78
0.58
0.77
65.2%
99.6%
0.75
Table 5. Agreement between ophthalmologists graders with the adjudicated reference standard on the Validation dataset. Sensitivity and specificity metrics are for moderate or worse DR
and referable DME for each grader. Agreement between the adjudicated grade and the 5-point
scale is also measured by the quadratic-weighted kappa. Since DME classification is binary, the
quadratic-weighted kappa and kappa yield the same number.
11
Artifact vs not
Extent of Lesions
Hemorrhage vs MA
Hemorrhage vs not
IRMA vs not
Missed hemorrhage
Missed MA
Missed NVD/NVE
PRP vs not
Other
Total
Adjudication grade (Retina Specialists) minus
Majority decision grade (Ophthalmologists)
-4
-2
-1
1
2
Total
5
28
1
5
39
1
16
9
26
1
13
13
3
30
4
4
11
19
1
1
2
2
6
63
69
2
1
3
1
1
1
3
1
1
1
13
65
92
22
193
Table 6. Reasons for difference between adjudication of retinal specialist and majority decision
from ophthalmologist graders. Disagreements between the adjudicated consensus and majority
decision were examined and characterized by a retinal specialist. Positive numbers denote that
the adjudication grade was more than the majority decision of ophthalmologist grade, and viceversa for negative numbers.
Majority decision ground truth
Adjudicated ground truth
Original
Original model
Original Original model Our full
model
+ high res
model
+ high res
model
Moderate DR+ 0.991
0.990
0.930
0.956
0.986
Referable DME 0.989
0.994
0.971
0.977
0.991
Table 7. Differences in the final AUC observed for algorithms trained in the image resolution algorithm selection experiments. “Original model” refers to the model and training data of Gulshan
et al., “Original model + high res” uses higher resolution input images (779 x 779 pixels), and
“Our full model” incorporates all changes described in the Algorithmic Improvements section in
the main text.
12
Algorithm Grade
No Mild Moderate Severe Proliferative
No
1,356
74
44
1
3
Mild
7
43
74
0
1
Adjudicated
Moderate
3
3
98
36
4
Consensus
Severe
0
0
2
47
1
Proliferative
0
0
0
4
12
Algorithm Grade
Not Referable Referable DME
Adjudicated Not Referable
1,632
65
9
107
Consensus Referable DME
DR
Algorithm
Sensitivity
Specificity
97.1%
92.3%
DME
Quadraticweighted kappa
0.84
Sensitivity
Specificity
kappa
92.2%
96.2%
0.72
Table 8. Comparison of algorithm grade versus adjudicated grades from retina specialists. Confusion matrix for DR and DME where the grade is determined by either the algorithm or adjudicated consensus of the retinal specialists. Sensitivity and specificity metrics reported are for
moderate or worse DR and referable DME.
13
No or Mild Diabetic Retinopathy (n=1603)
Moderate or Worse Diabetic Retinopathy (n=210)
100%
Fraction of Images
Fraction of Images
100%
80%
60%
40%
20%
80%
60%
40%
20%
0%
0%
0%
20%
40%
60%
80%
100%
0%
Agreement with Adjudicated Grade
No DME (n=1697)
40%
60%
80%
100%
Referrable Diabetic Macular Edema (n=116)
100%
Fraction of Images
Fraction of Images
100%
20%
Agreement with Adjudicated Grade
80%
60%
40%
20%
80%
60%
40%
20%
0%
0%
0%
20%
40%
60%
80%
100%
0%
Agreement with Adjudicated Grade
20%
40%
60%
80%
100%
Agreement with Adjudicated Grade
Fig. 1. Grader agreement based on the adjudicated referability of DR and DME. Graders include
all three retina specialists and three ophthalmologists.
Mild or Worse DR, Majority Vote Reference Standard
1.00
0.99
0.99
0.98
0.98
0.97
0.97
0.96
0.96
AUC
AUC
1.00
0.95
0.95
0.94
0.94
0.93
0.93
0.92
0.92
0.91
Mild or Worse DR, Adjudicated Reference Standard
0.91
300
400
500
600
700
800
Input Image Resolution
300
400
500
600
700
800
Input Image Resolution
Fig. 2. Image resolution input to model vs AUC for mild and above DR. Left: Using majority
vote of all ophthalmologists and retinal specialists as the reference standard. Right: Using the
adjudicated grade as a reference standard. Shaded areas represent a 95% confidence interval as
measured via bootstrapping.
14
Mild or Worse Diabetic Retinopathy, AUC: 0.986
100
80
100
90
60
80
70
60
40
50
60
30
80
70
60
40
50
40
20
100
90
Sensitivity, %
Sensitivity, %
80
Moderate or Worse Diabetic Retinopathy, AUC: 0.986
100
40
20
0
2
4
6
8
30
10
0
0
0
20
40
60
80
100
0
20
4
40
1 - Specificity, %
6
8
10
80
80
80
70
60
50
30
100
60
80
70
60
40
50
40
20
40
20
0
2
4
6
100
90
Sensitivity, %
90
40
80
Proliferative Diabetic Retinopathy, AUC: 0.998
100
100
60
60
1 - Specificity, %
Severe or Worse Diabetic Retinopathy, AUC: 0.995
100
Sensitivity, %
2
0
8
30
10
0
0
2
4
6
8
10
0
0
20
40
60
80
100
0
20
1 - Specificity, %
60
80
100
Referrable Diabetic Macular Edema, AUC: 0.991
100
80
Sensitivity, %
40
1 - Specificity, %
100
90
60
80
70
60
40
50
40
20
30
0
2
4
6
8
10
0
0
20
40
60
80
100
1 - Specificity, %
Fig. 3. Comparison of the algorithm, ophthalmologists, and retinal specialists using the adjudicated reference standard at various DR severity thresholds and DME. The algorithms performance
is the blue curve. The three retina specialists are represented in shades of orange/red and the three
ophthalmologists are in shades of blue. n=1,813 fully gradable images.
15
No
No
1,472
Mild
65
Adjudicated
Moderate
26
Consensus
Severe
0
Proliferative
0
Majority of retinal specialist
and ophthalmologist grading
Mild Moderate Severe Proliferative
3
3
0
0
55
5
0
0
6
109
3
0
0
10
40
0
0
1
2
13
Majority of retinal specialist
and ophthalmologist grading
Not Referable Referable DME
Adjudicated Not Referable
1,695
2
23
93
Consensus Referable DME
Table S1. Comparison of combined retina specialist and ophthalmologist grades versus adjudicated grades from retina specialists. Confusion matrix for DR and DME where the grade is
determined by either the majority decision of all six retina specialists and ophthalmologists or is
the adjudicated consensus of the retinal specialists.
16
Fig. S1. STARD diagram for DR.
17
Fig. S2. STARD diagram for DME.
18
| 1cs.CV
|
J. Comput.
Phys.
Journal of Computational Physics 00 (2017) 1–33
arXiv:1708.08741v1 [cs.CE] 29 Aug 2017
Coupled Multiphysics Simulations of Charged Particle
Electrophoresis for Massively Parallel Supercomputers
Dominik Bartuschata,∗, Ulrich Rüdea,b
a Lehrstuhl
für Systemsimulation, Friedrich-Alexander Universität Erlangen-Nürnberg, Cauerstrasse 11, 91058 Erlangen,
Germany
b Parallel Algorithms Group, CERFACS, 42 Avenue Gaspard Coriolis, 31057 Toulouse, France
Abstract
The article deals with the multiphysics simulation of electrokinetic flows. When charged particles are immersed in a
fluid and are additionally subjected to electric fields, this results in a complex coupling of several physical phenomena.
In a direct numerical simulation, the dynamics of moving and geometrically resolved particles, the hydrodynamics of
the fluid, and the electric field must be suitably resolved and their coupling must be realized algorithmically. Here
the two-relaxation-time variant of the lattice Boltzmann method is employed together with a momentum-exchange
coupling to the particulate phase. For the electric field that varies in time according to the particle trajectories,
a quasistatic continuum model and its discretization with finite volumes is chosen. This field is coupled to the
particulate phase in the form of an acceleration due to electrostatic forces and conversely via the respective charges
as boundary conditions for the electric potential equation. The electric field is also coupled to the fluid phase by
modeling the effect of the ion transport on fluid motion. With the multiphysics algorithm presented in this article,
the resulting multiply coupled, interacting system can be simulated efficiently on massively parallel supercomputers.
This algorithm is implemented in the waLBerla framework, whose modular software structure naturally supports
multiphysics simulations by allowing to flexibly combine different models. The largest simulation of the complete
system reported here performs more than 70 000 time steps on more than five billion (5 × 109 ) mesh cells for both
the hydrodynamics, as represented by a D3Q19 lattice Boltzmann automaton, and the scalar electric field. The
computations are executed in a fully scalable fashion on up to 8192 processor cores of a current supercomputer.
c 2017 Published by Elsevier Ltd.
Keywords: Parallel simulation; Electrokinetic flow; Electrophoresis; Fluid-particle interaction; MPI.
1. Introduction
1.1. Motivation
The motion of charged particles in fluids under the influence of electric fields occurs in a wide range
of industrial, medical, and biological processes. When the charged particles are immersed in liquids, their
migration caused by electric fields is termed electrophoresis. Due to the complex interplay of the physical
effects involved in such particle-laden electrokinetic flows, numerical simulations are required to analyze,
∗ Corresponding
author
Email address: [email protected] (Dominik Bartuschat)
1
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
2
predict, and optimize the behavior of these processes. To this end, we present a parallel multiphysics
algorithm for direct numerical simulations of electrophoretic particle motion.
Industrial applications that involve electrophoretic effects are electrofiltration [1, 2, 3] and electrodewatering [4]. Moreover, electrophoresis is utilized in electrophoretic deposition techniques for fabricating
advanced materials [5] and especially ceramic coatings [6, 7] in material science. Electrophoresis and electric
fields are also applied in many medical and biological applications. The trend towards miniaturization of
analysis processes has lead to the development of micro total analysis systems. Due to their high portability,
reduced costs, fast operation, and high sensitivity [8, 9], the design of such lab-on-a-chip systems has been a
highly active area of research for many years. These microfluidic systems require only small samples of liquid
and particles, which are transported, manipulated, and analyzed in structures of length scales from several
nm to 100 µm. Therefore, microfluidic separation and sorting of particles and cells are important steps of
diagnostics in such systems [10, 11]. Many of the employed techniques utilize electric fields to manipulate,
separate, and sort biological particles and macromolecules [8], such as cells [9, 10] or DNA [12].
At the small scales of microfluidic analysis systems, flow measurements are difficult or even impossible. Moreover, the complex coupling of hydrodynamic and electrostatic effects involved in electrophoretic
processes make predictions of electrophoretic motion challenging, especially for large numbers of particles.
Therefore, numerical simulations are essential to aid the design and optimization of electrophoretic systems.
The different physical effects in electrophoretic deposition can be better understood from insight gained
in simulations. By means of such simulations, electrophoretic sorting in lab-on-a-chip systems can be optimized for maximal throughput, sorting efficiency, and sorting resolution. A review of simulation methods for
electrophoretic separation of macromolecules is given in [13]. Also industrial applications of electrophoretic
deposition can be optimized with the help of simulations, as presented in [14] for a coating process.
1.2. Multiphysics Coupling Strategy
For simulations of electrokinetic flows with electrophoretic particle motion, the coupling between three
system components must be modeled: charged objects, fluid flow, and electric effects. The interacting
physical effects are sketched in Fig. 1. In the simulation method introduced in this article, the motion of the
ce
on
io
ty
ns
io
ot
e
si
m
rg
n
de
n
n
a
ch
r
fo
a
st
e
rc
fo
io
el
tro
ec
tic
Electroquasi-statics
object motion
Rigid body
dynamics
Fluid
dynamics
hydrodynamic force
Figure 1: Coupled physical effects of electrophoresis simulated with waLBerla and pe.
rigid, charged particles is modeled with Newtonian mechanics. The motion of the surrounding fluid, which
exerts hydrodynamic forces on the particles, is described by the incompressible Navier-Stokes equation. To
capture fluid-particle interactions, the hydrodynamic forces and the influence of the particle motion on the
fluid are modeled, based on the momentum exchange between fluid and particles. In this way, long-range
hydrodynamic interactions between individual particles and between particles and walls are recovered.
Moreover, electrostatic forces exerted by applied electric fields on the charged particles are modeled, which
cause the electrophoretic motion. The varying positions of the charged particles in return affect the electric
2
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
3
potential distribution in the simulation domain, based on their surface charge. Such a charge is carried by
most biomolecules such as cells, proteins, and DNA [15]. In fact, most substances acquire surface charges
when they get into contact with an aqueous medium [16] or electrolyte solution [17].
Electrostatic forces additionally act on the ions in the fluid and affect the fluid motion via body forces
present in locations of net charge. This net charge originates from the repulsion of co-ions in the fluid of
the same polarity as the surface charge and from the attraction of counter-ions. The ions in the fluid are
transported with the fluid flow, which in turn alters the electric potential distribution. In general, the motion
of ions in the fluid due to an electric field is governed by the Nernst-Planck equation, an advection-diffusion
equation that employs a continuum description and treats the ions as point charges. For flows in which
diffusion strongly dominates over advection and therefore quasi-thermodynamic equilibrium can be assumed
to hold, as considered in this paper, the electric potential due to the ion charge distribution is governed by
the Poisson-Boltzmann equation.
The region of the particle’s surface charge and of excess counter-ions in the fluid is denoted as electric
double layer (EDL). According to the Stern model [18] employed in this article, this double layer comprises
a region of ions attached to the surface and a diffuse part in which the ion concentration follows a Boltzmann distribution. At the surface of shear between the particle and the surrounding diffuse double layer,
the characteristic ζ-potential is defined. The employed equilibrium considerations based on the PoissonBoltzmann equation capture the dominant retardation effect of electrophoretic motion. This effect describes
the retardation of the charged particle motion by the action of the applied electric field on the opposite net
charge in the surrounding EDL. At high ζ-potentials, additionally the weaker relaxation effect occurs [16]
that is caused by a distortion of the EDL and can be captured by the Nernst-Planck equation.
In this article, we present an efficient parallel multiphysics simulation algorithm for electrophoresis on
a fixed Eulerian grid with a Lagrangian representation of moving particles. The particles are represented
by the physics engine pe [19, 20] as geometrically fully resolved three-dimensional objects. Dependent
on the electrostatic and hydrodynamic forces acting on the particles, the pe computes their trajectories
by rigid body dynamics and additionally resolves particle collisions. The pe is coupled to waLBerla
[21, 22, 23], a massively parallel simulation framework for fluid flow applications that employ the lattice
Boltzmann method (LBM) [24, 25]. By means of a modular software structure that avoids dependencies
between modules, functionality from different modules can be combined flexibly. To model fluid-particle
interactions, the LBM momentum exchange method [26, 27] implemented in waLBerla [28] is applied. For
the electrophoresis simulations, the LBM is performed with the two-relaxation time collision operator [29, 30]
and an appropriate forcing term for the electric body force due to the ions in the fluid. The electric potential
is represented by the Debye-Hückel approximation of the Poisson-Boltzmann equation that is discretized
with finite volumes whose mesh structure naturally conforms to the lattice Boltzmann (LB) grid. This
discretization also facilitates accommodating variable and discontinuous dielectricity values that vary in
time according to the particle positions, as required for simulating dielectrophoretic effects. By means of the
waLBerla solver module introduced in [31] together with the cell-centered multigrid solver implemented
therein, the resulting linear system of equations is solved. Since the counter-ions lead to a quicker decay of
the electric potential compared to the long-range electric potentials modeled in [31], the parallel successive
over-relaxation (SOR) method implemented in this module is an adequate choice.
In a previous article, we have shown that the implemented fluid-particle interaction algorithm for arbitrarily shaped objects efficiently recovers hydrodynamic interactions also for elongated particles [32]. Moreover,
we presented a parallel multiphysics algorithm for charged particles in the absence of ions in the fluid in
[31]. We have therein shown that several millions of charged particles with long-range hydrodynamic and
electrostatic interactions can be simulated with excellent parallel performance on a supercomputer. The
present paper extends these simulation algorithms by also considering ions in the EDL around the particles and their effect on the fluid motion, and presenting suitable parallel coupling techniques. Together
with the full four-way coupling of the fluid-particle interaction [31], the coupling with the quasi-equilibrium
representation of the electric effects results in a 7.5-way interaction.
3
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
4
1.3. Related Work
In the following, we give an overview of numerical methods for simulations of electrophoretic phenomena
that have been developed for different resolution levels. At the coarsest modeling scale, both fluid and
solid phase are described by an Eulerian approach. These continuum models represent the charged species
in terms of concentrations and are well suited for simulations with large numbers of unresolved particles.
In [33] two-dimensional electrophoresis simulations of biomolecule concentrations are presented. These
finite element simulations consider the effect of reactive surfaces and the electric double layer is included
through the biomolecule charge. Also three-dimensional parallel simulations of electrophoretic separation
with continuum approaches have been reported. The finite element simulations in [34] consider the buffer
composition and the ζ-potential at channel walls. In [35, 36] mixed finite element and finite difference
simulations of protein separation were performed on up to 32 processes.
At the finest level of resolution, fluid, ions, and particles are simulated by Lagrangian approaches.
These explicit solvent methods [37] typically apply coarse-grained molecular dynamics (MD) models to
describe the motion of fluid molecules and incorporate Brownian motion [38]. The mesoscale dissipative
particle dynamics method is applied in [39] to simulate electrophoresis of a polyelectrolyte in a nanochannel.
Another explicit solvent method is presented in [40] for simulating DNA electrophoresis, modeling DNA as
a polymer. In both methods the polymer is represented by bead-spring chains with beads represented by
a truncated Lennard-Jones potential and connected by elastic spring potentials. Explicit solvent models,
however, are computationally very expensive, especially for large numbers of fluid molecules due to pairwise
interactions [40]. Moreover, the resolution of solvent, macromolecules, and ions on the same scale limits the
maximal problem sizes that can be simulated [41]. Also the mapping of measurable properties from colloidal
suspensions to these particle-based methods is problematic [37].
The high computational effort is significantly reduced in implicit particle-based methods that incorporate
hydrodynamic interactions into the inter-particle forces. Such methods are applied in [37] and [42] to simulate
electrophoretic deposition under consideration of Brownian motion and van der Waals forces. Nevertheless,
these methods are restricted to few particle shapes and hydrodynamic interactions in Stokes flow.
Euler-Lagrange methods constitute the intermediate level of resolution. These approaches employ Eulerian methods to simulate the fluid phase, whereas the motion of individual particles is described by Newtonian mechanics. For simulations of particles in steady-state motion, the resolved particles can be modeled
as fixed while the moving fluid is modeled by an Eulerian approach. In [43] the finite volume method is
applied to simulate electrophoresis of up to two stagnant toroids in a moving fluid, employing the Hückel
approximation for a fixed ζ-potential and different electrical double layer thicknesses. The steady-state electrophoretic motion of particles with low surface potentials under a weak applied electric field in a charged
cylindrical pore is simulated in [44] for a single cylinder and in [45] for two identical spheres. In both cases,
a two-dimensional simulation with a finite element method for Stokes flow and Hückel approximation is
performed, exploiting the axial symmetry of the problem.
For electrophoresis at steady state perturbation approaches can be employed that are based on the assumption that the double layer is only slightly distorted from the equilibrium distribution for weak applied
electric fields (w.r.t. the field in the EDL). In addition to the equilibrium description based on the PoissonBoltzmann equation, small perturbations in the equilibrium EDL are considered in terms of linear correction
terms in the applied electric field for the ion distribution and the electric potential (see e. g. [46]). Using a
perturbation approach with finite elements, the electrophoresis of two identical spheres along the symmetry
axis of a cylindrical domain at pseudo steady-state were studied in [47]. Additionally to the hydrodynamic
and electric interactions, these axisymmetric simulations consider van der Waals forces for particles in close
proximity. In [48] a perturbation approach is applied to simulate a single colloid in a rest frame with periodic boundary conditions. The zeroth-order perturbation corresponds to the Poisson-Boltzmann equation,
which is solved by a constrained variational approach suggested in [49]. For the first-order perturbation,
the stationary Stokes equation is solved by a surface element method, the convection-diffusion for ionic
concentrations by a finite volume solver, and the Poisson equation by a fast Fourier Transform method.
More sophisticated Euler-Lagrange methods include direct numerical simulation (DNS) models that
represent the moving particles as geometrically fully resolved objects. These methods for particulate flows
4
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
5
include approaches with body-fitted moving meshes and fixed meshes. Moving meshes can be represented by
the Arbitrary-Lagrangian-Eulerian (ALE) formulation [50, 51] that employs moving unstructured meshes for
fluid-particle interaction problems. Such an ALE method is applied in [52] to simulate electrophoresis of a
single particle surrounded by a thin electrical double layer. The moving-mesh techniques require re-meshing
when the distortion of the updated mesh becomes too high, and subsequent projection of the solution onto
the new mesh. This overhead is circumvented in fixed-mesh techniques that allow the use of regular grids
and therefore the application of efficient solvers. The fluid particle dynamics method [53] falls into the latter
category, solving the Navier-Stokes and continuity equation on a fixed lattice, and representing the moving
solid particles by fluid particles of very high viscosity. By means of a concentration field that represents the
particle distribution, the particles affect the fluid viscosity and, together with forces acting on the particles,
the body force term of the Navier-Stokes equation. The rigid particles are geometrically modeled by the
Lennard-Jones potential and their motion is described by Newtonian mechanics [53]. This technique is
applied to simulate electrophoretic deposition of two charged particles in [54] and electrophoretic separation
in [55]. In these simulations the electrostatic interactions of the particles are modeled in terms of the body
force field, together with the advection and diffusion of ions and the resulting effect of the applied electric field
on the fluid motion. With this method, however, the particle rigidity is imposed by very high viscosity values
that restrict the time step size [56] and the Lennard-Jones potential restricts the particles shapes to spheres.
The smoothed profile method (SPM) [56] circumvents this time-step constraint by directly modeling the
particles as solid objects. Inside the particles and at solid-fluid boundaries that are represented by diffuse
interfaces, a body force is imposed on the fluid to model the effect of the particle motion on the fluid.
The fluid is again modeled on a fixed Cartesian grid and the particle motion with Newtonian mechanics,
where particle overlaps are typically prevented by a truncated Lennard-Jones potential. With this method,
electrophoresis of charged spherical particles is simulated in [57] for a constant, uniform electric field and
in [58] for an oscillating electric field. In both articles, the ion number concentration is modeled by an
advection-diffusion equation to recover non-equilibrium double layer effects. The SPM is also applied in [59]
to simulate electrophoresis of single cylinders and microtubules, employing the equilibrium representation of
the EDL. A further fixed-mesh technique is the immersed boundary method [60, 61, 62], where the rigid body
motion is imposed on the flow by body forces applied at the particle boundaries. This method, combined
with a finite volume method for solving the steady-state Poisson-Nernst-Planck equation system, is applied
in [63] to simulate the electrophoretic motion of up to three spherical particles in a two-dimensional setup.
Lattice-Boltzmann based methods are very well suited for parallel direct numerical simulations of fluidparticle interactions on fixed Cartesian grids. Both the Lagrangian particles and ions are often explicitly
modeled by molecular dynamics approaches that represent the rigid objects by repulsive potentials. In [64]
the electrophoresis of a colloidal sphere immersed in a fluid with counter-ions is simulated, modeling the
solvent by a lattice Boltzmann method. The charged sphere modeled with molecular dynamics is represented
by a raspberry model that comprised several beads connected by the finitely extensible nonlinear-elastic
(FENE) potential. Using a modified raspberry model with two spherical shells of beads solidly attached
to a larger spherical particle, this method is extended in [65] to simulate the electrophoresis of a spherical
Janus particle. The partially uncharged particle is surrounded by anions and cations represented by charged
beads. Electrophoresis simulations for a single highly charged spherical macro-ion in an electrolyte solution
with explicitly modeled positive and negative micro-ions are presented in [66]. Since the coupling of fluid
and macro-ion is performed via several particle boundary points, a single spherical particle is sufficient
to represent the macro-ion. A similar LB-MD method is applied in [67] to simulate the stretching of a
charged polyelectrolyte between parallel plates. The polyelectrolyte immersed in a liquid with explicitly
modeled counter- and co-ions is modeled by beads bonded together by the FENE potential. In all these
LB-MD simulations with explicitly modeled ions, hydrodynamic interactions are simulated with the LBM
and thermal fluctuations are added to both the fluid and the MD objects. The high computational effort
for modeling each individual ion by means of molecular dynamics, however, restricts the maximum feasible
problem size. With these approaches, only a limited number of ions per colloidal particle can be simulated
and the colloid radius is typically restricted to one order of magnitude larger than the ion size [48]. Therefore,
approaches based on continuum descriptions of the suspended ions are more practical for simulations of many
or for larger charged particles.
5
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
6
Alternatively to the continuum approach based on the Poisson-Boltzmann equation employed in this
article, electrophoresis can be simulated with the link-flux method [68] that models the advection and
diffusion of ions in terms of Nernst-Planck equations. The link-flux method employs the LBM for fluid
dynamics and models ion motion in terms of fluxes between lattice cells. In [69] this method is compared to
a LB-Poisson-Boltzmann approach for a fixed spherical particle in a periodic three-dimensional domain. The
aim is to examine the influence of particle motion and counter-ion concentration on the ζ-potential, leading to
the conclusion that for weakly perturbing electric fields or low Péclet numbers the equilibrium and dynamic
ζ-potentials are indistinguishable. The link-flux method is extended in [41] to support moving particles in
combination with the LB momentum exchange method. To ensure charge conservation in electrophoresis
simulations, appropriate moving boundary conditions for the solute fluxes are introduced. This method is
verified in [41] by electrophoresis simulations of up to eight particles.
1.4. Objectives and Outline
The primary goal of this paper is the introduction of a parallel multiphysics algorithm for electrophoresis
simulations together with validations of the physical correctness of the coupled algorithm for different particle
sizes. For this algorithm, waLBerla is augmented by an efficient boundary handling method that is able to
treat electric potential boundary conditions on the moving particles. Moreover, a joint parameterization for
the different coupled numerical methods is introduced. To achieve excellent computational performance, a
matrix-free representation of the linear system based on a stencil paradigm is used in the solver module [31].
For the linear Debye-Hückel approximation it is systematically exploited that these stencils are almost
uniformly identical throughout the simulation domain. The validation runs were performed on up to 8192
parallel processes of a modern supercomputer. Moreover, simulation results for the electrophoretic motion of
a single particle in a microchannel are presented, including visualizations of the electric potential distribution
and of the resulting flow field around the particle.
The equilibrium considerations in the present paper recover the predominant retardation effect due to
an opposing electrostatic force on the net opposite charge in the electrical double layer that counteracts the
particle motion. For the presented method, a computationally cheap and flexible SOR method is sufficient
to solve the electric potential equations. With our approach we aim for simulations of millions of charged
particles as in [31]. For these large numbers of particles the dynamics of an electrical double layer as in [41]
is computationally too expensive, even on modern supercomputers.
This paper is structured as follows: The physical background of fluid-particle interactions and electrophoresis are described in Sec. 2 and Sec. 3, respectively. In Sec. 4, the employed LB-momentum exchange
method for fluid-particle interactions is outlined, together with the finite volume discretization and the common parameterization concept for the coupled multiphysics methods. Then the extension of the waLBerla
framework for the electrophoresis algorithm is described in Sec. 5. Finally validation results for the electrophoretic motion of a spherical particle and visualizations of the resulting flow field and electric potential
distribution are presented in Sec. 6 before conclusions are drawn in Sec. 7.
2. Fluid-Particle Interaction
The macroscopic description of fluid behavior is based on the continuum hypothesis (cf. Batchelor [70])
that allows to consider a fluid as continuum, irrespective of the underlying molecular structure. In this case,
fluid properties can be represented by macroscopic quantities like density ρf , velocity ~u, and pressure p,
as functions of space and time. In terms of these quantities, fluid dynamics is described by conservation
laws for mass, momentum, and energy. In this article isothermal flows are considered, and therefore the
energy equation does not have to be solved. Moreover, non-continuum effects that become relevant for gas
flows at very small scales [71] are assumed to be negligible. Therefore, no-slip boundary conditions are
assumed to hold at solid-fluid interfaces, and slip velocities due to non-continuum Knudsen layer effects are
not considered.
Conservation of mass is described by the continuity equation. This equation can be derived by considering
a fixed control volume in the fluid relative to a stationary observer (Eulerian view). For incompressible fluids
6
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
7
the density of a fluid element is not affected by pressure changes [70, 72]. Then the density is spatially and
temporally constant (i. e. ρf = const), and the continuity equation reads as
∇· ~u = 0.
(1)
Conservation of momentum in a viscous, compressible fluid can be described in terms of the momentum
flux density tensor Π [72]. The temporal change of momentum in a control volume is balanced by the net
momentum flux through the surface of this volume and by external body forces f~b acting on the volume as
∂ (ρf ~u)
= −∇· Π + f~b .
∂t
(2)
The second-order tensor Π comprises a term for the convective transport of momentum and the total stress
tensor σ for the momentum transfer due to pressure and viscosity
Π = ρf ~u~u> − σ.
(3)
The total stress tensor σ can be decomposed into a part representing normal stresses related to the pressure
and a viscous part related to shear stresses. For incompressible Newtonian fluids, the stress tensor reads as
(4)
σ = −pI + µf ∇~u + (∇~u)> ,
where the first term with the second-rank identity tensor I contains the thermodynamic pressure p defined
according to Landau & Lifshitz [72] as used in the LBM literature [73]. The second term with dynamic
viscosity µf represents the shear stresses that are proportional to the rate of deformation [74] and result from
molecular transport of momentum [75]. With this stress tensor the incompressible Navier-Stokes equation
results from Eqn. (2), together with basic vector calculus and the continuity equation for compressible
fluids [76], as
∂~u
ρf
+ (~u · ∇)~u =
−∇p
+ µf ∆ ~u +
f~b
.
(5)
| {z }
|{z}
∂t
| {z }
|
{z
} pressure stress viscous stress external body force
inertial forces
This equation describes the balance of momentum change and the net force acting on a control volume
in terms of Newton’s second law. The left-hand side represents inertial forces acting on a fluid volume.
It comprises a term for local change of velocity and a term for convective acceleration i. e. the change of
velocity in space [77]. The right-hand side represents surface forces and body forces acting on the fluid
volume. Surface forces are short-range forces acting on the surface of the fluid element and are equivalent
to stress in the fluid [70]. Body forces, such as gravity or electrostatic force, act on the center of mass and
are represented by the force per unit volume of fluid element (or force density) f~b .
An important dimensionless quantity to characterize fluid flows is the Reynolds number Re = UνfL , with
µ
kinematic viscosity νf = ρff , characteristic velocity U , and characteristic length scale L. Flows in the regime
of creeping motion, where Re 1 and thus inertial forces are negligible, are termed Stokes flow.
The Stokes equations for incompressible Newtonian fluids resulting from Eqns. (5) and (1) read as
−∇p + µf ∆ ~u + f~b = 0,
∇· ~u = 0.
(6)
These equations are linear in both, velocity and pressure. Due to the linearity the superposition principle
holds, which is often utilized for fluid-particle interaction in Stokes flow and is employed for the validations
in Sec. 6.3.
Particles immersed in a fluid experience a force in case of relative fluid motion or a pressure gradient
in the fluid (e. g. due to gravity). This force exerted by the surrounding fluid can be calculated from the
7
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
8
stresses in the fluid next to the particle by integrating the stress tensor σ over the particle surface Γp [74, 46]
Z
F~part = σ · ~n dA.
(7)
Γp
Here, dA denotes the surface area elements and ~n the associated normal vector pointing into the fluid.
For simple cases, such as spherical bodies moving in unbounded Stokes flow, or single bodies moving in a
confined domain, analytical solutions for the particle motion are known as described in the following. More
complex fluid-structure interaction problems must be solved numerically, e. g. by the LBM.
For the computation of particle motion in incompressible fluids, hydrostatic effects that result e. g. from
gravitation acting on the fluid, do not have to be considered explicitly in the momentum equation (if the
buoyancy and gravitational force are directly applied to the particle). In this case, the force exerted by the
fluid on the particle according to Eqns. (7) and (4) is the drag force given by
Z
Z
~
~
~
Fd = − p dA + µf ∇~u + (∇~u)> dA,
Γp
Γp
where p is the hydrodynamic part of the total pressure.
The resistance to the motion of a sphere in an unbounded fluid at very low Reynolds numbers can be
calculated analytically from the above expression for the drag force and the Stokes equations (6) that govern
the fluid flow w.r.t. the imposed boundary conditions (BCs). For a sphere located at ~x = ~0, the unbounded
fluid is represented by the BC
~u → 0 as ~x → ∞
imposed on the fluid velocity in the Stokes equations (6). The resulting drag force acting on a rigid sphere
~ was derived by Stokes [78] as
of radius R that moves at constant velocity U
~.
F~d = −6πµf RU
(8)
This equation is commonly referred to as Stokes’ law.
For a sphere moving in a fluid subject to a constant force F~ , Stokes’ law relates the terminal steady-state
velocity of the sphere to the drag force exerted by the fluid. Such a constant force may e. g. be the Coulomb
force that acts on a charged particle in an electric field. The terminal sphere velocity is then obtained from
the balance of the external force and the drag force F~ + F~d = ~0 as
~ =
U
1
F~ .
6πµf R
(9)
A particle moving in a confined domain experiences a retardation caused by surrounding walls. Consequently, the drag force on a sphere that moves in a viscous fluid limited by walls is higher than the force
according to Stokes’ law. The effect of walls on a moving particle in Stokes regime can be determined by
means of the method of reflections, as described in detail in [74]. Happel & Bart [79] employed this method
to obtain a first-order correction to the drag force on a sphere settling in a long square duct with no-slip
walls. Miyamura et al. [80] found polynomial expressions for the increased drag by fitting the coefficients to
experimentally obtained settling velocities of spheres in different confining geometries. The correctness of
the wall effect recovered in LBM simulations with the fluid-particle interaction algorithm employed in this
article was verified in [76] against these expressions.
3. Electrokinetic Flows
The transport of ions in fluids subject to electric fields that occurs in electrokinetic flows can be modeled
by means of a continuum theory, similar to the description of fluid dynamics by the Navier-Stokes equation.
8
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
9
Instead of modeling individual ions and their interactions, local ion concentrations ni and fluxes ~ji of the
different ionic species i are considered. Based on these macroscopic quantities the ion transport in dilute
electrolyte can be described by the Nernst-Planck equation combined with the law for the conservation of
ionic species in a solution
∂ni
= −∇· ~ji
(10)
∂t
in the absence of chemical reactions. Here ni denotes the number density or concentration that is related
to the molar concentration as ci = NnAi with Avogadro’s number NA . The total ionic flux ~ji of species i
comprises an advective flux with a common mass average velocity ~u for all species and fluxes relative to the
advective flux due to diffusion and electric migration [46]. This relation is expressed by the Nernst-Planck
equation
~ji =
ni ~u
− Di ∇ni − ni µ∗i ∇Φ,
(11)
|{z}
| {z }
| {z }
advective flux
diffusive flux
migration flux
where Di and
represent the spatially homogeneous diffusion coefficient and ionic mobility of species i,
respectively, and ∇Φ the local electric potential gradient. The ionic mobility is defined as µ∗i =: DkiBzTi e ,
where e denotes the elementary charge, zi the valence of a given ionic species, kB the Boltzmann constant
and T the temperature.
To model the influence of the charged ions on the electric potential governed by the Poisson equation
µ∗i
− ∆ Φ(~x) =
ρe
εe
(12)
for spatially uniform fluid permittivity εe , the ion charge distribution is considered in terms of the local
mean macroscopic charge density as
X
ρe =
e zi ni .
(13)
i
The Poisson-Nernst-Planck equation system Eqns. (10)–(12) is highly nonlinear, and solving the overall system is computationally very expensive, especially for electrophoresis of many particles. Therefore
the problem is simplified by restriction to equilibrium considerations based on the Boltzmann distribution
that capture the dominant electrophoretic effects. The resulting Poisson-Boltzmann equation holds for
(quasi-)thermodynamic equilibrium when the ion distribution is not affected by fluid flow or by externally
applied electric fields. Therefore the electric potential ψ resulting from the non-uniform ion distribution in
the EDL is considered in the following, instead of the total electric potential Φ = ψ + ϕ that additionally
comprises the potential ϕ of the externally applied electric field.
The Boltzmann distribution for ions can be derived from the Nernst-Planck equation, as outlined in [46].
Considering the Nernst-Planck equation (11) in one dimension and at equilibrium, i. e., for zero macroscopic
fluid velocity u = 0 and ionic flux ji = 0, results in
d ni
zi e d ψ
= −ni
dx
kB T dx
(14)
for the above definition of µ∗i . In this case, the Péclet number P e = UDL for mass transfer relating the
advection rate to diffusion rate becomes zero. Here U is the fluid speed, L a characteristic length scale, and
D the diffusion coefficient. Applying the chain rule to the left-hand side of Eqn. (14) and integrating from
a reference point in the bulk with potential ψ∞ and concentration ni∞ , yields
zi e(ψ − ψ∞ )
kB T
ni = ni∞ e
.
−
(15)
Setting the reference potential ψ∞ in the electroneutral bulk solution to zero recovers the Boltzmann distribution with the number density ni∞ at the location of the neutral state.
9
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
10
From Poisson’s equation (12) with net charge density according to Eqn. (13) and the obtained Boltzmann
distribution, the Poisson-Boltzmann equation follows as
zi e ψ
−
e X
− ∆ψ =
zi ni∞ e kB T ,
εe i
(16)
relating the electric potential ψ to the ion concentrations at equilibrium. For binary, symmetric electrolyte
solutions comprising two species of valence z = −z− = z+ , the Poisson-Boltzmann equation takes the form
2 z e n∞
zeψ
− ∆ψ = −
sinh
.
(17)
εe
kB T
For low ζ-potentials compared to the thermal voltage kB T /e, the term kzBe ψT in Eqn. (17) becomes
smaller than unity. At room temperature this is fulfilled for ζ 25.7zmV [81]. In this case the approximation
sinh(x) ≈ x is accurate, up to a small error of order O(x3 ) (cf. Taylor’s expansion). With this linearization,
the symmetric Poisson-Boltzmann equation simplifies to the Debye-Hückel approximation (DHA)
− ∆ψ = −
2 e2 z 2 n∞
ψ = −κ2 ψ.
εr ε0 kB T
(18)
This equation was originally derived by Debye & Hückel [82] for strong electrolytes [81]. The parameter κ,
defined by
r
εr ε0 kB T
κ :=
,
(19)
2 e2 z 2 n∞
is commonly referred to as Debye-Hückel parameter. Moreover, the charge density in the fluid is then given
by
ρe (ψ) = −κ2 εe ψ.
(20)
In this article, we consider spherical particles with uniform ζ-potential distribution as depicted in Fig. 2.
The electric potential ψ for such a particle of radius R is represented by the Debye-Hückel equation in
E∞
r
ψ =ζ
~u = −U
θ
R
x
λD
~ex
Figure 2: Electrophoresis setup of a stationary (negatively) charged sphere of radius R, surrounded by a double layer and
subject to an applied electric field in opposing fluid flow. Similar to [16].
spherical-polar coordinates as
1 d
r2 dr
r2
dψ
dr
10
= κ2 ψ,
(21)
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
11
with radial distance r from the sphere center and subject to the Dirichlet BCs
ψ=ζ
ψ→0
at
as
r = R,
r → ∞.
(22)
Solving this equation subject to these BCs results in the electric potential distribution in the surrounding
EDL (and beyond) as [81]
R
ψ(r) = ζ e−κ(r−R)
for r ≥ R.
(23)
r
For the Debye-Hückel approximation, an analytical solution for the electrophoretic velocity of a single
spherical particle at steady state and arbitrary EDL thickness has been derived by Henry [83]. Initially the
problem was formulated to account for finite conductivities of particle and medium, and the potential in
the EDL was described by the Poisson-Boltzmann equation. The final results, however, were provided for
insulating spheres, and sufficiently low ζ-potentials for the Debye-Hückel approximation to hold. Therefore
the Debye-Hückel approximation is employed in the following to represent the ion distribution around the
particles under electrophoretic motion.
Instead of modeling a sphere moving at terminal speed U , Henry [83] considered the analogous problem of
a stationary sphere in a steadily moving liquid. To this end, the opposing velocity −U was imposed on the
overall system by setting ~u r→∞ = −U ~ex in the liquid far from the particle. As shown in Fig. 2, a sphericalpolar coordinate system fixed at the particle center with radial distance r and polar angle θ was used. Under
the assumption that the electric potential in the EDL is not distorted from its equilibrium distribution by
the applied field and the fluid flow, the potentials ϕ and ψ were linearly superimposed. Therefore, the
electric potential ψ in the diffuse double layer is described by the Poisson-Boltzmann equation and the
applied potential ϕ by a Laplace equation. The BCs for the Laplace equation applied by Henry represent
the insulating particle by homogeneous Neumann BCs at the particle surface and impose the applied field
by the inhomogeneous Neumann condition ∂ϕ/∂x r→∞ = −E∞ . For the Poisson-Boltzmann equation, the
ζ-potential at the hydrodynamic radius R and the decaying potential were imposed as given in Eqn. (22).
Making use of the equations for the electric potential, the Stokes equations for steady-state creeping flow
with body force term on the right-hand side
−µf ∆ u + ∇p
= −ρe ∇ (ϕ + ψ)
∇· ~u =
(24)
0,
(25)
were solved by Henry [83]. In addition to the BC imposing the opposing velocity far from the particle to
bring the whole system to rest, the no-slip condition ~u r=R = 0 was applied at the particle surface.
From the flow field around the particle, the force acting on the particle was obtained by integrating
the normal stresses over the sphere surface. To the resulting force that comprises Stokes drag and electric
components, the electrostatic force on the particle due to its fixed surface charge was added. The total force
must vanish at steady motion and was thus equated with zero, resulting in the electrophoretic velocity
ZR
ZR
ψ
ψ
ε
~ ext
~ EP = e ψR + R3 5R2
dr − 2
dr E
(26)
U
µf
r6
r4
∞
∞
|
{z
}
=ζ f (κR)
obtained by Henry for an insulating particle. The function f (κR) introduced in [83] is usually referred to
as Henry’s function. In [84] the following expression is derived that approximates the integral equations as
f (κR) =
2
1 +
3
1
2 1+
11
2.5
κR (1 + 2e−κR )
3 ,
(27)
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
12
with a relative error below 1 % for all values of κR [84].
In terms of Henry’s function f (κR) [83], the electrophoretic mobility of a spherical, non-conducting
particle reads as
εe
ζ f (κR) .
(28)
µEP =
µf
This solution is correct to the first order of the ζ-potential, since the relaxation effect is neglected [85]. With
the definition of the electrophoretic mobility µEP := UEP /Eext as electrophoretic speed per unit applied field
the electrophoretic velocity of the particle can be obtained.
Henry’s analytical solution for the electrophoretic velocity of a spherical particle of radius R in an
unbounded electrolyte solution of dynamic viscosity µf with Debye-Hückel parameter κ, subject to an
~ ext reads as
applied field of strength E
~ EP = 2εe ζ
U
1 +
3µf
1
h
2 1+
2.5
κR(1+2e−κR )
~
i3 E
ext ,
(29)
according to the expression for the electrophoretic mobility of Ohshima given in Eqn. (27).
For the electrophoresis simulations in this article, the particle charge must be known to compute the
electrostatic force on the particle. Analytical solutions for electrophoretic motion such as Henry’s equation,
however, are typically given in terms of the ζ-potential, which is defined at the slip surface between the
compact and diffuse EDL layer. Since the particle charge is acquired as a surface charge, for a given ζpotential the surface charge (density) enclosed by the slip surface is therefore needed in the simulations.
The surface charge density is hereby obtained from the overall surface charge bound at the fluid-particle
interface and in the Stern layer. This approach is justified by the fact that the electric potential at the Stern
surface and the ζ-potential can in general be assumed to be identical [86].
The relation of the surface density σs to the ζ-potential is obtained from the Neumann BC on the surface
of the insulating particle [85]
dψ
σs = −εe
(30)
d~r r=R
in case these electrical properties do not vary in angular direction. This condition holds for insignificant
permittivity of the insulating particle compared to the fluid permittivity εe . Alternatively, this relation can
be derived from the electroneutrality condition [46]. With the spatial distribution of ψ around the spherical
particle according to Eqn. (23) the ζ–σs relationship follows from Eqn. (30) as
qs
1 + κR
σs =
=
ζ
ε
.
(31)
e
4πR2
R
For a spherical particle with an EDL potential ψ described by the spherical symmetric Poisson-Boltzmann
equation, the more general ζ–σs relationship
v
u
zeζ
u
8 ln cosh
zeζ u
2
1
2εe κkB T
1
4kB T
u
+
u1 +
sinh
(32)
σs =
2
zeζ
zeζ
ze
κR
2kB T t
(κR)
2
2
cosh
sinh
4kB T
2kB T
for 1-1 and 2-1 electrolyte solutions is derived in [87]. The relative error of this approximation w.r.t. the
exact numerical results computed by [88] is below 1 % for 0.5 ≤ κR < ∞ [85]. This relationship is applied in
the electrophoresis simulation validation in Sec. 6.3 to compute the particle charge for a given ζ-potential.
The applied ζ–σs relationship is derived for electric potentials governed by the spherical Poisson-Boltzmann
equation and is thus more general than the ζ–σs relationship (31) for the Debye-Hückel approximation. For
the simulation parameters used in Sec. 6.2, the deviation of σs for the general relationship w.r.t. the exact
value of the Debye-Hückel approximation is about 0.2 % and hence negligible.
12
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
13
4. Numerical Modeling
4.1. Lattice Boltzmann Method with Forcing
The LBM is a mesoscopic method for the numerical simulation of fluid dynamics based on kinetic theory
of gases. This method statistically describes the dynamics of ensembles of fluid molecules in terms of particle
distribution functions (PDFs) that represent the spatial and velocity distribution of molecules in phase space
over time. The temporal and spatial variation of PDFs, balanced by molecular collisions, is described by
the Boltzmann equation. The solution of this equation converges towards the Maxwell-Boltzmann velocity
distribution of molecules in local thermodynamic equilibrium. For small deviations from this equilibrium,
the Navier-Stokes equations can be derived from the Boltzmann equation by means of a Chapman-Enskog
expansion (cf. [89, 90]).
In the LBM, the phase space is discretized into a Cartesian lattice Ωdx ⊂ RD of dimension D with
spacing dx, and a set of Q discrete velocities ~cq ∈ RD , q ∈ {1, . . . , Q}. Moreover, time is discretized as
cq are chosen
Tdt = {tn : n = 0, 1, 2, . . .} ⊂ R+
0 , with a time increment of dt = tn+1 − tn . The velocities ~
such that within a time increment, molecules can move to adjacent lattice sites or stay at a site. Associated
with each of these velocities is a PDF fq : Ωdx × Tdt 7→ R. A forward-difference discretization in time and
an upwind discretization in space [91] result in the discrete lattice Boltzmann equation
fq (~xi + ~cq dt, tn + dt) − fq (~xi , tn ) = dtCq + dtFq ,
(33)
with lattice site positions ~xi , discrete collision operator dtCq , and discrete body-force term Fq . This equation
describes the advection of PDFs between neighboring lattice sites and subsequent collisions.
In general, the collision operator can be represented in terms of the collision matrix S as dtCq =
P
eq
(cf. [92]), with the vector f~ := (f0 , f1 , . . . , fQ−1 )> of the PDFs fq , and f~eq of the equij Sqj fj − fj
librium distributions fqeq . The latter are obtained from a low Mach number expansion of the MaxwellBoltzmann distribution [89]. With a representation of the macroscopic fluid density as ρf = ρ0 + δρ in
terms of a reference density ρ0 and a density fluctuation δρ, the equilibrium distribution function for the
incompressible LBM is derived in [93] as
(~u · ~u)
(~cq · ~u) (~cq · ~u)2
+
−
,
(34)
fqeq (ρ0 , ~u) = wq ρf + ρ0
c2s
2c4s
2c2s
where ‘·’ denotes the standard Euclidean scalar product. This distribution function depends on ρf , the
macroscopic fluid velocity ~u, and lattice-model dependent weights wq . At each instant of time, ρf and ~u are
given by moments of PDFs as
P
ρf (~xi , t) = fq (~xi , t),
qP
(35)
~cq fq (~xi , t).
~u(~xi , t) = ρ10
q
c2s ρf (~xi , t)
according to the equation of √
state for an ideal
Moreover, the pressure p is given as p(~xi , t) =
gas. We employ the D3Q19 model of [94] with thermodynamic speed of sound cs = c/ 3 for the lattice
velocity c = dx/dt. For this model, the weights wq are: w1 = 1/3, w2,...,7 = 1/18, and w8,...,19 = 1/36. As
discussed in [93], fqeq recovers the incompressible Navier-Stokes equation with at least second-order accuracy
u|
2
of the Mach number Ma := |~
cs , O(Ma ). In LBM simulations, the kinematic fluid viscosity νf is generally
determined by a dimensionless relaxation time τ of the collision operator as
1 2
c dt.
(36)
ν= τ−
2 s
As shown in [95], with this definition the LBM is second-order accurate in space and time.
Among the different collision operators available for the LBM, we employ the two-relaxation-time (TRT)
collision operator of [29, 30]
X
Sqj fj − fjeq = λe fqe − fqeq,e + λo fqo − fqeq,o ,
(37)
j
13
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
14
with the relaxation parameters λe and λo for even- and odd-order non-conserved moments, respectively.
Alternative collision operators either have disadvantages regarding stability and accuracy [96] such as the
BGK model [97] or are computationally more costly such as the MRT operator [92] or the cumulant operator
[98]. To ensure stability, both relaxation parameters should be within the interval ] − 2, 0[, cf. [99, 30].
The even relaxation parameter is related to the dimensionless relaxation time by λe = − τ1 and therefore
determines the fluid viscosity. The free parameter λo is set to λo = −8(2 − 1/τ )/(8 − 1/τ ) in this article,
which prevents τ -dependent boundary locations for no-slip BCs as they arise for the BGK operator. Instead,
walls aligned with the lattice dimensions are fixed half-way between two lattice sites, as shown in [100]. For
the TRT operator, the PDFs are decomposed as fq = fqe + fqo into their even and odd components
fqe = 12 (fq + fq̄ )
fqo = 12 (fq − fq̄ )
and
and
fqeq,e = 12 (fqeq + fq̄eq )
fqeq,o = 12 (fqeq − fq̄eq ),
with ~cq = −~cq̄ . The local equilibrium distribution function in Eqn. (34) is then given by
ρ0
ρ0
u · ~u) + 2c
cq · ~u)2
fqeq,e = wq ρf − 2c
2 (~
4 (~
s
s
fqeq,o = wq ρc20 (~cq · ~u).
(38)
(39)
s
At each time step tn ∈ Tdt the lattice Boltzmann method performs a collide– and a stream step
f˜q (~xi , tn ) = fq (~xi , tn ) +λe [fqe (~xi , tn ) − fqeq,e (~xi , tn )]
+λo [fqo (~xi , tn ) − fqeq,o (~xi , tn )]
(40)
fq (~xi + ~eq , tn + dt) = f˜q (~xi , tn ) + dtFq ,
(41)
where f˜q denotes the post-collision state and ~eq = ~cq dt a discrete lattice direction.
In the stream step, the product of dt and the forcing term Fq is added to the post-collision PDFs. The
term Fq considers the external effect of body forces acting on the fluid. In this article, the discrete forcing
term according to Luo [101] is employed as
(~cq − ~u) (~cq · ~u)
~
+
~
c
Fq = w q
(42)
q · fb ,
c2s
c4s
with f~b representing the electrical body force per unit volume. The forcing terms lead to an additional
term in the continuity equation that arises for spatially varying external forces, as shown in [102, 103]. This
additional term can be removed by incorporating the external force in the momentum density definition as
!
1 X
dt ~
~u =
fq~cq + fb .
(43)
ρ0
2
q
Thus, [102] suggest to use the forcing term (42) in combination with the modified momentum density
definition in Eqn. (43) for the BGK. We therefore use this forcing term with second-order accuracy, together
with the modified momentum density for the resulting velocity ~u.
To increase the computational efficiency of the implementation, the compute-intensive collide step and
the memory-intensive stream step are fused to a stream-collide step. In the simulations presented in this
article, no-slip and free-slip BCs are applied as described in [32].
4.2. Momentum Exchange Approach
To model the fluid-particle interaction and the resulting hydrodynamic interactions of the particles, the
momentum exchange approach suggested by [27] is employed in this article. This approach computes the
momentum exchange between the fluid and the suspended rigid particles from PDFs surrounding these
solid objects. The implementation of this approach in waLBerla has recently been applied to simulate
fluid-particle interactions also at Reynolds numbers beyond the Stokes regime, as presented in [104, 105, 106].
14
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
15
For the momentum exchange approach, the solid particles are mapped onto the lattice by considering
each cell whose center is overlapped by a particles a moving obstacle cell. All other lattice cells are fluid cells,
resulting in a staircase approximation of the particle surfaces. These surfaces are represented by surface cells
indicated by subscript s in the following. On the fluid cells denoted by subscript F , the LBM is applied. To
model the momentum transfer from the particles to the fluid, the velocity bounce-back BC
ωq
fq̄ (~xF , tn + dt) = f˜q (~xF , tn ) − 2 2 ρ0~cq · ~us
cs
(44)
is applied at fluid cells with position ~xF = ~xs + ~eq̄ adjacent to a surface cell at ~xs . This boundary condition
introduced in [26] matches the fluid velocity to the local particle surface velocity ~us .
From the sum of all force contributions due to the momentum transfer from fluid cells to neighboring
surface cells (cf. [32]), the overall hydrodynamic force on the particle can be obtained according to [27] as
X X
dx3
ωq
F~h =
2f˜q (~xF , tn ) − 2 2 ρ0~cq · ~us ~cq
.
(45)
cs
dt
s
q∈Ds
Here, Ds is the set of direction indices q in which a given particle surface cell s is accessed from an adjacent
~ h is given by substituting ~cq × (~xs − ~xC ) for the last ~cq term in
fluid cell. Analogously, the overall torque M
Eqn. (45), with ~xC representing the particle’s center of mass.
The mapping of the solid particles onto the lattice results in fluid cells appearing and disappearing as
the particles move. Therefore, at uncovered lattice sites the PDFs must be reconstructed. We set the PDFs
at those fluid cells to the equilibrium distribution f eq (ρ0 , ~us (~xs (tn − dt)) according to Eqn. (34) dependent
on the local particle surface velocity from the previous time step.
4.3. Finite Volume Discretization for Electric Potential Equations
To solve the Debye-Hückel approximation a cell-centered finite volume scheme is applied on the Cartesian
lattice Ωdx introduced in Sec. 4.1. Associated with this lattice of spacing dx that represents the computational
domain Ω ⊂ R3 is the three-dimensional cell-centered grid Gdx defined (cf. [107]) as
j − 1/2
V
V
Gdx := ~xi ∈ Ω ~xi = k − 1/2 dx i ∈ i , (j, k, m) ∈ J
,
(46)
m − 1/2
with Ωdx = Ω ∩ Gdx . For
V indexing of lattice cells by tuples (j, k, m) of cell indices in the three spatial
dimensions, the index set J := {(j, k, m) | j = 1, . . . , lx ; k = 1, . . . , ly ; m = 1, . . . , lz } is introduced.
Here
V
lx , ly , and lz represent the numbers
of
cells
in
x,
y,
and
z-direction,
respectively.
The
set
is
related
J
V
to the
V setVof single cell indices i := {i | i = 1, . . . , lx · ly · lz } used for the LBM by a bijective mapping
g : i → J , according to Eqn. (46).
The finite volume discretization of the Debye-Hückel approximation Eqn. (18) includes volume integration
over each lattice cell Ωi = Ωklm := [~xj−1,k,m , ~xj,k,m ] × [~xj,k−1,m , ~xj,k,m ] × [~xj,k,m−1 , ~xj,k,m ] and applying
the divergence theorem to the Laplace operator ∆ = ∇· ∇, resulting in
I
Z
2
~
−
∇ψ(~x) dΓi + κ
ψ(~x) d~x = 0 ∀ Ωi ∈ Ωdx .
(47)
∂Ωi
Ωi
Here, ∂Ωi denotes the closed surface of the cell, and ~Γi is a surface directed element. The cell surface
consists of six planar faces with constant outward unit normal vectors ~niq (q = 1, . . . , 6). Therefore, the
surface integral can be decomposed into a sum of integrals [108] as
I
−
∂Ωi
∇ψ(~x) d~Γi = −
Z
6
X
q=1 ∂Ω
iq
15
∇ψ(~x) · ~niq dΓiq ,
(48)
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
16
where q represents LBM direction indices introduced in Sec. 4.1, and ∂Ωiq is the common face with the neighboring cell in direction q. The gradients ∇ψ(~xi ) · ~niq in normal direction of the faces ∂Ωiq are approximated
at the face centers by central differences of ψ from a neighboring and the current cell as
∇ψ(~xi ) · ~niq
~
xi + 12 ~
eq
≈
ψ(~xi + ~eq ) − ψ(~xi )
.
dx
(49)
Here, ~eq represents the corresponding lattice direction introduced in Sec. 4.1.
Substituting the approximation of the normal fluxes across the surfaces of area dΓiq = dx2 , Eqn. (49) into
Eqn. (48) results in
I
6
X
ψ(~xi + ~eq ) − ψ(~xi ) 2
−
dx .
(50)
∇ψ(~x) d~Γi ≈ −
dx
q=1
∂Ωi
Applying the above finite volume discretization to the linear term of the Debye-Hückel approximation
results in
Z
κ2
ψ(~x) d~x ≈ κ2 ψ(~xi ) dx3 .
(51)
Ωi
DHA
This additional term enters the central element of the resulting seven-point stencil Ξdx
as
0
−1
0
0
0
0
0
0
0
1
−1 6 + κ2 dx2 −1 0 −1 0 ,
0
−1
0
dx2
0
0
0
0
0
0
0
−1
0
(52)
and the right-hand side for each unknown is zero.
4.4. Parametrization for Electrokinetic Flows
Numerical simulations are typically performed in terms of dimensionless parameters. To ensure consistently good numerical accuracy independent of the simulated scales, the physical quantities are mapped to a
computationally reasonable numerical value range. In LBM simulations usually the quantities are expressed
in lattice units. Therefore, physical values must be converted to lattice values before the simulation and
vice versa to obtain physical values from simulation results. In the following the lattice unit (LU) system
employed in this article is presented, providing a common parameterization also for further (electrokinetic)
flow scenarios [76].
Physical simulation parameters are given in terms of the international system of quantities (ISQ) associated to the international system of units (Système International d’Unités, SI [109]) [110]. The value of a
quantity is defined as product of a numerical value and a unit, e. g. 10−6 m. The SI system comprises the
base units displayed in Tab. 1, together with the corresponding mutually independent base quantities length,
time, mass, electric current, thermodynamic temperature, amount of substance, and luminous intensity.
Moreover, derived units such as N = kgs2m (Newton) are defined in the SI system as products of powers of
base units [109].
For LBM simulations the base quantities are length, time, and mass density. The corresponding lattice
Table 1: Physical and LBM base quantities for electrokinetic simulations.
kind of quantity
length
time
mass
electr.
thermodyn.
chem.
photometr.
ISQ quantity
SI unit
x
m
t
s
m
kg
I
A
T
K
n
mol
Iv
cd
LBM quantity
lattice unit
x
dx
t
dt
ρ
ρ0
Φ
V
–
–
–
–
–
–
16
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
17
base units are the (physical) lattice spacing dx, time increment dt, and fluid reference density ρ0 . Therefore,
the numerical values of these quantities in lattice units become unity, as shown in Tab. 2. The lattice
parameters representing these dimensionless numerical values in LUs are indicated with subscript L in
the following, e. g. dxL for the lattice spacing. Performing LBM computations on such normalized lattice
parameters saves numerical operations: Additional scaling factors are avoided, e. g. in the LBM streamcollide step, when computing the equilibrium distribution function given in Eqn. (34) (cf. cL = dxL /dtL = 1)
or the macroscopic velocity according to Eqn. (35) (cf. ρ0,L = 1).
For the electrokinetic simulations, the electric potential Φ is chosen as electric base quantity. This quantity, however, is not scaled on the lattice but keeps its numerical value that typically lies in a range not too
far from unity. In contrast to the SI system, the LU system for electrokinetic simulations requires no base
units corresponding to the thermodynamic temperature T or the amount of substance n:
In the simulations, temperature appears only in combination with the Boltzmann constant as energy,
i. e., E = kB T , with the derived unit [E] = [m] [x]2 /[t]2 . With the relation of mass and mass density
[m] = [ρ0 ] [x]3 , this unit can be represented in lattice base units (see Tab. 1).
Moreover, by representing the molar concentration with unit [ci ] = mol
l in terms of the number density as
1
ni = ci NA with Avogadro’s number NA = 6.022 14 · 1023 mol
, the unit ‘mol’ cancels out, yielding [ni ] = [x]−3 .
With the choice of the potential Φ as base quantity, the electric current I becomes a derived quantity
in the LU system. The unit of I can be derived from the energy in electrical units [E] = [I] [Φ] [t] and the
above definition of energy in terms of lattice units [E] = [ρ0 ] [x]5 /[t]2 . Equating both relations results in the
derived lattice unit
[ρ0 ] [x]5
.
(53)
[I] =
[Φ] [t]3
In Tab. 2 different physical quantities and their SI units are displayed, as well as their representation by
(dimensionless) lattice parameters and their lattice units. The conversion of physical quantities to lattice
Table 2: Relation of physical quantities and lattice parameters for electrokinetic simulations.
physical quantity
SI unit
dx (lattice spacing)
dt (time increment)
m
s
ρ0 (fluid density)
Φ (electr. potential)
kg
m3
L (length)
ν (kinem. viscosity)
~u (velocity)
m (mass)
F~ (force)
I (electr. current)
V
m
m2
s
m
s
kg
kg m
s2
A
lattice parameter
dxL
dtL
ρ0,L
ΦL
LL
νL
~uL
mL
F~L
IL
e (elem. charge)
As
eL
ε0 (vacuum permittivity)
~ (electr. field )
E
As
Vm
V
m
ε0,L
~L
E
J
EL
E (energy)
numerical value
1
dx dx (= 1)
1
dt dt (= 1)
1
ρ0 ρ0 (= 1)
1
V Φ
1
dx L
dt
ν
dx2
dt
u
dx ~
1
m
ρ0 dx3
dt2
F~
ρ0 dx4
3
V dt
I
ρ0 dx5
V dt2
e
ρ0 dx5
V dt2
ε
ρ0 dx5 0
dx ~
V E
dt2
E
ρ0 dx5
lattice unit
dx
dt
ρ0
V
dx
dx2
dt
dx
dt
ρ0 dx3
ρ0 dx4
dt2
ρ0 dx5
dt3 V
ρ0 dx5
dt2 V
ρ0 dx5
V dt2
V
dx
ρ0 dx5
dt2
units requires their division by powers of the LBM base quantities with the corresponding SI units. Since the
potential Φ has the same numerical value in physical and lattice units, the LBM base unit of the potential is
dx5
simply ‘1 V’. Therefore, the derived lattice unit ‘Ampere’ for the electric current is given by A = ρV0 dt
3 (see
Eqn. (53)). The corresponding scale factors for converting the physical parameters (e. g. ν) to the associated
17
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
18
lattice parameters (e. g. νL ) are shown in Tab. 2. Multiplication with the inverse scale factors converts the
lattice parameters back to physical quantities.
5. Extension of waLBerla for Electrophoresis Simulations
Electrophoresis simulations require the mutual coupling of fluid dynamics, rigid body dynamics, and
electro-statics, as shown in Fig. 1. In addition to electrostatic and hydrodynamic interactions, the applied
field acts on the EDL charge and thereby affects fluid flow and particle motion. For the electrophoresis
algorithm presented below, the equilibrium description of the EDL potential in terms of the linear DebyeHückel equation is employed. Therefore, the predominant retardation effect is recovered in the simulations.
The applied potential ϕ and the EDL potential ψ are linearly superimposed (cf. Henry’s equation, Sec. 3),
which is valid for weak applied fields when the EDL distortion by the field is negligible. Thus, the applied
electric field can be imposed directly, without solving the associated Laplace equation. In the following the
main concepts of the waLBerla framework are described, together with the functionality for electrophoresis
simulations implemented therein.
5.1. Design Concepts of waLBerla
WaLBerla is a framework for massively parallel multiphysics simulations with a MPI-based distributed
memory parallelization that is specifically designed for supercomputers. The main software design goals of
this framework are flexibility to combine models of different effects, extensibility to allow the incorporation
of further effects and details, e. g., for electrokinetic flows, and generality to support further applications [76].
These goals are reached by integrating the coupled simulation models into waLBerla in a modular fashion
that avoids unnecessary dependencies between the modules. This way, the modules can be augmented
by more sophisticated models or models tailored to a certain application, and functionality from different
modules can be combined flexibly. The modular code structure also provides excellent maintainability, since
modifications of the code in one module do not affect other modules. Developers can therefore efficiently
locate faulty modules and find bugs inside these modules, also by systematically utilizing automatic tests.
In addition to modules, the waLBerla code structure comprises a core for sequence control that initializes data structures, performs the time stepping, and finalizes the simulation on each parallel process.
By means of applications, multiphysics simulations can be defined by assembling the associated functionality and coupled models from the modules. The coupling strategy for multiphysics simulations is based
on accessing mutually dependent data structures (see [31] for more details). These data strucutres are
defined in the modules that implement models for the different physical effects. Also infrastructural and
utility functionality is encapsulated in modules, e. g., for domain setup, MPI communication, BC handling,
parameterization, or simulation data output. For parallel simulations the discretized simulation domain is
decomposed into equally sized blocks of cells that can be assigned to different parallel processes. In this block
concept, each block contains a layer of surrounding ghost cells that is needed for BC treatment and parallelization. For parallelization, cell data of neighboring processes is copied to the ghost layer, dependent on
the data dependencies of the unknowns located on a given block. Moreover, metadata of a block specifies its
location in the simulation domain or its rank for MPI communication. The communication concept provides
a simple and flexible communication mechanism tailored to simulations on Cartesian grids and facilitates
various communication patterns. Individual work steps of a simulation algorithm are specified as sweeps that
are executed on a block-parallel level. The sweep concept defines a structure in which callable objects (i. e.
kernels) implemented in the modules can be specified at compile time. By means of dynamic application
switches, specific kernels tailored to a given computer architecture or implementing a desired model variant,
can be selected at run time. For time-dependent simulations, the sweeps are organized in a timeloop that
specifies the order in which the individual work steps are executed at each time step. To facilitate iterative
solvers, sweeps can also be nested to repeatedly perform a grid traversal until a termination criterion is met.
The boundary condition concept for handling multiple physical fields, numerical methods, and governing
equations, introduced in [31], is applied in this article for moving obstacles with electric BCs. This concept
relies on flags to indicate for each boundary cell the kind of boundary treatment that to is be performed,
18
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
19
with an individual BC flag for each condition. Moreover, cells adjacent to a boundary are indicated with a
nearBC flag and non-boundary cells with a nonBC flag. Individual sets of these flags are defined for each
governing equation. Due to specific LBM requirements, the boundary handling is performed such that the
BCs are fulfilled when a boundary cell is accessed from a nearBC cell in the susequent sweep. The abstract
boundary handling is implemented in the bc module and provides functionality for all BCs are handled
either as direct or direction-dependent treatment. In direction-dependent treatment the BC value is set
at a boundary cell dependent on the value at a neighboring cell, whereas in direct BC treatment the BC
value is directly set at a boundary cell [31]. The actual boundary handling functionality is implemented in
corresponding BC classes whose functions are executed when the associated BC flag is found.
The parameterization concept for multiphysics simulations introduced in [76] is based on the conversion
of physical parameters to lattice units before the simulation, as described in Sec. 4.4. This approach ensures
consistent parameters in all waLBerla modules, independent of the underlying physics. Individual modules
can therefore be developed independently w.r.t. the common lattice unit system.
Simulation parameters and BCs are typically provided to waLBerla through an input file. To ensure
a consistent parameter set and a correct mapping of the physical quantities to lattice units, the class
PhysicalCheck has been introduced in waLBerla in [111]. This class checks the simulation parameter set
specified in the input file for completeness and physical validity and converts the parameters to lattice units.
Since the quantities are converted based on the SI system, the unit Ampere is re-defined for PhysicalCheck
according to Eqn. (53) to support simulations including electric effects.
5.2. Overview of waLBerla Modules for Electrophoresis Simulations
In the following an overview of the modules relevant for electrophoresis simulations is given. For fluid
simulations with the LBM, the lbm module implements various kernels for the stream-collide step with the
different collision operators and forcing terms described in Sec. 4.1. The classes for treating the corresponding
BCs are provided by the associated lbm bc module. In the lbm module block-fields of cells are provided for the
PDFs, the velocity, the density, and an external force. The external force field is used for coupling the LBM
to other methods, e. g. via the forces exerted by electric fields on the EDL and the fluid in electrophoresis
simulations. The PDF and velocity field are accessed by moving obstacle module functions for the simulation
of moving particles.
The moving obstacle module facilitates simulations of fluid-particle interactions with the momentum
exchange method by implementing kernels for moving obstacle sweeps and providing the corresponding
data structures. This module furthermore provides setup functions for initializing and connecting the pe
to waLBerla. For the moving obstacle handling, an obstacle-cell relation field is provided that stores for
each lattice cell overlapped by a pe object a pointer to this object. Moreover, from the lbm module the PDF
source field is accessed for the hydrodynamic force computation and the reconstruction of PDFs. In the lbm
velocity field, body velocities are stored and accessed in the moving boundary treatment of the LBM.
For the computation of the electric potential distribution, the lse solver module described in Sec. 5.3 is
employed. This module has been implemented for solving large sparse linear systems of equations as they
arise from the discretization of the electric potential equations (see Sec. 4.3). The corresponding BC classes
are implemented in the pot bc module described in Sec. 5.4. The data structures defined in the lse solver
module, accessed by the application and other modules, include the stencil field representing the system
matrix as well as the scalar fields for the solution and for the RHS.
The electrokin flow module was designed for facilitating electrokinetic flow simulations. This module
provides kernels for coupling the involved methods as well as the setup and parameterization of simulations
such as electrophoresis. The setup includes initializing the stencils and RHS from the lse solver module
according to the finite volume discretization of the Debye-Hückel equation presented in Sec. 4.3. The kernels
for imposing the electric potential BCs on the moving particles and for computing the electrostatic forces
on fluid and particles are described in Sec. 5.5 and Sec. 5.6, respectively. Finally, the coupled algorithm for
electrophoresis provided by the electrokin flow module is presented in Sec. 5.7.
19
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
20
5.3. Solver for Linear Elliptic PDEs with Moving Boundaries
The solver module lse solver in waLBerla for large linear systems of equations has been designed as
an efficient and robust black-box solver on block-structured grids [112, 31, 76]. To specify the problem to
be solved, the application sets up the system matrix, right-hand side, and BC handling.
In the lse solver module, solver sweeps are pre-defined that perform the iterations at the position where
they are added to the timeloop. For a given simulation setup, the employed solver is selected via the input
file where also the solver parameters and BCs are specified. An iterative solver requires a nested sweep
that is executed until a specific convergence criterion is satisfied. For all implemented solvers, these sweeps
share a common structure. This structure is displayed for the SOR sweep solveTimeVaryingBCSOR for moving
boundaries and linear PDEs such as the Debye-Hückel approximation in Fig. 3 as an activity diagram. For
Adapt stencils
and RHS to BCs
Maximum iterations reached?
yes
no
Compute
residual and norm
Termination criterion met?
yes
no
Update ‘red’
unknowns
Update ‘black’
unknowns
Figure 3: Activity diagram for SOR sweep solveTimeVaryingBCSOR with time varying boundary conditions (BCs),
e. g., due to moving particles
the employed discretization of the electric potential equations, the system matrix is represented by D3Q7
stencils. These stencils comprise an entry for each, the center and the six cardinal directions.
In the setup function for this SOR sweep, the communication for the ghost layer exchange of the solution
field in the initialization phase is set up first. Then the SOR solver sweep is added to the timeloop, and the
kernels for relaxation, communication, and BC treatment are specified as solver sub-sweeps. For parallel
execution, the SOR algorithm is implemented in red-black order.
The filled circle at the top of the diagram in Fig. 3 indicates the starting point of the sweep in the
timeloop. At the beginning of the sweep solveTimeVaryingBCSOR for moving boundaries, the stencils are
constructed to incorporate the BCs according to the present boundary locations. Furthermore, the RHS
is adapted to these BCs before the solver iterations start. For this purpose, a sub-sweep with a kernel
for re-setting the stencils and RHS is executed before the iteration loop, followed by the BC treatment
functions adaptStencilsBC and adaptRHSBC described in Sec. 5.4. Then the standard parallel Red-Black SOR
sweep is performed until the termination criterion is met. This sweep comprises a sub-sweep for computing
the residual and its L2 -norm for the termination criterion, as well as two solver sub-sweeps for the SOR
update of the ‘red’ and the ‘black’ unknowns, respectively. In these sub-sweeps, the quasi-constant stencil
optimization technique introduced in [31] is employed. Based on the residual L2 -norm, the termination
20
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
21
criterion for the simulations performed in this article is provided as residual reduction factor RESRF w.r.t.
the initial norm of the simulation.
For multigrid solvers as applied in [31] to charged particle simulations in absence of ions in the fluid, the
red-black update sub-sweeps in Fig. 3 are replaced by solver sub-sweeps of a V-cycle. To apply the SOR
sweep to varying stencils of linearized PDEs, such as the symmetric Poisson-Boltzmann equation, only the
sub-sweep for adapting the stencils and RHS to the BCs is performed at the beginning of each iteration
instead of once before the iterations start (see [76]).
5.4. Boundary Condition Handling for Solver Module
The implicit BC handling used and initiated by the solver module has been introduced in [31]. This
boundary handling is based on incorporating the BCs into the stencils and right-hand side of the finite
volume discretization. That way, at an iterative update of a near-boundary value, the method implicitly
uses the new values for the BCs. For Dirichlet boundaries, the boundary values are linearly extrapolated to
the boundary cell and for Neumann BCs the boundary values are approximated by central differences. For
both the stencils and the right-hand side, a direction-dependent BC treatment is used.
The functions for this BC treatment are implemented in the pot bc module. This module employs own
nonBC and nearBC flags for the BC handling of scalar potentials. Moreover, for each BC class in this
module an associated BC flag is defined. For the employed cell-centered discretization, the module contains
one class for each, Neumann and Dirichlet BCs.
For incorporating the BCs into the stencils the kernel adaptStencilsBC is implemented. This kernel
iterates over all lattice cells to find scalar potential nearBC cells. At each cell with nearBC flag, the kernel
employs the D3Q7 stencil directions to iterate over the neighboring cells. In directions of a cell with scalar
potential BC flag, the stencil entry of the nearBC cell, associated with the direction of the BC flag, is
adapted accordingly.
The function adaptRHSBC employs the standard boundary handling of the bc module to invoke the pot bc
kernels for adapting the RHS to the BCs. To this end, the direction-dependent BC treatment kernels in
the corresponding BC classes implement the RHS adaption depending on the BC value. The BC value
is specified in the input file for static BCs, or in a previous time step for BCs of moving particles (see
Sec. 5.5). To facilitate such complex boundaries, the BC classes store the BC values and the corresponding
boundary cell ranges. The latter are stored in a memory-efficient way either as cell intervals or as cell vectors.
Moreover, to allow the computation of scalar potential gradients directly from the solution field, the BC
values are set in this field at boundary cells when the RHS is adapted. From these BC values and from the
solution at the nearBC cell, the value at the boundary cell required for the gradient can be extrapolated.
5.5. Electric Potential Boundary Condition Handling for Moving Particles
Prior to the EDL potential distribution computation by the lse solver module, uniform ζ-potentials
or surface charge densities are imposed at the moving particles by means of scalar potential BCs. These
electrical surface properties are specified in the input file for different particle types with a common uid
(unique identifier) defined in the pe. To this end, the sweep function setPotBC_ChargParticles is implemented
in the electrokin flow module that maps the charged particles onto the lattice and sets the electric potential
BC values and the associated BC handling flags at the corresponding cells. The function first overwrites the
values of all cells in the RHS field with zero to remove the values from the previous BC treatment. Then
the mapping is performed for all movable rigid bodies located in the subdomain of each process. For each
particle the mapping is conducted in an extended axis-aligned bounding box that surrounds this rigid body
and the associated nearBC flags. The mapping is realized in three steps:
1) First, all scalar potential nearBC and BC flags from the previous time step are removed, and the
scalar potential nonBC flags are set. Moreover, the previous BC values and the associated cells in the
BC class instances are removed.
2) Then, for particles with prescribed electric BCs, the BC handling for the lse solver module (see Sec. 5.4)
is prepared. For each rigid body with a uid for which a surface property is specified in the input file,
the associated BC is obtained. The cells overlapped by this particle are gathered and are added
21
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
22
together with the BC value to the corresponding BC class instance. Moreover, the BC flag is set at
the overlapped cells.
3) Finally, the nearBC flag is set at all cells adjacent to a BC cell.
Each step is performed for all bodies on a process before the next step begins, to prevent that for particles
with overlapping bounding boxes the flags from a previous step are overwritten.
5.6. Computing Electrostatic Forces on Fluid and Particles
The electric forces acting on the ions in the fluid are incorporated into the incompressible NavierStokes equation by the body force term f~b = −ρe (ψ) ∇ (ϕ + ψ), which coincides with the corresponding
term employed in Henry’s solution (see Sec. 3). Due to the linear superposition of the electric potential
~ ext = −∇ϕ is
components, the gradient is applied to both components separately. Since the applied field E
given, only the EDL potential gradient must be computed.
For the computation of the electric field due to the EDL, the electrokin flow module provides a kernel
that performs the gradient computation at all scalar potential nonBC cells. The gradient of the electric
potential is computed, as previously introduced in [31], by means of finite differences that provide O(dx2 )
accuracy. Where possible, an isotropy-preserving D3Q19 stencil is used (cf. [113]) instead of a D3Q7 stencil.
With the LBM D3Q19 stencil, the gradient can be computed using wq -weighted differences of neighboring
values in 18 directions ~eq as
19
~eq
1 X
wq ψ(~xb + ~eq ) · 2 .
(54)
∇ψ(~xb ) ≈
w1 q=2
dx
At nearBC cells the D3Q7 stencil is applied to compute the gradient of ψ from the BC values stored at
particle and boundary cells in the BC treatment (see Sec. 5.4). The obtained electric field is stored in a field
of cells that is accessed in the body force computation.
For the computation of the electric body force and of the electrostatic force exerted on the particles, a
further kernel is implemented in the electrokin flow module:
The kernel first iterates on each parallel process over all lattice cells to compute the body force at scalar
potential nonBC cells. This force is computed as product of charge density ρe (ψ) and total electric field
~ total = E
~ ext − ∇ψ. The relation of charge density and EDL potential follows Eqn. (20). The obtained
E
electric body force is written to the external force field of the lbm module that is accessed by the LBM
kernels with forcing. Then the kernel iterates over all non-fixed particles residing on the current parallel
process to compute the electrostatic force. For each of these particles the force is computed from the particle
~ ext and is then added to the particle.
charge and the applied field as F~C = qe E
5.7. Algorithm for Electrophoresis
The overall parallel algorithm for electrophoresis simulations with waLBerla is shown in Alg. 1. After
the setup and initialization phase, the electric BCs for the EDL potential are set at the moving charged
particles by means of setPotBC_ChargParticles at each time step. Then the Debye-Hückel approximation is
solved by means of the SOR sweep solveTimeVaryingBCSOR. The iterations are performed until the specified
termination criterion for the residual is met.
From the obtained EDL potential distribution and the applied field the electric body force exerted on the
fluid is computed, as described in Sec. 5.6. First the kernel computing the electric field caused by the EDL is
applied. Then the kernel for the electric body force computation from the charge density distribution in the
fluid and the total electric field is invoked. This kernel additionally applies the electrostatic force exerted
by the applied field to the particles.
Then the rigid body mapping sweep described in Sec. 4.2 is performed, imposing the particle velocities
for the subsequent LBM sweep. In that parallel sweep, an LBM kernel with fused stream-collide step and
forcing term (see Sec. 4.1) is employed to compute the fluid motion influenced by the moving particles and
by the electrostatic force exerted on ions in the EDL.
After the LBM sweep, the hydrodynamic forces on the particles are computed by the momentum exchange
method. The obtained hydrodynamic force contributions and the electrostatic forces are then aggregated by
22
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
23
Algorithm 1: Electrophoresis Algorithm
foreach time step, do
// solve Debye-Hückel approximation (DHA):
set electric BCs of particles
while residual ≥ tol do
apply SOR iteration to DHA
// couple potential solver and LBM:
begin
compute electric field due to EDL
compute charge density in fluid
compute electric body force
apply electrostatic force to particles
// solve lattice Boltzmann equation with forcing, considering particle velocities:
set velocity BCs of particles
begin
perform fused stream-collide step
// couple potential solver and LBM to pe:
begin
apply hydrodynamic force to particles
pe moves particles depending on forces
the pe. From the resulting forces and torques, the new particle velocities and positions are computed in the
subsequent pe simulation step by the PFFD algorithm [114] that additionally resolves rigid body collisions.
6. Electrophoresis Simulations
In the following, the electric potential computation is validated for a sphere with uniform surface charge
surrounded by an EDL. This sphere is placed in a micro-channel subject to an applied electric field. Moreover
the flow field caused by the electrophoretic motion of the sphere in the micro-channel is visualized, together
with the electric potential and the surrounding ions to qualitatively show the correctness of the simulations.
Then the electrophoretic velocity and the retardation by the counter-ions in the EDL is validated w.r.t.
Henry’s solution for different sphere sizes and values of κR.
The simulation setup for the validation experiments is depicted in Fig. 4. A sphere is placed on the
2
L x/
BC
Fy
BC
C
B
z
x
y
BC
BC
Lz/2
Lz/2
2
L x/
C
B
Ly
Figure 4: Setup for electrophoresis of spheres in square duct with different BCs.
longitudinal axis of a cuboid domain of size Lx × Ly × Lz at an initial position of yinit , and an electrostatic
force in y-direction acts on the sphere. The sphere is suspended in a symmetric aqueous electrolyte solution
23
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
24
kg
of kinematic viscosity νf = 1.00 · 10−6 ms , density ρf = 1.00 · 103 m
3 , permittivity εe = 78.54 · ε0 , and ions
of valence z = 1 at a temperature of T = 293 K.
In all simulations, the EDL thickness λD is in the order of the particle diameter. The electric potential
around the sphere due to the EDL is represented by the Debye-Hückel approximation. For this approximation, analytical solutions are known for the potential distribution and for the electrophoretic velocity.
The LBM is employed with TRT operator and second-order forcing term by Luo and re-defined momentum
density (see Sec. 4.1) in all simulations. The linear system of equations resulting from the finite volume
discretization of the Debye-Hückel equation is solved by the SOR method that is sufficient for the quickly
decaying electric potential due to the counter-ions in the EDL. For the SOR, a relaxation parameter of
ωSOR = 1.7 is applied in all simulations.
2
6.1. Electric Potential in an EDL around a Sphere
To validate the computation of the EDL potential ψ around a charged particle, a sphere of radius
R with uniform ζ-potential is simulated in a large domain. The analytical solution of the Debye-Hückel
equation (21) representing the EDL potential around the sphere is given by Eqn. (23). For the validation,
a spherical particle of radius RL = 12 with initial position yinit = 64 dx is chosen and a domain size of
128 dx × 256 dx × 128 dx. The lattice spacing dx is displayed in Tab. 3 with the simulation parameters as
employed related to the electric potential. The ζ-potential is chosen sufficiently low to approximate the
Table 3: Parameters for sphere with uniform surface charge in micro-channel.
c∞ / mol
l
ζ/mV
−10.0
5.00 · 10
1
κ/ m
−6
7.41 · 10
dx/m
6
10 · 10
λD,L
−9
13.49
Poisson-Boltzmann equation by the Debye-Hückel approximation. To obtain the displayed values of κ and
of the characteristic EDL thickness λD , a symmetric aqueous electrolyte solution with ions of the valence
z and the bulk concentration c∞ shown in Tab. 3 is simulated. The EDL thickness is greater than the 12
lattice sites required for a sufficient resolution as observed in the electro-osmotic flow simulations in [76].
Despite its quick decay, the analytical solution of the electric potential at the domain boundaries differs from zero. Thus, the values of ψ at these boundaries are set to the analytical solution by means of
Dirichlet conditions. To solve the Debye-Hückel equation subject to these BCs, a residual reduction factor
of RESRF = 2 · 10−7 is employed as termination criterion for the SOR method.
The analytical (ψ) and numerical (ψ ∗ ) solution at the initial particle position are depicted in Fig. 5 along
ψ/V
0
·10−3
Analytical
solution
−5
−10
Numerical
solution
0
20
40
60
xL
80
100
120
Figure 5: Analytical and numerical solution for EDL potential of sphere with uniform surface charge.
24
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
25
a line in x-direction through the sphere center. Both graphs agree very well, showing the correctness of the
finite volume discretization and the applied SOR solver, as well as the boundary handling at the particle
surface. Inside the insulating particle, the electric potential is not computed explicitly, due to the uniform
surface potential and the resulting symmetric distribution of ψ in the sphere. The electrostatic force needed
to compute the particle motion is computed directly from the applied field and the particle charge, instead
of the electric potential gradient as in [31].
6.2. Electrophoresis of a Sphere in a Micro-Channel
The application of the external electric field in the micro-channel setup described in Sec. 6.1 gives rise to
an electrophoretic motion of the sphere. For the simulation of this scenario, the parameters listed in Tab. 4
are employed in addition to the parameters in Tab. 3. The particle is suspended in an electrolyte solution
Table 4: Parameters for electrophoretic motion of sphere in micro-channel.
V
Ey / m
qs /A s
−18
−19.9 · 10
−4.7 · 10
dt/s
6
−12
200 · 10
UEP,L
4.49 · 10−3
with the parameters introduced at the beginning of Sec. 6. From the ζ-potential and the parameters in
Tab. 3, the sphere’s charge qs in Tab. 4 is obtained from the surface charge density σs according to the ζ–σs
relationship (32), multiplied by the surface area of the sphere, as qs = 4πR2 σs [81].
The high electric field strength Ey in Tab. 4 is chosen to keep the number of simulation time steps at a
minimum. Due to the applied field and the resulting electrostatic force FC,y = 933 · 10−12 N the particle
moves in y-direction, retarded by the channel walls and the opposing force on the EDL. For the chosen
parameters, the terminal particle speed of UEP = 224.5 mm
s is obtained for free space according to Henry’s
solution (see Eqn. (29)), corresponding to a particle Reynolds number of Rep,d = 0.054.
In the simulation, periodic BCs are applied in y-direction. At all other walls, no-slip conditions are
applied for the LBM and homogeneous Neumann conditions for the electric potential. At each time step,
the Debye-Hückel equation is solved by SOR with the termination criterion from Sec. 6.1. The LBM is
run with τ = 6.5, resulting in the time increment dt listed in Tab. 4 for the chosen viscosity νf and dx
(cf. Eqn. (36)). With dt and dx, the electrophoretic speed in lattice units UEP,L attains the values given
in Tab. 4. Gravitational effects are neglected in the simulation to ensure that the particle motion is driven
kg
solely by electric forces. Thus, the employed density ρp = 1195 m
3 of the particle only has an impact on the
time required to reach steady-state.
In Fig. 6, the results of the electrophoresis simulation are visualized at different time steps. The EDL
potential ψ in the y-z plane through the domain center is displayed, together with semi-transparent equipotential surfaces representing the excess counter-ions in the EDL. The flow field around the moving sphere is
visualized in the x-z plane through the domain center. Arrows of uniform length indicate the flow direction,
while the velocity magnitude is represented by the shown color-scale and by twelve white isosurface contour
lines with logarithmic intervals in the range of 13.67 · 10−3 ms to 6.08 · 10−6 ms .
The flow field around the moving particle shown in Fig. 6(a) is fully developed after 5001 time steps, and
the particle has already attained its terminal velocity. Due to the periodicity of the domain, a channel flow in
axial direction has emerged from the particle motion, as indicated by the contour lines. Moreover, a vortex
has formed between the sphere and the surrounding no-slip walls. The flow field moves with the particle
that translates along the channel centerline (see Fig. 6(b)). Because of the equilibrium representation of the
EDL, the distribution of ψ is at all time steps symmetric w.r.t. the sphere center, almost up to the boundary.
6.3. Validation of Electrophoretic Motion of a Sphere
To quantitatively validate the overall electrophoresis simulation algorithm, the electrophoretic velocity
of spheres with uniform surface charge is compared to Henry’s solution Eqn. (29) for a spherical particle in
an unbounded electrolyte solution.
25
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
26
(a) Results after 5001 time steps.
(b) Results after 30 001 time steps.
Figure 6: Electrophoresis of spherical particle in micro-channel with insulating no-slip walls. Visualization of flow field in x-y
plane, EDL potential in y-z plane, and ion charge distribution around charged particle.
In the simulation experiments, a sphere with uniform ζ-potential is moving under the influence of an
~ ext in a large domain filled with an electrolyte solution. The simulations are
applied field of strength E
performed for different values of κR by varying the sphere radius while keeping the EDL thickness constant.
For validation, the relative deviation ∆r U of the obtained terminal sphere velocity from the theoretical
velocity in Eqn. (29) is evaluated.
The simulation parameters are chosen such that the electrophoretic motion is in the Stokes regime, and
thus the nonlinear inertial term of Navier-Stokes equation can be neglected. Moreover, the EDL potential
distribution is governed by the linear Debye-Hückel approximation. Therefore, the superposition principle is
assumed to hold. Thus, from the obtained relative deviation from the analytical solution in an unbounded
domain ∆r U , the relative deviation due to wall effects and volume mapping errors ∆r UStokes will be subtracted. These relative deviations were examined in [76] for several domain sizes and sphere radii. For no-slip
BCs, the wall effect was shown to comply with analytical and experimental results. In the experiments,
domain sizes are used for which the relative deviations ∆r UStokes for free-slip BCs are close to −3 %.
For all simulations, the parameters in Tab. 5 are used. In each experiment, a sphere is suspended in a
symmetric aqueous electrolyte solution with the parameters introduced at the beginning of Sec. 6 and a bulk
concentration of c∞ = 1.60 · 10−5 mol
l . These parameters result in the Debye-Hückel parameter κ shown in
Tab. 5 and, together with the displayed lattice spacing dx, in the EDL thickness λD of approximately 15
26
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
27
Table 5: Parameters for electrophoretic velocity validation w.r.t. Henry’s solution.
ζ/mV
1
κ/ m
V
Ey / m
dx/m
λD,L
10.0
13.3 · 106
99.0 · 106
5.00 · 10−9
15.08
lattice sites. The ζ-potential has the same absolute value as in the electrophoresis simulations in a microchannel from Sec. 6.2. Moreover, as in Sec. 6.2, a high electric field strength Ey is chosen to keep the number
of simulation time steps low. The resulting electrostatic forces FC = qs Ey for the different sphere radii
are displayed in Tab. 6, together with the electrophoretic charges qs on the spheres. These surface charges
associated with the ζ-potential are obtained from the general ζ–σs relationship (32) as in Sec. 6.2. For all
simulation parameters, the maximum relative deviation of σs for the general relationship from the exact
value for the Debye-Hückel approximation is below 0.2 %. These charges are applied as particle charges to
drive the spheres by the electrostatic force due to the electric field.
For the different sphere radii and the associated values of κR, the electrophoretic velocities according
to Henry’s solution are displayed in Tab. 6. These velocities correspond to particle Reynolds numbers Rep,d
from 0.018 to 0.057 for the particle diameters of 40 nm to 120 nm.
EM
is introduced.
To quantify the retardation by the opposing force on the EDL, the variable EPRet = UEPU−U
EM
This variable represents the relative deviation of the electrophoretic velocity of a particle with charge qs in
presence of the electric double layer from the migration velocity UEM of a particle with the same charge in
absence of surrounding ions. For the examined sphere radii, this retardation is in the range of 20 % to 42 %.
Table 6: Electrophoresis parameters and domain sizes dependent on sphere size. Shown are the sphere radii RL in lattice units,
the corresponding charges qs , and the electrostatic forces FC . For relation of sphere radius to EDL thickness κR, theoretical
electrophoretic velocities UEP , Reynolds numbers Rep,d , and electrophoretic retardation parameter EPRet are given. Listed in
lower part are domain sizes per dimension Lx,y,z , initial sphere positions yinit , process numbers per dimension #procx,y,z , and
relative deviations of sphere velocities in free-slip domain from Stokes velocity ∆r UStokes .
4
6
8
9
12
qs /(10
A s)
FC /(10−12 N)
2.21
219
3.67
363
5.36
530
6.29
622
9.43
934
κR
UEP /(10−3
Rep,d
EPRet /%
0.265
461
0.018
-20.6
0.398
464
0.028
-27.7
0.530
466
0.037
-33.6
0.597
468
0.042
-36.2
0.796
471
0.057
-42.8
864
1248
392
1280
1536
588
1632
2048
784
1632
2048
882
1632
2048
1176
8 × 16 × 8
8 × 16 × 16
16 × 16 × 32
16 × 16 × 32
16 × 16 × 32
-3.04
-2.65
-2.98
-2.89
-3.00
RL
−18
m
s )
Lx,z /dx
Ly /dx
yinit /dx
#procx,y,z
∆r UStokes /%
Then the domain sizes for which the relative deviation from Stokes velocity is about ∆r UStokes = −3 %
are listed in Tab. 6, together with the initial position yinit in movement direction that correspond to 98 × RL .
The LBM with TRT collision operator is employed with the relaxation time τ = 6, resulting in the time
increment dt = 45.8 · 10−12 s. As in the micro-channel electrophoresis simulations, gravitational are effects
kg
neglected, and the insulating particles have a density of ρp = 1195 m
3.
For the electric potential, homogeneous Neumann BCs are applied at the walls in x- and z-direction as
27
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
28
= 0. To improve convergence [31], homogeneous Dirichlet BCs are applied at the walls
in y-direction as Φ Γ ∪Γ = 0 V. The particle is located at sufficient distance from these walls for the
N
S
EDL potential to decay to approximately zero, and therefore these BCs do not affect the electric potential
distribution. As termination criterion of the SOR solver for the Debye-Hückel equation, a residual reduction
factor of RESRF = 1 · 10−6 is used.
The parallel simulations are run on SuperMUC1 of the Leibniz Supercomputing Centre LRZ in Garching
(Germany) on 64 to 512 nodes with the numbers of processes listed in Tab. 6. Within the execution time
of 37 h to 48 h, a number of 70 296 to 74 000 time steps are performed, and the spheres cover a distance
of about 300 dx. The high numbers of time steps are chosen to ensure that the spheres reach steady-state
motion, and that large numbers of sampling values are available to compute the terminal velocities.
The sphere velocities are sampled every 20 time steps during the simulation. In the second half of the
simulation, the terminal particle velocity is reached. Thus, the mean particle velocity U ∗ and the velocity
U ∗ −U ∗
fluctuations δU = maxU ∗ min due to volume mapping effects are computed from the last 50 % output values.
In the considered range of time steps additionally the number of time steps between two SOR calls and the
number of SOR iterations is monitored. The average number of time steps between two SOR calls decreases
from ∆TSSOR = 24 to ∆TSSOR = 3 as RL increases from 4 to 12. Likewise, the average number of SOR
iterations per solver call decreases from 451 iterations to 198 iterations for the respective sphere radii.
The obtained terminal velocity in lattice units UL∗ and the fluctuations are displayed in Tab. 7. As expected,
∂Φ
∂~
n ΓW ∪ΓE ∪ΓB ∪ΓT
Table 7: Simulation results of electrophoretic velocity validation for different sphere sizes. Shown are the theoretical velocities
∗ and fluctuations δ , the relative deviations ∆ U of
UEP,L for unbounded domains in lattice units, the obtained velocities UL
r
U
∗ from U
UL
EP,L , and the relative deviations ∆r UEP corrected by hydrodynamic wall and mapping effects.
4
6
8
9
12
UEP,L /10
4.227
4.249
4.274
4.286
4.320
UL∗ /10−3
4.119
2.52
-2.56
0.5
4.144
1.52
-2.48
0.2
4.120
0.83
-3.59
-0.6
4.135
0.68
-3.51
-0.6
4.151
0.29
-3.91
-0.9
RL
−3
δU /%
∆r U/%
∆r UEP /%
the fluctuations decrease with increasing sphere resolution. Moreover, the relative deviation of the obtained
velocity U ∗ from the theoretical electrophoretic velocity UEP given by ∆r U = (U ∗ − UEP )/UEP is listed
in Tab. 7. For all examined sphere radii, the obtained velocities in the confined domain are by 2.5 % to
3.9 % lower than the theoretical values of UEP for a particle in an unbounded electrolyte solution. The
effect of the confinement on the particle velocity is deducted by subtracting the relative deviation from
Stokes velocity ∆r UStokes in Tab. 6 for the corresponding domain sizes from the relative electrophoretic
velocity deviation ∆r U obtained in the electrophoresis simulation. From the resulting relative deviations
∆r UEP = ∆r U − ∆r UStokes , the inaccuracies due to electric effects in the electrophoresis simulation are
assessed. As can be seen from the values of ∆r UEP in Tab. 7, the simulations results agree with the theoretical
values with relative deviations of less than 1 %.
7. Conclusion
In this article, a coupled multiphysics algorithm for parallel simulations of the electrophoretic motion
of geometrically resolved particles in electrolyte solutions is presented. The physical effects are simulated
by means of the lattice Boltzmann method for fluid dynamics, a physics engine for rigid body dynamics,
and a scalar iterative solver for electric potentials. These components are integrated into the parallel
software framework waLBerla to simulate the electrophoretic motion of fully resolved charged particles in
1 www.lrz.de/services/compute/supermuc/
28
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
29
microfluidic flow. The simulations include fluid-particle and electrostatic particle interactions. Additionally
electric effects on ions around the homogeneously charged particles are recovered.
The current work is an extension of [31], where the electrical migration of charged particles without
ions in the fluid was validated and excellent parallel performance was shown for more than seven millions
of interacting charged particles. In the present article, the opposite net charge in the electric double layer
(EDL) around the charged particles due to ions in the fluid is considered, together with its effect on the fluid
motion that counteracts particle motion. To this end, the electric potential distribution in the fluid due to
the EDLs is computed that causes an electric body force on the fluid. This quasi-equilibrium distribution
recovers the motion of ions in the fluid along with the charged particles while neglecting EDL distortion.
The overall electrophoresis algorithm is introduced and an overview of the coupled functionality implemented in the involved waLBerla modules is given. For the simulations, a solver sweep for time-varying
boundary conditions has been developed that is presented here for the parallel SOR method employed to
solve the EDL potential equation. Based on the multiphysics boundary handling concept [31] an efficient
parallel algorithm is implemented to impose electric potential boundary conditions on the moving particles. These methods can also be employed for other governing equations with spatially varying boundary
conditions that model physical effects different from electric fields. The presented parallel electrophoresis simulations are also facilitated by a joint parameterization concept for the different coupled governing
equations and numerical methods implemented in waLBerla. This concept is based on lattice Boltzmann
requirements and is applicable and extensible to further multiphysics simulations.
For the electrophoresis simulations in this article, the electric potential in the double layer is shown to
coincide with analytical solutions. The obtained terminal electrophoretic velocities comply with analytical
solutions for different proportions of the particle radii to double layer thickness. These validation results
verify the correctness of the implementation and the coupling of the different methods. Moreover, the
observed relative errors in the modeling of electric effects are below 1 %. The retardation effect caused
by the presence of the EDL is shown to be significant for the examined sphere radii, reducing the sphere
velocity up to 42 %. For the electrophoretic motion in a micro-channel, the flow field and the electric potential
distribution are visualized, including the ion charge distribution in the EDL surrounding the particle.
The presented algorithm can be applied to find design parameters in industrial and medical applications,
e. g., for optimal separation efficiency of charged biological particles in lab-on-a-chip devices, depending on
fluid, particle, and electrolyte properties. Our algorithms were shown to correctly recover fluid-particle
interactions for elongated particles in [32]. These future simulations may therefore include suspended,
possibly charged particles of various shapes including spheres, spherocylinders, and particles of more complex
shapes, e. g., to represent different biological particles. Also pairwise van-der-Waals forces can be added
easily, to facilitate simulations of electrophoretic deposition in material science applications.
The electrophoresis algorithm introduced here is well suited for massively parallel simulations. In the
current implementation of this algorithm, the EDL thickness is restricted to values in the order of the particle
radius. Therefore, adaptive lattice refinement as in [23] may be employed to allow for thinner double layers
relative to the particle size. For the incorporation of transient effects in the simulations including EDLs,
the link-flux method implemented into waLBerla in [115] and [41] may be employed. This method was
extended in [41] to simulate electrophoresis, enabling the simulation of non-equilibrium ion distributions in
the EDL. Due to the higher computational complexity of the link-flux method compared to the equilibrium
approach in this article, the maximum number of particles will be lower than in our approach.
Acknowledgements
The authors are grateful to the LRZ for providing the computational resources on SuperMUC.
References
[1] K. J. Ptasinski, P. J. A. M. Kerkhof, Electric Field Driven Separations: Phenomena and Applications, Sep. Sci. Technol.
27 (8-9) (1992) 995–1021. doi:10.1080/01496399208019021.
29
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
30
[2] S. Zhang, R. Tan, K. Neoh, C. Tien, Electrofiltration of Aqueous Suspensions, J. Colloid Interface Sci. 228 (2) (2000)
393–404. doi:10.1006/jcis.2000.6966.
[3] Y.-H. Weng, K.-C. Li, L. H. Chaung-Hsieh, C. Huang, Removal of humic substances (HS) from water by electromicrofiltration (EMF), Water Res. 40 (9) (2006) 1783–1794. doi:10.1016/j.watres.2006.02.028.
[4] A. Mahmoud, J. Olivier, J. Vaxelaire, A. Hoadley, Electrical field: A historical review of its application and contributions
in wastewater sludge dewatering, Water Res. 44 (8) (2010) 2381–2407. doi:10.1016/j.watres.2010.01.033.
[5] L. Besra, M. Liu, A review on fundamentals and applications of electrophoretic deposition (EPD), Prog. Mater. Sci.
52 (1) (2007) 1 – 61. doi:10.1016/j.pmatsci.2006.07.001.
[6] I. Zhitomirsky, Cathodic electrodeposition of ceramic and organoceramic materials. Fundamental aspects, Adv. Colloid
Interface Sci. 97 (1–3) (2002) 279–317. doi:10.1016/S0001-8686(01)00068-9.
[7] P. Sarkar, D. De, T. Uchikochi, L. Besra, Electrophoretic Deposition (EPD): Fundamentals and Novel Applications in
Fabrication of Advanced Ceramic Microstructures, Springer, 2012, Ch. 5, pp. 181–215. doi:10.1007/978-1-4419-9730-2_
5.
[8] Y. Kang, D. Li, Electrokinetic motion of particles and cells in microchannels, Microfluid Nanofluid. 6 (4) (2009) 431–460.
doi:10.1007/s10404-009-0408-7.
[9] A. A. S. Bhagat, H. Bow, H. W. Hou, S. J. Tan, J. Han, C. T. Lim, Microfluidics for cell separation, Med. Biol. Eng.
Comput. 48 (10) (2010) 999–1014. doi:10.1007/s11517-010-0611-4.
[10] D. R. Gossett, W. M. Weaver, A. J. Mach, S. C. Hur, H. T. K. Tse, W. Lee, H. Amini, D. Di Carlo, Label-free
cell separation and sorting in microfluidic systems, Anal. Bioanal. Chem. 397 (8) (2010) 3249–3267. doi:10.1007/
s00216-010-3721-9.
[11] N. Pamme, Continuous flow separations in microfluidic devices, Lab Chip 7 (2007) 1644–1659. doi:10.1039/B712784G.
[12] D. G. Hert, C. P. Fredlake, A. E. Barron, Advantages and limitations of next-generation sequencing technologies: A
comparison of electrophoresis and non-electrophoresis methods, Electrophoresis 29 (23) (2008) 4618–4626. doi:10.1002/
elps.200800456.
[13] G. W. Slater, C. Holm, M. V. Chubynsky, H. W. de Haan, A. Dube, K. Grass, O. A. Hickey, C. Kingsburry, D. Sean,
T. N. Shendruk, L. Zhan, Modeling the separation of macromolecules: A review of current computer simulation methods,
Electrophoresis 30 (5) (2009) 792–818. doi:10.1002/elps.200800673.
[14] F. Keller, H. Nirschl, W. Dörfler, E. Woldt, Efficient numerical simulation and optimization in electrophoretic deposition
processes, J. Eur. Ceram. Soc. 35 (9) (2015) 2619–2630. doi:10.1016/j.jeurceramsoc.2015.02.031.
[15] D. Sheehan, Physical Biochemistry: Principles and Applications, John Wiley & Sons, 2013.
[16] R. F. Probstein, Physicochemical Hydrodynamics: An Introduction, 2nd Edition, Butterworths series in chemical engineering, Butterworth Publishers, 1989.
[17] H. C. Chang, L. Y. Yeo, Electrokinetically Driven Microfluidics and Nanofluidics, Cambridge Univ. Press, 2009.
[18] O. Stern, ZUR THEORIE DER ELEKTROLYTISCHEN DOPPELSCHICHT, Z. Elektrochem. 30 (21-22) (1924) 508–
516. doi:10.1002/bbpc.192400182.
[19] T. Preclik, U. Rüde, Ultrascale Simulations of non-smooth granular dynamics, Comp. Part. Mech. 2 (2) (2015) 173–196.
doi:10.1007/s40571-015-0047-6.
[20] K. Iglberger, U. Rüde, Massively Parallel Rigid Body Dynamics Simulation, Comp. Sci. Res. Dev. 23 (3-4) (2009) 159–167.
doi:10.1007/s00450-009-0066-8.
[21] C. Feichtinger, S. Donath, H. Köstler, J. Götz, U. Rüde, WaLBerla: HPC software design for computational engineering
simulations, J. Comput. Sci. 2 (2) (2011) 105–112. doi:10.1016/j.jocs.2011.01.004.
[22] C. Godenschwager, F. Schornbaum, M. Bauer, H. Köstler, U. Rüde, A Framework for Hybrid Parallel Flow Simulations
with a Trillion Cells in Complex Geometries, in: Proc. Int. Conf. on High Performance Computing, Networking, Storage
and Analysis, SC ’13, ACM, 2013, pp. 35:1–35:12. doi:10.1145/2503210.2503273.
[23] F. Schornbaum, U. Rüde, Massively Parallel Algorithms for the Lattice Boltzmann Method on NonUniform Grids, SIAM
J. Sci. Comput. 38 (2) (2016) C96–C126. arXiv:1508.07982, doi:10.1137/15M1035240.
[24] S. Chen, G. D. Doolen, Lattice Boltzmann Method for Fluid Flows, Annu. Rev. Fluid Mech. 30 (1) (1998) 329–364.
doi:10.1146/annurev.fluid.30.1.329.
[25] C. K. Aidun, J. R. Clausen, Lattice-Boltzmann method for complex flows, Annu. Rev. Fluid Mech. 42 (1) (2010) 439–472.
doi:10.1146/annurev-fluid-121108-145519.
[26] A. J. C. Ladd, Numerical simulations of particulate suspensions via a discretized Boltzmann equation. Part 1. Theoretical
foundation, J. Fluid Mech. 271 (1994) 285–309. doi:10.1017/S0022112094001771.
[27] N. Q. Nguyen, A. J. C. Ladd, Lubrication corrections for lattice-Boltzmann simulations of particle suspensions, Phys.
Rev. E 66 (4) (2002) 046708. doi:10.1103/PhysRevE.66.046708.
[28] J. Götz, K. Iglberger, M. Stürmer, U. Rüde, Direct Numerical Simulation of Particulate Flows on 294912 Processor
Cores, in: Proc. 2010 ACM/IEEE Proc. Int. Conf. for High Performance Computing, Networking, Storage and Analysis,
SC ’10, IEEE, 2010, pp. 1–11. doi:10.1109/SC.2010.20.
[29] I. Ginzburg, J.-P. Carlier, C. Kao, Lattice Boltzmann approach to Richards’ equation, in: W. G. G. C. T. Miller,
M. W. Farthing, G. F. Pinder (Eds.), Computational Methods in Water Resources, Vol. 55 of Developments in Water
Science, Elsevier, 2004, pp. 583–595. doi:10.1016/S0167-5648(04)80083-2.
[30] I. Ginzburg, F. Verhaeghe, D. d’Humières, Two-relaxation-time lattice Boltzmann scheme: About parametrization,
velocity, pressure and mixed boundary conditions, Commun. Comput. Phys. 3 (2) (2008) 427–478.
[31] D. Bartuschat, U. Rüde, Parallel Multiphysics Simulations of Charged Particles in Microfluidic Flows, J. Comput. Sci.
8 (0) (2015) 1–19. doi:10.1016/j.jocs.2015.02.006.
[32] D. Bartuschat, E. Fischermeier, K. Gustavsson, U. Rüde, Two Computational Models for Simulating the Tumbling
30
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
31
Motion of Elongated Particles in Fluids, Comput. Fluids 127 (2016) 17–35. doi:10.1016/j.compfluid.2015.12.010.
[33] S. Jomeh, M. Hoorfar, Study of the effect of electric field and electroneutrality on transport of biomolecules in microreactors, Microfluid Nanofluid. 12 (1) (2012) 279–294. doi:10.1007/s10404-011-0871-9.
[34] P. Kler, C. Berli, F. Guarnieri, Modeling and high performance simulation of electrophoretic techniques in microfluidic
chips, Microfluid Nanofluid. 10 (1) (2011) 187–198. doi:10.1007/s10404-010-0660-x.
[35] M. Chau, T. Garcia, P. Spiteri, Asynchronous grid computing for the simulation of the 3D electrophoresis coupled
problem, Adv. Eng. Softw. 60–61 (2013) 111–121. doi:10.1016/j.advengsoft.2012.11.010.
[36] M. Chau, P. Spiteri, H. C. Boisson, Parallel numerical simulation for the coupled problem of continuous flow electrophoresis, Int. J. Numer. Meth. Fluids 55 (10) (2007) 945–963. doi:10.1002/fld.1502.
[37] B. Giera, L. A. Zepeda-Ruiz, A. J. Pascall, J. D. Kuntz, C. M. Spadaccini, T. H. Weisgraber, Mesoscale particle-based
model of electrophoresis, J. Electrochem. Soc. 162 (11) (2015) D3030–D3035. doi:10.1149/2.0161511jes.
[38] R. G. M. van der Sman, Simulations of confined suspension flow at multiple length scales, Soft Matter 5 (2009) 4376–4387.
doi:10.1039/B915749M.
[39] J. Smiatek, F. Schmid, Mesoscopic Simulations of Electroosmotic Flow and Electrophoresis in Nanochannels, Comput.
Phys. Commun. 182 (9) (2011) 1941 – 1944, computer Physics Communications Special Edition for Conference on
Computational Physics Trondheim, Norway, June 23-26, 2010. doi:10.1016/j.cpc.2010.11.021.
[40] R. Wang, J.-S. Wang, G.-R. Liu, J. Han, Y.-Z. Chen, Simulation of DNA electrophoresis in systems of large number
of solvent particles by coarse-grained hybrid molecular dynamics approach, J. Comput. Chem. 30 (4) (2009) 505–513.
doi:10.1002/jcc.21081.
[41] M. Kuron, G. Rempfer, F. Schornbaum, M. Bauer, C. Godenschwager, C. Holm, J. de Graaf, Moving charged particles
in lattice Boltzmann-based electrokinetics, J. Chem. Phys. 145 (21) (2016) 214102. doi:10.1063/1.4968596.
[42] J. S. Park, D. Saintillan, Direct Numerical Simulations of Electrophoretic Deposition of Charged Colloidal Suspensions,
in: R. C. A. R. Boccaccini, O. Van der Biest, J. Dickerson (Eds.), Key Engineering Materials, Vol. 507, Trans Tech Publ,
2012, pp. 47–51. doi:10.4028/www.scientific.net/KEM.507.47.
[43] J.-P. Hsu, C.-H. Chou, C.-C. Kuo, S. Tseng, R. Wu, Electrophoresis of an arbitrarily oriented toroid in an unbounded
electrolyte solution, Colloids Surf., B 82 (2) (2011) 505–512. doi:10.1016/j.colsurfb.2010.10.009.
[44] J.-P. Hsu, M.-H. Ku, Boundary effect on electrophoresis: finite cylinder in a cylindrical pore, J. Colloid Interface Sci.
283 (2) (2005) 592–600. doi:10.1016/j.jcis.2004.09.004.
[45] J.-P. Hsu, L.-H. Yeh, Electrophoresis of Two Identical Rigid Spheres in a Charged Cylindrical Pore, J. Phys. Chem. B
111 (10) (2007) 2579–2586. doi:10.1021/jp068407z.
[46] J. H. Masliyah, S. Bhattacharjee, Electrokinetic and Colloid Transport Phenomena, John Wiley & Sons, Inc., 2006.
[47] S. Tseng, C.-H. Huang, J.-P. Hsu, Electrophoresis of two spheres: Influence of double layer and van der Waals interactions,
J. Colloid Interface Sci. 451 (2015) 170–176. doi:10.1016/j.jcis.2015.03.060.
[48] R. Schmitz, V. Starchenko, B. Dünweg, Computer simulation of electrokinetics in colloidal systems, Eur. Phys. J. Spec.
Topics 222 (11) (2013) 2873–2880. doi:10.1140/epjst/e2013-02063-2.
[49] M. Baptista, R. Schmitz, B. Dünweg, Simple and robust solver for the Poisson-Boltzmann equation, Phys. Rev. E 80
(2009) 016705. doi:10.1103/PhysRevE.80.016705.
[50] H. H. Hu, N. A. Patankar, M. Y. Zhu, Direct Numerical Simulations of Fluid-Solid Systems Using the Arbitrary
Lagrangian-Eulerian Technique, J. Comput. Phys. 169 (2) (2001) 427–462. doi:10.1006/jcph.2000.6592.
[51] T. J. R. Hughes, W. K. Liu, T. K. Zimmermann, Lagrangian-Eulerian finite element formulation for incompressible
viscous flows, Compu. Meth. Appl. Mech. and Engin. 29 (3) (1981) 329–349. doi:10.1016/0045-7825(81)90049-9.
[52] C. Ye, D. Li, 3-D transient electrophoretic motion of a spherical particle in a T-shaped rectangular microchannel, J.
Colloid Interface Sci. 272 (2) (2004) 480–488. doi:10.1016/j.jcis.2003.11.014.
[53] H. Tanaka, T. Araki, Simulation Method of Colloidal Suspensions with Hydrodynamic Interactions: Fluid Particle
Dynamics, Phys. Rev. Lett. 85 (2000) 1338–1341. doi:10.1103/PhysRevLett.85.1338.
[54] H. Kodama, K. Takeshita, T. Araki, H. Tanaka, Fluid particle dynamics simulation of charged colloidal suspensions, J.
Phys. Condens. Matter 16 (10) (2004) L115–L123. doi:10.1088/0953-8984/16/10/L01.
[55] T. Araki, H. Tanaka, Physical principle for optimizing electrophoretic separation of charged particles, Europhys. Lett.
82 (1) (2008) 18004. doi:10.1209/0295-5075/82/18004.
[56] Y. Nakayama, R. Yamamoto, Simulation method to resolve hydrodynamic interactions in colloidal dispersions, Phys.
Rev. E 71 (2005) 036707. doi:10.1103/PhysRevE.71.036707.
[57] K. Kim, Y. Nakayama, R. Yamamoto, Direct Numerical Simulations of Electrophoresis of Charged Colloids, Phys. Rev.
Lett. 96 (2006) 208302. doi:10.1103/PhysRevLett.96.208302.
[58] C. Shih, R. Yamamoto, Dynamic electrophoresis of charged colloids in an oscillating electric field, Phys. Rev. E 89 (2014)
062317. doi:10.1103/PhysRevE.89.062317.
[59] X. Luo, A. Beskok, G. E. Karniadakis, Modeling Electrokinetic Flows by the Smoothed Profile Method, J. Comput.
Phys. 229 (10) (2010) 3828 – 3847. doi:10.1016/j.jcp.2010.01.030.
[60] C. S. Peskin, Flow patterns around heart valves: A numerical method, J. Comput. Phys. 10 (2) (1972) 252–271. doi:
10.1016/0021-9991(72)90065-4.
[61] R. Mittal, G. Iaccarino, Immersed boundary methods, Annu. Rev. Fluid Mech. 37 (2005) 239–261. doi:10.1146/annurev.
fluid.37.061903.175743.
[62] M. Uhlmann, An immersed boundary method with direct forcing for the simulation of particulate flows, J. Comput.
Phys. 209 (2) (2005) 448–476. doi:10.1016/j.jcp.2005.03.017.
[63] S. Kang, Direct simulations on the electrophoretic motion of multiple charged particles using an immersed boundary
method, Comput. Fluids 73 (2013) 10–23. doi:10.1016/j.compfluid.2012.12.005.
31
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
32
[64] V. Lobaskin, B. Dünweg, C. Holm, Electrophoretic mobility of a charged colloidal particle: a computer simulation study,
J. Phys. Condens. Matter 16 (38) (2004) S4063–S4073. doi:10.1088/0953-8984/16/38/021.
[65] T. Y. Molotilin, V. Lobaskin, O. I. Vinogradova, Electrophoresis of Janus particles: A molecular dynamics simulation
study, J. Chem. Phys. 145 (24) (2016) 244704. doi:10.1063/1.4972522.
[66] A. Chatterji, J. Horbach, The role of effective charges in the electrophoresis of highly charged colloids, J. Phys. Condens.
Matter 22 (49) (2010) 494102. doi:10.1088/0953-8984/22/49/494102.
[67] O. A. Hickey, C. Holm, J. Smiatek, Lattice-Boltzmann simulations of the electrophoretic stretching of polyelectrolytes:
The importance of hydrodynamic interactions, J. Chem. Phys. 140 (16) (2014) 164904. doi:10.1063/1.4872366.
[68] F. Capuani, I. Pagonabarraga, D. Frenkel, Discrete solution of the electrokinetic equations, J. Chem. Phys. 121 (2) (2004)
973–986. doi:10.1063/1.1760739.
[69] G. Giupponi, I. Pagonabarraga, Determination of the zeta potential for highly charged colloidal suspensions, Phil. Trans.
R. Soc. A: Mathematical, Physical and Engineering Sciences 369 (2011) 2546–2554. doi:10.1098/rsta.2011.0024.
[70] G. K. Batchelor, An Introduction to Fluid Dynamics, Cambridge Univ. Press, 1979.
[71] M. A. van der Hoef, M. Ye, M. van Sint Annaland, A. T. Andrews, S. Sundaresan, J. A. M. Kuipers, Multi-Scale
Modeling of Gas-Fluidized Beds, in: G. B. Marin (Ed.), Computational Fluid Dynamics, Vol. 31 of Advances in Chemical
Engineering, Elsevier, 2006, pp. 65–149. doi:10.1016/S0065-2377(06)31002-2.
[72] L. D. Landau, E. M. Lifshitz, Fluid Mechanics, Second Edition: Volume 6 (Course of Theoretical Physics), 2nd Edition,
Vol. 6 of Course of theoretical physics / by L. D. Landau and E. M. Lifshitz, Vol. 6, Butterworth-Heinemann, 1987.
[73] P. J. Dellar, Bulk and shear viscosities in lattice Boltzmann equations, Phys. Rev. E 64 (3) (2001) 031203. doi:
10.1103/PhysRevE.64.031203.
[74] J. Happel, H. Brenner, Low Reynolds number hydrodynamics: with special applications to particulate media, Vol. 1,
Springer, 1983.
[75] F. Durst, Grundlagen Der Strömungsmechanik, Springer, 2006.
[76] D. Bartuschat, Direct Numerical Simulation of Particle-Laden Electrokinetic Flows on High Performance Computers,
Ph.D. thesis, University of Erlangen-Nürnberg (2016).
URL https://opus4.kobv.de/opus4-fau/frontdoor/index/index/docId/7298
[77] J. Anderson, Computational Fluid Dynamics, Computational Fluid Dynamics: The Basics with Applications, McGrawHill Education, 1995.
[78] G. G. Stokes, On the effect of the internal friction of fluids on the motion of pendulums, Vol. 9, Pitt Press, 1851.
[79] J. Happel, E. Bart, The settling of a sphere along the axis of a long square duct at low Reynolds’ number, Appl. Sci.
Res. 29 (1) (1974) 241–258. doi:10.1007/BF00384149.
[80] A. Miyamura, S. Iwasaki, T. Ishii, Experimental wall correction factors of single solid spheres in triangular and square
cylinders, and parallel plates, Int. J. Multiph. Flow 7 (1) (1981) 41–46. doi:10.1016/0301-9322(81)90013-6.
[81] R. J. Hunter, Zeta Potential in Colloid Science: Principles and Applications, Academic Press, London, 1981.
[82] P. Debye, E. Hückel, Zur Theorie der Elektrolyte. I. Gefrierpunktserniedrigung und verwandte Erscheinungen, Phys. Z
24 (9) (1923) 185–206.
[83] D. C. Henry, The cataphoresis of suspended particles. Part I. The equation of cataphoresis, Proc. R. Soc. Lond. A
133 (821) (1931) 106–129. doi:10.1098/rspa.1931.0133.
[84] H. Ohshima, A Simple Expression for Henry’s Function for the Retardation Effect in Electrophoresis of Spherical Colloidal
Particles, J. Colloid Interface Sci. 168 (1) (1994) 269–271. doi:10.1006/jcis.1994.1419.
[85] H. Ohshima (Ed.), Theory of Colloid and Interfacial Electric Phenomena, Vol. 12 of Interface Science and Technology,
Elsevier, 2006. doi:10.1016/S1573-4285(06)80022-0.
[86] D. J. Shaw, Introduction to Colloid and Surface Chemistry, 4th Edition, Butterworth-Heinemann, 1992.
[87] H. Ohshima, T. W. Healy, L. R. White, Accurate analytic expressions for the surface charge density/surface potential
relationship and double-layer potential distribution for a spherical colloidal particle, J. Colloid Interface Sci. 90 (1) (1982)
17–26. doi:10.1016/0021-9797(82)90393-9.
[88] A. L. Loeb, J. T. G. Overbeek, P. H. Wiersema, The Electrical Double Layer Around a Spherical Colloid Particle, J.
Electrochem. Soc. 108 (12) (1961) 269C. doi:10.1149/1.2427992.
[89] D. Hänel, Molekulare Gasdynamik, Springer, 2004.
[90] S. Chapman, T. G. Cowling, The Mathematical Theory of Non-Uniform Gases, 3rd Edition, Cambridge Univ. Press,
1990.
[91] D. A. Wolf-Gladrow, Lattice-Gas Cellular Automata and Lattice Boltzmann Models: An Introduction, no. 1725 in
Lattice-gas Cellular Automata and Lattice Boltzmann Models: An Introduction, Springer, 2000.
[92] D. d’Humières, Generalized lattice-Boltzmann equations, in: Rarefied Gas Dynamics: Theory and Simulations, Vol. 159
of Prog. Astronaut. Aeronaut., AIAA, 1992, pp. 450–458. doi:10.2514/5.9781600866319.0450.0458.
[93] X. He, L.-S. Luo, Lattice Boltzmann Model for the Incompressible Navier-Stokes Equation, J. Stat. Phys. 88 (3) (1997)
927–944. doi:10.1023/B:JOSS.0000015179.12689.e4.
[94] Y. H. Qian, D. d’Humières, P. Lallemand, Lattice BGK Models for Navier-Stokes Equation, Europhys. Lett. 17 (6) (1992)
479. doi:10.1209/0295-5075/17/6/001.
[95] J. D. Sterling, S. Chen, Stability Analysis of Lattice Boltzmann Methods, J. Comput. Phys. 123 (1) (1996) 196–206.
doi:10.1006/jcph.1996.0016.
[96] L.-S. Luo, W. Liao, X. Chen, Y. Peng, W. Zhang, Numerics of the lattice Boltzmann method: Effects of collision models
on the lattice Boltzmann simulations, Phys. Rev. E 83 (5) (2011) 056710. doi:10.1103/PhysRevE.83.056710.
[97] P. L. Bhatnagar, E. P. Gross, M. Krook, A Model for Collision Processes in Gases. I. Small Amplitude Processes in
Charged and Neutral One-Component Systems, Phys. Rev. 94 (3) (1954) 511–525. doi:10.1103/PhysRev.94.511.
32
D. Bartuschat and U. Rüde / Journal of Computational Physics 00 (2017) 1–33
33
[98] M. Geier, M. Schönherr, A. Pasquali, M. Krafczyk, The cumulant lattice Boltzmann equation in three dimensions: Theory
and validation, Comput. Math. Appl. 70 (4) (2015) 507–547. doi:10.1016/j.camwa.2015.05.001.
[99] F. J. Higuera, S. Succi, R. Benzi, Lattice Gas Dynamics with Enhanced Collisions, Europhys. Lett. 9 (4) (1989) 345.
doi:10.1209/0295-5075/9/4/008.
[100] I. Ginzburg, F. Verhaeghe, D. d’Humières, Study of Simple Hydrodynamic Solutions with the Two-Relaxation-Times
Lattice Boltzmann Scheme, Commun. Comput. Phys. 3 (3) (2008) 519–581.
[101] L.-S. Luo, Analytic Solutions of Linearized Lattice Boltzmann Equation for Simple Flows, J. Stat. Phys. 88 (3/4) (1997)
913–926. doi:10.1023/B:JOSS.0000015178.19008.78.
[102] A. J. C. Ladd, R. Verberg, Lattice-Boltzmann Simulations of Particle-Fluid Suspensions, J. Stat. Phys. 104 (5) (2001)
1191–1251. doi:10.1023/A:1010414013942.
[103] Z. Guo, C. Zheng, B. Shi, Discrete lattice effects on the forcing term in the lattice Boltzmann method, Phys. Rev. E
65 (4) (2002) 046308. doi:10.1103/PhysRevE.65.046308.
[104] C. Rettinger, U. Rüde, A comparative study of fluid-particle coupling methods for fully resolved lattice Boltzmann
simulations, Comput. Fluids 154 (2017) 74–89. doi:10.1016/j.compfluid.2017.05.033.
[105] E. Fattahi, C. Waluga, B. Wohlmuth, U. Rüde, M. Manhart, R. Helmig, Lattice Boltzmann methods in porous media
simulations: From laminar to turbulent flow, Comput. Fluids 140 (2016) 247 – 259. doi:10.1016/j.compfluid.2016.
10.007.
[106] S. Bogner, S. Mohanty, U. Rüde, Drag correlation for dilute and moderately dense fluid-particle systems using the lattice
Boltzmann method, Int. J. Multiph. Flow 68 (2015) 71–79. doi:10.1016/j.ijmultiphaseflow.2014.10.001.
[107] P. Wesseling, Introduction To Multigrid Methods, R. T. Edwards, 2004.
[108] P. Knabner, L. Angermann, Numerical Methods for Elliptic and Parabolic Partial Differential Equations, Texts in Applied
Mathematics, Springer, 2003.
[109] B. I. des Poids et Mesures, The International System of Units (SI), 8th Edition, BIPM, 2006 [cited Oct. 2015].
URL http://www.bipm.org/en/publications/si-brochure/
[110] I. 80000-1:2009, Quantities and units – Part 1: General, ISO/IEC, 2009 [cited Oct. 2015].
URL https://www.iso.org/obp/ui/#iso:std:30669:en
[111] S. Donath, Wetting Models for a Parallel High-Performance Free Surface Lattice Boltzmann Method, Ph.D. thesis,
University of Erlangen-Nürnberg (2011).
URL http://www.dr.hut-verlag.de/978-3-8439-0066-9.html
[112] D. Bartuschat, D. Ritter, U. Rüde, Parallel multigrid for electrokinetic simulation in particle-fluid flows, in: High
Performance Computing and Simulation (HPCS). Madrid 2012, IEEE, 2012, pp. 374–380. doi:10.1109/HPCSim.2012.
6266940.
[113] R. Ramadugu, S. P. Thampi, R. Adhikari, S. Succi, S. Ansumali, Lattice differential operators for computational physics,
Europhys. Lett. 101 (5) (2013) 50006. doi:10.1209/0295-5075/101/50006.
[114] K. Iglberger, Software design of a massively parallel rigid body framework, Ph.D. thesis, University of Erlangen-Nürnberg
(2010).
URL http://www.dr.hut-verlag.de/978-3-86853-736-9.html
[115] L. Hufnagel, Transient simulation of electric double layers in electrokinetic flows with a coupled lattice Boltzmann
Link-Flux method, Bachelor thesis, Lehrstuhl für Informatik 10 (Systemsimulation), Universität Erlangen-Nürnberg
(November 2014).
33
| 5cs.CE
|
HIGH EFFICIENCY COMPRESSION FOR OBJECT DETECTION
Hyomin Choi and Ivan V. Bajić
School of Engineering Science, Simon Fraser University, Burnaby, BC, Canada
arXiv:1710.11151v2 [eess.IV] 16 Feb 2018
ABSTRACT
Image and video compression has traditionally been tailored
to human vision. However, modern applications such as visual analytics and surveillance rely on computers “seeing”
and analyzing the images before (or instead of) humans. For
these applications, it is important to adjust compression to
computer vision. In this paper we present a bit allocation and
rate control strategy that is tailored to object detection. Using the initial convolutional layers of a state-of-the-art object
detector, we create an importance map that can guide bit allocation to areas that are important for object detection. The
proposed method enables bit rate savings of 7% or more compared to default HEVC, at the equivalent object detection rate.
Index Terms— Bit allocation, rate control, HEVC, object
detection, YOLO
1. INTRODUCTION
Human perceptual quality has always been among the main
guiding principles of image and video compression. This influence can be seen throughout the history of development of
image and video codecs: from perceptually-optimized quantization matrices in JPEG [1] to the perceptual rate control [2,
3] for High Efficiency Video Coding (HEVC) [4]. However,
modern multimedia applications do not have humans as the
only users. In many cases, for example surveillance and visual analytics, computers must “see” and examine images or
video before humans do. Often, the first step of computer
vision would be to detect objects, after which higher-level analytics such as activity recognition or anomaly detection can
be performed.
Despite its importance for these applications, image and
video coding tailored to computer (as opposed to human) vision has been largely unexplored. Among the few studies
to tackle this topic is gradient-preserving quantization [5],
which attempts to adjust quantiztion in image compression in
order to preserve gradient information. The motivation is that
gradients are useful features in a number of computer vision
problems, so well-preserved gradients will likely improve the
accuracy of the vision pipeline. Another recent work [6] develops a rate control scheme for H.264/AVC video coding that
preserves SIFT [7] and SURF [8] features, which have also
been found useful in many computer vision problems. These
studies ([5, 6]) have proposed ways to preserve well-known
handcrafted features through the compression process, without focusing on any particular problem. However, the recent
trend in computer vision has been away from handcrafted features and towards learnt features, especially the features learnt
by deep neural networks (DNNs) [9] for specific problems.
In this paper we develop a bit allocation and rate control
method that improves object detection of a DNN-based stateof-the-art object detector called YOLO9000 [10]. We utilize
the outputs of the initial convolutional layers of this detector to create the importance map, which is used to guide bit
allocation towards regions that are important for object detection. The resulting strategy offers significant bit savings of
7% or more compared to the default HEVC at the equivalent
object detection rate. For the same bitrate, the proposed strategy offers more accurate object detection and classification
compared to the default HEVC.
The paper is organized as follows. Section 2 describes the
creation of object importance maps from the outputs of convolutional layers, and presents the related bit allocation and
rate control strategies. Section 3 presents the experimental
results and Section 4 concludes the paper.
2. PROPOSED METHODS
2.1. Background
In a convolutional neural network, convolutional layers compute cross-correlation between the input and a set of filters [9]. The cross-correlation is usually followed by maxpooling, which selects the local maximum within each small
window of the cross-correlation output. Large values therefore tend to propagate through the network towards the final
layers, where they contribute to the final output. It is important to appreciate that filter coefficients are computed during
the training process to maximize the performance on a given
task. Hence, DNN-based object detectors have filters whose
coefficients have been tuned to extract the features relevant
to detecting the objects that the network was trained on. And
because max-pooling suppresses small outputs, it follows that
large outputs are the ones that are relevant for detection.
The input size of the YOLO9000 object detector [10] is
fixed at 416 × 416. If an input image has different resolution,
say W ×H, the image resolution is first scaled (while keeping
Fig. 2. The proposed object detection-friendly compression
framework
(a)
(b)
(c)
Fig. 1. Examples of (a) test images and outputs of selected
filters in the (b) first and (c) third convolutional layers
where l indicates the layer, x and y are spatial coordinates,
and n is the filter index in the given layer. Then, all clamped
outputs are stacked in a tensor
h
i
(1) (1)
(2) (2)
(N ) (N )
vl (x, y) = αl ψ̂l (x, y) , αl ψ̂l (x, y) , . . . , αl ψ̂l (x, y)
the aspect ratio) and centered so that it fits the input. The
scaling constants for various layers are Sl = Cl /max{W, H},
where Cl is the spatial dimension of layer l, so C1 = 416,
C2 = 208, etc.
The first convolutional layer employs 32 filters with kernel size 3 × 3, and produces 32 outputs. This is followed by
max-pooling over 2 × 2 windows. The subsequent convolutional layers operate on the previous layer’s outputs. There
are total 32 layers in the YOLO9000 architecture. Fig. 1
shows several input images, and the corresponding outputs
of several filters in the first and third convolutional layer. The
brighter pixels in the output indicate the higher correlation
with the associated filter. As seen in this figure, even the early
layers in the convolutional network are able to provide some
information about the objects, although precise object location and class is not available until upper layers of the network
complete their processing.
Based on this reasoning, we propose an object detectionfriendly compression framework shown in Fig. 2. The input
image is processed by the initial convolutional layers of the
object detector. The filters in each layer can run in parallel,
so this process is highly parallelizable. From the resulting
filter outputs, we construct an object importance map, which
guides bit allocation and rate control in HEVC. The resulting image turns out to be more object detection-friendly, as
demonstrated in Section 2.2.
2.2. Object importance map
The object importance map is meant to indicate how important is each pixel to object detection. The YOLO9000 architecture employs leaky activation, which means that layer outputs can be negative. We first clamp the outputs to the range
[0, 1] as
(n)
ψ̂l
n
n
oo
(n)
(x, y) = max 0, min 1, ψl (x, y)
(1)
(2)
(n)
where αl is a weight factor for the n-th filter in layer l.
The weights are meant to indicate how informative is a
particular filter’s output for a given input image. Ideally, the
filter’s output would be high near the objects of interest and
low elsewhere. As seen in Fig. 1, filters’ outputs are not
equally informative about the objects in the image. Moreover,
a certain filter may be very informative on one image, and not
very informative on another image, which means that weights
should be adapted from image to image. We experimented
with entropy of the filter output as a guide to set weights
(lower entropy inducing higher weight), but eventually settled for a simpler approach that gave slightly better results. In
particular, we set the weight as 1 minus the average clamped
output:
(n)
αl
=1−
H
l −1 W
l −1
X
X
1
(n)
ψ̂ (x, y)
Wl · Hl y=0 x=0 l
(3)
where Wl = W/Sl and Hl = H/Sl are the width and height
of the filter’s output on the particular image at level l. When
the filter produces high responses across the entire image
(i.e., it is not very informative), the average is high, so its
weight becomes low. If the filter’s output is low on average,
its weight becomes high. Therefore, vl (x, y) in Eq. (2) will
be high only when the filter is informative (high weight) and
has a high response at the particular (x, y). Finally, we take
the `2 norm of vl (x, y), Ol (x, y) = kvl (x, y)k2 , and then
normalize Ol (x, y) by linearly mapping it to the range [0, 1]
to produce the final importance map Õl (x, y). Figure 3 shows
several importance maps generated from the first and third
layer on different images.
2.3. Bit allocation and rate control
The proposed bit allocation makes use of the object importance map Õl (x, y) to decide how to spend bits.
importance of that block
IF (i, j)
w (i, j) = P P
IF (i, j)
(6)
where again the double summation is over all valid (i, j).
Then the total target bits for the block can be computed as
Blk
Tbits
(a)
(b)
Fig. 3. The object importance maps combined using the outputs of the (a) first and (b) third layer
(i, j) =
Lbits (i, j) +
Lblk (i,j)(Lbits (i,j)−LEst
bits (i,j))
SW
w(i, j)
(7)
where Lbits (i, j) are the remaining bits in the bit budget before coding block (i, j), Lblk (i, j) is the number of remaining
blocks to be coded, including (i, j), and SW is the sliding
window size used to smooth out the bit variation (we use the
default SW = 4 from the HM16.12 [13]). LEst
bits (i, j) is an
estimate of the bits that will actually be used for coding the
block (i, j) and all subsequent blocks, computed as
XX
LEst
bppp (u, v) · Npixels (u, v) (8)
bits (i, j) =
u≥i v≥j
First, the pixel-wise importance Õl (x, y) is converted into
block-wise importance I(i, j). The size of the block is the
size of the corresponding coding unit scaled by Sl . Then
I(i, j) is computed by summing Õl (x, y) within the corresponding block and dividing by the total sum over the importance map. From here on, we use (i, j) as the coordinates of
the top-left corner of the block.
We then calculate the initial coarse estimates of the bits
per pixel (bppcoarse ) for each block as
bppcoarse (i, j) =
1
Npixels (i, j)
· I (i, j) · Tbits
3. EXPERIMENTS
(4)
where Npixels (i, j) is the number of pixels in the block whose
top-left corner is at (i, j) and Tbits represents the number of
target bits for the image to be coded. In Eq. (4), the calculated
bppcoarse (i, j) could possibly be zero, which turns out to be
harmful in subsequent encoding. In order to refine this coarse
estimate, we run the R-λ model [11, 12], which is default
rate control model in the HEVC reference software [13].
Specifically, by inputting Tbits and bppcoarse (i, j) into the
R-λ model, we compute the slice/picture-level QPs and the
block-level preliminary quantization parameter QPp (i, j).
Note that QPp (i, j) is bounded by QPs ± 2 by default, however we extend the range to QPs ± 3. Then, the preliminary
bits per pixel bppp (i, j) is computed by inverting the R-λ
model with QPp (i, j) as the input. Finally, QPp (i, j) is
incremented by 1 if I (i, j) = 0.
The final block importance is computed as
bppp (i, j)
IF (i, j) = P P
bppp (i, j)
where the summations are starting at (i, j) and going over all
subsequent valid block indices (u, v).
Finally, we estimate the actual QP values QPa by inBlk
putting Tbits
(i, j)/Npixels (i, j) to the R-λ model. The resulting QPa is bounded by QPs ± 2 by default. However, we
shift the bound upward as [QPs , QPs + 4] if QPp ≥ QPs + 3.
This has the potential to save bits in less important regions.
(5)
where the double summation is over all valid (i, j). Using
this, we compute the weight for each block as the normalized
In this section, we assess the performance of the proposed bit
allocation and rate control scheme in terms of its effect on
object detection. The proposed methods were implemented
in HEVC reference software HM16.12 [13]. The YOLO9000
model in the Darknet framework [14] is used for the object
detection performance evaluation.
Bjøntegaard Delta (BD) [15] is a standard measurement
method for evaluating compression performance. It compares
the average bit rates of two coding methods at the equivalent
quality metric. Usually, the quality metric is Peak Signal-toNoise Ratio (PSNR) and we refer to this measurement as BD
bitrate for PSNR (BD-BR-PSNR). However, it is also possible to use quality metrics other than PSNR in the BD analysis. Specifically, since our goal is to compare object detection
performance between methods, instead of PSNR we use the
standard object detection accuracy metric called mean Average Precision (mAP) [16]. The mAP is in the range [0, 1]. By
computing BD over rate vs. mAP curves, we can obtain the
average bit rate saving (or increment) that one compression
method would have over another at the equivalent mAP. We
call this metric BD bitrate for mAP (BD-BR-mAP).
For testing, we employ the widely used PASCAL VOC
2007 dataset [16], which has 9963 images out of which 4952
HM (QP=22)
HM-RC
Proposed
HM (QP=22)
HM-RC
Proposed
Increasing correct detections
Decreasing false detections
Increasing correct detections
Alternative object class
Fig. 4. Examples illustrating different object detection/classification produced by YOLO9000 on images encoded by HM with
QP=22, HM-RC at the same bitrate, and the proposed method at the same bitrate, with importance maps computed from the
third convolutional layer.
are test images. The images are annotated with 20 different object classes, such as aeroplane, bicycle, bird, and so
on. For encoding, 16×16 CTU is adopted and RDOQ tool is
off, but other coding parameters follow the common HEVC
test conditions [17] of the Main Still Picture Profile [18]. We
first encode each test image using the default HM with QP∈
{22, 27, 32, 37}. The resulting bits are used as target bits
for the default HM rate control (HM-RC) and our proposed
method. For the proposed method we construct the importance maps from the outputs of the first, third, and seventh
layer, in order to examine the behaviour of the system with
different importance maps.
maps can successfully guide bit allocation towards regions
that are most relevant for object detection. Fig. 4 shows a few
examples where YOLO9000 produces different detections on
the images encoded by HM, HM-RC, and proposed methods.
Although the images look very similar visually, detection on
images encoded by the proposed method is the most accurate. This again illustrates that importance maps successfully
guide bit allocation towards regions that are most relevant for
object detection.
All encoded images are then decoded and fed to the
YOLO9000 object detector. mAP is computed by comparing
detector’s output with the ground truth. Table 1 shows various
comparisons among the three tested codecs: HM, HM-RC
and proposed. For the rate control accuracy, ∆bpp is the mean
absolute difference (MAD) in bits per pixel (bpp) between the
output bits of HM and the two rate control methods (HM-RC
and proposed) across all images. HM-RC shows averaged
∆bpp = 0.0483, while our rate control gave smaller deviation
in each of the three cases, with importance map computed
from the seventh layer being the most accurate. In terms of
BD-BR-PSNR, both HM-RC and our proposed method have
lower performance (positive BD-BR-PSNR) compared to the
default HM, since they both deviate from the optimal ratedistortion allocation in order to achieve different objectives.
However, our method achieves significant advantage in BDBR-mAP over both HM and HM-RC, which was the main
design objective. In particular, with the importance map computed from the output of the third convolutional layer, 7.32%
bit reduction is achieved over HM, and 8.23% reduction over
HM-RC, at an equivalent mAP. This shows that importance
We proposed a novel bit allocation and rate control strategy
whose goal was to improve object detection after decoding.
Using the outputs of the initial convolutional layers of a stateof-the-art object detector, the proposed algorithm successfully
achieved efficient bit control and improved object detection
performance over the default HEVC implementations. The
proposed strategy can be used in many applications where
computers “see” and analyze the data before (or instead of)
humans.
4. CONCLUSION
Table 1. Performance comparison among HM, HM-RC, and
proposed method
Test cases
∆bpp
σbpp
BD-BR-PSNR
BD-BR-mAP
HM vs. HM-RC
HM vs. Ours w/ 1st L.
HM vs. Ours w/ 3rd L.
HM vs. Ours w/ 7th L.
HM-RC vs. Ours w/ 1st L.
HM-RC vs. Ours w/ 3rd L.
HM-RC vs. Ours w/ 7th L.
0.0483
0.0385
0.0372
0.0232
-
0.1187
0.1113
0.1094
0.1086
-
3.08%
7.10%
7.15%
6.96%
3.82%
3.87%
3.68%
1.67%
-3.90%
-7.32%
-6.33%
-5.31%
-8.23%
-7.10%
5. REFERENCES
[1] ITU-T Rec. T.81 and ISO/IEC 10918-1, “Information technology - digital compression and coding of
continuous-tone still images,” 1992.
[2] H. Zeng, K. N. Ngan, and M. Wang, “Perceptual adaptive lagrangian multiplier for high efficiency video coding,” in Picture Coding Symposium, Dec. 2013.
[3] A. Yang, H. Zeng, L. Ma, J. Chen, C. Cai, and K.-K. Ma,
“A perceptual-based rate control for HEVC,” in 6th Int.
Conf. Image Processing Theory Tools and Applications,
Dec. 2016.
[4] G. J. Sullivan, J.-R. Ohm, W.-J. Han, and T. Wiegand,
“Overview of the high efficiency video coding (HEVC)
standard,” IEEE Trans. Circuits Syst. Video Technol.,
vol. 22, pp. 1649–1668, Dec. 2012.
[5] M. Makar, H. Lakshman, V. Chandrasekhar, and
B. Girod, “Gradient preserving quantization,” in Proc.
IEEE ICIP’12, Sep. 2012, pp. 2505–2508.
[6] J. Chao, R. Huitl, E. Steinbach, and D. Schroeder, “A
novel rate control framework for SIFT/SURF feature
preservation in H.264/AVC video compression,” IEEE
Trans. Circuits Syst. Video Technol., vol. 25, no. 6, pp.
958–972, Jun. 2015.
[7] D. Lowe,
“Distinctive image feature from scaleinvariant keypoints,” Int. J. Computer Vision, vol. 60,
no. 2, pp. 91–110, Nov. 2004.
[8] H. Bay, T. Tuytelaars, and L. V. Gool, “Surf: Speeded
up robust features,” in Proc. European Conf. Computer
Vision, May. 2006, pp. 404–417.
[9] Ian Goodfellow, Yoshua Bengio, and Aaron Courville,
Deep Learning, MIT Press, 2016.
[10] J. Redmon and A. Farhadi, “YOLO9000: better, faster,
stronger,” in IEEE Conf. Computer Vision and Pattern
Recognition, Jul. 2017.
[11] B. Li, H. Li, L. Li, and J. Zhang, “λ domain rate control algorithm for high efficiency video coding,” IEEE
Trans. Image Processing, vol. 23, no. 9, pp. 3841–3854,
Sep. 2014.
[12] B. Li, J. Xu, D. Zhang, and H. Li, “QP refinement according to Lagrange multiplier for high efficiency video
coding,” in Proc. IEEE Int. Symp. Circuits Syst., May
2013, pp. 477–480.
[13] “HEVC
reference
software
(HM
16.12),”
https://hevc.hhi.fraunhofer.de/trac/
hevc/browser/tags/HM-16.12,
Accessed:
2017-05-27.
[14] J. Redmon, “Darknet: Open source neural networks
in C.,” http://pjreddie.com/darknet/, 20132017, Accessed: 2017-10-19.
[15] G. Bjøntegaard, “VCEG-M33: Calculation of average
PSNR differences between RD curves,” in Video Coding
Experts Group (VCEG), Apr. 2001.
[16] M. Everingham, L. Van Gool, C. K. I.
Williams, J. Winn, and A. Zisserman,
“The
PASCAL
Visual
Object
Classes
Challenge
2007 (VOC2007) Results,”
http://www.pascalnetwork.org/challenges/VOC/voc2007/workshop/index.html.
[17] F. Bossen, “Common HM test conditions and software reference configurations,” in ISO/IEC JTC1/SC29
WG11 m28412, JCTVC-L1100, Jan. 2013.
[18] T. Nguyen and D. Marpe, “Objective performance evaluation of the HEVC Main Still Picture profile,” IEEE
Trans. Circuits Syst. Video Technol., vol. 25, no. 5, May
2015.
| 1cs.CV
|
arXiv:1803.02604v1 [math.GR] 7 Mar 2018
On certain semigroups of partial contractions of a
finite chain
Abdullahi Umar and M. M. Zubairu
∗
Department of Mathematics, The Petroleum Institute, Sas Nakhl,
Khalifa University of Science and Technology, P. O. Box 2533 Abu Dhabi, UAE
[email protected]
Department of Mathematics, Bayero University Kano, P. M. B. 3011 Kano Nigeria
[email protected]
March 8, 2018
Abstract
Let [n] = {1, 2, . . . , n} be a finite chain and let Pn be the semigroup
of partial transformations on [n]. Let CP n = {α ∈ Pn : (f or all x, y ∈
Dom α) |xα − yα| ≤ |x − y|} be the subsemigroup of partial contraction
mappings on [n]. We have shown that the semigroup CP n and some of its
subsemigroups are nonregular left abundant semigroups for all n but not
right abundant for n ≥ 4.
2010 Mathematics Subject Classification. 20M20.
1
Introduction and Preliminaries
Let [n] = {1, 2, . . . , n} be a finite chain, a map α which has domain and image
both subsets of [n] is said to be a transformation. A transformation α whose
domain is a subset of [n] (i. e., Dom α ⊆ [n]) is said to be partial. The
collection of all partial transformations of [n] is known as semigroup of partial
transformations, usually denoted by Pn . A map α ∈ Pn is said to be order
preserving (resp., order reversing) if (for all x, y ∈ Dom α) x ≤ y implies
xα ≤ yα (resp. xα ≥ yα); is order decreasing if (for all x ∈ Dom α) xα ≤ x; an
isometry (i. e., distance preserving) if (for all x, y ∈ Dom α) |xα−yα| = |x−y|;
a contraction if (for all x, y ∈ Dom α) |xα − yα| ≤ |x − y|. Let CP n = {α ∈
Pn : (f or all x, y ∈ Dom α) |xα − yα| ≤ |x − y|} and OCP n = {α ∈ CP n :
∗ Corresponding
Author.
Email: [email protected]
1
(f or all x, y ∈ Dom α) x ≤ y implies xα ≤ yα} be the subsemigroups of partial
contractions and of order preserving partial contractions of [n], respectively.
Further, the collection of all order preserving or order reversing partial contractions denoted by ORCP n is a subsemigroup of ORP n (where ORP n denotes the semigroup of order preserving or order reversing partial transformations of [n]).
In 2013, Zhao and Yang [23] characterized the Green’s relations on the subsemigroup OCP n of CP n (OCP n is the semigroup of order preserving partial
contractions on [n]). Recently, Ali et al. [2] obtained a necessary and sufficient
condition for an element in CP n to be regular and also described all its Green’s
equivalences. Most of the results concerning regularity and Green’s relations for
some subsemigroups of CP n can be deduced from the results obtained in that
paper. Zhao and Yang [23] have shown that the semigroup OCP n (n > 2) is
nonregular. Similarly, Ali et al. [2] have shown that the semigroups CP n and
ORCP n are nonregular for n > 2. Thus, there is a need to identify the class
of semigroups to which they belong, for example, whether they are abundant
semigroups. Therefore, this paper is a natural sequel to Ali et al. [2].
This section includes a brief introduction giving some basic definitions and
introducing some new concepts. In section 2, we characterize all the starred
Green’s relations on the semigroups CP n , ORCP n and OCP n and show that
D∗ = J ∗ [5]. In section 3, we show that the collection of all strongly regular
elements of ORCP n is a subsemigroup and in section 4, we show that the
semigroups CP n , ORCP n and OCP n are left abundant for all n but not right
abundant for n ≥ 4.
For standard concepts in semigroup theory, we refer the reader to Howie [12]
and Higgins [11].
Let α be element of CP n . Let Dom α, Im α, h (α) and F (α) denote, the
domain of α, image of α, | Im α| and {x ∈ Dom α : xα = x} (i. e., the set
of fixed points of α), respectively. For α, β ∈ CP n , the composition of α and β
is defined as x(α ◦ β) = ((x)α)β for any x in Dom α. Without ambiguity, we
shall be using the notation αβ to denote α ◦ β.
Next, given any transformation α in ORCP n , the domain of α is partitioned
into p − blocks by the relation ker α = {(x, y) ∈ [n] × [n] : xα = yα} and so as
in [13], α can be expressed as
A1 A2 . . . Ap
∈ Pn (1 ≤ p ≤ n),
(1)
α=
x1 x2 . . . xp
where, Ai (1 ≤ i ≤ p) are equivalence classes under the relation ker α, i. e.,
Ai = xi α−1 (1 ≤ i ≤ p) and further Ker α is ordered under the usual ordering,
i. e., Ker α = {A1 < A2 < . . . < Ap }. Thus for the rest of the content of the
paper we shall consider α to be as expressed in (1) unless otherwise specified.
Now, let Ker α = (Ai )i∈[p] = {A1 , A2 , . . . , Ap } be the partition of Dom α. A
subset Tα of [n] is said to be a transversal of the partition Ker α if |Ai ∩Tα | = 1
(1 ≤ i ≤ p). A transversal Tα is said to be relatively convex if for all x, y ∈ Tα
with x ≤ y and if x ≤ z ≤ y (z ∈ Dom α), then z ∈ Tα . Notice that every convex
2
transversal is necessarily relatively convex but not vice-versa. A transversal Tα
is said to be admissible if and only if the map Ai 7→ ti (ti ∈ Tα , i ∈ {1, 2, . . . , p})
is a contraction, see [2]. Notice that every convex transversal is admissible but
not vice-versa.
Let S be a semigroup and A be a subset of S, hAi denotes the semigroup
generated by A and is a subsemigroup of S. If hAi = S then A is said to generate
S and also hAi = A if and only if A is a subsemigroup of S. An element a in a
semigroup S is said to be an idempotent if and only if a2 = a. As usual E(S)
denotes the set of all idempotents in S. It is well known that an element α in
Pn is idempotent if and only if Im α = F (α). (Equivalently, α is idempotent
if and only if xi ∈ Ai for 1 ≤ i ≤ p, that is to say the blocks in Ker α are
stationary.)
Next, we quote some basic lemmas from [2] which would be useful for some
of the subsequent results.
Lemma 1.1 ([2], Lemma 1.3). For n ≥ 4, let α ∈ CP n such that there exists
k ∈ {2, . . . , p − 1} (3 ≤ p ≤ n) with |Ak | ≥ 2. If Ai < Aj (i < j) for all
i, j ∈ {1, 2, . . . , p} then the partition Ker α = {A1 , A2 , . . . , Ap } of Dom α has
no relatively convex transversal.
Corollary 1.2 ([2], Lemma 1.4). For n ≥ 4, let α ∈ ORCP n such that there
exists k ∈ {2, . . . , p − 1} (p ≥ 3) with |Ak | ≥ 2. Then the partition Ker α =
{A1 , A2 , . . . , Ap } of Dom α has no relatively convex transversal.
Lemma 1.3 ([2], Lemma 1.5). Let α ∈ CP n such that Ai < Aj for all i < j in
{1, 2, . . . , p} (p ≥ 3). If |Ai | = 1 for all 2 ≤ i ≤ p − 1 then the partition Ker α
of Dom α has an admissible transversal Tα .
Lemma 1.4 ([2] Lemma 1.8). Let α ∈ CP n and let A be a convex subset of
Dom α. Then Aα is convex.
Lemma 1.5 ([2], Corollary 4.2). Let α ∈ ORCP n . If | Im α| ≥ 3, then α is
regular if and only if either min Ap − xp = max A1 − x1 = d and Ai = {xi + d}
or min Ap − x1 = max A1 − xp = d and Ai = {xp−i+1 + d}, for i = 2, . . . , p − 1.
Equivalently, if Tα is admissible then α is regular if and only if α|Tα is an
isometry.
2
Starred Green’s relations
Let S be a semigroup. A relation L∗ defined as (∀ a, b ∈ S) aL∗ b if and only
if a, b are related by the Green’s L∗ relation in some oversemigroup of S, is
known as the starred Green’s L relation. The relation R∗ is defined dually,
while the relation D∗ is defined as the join of the relations L∗ and R∗ . The
intersection of L∗ and R∗ is denoted by H∗ . A semigroup S is said to be left
abundant (resp., right abundant ) if each L∗ −class (resp., R∗ −class) contains an
idempotent, it is called abundant if each L∗ −class and R∗ −class of S contains
an idempotent. An abundant semigroup in which the set E(S), of idempotents
3
of S is a subsemigroup of S is called quasi adequate and if E(S) is commutative
then S is called adequate [4, 5].
Many nonregular classes of transformation semigroups were shown to be
either abundant or adequate, for example see [18, 19, 20, 21, 15, 16, 22, 14].
Recently, AlKharousi et al. have shown that the semigroup OCI n , of all order
preserving one to one contraction maps of a finite chain is adequate [3]. In this
section we are going to show that the semigroups CP n , OCP n , CT n and OCT n
are all left abundant (for all n) but not right abundant for n ≥ 4.
We shall use the following notation from ([12], Chapter 2). If U is a subsemigroup of a semigroup S then aLU b means that there exist u, v ∈ U 1 such
that ua = b and vb = a, while aLS b means that there exist x, y ∈ S 1 such that
xa = b and yb = a. Similarly, for the relation R.
Some of the earlier results concerning starred Green’s relations on a transformation semigroup were obtained by Umar [18, 19, 20, 21], where he described
all the starred relations on the semigroups of order decreasing full and of order decreasing partial one-one transformations of a chain, these papers marked
the beginning of the study of these relations on a transformation semigroup.
Recently, Garba et al. characterized these relations on the semigroup of full
contraction maps and of order preserving full contraction maps of a finite chain:
CT n and OCT n , respectively [7]. In this section, we characterize these relations
on the more general semigroup of partial contractions CP n and its subsemigroups of order preserving or order reversing partial contraction maps of a finite
chain ORCP n , and of order preserving partial contraction maps of a finite chain
OCP n , respectively. We equally show that the relations D∗ and J ∗ coincide on
these semigroups.
To begin our investigation let us start with the following. The relations
L∗ and R∗ have the following characterizations as described in ([12], Exercise
2.6.7-9) or as described in [4].
L∗ = {(a, b) : (∀x, y ∈ S 1 ) ax = ay ⇔ bx = by}
(2)
R∗ = {(a, b) : (∀x, y ∈ S 1 ) xa = ya ⇔ xb = yb}
(3)
and
We next give the characterizations of these relations on the semigroups CP n ,
ORCP n and OCP n as follows: Let S be a semigroup in {CP n , ORCP n , OCP n }.
Theorem 2.1. Let α, β ∈ S. Then
(i) αL∗ β if and only if Im α = Im β.
(ii) αR∗ β if and only if ker α = ker β.
(iii) αH∗ β if and only if Im α = Im β and ker α = ker β.
(iv) αD∗ β if and only if | Im α| = | Im β|.
∗
Proof. (i) Let α, β be elements in S ∈ {CP n , ORCP
n , OCP n } suchthat αL β
x1 x2 . . . xp
.
and Im α = {x1 , x2 , . . . , xp }. Further, let γ =
x1 x2 . . . xp
4
Then clearly γ ∈ S and
x1 x2 . . . xp
α.
= α · 1[n] (by (2))
x1 x2 . . . xp
x1 x2 . . . xp
= β · 1[n]
⇔β·
x1 x2 . . . xp
which implies that Im β ⊆ {x1 , x2 , . . . , xp } = Im α. Similarly in the same
manner we can have Im α ⊆ Im β. Thus, Im α = Im β.
Conversely, suppose that Im α = Im β. Then by ([12], Excercise 2.6, 17)
αLPn β and it follows from definition that αL∗ β. Thus, the result follows.
(ii) Suppose that α, β ∈ S and αR∗ β. Then (x, y) ∈ ker α if and only if
Dom α
Dom α
xα = yα ⇔
◦α=
◦ α (by (3))
x
y
Dom α
Dom α
⇔
◦β =
◦ β.
x
y
⇔ xβ = yβ
⇔ (x, y) ∈ ker β.
Hence ker α = ker β.
Conversely, suppose that ker α = ker β. Then by ([12], Excercise 2.6, 17)
αRPn β and it follows from definition that αR∗ β. (iii) This follows from (i) and
(ii).
(iv) Suppose that αD∗ β. Then by ([12], Proposition 1.5.11) there exist elements γ1 , γ2 , . . . , γ2n−1 ∈ S such that αL∗ γ1 , γ1 R∗ γ2 , γ2 L∗ γ3 , . . . , γ2n−1 R∗ β
for some n ∈ N. Thus, by (i) and (ii) we have Im α = Im γ1 , ker γ1 = ker γ2 ,
Im γ2 = Im γ3 . . . , ker γ2n−1 = ker β. This implies that | Im α| = | Im γ1 | =
| Dom γ1 / ker γ1 | = | Dom γ2 / ker γ2 | = . . . = | Dom γ2n−1 / ker γ2n−1 | =
| Dom β/ ker β| = | Im β|.
Conversely, suppose that | Im α| = | Im β| where
B1 B2 . . . Bp
A1 A2 . . . Ap
(p ≤ n) (4)
and β =
α=
y1 y2 . . . yp
x1 x2 . . . xp
such that the map xi 7→
yi (i = 1, 2, . . . , p) is an isometry. Then the map
B1 B2 . . . Bp
is in S. Moreover, by (i) and (ii), it follows that
γ =
x1 x2 . . . xp
αL∗ γ and γR∗ β. Thus, by ([12] Proposition 1.5.11) we have αD∗ γ and the
proof is complete.
A left (resp. right) ∗−ideal of a semigroup S is defined as the left (rep.
right) ideal of S for which L∗a ⊆ I (resp. Ra∗ ⊆ I) for all a ∈ I. A subset
I of a semigroup S is a ∗−ideal if it is both left and right ∗−ideal of S. The
principal J ∗ ∗−ideal, J ∗ (a), generated by a ∈ S is the intersection of all ∗−ideal
of S containing a, where the relation J ∗ is defined as: aJ ∗ b if and only if
J ∗ (a) = J ∗ (b) for all a, b ∈ S. We now recognize the following lemma from
Fountain [5].
5
Lemma 2.2 ( [5]). Let a,b be elements of a semigroup S. Then b ∈ J ∗ (a) if and
only if there are elements a0 , a1 , . . . , an ∈ S, x1 , x2 , . . . , xn , y1 , y2 , . . . , yn ∈ S 1
such that a = a0 , b = an and (ai , xi ai−1 yi ) ∈ D∗ for i = 1, 2, . . . , n.
As in [19], we immediately have:
Lemma 2.3. Let S be in {CP n , ORCP n , OCP n }. Then for α, β ∈ S, α ∈
J ∗ (β) implies | Im α| ≤ | Im β|.
Proof. Let α ∈ J ∗ (β), then by lemma(2.2), there exist η0 , η1 . . . , ηn ∈ S,
ρ1 , ρ2 , . . . , ρn , τ1 , τ2 , . . . , τn ∈ S 1 such that β = η0 , α = ηn and (ηi , ρi ηi−1 τi ) ∈
D∗ for i = 1, 2, . . . , n. Thus, by Theorem(2.1)(iv), it implies that
| Im ηi | = | Im ρi ηi−1 τi | ≤ | Im ηi | f or i = 1, 2, . . . , n,
which implies that | Im α| ≤ | Im β|.
Notice that, D∗ ⊆ J ∗ and together with Lemma (2.3) we have:
Corollary 2.4. On the semigroups CP n , ORCP n or OCP n we have D∗ = J ∗ .
We now are going to show in the next lemma that if S ∈ {CP n , OCP n , ORCP n }
then S is left abundant.
Lemma 2.5. Let S ∈ {CP n , OCP n , ORCP n }. Then S is left abundant.
A1 A2 . . . Ap
Proof. Let α ∈ S and L∗α be an L∗ −class of α in S, where α =
x1 x2 . . . xp
x1 x2 . . . xp
. Clearly γ 2 = γ ∈ S and
(1 ≤ p ≤ n). Define γ =
x1 x2 . . . xp
Im α = Im γ, therefore by Theorem(2.1)(i), αL∗ γ, which means that γ ∈ L∗α .
Thus, S is left abundant, as required.
Lemma 2.6. Let S ∈ {CP n , ORCP n , OCP n }. Then for n ≥ 4, S is not right
abundant.
1 {2, 3} 4
Proof. Let n = 4 and consider α =
. Clearly α is in S and
1
2
3
1 {2, 3} 4
1 {2, 3} 4
1 {2, 3} 4
1 {2, 3} 4
∗
Rα
=
,
,
,
,
1
2
3
3
2
1
2
3
4
4
3
2
which has no idempotent element.
Remark 2.7. Let S ∈ {CP n , ORCP n , OCP n }. Then for 1 ≤ n ≤ 3, S is right
abundant.
6
3
On strongly regular elements in ORCP n
In 2018, Ali et al. [2] characterized the regular elements in ORCP n and showed
that the semigroup ORCP n is not regular. We begin the section with a remark
concerning the product of regular elements in ORCP n .
Remark 3.1. Product of regular
elements
in ORCP
n or OCP nis not neces1 3
1 {2, 3}
sarily regular. Consider α =
and β =
∈ ORCP 3 .
1 3
1
2
1 3
Then αβ =
is not regular.
1 2
Next, we recognize the following result (due to Hall ([10], Proposition 1.))
which is crucial in proving some of the results below.
Proposition 3.2. ([10], Proposition 1.) Let S be an arbitrary semigroup. Then
the following are equivalent:
(i) For all idempotents e and f of S, the element ef is regular;
(ii) Reg(S) is regular subsemigroup;
(iii) hE(S)i is a regular semigroup.
We now introduce the following definition. A regular element α in ORCP n is
said to be strongly regular if and
only ifKer α has a convex transversal Tα . For
1 3
example, the contraction α =
∈ ORCP 3 is regular but not strongly
3 1
regular,
= {1, 3} is not convex. On the other hand, the contraction
since Tα
1 {2, 3}
β=
∈ ORCP 3 is strongly regular, since Tβ = {1, 2} is convex.
3
4
Now let SReg(ORCP n ) denote the set of all strongly regular elements in
ORCP n . Then we have the following two lemmas:
Lemma 3.3.Let ǫ be an idempotent element in SReg(ORCP
n ). Then α can be
A1
a + 2 a+ 3 ... a+ p − 1
Ap
expressed as
, where a + 1 =
a + 1 a + 2 a+ 3 ... a+ p − 1 a+ p
max A1 , a + p = min Ap .
Proof. Let ǫ ∈ SReg(ORCP n ) be of height p. Then by the contrapositive of
Lemma(1.2) we see that Ker ǫ = {A1 < {a + 2} < . . . < {a + p − 1} < Ap },
and so Tǫ is convex. Thus by Lemma (1.4), Tǫ ǫ = Im ǫ is convex and hence
A1
a+ 2 a + 3 ... a+ p − 1
Ap
.
ǫ=
x + 1 x + 2 x + 3 ... x + p − 1 x + p
However, since ǫ is an idempotent then the blocks of Ker ǫ are stationary i.
e., x + 1 ∈ A1 , x + p ∈ Ap , and x + i = a + i (i = 2, . . . , p − 1), which
implies x = a. However, since ǫ is regular then it follows by Lemma(1.5) that
max A1 − (a + 1) = min Ap − (a + p) = 0, showing that max A1 = a and
min Ap = a + p, as required.
7
A1
a+ 2 ... a + p − 1
Ap
and
Lemma 3.4. Let ǫ =
a+ 1 a+ 2 ... a + p − 1 a + p
B1
b + 2 . . . b + s − 1 Bs
τ =
be two idempotent elements in
b + 1 b + 2 ... b + s− 1 b + s
SReg(ORCP n ) with p, s = 1, 2, . . . , n. Then ǫτ is strongly regular.
Proof. Let c = max{a + 1, b + 1} and d = min{a + p, b + s}. Suppose also that
F (ǫ) ∩ F (τ ) 6= ∅ and the blocks of the product ǫτ are D1 , D2 , . . . , Dk , where
k ≤ min{p, s}. Then clearly, c ≤ d. Thus we shall consider four cases:
(i) If a + 1 = c and a + p = d then b + 1 ≤ a + 1 and a + p ≤ b + s. Using
convexity, it is now not difficult to see that D1 = A1 ∪ [b + 1, a + 1],
Di = {a + i} (i = 2, . . . , k − 1) and Dk = [a + p, b + s] ∪ Ap . Moreover,
D1 ǫτ = a + 1 = max D1 and Dk ǫτ = a + p = min Dk . Hence ǫτ is a
strongly regular idempotent.
The other three cases listed below are handled similarly.
(ii) If a + 1 = c and b + s = d;
(iii) If b + 1 = c and a + p = d;
(iv) If b + 1 = c and b + s = d.
Now as a consequence, from Proposition(3.2) and Lemma(3.4), we readily
have:
Theorem 3.5. Let S = SReg(ORCP n ). Then S is a regular subsemigroup of
ORCP n .
Acknowledgements. The second named author would like to thank Bayero
University and TET Fund for financial support. He would also like to thank
The Petroleum Institute, Khalifa University of Science and Technology for hospitality during his 3-month research visit to the institution.
References
[1] Adeshola, A. D. and Umar, A. Combinatorial results for certain semigroups
of order-preserving full contraction mappings of a finite chain. J. Comb.
Maths. and Comb. Computing. To appear.
[2] Ali, B. Umar and A. Zubairu, M. M. Regularity and Green’s relations on
the semigroup of partial contractions of a finite chain. Submitted.
[3] AlKharousi, F., Garba, U. G., Ibrahim, M. J., Imam, A. T. and Umar,
A. On the semigroup of finite order-preserving partial injective contraction
mappings Submitted.
8
[4] Fountain, J. B. Adequate Semigroups. Proc. Edinb. Math. Soc. 22 (1979),
113–125.
[5] Fountain, J. B. Abundant Semigroups. Proc. Lond. Math. Soc. 44 (1982),
103–129.
[6] Garba G. U. Idempotents in partial transformation semifroups. Proc. Roy.
Soc. Edinb. 116 A (1990), 359–366.
[7] Garba, G. U., Ibrahim, M. J. and Imam, A. T. On certain semigroups of
full contraction maps of a finite chain. Turk. J. Math. 41 (2017) 500-507.
[8] Green, J. A. On the structure of semigroups, Ann. of Math. (2) 54 (1951),
163–172.
[9] Ganyushkin, O. and Mazorchuk, V. Classical Finite Transformation Semigroups. Springer−Verlag: London Limited (2009).
[10] Hall, P. Some properties of local subsemigroups inherited by larger subsemigroups. Semigroup Forum 1982, 35–49.
[11] Higgins, P. M. Techniques of semigroup theory. Oxford university Press
(1992).
[12] Howie, J. M. Fundamental of semigroup theory. London Mathematical Society, New series 12. The Clarendon Press, Oxford University Press, 1995.
[13] Howie, J. M., Robertson, E. F. and Schein, B. M. A combinatorial property of finite full transformation semigroups. Proc. Roy. Soc. Edinb. 109A
(1988), 319-328.
[14] Mendes-Gonalves, Suzana. Green’s relations, regularity and abundancy for
semigroups of quasi-onto transformations. Semigroup Forum 91 (2015), no.
1, 39-52.
[15] Pei, Huisheng and Zhou, Huijuan. Abundant semigroups of transformations
preserving an equivalence relation. Algebra Colloq. 18 (2011), no. 1, 77-82.
[16] Sun, Lei. A note on abundance of certain semigroups of transformations
with restricted range. Semigroup Forum 87 (2013), no. 3, 681-684.
[17] Umar, A. and Al-Kharousi, F. Studies in semigroup of contraction mappings of a finite chain. The Research Council of Oman Research grant
proporsal No. ORG/CBS/12/007, 6th March 2012.
[18] Umar, A. On the semigroups of order-decreasing finite full transformations.
Proc. Roy. Soc. Edinb. Sect. A 120 (1992), no. 1-2, 129–142.
[19] Umar, A. On the semigroups of partial one-one order-decreasing finite
transformations. Proc. Roy. Soc. Edinb. Sect. A 123 (1993), 355–363.
9
[20] Umar, A. On certain infinite semigroups of order-decreasing transformations I Comm. Algebra 25 (1997), 2989–2999.
[21] Umar, Abdullahi. On certain infinite semigroups of order-increasing transformations II. Arab. J. Sci. Eng. Sect. A Sci. 28 (2003), 203–210.
[22] Zhao, Ping. On the semigroups of order-preserving and A -decreasing finite
transformations. Algebra Colloq. 21 (2014), no. 4, 653-662.
[23] Zhao, P. and Yang, M. Regularity and Green’s relations on semigroups of
transformation preserving order and compression. Bull. Korean Math. Soc.
49 (2012), No. 5, 1015–1025.
10
| 4math.GR
|
Deep Visual Domain Adaptation: A Survey
arXiv:1802.03601v1 [cs.CV] 10 Feb 2018
Mei Wang, Weihong Deng
School of Information and Communication Engineering,
Beijing University of Posts and Telecommunications, Beijing, China.
[email protected], [email protected]
Abstract
time.
Fortunately, the big data era makes a large amount of
data available for other domains and tasks. For instance,
although large-scale labeled video databases that are publicly available only contain a small number of samples, statistically, the YouTube face dataset (YTF) consists of 3.4K
videos. The number of labeled still images is more than
sufficient [85]. Hence, skillfully using the auxiliary data for
the current task with scarce data will be helpful for realworld applications. However, there is always a distribution
change or domain shift between two domains that can degrade the performance, as shown in Fig. 1. Mimicking the
human vision system, domain adaptation (DA) is a particular case of transfer learning (TL) that utilizes labeled data in
one or more relevant source domains to execute new tasks
in a target domain. Over the past decades, various shallow
DA methods have been proposed to solve a domain shift
between the source and target domains. The common algorithms for shallow DA can mainly be categorized into
two classes: instance-based DA [5, 12] and feature-based
DA [30, 66, 23, 65]. The first class reduces the discrepancy by reweighting the source samples, and it trains on the
weighted source samples. For the second class, a common
shared space is generally learned in which the distributions
of the two datasets are matched.
Deep domain adaption has emerged as a new learning technique to address the lack of massive amounts of
labeled data. Compared to conventional methods, which
learn shared feature subspaces or reuse important source
instances with shallow representations, deep domain adaption methods leverage deep networks to learn more transferable representations by embedding domain adaptation
in the pipeline of deep learning. There have been comprehensive surveys for shallow domain adaption, but few
timely reviews the emerging deep learning based methods.
In this paper, we provide a comprehensive survey of deep
domain adaptation methods for computer vision applications with four major contributions. First, we present a taxonomy of different deep domain adaption scenarios according to the properties of data that define how two domains
are diverged. Second, we summarize deep domain adaption
approaches into several categories based on training loss,
and analyze and compare briefly the state-of-the-art methods under these categories. Third, we overview the computer vision applications that go beyond image classification, such as face recognition, semantic segmentation and
object detection. Fourth, some potential deficiencies of current methods and several future directions are highlighted.
Recently, neural-network-based deep learning approaches have achieved many inspiring results in visual categorization applications, such as image classification [53],
face recognition [90], and object detection [28]. Simulating the perception of the human brain, deep networks
can represent high-level abstractions by multiple layers of
non-linear transformations. Existing deep network architectures [60] include convolutional neural networks (CNNs)
[53, 84, 88, 36], deep belief networks (DBNs) [38], and
stacked autoencoders (SAEs) [99], among others. As the architecture becomes deeper and wider, deep networks need
more training data to avoid overfitting, which also results
in the need for transferring knowledge from one domain
to another. Although some studies have shown that deep
networks can learn more transferable representations that
1. INTRODUCTION
Over the past few years, machine learning has achieved
great success and has benefited real-world applications. It
has mainly owed to large-scale datasets under the major assumption that the training and testing data have the same
feature space or distribution. However, a mismatch between
the feature space and distribution always exists in realworld applications due to many factors (e.g., illumination,
pose, and image quality). In addition, collecting and annotating testing datasets for every new task and domain are
extremely expensive and time-consuming processes. Sufficient training data may not always be available, which typically degrades the performance of learning methods at test
1
Figure 1. (a) Some object images from the ”Bike” and ”Laptop” categories in Amazon, DSLR, Webcam, and Caltech-256 databases. (b)
Some digit images from MNIST, USPS, and SVHN databases. (c) Some face images from LFW, BCS and CUFS databases. Realworld
computer vision applications, such as face recognition, must learn to adapt to distributions specific to each domain.
disentangle the exploratory factors of variations underlying
the data samples and group features hierarchically in accordance with their relatedness to invariant factors, Donahue et
al. [16] showed that a domain shift still affects their performance. The deep features would eventually transition from
general to specific, and the transferability of the representation sharply decreases in higher layers. Therefore, recent
work has addressed this problem by deep DA, which combines deep learning and DA.
of data that define how two domains are diverged. 2) extending Csurka’s work, we improve and detail the three subsettings (training with classification loss, discrepancy loss
and adversarial loss) and summarize different approaches
used in different DA scenes. 3) Considering the distance
of the source and target domains, multi-step DA methods
are studied and categorized into hand-crafted, feature-based
and representation-based mechanisms. 4) We provide a survey of many computer vision applications, such as image
classification, face recognition, style translation, object detection, semantic segmentation and person re-identification.
The remainder of this survey is structured as follows. In
Section II, we first define some notations, and then we categorize deep DA into different settings (given in Fig. 2). In
the next three sections, different approaches are discussed
for each setting, which are given in Table 1 and Table 2 in
detail. Then, in Section VI, we introduce some successful computer vision applications of deep DA. Finally, the
conclusion of this paper and discussion of future works are
presented in Section VII.
There have been other surveys on TL and DA over the
past few years [67, 80, 14, 68, 109, 13]. Pan et al. [67] categorized TL under three subsettings, including inductive TL,
transductive TL, and unsupervised TL, but they only studied
homogeneous feature spaces. Shao et al. [80] categorized
TL techniques into feature-representation-level knowledge
transfer and classifier-level knowledge transfer. The survey written by Patel [68] only focused on DA, a subtopic of
TL. [14] discussed 38 methods for heterogeneous TL that
operate under various settings, requirements, and domains.
Zhang et al. [109] were the first to summarize several transferring criteria in detail from the concept level. These five
surveys mentioned above only cover the methodologies on
shallow TL or DA. The work presented by Csurka et al.
[13] briefly analyzed the state-of-the-art shallow DA methods and categorized the deep DA methods into three subsettings based on training loss: classification loss, discrepancy
loss and adversarial loss. However, Csurka’s work mainly
focused on shallow methods, and it only discussed deep DA
in image classification applications.
2. Overview
2.1. Notations and Definitions
In this section, we introduce some notations and definitions that are used in this survey. The notations and definitions match those from the survey papers by [67, 13] to
maintain consistency across surveys. A domain D consists
of a feature space X and a marginal probability distribution
P (X), where X = {x1 , ..., xn } ∈ X . Given a specific
domain D = {X , P (X)}, a task T consists of a feature
space Y and an objective predictive function f (·), which
can also be viewed as a conditional probability distribution
In this paper, we focus on analyzing and discussing deep
DA methods. Specifically, the key contributions of this survey are as follows: 1) we present a taxonomy of different
deep domain adaption scenarios according to the properties
2
P (Y |X) from a probabilistic perspective. In general, we
can learn P (Y |X) in a supervised manner from the labeled
data {xi , yi }, where xi ∈ X and yi ∈ Y.
Assume that we have two domains: the training dataset
with sufficient labeled data is the source domain Ds =
{X s , P (X)s }, and the test dataset with a small amount
of labeled data or no labeled data is the target domain
Dt = {X t , P (X)t }. We see that the partially labeled part,
Dtl , and the unlabeled parts, Dtu , form the entire target domain, that is, Dt = Dtl ∪ Dtu . Each domain is together
with its task: the former is T s = {Y s , P (Y s |X s )}, and
the latter is T t = {Y t , P (Y t |X t )}. Similarly, P (Y s |X s )
can be learned from the source labeled data {xsi , yis }, while
tl
P (Y t |X t ) can be learned from labeled target data {xtl
i , yi }
tu
and unlabeled data {xi }.
able when training the network.
• In the heterogeneous DA setting, the feature spaces between the source and target domains are nonequivalent
(X s 6= X t ), and the dimensions may also generally
differ (ds 6= dt ).
Similar to the homogeneous setting, the heterogeneous DA setting can also be divided into supervised,
semi-supervised and unsupervised DA.
All of the above DA settings assumed that the source
and target domains are directly related; thus, transferring
knowledge can be accomplished in one step. We call them
one-step DA. In reality, however, this assumption is occasionally unavailable. There is little overlap between the two
domains, and performing one-step DA will not be effective.
Fortunately, there are some intermediate domains that are
able to draw the source and target domains closer than their
original distance. Thus, we use a series of intermediate
bridges to connect two seemingly unrelated domains and
then perform one-step DA via this bridge, named multi-step
(or transitive) DA [91, 92]. For example, face images and
vehicle images are dissimilar between each other due to different shapes or other aspects, and thus, one-step DA would
fail. However, some intermediate images, such as ’football
helmet’, can be introduced to be an intermediate domain
and have a smooth knowledge transfer. Fig. 3 shows the
differences between the learning processes of one-step and
multi-step DA techniques.
2.2. Different Settings of Domain Adaption
The case of traditional machine learning is Ds = Dt and
T = T t . For TL, Pan et al. [67] summarized that the
differences between different datasets can be caused by domain divergence Ds 6= Dt (i.e., distribution shift or feature
space difference) or task divergence T s 6= T t (i.e., conditional distribution shift or label space difference), or both.
Based on this summary, Pan et al. categorized TL into three
main groups: inductive, transductive and unsupervised TL.
According to this classification, DA methods are transductive TL solutions with the assumption that the tasks are
the same, i.e., T s = T t , and the differences are only caused
by domain divergence, Ds 6= Dt . Therefore, DA can be
split into two main categories based on different domain
divergences (distribution shift or feature space difference):
homogeneous and heterogeneous DA. Then, we can further
categorize DA into supervised, semi-supervised and unsupervised DA in consideration of labeled data of the target
domain. The classification is given in Fig. 2.
• In the homogeneous DA setting, the feature spaces
between the source and target domains are identical
(X s = X t ) with the same dimension (ds = dt ).
Hence, the source and target datasets are generally
different in terms of data distributions (P (X)s 6=
P (X)t ).
In addition, we can further categorize the homogeneous DA setting into three cases:
s
3. Approaches of Deep Domain Adaption
In a broad sense, deep DA is a method that utilizes a
deep network to enhance the performance of DA. Under
this definition, shallow methods with deep features can be
considered as a deep DA approach. DA is adopted by shallow methods, whereas deep networks only extract vectorial
features and are not helpful for transferring knowledge directly [16, 41, 72, 64, 110]. This approach reliably outperforms current state-of-the-art approaches based on traditional hand-crafted features because sufficient representational and transferable features can be extracted through
deep networks, which can work better on discrimination
tasks [16].
In a narrow sense, deep DA is based on deep learning
architectures designed for DA and can obtain a firsthand
effect from deep networks via back-propagation. The intuitive idea is to embed DA into the process of learning representation and to learn a deep feature representation that is
both semantically meaningful and domain invariant. With
the ”good” feature representations, the performance of the
target task would improve significantly. In this paper, we
focus on the narrow definition and discuss how to utilize
deep networks to learn ”good” feature representations with
1. In the supervised DA, a small amount of labeled
target data, Dtl , are present. However, the labeled data are commonly not sufficient for tasks.
2. In the semi-supervised DA, both limited labeled
data, Dtl , and redundant unlabeled data, Dtu ,
in the target domain are available in the training stage, which allows the networks to learn the
structure information of the target domain.
3. In the unsupervised DA, no labeled but sufficient
unlabeled target domain data, Dtu , are observ3
Figure 2. An overview of different settings of domain adaption
Figure 3. Different learning processes between (a) traditional machine learning, (b) one-step domain adaption and (c) multi-step domain
adaption [67].
Kullback-Leibler (KL) divergence [114] and H divergence, among others.
• Architecture Criterion: aims at improving the ability
of learning more transferable features by adjusting the
architectures of deep networks. The techniques that
are proven to be cost effective include adaptive batch
normalization (BN) [58, 46, 57], weak-related weight
[78], domain-guided dropout [102], and so forth.
• Geometric Criterion: bridges the source and target domains according to their geometrical properties. This
criterion assumes that the relationship of geometric
structures can reduce the domain shift [10].
The second case can be referred to as an adversarialbased deep DA approach [19]. In this case, a domain
discriminator that classifies whether a data point is drawn
from the source or target domain is used to encourage domain confusion through an adversarial objective to minimize the distance between the empirical source and target
mapping distributions. Furthermore, the adversarial-based
deep DA approach can be categorized into two cases based
on whether there are generative models.
• Generative Models: combine the discriminative model
with a generative component in general based on generative adversarial networks (GANs). One of the typical cases is to use source images, noise vectors or both
to generate simulated samples that are similar to the
target samples and preserve the annotation information
extra training criteria.
3.1. Categorization of one-step domain adaption
In one-step DA, the deep approaches can be summarized into three cases, which refers to [13]. Table 1 shows
these three cases and brief descriptions. The first case is
the discrepancy-based deep DA approach, which assumes
that fine-tuning the deep network model with labeled or unlabeled target data can diminish the shift between the two
domains. Class criterion, statistic criterion, architecture criterion and geometric criterion are four major techniques for
performing fine-tuning.
• Class Criterion: uses the class label information as
a guide for transferring knowledge between different
domains. When the labeled samples from the target
domain are available in supervised DA, soft label and
metric learning are always effective [95, 40, 34, 70, 45,
37]. When such samples are unavailable, some other
techniques can be adopted to substitute for class labeled data, such as pseudo labels [63, 111, 104] and
attribute representation [22, 95].
• Statistic Criterion: aligns the statistical distribution
shift between the source and target domains using
some mechanisms. The most commonly used methods for comparing and reducing distribution shift are
maximum mean discrepancy (MMD) [62, 104, 61,
63, 97, 25], correlation alignment (CORAL) [87, 71],
4
Table 1. Different Deep Approaches to One-Step DA
One-step DA
Approaches
Brief Description
Subsettings
Discrepancybased
fine-tuning the deep network with
labeled or unlabeled target data to
diminish the domain shift
Adversarialbased
using domain discriminators to
encourage domain confusion through
an adversarial objective
Reconstructionbased
using the data reconstruction as an
auxiliary task to ensure feature
invariance
class criterion [95, 40, 34, 70]
[45, 37, 63, 111, 104, 22, 95, 21]
statistic criterion [62, 104, 61]
[63, 97, 25, 87, 71, 114]
architecture criterion [58, 46]
[57, 78, 102, 73]
geometric criterion [10]
generative models
[59, 3, 48]
non-generative models [96]
[95, 19, 18, 94, 69]
encoder-decoder reconstruction
[4, 26, 24, 114]
adversarial reconstruction
[105, 113, 50]
Table 2. Different Deep Approaches to Multi-Step DA
Multi-step Approaches
Hand-crafted
Instance-based
Representation-based
Brief Description
users determine the intermediate domains based on experience [103]
selecting certain parts of data from the auxiliary datasets to
compose the intermediate domains [92, 10]
freeze weights of one network and use their intermediate
representations as input to the new network [79]
3.2. Categorization of multi-step domain adaption
of the source domain [59, 3, 48].
• Non-Generative Models: rather than generating models with input image distributions, the feature extractor learns a discriminative representation using the labels in the source domain and maps the target data
to the same space through a domain-confusion loss,
thus resulting in the domain-invariant representations
[96, 95, 19, 18, 94].
The third case can be referred to as a reconstructionbased DA approach, which assumes that the data reconstruction of the source or target samples can be helpful for
improving the performance of DA. The reconstructor can
ensure both specificity of intra-domain representations and
indistinguishability of inter-domain representations.
• Encoder-Decoder Reconstruction: by using stacked
autoencoders (SAEs), encoder-decoder reconstruction
methods combine the encoder network for representation learning with a decoder network for data reconstruction [4, 26, 24, 114].
• Adversarial Reconstruction: the reconstruction error is
measured as the difference between the reconstructed
and original images within each image domain by
a cyclic mapping obtained via a GAN discriminator,
such as dual GAN [105], cycle GAN [113] and disco
GAN [50].
In multi-step DA, we first determine the intermediate domains that are more related with the source and target domains than their direct connection. Second, the knowledge
transfer process will be performed between the source, intermediate and target domains by one-step DA with less
information loss. Thus, the key of multi-step DA is how
to select and utilize intermediate domains; additionally, it
can fall into three categories referring to [67]: hand-crafted,
feature-based and representation-based selection mechanisms.
• Hand-Crafted: users determine the intermediate domains based on experience [103].
• Instance-Based: selecting certain parts of data from
the auxiliary datasets to compose the intermediate domains to train the deep network [92, 10].
• Representation-Based: transfer is enabled via freezing
the previously trained network and using their intermediate representations as input to the new one [79].
4. HOMOGENEOUS DOMAIN ADAPTION
As mentioned in Section 2.1, the data in the target domain have three types regardless of homogeneous or heterogeneous DA: 1) supervised DA with labeled data, 2)
semi-supervised DA with labeled and unlabeled data and
5
Table 3. Different Approaches used in Different Domain Adaption Settings
Discrepancy-based
Adversarial-based
Reconstruction-based
Class Criterion
Statistic Criterion
Architecture Criterion
Geometric Criterion
Generative Model
Non-Generative Model
Encoder-Decoder Model
Adversarial Model
3) non-supervised DA with unlabeled data. The second setting is able to be accomplished by combining the methods
of setting 1 and setting 3; thus, we only focus on the first
and third settings in this paper. The cases where the different approaches are mainly used for each DA setting are
shown in Table 3. As shown, more work is focused on
unsupervised scenes because supervised DA has its limitations. When only few labeled data in the target domain are
available, using the source and target labeled data to train
parameters of models typically results in overfitting to the
source distribution. In addition, the discrepancy-based approaches have been studied for years and produced more
methods in many research works, whereas the adversarialbased and reconstruction-based approaches are a relatively
new research topic but have recently been attracting more
attention.
Supervised DA
√
√
√
Unsupervised DA
√
√
√
√
√
√
Figure 4. The average accuracy over the validation set for a network trained with different strategies. Baseline B: the network is
trained on dataset B. 2) BnB: the first n layers are reused from
baseline B and frozen. The higher layers are trained on dataset
B. 3) BnB+: the same as BnB but where all layers are fine-tuned.
4) AnB: the first n layers are reused from the network trained on
dataset A and frozen. The higher layers are trained on dataset B. 5)
AnB+: the same as AnB but where all layers are fine-tuned [107].
4.1. One-step Domain Adaption
4.1.1 Discrepancy-Based Approaches
P
training loss, L = − N
i=0 yi log ŷi (ŷi are the softmax predictions of the model, which represent class probabilities)
[95, 40, 34, 70, 45, 100]. To extend this, Hinton et al. [37]
modified the softmax function to soft label loss:
Yosinski et al.[107] proved that transferable features learned
by deep networks have limitations due to fragile coadaptation and representation specificity and that finetuning can enhance generalization performance. Finetuning (can also be viewed as a discrepancy-based deep
DA approach) is to train a base network with source data
and then directly reuse the first n layers to conduct a target network. The remaining layers of the target network
are randomly initialized and trained with loss based on discrepancy. During training, the first n layers of the target
network can be fine-tuned or frozen depending on the size
of the target dataset and its similarity to the source dataset
[11]. Some common rules of thumb for navigating the 4
major scenarios are given in Table 4.
• Class Criterion
The class criterion is the most basic training loss in deep
DA. After pre-training the network with source data, the remaining layers of the target model use the class label information as a guide to train the network. Hence, a small number of labeled samples from the target dataset is assumed to
be available.
Ideally, the class label information is given directly in supervised DA. Most work commonly uses the negative loglikelihood of the ground truth class with softmax as their
qi = P
exp(zi /T )
j (exp(zj /T ))
(1)
where z i is the logit output computed for each class. T
is a temperature that is normally set to 1 in standard softmax, but it takes a higher value to produce a softer probability distribution over classes. By using it, much of the
information about the learned function that resides in the
ratios of very small probabilities can be obtained. For example, when recognizing digits, one version of 2 may obtain a probability of 106 of being a 3 and 109 of being a
7; in other words, this version of 2 looks more similar to
3 than 7. Inspired by Hinton, [95] fine-tuned the network
by simultaneously minimizing the domain confusion loss
(belonging to adversarial-based approaches, which will be
presented in Section 4.1.2) and soft label loss. Using soft
labels rather than hard labels can preserve the relationships
between classes across domains. Gebru et al. [22] modified
existing adaptation algorithms based on [95] and utilized
6
Table 4. Some Common Rules of Thumb for Deciding Fine-tuned or Frozen in the First n Layers. [11]
The Distance
between
Source and Target
Low
Medium
High
The Size of Target Dataset
Low
Medium
Freeze
Try Freeze or Tune
Try Freeze or Tune
Tune
Try Freeze or Tune
Tune
soft label loss at the fine-grained class level Lcsof t and attribute level Lasof t .
High
Tune
Tune
Tune
humans can identify unseen classes given only a high-level
description. For instance, when provided the description
”tall brown animals with long necks”, we are able to recognize giraffes. To imitate the ability of humans, [54] introduced high-level semantic attributes per class. Assume
that ac = (ac1 , ..., acm ) is the attribute representation for
class c, which has fixed-length binary values with m attributes in all the classes. The classifiers provide estimates
of p(am |x) for each attribute am . In the test stage, each
target class y obtains its attribute vector ay in a deterministic way, i.e., p(a|y) = [[a = ay ]]. By applying Bayes rule,
p(y)
y
p(y|a) = p(a
y ) [[a = a ]], the posterior of a test class can be
calculated as follows:
p(y|x) =
X
p(y|a)p(a|x) =
a∈{0,1}M
(3)
Gebru et al. [22] drew inspiration from these works and
leveraged attributes to improve performance in the DA of
fine-grained recognition. There are multiple independent
softmax losses that simultaneously perform attribute and
class level to fine-tune the target model. To prevent the independent classifiers from obtaining conflicting labels with
attribute and class level, an attribute consistency loss is also
implemented.
Occasionally, when fine-tuning the network in unsupervised DA, a label of target data, which is called a pseudo
label, can preliminarily be obtained based on the maximum
posterior probability. Yan et al. [104] initialized the target model using the source data and then defined the class
posterior probability p(yjt = c|xtj ) by the output of the target model. With p(yjt = c|xtj ), they assigned pseudo-label
ybt to xt by ybt = arg max p(y t = c|xt ). The deep trans-
Figure 5. Deep DA by combining domain confusion loss and soft
label loss [95].
In addition to softmax loss, there are other methods that
can be used as training loss to fine-tune the target model in
supervised DA. For example, given a pair of samples from
the source and target domains (such as an RGB image and
its paired depth image), [34] minimized the difference between the representations using a Euclidean loss. Embedding metric learning in deep networks is another method
that can make the distance of samples from different domains with the same labels be closer while those with different labels are far away. Deep transfer metric learning
was proposed by [45], which applies the marginal Fisher
analysis criterion and MMD criterion (described in Statistic
Criterion) to minimize their distribution difference:
(M)
(M)
min J = Sc(M) − αSb + βDts X s , X t
+γ
M
X
m=1
( W (m)
2
F
+ b(m)
2
2
)
M
p(y) Y
p(aym |x)
p(ay ) m=1
j
j
j
c
j
j
fer network (DTN) [111] used some base classifiers, e.g.,
SVMs and MLPs, to obtain the pseudo labels for the target
samples to estimate the conditional distribution of the target
samples and match both the marginal and the conditional
distributions with the MMD criterion. When casting the
classifier adaptation into the residual learning framework,
[63] used the pseudo label to build the conditional entropy
E(Dt , f t ), which ensures that the target classifier f t fits the
target-specific structures well.
• Statistic Criterion
Although some discrepancy-based approaches search for
pseudo labels, attribute labels or other substitutes to labeled target data, more work focuses on learning domain-
(2)
where α, β and γ are regularization parameters and W (m)
and b(m) are the weights and biases of the mth layer of the
(M)
network. Dts (X s , X t ) is the MMD between representations of the source and target domains. Sc and Sb define the
intra-class compactness and the interclass separability.
However, what can we do if there is no class label information in the target domain directly? As we all know,
7
invariant representations via minimizing the domain distribution discrepancy in unsupervised DA.
MMD is an effective metric for comparing the distributions between two datasets by a kernel two-sample test [2].
Given two distributions s and t, the MMD is defined as follows:
M M D2 (s, t) =
sup
Exs ∼s [φ(xs )] − Ext ∼s [φ(xt )]
kφkH ≤1
distributions are matched based on MMD. The shared feature extraction layer learns a subspace to match the marginal
distributions of the source and the target samples, and the
discrimination layer matches the conditional distributions
by classifier transduction. In addition to adapting features
using MMD, residual transfer networks (RTNs) [63] added
a gated residual layer for classifier adaptation. More recently, [104] proposed a weighted MMD model that introduces an auxiliary weight for each class in the source domain when the class weights in the target domain are not
the same as those in the source domain.
If φ is a characteristic kernel (i.e., Gaussian kernel or
Laplace kernel), MMD will compare all the orders of statistic moments. In contrast to MMD, CORAL [86] learned a
linear transformation that aligns the second-order statistics
between domains. Sun et al. [87] extended CORAL to deep
neural networks (deep CORAL) with a nonlinear transformation.
1
2
(7)
LCORAL = 2 kCS − CT kF
4d
where k · k2F denotes the squared matrix Frobenius norm.
CS and CT denote the covariance matrices of the source
and target data, respectively.
By the Taylor expansion of the Gaussian kernel, MMD
can be viewed as minimizing the distance between the
weighted sums of all raw moments [56]. The interpretation of MMD as moment matching procedures motivated
Zellinger et al. [108] to match the higher-order moments of
the domain distributions, which we call central moment discrepancy (CMD). An empirical estimate of the CMD metric
for the domain discrepancy in the activation space [a, b]N is
given by
2
H
(4)
where φ represents the kernel function that maps the original data to a reproducing kernel Hilbert space (RKHS) and
kφkH ≤ 1 defines a set of functions in the unit ball of
RKHS H.
Based on the above, Ghifary et al. [25] proposed a model
that introduced the MMD metric in feedforward neural networks with a single hidden layer. The MMD metric is computed between representations of each domain to reduce the
distribution mismatch in the latent space. The empirical estimate of MMD is as follows:
M
N
1 X
1 X
M M D (Ds , Dt ) =
φ(xsi )−
φ(xtj )
M i=1
N j=1
2
2
H
(5)
Subsequently, Tzeng et al. [97] and Long et al. [61]
extended MMD to a deep CNN model and achieved great
success. The deep domain confusion network (DDC) by
Tzeng et al. [97] used two CNNs for the source and target
domains with shared weights. The network is optimized for
classification loss in the source domain, while domain difference is measured by an adaptation layer with the MMD
metric.
L=LC (X L , y) + λM M D2 (X s X t )
(6)
CM DK (X s , X t ) =
where the hyperparameter λ is a penalty parameter.
LC (X L , y) denotes classification loss on the available labeled data, X L , and the ground-truth labels, y.
M M D2 (X s X t ) denotes the distance between the source
and target data. DDC only adapts one layer of the network,
resulting in a reduction in the transferability of multiple layers. Rather than using a single layer and linear MMD, Long
et al. [61] proposed the deep adaptation network (DAN) that
matches the shift in marginal distributions across domains
by adding multiple adaptation layers and exploring multiple kernels, assuming that the conditional distributions remain unchanged. However, this assumption is rather strong
in practical applications; in other words, the source classifier cannot be directly used in the target domain. To make
it more generalized, a joint adaptation network (JAN) [62]
aligns the shift in the joint distributions of input features and
output labels in multiple domain-specific layers based on a
joint maximum mean discrepancy (JMMD) criterion. [111]
proposed DTN, where both the marginal and the conditional
+
K
X
k=2
1
E(X s ) − E(X t )
(b − a)
(8)
1
|b − a|
2
s
k
t
Ck (X ) − Ck (X )
k
2
where Ck (X) = E((x − E(X)) is the vectorP
of all k th 1
order sample central moments and E(X) = |X| x∈X x is
the empirical expectation.
• Architecture Criterion
Some other methods optimize the architecture of the network to minimize the distribution discrepancy. This adaptation behavior can be achieved in most deep DA models,
such as supervised and unsupervised settings.
Rozantsev et al. [78] considered that the weights in corresponding layers are not shared but related by a weight regularizer rw (·) to account for the differences between the two
domains. The weight regularizer rw (·) can be expressed as
the exponential loss function:
2
rw (θjs , θjt ) = exp θjs − θjt
−1
(9)
8
Figure 6. Different approaches with the MMD metric. (a) The deep adaptation network (DAN) architecture [61], (b) the joint adaptation
network (JAN) architecture [62] and (c) the residual transfer network (RTN) architecture [63].
where θjs and θjt denote the parameters of the j th layer of the
source and target models, respectively. To further relax this
restriction, they allow the weights in one stream to undergo
a linear transformation:
rw (θjs , θjt ) = exp( aj θjs + bj − θjt
2
)−1
layers with instance normalization (IN) layers, where µ(x)
and σ(x) are computed independently for each channel and
each sample, the performance of DA can be further improved.
Occasionally, neurons are not effective for all domains
because of the presence of domain biases. For example,
when recognizing people, the target domain typically contains one person centered with minimal background clutter,
whereas the source dataset contains many people with more
clutter. Thus, the neurons that capture the features of other
people and clutter are useless. Domain-guided dropout was
proposed by [102] to solve the problem of multi-domain
adaptation, and it mutes non-related neurons for each domain. Rather than assigning dropout with a specific dropout
rate, it depends on the gain of the loss function of each neuron on the domain sample when the neuron is removed.
(10)
where aj and bj are scalar parameters that encode the linear
transformation. The work of Shu et al. [83] is similar to [78]
using weakly parameter-shared layers. The penalty term Ω
controls the relatedness of parameters.
Ω=
L
X
(l)
(l)
( WS − WT
i=1
(l)
(l)
2
F
(l)
(l)
+ bS − bT
(l)
2
F
)
(11)
(l)
L
where {WS , bS }L
l=1 and {WT , bT }l=1 are the parameth
ters of the l layer in the source and target domains, respectively.
Li et al. [58] hypothesized that the class-related knowledge is stored in the weight matrix, whereas domain-related
knowledge is represented by the statistics of the batch normalization (BN) layer [47]. BN normalizes the mean and
standard deviation for each individual feature channel such
that each layer receives data from a similar distribution, irrespective of whether it comes from the source or the target
domain. Therefore, Li et al. used BN to align the distribution for recomputing the mean and standard deviation in the
target domain.
x − µ(X t )
t
+β
(12)
BN (X ) = λ
σ(X t )
si = L(g(x)\i ) − L(g(x))
(13)
where L is the softmax loss function and g(x)\i is the feature vector after setting the response of the ith neuron to
zero.
• Geometric Criterion
The geometric criterion mitigates the domain shift by integrating intermediate subspaces on a geodesic path from
the source to the target domains. A geodesic flow curve is
constructed to connect the source and target domains on the
Grassmannian. The source and target subspaces are points
on a Grassmann manifold. By sampling a fixed [33] or infinite number of subspaces along the geodesic [31], we can
form the intermediate subspaces to help to find the correlations between domains. Then, both source and target data
are projected to the obtained intermediate subspaces to align
the distribution.
Inspired by the intermediate representations on the
geodesic path, Chopra et al. [10] proposed a model called
where λ and β are parameters learned from the target data
and µ(x) and σ(x) are the mean and standard deviation
computed independently for each feature channel. Furthermore, Ulyanov et al. [98] found that when replacing BN
9
Figure 7. The two-stream architecture with related weight [78].
deep learning for domain adaption by interpolating between
domains (DLID). DLID generates intermediate datasets,
starting with all the source data samples and gradually replacing source data with target data. Each dataset is a single
point on an interpolating path between the source and target domains. Once intermediate datasets are generated, a
deep nonlinear feature extractor using the predictive sparse
decomposition is trained in an unsupervised manner.
4.1.2 Adversarial-Based Approaches
Figure 8. Generalized architecture for adversarial domain adaptation. Existing adversarial adaptation methods can be viewed as instantiations of a framework with different choices regarding their
properties. [96]
Recently, great success has been achieved by the GAN
method [32], which estimates generative models via an adversarial process. GAN consists of two models: a generative model G that extracts the data distribution and a discriminative model D that distinguishes whether a sample is
from G or training datasets by predicting a binary label. The
networks are trained on the label prediction loss in a minimax fashion: simultaneously optimizing G to minimize the
loss while also training D to maximize the probability of
assigning the correct label:
min max V (D, G) = Ex∼pdata (x) [log D(x)]
G
D
with labels are used to train the target model as if no DA
were required. Adversarial-based approaches with generative models are able to learn such a transformation in an
unsupervised manner based on GAN.
The core idea of CoGAN [59] is to generate synthetic
target data that are paired with synthetic source ones. It
consists of a pair of GANs: GAN1 for generating source
data and GAN2 for generating target data. The weights of
the first few layers in the generative models and the last few
layers in the discriminative models are tied. This weightsharing constraint allows CoGAN to achieve a domaininvariant feature space without correspondence supervision.
A trained CoGAN can adapt the input noise vector to paired
images that are from the two distributions and share the labels. Therefore, the shared labels of synthetic target samples can be used to train the target model.
(14)
+Ez∼pz (z) [log(1 − D(G(z)))]
In DA, this principle has been employed to ensure that
the network cannot distinguish between the source and
target domains. [96] proposed a unified framework for
adversarial-based approaches and summarized the existing
approaches according to whether to use a generator, which
loss function to employ, or whether to share weights across
domains. In this paper, we only categorize the adversarialbased approaches into two subsettings: generative models
and non-generative models.
• Generative Models
Synthetic target data with ground-truth annotations are
an appealing alternative to address the problem of a lack of
training data. First, with the help of source data, generators
render unlimited quantities of synthetic target data, which
are paired with synthetic source data to share labels or appear as if they were sampled from the target domain while
maintaining labels, or something else. Then, synthetic data
Figure 9. The CoGAN architecture. [59]
More work focuses on generating synthetic data that
10
are similar to the target data while maintaining annotations. Yoo et al. [106] transferred knowledge from the
source domain to pixel-level target images with GANs. A
domain discriminator ensures the invariance of content to
the source domain, and a real/fake discriminator supervises
the generator to produce similar images to the target domain. Shrivastava et al. [82] developed a method for
simulated+unsupervised (S+U) learning that uses a combined objective of minimizing an adversarial loss and a selfregularization loss, where the goal is to improve the realism
of synthetic images using unlabeled real data. In contrast to
other works in which the generator is conditioned only on a
noise vector or source images, Bousmalis et al. [3] proposed
a model that exploits GANs conditioned on both. The classifier T is trained to predict class labels of both source and
synthetic images, while the discriminator is trained to predict the domain labels of target and synthetic images. In addition, to expect synthetic images with similar foregrounds
and different backgrounds from the same source images, a
content similarity is used that penalizes large differences
between source and synthetic images for foreground pixels
only by a masked pairwise mean squared error [17]. The
goal of the network is to learn G, D and T by solving the
optimization problem:
ters of the target model are initialized by the pre-trained
source one. This is more flexible because of allowing more
domain-specific feature extractions to be learned. ADDA
minimizes the source and target representation distances
through iteratively minimizing these following functions,
which is most similar to the original GAN:
min Lcls (X s , Y s ) =
M s ,C
−E
(xs ,y s )∼(X s ,Y s )
D
1[k=ys ] log C(M s (xs ))
k=1
min LadvD (X s ,X t , M s , M t ) =
D
− E(xs )∼(X s ) [log D(M s (xs ))]
− E(xt )∼(X t ) [log(1 − D(M t (xt )))]
min LadvM (M s , M t ) =
M s ,M t
− E(xt )∼(X t ) [log D(M t (xt ))]
(16)
where the mappings M s and M t are learned from the
source and target data, X s and X t . C represents a classifier working on the source domain. The first classification
loss function Lcls is optimized by training the source model
using the labeled source data. The second function LadvD is
minimized to train the discriminator, while the third function LadvM is learning a representation that is domain invariant.
Tzeng et al. [95] proposed adding an additional domain
classification layer that performs binary domain classification and designed a domain confusion loss to encourage its
prediction to be as close as possible to a uniform distribution over binary labels. Unlike previous methods that match
the entire source and target domains, Cao et al. introduced
a selective adversarial network (SAN) [6] to address partial transfer learning from large domains to small domains,
which assumes that the target label space is a subspace of
the source label space. It simultaneously avoids negative
transfer by filtering out outlier source classes, and it promotes positive transfer by matching the data distributions in
the shared label space via splitting the domain discriminator
into many class-wise domain discriminators.
min max V (D, G) = αLd (D, G)
G,T
K
X
(15)
+βLt (T, G) + γLc (G)
where α, β, and γ are parameters that control the trade-off
between the losses. Ld , Lt and Lc are the adversarial loss,
softmax loss and content-similarity loss, respectively.
• Non-Generative Models
The key of deep DA is learning domain-invariant representations from source and target samples. With these representations, the distribution of both domains can be similar
enough such that the classifier is fooled and can be directly
used in the target domain even if it is trained on source samples. Therefore, whether the representations are domainconfused or not is crucial to transferring knowledge. Inspired by GAN, domain confusion loss, which is produced
by the discriminator, is introduced to improve the performance of deep DA without generators.
The domain-adversarial neural network (DANN) [18] integrates a gradient reversal layer (GRL) into the standard
architecture to ensure that the feature distributions over the
two domains are made similar. The network consists of
shared feature extraction layers and two classifiers. DANN
minimizes the domain confusion loss (for all samples) and
label prediction loss (for source samples) while maximizing
domain confusion loss via the use of the GRL. In contrast
to the above methods, the adversarial discriminative domain
adaptation (ADDA) [96] considers independent source and
target mappings by untying the weights, and the parame-
4.1.3 Reconstruction-Based Approaches
In DA, the data reconstruction of source or target samples is
an auxiliary task that simultaneously focuses on creating a
shared representation between the two domains and keeping
the individual characteristics of each domain.
• Encoder-Decoder Reconstruction
The basic autoencoder framework [1] is a feedforward
neural network that includes the encoding and decoding
processes. The autoencoder first encodes an input to some
11
Figure 10. The model that exploits GANs conditioned on noise vector and source images. [3]
proposed in [26] learns a shared encoding representation
that provides useful information for cross-domain object
recognition. DRCN is a CNN architecture that combines
two pipelines with a shared encoder. After a representation is provided by the encoder, the first pipeline, which
is a CNN, works for supervised classification with source
labels, whereas the second pipeline, which is a deconvolutional network, optimizes for unsupervised reconstruction
with target data.
min λLc ({θenc , θlab }) + (1 − λ)Lr ({θenc , θdec }) (17)
Figure 11. The domain-adversarial neural network (DANN) architecture. [18]
where λ is a hyper-parameter that controls the trade-off between classification and reconstruction. θenc , θdec and θlab
denote the parameters of the encoder, decoder and source
classifier, respectively. Lc is cross-entropy loss for classi2
fication, and Lr is squared loss k x − fr (x) k2 for reconstruction in which fr (x) is the reconstruction of x.
Domain separation networks (DSNs) [4] explicitly and
jointly model both private and shared components of the
domain representations. A shared-weight encoder learns
to capture shared representations, while a private encoder
is used for domain-specific components in each domain.
Additionally, a shared decoder learns to reconstruct the input samples by both the private and shared representations.
Then, a classifier is trained on the shared representation. By
partitioning the space in such a manner, the shared representations will not be influenced by domain-specific representations such that a better transfer ability can be obtained.
Zhuang et al. [114] proposed transfer learning with
deep autoencoders (TLDA), which consists of two encoding layers. The distance in distributions between domains is
minimized with KL divergence in the embedding encoding
layer, and label information of the source domain is encoded
using a softmax loss in the label encoding layer. Ghifary et
al. [24] extended the autoencoder into a model that jointly
hidden representation, and then it decodes this hidden representation back to a reconstructed version. The DA approaches based on encoder-decoder reconstruction typically
learn the domain-invariant representation by a shared encoder and maintain the domain-special representation by a
reconstruction loss in the source and target domains.
Xavier and Bengio [29] proposed extracting a high-level
representation based on stacked denoising autoencoders
(SDA) [99]. By reconstructing the union of data from various domains with the same network, the high-level representations can represent both the source and target domain
data. Thus, a linear classifier that is trained on the labeled
data of the source domain can make predictions on the target domain data with these representations. Despite their
remarkable results, SDAs are limited by their high computational cost and lack of scalability to high-dimensional features. To address these crucial limitations, Chen et al. [7]
proposed the marginalized SDA (mSDA), which marginalizes noise with linear denoisers; thus, parameters can be
computed in closed-form and do not require stochastic gradient descent.
The deep reconstruction classification network (DRCN)
12
Figure 12. The Adversarial discriminative domain adaptation (ADDA) architecture. [96]
Figure 13. The deep reconstruction classification network (DRCN) architecture. [26]
or G(Y ) ≈ X) by an adversarial loss and how well the
original input is reconstructed after a sequence of two generations (F (G(X)) ≈ X or G(F (Y )) ≈ Y ) by a cycle
consistency loss (reconstruction loss). Thus, the distribution of images from G(X) (or F (Y )) is indistinguishable
from the distribution Y (or X).
learns two types of data-reconstruction tasks taken from related domains: one is self-domain reconstruction, and the
other is between-domain reconstruction.
• Adversarial Reconstruction
Dual learning was first proposed by Xia et al. [35] to
reduce the requirement of labeled data in natural language
processing. Dual learning trains two ”opposite” language
translators, e.g., A-to-B and B-to-A. The two translators
represent a primal-dual pair that evaluates how likely the
translated sentences belong to the targeted language, and
the closed loop measures the disparity between the reconstructed and the original ones. Inspired by dual learning,
adversarial reconstruction is adopted in deep DA with the
help of dual GANs.
Zhu et al. [113] proposed a cycle GAN that can translate
the characteristics of one image domain into the other in the
absence of any paired training examples. Compared to dual
learning, cycle GAN uses two generators rather than translators, which learn a mapping G : X → Y and an inverse
mapping F : Y → X. Two discriminators, DX and DY ,
measure how realistic the generated image is (G(X) ≈ Y
LGAN (G, DY , X, Y ) = Ey∼pdata (y) [log DY (y)]
+Ex∼pdata (x) [log(1 − DY (G(x)))]
Lcyc (G, F ) = Ex∼data(x) [kF (G(x)) − xk1 ]
+Ey∼data(y) [kG(F (y)) − yk1 ]
(18)
where LGAN is the adversarial loss produced by discriminator DY with mapping function G : X → Y . Lcyc is the
reconstruction loss using L1 norm.
The dual GAN [105] and the disco GAN [50] were proposed at the same time, where the core idea is similar to
cycle GAN. In dual GAN, the generator is configured with
skip connections between mirrored downsampling and upsampling layers [76, 48], making it a U-shaped net to share
low-level information (e.g., object shapes, textures, clutter,
13
Figure 14. The cycle GAN architecture. [113]
and so forth). For discriminators, the Markovian patchGAN [55] architecture is employed to capture local highfrequency information. In disco GAN, various forms of distance functions, such as mean-square error (MSE), cosine
distance, and hinge loss, can be used as the reconstruction
loss, and the network is applied to translate images, changing specified attributes including hair color, gender and orientation while maintaining all other components.
clearly be crawled as intermediate domain data.
With the common sense that nighttime light intensities
can be used as a proxy for economic activity, Xie et al.
[103] transferred knowledge from daytime satellite imagery
to poverty prediction with the help of some nighttime light
intensity information as an intermediate domain.
4.2.2 Instance-Based Approaches
In other problems where there are many candidate intermediate domains, some automatic selection criterion should be
considered. Similar to the instance-transfer approaches proposed by Pan [67], because the samples of the source domain cannot be used directly, the mixture of certain parts of
the source and target data can be useful for constructing the
intermediate domain.
Tan et al. [92] proposed distant domain transfer learning (DDTL), where long-distance domains fail to transfer
knowledge by only one intermediate domain but can be related via multiple intermediate domains. DDTL gradually
selects unlabeled data from the intermediate domains by
minimizing reconstruction errors on the selected instances
in the source and intermediate domains and all the instances in the target domain simultaneously. With removal
of the unrelated source data, the selected intermediate domains gradually become closer to the target domain from
the source domain:
4.1.4 Hybrid Approaches
To obtain better performance, some of the aforementioned
methods have been used simultaneously. [95] combined
a domain confusion loss and a soft label loss, while
[63] used both statistic (MMD) and architecture criteria
(adapt classifier by residual function) for unsupervised DA.
[104] introduced class-specific auxiliary weights assigned
by the pseudo-labels into the original MMD. In DSNs [4],
encoder-decoder reconstruction approaches separate representations into private and shared representations, while the
MMD criterion or domain confusion loss is helpful to make
the shared representations similar and soft subspace orthogonality constraints ensure dissimilarity between the private
and shared representations. [78] used the MMD between
the learned source and target representations and also allowed the weights of the corresponding layers to differ.
[114] learned domain-invariant representations by encoderdecoder reconstruction approaches and the KL divergence.
J1 (fe , fd , vS , vT ) =
4.2. Multi-Step Domain Adaption
For multi-step DA, the selection of the intermediate domain is problem specific, and different problems may have
different strategies.
4.2.1 Hand-Crafted Approaches
nS
1 X
v i x̂i − xiS
nS i=1 S S
+
nI
1 X
v i x̂i − xiI
nI i=1 I I
+
nT
1 X
x̂i − xiT
nT i=1 T
2
2
2
2
2
2
+ R(vS , vT )
(19)
Occasionally, the intermediate domain can be selected by
experience, that is, it is decided in advance. For example,
when the source domain is image data and the target domain is composed of text data, some annotated images will
where x̂iS , x̂iT and x̂iI are reconstructions of source data S i ,
target data T i and intermediate data I i based on the autoencoder, respectively, and fe and fd are the parameters of the
14
⊤
encoder and decoder, respectively. vS = (vS1 , ..., vSnS ) and
⊤
vI = (vI1 , ..., vInI ) , vSi , vIi ∈ 0, 1 are selection indicators
for the ith source and intermediate instance, respectively.
R(vS , vT ) is a regularization term that avoids all values of
vS and vI being zero.
The DLID model [10] mentioned in Section 4.1.1 (Geometric Criterion) constructs the intermediate domains with
a subset of the source and target domains, where source
samples are gradually replaced by target samples.
from the text domain to the image domain. DTNs take
paired data, such as text and image, as input to two SAEs,
followed by weakly parameter-shared network layers at the
top, which were mentioned in the architecture criterion of
Section 4.1.1. Chen et al. [8] proposed transfer neural
trees (TNTs), which consist of two stream networks to learn
a domain-invariant feature representation for each modality. Then, a transfer neural decision forest (Transfer-NDF)
[77, 52] is used with stochastic pruning for adapting representative neurons in the prediction layer.
4.2.3 Representation-Based Approaches
6. Application of Deep Domain Adaption
Representation-based approaches freeze the previously
trained network and use their intermediate representations
as input to the new network. Rusu et al. [79] introduced
progressive networks that have the ability to accumulate and
transfer knowledge to new domains over a sequence of experiences. To avoid the target model losing its ability to
solve the source domain, they constructed a new neural network for each domain, while transfer is enabled via lateral
connections to features of previously learned networks. In
the process, the parameters in the latest network are frozen
to remember knowledge of intermediate domains.
Deep DA techniques have recently been successfully applied in many real-world applications, including image classification, object recognition, face recognition, object detection, style translation, and so forth. In this section, we
present different application examples using various visual
deep DA methods. Because the information of commonly
used datasets for evaluating the performance is provided in
[109] in detail, we do not introduce it in this paper.
6.1. Image Classification
Because image classification is a basic task of computer vision applications, most of the algorithms mentioned
above were originally proposed to solve such problems.
Therefore, we do not discuss this application repeatedly,
but we show how much benefit deep DA methods for image classification can bring. Because different papers often
use different parameters, experimental protocols and tuning
strategies in the preprocessing steps, it is quite difficult to
perform a fair comparison among all the methods directly.
Thus, similar to the work of Pan [67], we show the comparison results between the proposed deep DA methods and
non-adaption methods using only deep networks. A list of
simple experiments taken from some published deep DA
papers are presented in Table 5.
In [62], [108], and [95], the authors used the Office31 dataset1 as one of the evaluation data sets, as shown in
Fig. 1(a). The Office dataset is a computer vision classification data set with images from three distinct domains:
Amazon (A), DSLR (D), and Webcam (W). The largest
domain, Amazon, has 2817 labeled images and its corresponding 31 classes, which consist of objects commonly
encountered in office settings. By using this dataset, previous works can show the performance of methods across
all six possible DA tasks. [62] showed comparison experiments among the standard AlexNet [53], the DANN method
[18], and the MMD algorithm and its variations, such as
DDC [97], DAN [61], JAN [62] and RTN [63]. Zellinger et
al. [108] evaluated their proposed CMD algorithm in comparison to other discrepancy-based methods (DDC, deep
Figure 15. The progressive network architecture. [79]
5. Heterogeneous Domain Adaption
In heterogeneous DA, the feature spaces of the source
and target domains are not the same, Xs 6= Xt, and the dimensions of the feature spaces may also differ. Most heterogeneous DA with shallow methods fall into two categories:
the first learns feature transformations to project the source
and target features onto a common subspace, and the second
transforms one of them to align with the other. However, for
deep methods, the network generally shares or reuses the
first n layers between the source and target domains, which
limits the feature spaces of the input to the same dimension.
Therefore, little work focuses on heterogeneous DA with
deep methods.
[83] proposed weakly shared DTNs to transfer labeled
information across heterogeneous domains, particularly
1
15
https://cs.stanford.edu/∼jhoffman/domainadapt/
Table 5. Comparison between Transfer Learning and Non-Adaption Learning Methods
Data Set
(reference)
Source vs. Target
Office-31 Dataset
ACC (unit:%)[62]
Office-31 Dataset
ACC (unit:%)[108]
Office-31 Dataset
ACC (unit:%)[95]
MNIST, USPS,
and SVHN
digits datasets
ACC (unit:%)[96]
A vs. W
D vs. W
W vs. D
A vs. D
D vs. A
W vs. A
Avg
A vs. W
D vs. W
W vs. D
A vs. D
D vs. A
W vs. A
Avg
A vs. W
D vs. W
W vs. D
A vs. D
D vs. A
W vs. A
Avg
M vs. U
U vs. M
S vs. M
Baselines
Deep Domain Adaption Methods
AlexNet
61.6±0.5
95.4±0.3
99.0±0.2
63.8±0.5
51.1±0.6
49.8±0.4
70.1
AlexNet
61.6
95.4
99.0
63.8
51.1
49.8
70.1
DDC
61.8±0.4
95.0±0.5
98.5±0.4
64.4±0.3
52.1±0.6
52.2±0.4
70.6
Deep CORAL
66.4
95.7
99.2
66.8
52.8
51.5
72.1
DAN
68.5
96.0±0.3
99.0±0.3
67.0±0.4
54.0±0.5
53.1±0.5
72.9
CMD
77.0±0.6
96.3±0.4
99.2±0.2
79.6±0.6
63.8±0.7
63.3±0.6
79.9
RTN
73.3±0.3
96.8±0.2
99.6±0.1
71.0±0.2
50.5±0.3
51.0±0.1
73.7
DLID
51.9
78.2
89.9
-
AlexNet
DLID
DANN
Soft Labels
56.5±0.3
92.4±0.3
93.6±0.2
64.6±0.4
47.6±0.1
42.7±0.1
66.2
VGG-16
75.2±1.6
57.1±1.7
60.1±1.1
51.9
78.2
89.9
DANN
77.1±1.8
73.0±2.0
73.9
53.6±0.2
71.2±0.0
83.5±0.0
CoGAN
91.2±0.8
89.1±0.8
-
82.7±0.7
95.9±0.6
98.3±0.3
84.9±1.2
66.0±0.5
65.2±0.6
82.17
ADDA
89.4±0.2
90.1±0.8
76.0±1.8
DANN
73.0±0.5
96.4±0.3
99.2±0.3
72.3±0.3
53.4±0.4
51.2±0.5
74.3
DANN
73
96.4
99.2
Confusion
+Soft
82.7±0.8
95.7±0.5
97.6±0.2
86.1±1.2
66.2±0.3
65.0±0.5
82.22
be caused by poses, resolution, illuminations, expressions,
and modality. Kan et al. [49] proposed a bi-shifting autoencoder network (BAE) for face recognition across view
angle, ethnicity, and imaging sensor. In BAE, source domain samples are shifted to the target domain, and sparse
reconstruction is used with several local neighbors from the
target domain to ensure its correction, and vice versa. Single sample per person domain adaptation network (SSPPDAN) in [43] generates synthetic images with varying poses
to increase the number of samples in the source domain and
bridges the gap between the synthetic and source domains
by adversarial training with a GRL in real-world face recognition. [85] improved the performance of video face recognition by using an adversarial-based approach with largescale unlabeled videos, labeled still images and synthesized
images. Considering that age variations are difficult problems for smile detection and that networks trained on the
current benchmarks do not perform well on young children,
Xia et al. [101] applied DAN [61] and JAN [62] (mentioned
CROAL [87], DLID [10], AdaBN [58]) and the adversarialbased method DANN. [95] proposed an algorithm combining soft label loss and domain confusion loss, and they also
compared them with DANN and DLID under a supervised
DA setting.
In [96], MNIST2 (M), USPS3 (U), and SVHN4 (S) digit
datasets (shown in Fig. 1(b)) are used for a cross-domain
hand-written digit recognition task, and the experiment
showed the comparison results on some adversarial-based
methods, such as DANN, CoGAN [59] and ADDA [96],
where the baseline is VGG-16 [84].
6.2. Face Recognition
The performance of face recognition significantly degrades when there are variations in the test images that are
not present in the training images. The dataset shift can
2
JAN
75.2±0.4
96.6±0.2
99.6±0.1
72.8±0.3
57.5±0.2
56.3±0.2
76.3
AdaBN
74.2
95.7
99.8
73.1
59.8
57.4
76.7
Domain
Confusion
82.8±0.9
95.6±0.4
97.5±0.2
85.9±1.1
66.2±0.4
64.9±0.5
82.13
http://yann.lecun.com/exdb/mnist/
3 http://statweb.stanford.edu/∼tibs/ElemStatLearn/data.html
4 http://ufldl.stanford.edu/housenumbers/
16
in Section 4.1.1) to two baseline deep models, i.e., AlexNet
and ResNet, to transfer the knowledge from adults to infants.
ages with the help of virtual ones. It uses the global label
distribution loss of the images and local label distribution
loss of the landmark superpixels in the target domain to
effectively regularize the fine-tuning of the semantic segmentation network. Chen et al. [9] proposed a framework
for cross-city semantic segmentation. The framework assigns pseudo labels to pixels/grids in the target domain and
jointly utilizes global and class-wise alignment by domain
adversarial learning to minimize domain shift.
6.3. Object Detection
Recent advances in object detection are driven by regionbased convolutional neural networks (R-CNNs [28], fast RCNNs [27] and faster R-CNNs [74]). They are composed
of a window selection mechanism and classifiers that are
pre-trained labeled bounding boxes by using the features
extracted from CNNs. At test time, the classifier decides
whether a region obtained by sliding windows contains the
object. Although the R-CNN algorithm is effective, a large
amount of bounding box labeled data is required to train
each detection category. To solve the problem of lacking labeled data, considering the window selection mechanism as
being domain independent, deep DA methods can be used
in classifiers to adapt to the target domain.
Because R-CNNs train classifiers on regions just like
classification, weak labeled data (such as image-level class
labels) are directly useful for the detector. Most works
learn the detector with limited bounding box labeled data
and massive weak labeled data. The large-scale detection
through adaptation (LSDA) [39] trains a classification layer
for the target domain and then uses a pre-trained source
model along with output layer adaptation techniques to update the target classification parameters directly. Rochan et
al. [75] used word vectors to establish the semantic relatedness between weak labeled source objects and target objects
and then transferred the bounding box labeled information
from source objects to target objects based on their relatedness. Extending [39] and [75], Tang et al. [93] transferred
visual (based on the LSDA model) and semantic similarity
(based on work vectors) for training an object detector on
weak labeled category.
6.5. Image-to-Image Translation
Image-to-image translation has recently achieved great
success with deep DA, and it has been applied to various
tasks, such as style transferring and generating photographs
from sketches or from attribute and semantic layouts. More
approaches of image-to-image translation use a dataset of
paired images and incorporate a DA algorithm into generative networks. Isola et al. [48] proposed the pix2pix framework, which uses a conditional GAN to learn a mapping
from source to target images. Tzeng et al. [94] utilized
domain confusion loss and pairwise loss to adapt from simulation to real-world data in a PR2 robot. However, several other methods also address the unpaired setting, such
as CoGAN [59], cycle GAN [113], dual GAN [105] and
disco GAN [50]. [89] employs a compound loss function
that consists of a multiclass GAN loss, a regularizing component and an f-constancy component to transfer unlabeled
face images to emoji images.
Matching the statistical distribution by fine-tuning a deep
network is another way to achieve image-to-image translation. Gatys et al. [20] fine-tuned the CNN to achieve DA
by the total loss, which is a linear combination between
the content and the style loss, such that the target image
is rendered in the style of the source image maintaining
the content. The content loss minimizes the mean squared
difference of the feature representation between the original image and generated image in higher layers, while the
style loss minimizes the element-wise mean squared difference between the Gram matrix of them on each layer. [57]
demonstrated that matching the Gram matrices of feature
maps is equivalent to minimizing the MMD. Rather than
MMD, [71] proposed a deep generative correlation alignment network (DGCAN) that bridges the domain discrepancy between CAD synthetic and real images by applying
the content and CORAL losses to different layers.
6.4. Semantic Segmentation
Fully convolutional network models (FCNs) for dense
prediction have proven to be successful for evaluating semantic segmentation, but their performance will also degrade under domain shifts. Therefore, some work has also
explored using weak labels to improve the performance of
semantic segmentation. Hong et al. [44] used a novel
encoder-decoder architecture with attention model by transferring weak class labeled knowledge in the source domain,
while [51, 81] transferred weak object location knowledge.
Limited attention has also been given to deep unsupervised DA in semantic segmentation. Hoffman et al. [42]
first introduced it, in which global domain alignment is performed using FCNs with adversarial-based training, while
transferring spatial layout is achieved by leveraging classaware constrained multiple instance loss. Zhang et al.
[112] enhanced the segmentation performance on real im-
6.6. Person Re-identification
In the community, person re-identification (re-ID) has
become increasingly popular. When given video sequences
of a person, person re-ID recognizes whether this person
has been in another camera to compensate for the limitations of fixed devices. Recently, deep DA methods have
been used in re-ID when models trained on one dataset are
17
Figure 16. The single sample per person domain adaptation network (SSPP-DAN) architecture. [43]
tasks.
Deep DA is classified as homogeneous DA and heterogeneous DA, and it can be further divided into supervised,
semi-supervised and unsupervised settings. The first setting
is the simplest but is generally limited due to the need for
labeled data; thus, most previous works focused on unsupervised cases. Semi-supervised deep DA is a hybrid method
that combines the methods of the supervised and unsupervised settings.
Furthermore, the approaches of deep DA can be classified into one-step DA and multi-step DA considering the
distance of the source and target domains. When the distance is small, one-step DA can be used based on training loss. It consists of the discrepancy-based approach, the
adversarial-based approach, and the reconstruction-based
approach. When the source and target domains are not directly related, multi-step (or transitive) DA can be used.
The key of multi-step DA is to select and utilize intermediate domains, thus falling into three categories, including
hand-crafted, feature-based and representation-based selection mechanisms.
In addition, the network of deep DA generally shares or
reuses the first n layers between the source and target domains, which limits the feature spaces of the input to the
same dimension. Therefore, most existing algorithms focus
on homogeneous deep DA, which assumes that the feature
spaces between the source and target domains are the same.
However, this assumption may not be true in many applications. We expect to transfer knowledge without this severe
limitation and take advantage of existing datasets to help
with more tasks. Heterogeneous deep DA may attract increasingly more attention in the future.
Finally, deep DA techniques have been successfully applied in many real-world applications, including image classification, face recognition, and style translation. We have
also found that only a few papers address adaptation beyond
classification and recognition, such as object detection, semantic segmentation and person re-identification. How to
Figure 17. The architecture of pixel-level adversarial and
constraint-based adaptation. [42]
directly used on another. Xiao et al. [102] proposed the
domain-guided dropout algorithm to discard useless neurons for re-identifying persons on multiple datasets simultaneously. Inspired by cycle GAN and Siamese network, the
similarity preserving generative adversarial network (SPGAN) [15] translates the labeled source image to the target
domain, preserving self similarity and domain-dissimilarity
in an unsupervised manner, and then it trains re-ID models
with the translated images using supervised feature learning
methods.
7. Conclusion
In a broad sense, deep DA is utilizing deep networks to
enhance the performance of DA, such as shallow DA methods with features extracted by deep networks. In a narrow
sense, deep DA is based on deep learning architectures designed for DA and optimized by back propagation. In this
survey paper, we focus on this narrow definition, and we
have reviewed deep DA techniques on visual categorization
18
achieve these tasks with no or a very limited amount of data
is probably one of the main challenges that should be addressed by deep DA in the next few years.
[13] G. Csurka. Domain adaptation for visual applications: A
comprehensive survey. arXiv preprint arXiv:1702.05374,
2017.
[14] O. Day and T. M. Khoshgoftaar. A survey on heterogeneous
transfer learning. Journal of Big Data, 4(1):29, 2017.
[15] W. Deng, L. Zheng, G. Kang, Y. Yang, Q. Ye, and
J. Jiao. Image-image domain adaptation with preserved
self-similarity and domain-dissimilarity for person reidentification. arXiv preprint arXiv:1711.07027, 2017.
[16] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang,
E. Tzeng, and T. Darrell. Decaf: A deep convolutional
activation feature for generic visual recognition. In International conference on machine learning, pages 647–655,
2014.
[17] D. Eigen, C. Puhrsch, and R. Fergus. Depth map prediction
from a single image using a multi-scale deep network. In
Advances in neural information processing systems, pages
2366–2374, 2014.
[18] Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on
Machine Learning, pages 1180–1189, 2015.
[19] Y. Ganin, E. Ustinova, H. Ajakan, P. Germain,
H. Larochelle, F. Laviolette, M. Marchand, and V. Lempitsky. Domain-adversarial training of neural networks.
Journal of Machine Learning Research, 17(59):1–35,
2016.
[20] L. A. Gatys, A. S. Ecker, and M. Bethge. Image style transfer using convolutional neural networks. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 2414–2423, 2016.
[21] W. Ge and Y. Yu. Borrowing treasures from the wealthy:
Deep transfer learning through selective joint fine-tuning.
arXiv preprint arXiv:1702.08690, 2017.
[22] T. Gebru, J. Hoffman, and L. Fei-Fei. Fine-grained recognition in the wild: A multi-task domain adaptation approach.
arXiv preprint arXiv:1709.02476, 2017.
[23] M. Gheisari and M. S. Baghshah. Unsupervised domain
adaptation via representation learning and adaptive classifier learning. Neurocomputing, 165:300–311, 2015.
[24] M. Ghifary, W. Bastiaan Kleijn, M. Zhang, and D. Balduzzi. Domain generalization for object recognition with
multi-task autoencoders. In Proceedings of the IEEE international conference on computer vision, pages 2551–2559,
2015.
[25] M. Ghifary, W. B. Kleijn, and M. Zhang. Domain adaptive neural networks for object recognition. In Pacific Rim
International Conference on Artificial Intelligence, pages
898–904. Springer, 2014.
[26] M. Ghifary, W. B. Kleijn, M. Zhang, D. Balduzzi, and
W. Li. Deep reconstruction-classification networks for unsupervised domain adaptation. In European Conference on
Computer Vision, pages 597–613. Springer, 2016.
[27] R. Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440–1448,
2015.
[28] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
8. Acknowledgements
This work was partially supported by the National
Natural Science Foundation of China under Grant Nos.
61573068, 61471048, and 61375031, and Beijing Nova
Program under Grant No. Z161100004916088.
References
[1] Y. Bengio. Learning deep architectures for ai. Foundations
and Trends in Machine Learning, 2(1):1–127, 2009.
[2] K. M. Borgwardt, A. Gretton, M. J. Rasch, H.-P. Kriegel,
B. Schölkopf, and A. J. Smola. Integrating structured biological data by kernel maximum mean discrepancy. Bioinformatics, 22(14):e49–e57, 2006.
[3] K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and
D. Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. arXiv preprint
arXiv:1612.05424, 2016.
[4] K. Bousmalis, G. Trigeorgis, N. Silberman, D. Krishnan,
and D. Erhan. Domain separation networks. In Advances
in Neural Information Processing Systems, pages 343–351,
2016.
[5] L. Bruzzone and M. Marconcini. Domain adaptation problems: A dasvm classification technique and a circular validation strategy. IEEE transactions on pattern analysis and
machine intelligence, 32(5):770–787, 2010.
[6] Z. Cao, M. Long, J. Wang, and M. I. Jordan. Partial transfer
learning with selective adversarial networks. arXiv preprint
arXiv:1707.07901, 2017.
[7] M. Chen, Z. Xu, K. Weinberger, and F. Sha. Marginalized denoising autoencoders for domain adaptation. arXiv
preprint arXiv:1206.4683, 2012.
[8] W.-Y. Chen, T.-M. H. Hsu, Y.-H. H. Tsai, Y.-C. F. Wang,
and M.-S. Chen. Transfer neural trees for heterogeneous
domain adaptation. In European Conference on Computer
Vision, pages 399–414. Springer, 2016.
[9] Y.-H. Chen, W.-Y. Chen, Y.-T. Chen, B.-C. Tsai, Y.-C. F.
Wang, and M. Sun. No more discrimination: Cross
city adaptation of road scene segmenters. arXiv preprint
arXiv:1704.08509, 2017.
[10] S. Chopra, S. Balakrishnan, and R. Gopalan. Dlid: Deep
learning for domain adaptation by interpolating between
domains. In ICML workshop on challenges in representation learning, volume 2, 2013.
[11] B. Chu, V. Madhavan, O. Beijbom, J. Hoffman, and T. Darrell. Best practices for fine-tuning visual classifiers to new
domains. In Computer Vision–ECCV 2016 Workshops,
pages 435–442. Springer, 2016.
[12] W.-S. Chu, F. De la Torre, and J. F. Cohn. Selective transfer
machine for personalized facial action unit detection. In
Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 3515–3522, 2013.
19
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43] S. Hong, W. Im, J. Ryu, and H. S. Yang. Sspp-dan: Deep
domain adaptation network for face recognition with single
sample per person. arXiv preprint arXiv:1702.04069, 2017.
[44] S. Hong, J. Oh, H. Lee, and B. Han. Learning transferrable
knowledge for semantic segmentation with deep convolutional neural network. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
3204–3212, 2016.
[45] J. Hu, J. Lu, and Y.-P. Tan. Deep transfer metric learning.
In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 325–333, 2015.
[46] X. Huang and S. Belongie. Arbitrary style transfer in realtime with adaptive instance normalization. arXiv preprint
arXiv:1703.06868, 2017.
[47] S. Ioffe and C. Szegedy. Batch normalization: Accelerating
deep network training by reducing internal covariate shift.
In International Conference on Machine Learning, pages
448–456, 2015.
[48] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Imageto-image translation with conditional adversarial networks.
arXiv preprint arXiv:1611.07004, 2016.
[49] M. Kan, S. Shan, and X. Chen. Bi-shifting auto-encoder
for unsupervised domain adaptation. In Proceedings of the
IEEE International Conference on Computer Vision, pages
3846–3854, 2015.
[50] T. Kim, M. Cha, H. Kim, J. Lee, and J. Kim. Learning to
discover cross-domain relations with generative adversarial
networks. arXiv preprint arXiv:1703.05192, 2017.
[51] A. Kolesnikov and C. H. Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In European Conference on Computer Vision,
pages 695–711. Springer, 2016.
[52] P. Kontschieder, M. Fiterau, A. Criminisi, and S. Rota Bulo.
Deep neural decision forests. In Proceedings of the IEEE
International Conference on Computer Vision, pages 1467–
1475, 2015.
[53] 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.
[54] C. H. Lampert, H. Nickisch, and S. Harmeling. Learning to detect unseen object classes by between-class attribute transfer. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 951–
958. IEEE, 2009.
[55] C. Li and M. Wand. Precomputed real-time texture synthesis with markovian generative adversarial networks. In
European Conference on Computer Vision, pages 702–716.
Springer, 2016.
[56] Y. Li, K. Swersky, and R. Zemel. Generative moment
matching networks. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pages
1718–1727, 2015.
[57] Y. Li, N. Wang, J. Liu, and X. Hou. Demystifying neural
style transfer. arXiv preprint arXiv:1701.01036, 2017.
[58] Y. Li, N. Wang, J. Shi, J. Liu, and X. Hou. Revisiting
batch normalization for practical domain adaptation. arXiv
preprint arXiv:1603.04779, 2016.
segmentation. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 580–587,
2014.
X. Glorot, A. Bordes, and Y. Bengio. Domain adaptation
for large-scale sentiment classification: A deep learning approach. In Proceedings of the 28th international conference
on machine learning (ICML-11), pages 513–520, 2011.
B. Gong, K. Grauman, and F. Sha. Connecting the
dots with landmarks: Discriminatively learning domaininvariant features for unsupervised domain adaptation. In
International Conference on Machine Learning, pages
222–230, 2013.
B. Gong, Y. Shi, F. Sha, and K. Grauman. Geodesic flow
kernel for unsupervised domain adaptation. In Computer
Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 2066–2073. IEEE, 2012.
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio.
Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014.
R. Gopalan, R. Li, and R. Chellappa. Domain adaptation
for object recognition: An unsupervised approach. In Computer Vision (ICCV), 2011 IEEE International Conference
on, pages 999–1006. IEEE, 2011.
S. Gupta, J. Hoffman, and J. Malik. Cross modal distillation for supervision transfer. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 2827–2836, 2016.
D. He, Y. Xia, T. Qin, L. Wang, N. Yu, T. Liu, and W.-Y.
Ma. Dual learning for machine translation. In Advances
in Neural Information Processing Systems, pages 820–828,
2016.
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.
G. Hinton, O. Vinyals, and J. Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531,
2015.
G. E. Hinton, S. Osindero, and Y.-W. Teh. A fast learning algorithm for deep belief nets. Neural computation,
18(7):1527–1554, 2006.
J. Hoffman, S. Guadarrama, E. S. Tzeng, R. Hu, J. Donahue, R. Girshick, T. Darrell, and K. Saenko. Lsda: Large
scale detection through adaptation. In Advances in Neural
Information Processing Systems, pages 3536–3544, 2014.
J. Hoffman, S. Gupta, J. Leong, S. Guadarrama, and T. Darrell. Cross-modal adaptation for rgb-d detection. In
Robotics and Automation (ICRA), 2016 IEEE International
Conference on, pages 5032–5039. IEEE, 2016.
J. Hoffman, E. Tzeng, J. Donahue, Y. Jia, K. Saenko, and
T. Darrell. One-shot adaptation of supervised deep convolutional models. arXiv preprint arXiv:1312.6204, 2013.
J. Hoffman, D. Wang, F. Yu, and T. Darrell. Fcns in the
wild: Pixel-level adversarial and constraint-based adaptation. arXiv preprint arXiv:1612.02649, 2016.
20
[76] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and
Computer-Assisted Intervention, pages 234–241. Springer,
2015.
[77] S. Rota Bulo and P. Kontschieder. Neural decision forests
for semantic image labelling. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition,
pages 81–88, 2014.
[78] A. Rozantsev, M. Salzmann, and P. Fua. Beyond sharing weights for deep domain adaptation. arXiv preprint
arXiv:1603.06432, 2016.
[79] A. A. Rusu, N. C. Rabinowitz, G. Desjardins, H. Soyer,
J. Kirkpatrick, K. Kavukcuoglu, R. Pascanu, and R. Hadsell.
Progressive neural networks.
arXiv preprint
arXiv:1606.04671, 2016.
[80] L. Shao, F. Zhu, and X. Li. Transfer learning for visual
categorization: A survey. IEEE transactions on neural networks and learning systems, 26(5):1019–1034, 2015.
[81] W. Shimoda and K. Yanai. Distinct class-specific saliency
maps for weakly supervised semantic segmentation. In European Conference on Computer Vision, pages 218–234.
Springer, 2016.
[82] A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang,
and R. Webb. Learning from simulated and unsupervised images through adversarial training. arXiv preprint
arXiv:1612.07828, 2016.
[83] X. Shu, G.-J. Qi, J. Tang, and J. Wang. Weakly-shared
deep transfer networks for heterogeneous-domain knowledge propagation. In Proceedings of the 23rd ACM international conference on Multimedia, pages 35–44. ACM,
2015.
[84] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014.
[85] K. Sohn, S. Liu, G. Zhong, X. Yu, M.-H. Yang, and
M. Chandraker. Unsupervised domain adaptation for
face recognition in unlabeled videos. arXiv preprint
arXiv:1708.02191, 2017.
[86] B. Sun, J. Feng, and K. Saenko. Return of frustratingly easy
domain adaptation. In AAAI, volume 6, page 8, 2016.
[87] B. Sun and K. Saenko. Deep coral: Correlation alignment
for deep domain adaptation. In Computer Vision–ECCV
2016 Workshops, pages 443–450. Springer, 2016.
[88] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. In Proceedings of the
IEEE conference on computer vision and pattern recognition, pages 1–9, 2015.
[89] Y. Taigman, A. Polyak, and L. Wolf.
Unsupervised cross-domain image generation. arXiv preprint
arXiv:1611.02200, 2016.
[90] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. Deepface:
Closing the gap to human-level performance in face verification. In Proceedings of the IEEE conference on computer
vision and pattern recognition, pages 1701–1708, 2014.
[59] M.-Y. Liu and O. Tuzel. Coupled generative adversarial
networks. In Advances in neural information processing
systems, pages 469–477, 2016.
[60] W. Liu, Z. Wang, X. Liu, N. Zeng, Y. Liu, and F. E. Alsaadi. A survey of deep neural network architectures and
their applications. Neurocomputing, 234:11–26, 2017.
[61] M. Long, Y. Cao, J. Wang, and M. Jordan. Learning transferable features with deep adaptation networks. In International Conference on Machine Learning, pages 97–105,
2015.
[62] M. Long, J. Wang, and M. I. Jordan. Deep transfer
learning with joint adaptation networks. arXiv preprint
arXiv:1605.06636, 2016.
[63] M. Long, H. Zhu, J. Wang, and M. I. Jordan. Unsupervised domain adaptation with residual transfer networks. In
Advances in Neural Information Processing Systems, pages
136–144, 2016.
[64] H. V. Nguyen, H. T. Ho, V. M. Patel, and R. Chellappa.
Dash-n: Joint hierarchical domain adaptation and feature learning. IEEE Transactions on Image Processing,
24(12):5479–5491, 2015.
[65] S. Pachori, A. Deshpande, and S. Raman. Hashing in the
zero shot framework with domain adaptation. Neurocomputing, 2017.
[66] S. J. Pan, I. W. Tsang, J. T. Kwok, and Q. Yang. Domain
adaptation via transfer component analysis. IEEE Transactions on Neural Networks, 22(2):199–210, 2011.
[67] S. J. Pan and Q. Yang. A survey on transfer learning.
IEEE Transactions on knowledge and data engineering,
22(10):1345–1359, 2010.
[68] V. M. Patel, R. Gopalan, R. Li, and R. Chellappa. Visual
domain adaptation: A survey of recent advances. IEEE signal processing magazine, 32(3):53–69, 2015.
[69] K.-C. Peng, Z. Wu, and J. Ernst. Zero-shot deep domain
adaptation. arXiv preprint arXiv:1707.01922, 2017.
[70] X. Peng, J. Hoffman, X. Y. Stella, and K. Saenko. Fine-tocoarse knowledge transfer for low-res image classification.
In Image Processing (ICIP), 2016 IEEE International Conference on, pages 3683–3687. IEEE, 2016.
[71] X. Peng and K. Saenko. Synthetic to real adaptation with
deep generative correlation alignment networks. arXiv
preprint arXiv:1701.05524, 2017.
[72] A. Raj, V. P. Namboodiri, and T. Tuytelaars. Subspace
alignment based domain adaptation for rcnn detector. arXiv
preprint arXiv:1507.05578, 2015.
[73] S.-A. Rebuffi, H. Bilen, and A. Vedaldi. Learning multiple visual domains with residual adapters. arXiv preprint
arXiv:1705.08045, 2017.
[74] S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91–99, 2015.
[75] M. Rochan and Y. Wang. Weakly supervised localization
of novel objects using appearance transfer. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, pages 4315–4324, 2015.
21
[106] D. Yoo, N. Kim, S. Park, A. S. Paek, and I. S. Kweon.
Pixel-level domain transfer. In European Conference on
Computer Vision, pages 517–532. Springer, 2016.
[107] J. Yosinski, J. Clune, Y. Bengio, and H. Lipson. How
transferable are features in deep neural networks?
In
Advances in neural information processing systems, pages
3320–3328, 2014.
[108] W. Zellinger, T. Grubinger, E. Lughofer, T. Natschläger,
and S. Saminger-Platz.
Central moment discrepancy
(cmd) for domain-invariant representation learning. arXiv
preprint arXiv:1702.08811, 2017.
[109] J. Zhang, W. Li, and P. Ogunbona. Transfer learning for
cross-dataset recognition: A survey. 2017.
[110] L. Zhang, Z. He, and Y. Liu. Deep object recognition across
domains based on adaptive extreme learning machine. Neurocomputing, 239:194–203, 2017.
[111] X. Zhang, F. X. Yu, S.-F. Chang, and S. Wang. Deep
transfer network: Unsupervised domain adaptation. arXiv
preprint arXiv:1503.00591, 2015.
[112] Y. Zhang, P. David, and B. Gong. Curriculum domain
adaptation for semantic segmentation of urban scenes. In
The IEEE International Conference on Computer Vision
(ICCV), volume 2, page 6, 2017.
[113] 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.
[114] F. Zhuang, X. Cheng, P. Luo, S. J. Pan, and Q. He. Supervised representation learning: Transfer learning with deep
autoencoders. In IJCAI, pages 4119–4125, 2015.
[91] B. Tan, Y. Song, E. Zhong, and Q. Yang. Transitive transfer
learning. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, pages 1155–1164. ACM, 2015.
[92] B. Tan, Y. Zhang, S. J. Pan, and Q. Yang. Distant domain
transfer learning. In AAAI, pages 2604–2610, 2017.
[93] Y. Tang, J. Wang, B. Gao, E. Dellandréa, R. Gaizauskas,
and L. Chen. Large scale semi-supervised object detection
using visual and semantic knowledge transfer. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2119–2128, 2016.
[94] E. Tzeng, C. Devin, J. Hoffman, C. Finn, P. Abbeel,
S. Levine, K. Saenko, and T. Darrell. Adapting deep visuomotor representations with weak pairwise constraints.
CoRR, vol. abs/1511.07111, 2015.
[95] E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko. Simultaneous deep transfer across domains and tasks. In Proceedings of the IEEE International Conference on Computer Vision, pages 4068–4076, 2015.
[96] E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell. Adversarial discriminative domain adaptation. arXiv preprint
arXiv:1702.05464, 2017.
[97] E. Tzeng, J. Hoffman, N. Zhang, K. Saenko, and T. Darrell. Deep domain confusion: Maximizing for domain invariance. arXiv preprint arXiv:1412.3474, 2014.
[98] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Improved texture networks: Maximizing quality and diversity in feedforward stylization and texture synthesis. arXiv preprint
arXiv:1701.02096, 2017.
[99] P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P.A. Manzagol. Stacked denoising autoencoders: Learning
useful representations in a deep network with a local denoising criterion. Journal of Machine Learning Research,
11(Dec):3371–3408, 2010.
[100] X. Wang, X. Duan, and X. Bai. Deep sketch feature for
cross-domain image retrieval. Neurocomputing, 207:387–
397, 2016.
[101] Y. Xia, D. Huang, and Y. Wang. Detecting smiles of young
children via deep transfer learning. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 1673–1681, 2017.
[102] T. Xiao, H. Li, W. Ouyang, and X. Wang. Learning deep
feature representations with domain guided dropout for person re-identification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages
1249–1258, 2016.
[103] M. Xie, N. Jean, M. Burke, D. Lobell, and S. Ermon.
Transfer learning from deep features for remote sensing and
poverty mapping. 2015.
[104] H. Yan, Y. Ding, P. Li, Q. Wang, Y. Xu, and W. Zuo. Mind
the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. arXiv preprint
arXiv:1705.00609, 2017.
[105] Z. Yi, H. Zhang, P. T. Gong, et al. Dualgan: Unsupervised
dual learning for image-to-image translation. arXiv preprint
arXiv:1704.02510, 2017.
22
| 1cs.CV
|
Safe Sequential Path Planning of Multi-Vehicle Systems via
Double-Obstacle Hamilton-Jacobi-Isaacs Variational Inequality
arXiv:1412.7223v2 [cs.MA] 21 Mar 2016
Mo Chen, Jaime F. Fisac, Shankar Sastry, Claire J. Tomlin
Abstract— We consider the problem of planning trajectories
for a group of N vehicles, each aiming to reach its own
target set while avoiding danger zones of other vehicles. The
analysis of problems like this is extremely important practically,
especially given the growing interest in utilizing unmanned
aircraft systems for civil purposes. The direct solution of this
problem by solving a single-obstacle Hamilton-Jacobi-Isaacs
(HJI) variational inequality (VI) is numerically intractable
due to the exponential scaling of computation complexity with
problem dimensionality. Furthermore, the single-obstacle HJI
VI cannot directly handle situations in which vehicles do not
have a common scheduled arrival time. Instead, we perform
sequential path planning by considering vehicles in order
of priority, modeling higher-priority vehicles as time-varying
obstacles for lower-priority vehicles. To do this, we solve a
double-obstacle HJI VI which allows us to obtain the reachavoid set, defined as the set of states from which a vehicle
can reach its target while staying within a time-varying state
constraint set. From the solution of the double-obstacle HJI VI,
we can also extract the latest start time and the optimal control
for each vehicle. This is a first application of the double-obstacle
HJI VI which can handle systems with time-varying dynamics,
target sets, and state constraint sets, and results in computation
complexity that scales linearly, as opposed to exponentially, with
the number of vehicles in consideration.
I. I NTRODUCTION
Consider a group of autonomous vehicles trying to perform a task or reach a goal which may be time-varying in
their joint state space, while avoiding obstacles and other
vehicles. Providing safety and performance guarantees for
such a multi-agent autonomous system (MAAS) is very
relevant practically: Recently, there has been a growing
interest in using unmanned aerial vehicles (UAVs) for civil
applications, as companies like Amazon and Google are
looking in the near future to send UAVs into the airspace
to deliver packages [1], [2]. Government agencies such as
the Federal Aviation Administration (FAA) and National
Aeronautics and Space Administration (NASA) of the United
States are also expressing growing interest in analyzing these
problems in order to prevent airspace conflicts that could
arise with the introduction of potentially many UAVs in an
urban environment [3]. In addition, UAVs can be used not
only to deliver packages quickly, but in any situation where
fast response is desired. For example, UAVs can provide
This work has been supported in part by NSF under CPS:ActionWebs
(CNS-931843), by ONR under the HUNT (N0014-08-0696) and SMARTS
(N00014-09-1-1051) MURIs and by grant N00014-12-1-0609, by AFOSR
under the CHASE MURI (FA9550-10-1-0567). The research of J.F. Fisac
has received funding from the “la Caixa” Foundation.
All authors are with the Department of Electrical Engineering and
Computer Sciences, University of California, Berkeley. {mochen72, jfisac,
sastry, tomlin}@eecs.berkeley.edu
emergency supplies to disaster-struck areas that are otherwise
difficult to reach [4].
In general, MAASs are difficult to analyze due to their
inherent high dimensionality. MAASs also often involve
aspects of cooperation and asymmetric goals among the
vehicles or teams of vehicles, making their analysis particularly interesting. MAASs have been explored extensively in
the literature. Some researchers have done work on multivehicle path planning in the presence of other unknown
vehicles or moving entities with assumptions on their specific
control strategies [5]. In a number of formulations for safe
multi-vehicle navigation, these assumed strategies induce
velocity obstacles that vehicles must avoid to maintain safety
[6], [7]. Researchers have also used potential functions to
perform collision avoidance while maintaining formation
given a predefined trajectory [8], [9]. However, these bodies
of work have not considered trajectory planning and collision
avoidance simultaneously.
One well-known technique for optimal trajectory planning
under disturbances or adversaries is reachability analysis, in
which one computes the reach-avoid set, defined as the set
of states from which the system can reach a target set while
remaining within a state constraint set for all time. For reachability of systems of up to five dimensions, single-obstacle
Hamilton-Jacobi-Isaacs (HJI) variational inequalities (VI)
[10], [11] have been used in situations where obstacles and
target sets are static. Another HJI VI formulation [12] is able
to handle problems with moving target sets with no obstacles.
A major practical appeal of the above approaches stems
from the availability of modern numerical tools such as [10],
[13], [14], [15], which can efficiently solve HJI equations
when the problem dimension is low. These numerical tools
have been successfully used to solve a variety of differential games, path planning problems, and optimal control
problems[10], [16], [17]. Despite the power of the previous
HJI formulations, the approaches become numerically intractable very quickly as the number of vehicles in the system
is increased. This is because the numerical computations are
done on a grid in the joint state space of the system, resulting
in an exponential scaling of computation complexity with
respect to the dimensionality of the problem. Furthermore,
state constraint sets, while useful for modeling unsafe vehicle
configurations, are required to be time-invariant in [10],
[11], [18]. To solve problems involving time-varying state
constraints, [19] proposed to augment the state space with
time; however, this process introduces an extra state space
dimension, resulting in added computation complexity.
Recently, [20] presented a double-obstacle HJI VI which
handles problems in which the dynamics, target sets, and
state constraint sets are all time-varying, and provided a numerical implementation based on well-known schemes. The
formulation does not introduce any additional computation
overhead compared to the above-mentioned techniques, yet
it still maintains the same guarantees on the system’s safety
and performance. In this paper, we provide a first application
of the theory presented in [20]. As a point of clarification,
“obstacles” in the context of HJI VIs refer to the effective
constraints in the HJI VI, while obstacles in the state space
represent physical obstacles that vehicles must avoid.
Our contributions are as follows. First, we formulate
a multi-vehicle collision avoidance problem involving N
autonomous vehicles. Each vehicle seeks to get to its own
target sets while avoiding obstacles and collision with all
other vehicles. To reduce the problem complexity to make
the problem tractable, we assign a priority to each vehicle,
and model higher-priority vehicles as time-varying obstacles
that need to be avoided. We then utilize the double-obstacle
HJI VI proposed in [20] to compute reach-avoid sets to plan
trajectories for vehicles in order of priority. This way, we
are able to offer a tractable solution that scales linearly, as
opposed to exponentially, with the number of vehicles. We
demonstrate the scalability of our approach in a four-vehicle
system.
II. P ROBLEM F ORMULATION
Consider N vehicles Pi , i = 1 . . . , N , each trying to reach
one of N target sets Ti , i = 1 . . . , N , while avoiding obstacles and collision with each other. Each vehicle i has states
xi ∈ Rni and travels on a domain Ω = Ωobs ∪ Ωf ree ∈ Rp ,
where Ωobs represents the obstacles that each vehicle must
avoid, and Ωf ree represents all other states in the domain
on which vehicles can move. Each vehicle i = 1, 2, . . . , N
STA
moves with the following dynamics for t ∈ [tEST
i , ti ]:
ẋi = fi (t, xi , ui ),
0
xi (tLST
i ) = xi
(1)
where x0i represents the initial condition of vehicle i, and
ui (·) represents the control function of vehicle i. In general,
fi (·, ·, ·) depends on the specific dynamic model of vehicle i,
and need not be of the same form across different vehicles.
Denote pi ∈ Rp the subset of the states that represent the
position of the vehicle. Given p0i ∈ Ωf ree , we define the
admissible control function set for Pi to be the set of all
control functions such that pi (t) ∈ Ωf ree ∀t ≥ tLST
i . Denote
P
the joint state space of all vehicles x ∈ Rn where n = i ni ,
and their joint control u.
We assume that the control functions ui (·) are drawn from
STA
1
the set Ui := {ui : [tEST
i , ti ] → Ui , measurable } where
nu
Ui ∈ R i is the set of allowed control inputs. Furthermore,
we assume fi (t, xi , ui ) is bounded, Lipschitz continuous
in xi for any fixed t, ui , and measurable in t, ui for each
1 A function f : X → Y between two measurable spaces (X, Σ )
X
and (Y, ΣY ) is said to be measurable if the preimage of a measurable set
−1
in Y is a measurable set in X, that is: ∀V ∈ ΣY , f (V ) ∈ ΣX , with
ΣX , ΣY σ-algebras on X,Y .
𝜕Ω
𝑃3
𝒯2
Ω
Ω𝑜𝑏𝑠
𝒯1
𝑅𝐶
Obstacle
Targets
Vehicle
Danger zone
𝑃1
𝒯3
𝑃2
Fig. 1: An illustration of the problem formulation with three
vehicles. Each vehicle Pi seeks to reach its target set Ti by
time t = tSTA
i , while avoiding physical obstacles Ωobs and
the danger zones of other vehicles.
xi . Therefore given any initial state x0i and any control
function ui (·), there exists a unique, continuous trajectory
xi (·) solving (1) [21].
The goal of each vehicle i is to arrive at Ti ⊂ Rni
at or before some scheduled time of arrival (STA) tSTA
in
i
minimum time, while avoiding obstacles and danger with all
other vehicles. The target sets Ti can be used to represent
desired kinematic quantities such as position and velocity
and, in the case of non-holonomic systems, quantities such
as heading angle. tEST
can be interpreted as the earliest start
i
time (EST) of vehicle i, before which the vehicle may not
depart from its initial state. Further, we define tLST
i , the latest
(acceptable) start time (LST) for vehicle i. Our problem can
now be thought of as determining the LST tLST
for each
i
vehicle to get to Ti at or before the STA tSTA
i , and finding
a control to do this safely. If the LST is before the EST
tLST
< tEST
i
i , then it is infeasible for vehicle i to arrive at
LST
Ti at or before the STA tSTA
and tEST
is
i . Comparing ti
i
feasibility problem that may arise in practice; however, for
simplicity of presentation, we will assume that tEST
≤ tLST
i
i ∀i.
Danger is described by sets Dij (xj ) ⊂ Ω. In general, the
definition of Dij depends on the conditions under which vehicles i and j are considered to be in an unsafe configuration,
given the state of vehicle j. Here, we define danger to be
the situation in which the two vehicles come within a certain
radius RC of each other: Dij (xj ) = {xi : kpi − pj k2 ≤
RC }. Such a danger zone is also used by the FAA [22]. An
illustration of the problem setup is shown in Figure 1.
In general, the above problem must be analyzed in the joint
state space of all vehicles, making the solution intractable.
In this paper, we will instead consider the problem of
performing path planning of the vehicles in a sequential
manner. Without loss of generality, we consider the problem
of first fixing i = 1 and determining the optimal control for
vehicle 1, the vehicle with the highest priority. The resulting
optimal control u1 sends vehicle 1 to T1 in minimum time.
Then, we plan the minimum time trajectory for each of
the vehicles 2, . . . , N , in decreasing order of priority, given
the previously-determined trajectories for higher-priority vehicles 1, . . . , i − 1. We assume that all vehicles have complete information about the states and trajectories of higherpriority vehicles, and that all vehicles adhere to their planned
trajectories. Thus, in planning its trajectory, vehicle i treats
higher-priority vehicles as known time-varying obstacles.
With the above sequential path planning (SPP) protocol
and assumptions, our problem now reduces to the following
for vehicle i. Given xj (·), j = 1, . . . , i − 1, determine ui (·)
that maximizes tLST
and such that xi (τ ) ∈ Ti , τ ≤ tSTA
i
i .
III. S OLUTION VIA DOUBLE - OBSTACLE HJI VI AND SPP
One direct way of solving the problem formulated in Section II is by solving a single-obstacle HJI VI [10], [11], [23],
[24]. In this approach, one considers the joint time-invariant
dynamics of the entire system, f (x, u), and defines the static
goal set and the static avoid set in the joint state space of all
vehicles. The goal set encodes the joint states representing all
vehicles being at their target sets, and the avoid set encodes
the joint states representing all unsafe configurations. These
sets are defined as sub-zero level sets of appropriate implicit
surface functions s(x) where x ∈ S ⇔ s(x) ≤ 0. Having
defined the implicit surface functions, the HJI VI (2) is then
solved backwards in time with the implicit surface function
representing the terminal set l(x) as the initial condition and
the implicit surface function representing the avoid set a(x)
as an effective constraint:
max Dt V + min [0, H (x, Dx V )] , −a(x) − V (x, t) = 0,
V (x, 0) = l(x)
(2)
with the optimal Hamiltonian H (x, p) = minu∈U p·f (x, u).
The solution V (x, t) is the implicit surface function representing the reach-avoid set RA(t), which defines the set
of states from which the system has a control to drive the
state at time t to the goal set L at time 0 while staying out
of the avoid set A at all times. Note that the joint dynamics,
goal set, and avoid set must be time-invariant. Time-varying
dynamics and sets can be treated by augmenting the state
space with time as an auxiliary state [19]; however, this state
augmentation comes at a large computational expense.
The direct solution described above has been successfully
used to solve a number of problems involving up to a pair
of vehicles [10], [16], [17], [25]. However, since numerical
methods for solving a PDE or a VI involve gridding up the
state space, the computation complexity scales exponentially
with the number of dimensions in the joint state. This
makes the single-obstacle HJI VI inapplicable for problems
involving three or more vehicles. Therefore, instead of
solving aPsingle-obstacle HJI VI in the joint state space in
Rn = R i ni , we will consider the problem in in Rni and
solve a sequence of double-obstacle HJI VIs introduced in
[20]. By doing so, we take advantage of the fact that timevarying targets, obstacles, and dynamics can be handled by
the double-obstacle HJI VIs (but not by the single-obstacle
HJI VI without incurring significant computational expense),
making the analysis of the problem tractable. Furthermore,
even if the dimensionality of the problem is sufficiently low
for computing a numerical solution to the single-obstacle HJI
VI, its inability to handle time-varying systems would still
limit us to only consider problems in which the required time
of arrival is common across all vehicles: tSTA
= tSTA ∀i.
i
We first describe the framework for computing reach-avoid
sets with arbitrary terrain, domain, moving obstacles, and
moving target sets based on [20]. As with the single-obstacle
HJI VI, sets are defined as sub-zero level sets of implicit
surface functions; however, crucially, these implicit surface
functions can be time-varying in the double-obstacle HJI VI
without increasing computational complexity. Being able to
compute reach-avoid sets with moving obstacles allows us to
overcome the computational intractability described above
by sequentially performing path planning for one vehicle
at a time in order of priority, while treating higher-priority
vehicles as moving obstacles. The target set is defined in the
same way as in the single-obstacle HJI VI; the avoid set is by
convention defined as the complement of the state constraint
set in the double-obstacle HJI VI.
A. Reachability via HJI VI
We first state the result given in [20], and then specialize
the result to the problem formulation given in Section II.
Consider a general nonlinear system describing the state
evolution of two players in a differential game for t ∈ [0, T ].
ẋ(t) = f (t, x, u, d),
x(0) = x
(3)
where x is the joint state, u is the control input for player 1,
and d is the control input for player 2. Their joint dynamics f
is assumed to be bounded, Lipschitz continuous in x for any
fixed u, d and t, and measurable in t, u, d for each x. Given
control functions u(·), d(·), there exists a unique trajectory
φu,d
x ((τ ), τ ) [21]. Player 1 wishes to minimize, and player 2
wishes to maximize the following cost functional:
V t, x, u(·), d(·)
u,d
= min max l(φu,d
x(0) (τ ), τ ), max g(φx(0) (s), s)
τ ∈[t,T ]
s∈[t,τ ]
(4)
The value of the game is thus given by
V (x, t) := sup inf V t, x, u(·), δ[u](·)
δ[u](·) u(·)
(5)
where player 2 chooses a nonanticipative strategy d(·) =
δ[u](·), under which the control signal d(t) is chosen in
response to player 1’s control function up to time t, u(τ ), τ ≤
t [18]. The value of the game characterizes reach-avoid set,
or all the states from which player 1 can reach the target L
encoded by the implicit surface function l(x, t), while staying
within some state constraint set G encoded by the implicit
surface function g(x, t), despite the adversarial actions of
player 2. The value function is the unique viscosity solution
[26] to the following single-obstacle HJI VI [20]:
max
n
min Dt V + H (x, Dx V, t) , l(x, t) − V (x, t)
o
g(x, t) − V (x, t) = 0, t ∈ [0, T ], x ∈ Rn (6)
V (x, T ) = max l(x, T ), g(x, T ) , x ∈ Rn
The proof is given in [20] and is based on viscosity
solution theory [27], [28].
Now consider the system with dynamics given by (1).
Given a time-varying target set Ti (t) and obstacle Ai (t) that
vehicle i must avoid, we define implicit surface functions
l(xi , t), g(xi , t) such that xi ∈ Ti (t) ⇔ li (xi , t) ≤ 0, xi ∈
/
Ai (t) ⇔ gi (x, t) ≤ 0. Now, the problem formulated in
Section II becomes one in which vehicle i chooses a control
function ui (·) to minimize the following cost functional:
Vi t, xi , ui (·)
= min max li (xi (τ ), τ ), max gi (xi (s), s)
τ ∈[t,T ]
(7)
s∈[t,τ ]
ui (t) = arg min Hi (t, Dxi V (xi , t), V (xi , t))
Note here, we have an optimal control problem involving
only one vehicle and no adversary (given gi (xi (s), s)),
unlike in the case of the HJI VI (6). Now, specializing
(6) to our optimal control problem, the value function that
characterizes the reach-avoid set RAi (t) is Vi (xi , t), where
xi ∈ RAi (t) ⇔ Vi (xi , t) ≤ 0. Vi (xi , t) is the viscosity
solution [26] of the HJI VI
max min{Dt Vi + Hi (xi , Dxi Vi , t) , li (xi , t) − Vi (xi , t)}
STA
ni
gi (xi , t) − Vi (xi , t) = 0, t ∈ [tEST
i , ti ], xi ∈ R
STA
STA
ni
Vi (xi , tSTA
i ) = max li (xi , ti ), gi (xi , ti ) , xi ∈ R
(8)
where the Hamiltonian Hi (t, xi , p) and optimal control ui
are given by
Hi (t, xi , p) = min p · fi (t, xi , ui )
ui ∈Ui
(9)
u∗i = arg min Hi (t, xi , p)
ui
B. Sequential Path Planning
In order to use (8) to perform SPP, we first define
the moving obstacles induced by higher-priority vehicles.
Specifically, for vehicle i, we define the moving obstacles
Oji (t) induced by vehicles j = 1, . . . , i−1, given their known
trajectories xj (·), to be Oji (t) := {xi : pi ∈ Dij (xj (t)).
Each vehicle i must avoid being in Oji (t) for each j =
1, . . . , i − 1 and for all time t, as well as avoid being in
static obstacles Ωobs in the domain. Therefore, for the ith
vehicle, we compute the reach-avoid set with the following
time-varying avoid set Ai (t) and goal set Li (t):
Ai (t) := {xi : pi ∈ Ωobs } ∪
[
j=1,...,i−1
Oji (t)
constraint set in the HJI VI is defined as the complement
of the avoid set, Aci (t), and is represented by the implicit
surface function g(xi , t), where g(xi , t) ≤ 0 ⇔ xi ∈
/ Ai (t).
For both li (xi , t) and g(xi , t), we use the signed distance
function (in xi ) to the sets Li (t) and Aci (t), respectively.
Now, we can solve the double-obstacle HJI VI (8). The
solution V (xi , t) represents the reach-avoid set RA(t):
V (xi , t) ≤ 0 ⇔ xi (t) ∈ RA(t). RA(t) is the set of states
at starting time t from which vehicle i can arrive at Ti at or
before time tSTA
while avoiding obstacles and danger zones
i
of all higher-priority vehicles j = 1, . . . , i − 1.
Alternatively, given an initial state x0i , we can solve (8)
to some tLST
= inf{t : x0i ∈ RA(t)}. This represents the
i
latest time that vehicle i must depart from its initial position
in order to reach Ti while avoiding obstacles and all danger
zones of higher-priority vehicles j = 1, . . . , i − 1.
The optimal control is given by
(10)
Li (t) := Ti , t ≤ tSTA
i
The goal set is represented by the implicit surface function
li (x, t), where li (xi , t) ≤ 0 ⇔ xi (t) ∈ Li (t). The state
(11)
Observe that since each vehicle i is guaranteed to be safe
with respect to higher priority vehicles j = 1, . . . , i − 1, the
safety of all vehicles, including lower-priority vehicles, can
also be guaranteed.
IV. R ESULTS : F OUR V EHICLES WITH C ONSTRAINED
T URN R ATE
Consider four vehicles with states xi = [xi , yi , θi ]> modeled using a horizontal kinematics model with the following
STA
dynamics for t ∈ [tEST
i , ti ], i = 1, 2, 3, 4:
ẋi = vi cos(θi )
0
xi (tEST
i ) = xi
ẏi = vi sin(θi )
(12)
|ωi | ≤ ω̄i
θ̇i = ωi
where (xi , yi ) is the position of vehicle i, θi is the heading
of vehicle i, and vi is the speed of vehicle i. The control
input ui of vehicle i is the turning rate ωi , whose absolute
value is bounded by ω̄i . For illustration, we chose ω̄i = 1∀i
and assume vi = 1 is constant; however, our method can
easily handle the case in which ω̄i differ across vehicles and
vi is a control input. Optimizing the Hamiltonian associated
with vehicle i, Hi (t, Dxi Vi (xi , t), Vi (xi , t)), we can obtain
the optimal control
ωi (t) = −ω̄i
Dθi Vi (xi , t)
|Dθi Vi (xi , t)|
(13)
The vehicles have initial conditions and STA as follows:
x01 = (−0.5, 0, 0),
tSTA
=0
1
x02
x03
x04
= (0.5, 0, π),
tSTA
= 0.2
2
= (−0.6, 0.6, 7π/4) ,
tSTA
= 0.4
3
= (0.6, 0.6, 5π/4) ,
tSTA
= 0.6
4
(14)
The target sets Ti of the vehicles are all 4 circles of
radius 0.1 in the domain. The centers of the target sets are at
(0.7, 0.2), (−0.7, 0.2), (0.7, −0.7), (−0.7, −0.7) for vehicles
i = 1, 2, 3, 4, respectively. The obstacles are rectangles near
Initial Setup
Obstacle
Target 1
Target 2
Target 3
Target 4
Position/Heading 1
Danger zone 1
Position/Heading 2
Danger zone 2
Position/Heading 3
Danger zone 3
Position/Heading 4
Danger zone 4
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-0.5
0
0.5
Fig. 2: Initial configuration of the four-vehicle example.
the middle of the domain. The setup for this example is
shown in Figure 2.
The joint state space of this system is twelve-dimensional,
intractable for analysis using the single-obstacle HJI VI (2).
Therefore, we will repeatedly solve the double-obstacle HJI
VI (8) to compute the reach-avoid sets from targets Ti
for vehicles 1, 2, 3, 4, in that order, with moving obstacles
induced by vehicles j = 1, . . . , i − 1. We will also obtain
tLST
i , i = 1, 2, 3, 4, the LSTs for each vehicle in order to
reach Ti by tSTA
i .
Figures 3, 4, and 5 show the results. Since the state space
of each vehicle is 3D, the reach-avoid set is also 3D. To
visualize the results, we slice the reach-avoid sets at the
initial heading angles θi0 . Figure 3 shows the 2D reach-avoid
=
= −1.12, tLST
set slices for each vehicle at its LSTs tLST
2
1
LST
LST
−0.94, t3 = −1.48, t4 = −1.44 determined from our
method. The obstacles in the domain Ωobs and the obstacles
induced by other vehicles inhibit the evolution of the reachavoid sets, carving out thin “channels” that separate the
reach-avoid set into different “islands”. One can see how
these channels and islands form by examining the time
evolution of the reach-avoid set, shown in Figure 4 for
vehicle 3.
Finally, Figure 5 shows the resulting trajectories of the
four vehicles. The subplot labeled t = −0.55 shows all four
vehicles in close proximity without collision: each vehicle is
outside of the danger zone of all other vehicles. The actual
arrival times of vehicles i = 1, 2, 3, 4 are 0, 0.19, 0.34, 0.31,
respectively. It is interesting to note that for some vehicles,
the actual arrival times are earlier than the STAs tSTA
i ,i =
1, 2, 3, 4. This is because in order to arrive at the target by
tSTA
i , these vehicles must depart early enough to avoid major
delays resulting from the induced obstacles of other vehicles;
these delays would have lead to a late arrival if vehicle i
departed after tLST
i .
V. C ONCLUSION
We have presented a problem formulation that allows us
to consider the multi-vehicle trajectory planning problem in
a tractable way by planning trajectories for vehicles in order
of priority. In order to do this, we modeled higher-priority
vehicles as time-varying obstacles. We then solved a doubleobstacle HJI VI to obtain the reach-avoid set for each vehicle.
The reach-avoid set characterizes the region from which each
1
Vehicle 1, t=t i1 =-1.12
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
1
0
1
Vehicle 3, t=t i3 =-1.48
-1
-1
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
0
1
-1
-1
Vehicle 2, t=t i2 =-0.94
0
1
Vehicle 4, t=t i4 =-1.44
0
1
Obstacle
Targets
Initial pos. and heading
Reach-avoid set
Danger zones
Fig. 3: Reach-avoid sets at t = tLST
for vehicles 1, 2, 3, 4,
i
sliced at initial headings θi0 . Black arrows indicate direction
of obstacle motion. Due to the turn rate constraint, the
presence of static obstacles Ωobs and time-varying obstacles
induced by higher-priority vehicles Oji (t) carves “channels”
in the reach-avoid set, dividing it up into multiple “islands”.
vehicle is guaranteed to arrive at its target within a time
horizon, while avoiding collision with obstacles and higherpriority vehicles. The solution also gives each vehicle a latest
start time as well as the optimal control which guarantees that
each vehicle safely reaches its target on time.
R EFERENCES
[1] Amazon.com, Inc. (2014) Amazon prime air. [Online]. Available:
http://www.amazon.com/b?node=8037720011
[2] J. Stewart. (2014) Google tests drone deliveries in Project Wing trials.
[Online]. Available: http://www.bbc.com/news/technology-28964260
[3] Jointed Planning and Development Office (JPDO), “Unmanned aircraft
systems (UAS) comprehensive plan – a report on the nation’s UAS
path forward,” Federal Aviation Administration, Tech. Rep., Sep 2013.
[4] W. M. Debusk, “Unmanned aerial vehicle systems for disaster relief:
Tornado alley,” in Infotech@Aerospace Conferences, 2010.
[5] G. C. Chasparis and J. Shamma, “Linear-programming-based multivehicle path planning with adversaries,” in Proceedings of American
Control Conference, June 2005.
[6] P. Fiorini and Z. Shillert, “Motion planning in dynamic environments
using velocity obstacles,” International Journal of Robotics Research,
vol. 17, pp. 760–772, 1998.
[7] J. van den Berg, M. Lin, and D. Manocha, “Reciprocal velocity obstacles for real-time multi-agent navigation,” in Robotics and Automation,
2008. ICRA 2008. IEEE International Conference on, May 2008, pp.
1928–1935.
[8] R. Olfati-Saber and R. M. Murray, “Distributed cooperative control of
multiple vehicle formations using structural potential functions,” in in
IFAC World Congress, 2002.
1
t=0.40
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
1
0
1
t=-0.88
-1
-1
1
0.5
0.5
0
0
-0.5
-0.5
-1
-1
0
1
-1
-1
t=-0.27
0
t=-1.15
0.5
0.5
0
0
-0.5
-0.5
1
-0.5
0
0.5
-0.5
t=0
t=-1.48
0
t=-0.55
1
Obstacle
Targets
Initial pos. and heading
Reach-avoid set
Danger zones
Fig. 4: Time evolution of the reach-avoid set for vehicle
3, sliced at its initial heading θ30 = 7π
4 . Black arrows
indicate direction of obstacle motion. Initially, the reachavoid set grows unobstructed by obstacles, as shown in
the top subplots. Then, in the bottom subplots, the static
obstacles Ωobs and the induced obstacles of vehicles 1 and
2, O13 , O23 , carve out “channels” in the reach-avoid set.
[9] Y.-L. Chuang, Y. Huang, M. D’Orsogna, and A. Bertozzi, “Multivehicle flocking: Scalability of cooperative control algorithms using
pairwise potentials,” in Robotics and Automation, 2007 IEEE International Conference on, April 2007, pp. 2292–2299.
[10] I. Mitchell, A. Bayen, and C. Tomlin, “A time-dependent HamiltonJacobi formulation of reachable sets for continuous dynamic games,”
IEEE Transactions on Automatic Control, vol. 50, no. 7, July 2005.
[11] O. Bokanowski, N. Forcadel, and H. Zidani, “Reachability and minimal times for state constrained nonlinear problems without any controllability assumption,” SIAM Journal on Control and Optimization,
pp. 1–24, 2010.
[12] E. Barron and H. Ishii, “The Bellman equation for minimizing the
maximum cost,” Nonlinear Analysis: Theory, Methods & Applications,
1989.
[13] J. A. Sethian, “A fast marching level set method for monotonically
advancing fronts,” Proceedings of the National Academy of Sciences,
vol. 93, no. 4, pp. 1591–1595, 1996.
[14] S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit
Surfaces. Springer-Verlag, 2002, ISBN: 978-0-387-95482-0.
[15] I. Mitchell, A Toolbox of Level Set Methods, 2009, http://people.cs.
ubc.ca/∼mitchell/ToolboxLS/index.html.
[16] J. Ding, J. Sprinkle, S. S. Sastry, and C. J. Tomlin, “Reachability
calculations for automated aerial refueling,” in IEEE Conference on
Decision and Control, Cancun, Mexico, 2008.
[17] H. Huang, J. Ding, W. Zhang, and C. Tomlin, “A differential game
approach to planning in adversarial scenarios: A case study on
capture-the-flag,” in Robotics and Automation (ICRA), 2011 IEEE
International Conference on, 2011, pp. 1451–1456.
[18] I. Mitchell, “Application of level set methods to control and reachability problems in continuous and hybrid systems,” Ph.D. dissertation,
Stanford University, 2002.
0.5
0.5
0
0
-0.5
-0.5
-0.5
0
0
0.5
t=0.34
0.5
-0.5
0
0.5
Obstacle
Targets
Positions, Headings
Trajectories
Danger Zones
Fig. 5: The planned trajectories of the four vehicles. In
the left top subplot, only vehicles 3 (green) and 4 (purple)
have started moving, showing tLST
is not common across
i
the vehicles. Right top subplot: all vehicles have come
within very close proximity, but none is in the danger zone
another. Left bottom subplot: vehicle 1 (blue) arrives at T1 at
t = 0. Right bottom subplot: all vehicles have reached their
destination, some ahead of the STA tSTA
i .
[19] O. Bokanowski and H. Zidani, “Minimal time problems with moving
targets and obstacles,” 18th IFAC World Congress, 2011.
[20] J. F. Fisac, M. Chen, C. J. Tomlin, and S. S. Sastry, “Reach-Avoid
Problems with Time-Varying Dynamics, Targets and Constraints,” in
18th International Conference on Hybrid Systems: Computation and
Controls, 2015.
[21] E. A. Coddington and N. Levinson, Theory of ordinary differential
equations. Tata McGraw-Hill Education, 1955.
[22] M. L. C. Mike M. Paglione and H. F. Ryan, “Generic metrics for the
estimation of the prediction accuracy of aircraft to aircraft conflicts by
a strategic conflict probe tool,” Air Traffic Control Quarterly, 1999.
[23] K. Margellos and J. Lygeros, “Hamilton-Jacobi Formulation for
Reach-Avoid Differential Games,” IEEE Transactions on Automatic
Control, vol. 56, no. 8, Aug 2011.
[24] K. Margellos and J. Lygeros, “Toward 4-D Trajectory Management
in Air Traffic Control: A Study Based on Monte Carlo Simulation
and Reachability Analysis,” IEEE Transactions on Control Systems
Technology, vol. 21, no. 5, Sept 2013.
[25] M. Chen, Z. Zhou, and C. Tomlin, “Multiplayer reach-avoid games via
low dimensional solutions and maximum matching,” in Proceedings
of the American Control Conference, 2014.
[26] M. G. Crandall, L. C. Evans, and P. L. Lions, “Some properties of
viscosity solutions of hamilton-jacobi equations,” Transactions of the
American Mathematical Society, vol. 282, no. 2, p. 487, Apr. 1984.
[27] L. C. Evans and P. E. Souganidis, “Differential games and representation formulas for solutions of Hamilton-Jacobi-Isaacs equations,”
Indiana University Mathematics Journal, vol. 33, no. 5, 1984.
[28] E. Barron, “Differential Games with Maximum Cost,” Nonlinear
analysis: Theory, methods & applications, pp. 971–989, 1990.
| 3cs.SY
|
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
arXiv:1510.07804v1 [math.AG] 27 Oct 2015
JAROD ALPER
Abstract. This article contains a slightly expanded version of the lectures
given by the author at the summer school “Algebraic stacks and related
topics” in Mainz, Germany from August 31 to September 4, 2015. The
content of these lectures is purely expository and consists of two main goals.
First, we provide a treatment of Artin’s approximation and algebraization
theorems following the ideas of Conrad and de Jong which rely on a deep
desingularization result due to Néron and Popescu. Second, we prove that
under suitable hypotheses, algebraic stacks are étale locally quotients stacks
in a neighborhood of a point with a linearly reductive stabilizer.
Contents
Introduction
Lecture 1: Artin approximation
1.1. Néron–Popescu desingularization
1.2. Artin approximation
1.3. Alternative formulations of Artin approximation
1.4. A first application of Artin approximation
1.5. Proof of Artin approximation
1.6. Categories fibered in groupoids
Lecture 2: Artin algebraization
2.1. Conrad–de Jong approximation
2.2. Artin algebraization
2.3. Algebraic stacks
2.4. Artin’s axioms
2.5. A more refined version of Artin’s axioms
Lecture 3: The geometry of quotient stacks
3.1. Quotient stacks
3.2. A summary of known results on quotient stacks
3.3. The local quotient structure of algebraic stacks
3.4. Ingredients in the proof of Theorem 3.16
Lecture 4: A Luna étale slice theorem for algebraic stacks and
applications
4.1. Proof of Theorem 3.16
4.2. Equivariant Artin algebraization
4.3. Proof of Theorem 3.18
4.4. Applications
References
1
2
3
3
4
5
6
7
8
12
12
16
18
19
21
25
25
27
28
30
32
32
33
34
35
38
2
ALPER
Introduction
The goal of these lectures is twofold:
(1) Discuss Artin’s approximation and algebraization theorems. We will
in fact prove that both theorems follow from a deep desingularization
theorem due to Néron and Popescu. This approach follows the ideas
of Conrad and de Jong.
(2) Prove that “algebraic stacks with linearly reductive stabilizers at closed
points are étale locally quotient stacks.” See Theorem 3.18 for a precise
statement. This theorem was established by the author, Hall and Rydh
in [AHR15].
These two goals are connected in the sense that Artin’s theorems will be
one of the key ingredients in establishing the main theorem expressed in the
second goal. In fact, the proof of Theorem 3.18 will rely on an equivariant
generalization of Artin’s algebraization theorem. Perhaps more importantly
though, both these goals shed light on the local structure of algebraic stacks.
Artin’s approximation and algebraization theorems together with Artin’s criterion for algebraicity instruct us on how we should think of the local structure
of algebraic stacks. The main theorem in the second goal yields a more refined
and equivariant understanding of the local structure of algebraic stacks in the
case that the stabilizers of the closed points have linearly reductive stabilizers, a property that is often satisfied for algebraic stacks appearing in moduli
theory.
Acknowledgements. We thank Ariyan Javanpeykar and Ronan Terpereau
for organizing the summer school “Algebraic stacks and related topics” and we
thank the attentive audience of this summer school for providing worthwhile
feedback. In particular, we thank Pieter Belmans for providing a number
of corrections of the first draft of these notes. Finally, we thank Jack Hall
and David Rydh for extensive and useful suggestions on the content of these
lectures.
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
3
Lecture 1: Artin approximation
The goal of this first lecture is to discuss Artin approximation and provide a few applications. While we will not give a complete and self-contained
proof of Artin approximation, we will show how it follows from Néron–Popescu
desingularization, which is a very deep and difficult result.
Throughout these lectures, k will denote an algebraically closed field. Although almost every statement we make can be generalized to an arbitrary
base scheme S with the arguments being essentially identical, it is nevertheless my belief that working over a fixed algebraically closed field k makes the
geometric content of the statements more transparent. Once the geometric
content is digested over the field k, the motivated reader will have no problem
stating and proving the analogous statements over an arbitrary base scheme.
1.1. Néron–Popescu desingularization.
Definition 1.1. A ring homomorphism A → B of noetherian rings is called
geometrically regular if A → B is flat and for every prime ideal p ⊂ A and
every finite field extension k(p) → k ′ (where k(p) = Ap /p), the fiber B ⊗A k ′
is regular.
Remark 1.2. It is important to note that A → B is not assumed to be of finite
type. In the case that A → B is a ring homomorphism (of noetherian rings)
of finite type, then A → B is geometrically regular if and only if A → B is
smooth (i.e. Spec B → Spec A is smooth).
Remark 1.3. It can be shown that it is equivalent to require the fibers B ⊗A k ′
to be regular only for inseparable field extensions k(p) → k ′ . In particular,
in characteristic 0, A → B is geometrically regular if it is flat and for every
prime ideal p ⊂ A, the fiber B ⊗A k(p) is regular.
We will accept the following result as a black box. The proof is difficult.
Black Box 1 (Néron–Popescu desingularization). Let A → B be a ring
homomorphism of noetherian rings. Then A → B is geometrically regular
if and only if B = lim Bλ is a direct limit of smooth A-algebras.
−→
Remark 1.4. This was result was proved by Néron in [Nér64] in the case that A
and B are DVRs and in general by Popescu in [Pop85], [Pop86], [Pop90]. We
recommend [Swa98] and [Stacks, Tag 07GC] for an exposition on this result.
Example 1.5. If l is a field and ls denotes its separable closure, then l → ls is
geometrically regular. Clearly, ls is the direct limit of separable field extensions
l → l′ (i.e. étale and thus smooth l-algebras). If l is a perfect field, then any
field extension l → l′ is geometrically regular—but if l → l′ is not algebraic,
it is not possible to write l′ is a direct limit of étale l-algebras. On the other
hand, if l is a non-perfect field, then l → l is not geometrically regular as the
geometric fiber is non-reduced and thus not regular.
4
ALPER
In order to apply Néron–Popescu desingularization, we will need the following result, which we will also accept as a black box. The proof is substantially
easier than Néron–Popescu’s result but nevertheless requires some effort.
Black Box 2. If S is a scheme of finite type over k and s ∈ S is a k-point,
b S,s is geometrically regular.
then OS,s → O
Remark 1.6. See [EGA, IV.7.4.4] or [Stacks, Tag 07PX] for a proof.
b is
Remark 1.7. A local ring A is called a G-ring if the homomorphism A → A
geometrically regular. We remark that one of the conditions for a scheme S
to be excellent is that every local ring is a G-ring. Any scheme that is finite
type over a field or Z is excellent.
1.2. Artin approximation. Let S be a scheme and consider a contravariant
functor
F : Sch/S → Sets
where Sch/S denotes the category of schemes over S. An important example of
a contravariant functor is the functor representing a scheme: if X is a scheme
over S, then the functor representing X is:
(1.1)
hX : Sch/S → Sets,
(T → S) 7→ HomS (T, X).
We say that F is limit preserving if for every direct limit lim Bλ of OS -algebras
−→
Bλ (i.e. a direct limit of commutative rings Bλ together with morphisms
Spec Bλ → S), the natural map
lim F (Spec Bλ ) → F (Spec lim Bλ )
−→
−→
is bijective. This should be viewed as a finiteness condition on the functor
F . For instance, the functor hX from (1.1) is limit preserving if and only if
X → S is locally of finite presentation (equivalently locally of finite type if S
is noetherian).
Recall that Yoneda’s lemma asserts that for an S-scheme T , there is a
natural bijection between F (T ) and the set Hom(T, F ) of natural transformation of functors (where we are abusing notation for writing T rather than its
representable functor hT ). Moreover, we will consistently abuse notation by
conflating objects ξ ∈ F (T ) and morphisms (i.e. natural transformations of
functors) ξ : T → F .
Theorem 1.8 (Artin approximation). Let S be a scheme of finite type over k
and let
F : Sch/S → Sets
be a limit preserving contravariant functor. Let s ∈ S be a k-point and
b S,s ). For any integer N ≥ 0, there exist an étale morphism
ξb ∈ F (Spec O
(S ′ , s′ ) → (S, s) and an element ξ ′ ∈ F (S ′ ) such that the restrictions of ξb
+1
+1 ∼
and ξ ′ to Spec(OS,s /mN
) are equal (under the identification OS,s /mN
=
s
s
N +1
OS ′ ,s′ /ms′ ).
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
5
Remark 1.9. This was proven in [Art69a, Cor. 2.2] in the more general case
that S is of finite type over a field or an excellent dedekind domain. In fact,
this theorem holds when the base scheme S is excellent and our proof below
works in this generality with only minor modifications.
Remark 1.10. It is not possible in general to find ξ ′ ∈ F (S ′ ) restricting to ξb
or even such that the restrictions of ξ ′ and ξb to Spec OS,s /mn+1
agree for all
s
n ≥ 0. For instance, F could be the functor hA1 representing the affine line
b S,s could be a non-algebraic power series.
A1 and ξb ∈ O
1.3. Alternative formulations of Artin approximation. Consider the
case that S = Spec A is an affine scheme of finite type over k and hX : Sch/S →
Sets is the functor representing the affine scheme X = Spec A[x1 , . . . , xn ]/
(f1 , . . . , fm ) of finite type over S. Restricted to the category of affine schemes
over S (or equivalently A-algebras), the functor is:
hX : AffSch/S → Sets
Spec R 7→ {a = (a1 , . . . , an ) ∈ R⊕n | fi (a) = 0 for all i}
Applying Artin approximation to this functor, we obtain:
Corollary 1.11. Let A be a finitely generated k-algebra and m ⊂ A be a maximal ideal. Let f1 , . . . , fm ∈ A[x1 , . . . , xn ] be polynomials. Let b
a = (b
a1 , . . . , b
an ) ∈
bm be a solution to the equations f1 (x) = · · · = fm (x) = 0. Then for every
A
N ≥ 0, there exist an étale ring homomorphism A → A′ , a maximal ideal
m′ ⊂ A′ over m, and a solution a′ = (a′1 , . . . , a′n ) ∈ A′⊕n to the equations
a mod mN +1 .
f1 (x) = · · · = fm (x) = 0 such that a′ ∼
=b
Remark 1.12. Although this corollary may seem weaker than Artin approximation, it is not hard to see that it in fact directly implies Artin approximation.
b S,s as a direct limit of finite type
Indeed, writing S = Spec A, we may write O
A-algebras and since F is limit preserving, we can find a commutative diagram
b S,s
Spec O
ξb
Spec A[x1 , . . . , xn ]/(f1 , . . . , fm )
ξ
F.
b ⊕n to the
The vertical morphism corresponds to a solution b
a = (b
a1 , . . . , b
an ) ∈ O
S,s
equations f1 (x) = · · · = fm (x) = 0. Applying Corollary 1.11 yields the desired
étale morphism (Spec A′ , s′ ) → (Spec A, s) and a solution a′ = (a′1 , . . . , a′n ) ∈
A′⊕n to the equations f1 (x) = · · · = fm (x) = 0 agreeing with b
a up to order N
(i.e. congruent modulo mN +1 ). This induces a morphism
ξ ′ : Spec A′ → Spec A[x1 , . . . , xn ]/(f1 , . . . , fm ) → F
b S,s → F to order N.
which agrees with ξb: Spec O
6
ALPER
Alternatively, we can state Corollary 1.11 using henselian rings. Recall that
a local ring (A, m) is called henselian if the following analogue of the implicit
function theorem holds: if f1 , . . . , fn ∈ A[x1 , . . . , xn ] and a = (a1 , . . . , an ) ∈
(A/m)⊕n isa solution to the equations f1 (x) = · · · = fn (x) = 0 modulo m and
∂fi
(a) i,j=1,...,n 6= 0, then there exists a solution a = (a1 , . . . , an ) ∈ A⊕n
det ∂x
j
to the equations f1 (x) = · · · = fn (x) = 0. Equivalently, if (A, m) is a local
k-algebra with A/m ∼
= k, then (A, m) is henselian if every étale homomorphism
′
′
(A, m) → (A , m ) of local rings with A/m ∼
= A′ /m′ is an isomorphism. Also,
if S is a scheme and s ∈ S is a point, one defines the henselization OhS,s of S
at s to be
Γ(S ′ , OS ′ )
OhS,s =
lim
−
→
′ ′
(S ,s )→(S,s)
where the direct limit is over all étale morphisms (S ′ , s′ ) → (S, s). In other
words, OhS,s is the local ring of S at s in the étale topology.
Corollary 1.13. Let (A, m) be a local henselian ring which is the henselization
of a scheme of finite type over k at a k-point.1 Let f1 , . . . , fm ∈ A[x1 , . . . , xn ].
b⊕n is a solution to the equations f1 (x) =
Suppose that b
a = (b
a1 , . . . , b
an ) ∈ A
· · · = fm (x) = 0. For any integer N ≥ 0, there exists a solution a =
(a1 , . . . , an ) ∈ A⊕n to the equations f1 (x) = · · · = fm (x) = 0 such that b
a∼
=a
N +1
mod m
.
1.4. A first application of Artin approximation. The next corollary
states an important fact which you may have taken for granted: if two schemes
are formally isomorphic, then they are isomorphic in the étale topology. First,
we recall that if X → Y is a morphism of schemes of finite type over k and
x ∈ X is a k-point, then X → Y is étale at x if and only if the induced
b Y,f (x) → O
b X,x of completions is an isomorphism.
homomorphism O
Corollary 1.14. Let X1 , X2 be schemes of finite type over k. Suppose x1 ∈
b X ,x ∼
b
X1 , x2 ∈ X2 are k-points such that O
Then there exists a
1 1 = OX2 ,x2 .
common étale neighborhood
(X ′ , x′ )
(1.2)
(X1 , x1 )
(X2 , x2 ) .
Proof. The functor
F : Sch/X1 → Sets,
(T → X1 ) 7→ Hom(T, X2 )
is limit preserving as it can be identified with the representable functor
HomX1 (−, X2 ×X1 ) corresponding to the finite type morphism X2 ×X1 → X1 .
b X ,x ∼
b
b
The isomorphism O
1 1 = OX2 ,x2 provides an element of F (Spec OX1 ,x1 ). By
1
More generally, this statement is true for any local henselian G-ring (see Remark 1.7)
and, in particular, any local henselian excellent ring.
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
7
applying Artin approximation with N = 1, we obtain a diagram as in (1.2)
with X ′ → X1 étale at x′ and such that OX2 ,x2 /m2x2 → OX ′ ,x′ /m2x′ is an isob X ,x , we can conclude
b X ′ ,x′ is abstractly isomorphic to O
morphism. But as O
2 2
b X ,x → O
b X ′ ,x′ induced by X ′ → X2 is an isomorthat the homomorphism O
2 2
phism.2
1.5. Proof of Artin approximation.
Proof of Artin approximation (Theorem 1.8). By Black Box 2, the morphism
b S,s is geometrically regular. By Black Box 1 (Néron–Popescu desinOS,s → O
b S,s = lim Bλ is a direct limit of smooth OS,s -algebras. Since
gularization), O
−→
b S,s and an
F is limit preserving, there exist λ, a factorization OS,s → Bλ → O
b S,s ) is ξ.
b
element ξλ ∈ F (Spec Bλ ) whose restriction to F (Spec O
Let B = Bλ and ξ = ξλ . Geometrically, we have a commutative diagram
ξb
b S,s
Spec O
g
Spec B
ξ
F
Spec OS,s
where Spec B → Spec OS,s is smooth. We claim that we can find a commutative diagram
S′
Spec B
(1.3)
Spec OS,s
where S ′ ֒→ Spec B is a closed immersion, (S ′ , s′ ) → (Spec OS,s , s) is étale, and
+1
the composition Spec OS,s /mN
→ S ′ → Spec B agrees with the restriction
s
b S,s → Spec B.3
of g : Spec O
To see this, observe that the B-module of relative differentials ΩB/OS,s is
locally free. After shrinking Spec B around the image of the closed point under
b S,s → Spec B, we may assume ΩB/O is free with basis db1 , . . . , dbn .
Spec O
S,s
2
The fact that we are using here is that if (A, m) is a local noetherian ring and φ : A → A
is a local homomorphism such that the induced map A/m2 → A/m2 is an isomorphism, then
φ is an isomorphism. Indeed, one can use Nakayama’s lemma to show that the inclusion
φ(m) ⊂ m is also surjective and then Nakayama’s lemma again to show that φ : A → A
is surjective. (See also Lemma 2.15 or [Har77, Lem. II.7.4].) Finally, we use the fact a
surjective endomorphism of noetherian rings is necessarily an isomorphism.
3
This is where the approximation occurs. It is not possible to find a morphism S ′ →
b S,s →
Spec B → Spec OS,s which is étale at a point s′ over s such that the composition Spec O
′
S → Spec B is equal to g.
8
ALPER
This induces a homomorphism OS,s [x1 , . . . , xn ] → B defined by xi 7→ bi and
provides a factorization
AnOS,s
Spec B
Spec OS,s
where Spec B → AnOS,s is étale. We may choose a lift of the composition
b S,s → OS,s /mN +1
OS,s [x1 , . . . , xn ] → B → O
s
to a morphism OS,s [x1 , . . . , xn ] → OS,s . This gives a section s : Spec OS,s →
AnOS,s and we define S ′ as the fibered product
S′
Spec OS,s
Spec B
s
AnOS,s .
ξ
This gives the desired Diagram 1.3. The composition ξ ′ : S ′ → Spec B −
→ F is
b
an element which agrees with ξ up to order N.
By “standard direct limit” methods, we may “smear out” the étale morphism (S ′ , s′ ) → (Spec OS,s , s) and the element ξ ′ : S ′ → F to find an étale
morphism (S ′′ , s′′ ) → (S, s) and an element ξ ′′ : S ′′ → F agreeing with ξb
up to order N. Since this may not be standard for everyone, we spell out
the details. Let Spec A ⊂ S be an open affine containing s. We may write
′
A , we
S ′ = Spec A′ and A′ = OS,s [y1 , . . . , yn ]/(f1′ , . . . , fm
). As OS,s = limg∈m
−→ / s g
′′
′′
can find an element g ∈
/ ms and elements f1 , . . . , fm ∈ Ag [y1 , . . . , yn ] restrict′
′′
ing to f1′ , . . . , fm
. Let S ′′ = Spec Ag [y1 , . . . , yn ]/(f1′′ , . . . , fm
) and s′′ ∈ S ′′
be the image of s′ under S ′ → S ′′ . Then S ′′ → S is étale at s′′ . As
′
A [y , . . . , yn ]/(f1′ , . . . , fm
) and F is limit preserving, we can,
A′ = limh∈m
−→ / s hg 1
after replacing g with hg, find an element ξ ′′ ∈ F (S ′′ ) restricting to ξ ′ and, in
particular, agreeing with ξb up to order N. Finally, we shrink S ′′ around s′′ so
that S ′′ → S is étale everywhere.
1.6. Categories fibered in groupoids. We now shift in the direction of
stacks by introducing categories fibered in groupoids.
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
9
Let S be a scheme. Let X be a category and p : X → Sch/S be a functor.
We visualize this data as
α
ξ1
X
ξ2
β
η2
p
f
T1
Sch/S
T2
where the greek letters ξ1 and ξ2 , η2 are objects in X over the S-schemes T1
and T2 , respectively, and the morphisms α, β are over f .
The main motivation for introducing categories fibered in groupoids rather
than simply working with functors as above is to have a nice framework to handle moduli spaces parameterizing objects with automorphisms (see Example
1.18).
Definition 1.15. A category fibered in groupoids over S is a category X together with a functor p : X → Sch/S such that:
(1) (“existence of pullbacks”) For every morphism of S-schemes T1 → T2
and object ξ2 of X over T2 , there exists an object ξ1 over T1 completing
the diagram
ξ1
T1
α
f
ξ2
T2
where the morphism α is over f (i.e. p(α) = f ).
(2) (“uniqueness and composition of pullbacks”) For all diagrams
ξ1
ξ2
ξ3
T2
T3
T1
there exists a unique arrow ξ1 → ξ2 over T1 → T2 filling in the diagram.
We will often simply write X for a category fibered in groupoids over S where
it is implicitly understood that part of the data is the functor p : X → Sch/S.
10
ALPER
Remark 1.16. Axiom (2) above implies that the pullback in Axiom (1) is
unique up to unique isomorphism. Often we write ξ2 |T1 to indicate a choice of
pullback of ξ2 under T1 → T2 .
If X is a category fibered in groupoids over Sch/S and T is an S-scheme,
we define the fiber category over T , denoted by X(T ), as the category whose
objects are objects of X over T and whose morphisms are morphisms of X over
the identity morphism idT : T → T . Axioms (1) and (2) imply that X(T ) is a
groupoid, i.e. all morphisms in X(T ) are isomorphisms.
Example 1.17. (Functors as categories fibered in groupoids)
A contravariant functor F : Sch/S → Sets may be viewed as a category
fibered in groupoids over Sch/S as follows. Let XF be the category whose
objects are pairs (T, ξ), where T is an S-scheme and ξ ∈ F (T ). A morphism
(T, ξ) → (T ′ , ξ ′ ) in the category XF is given by a morphism f : T → T ′ such
that the pullback of ξ ′ via f is ξ (i.e. F (f )(ξ ′) = ξ). The functor XF → Sch/S
takes (T, ξ) to T and a morphism (T, ξ) → (T ′ , ξ ′) to T → T ′ . It is easy to see
that XF is a category fibered in groupoids over S.
In this case, pullbacks (as defined in Definition 1.15(1)) are unique and each
fiber category XF (T ) is a setoid, i.e. a set with identity morphisms.
Example 1.18. (The moduli space Mg of smooth curves)
We define Mg as the category of pairs (T, C → T ) where T is a scheme over
k and C → T is a smooth proper morphism such that every geometric fiber is
a connected genus g curve. A morphism (T, C → T ) → (T ′ , C′ → T ′ ) is the
data of a cartesian diagram
C
C′
T.
T
The functor Mg → Sch/S takes (T, C → T ) to T and a diagram as above
to the morphism T → T ′ . It is easy to see that Mg is a category fibered in
groupoids over k. If C is a smooth curve over k, note that the morphisms from
C to C in the fiber category Mg (k) correspond to automorphisms of C.
We say that a category fibered in groupoids X over S is limit preserving if
for every direct limit lim Bλ of OS -algebras, the natural functor
−→
lim X(Spec Bλ ) → X(Spec lim Bλ )
−→
−→
is an equivalence of categories.
As before, we have a Yoneda’s lemma which asserts that if T is an S-scheme,
there is a natural equivalence of categories between X(T ) and the category
Hom(T, X) of morphisms T → X of categories fibered in groupoids over S.4
4
Here T is considered as the corresponding category fibered in groupoids (for instance by
first considering the corresponding functor as in (1.1) and then the corresponding category
fibered in groupoids via Example 1.17). A morphism X1 → X2 of categories fibered in
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
11
We now restate Artin’s approximation using categories fibered in groupoids
rather than functors.
Theorem 1.19 (Groupoid version of Artin approximation). Let S be a scheme
of finite type over k and X be a limit preserving category fibered in groupoids
b S,s . For any
over S. Let s ∈ S be a k-point and ξb an object of X over Spec O
′ ′
integer N ≥ 0, there exist an étale morphism (S , s ) → (S, s) and an element
′
b
ξ ′ of X over S ′ such that ξ|
Spec(OS,s /mN+1
) and ξ |Spec(OS ′ ,s′ /mN+1
) are isomorphic
s
s′
N
+1
(under the identification OS,s /mN +1 ∼
= OS ′ ,s′ /m ′ ).
s
s
Proof. The argument for the functorial case of Artin’s approximation goes
through essentially without change. Alternatively, it can be seen to follow
directly from the functorial version by considering the underlying functor
F : Sch/S → Sets where for an S-scheme T , the set F (T ) is the set of isomorphism classes of X(T ).
groupoids over S is a functor X1 → X2 such that the diagram
X1
X2
Sch/S
strictly commutes.
12
ALPER
Lecture 2: Artin algebraization
The goal of this lecture is to prove Artin algebraization. We follow the ideas
of Conrad and de Jong from [CJ02] (see also [Stacks, Tag 07XB]). Namely, we
will show how Artin approximation (Theorem 1.8) implies a stronger approximation result (Theorem 2.1), which we refer to as Conrad–de Jong approximation, and then show how this implies Artin algebraization (Theorem 2.12).
The logical flow of implications is:
Néron–Popescu desingularization
w
w
Artin approximation
w
w
Conrad–de Jong approximation
w
w
Artin algebraization
2.1. Conrad–de Jong approximation. In Artin approximation the initial
b S,s which is
data is an object over a complete local noetherian k-algebra O
assumed to be the completion of a finitely generated k-algebra at a maximal
ideal. We will now see that a similar approximation result still holds if this
latter hypothesis is dropped where one approximates both the complete local
ring and the object over this ring.
Recall first that if (A, m) is a local ring and M is an
L A-module, then the
associated graded module of M is defined as Grm M = n≥0 mn M/mn+1 M.
Theorem 2.1 (Conrad–de Jong approximation). Let X be a limit preserving
category fibered in groupoids over k. Let (R, m) be a complete local noetherian
k-algebra and let ξb be an object of X over Spec R. Then for every integer
N ≥ 0, there exist
(1) an affine scheme Spec A of finite type over k and a k-point u ∈ Spec A,
(2) an object ξA of X over Spec A,
+1
,
(3) an isomorphism αN +1 : R/mN +1 ∼
= A/mN
u
b
(4) an isomorphism of ξ|Spec(R/mN+1 ) and ξA |Spec(A/mN+1
) via αN +1 , and
u
∼
(5) an isomorphism Grm (R) = Grmu (A) of graded k-algebras.
Remark 2.2. This was proven in [CJ02] (see also [Stacks, Tag 07XB]).
The proof of this theorem will proceed by simultaneously approximating
equations and relations defining R and the object ξb of X over Spec R. The
statements (1)–(4) will be easily obtained as a consequence of Artin approximation. A nice insight of Conrad and de Jong is that condition (5) can be
ensured by Artin approximation, and moreover that this condition suffices to
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
13
imply the isomorphism of complete local k-algebras in Artin algebraization.
As such, condition (5) takes the most work to establish.
Proof of Conrad–de Jong approximation (Theorem 2.1). Since X is limit preserving and R is the direct limit over its finitely generated k-subalgebras, we
can find an affine scheme V = Spec B of finite type over k and an object ξV
of X over V together with a 2-commutative diagram
ξb
Spec R
V
ξV
X.
Let v ∈ V be the image of the maximal ideal m ⊂ R. After adding generators
b V,v → R →
to the ring B if necessary, we can assume that the composition O
2
R/m is surjective. A basic fact about complete local noetherian rings (Lemma
b V,v → R is surjective. The goal now is to simultaneously
2.15) implies that O
approximate over V the equations and relations defining the closed immersion
b V,v and the object ξ.
b In order to accomplish this goal, we
Spec R ֒→ Spec O
choose a resolution
(2.4)
b
b ⊕s −β→ O
b ⊕r −αb→ O
b V,v −→ R,
O
V,v
V,v
and consider the functor
F : (Sch/V ) → Sets
α
⊕s β
(T → V ) 7→ complexes O⊕r
T −→ OT −→ OT
which is easily checked to be limit preserving. The resolution in (2.4) yields
b V,v ). If we apply Artin approximation (Theorem 1.8),
an element of F (Spec O
we obtain an étale morphism (V ′ = Spec B ′ , v ′ ) → (V, v) and an element
(2.5)
α′
β′
(B ′⊕r −→ B ′⊕s −→ B ′ ) ∈ F (V ′ )
such that α′ , β ′ are equal to α
b, βb modulo mN +1 .
Let U = Spec A ֒→ Spec B ′ = V ′ be the closed subscheme defined by im β ′
and set u = v ′ ∈ U. We have an induced object
ξ
V
ξA : U ֒→ V ′ → V −→
X
of X over U. As R = coker βb and A = coker β ′ , we have an isomorphism
+1
b Spec(R/mN+1 ) and
together with an isomorphism of ξ|
R/mN +1 ∼
= A/mN
u
ξA |Spec(A/mN+1
) . This gives statements (1)–(4).
u
We are left to establish statement (5). First, we provide some motivation
for the technical definition (Definition 2.3) and lemma (Lemma 2.5) below.
To establish (5), we need to show that there are isomorphisms mn /mn+1 ∼
=
N +1 ∼
mnu /mn+1
.
For
n
≤
N,
this
is
guaranteed
by
the
isomorphism
R/m
=
u
+1
A/mN
.
On
the
other
hand,
for
n
≫
0,
this
can
be
seen
to
be
a
consequence
u
of the Artin–Rees lemma (see the proof of Lemma 2.5(3) below). We will need
14
ALPER
to handle the middle range of n and we accomplish this be controlling the
constant appearing in the Artin–Rees lemma.
We now establish statement (5) using Definition 2.3 and Lemma 2.5. We first
realize that before we applied Artin approximation, we could have increased
b Therefore, we are
N in order to guarantee that (AR)N holds for α
b and β.
free to assume this. Now statement (5) follows directly if we apply Lemma
b
b ⊕r −αb→ O
b ⊕s −β→ O
b V,v from (2.4) and the complex
2.5 to the exact complex O
V,v
α
b′
V,v
βb′
b ⊕r −→ O
b ⊕s −→ O
b V,v obtained by restricting (2.5) to F (Spec O
b V,v ).
O
V,v
V,v
The Artin–Rees condition.
Definition 2.3. Let (A, m) be a local noetherian ring. Let ϕ : M → N be a
morphism of finite A-modules. Let c ≥ 0 be an integer. We say that (AR)c
holds for ϕ if
ϕ(M) ∩ mn N ⊂ ϕ(mn−c M), ∀n ≥ c.
Remark 2.4. The Artin–Rees lemma (see [AM69, Prop. 10.9] or [Eis95, Lem.
5.1]) implies that (AR)c holds for ϕ if c is sufficiently large.
The following lemma from [CJ02, §3] (see also [Stacks, Tags 07VD and
07VF]) is straightforward to prove.
Lemma 2.5. Let (A, m) be a local noetherian ring. Let
α
β
L−
→M −
→N
and
α′
β′
L−
→M −
→N
be two complexes of finite A-modules. Let c be a positive integer. Assume that
(a) the first sequence is exact,
(b) the complexes are isomorphic modulo mc+1 , and
(c) (AR)c holds for α and β.
Then
(1) (AR)c holds for β ′ ,
(2) the second sequence is exact, and
(3) there exists an isomorphism Grm (coker β) → Grm (coker β ′ ) of Grm (A)modules.
Remark 2.6. Only conclusion (3) was used in the proof of Conrad–de Jong
approximation. Statements (1) and (2) are included as they serve as convenient
steps in the proof of (3). Indeed, it is the fact that (AR)c holds for β that
implies the containment “⊂” in (2.6). Likewise, the other containment “⊃”
will hold once we know (AR)c holds for β ′ .
Proof. We claim that (β ′)−1 (mn N) ⊂ α′ (L) + mn−c M for all n ≥ c. Let
m ∈ M such that β ′ (m) ∈ mn N. Suppose we can find l ∈ L such that
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
15
m0 = m − α′ (l) ∈ mr M with r < n − c. We will show that we can modify l in
order to increase r by 1. Note that
β(m0 ) = β ′ (m0 ) + (β − β ′ )(m0 )
= β ′ (m) + (β − β ′ )(m0 )
∈ mn N + mr+c+1 N = mr+c+1 N
Since (AR)c holds for β, we have β(m0 ) ∈ β(mr+1 M). So we can write β(m0 ) =
β(m1 ) with m1 ∈ mr+1 M. As the first complex is exact, we may write m0 −
m1 = α(l1 ) for l1 ∈ L. Note that m0 − m1 ∈ mr M. We now break the
argument into whether or not r ≥ c. If r ≥ c, then as (AR)c holds for α, we
have m0 − m1 ∈ α(mr−c L) and we may replace l1 with an element of mr−c L. In
this case, (α − α′ )(l1 ) ∈ mr−c · mc+1 M = mr+1 M. On the other hand, if r < c,
then we automatically have (α − α′ )(l1 ) ∈ mc+1 M ⊂ mr+1 M. Therefore,
m − α′ (l + l1 ) = m0 − α′ (l1 )
= m1 + (α − α′ )(l1 ) ∈ mr+1 M.
By induction, this establishes the claim.
For (1), the condition that β ′ (M) ∩ mn N ⊂ β ′ (mn−c M) follows directly from
the claim. For (2), observe that the claim coupled with Krull’s intersection
theorem implies that
\
\
(β ′ )−1 (0) = (β ′ )−1 ( mn N) ⊂
α′ (L) + mn−c M) = α′ (L)
n
n
which gives exactness of the second sequence.
For (3), for n ≥ 0, we have
Grm (coker β)n = mn N/(mn+1 N + β(M) ∩ mn N)
and a similar description of Grm (coker β ′ )n . To obtain an isomorphism
Grm (coker β) → Grm (coker β ′ ), it clearly suffices to show that
(2.6)
mn+1 N + β(M) ∩ mn N = mn+1 N + β ′ (M) ∩ mn N.
We first show the containment “⊂”. If n ≤ c, then the statement is clear as
β = β ′ mod mc+1 . If n > c, suppose x ∈ β(M) ∩ mn N. Since (AR)c holds
for β, we may write x = β(m) for m ∈ mn−c M. Let x′ = β ′ (m). Then
x − x′ = (β − β ′ )(m) ∈ mc+1 · mn−c M = mn+1 N. Since (AR)c also holds for β ′ ,
by symmetry we obtain the other containment “⊃”.
Remark 2.7. Conrad–de Jong approximation directly implies Artin approximab S,s and N ≥ 1. Apply Conrad–de Jong
tion (Theorem 1.19). Indeed, let R = O
approximation to X × S to obtain a finite type morphism (U = Spec A, u) →
+1 ∼
+1
(S, s), an object ξA of X over U, an isomorphism αN +1 : OS,s /mN
= OU,u /mN
s
u
b
and an isomorphism ξ|
) via αN +1 . We claim
) and ξA |Spec(OU,u /mN+1
Spec(OS,s /mN+1
u
s
b S,s /m2 → O
b U,u /m2 is an isomorthat U → S is étale at u. Since we know that O
s
u
b S,s → O
b U,u is surjective. But condition
phism, the induced homomorphism O
16
ALPER
(5) in Conrad–de Jong approximation gives isomorphisms
N +1
N +1
mN
→ mN
s /ms
u /mu
b S,s → O
b U,u is injective and thus an isomorfor every N. This implies that O
phism.
2.2. Artin algebraization. Artin algebraization has a stronger conclusion
than Artin approximation or Conrad–de Jong approximation in that no approximation is necessary. It guarantees the existence of an object ξA over a
pointed affine scheme (Spec A, u) of finite type over k which agrees with the
given effective formal deformation ξb to all orders. However, in order to enb This
sure this, it is necessary to impose a further condition on the object ξ.
condition is known as formal versality and is extremely natural (see Remarks
2.9-2.11).
Definition 2.8. Let X be a category fibered in groupoids over k. Let R be
a complete local noetherian k-algebra and x ∈ Spec R be the closed point.
We say that an object ξb of X over Spec R is formally versal at x if for every
commutative diagram
Spec k(x)
Spec B
(2.7)
Spec R
ξb
Spec B ′
X
where B ′ → B is a surjection of artinian k-algebras, there is a lift Spec B ′ →
Spec R filling in the above diagram.
Remark 2.9. In other words, the formal versality of ξb means that whenever
you have a surjection B ′ → B of artinian OS -algebras, an object η ′ of X over
b Spec B ∼
Spec B ′ , a morphism Spec B → Spec R, and an isomorphism α : ξ|
=
′
′
η |Spec B , there exist a morphism Spec B → Spec R and an isomorphism
α′ : ηb|Spec B′ ∼
= η ′ extending α.
b Spec k(x) be the restriction of ξb to the closed point. Then
Let ξ0 = ξ|
η : Spec B → X can be viewed as an infinitesimal deformation of ξ0 and
η ′ : Spec B ′ → X a further infinitesimal deformation of η. In colloquial language, the condition of formal versality implies that the family of objects
ξb: Spec R → X contains all infinitesimal deformations of ξ0 .
Remark 2.10. The lifting criterion in Diagram (2.7) above should remind the
reader of the formal lifting property for smooth morphisms. Indeed, if f : X →
Y is a morphism of finite type k-schemes, then it is a theorem of Grothendieck
[EGA, IV.17.14.2] that f is smooth at x if and only if the above lifting criterion
holds (replacing of course ξb: Spec R → X with f : X → Y ).
Remark 2.11. It is easy to see that the condition of formal versality of
b Spec R/mn+1 }
ξb: Spec R → X only depends on the compatible family {ξn = ξ|
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
17
of restrictions. Therefore, we can extend the above definition to compatible
families {ξn }.
Theorem 2.12 (Artin algebraization). Let X be a limit preserving category
fibered in groupoids over k. Let (R, m) be a complete local noetherian k-algebra,
and let ξb be a formally versal object of X over Spec R. There exist
(1) an affine scheme Spec A of finite type over k and a k-point u ∈ Spec A,
(2) an object ξA of X over Spec A,
∼ b
(3) an isomorphism α : R → A
mu of k-algebras, and
b Spec R/mn+1 ∼
(un(4) a compatible family of isomorphisms ξ|
= ξU |Spec A/mn+1
u
n+1 ∼
n+1
der the identification R/m
= A/mu ) for n ≥ 0.
Remark 2.13. This was first proven in [Art69b].
Remark 2.14. In the case that R is known to be the completion of a finitely
generated k-algebra, this theorem can be viewed as an easy consequence of
Artin approximation. Indeed, one applies Artin approximation with N = 1
and then uses the formal versality condition to obtain compatible ring homobmu which
morphisms R → A/mn+1
and therefore a ring homomorphism R → A
u
bmu are abstractly isomorphic, the
is an isomorphism modulo m2 . As R and A
bmu is an isomorphism (see Footnote 2) and the statehomomorphism R → A
ment follows. We leave the details to the reader but remark that this argument
is very analogous to the proof of Artin algebraization below. For the general
case, since we don’t know R is the completion of a finitely generated k-algebra,
we apply Conrad–de Jong approximation instead of Artin approximation.
Proof of Artin algebraization (Theorem 2.12). Applying Conrad–de Jong approximation with N = 1, we obtain an affine scheme Spec A of finite type over
k with a k-point u ∈ Spec A, an object ξA of X over Spec A, an isomorphism
b Spec(R/m2 ) → ξA |Spec(A/m2 ) ,
α2 : Spec A/m2u → Spec R/m2 , an isomorphism ι2 : ξ|
u
∼
and an isomorphism Grm (R) = Grmu (A) of graded k-algebras.
We claim that α2 and ι2 can be extended to a compatible family of morb
phisms αn : Spec A/mn+1
→ Spec R and isomorphisms ιn : ξ|
u
) →
Spec(A/mn+1
u
n+1 .
ξA |
By induction, suppose we are given αn and ιn . Since ξb is
Spec(A/mu
)
formally versal, there is a lift αn+1 : Spec A/mn+1 → Spec R filling in the
commutative diagram
Spec A/mnu
αn
Spec R
αn+1
Spec A/mn+1
u
ξA
ξb
X,
which establishes the claim.
bmu which is surBy taking the limit, we have a homomorphism α
b: R → A
2
jective as α
b restricts to the given isomorphism R/m → A/m2u (where we have
18
ALPER
used Lemma 2.15). On the other hand, for each n, we know that the k-vector
N +1
spaces mN /mN +1 and mN
have the same dimension. This implies that
u /mu
α
b is an isomorphism which finishes the proof.
In the above arguments, this fact was used several times.
Lemma 2.15. Let (A, mA ) and (B, mB ) be local noetherian complete rings. If
A → B is a local homomorphism such that A/m2A → B/m2B is surjective, then
A → B is surjective.
Proof. This follows from the following version of Nakayama’s lemma for complete localTrings (A, m): if M is a (not-necessarily finitely generated) A-module
such that k mk M = 0 and m1 , . . . , mn ∈ F generate M/mM, then m1 , . . . , mn
also generate M (see [Eis95, Exercise 7.2]).
2.3. Algebraic stacks. We now quickly introduce the notion of stacks and
algebraic stacks.
` We say that {Si → S} is an étale covering if each Si → S is étale and
Si → S is surjective. To simplify the notation, we set Sij := Si ×S Sj and
Sijk := Si ×S Sj ×S Sk .
Definition 2.16. A category X fibered in groupoids over a scheme S is a stack
over S if for any étale covering {Ti → T } of an S-scheme T , we have:
(1) (“morphisms glue”) For objects a, b in X over T and morphisms
φi : a|Ti → b over Ti → T such that φi |Tij = φj |Tij , then there exists a unique morphism φ : a → b over the identity with φ|Ti = φi .
Pictorially, we are requiring that a commutative diagram
a|Ti
Ti
φi
a|Tij
a
φ
b
over
Tij
T
φj
a|Tj
Tj
can be completed in a unique way by an arrow φ : a → b.
(2) (“objects glue”) For objects ai over Ti with isomorphisms αij : ai |Tij →
aj |Tij over idTij satisfying the cocycle condition αij ◦ αjk = αik on Tijk ,
then there exist a unique object a over T and isomorphisms φi : a|Ti →
ai over idTi such that αij ◦ φi |Tij = φj |Tij . Pictorially, we are requiring
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
19
that diagrams
ai
αij
ai |Tij −→ aj |Tij
Ti
a
aj
over
Tij
T
Tj
where the αij satisfy the cocycle condition can be filled in with an
object a over T .
Remark 2.17. These gluing conditions are extremely natural and should already be familiar to you. For instance, rather than the étale topology,
conS
sider the Zariski topology (i.e. only consider Zariski covers S = Si ) and
consider the category fibered in groupoids X over Spec Z parametrizing pairs
(T, F), where T is a scheme and F is a quasi-coherent sheaf on T . A morphism
(T ′ , F ′) → (T, F) in X is the data of a morphism f : T ′ → T together with
∼
an isomorphism F → f ∗ F ′ . Then the fact that X is a stack (in the Zariski
topology) is the basic fact that quasi-coherent sheaves and their morphisms
can be uniquely glued. In this case, X is also a stack in the étale (or even fppf)
topology.
Definition 2.18. Let S be a scheme. A stack X over S is algebraic if
(1) the diagonal ∆ : X → X ×S X is representable, and
(2) there exist a scheme U and a representable, smooth and surjective
morphism U → X.
Remark 2.19. The representability condition in (1) means that for every scheme
T and pair of morphisms ξ, η : T → X, the fiber product X ×∆,X×S X,(ξ,η) T is
an algebraic space. This in turn translates to the condition that the functor
IsomT (ξ, η) : Sch/T → Sets,
(T ′ → T ) 7→ MorX(T ′ ) (ξ|T ′ , η|T ′ )
is representable by an algebraic space, where MorX(T ′ ) (ξ|T ′ , η|T ′ ) denotes the
set of morphisms (which are necessarily isomorphisms) in the fiber category
X(T ′ ) of pullbacks of ξ and η to T ′ .
In (2), the condition that U → X is representable, smooth and surjective
means that for any morphism T → X where T is a scheme, the fiber product
T ×X U is an algebraic space and T ×X U → T is smooth and surjective. (In
fact, condition (2) can be shown to imply condition (1).)
2.4. Artin’s axioms. A spectacular application of Artin’s algebraization theorem is Artin’s local criterion for algebraicity of stacks. This is a foundational
result in the theory of algebraic stacks and was proved by Artin in the very
same paper [Art74] where he introduced algebraic stacks.
We first state a conceptual version of Artin’s axioms that can be proved
easily using only Artin algebraization.
20
ALPER
Theorem 2.20. (Artin’s axioms—first version) Let X be a stack over k. Then
X is an algebraic stack locally of finite type over k if and only if the following
conditions hold:
(0) (Limit preserving) The stack X is limit preserving.
(1) (Representability of the diagonal) The diagonal X → X × X is representable.
(2) (Existence of formal deformations) For every x : Spec k → X, there
exist a complete local noetherian k-algebra (R, m) and a compatible
family of morphism ξn : Spec R/mn+1 → X with x = ξ0 such that {ξn }
is formally versal (as defined in Remark 2.11).
(3) (Effectivity) For every complete local noetherian k-algebra (R, m), the
natural functor
X(Spec R) → lim X(Spec R/mn )
←−
is an equivalence of categories.
(4) (Openness of versality) For any morphism ξU : U → X where U is a
scheme of finite type over k and point u ∈ U such that ξU is formally
b U,u → X is formally
versal at u (i.e. the induced morphism Spec O
versal), then ξU is formally versal in an open neighborhood of u.
Proof. The “only if” direction is fairly straightforward and left to the reader.
For the “if” direction, we first remark that Condition (1) implies that any
morphism U → X from a scheme U is necessarily representable.
Let x : Spec k → X be a morphism. We need to find a commutative diagram
Spec k
u
U
x
X
where U is a scheme and U → X is smooth. Condition (2) and (3) guarantee
that there exists a complete local noetherian k-algebra (R, m) with R/m ∼
=k
together with a commutative diagram
Spec R
Spec k
x
ξb
X
such that ξb is formally versal. By Artin’s algebraization theorem, we can find
an affine scheme U = Spec A of finite type over k, a point u ∈ U, a morphism
b U,u yielding a commutative diagram
ξ : U → X, and an isomorphism R ∼
=O
Spec k
Spec R
U
ξb
x
ξ
X.
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
21
We know that U → X is formally versal at u and condition (4) implies that it
is formally versal in a neighborhood. But this implies that U → X is smooth
in a neighborhood of u.5
Remark 2.21. In practice, Condition (1) is often easy to verify directly. Alternatively one could also apply the theorem to the diagonal X → X × X (i.e.
to the sheaves IsomT (ξ, η) defined in Remark 2.19). Condition (2) is often a
consequence of Schlessinger and Rim’s theorem on existence of formally versal deformations [Sch68], [Rim80]. Condition (3) is often a consequence of
Grothendieck’s existence theorem [EGA, III.5.1.4].
In some simplified moduli problems, Condition (4) can be checked directly.
For instance, if for each point x : Spec k → X, the formal deformation space
(R, m) (as in Condition (2)) is regular (or more generally normal, geometrically
unibranch and free of embedded points), then Condition (4) is automatically
satisfied; see [Art69b, Thm. 3.9]. In more general moduli problems, Condition
(4) is often guaranteed as a consequence of a well-behaved deformation and
obstruction theory. This will be explained in the next section.
2.5. A more refined version of Artin’s axioms. We will now state a
refinement of Theorem 2.20 that is often easier to verify in practice. In order
to state the theorem succinctly, we need to introduce a bit of notation.
Let A be a finitely generated k-algebra and let M be a finite A-module.
Denote by A[M] the ring A ⊕ M defined by M 2 = 0. Let ξ : Spec A → X.
Denote by Def ξ (M) the set of isomorphism classes of diagrams
ξ
Spec A
X
η
Spec A[M].
Let Autξ (M) be the group of automorphisms of the trivial deformation
Spec A[M] → Spec A → X which restrict to the identity automorphism of
ξ.
We remark that Autξ (M) naturally has the structure of an A-module.6
Condition (1a) below implies that Def ξ (M) is naturally an A-module.7
5
This isn’t a trivial implication. The formal versality condition is only a condition on
lifting local artinian k-algebras but a theorem of Grothendieck [EGA, IV.17.14.2] implies
that this is sufficient to guarantee smoothness.
6
Indeed, for a ∈ A, the A-algebra homomorphism A[M ] → A[M ], a0 + m0 7→ a0 + am0
induces a morphism fa : Spec A[M ] → Spec A[M ] over Spec A which in turn induces a group
homomorphism Autξ (M ) → Autξ (M ) obtained by pulling back automorphisms along fa .
7
This may seem surprising since in condition (1a), the k-algebras A, A′ , B are only assumed to be artinian. However, as shown in [HR13], this is strong enough to imply that for
any finitely generated k-algebra A and finite A-module M , the natural map
X(Spec(A[M ⊕ M ])) → X(Spec A[M ]) ×X(Spec A) X(Spec A[M ])
22
ALPER
Note that Autξ (k) is the group of infinitesimal automorphism of ξ, and
Def ξ (k) is the set of isomorphism classes of deformations over the dual numbers
Spec k[ǫ] which can be thought of as the Zariski tangent space.
Theorem 2.22 (Artin’s axioms—refined version). Let X be a stack over k.
Then X is an algebraic stack locally of finite type over k if and only if the
following conditions hold:
(0) (Limit preserving) The stack X is limit preserving.
(1) (Existence of formal deformations)
(a) (Homogeneity) For every diagram
A
A′
B
A′ ×A B
of finitely generated local artinian k-algebras where A → A′ is
surjective, the natural functor
X(Spec(A′ ×A B)) → X(Spec A′ ) ×X(Spec A) X(Spec B)
is an equivalence of categories.
(b) (Finiteness of tangent spaces) For every object ξ : Spec k → X,
Autξ (k) and Def ξ (k) are finite dimensional k-vector spaces.
(2) (Effectivity) For every complete local noetherian k-algebra (R, m), the
natural functor
X(Spec R) → lim X(Spec R/mn )
←−
is an equivalence of categories.
(3) (Openness of versality)
(a) (Coherent deformation theory) For every finitely generated k-algebra
A, finite A-module M and object ξ : Spec A → X, the A-modules
Autξ (M) and Def ξ (M) are finite. Thus there are A-linear functors
Autξ : Finite A-mod → Finite A-mod
Def ξ : Finite A-mod → Finite A-mod.
(b) (Existence of an obstruction theory) For every finitely generated
k-algebra A and object ξ : Spec A → X, there exists an A-linear
functor
Obξ : Finite A-mod → Finite A-mod.
is an equivalence of categories. Therefore, addition M ⊕M → M induces an A-algebra homomorphism A[M ⊕ M ] → A[M ] and thus a functor X(Spec A[M ]) ×X(Spec A) X(Spec A[M ]) ∼
=
X(Spec A[M ⊕ M ]) → X(Spec A[M ]) which defines addition on Def ξ (M ). Multiplication
by an element a ∈ A induces an A-algebra homomorphism A[M ] → A[M ] and therefore a
functor X(Spec A[M ]) → X(Spec A[M ]) which defines multiplication by a ∈ A.
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
23
Moreover, for each surjection A′ → A with squarezero kernel I,
there exists an element oξ (A′ ) ∈ Obξ (I) such that there is an extension
Spec A
ξ
X
Spec A′
if and only if oξ (A′ ) = 0.
(c) (Constructibility) For every finitely generated k-algebra A, finite
A-module M supported on Ared and object ξ : Spec A → X, there
exists a dense open subset U ⊂ Spec A such that for every k-point
u ∈ U, the natural maps
Autξ (M) ⊗A k(u) → Autξ|u (M ⊗A k(u))
Def ξ (M) ⊗A k(u) → Def ξ|u (M ⊗A k(u))
Obξ (M) ⊗A k(u) → Obξ (M ⊗A k(u))
are isomorphisms.
Remark 2.23. An analogous statement is true after replacing Spec k with an
arbitrary excellent scheme after suitably modifying Conditions 1–3.
Artin proved a version of the above theorem in [Art74]. We will not give a
complete proof of this statement here and restrict ourselves to only making a
few comments. First, one reduces to the case that X → X × X is representable
by bootstrapping the below argument using automorphisms and deformations
(rather than deformations and obstructions) to conclude that the isomorphism
sheaves (as defined in Remark 2.19) are representable.
Conditions 1(a)–(b) above allow us to apply Schlessinger and Rim’s theorem
[Sch68], [Rim80] which guarantees the existence of formal deformation, i.e.
Condition 1 in Theorem 2.20 holds. This condition is often fairly easy to
check in practice for moduli problems.
In practice, Condition 2 is often established as a direct consequence of
Grothendieck’s Existence Theorem.
Conditions 3(a)–(c) can be shown to imply that formal versality is an open
condition, i.e. Condition 3 in Theorem 2.20 holds. To this end, it is necessary
to show that if ξ : U → X is a morphism which is formally versal at u ∈ U, then
for points v ∈ U in a sufficiently small neighborhood of u, any commutative
diagram of the form
Spec B
(2.8)
U
ξ
Spec B ′
X,
where B ′ → B is a surjection of artinian k-algebras and Spec B → U maps
to v, has an extension. This is a deformation problem. Vaguely speaking,
24
ALPER
conditions 3(a)–(c) imply that the deformation theory of U → X is controlled
in some sense by a coherent module, and formal versality at u implies that
this coherent module vanishes at u and thus in an open neighborhood.
Condition 3 certainly takes the most time and space to formulate. Moreover
the most difficult part of the proof of Theorem 2.22 is to show that Condition
3 implies openness of versality. Nevertheless it is quite easy to establish Condition 3 in practice. For many moduli problems, one shows that if M is a finite
A-module, then the A-modules Autξ (M), Def ξ (M), and Obξ (M) are naturally
identified with certain cohomology modules in which case Condition 3(c) can
be seen to follow from cohomology and base change. For example, if X is the
moduli space of smooth curves as in Example 1.18 and ξ corresponds to a curve
C → Spec A, then Autξ (M) = H0 (C, TC ⊗ M), Def ξ (M) = H1 (C, TC ⊗ M)
and Obξ (M) = H2 (C, TC ⊗ M) = 0. Here the constructibility condition in
3(c) follows directly from cohomology and base change applied at the generic
points of Spec A.
We recommend [Hal14] for a conceptual proof of Artin’s criterion. There
is some flexibility in how one precisely formulates Conditions 1 and 3. We
recommend [HR13] for a technical account of various formulations of Artin’s
axioms and in particular for a complete proof of the formulation given in
Theorem 2.22.
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
25
Lecture 3: The geometry of quotient stacks
In this lecture, we discuss a particularly important example of an algebraic
stack, namely the quotient stack [X/G], which arises from the action of an
algebraic group G on a scheme X. We will emphasize that the geometry of a
quotient stack [X/G] is nothing other than the G-equivariant geometry of X.
In this lecture, we discuss when a general algebraic stack is a quotient stack
and then turn to the main theorem of these lectures which gives conditions for
an algebraic stack to be étale locally a quotient stack.
3.1. Quotient stacks.
Definition 3.1. Let G be a smooth affine group scheme over k acting on a
scheme X over k. We define the quotient stack of X by G, denoted by [X/G],
g
π
to be the category of objects (T, P −
→ T, P −
→ X) where T is a scheme, P → T
is a principal G-bundle and P → X is a G-equivariant morphism. A morphism
g′
g
π
π′
(T, P −
→ T, P −
→ X) → (T ′ , P ′ −
→ T ′, P ′ −
→ X) in this category consists of a
commutative diagram
g
P′
P
π
T
g′
X
π′
T′
where the square is cartesian.
Remark 3.2. With the above hypotheses, [X/G] is an algebraic stack. Here
Axiom (2) of the definition of an algebraic stack is satisfied by the projection
X → [X/G]. The morphism X → [X/G] corresponds via Yoneda’s lemma
to the object of [X/G] over X defined by the trivial G-bundle G × X → X
together with the G-equivariant morphism σ : G × X → X (corresponding to
the action of G on X). The morphism X → [X/G] is a G-torsor.
Remark 3.3. In fact, in characteristic p, if G is not smooth, it can still be
shown that [X/G] is an algebraic stack.
Definition 3.4. We say that an algebraic stack X is a global quotient stack if
X∼
= [U/ GLn ] where U is an algebraic space with an action of GLn .
Remark 3.5. If G is an affine group scheme of finite type over k and U is
an algebraic space over k with an action of G, then [U/G] is a global quotient stack. To see this, choose a faithful representation G ⊂ GLn . Then
[U/G] ∼
= [(U ×G GLn )/ GLn ] where U ×G GLn = (U × GLn )/G (and here G
acts diagonally on U × GLn ).
Quotient stacks provide very important examples of algebraic stacks as their
geometry is particularly well understood. In fact, the geometry of a quotient
stack X = [X/G] is nothing other than the G-equivariant geometry of X. To
26
ALPER
justify this philosophy, we provide a dictionary between geometric concepts of
[X/G] and G-equivariant geometric concepts of X.
G-equivariant geometry of X
Geometry of X = [X/G]
a point x : Spec k → X
a G-orbit Gx ⊂ X
the automorphism group AutX(k) (x)
the stabilizer Gx
a function f ∈ Γ(X, OX )
a G-invariant function f ∈ Γ(X, OX )G
a morphism X → Y where Y is a a G-invariant morphism X → Y
scheme (or an algebraic space)
a line bundle on X
a line bundle on X with a G-action
a coherent OX -module
a coherent OX -module with a G-action
properties of the diagonal X → X × X
properties of the group action G×X →
X ×X
the tangent space TX,x
the normal
TX,x /TGx,x
space
to
the
orbit
Above, we used the notion of the tangent space of a stack at a point. Since
this will be important later, let’s define it precisely.
Definition 3.6. If X is an algebraic stack of finite type over k and x is a
k-point, then the tangent space of X at x, denoted by TX,x , is defined as the
set of isomorphism classes of extensions
Spec k
x
X
Spec k[ǫ]/ǫ2 .
Remark 3.7. The equivalence TX,x ∼
= TX,x /TGx,x only holds when the stabilizer
Gx is smooth.
Example 3.8. Consider the action of the multiplicative group Gm = Spec k[t]t
on An via multiplication. Then [An /Gm ] is an algebraic stack. The origin is
the only closed k-point of [An /Gm ] as all other k-points have the origin in their
closure. As [(An \ 0)/Gm ] = Pn−1 , there is an open substack Pn−1 ⊂ [An /Gm ].
Example 3.9. Consider the action of Gm on A2 via t·(x, y) = (tx, t−1 y). Then
the closed k-points of [A2 /Gm ] correspond to the origin together with the hyperbolas {xy = a} ⊂ A2 (i.e. the Gm -orbit of (a, 1) for a 6= 0). The two orbits
Gm (1, 0) and Gm (0, 1) both have the origin in their closure. Here the open substack [(A2 \ 0)/Gm ] ⊂ [A2 /Gm ] is isomorphic to the non-separated affine line
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
27
with a double origin. The morphism [A2 /Gm ] → A1 given by (x, y) 7→ xy gives
a bijective correspondence between closed k-points in [A2 /Gm ] and k-points in
A1 . Note that the fiber over the origin consists of 3 points corresponding to
the orbits in the union of the x and y-axes.
A central question is:
Question 3.10. When is an algebraic stack X a global quotient?
This question is very difficult. As we will see below, the question of whether
X is a global quotient is related to other global geometric properties of X.
3.2. A summary of known results on quotient stacks. We first make
a basic observation. In this discussion, we restrict ourselves to the case that
X is a quasi-separated algebraic stack of finite type over k. If X = [X/ GLn ]
is a global quotient, then X → X is a GLn -torsor and one can construct
[X × An / GLn ] which is a vector bundle over X with the property that each
stabilizer acts faithfully on the fiber. Conversely, given a vector bundle V → X
with this same property, then the frame bundle Frame(V) is an algebraic space
and Frame(V) → X is a GLn -torsor. We conclude that
X ∼
= [alg space/ GLn ]
(i.e. a global quotient)
there exists a vector bundle V on X such
that for every point x, the stabilizer Gx
acts faithfully on the fiber V ⊗ k(x).
In a similar spirit, a theorem due to Totaro [Tot04] (generalized by Gross
[Gro13] to the non-normal case) implies
X∼
= [quasi-affine/ GLn ]
X∼
= [affine/ GLn ]
⇐⇒
⇐⇒
X satisfies the resolution property (i.e.
every coherent OX -module is surjected
onto by some vector bundle)
char(k) = 0
⇐⇒
X satisfies the resolution property, has
affine diagonal, and Hi (X, F) = 0 for
every coherent OX -module F and i > 0.
The last case when X ∼
= [Spec A/ GLn ] provides quotient stacks of the simplest structure. In fact, it is useful to replace GLn with any linearly reductive
group scheme G. Recall that an affine group scheme G of finite type over k is
called linearly reductive if the functor
G-Rep → k-Vect,
V → VG
is exact (or equivalently every G-representation is completely reducible). If G
is a linearly reductive group (e.g. G = GLn in characteristic 0) acting on an
affine scheme X = Spec A, then there is an affine GIT quotient
[Spec A/G] → Spec A//G := Spec AG
whose geometry is very well understood.
We now mention two nice results from [EHKV01].
28
ALPER
Proposition 3.11. [EHKV01, Thm. 2.18] If X is a smooth and separated
Deligne-Mumford stack of finite type over k with generically trivial stabilizer,
then X is a global quotient.
Proposition 3.12. [EHKV01, Thm. 3.6] Let X be a noetherian scheme and
let X → X be a Gm -gerbe corresponding to α ∈ H2 (X, Gm ). Then X is a
global quotient if and only if α is in the image of the Brauer map Br(X) →
H2 (X, Gm ).
Remark 3.13. This latter proposition can be used to construct algebraic stacks
that are not global quotient stacks, including a non-separated Deligne-Mumford
stack and a normal (but non-smooth) algebraic stack with affine diagonal. See
[EHKV01, Examples 2.21 and 3.12].
The following two questions are completely open.
Question 3.14. Is every separated Deligne-Mumford stack a global quotient
stack?
Question 3.15. Does every smooth algebraic stack with affine diagonal satisfy
the resolution property?
The question of whether a given algebraic stack is a global quotient stack
appears very difficult and is related to both global geometric properties (such
as existence of vector bundles) as well as arithmetic questions (such as the
surjectivity of the Brauer map). Below we will attempt to address the simpler
question: when are algebraic stack étale locally quotient stacks?
3.3. The local quotient structure of algebraic stacks. Recall that k denotes an algebraically closed field of any characteristic. Also recall that we
denote by TX,x the tangent space of a stack X at x; see Definition 3.6.
Theorem 3.16. Let X be a quasi-separated algebraic stack, locally of finite
type over k, with affine stabilizers. Let x ∈ X be a smooth k-point with smooth
and linearly reductive stabilizer group Gx . Then there exist an affine and étale
morphism (U, u) → (TX,x //Gx , 0), and a cartesian diagram
f
[TX,x /Gx ], 0
(X, x)
[Spec A/Gx ], w
(TX,x //Gx , 0)
(U, u)
such that f is étale and induces an isomorphism of stabilizer groups at w.
Remark 3.17. This theorem was established in [AHR15, Thm. 1.1]. The theorem is true even if the stabilizer Gx is not smooth if one replaces the tangent
space TX,x with the normal space Nx = (I/I2 )∨ , where I ⊂ OU denotes the
sheaf of ideals defining x in an open substack U ⊂ X where x is a closed point.
In the case that the stabilizer is smooth, TX,x ∼
= Nx .
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
29
This theorem implies that X and [TX,x /Gx ] have a common étale neighborhood of the form [Spec A/Gx ].
In the case that x is not necessarily a smooth point of X, one can prove a
similar structure theorem:
Theorem 3.18. [AHR15, Thm. 1.2] Let X be a quasi-separated algebraic stack,
locally of finite type over an algebraically closed field k, with affine stabilizers.
Let x ∈ X be a k-point with a linearly reductive stabilizer. Then there exist an
affine scheme Spec A with an action of Gx , a k-point w ∈ Spec A fixed by Gx ,
and an étale morphism
f : [Spec A/Gx ], w → (X, x)
such that f induces an isomorphism of stabilizer groups at w.
These theorems justify the philosophy that quotient stacks of the form
[Spec A/G], where G is a linearly reductive group, are the building blocks of
algebraic stacks near points with linearly reductive stabilizers in the same way
that affine schemes are the building blocks of schemes (and algebraic spaces).
These theorems were known in the following special cases:
(1) X is a Deligne-Mumford stack [AV02].
(2) More generally, X has quasi-finite inertia (i.e. all stabilizers groups
are finite but not necessarily reduced). This follows from [AOV08,
Thm. 3.2] and [KM97, §4].
(3) X ∼
= [X/G] where G is a linearly reductive algebraic group acting on
an affine scheme X and Gx is smooth and linearly reductive. This case
is often referred to as Luna’s étale slice theorem [Lun73, p. 97]. We
will discuss the relation between the theorems above and Luna’s étale
slice theorem in Section 4.4.1. We do emphasize though that there is
still content in the theorems even in the case that X ∼
= [Spec A/G] as
the theorems provide étale neighborhoods which are quotient stacks of
affine schemes by the stabilizer.
(4) X ∼
= [X/G] where G is a smooth affine group scheme and X is a normal
scheme [AK15, §2.2].
(5) X = Mss
g,n is the moduli stack of semistable curves. This is the central result of [AK15], where it is also shown that f can taken to be
representable.
We mention here counterexamples to Theorems 3.16 and 3.18 if either the
linearly reductive hypothesis or the condition of affine stabilizers is weakened.
Example 3.19. Some reductivity assumption of the stabilizer Gx is necessary
in Theorem 3.18. For instance, consider the group scheme G = Spec k[x, y]xy+1 →
A1 = Spec k[x] (with multiplication defined by y 7→ xyy ′ + y + y ′), where the
generic fiber is Gm but the fiber over the origin is Ga . Let X = BG and
x ∈ X be the point corresponding to the origin. There does not exist an étale
morphism ([W/Ga ], w) → (X, x), where W is an algebraic space over k with
an action of Ga .
30
ALPER
Example 3.20. It is essential to require that the stabilizer groups are affine
in a neighborhood of x ∈ X. For instance, let X be a smooth curve and
E → X be a group scheme whose generic fiber is a smooth elliptic curve but
the fiber over a point x ∈ X is isomorphic to Gm . Let X = BE. There is no
étale morphism ([W/Gm ], w) → (X, x), where W is an affine k-scheme with an
action of Gm .
3.4. Ingredients in the proof of Theorem 3.16. There are four main
techniques employed in the proof of Theorem 3.16:
(1)
(2)
(3)
(4)
deformation theory,
coherent completeness,
Tannakian formalism, and
Artin approximation.
Deformation theory produces an isomorphism between the nth infinitesimal
neighborhood Tn of 0 in T = [TX,x /Gx ] and the nth infinitesimal neighborhood
Xn of x in X. It is not at all obvious, however, that the system of closed
morphisms {fn : Tn → X} algebraizes. We establish algebraization in two
steps.
The first step is effectivization. To accomplish this, we prove a result similar
in spirit to Grothendieck’s existence theorem [EGA, III.5.1.4], which we refer
to as coherent completeness. To motivate the definition, recall that if (A, m) is a
complete local noetherian ring, then Coh(Spec A) = limn Coh(Spec(A/mn+1 )).
←−
Here, if X is a noetherian scheme, then Coh(X) denotes the category of coherent OX -modules.
Black Box 3. (Coherent completeness) Let G be a linearly reductive affine
group scheme over an algebraically closed field k. Let Spec A be a noetherian affine scheme with an action of G, and let x ∈ Spec A be a k-point
fixed by G. Suppose that AG is a complete local ring. Let X = [Spec A/G]
and let Xn be the nth infinitesimal neighborhood of x. Then the natural
functor
(3.9)
Coh(X) → lim Coh Xn
←−
n
is an equivalence of categories.
Remark 3.21. This was proven in [AHR15, Thm. 3]. The proof is not tremendously difficult but does require some care.
Remark 3.22. Let Y = Spec AG and let Yn = Spec AG /(m ∩ AG )n+1 be the nth
nilpotent thickening of the image of x under Spec A → Spec AG . The above
theorem implies that
(3.10)
Coh(X) → lim Coh(X ×Y Yn )
←−
n
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
31
is an equivalence, which had been established in [GZB15]. We emphasize
that the above theorem is significantly stronger in that it involves families of
coherent sheaves defined only on the nilpotent thickenings of the closed point
x ∈ X rather the fiber X ×Y Y0 .
Example 3.23. Consider X = [AN /Gm ]. Then the theorem states that
Gm -equivariant sheaves on AN are equivalent to compatible families of Gm equivariant sheaves on Spec k[x1 , . . . , xN ]/(x1 , . . . , xN )n+1 . Meanwhile, the
equivalence (3.10) is trivial as Yn = Y = Spec k.
The other key ingredient in the proof of Theorem 3.16 is Tannakian formalism.
Black Box 4. (Tannakian formalism) Let X be an excellent stack and Y
be a noetherian algebraic stack with affine stabilizers. Then the natural
functor
Hom(X, Y) → Homr⊗,≃ Coh(Y), Coh(X)
is an equivalence of categories, where Homr⊗,≃ (Coh(Y), Coh(X)) denotes
the category whose objects are right exact monoidal functors Coh(Y) →
Coh(X) and morphisms are natural isomorphisms of functors.
Remark 3.24. In the above generality, this result was established in [HR14,
Thm. 1.1]. Other versions had been established in [Lur04] and [BH15].
This proves that morphisms between algebraic stacks Y → X are equivalent to symmetric monoidal functors Coh(X) → Coh(Y). Therefore, to prove
Theorem 3.16, we can combine Black Box 3 with Black Box 4 and the above
deformation-theoretic observations to show that the morphisms {fn : Tn → X}
b T //G ,0 . The morb → X, where T
b = TX,x ×T //G Spec O
effectivize to fb: T
x
x
X,x
X,x
b
phism f is then algebraized using Artin approximation over the GIT quotient
TX,x //Gx . We will give the details of this argument in the next lecture.
32
ALPER
Lecture 4: A Luna étale slice theorem for algebraic stacks
and applications
In this lecture, we will give proofs of Theorems 3.16 and 3.18. Recall that
these theorems assert roughly that any algebraic stack with affine stabilizers
is étale locally a quotient stack in a neighborhood of a point with a linearly
reductive stabilizer. The proof of Theorem 3.18 will rely on an equivariant
version of Artin approximation (Theorem 4.2). We will also give several applications of Theorems 3.16 and 3.18.
Throughout this lecture, we will use the notation that if X is an algebraic
stack over a field k and x ∈ X is a closed k-point, then Xn denotes the nth
nilpotent thickening of the inclusion of the residual gerbe BGx ֒→ X; more
precisely, if BGx ֒→ X is defined by the sheaf of ideals I, then Xn ֒→ X is
defined by In+1 . The point x is not included in the notation Xn but it should
always be clear from the context.
4.1. Proof of Theorem 3.16.
Proof of Theorem 3.16. We may assume that x ∈ X is a closed point. Define
the quotient stack T = [TX,x /Gx ]. Since Gx is linearly reductive and x ∈ X is
a smooth point, a simple deformation theory argument implies that there are
isomorphisms Xn ∼
= Tn .
Let T → T = TX,x //Gx be the morphism to the GIT quotient, and denote
b := Spec O
b T,0 ×T T is
by 0 ∈ T the image of the origin. The fiber product T
noetherian and a quotient stack of the form [Spec A/G] with AG a complete
b satisfies the hypotheses of Black Box 3. We have
local ring. Therefore, T
equivalences
b X) ≃ Homr⊗,≃ Coh(X), Coh(T)
b
Hom(T,
(Tannakian formalism)
≃ Homr⊗,≃ Coh(X), lim Coh Tn
(coherent completeness)
←−
≃ lim Homr⊗,≃ Coh(X), Coh Tn
←−
≃ lim Hom Tn , X
(Tannakian formalism).
←−
b → X filling in
Thus the morphisms Tn ∼
= Xn ֒→ X extend to a morphism T
the diagram
Xn ∼
= Tn
b
T
T
X
The functor
F : Sch/T → Sets,
b T,0
Spec O
T.
(T ′ → T ) 7→ T ′ ×T T → X / ∼
b → X yields an
is easily checked to be limit preserving. The morphism T
b T,0 . By Artin approximation, there exist an étale
element of F over Spec O
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
33
morphism (U, u) → (T, 0) where U is an affine scheme and a morphism (U ×T
b 0) → (X, x) to first order. Since U ×T T
T, (u, 0)) → (X, x) agreeing with (T,
is smooth at (u, 0) and X is smooth at x, and since U ×T T → X induces an
isomorphism of tangent spaces at (u, 0), the morphism U ×T T → X is étale
at (u, 0). After shrinking U suitably, the theorem is established.
Remark 4.1. The smoothness hypothesis of x ∈ X is used above to establish
the isomorphisms Tn ∼
= Xn as well as the étaleness of U ×T T → X. More
critically, though, it implies that lim Γ(Xn , OXn ) is the completion of a finitely
←−
b T,0 . If x ∈ X
generated k-algebra since this inverse limit is identified with O
is not smooth, there does not appear to be a direct way to establish that
lim Γ(Xn , OXn ) (which can be identified with the Gx -invariants of a miniversal
←−
deformation space of Gx ) is the completion of a finitely generated k-algebra.
Recall that Artin algebraization was a direct consequence of Artin approximation in the case that the complete local ring was known to be the completion of
a finitely generated algebra (see Remark 2.14). In a similar manner, one could
deduce Theorem 3.18 if we did know that lim Γ(Xn , OXn ) was the completion
←−
of a finitely generated algebra. In order to circumvent this problem, we will
establish an equivariant version of Artin algebraization.
4.2. Equivariant Artin algebraization.
Theorem 4.2 (Equivariant Artin algebraization). Let H be a linearly reductive affine group scheme over k. Let X be a limit preserving category fibered in
groupoids over k. Let Z = [Spec A/H] be a noetherian algebraic stack over k.
Suppose that AH is a complete local k-algebra. Let z ∈ Z be the unique closed
point. Let ξb: Z → X be a morphism that is formally versal at z ∈ Z.8 Then
there exist
(1) an algebraic stack W = [Spec B/H] of finite type over k and a closed
point w ∈ W;
(2) a morphism ξ : W → X;
b w , where W
b w is defined as the fiber product
(3) an isomorphism Z ∼
=W
bw
W
W
φ
b W,φ(w)
Spec O
8
W = Spec B H ;
This is a stacky extension of the notion of formal versality introduced in Definition 2.8.
Namely, it means that for every commutative diagram of solid arrows
Z0
T
Z
ξb
T′
′
X
where T ֒→ T is a closed immersion of local artinian stacks over k (i.e. noetherian algebraic
stacks over k with a unique point), there is a lift T ′ → Z filling in the above diagram.
34
ALPER
b Zn ∼
(4) a compatible family of isomorphisms ξ|
= ξ|Wn (under the identifica∼
tion Zn = Wn ) for n ≥ 0.
Remark 4.3. If one takes H to be the trivial group, one recovers precisely the
statement of Artin algebraization given in Theorem 2.12.
Proof. This can be proved in a similar fashion to Theorem 2.12 by appealing
to a stacky generalization of Theorem 2.1. See [AHR15, App. A].
4.3. Proof of Theorem 3.18.
Proof of Theorem 3.18. We may assume that x ∈ X is a closed point. Let
T = [TX,x /Gx ], T → T = TX,x //Gx be the morphism to the GIT quotient,
b := Spec O
b T,0 ×T T where 0 ∈ T is the image of the origin. Since Gx is
and T
linearly reductive, a simple deformation theory argument implies that there
are compatible closed immersions Xn ֒→ Tn . The ideals sheaves In defining
these closed immersions give a compatibly family {OXn /In } of coherent OXn b satisfies the hypotheses of Black Box 3, there exists an ideal
modules. Since T
sheaf I ⊂ OTb such that the surjection OTb → OTb /I extends the surjections
b extending
OXn → OXn /In . Therefore there exists a closed immersion Z ֒→ T
the given closed immersions Xn ֒→ Tn . This yields a commutative diagram
Xn
Z
Tn
b
T
ξb
X
T
b T,0
Spec O
T.
of solid arrows. Since Z also satisfies the hypotheses of Black Box 3 and the
nilpotent thickenings Zn are identified with Xn , the equivalences
Hom(Z, X) ≃ Homr⊗,≃ Coh(X), Coh(Z)
(Tannakian formalism)
≃ Homr⊗,≃ Coh(X), lim Coh Xn
(coherent completeness)
←−
≃ lim Homr⊗,≃ Coh(X), Coh Xn
←−
≃ lim Hom Xn , X
(Tannakian formalism).
←−
imply the existence of a morphism ξb: Z → X filling in the above diagram. One
can check easily that ξb: Z → X is formally versal. By applying equivariant
Artin algebraization with (Theorem 4.2) with H = Gx , we obtain a morphism
ξ : W = [Spec B/Gx ] → X where W is of finite type over k, a closed point
b w over X, where W
b w is defined as in the
w ∈ W and an isomorphism Z → W
b w commutes over X follows
statement of Theorem 4.2. (The fact that Z → W
b Zn ∼
from the compatible family of isomorphisms ξ|
= ξ|Wn and the Tannakian
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
35
formalism (Black Box 4) as X is a noetherian algebraic stack with affine stabilizers.) Finally, it is easy to see that ξ : W → X is étale at w which completes
the proof.
4.4. Applications. In this section, we provide a few applications of Theorems
3.16 and 3.18. We will not include the proofs and instead refer the reader to
[AHR15].
4.4.1. Application 1. Our first application is a generalization of Luna’s étale
slice theorem and can be viewed as a refinement of Theorems 3.16 and 3.18 in
the case that X = [X/G] is a quotient stack.
Theorem 4.4. Let X be a quasi-separated algebraic space, locally of finite
type over k, with an action of a smooth affine group scheme G over k. Let
x ∈ X be a k-point with a linearly reductive stabilizer Gx . Then there exist an
affine scheme W with an action of Gx which fixes a point w and an unramified
Gx -equivariant morphism (W, w) → (X, x) such that fe: W ×Gx G → X is
étale.9
Let Nx = TX,x /TGx,x be the normal space to the orbit at x; this inherits
a natural linear action of Gx . If x ∈ X is smooth, then it can be arranged
that there is an étale Gx -equivariant morphism W → Nx such that W//Gx →
Nx //Gx is étale and
Nx ×Gx G
W ×Gx G
fe
X
Nx //Gx
W//Gx
is cartesian.
Remark 4.5. The theorem above follows from Luna’s étale slice theorem [Lun73]
if X is affine. In this case, Luna’s étale slice theorem is stronger than Theorem 4.4 as it asserts additionally that W → X can be arranged to be a locally
closed immersion (which is obtained by choosing a Gx -equivariant section of
TX,x → Nx and then restricting to an open subscheme of the inverse image of
Nx under a Gx -equivariant étale morphism X → TX,x ). However, when W is
not normal, it is necessary to allow unramified neighborhoods (for instance,
consider the example of Gm acting on the nodal cubic).
4.4.2. Application 2. Our second application is a generalization of Sumihiro’s
theorem on torus actions.
Theorem 4.6. Let X be a quasi-separated algebraic space, locally of finite type
over k, with an action of a torus T . If x ∈ X is a k-point, then there exist a
T -equivariant étale neighborhood (Spec A, u) → (X, x).
9
Here, W ×Gx G denotes the quotient (W × G)/Gx . Note that there is an identification
of GIT quotients (W ×Gx G)//G ∼
= W//Gx .
36
ALPER
Remark 4.7. More generally, if X is a Deligne–Mumford stack (with the other
hypotheses remaining the same), one can show that there exist a reparameterization α : T → T and an étale neighborhood (Spec A, u) → (X, x) that is
equivariant with respect to α.
In the case that X is a normal scheme, Theorem 4.6 was proved by Sumihiro
in [Sum74, Cor. 2] in which case Spec A → X can be taken to be an open
neighborhood. In the example of the Gm -action on the nodal cubic, there
does not exist a Gm -invariant affine open neighborhood of the node. Thus, for
non-normal schemes, it is necessary in general to allow étale neighborhoods.
In fact, we can prove more generally:
Theorem 4.8. Let X be a quasi-separated algebraic space, locally of finite type
over k, with an action of an affine group scheme G of finite type over k. Let
x ∈ X be a k-point with linearly reductive stabilizer Gx . Then there exist an
affine scheme W with an action of G and a G-equivariant étale neighborhood
W → X of x.
4.4.3. Application 3. We now translate the conclusion of Theorem 3.18 in the
case that X is the stack of all (possibly singular) curves. By a curve, we mean
a proper scheme over k of pure dimension one.
Theorem 4.9. Let C be an n-pointed curve. Suppose that every connected
component of C is either reduced of arithmetic genus g 6= 1 or contains a
marked point. Suppose that Aut(C) is smooth and linearly reductive. Then
there exist an affine scheme W of finite type over k with an action of Aut(C)
fixing a k-point w ∈ W and a miniversal deformation
C
C
W
w
Spec k
of C ∼
= Cw such that there exists an action of Aut(C) on the total family C
compatible with the action of Aut(C) on W and Cw .
4.4.4. Application 4. In our final application, we will conclude that under
suitable hypotheses the coherent completion of a point of an algebraic stack
exists and moreover we will give equivalent conditions for algebraic stacks to
be étale locally isomorphic.
Let X be a noetherian algebraic stack over k with affine stabilizers and let
x ∈ X be a closed k-point. We say that (X, x) is a complete local stack if the
natural functor
Coh(X) → lim Coh Xn
←−
n
is an equivalence.
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
37
Theorem 4.10. Let X be a quasi-separated algebraic stack, locally of finite
type over k, with affine stabilizers. For any k-point x ∈ X with linearly reducb x, x
tive stabilizer Gx , there exist a complete local stack (X
b) and a morphism
b
η : (Xx , x
b) → (X, x) inducing isomorphisms of nth infinitesimal neighborhoods
b x , η) is unique up to unique 2-isomorphism.
of x
b and x. The pair (X
If (W = [Spec A/Gx ], w) → (X, x) is an étale morphism as in Theorem 3.18
and φ : W → W = Spec AGx is the morphism to the GIT quotient, then the
b x is constructed as the fiber product
completion X
bx
X
W
φ
b W,φ(w)
Spec O
W.
Example 4.11. The pair ([An /Gm ], 0) is a complete local stack where Gm
acts with weight 1 on all coordinates.
Example 4.12. If we let X = [A2 /Gm ] where Gm acts with weights (1, −1)
on the coordinates A2 = Spec k[x, y] and x ∈ X denotes the origin, then (X, x)
is not a complete local stack. The completion of X at x is given by the fiber
product
bx
[A2 /Gm ]
X
Spec k[[xy]]
Spec k[xy]
which is identified with the quotient stack [Spec(k[[xy]] ⊗k[xy] k[x, y])/Gm ].
The next result is a stacky generalization of Corollary 1.14, which was our
first application of Artin approximation.
Theorem 4.13. Let X and Y be quasi-separated algebraic stacks, locally of
finite type over k, with affine stabilizers. Suppose x ∈ X and y ∈ Y are kpoints with linearly reductive stabilizer group schemes Gx and Gy , respectively.
Then the following are equivalent:
(1) There exist compatible isomorphisms Xn → Yn .
bx → Y
by .
(2) There exists an isomorphism X
(3) There exist an affine scheme Spec A with an action of Gx , a point
w ∈ Spec A fixed by Gx , and a diagram of étale morphisms
[Spec A/Gx ]
f
X
g
Y
such that f (w) = x and g(w) = y, and both f and g induce isomorphisms of stabilizer groups at w.
38
ALPER
If, in addition, the points x ∈ X and y ∈ Y are smooth and if the stabilizers Gx
and Gy are smooth, then the conditions above are equivalent to the existence of
an isomorphism Gx → Gy of group schemes and an isomorphism TX,x → TY,y
of tangent spaces which is equivariant under Gx → Gy .
References
[AHR15]
J. Alper, J. Hall, and D. Rydh, A Luna étale slice theorem for algebraic stacks,
ArXiv e-prints (2015), arXiv:1504.06467.
[AK15]
Jarod Alper and Andrew Kresch, Equivariant versal deformations of semistable
curves, ArXiv e-prints (2015), arXiv:1510.03201.
[AM69]
M. F. Atiyah and I. G. Macdonald, Introduction to commutative algebra,
Addison-Wesley Publishing Co., Reading, Mass.-London-Don Mills, Ont., 1969.
[AOV08] D. Abramovich, M. Olsson, and A. Vistoli, Tame stacks in positive characteristic,
Ann. Inst. Fourier (Grenoble) 58 (2008), no. 4, 1057–1091.
[Art69a] M. Artin, Algebraic approximation of structures over complete local rings, Inst.
Hautes Études Sci. Publ. Math. (1969), no. 36, 23–58.
, Algebraization of formal moduli. I, Global Analysis (Papers in Honor of
[Art69b]
K. Kodaira), Univ. Tokyo Press, Tokyo, 1969, pp. 21–71.
[Art74]
, Versal deformations and algebraic stacks, Invent. Math. 27 (1974), 165–
189.
[AV02]
Dan Abramovich and Angelo Vistoli, Compactifying the space of stable maps, J.
Amer. Math. Soc. 15 (2002), no. 1, 27–75 (electronic).
[BH15]
B. Bhatt and D. Halpern-Leistner, Tannaka duality revisited, ArXiv e-prints
(2015), arXiv:1507.01925.
[CJ02]
B. Conrad and A. J. de Jong, Approximation of versal deformations, J. Algebra
255 (2002), no. 2, 489–515.
[EGA]
A. Grothendieck, Éléments de géométrie algébrique, I.H.E.S. Publ. Math. 4, 8,
11, 17, 20, 24, 28, 32 (1960, 1961, 1961, 1963, 1964, 1965, 1966, 1967).
[EHKV01] Dan Edidin, Brendan Hassett, Andrew Kresch, and Angelo Vistoli, Brauer
groups and quotient stacks, Amer. J. Math. 123 (2001), no. 4, 761–777.
[Eis95]
D. Eisenbud, Commutative algebra, Graduate Texts in Mathematics, vol. 150,
Springer-Verlag, New York, 1995, With a view toward algebraic geometry.
[Gro13]
P. Gross, Tensor generators on schemes and stacks, preprint, June 2013,
arXiv:1306.5418.
[GZB15] Anton Geraschenko and David Zureick-Brown, Formal GAGA for good moduli
spaces, Algebr. Geom. 2 (2015), no. 2, 214–230.
[Hal14]
J. Hall, Openness of versality via coherent functors, J. Reine. Angew. Math.
(2014), available online, http://dx.doi.org/10.1515/crelle-2014-0057.
[Har77]
Robin Hartshorne, Algebraic geometry, Springer-Verlag, New York-Heidelberg,
1977, Graduate Texts in Mathematics, No. 52.
[HR13]
J. Hall and D. Rydh, Artin’s criteria for algebraicity revisited, ArXiv e-prints
(2013), arXiv:1306.4599.
[HR14]
J. Hall and D. Rydh, Coherent Tannaka duality and algebraicity of Hom-stacks,
ArXiv e-prints (2014), arXiv:1405.7680.
[KM97]
S. Keel and S. Mori, Quotients by groupoids, Ann. of Math. (2) 145 (1997),
no. 1, 193–213.
[Lun73]
D. Luna, Slices étales, Sur les groupes algébriques, Bull. Soc. Math. France,
Mém., vol. 33, Soc. Math. France, Paris, 1973, pp. 81–105.
[Lur04]
J. Lurie, Tannaka Duality for Geometric Stacks, ArXiv e-prints (2004),
arXiv:math/0412266.
ARTIN ALGEBRAIZATION AND QUOTIENT STACKS
[Nér64]
[Pop85]
[Pop86]
[Pop90]
[Rim80]
[Sch68]
[Stacks]
[Sum74]
[Swa98]
[Tot04]
39
André Néron, Modèles minimaux des variétés abéliennes sur les corps locaux et
globaux, Inst. Hautes Études Sci. Publ.Math. No. 21 (1964), 128.
Dorin Popescu, General Néron desingularization, Nagoya Math. J. 100 (1985),
97–126.
D. Popescu, General Néron desingularization and approximation, Nagoya Math.
J. 104 (1986), 85–115.
Dorin Popescu, Letter to the editor: “General Néron desingularization and approximation”, Nagoya Math. J. 118 (1990), 45–53.
Dock Rim, Equivariant G-structure on versal deformations, Trans. Amer. Math.
Soc. 257 (1980), no. 1, 217–226.
Michael Schlessinger, Functors of Artin rings, Trans. Amer. Math. Soc. 130
(1968), 208–222.
The
Stacks
Project
Authors,
Stacks
Project,
http://stacks.math.columbia.edu.
H. Sumihiro, Equivariant completion, J. Math. Kyoto Univ. 14 (1974), 1–28.
Richard G. Swan, Néron-Popescu desingularization, Algebra and geometry
(Taipei, 1995), Lect. Algebra Geom., vol. 2, Int. Press, Cambridge, MA, 1998,
pp. 135–192.
B. Totaro, The resolution property for schemes and stacks, J. Reine Angew.
Math. 577 (2004), 1–22.
(Alper) Mathematical Sciences Institute, Australia National University,
Canberra, ACT 0200
E-mail address: [email protected]
| 0math.AC
|
ON THE DIMENSION OF CONTACT LOCI AND THE IDENTIFIABILITY
OF TENSORS
arXiv:1706.02746v3 [math.AG] 1 Dec 2017
EDOARDO BALLICO, ALESSANDRA BERNARDI AND LUCA CHIANTINI
Abstract. Let X ⊂ Pr be an integral and non-degenerate variety. Set n := dim(X). We prove
that if the (k + n − 1)-secant variety of X has (the expected) dimension (k + n − 1)(n + 1) − 1 < r
and X is not uniruled by lines, then X is not k-weakly defective and hence the k-secant variety
satisfies identifiability, i.e. a general element of it is in the linear span of a unique S ⊂ X
with ♯(S) = k. We apply this result to many Segre-Veronese varieties and to the identifiability
of Gaussian mixtures G1,d . If X is the Segre embedding of a multiprojective space we prove
identifiability for the k-secant variety (assuming that the (k + n − 1)-secant variety has dimension
(k + n − 1)(n + 1) − 1 < r, this is a known result in many cases), beating several bounds on the
identifiability of tensors.
1. Introduction
The computation of the decompositions of tensors and the related evaluation of their ranks
recently received great interest from people working in algebraic geometry. Indeed, a geometric
analysis on projective spaces of tensors, Segre embeddings and their secant varieties produced
currently several new results in the field.
From the geometric point of view, the study of decompositions and ranks can be turned to a
general problem. For any (irreducible) algebraic variety X which spans a projective space Pr , one
can define the X-rank rX (q) of a point q ∈ Pr as the minimal cardinality of a finite set S ⊂ X such
that q ∈ hSi, where h i denotes the linear span. Any set S for which the minimum is attained is a
minimal X-decomposition of q.
In tensor analysis, one interprets Pr as a space of tensors of given type (n1 + 1) × · · · × (ns + 1)
(mod scalars), and X as the locus of tensors of rank 1. In geometric terms, X is the multiprojective
product Pn1 × · · · × Pns , naturally embedded in Pr via the Segre map. Then, the notions of X-rank
and minimal decompositions match with the usual multilinear algebra notion. The set of tensors of
fixed rank k corresponds to an open subset of the k-th secant variety σk (X) of X. Properties of the
rank decompositions can be derived from geometric properties of X, which determine some features
of σk (X). Examples of such analysis are contained in several papers, e.g. see [8, 10, 25, 26, 36, 41],
and the book [38].
In this paper we study the problem of the identifiability of generic points q ∈ Pr of given X-rank
k. A point q is X-identifiable when the set S(q, X) of minimal X-decompositions of q is a singleton;
in any case the cardinality of S(q, X) is called the secant degree of q or the k-th secant degree of q.
In tensor analysis the identifiability of a tensor is fundamental for algorithms of data compression
and for the analysis of mixture models (see [33, 34, 35, 43]). We will say that X is k-identifiable if
a generic point of σk (X) is X-identifiable.
We first work over an algebraically closed field of characteristic 0. Our investigation is based on
the notion of weak defectivity introduced by the last author and C. Ciliberto in [25]. A projective
variety X is weakly k-defective if hyperplanes of Pr , which are tangent to X at k general points, are
indeed tangent along a positive dimensional subvariety. It turns out (see [26, Theorem 2.5]) that
when r > k(n + 1) − 1 then X is k-identifiable, unless it is weakly k-defective (unfortunately the
reverse is not true). Therefore the knowledge on the non-k weak defectivity provides an effective
way to determine geometrically the identifiability of generic points. This approach has been applied
to tensors and linear systems of tensors in a series of papers ([1, 16, 18, 19, 29, 30, 31]), where
identifiability problems are partially or totally solved.
A challenging task is therefore to get criterions that can establish identifiability (or not) in the
cases when weak defectivity arises. To this purpose there is a related notion that is very helpful,
meaning the tangential weak defectivity. Our new analysis starts with the following observation
1
2
CONTACT LOCI AND IDENTIFIABILITY
(see [27]). If X is a weakly k-defective variety of dimension n, then a general choice of a set A of
k points on X determines a positive dimensional tangential k-contact locus Γk (A): the tangency
locus of the span of the tangent spaces to X at the points of A. If for a general choice of A we
have that dim Γk (A) > 0 than X is tangential k-weakly defective. It turns out that the k-th secant
degree of a general element of X is related with the k-th secant degree of the contact locus Γk (A)
for a general choice of A (cf. Lemma 4.2, Corollary 4.3, [26, Theorem 2.4] and part (iii) of [27,
Proposition 3.9]). We will exploit in more details the relation in Section 4.
Moreover, when r ≥ (k + 1)(n + 1) − 1, then the weak k-defectivity of X implies the weak
(k + 1)-defectivity. Moreover, we prove that, under some geometric assumption on X, if we take a
generic subset A as above and let A′ be obtained by adding to A a generic point of X, then the
dimension of the tangential contact locus Γk+1 (A′ ) is strictly bigger than the dimension of Γk (A).
As the dimension of contact loci is bounded by n − 1, by repeating the previous construction we
obtain the following result.
Theorem. (see Theorem 5.1). Let X ⊂ Pr be an integral and non-degenerate n-dimensional variety and let k be a positive integer. Assume that X is not uniruled by lines and that dim σk+n−1 (X) =
(k + n − 1)(n + 1) − 1 < r. Then X is not weakly j-defective for every j ≤ k.
It follows from the previous analysis that, with the assumptions of Theorem 5.1, X is jidentifiable for every j ≤ k.
The result can be directly applied e.g. to Segre-Veronese embeddings of products of projective
spaces, provided that the multidegree (d1 , . . . , ds ) satisfies di > 1 for all i. When some degree di
is 1, as in the case of Segre embeddings (where di = 1 for all i), then Theorem 5.1 does not apply
directly. However, for the Segre product we can modify the argument of Theorem 5.1, and prove
that the same conclusion holds when for each i ∈ {1, . . . , s} there is j ∈ {1, . . . , s} \ {i} such that
nj = ni (see Theorem 5.3).
Combining Theorem 5.3 with the main result in [23], for the Segre embedding X of s ≥ 5 copies
of P1 , we find that X is not weakly k-defective, hence it is k-identifiable, whenever (k+s−1)(s+1) <
2s − 1 (see Corollary 5.4).
Corollary 5.4 improves our previous knowledge on the weak defectivity of Segre embeddings of
many copies of P1 , hence on the identifiability of generic binary tensors. Indeed it extends the range
2
−1
− s, which beats all previously known
where we know that k-identifiability holds up to k ≤ ss+1
bounds (see Section 4 of [20]) as s grows. Since identifiability cannot hold, for trivial dimensional
reasons, when k > s2 /(s + 1) − 1, then we are quite close to a complete analysis of the generic
identifiability of binary tensors. We also get almost optimal results for the Segre embedding of
(Pm )s , m > 1, see Corollary 5.5) (we use the very strong [9, Theorem 3.1]).
See Example 5.8 for the case of Gaussian mixtures G1,d .
When the assumptions of Theorem 5.1 does not hold, one cannot conclude the non weak kdefectivity of X. Yet, even if X is weakly k-defective, it can still be k-identifiable. We prove the
following result.
Theorem. (seePTheorem 5.9) Let X be the Segre embedding of a product Pn1 × · · · × Pnq in
Pr , with n =
ni = dim(X). Let k be a positive integer. Assume that dim σk+n−1 (X) =
(k + n − 1)(n + 1) − 1 < r (in particular X is not (k + n − 1)-defective). Then X is j-identifiable
for every j ≤ k.
It would be interesting to overcame the assumption “X not uniruled by lines ” of Theorem 5.3 in
other cases, apart the Segre varieties covered by Theorem 5.9, because it would expand the known
cases of identifiability (e.g. for Grassmannians see [6], [17], [21], [22]; for the tangential variety
of the Veronese varieties see [7]). We just recall that conjecturally only four Grassmannians have
some defective secant variety.
We stress that, in particular, Remark 5.10 suggests that if one is interested in the k-identifiability
of some specific Segre product then the computational procedure induced by Theorem 5.9 can be
faster than previously known direct procedures.
In the last section of the paper we work over an algebraically closed base field K of arbitrary
characteristic. To include varieties X defined in positive characteristic, we will assume that X is
not very strange (see Definition 6.1). Under this mild assumption, we prove the following.
CONTACT LOCI AND IDENTIFIABILITY
3
Theorem 1.1. Let M be an integral projective variety and L, R line bundles on M with L very
ample. Set n := dim M . Fix a linear subspace V ⊆ H 0 (L) such that the associated map j : M → Pr
is an embedding with image not very strange. Fix an integer k ≤ min{h0 (R), dim V − n − 2}. Let
W ⊆ H 0 (L ⊗ R) be the image of the multiplication map V ⊗ H 0 (R) → H 0 (L ⊗ R) and let X ⊂ Pr ,
r = dim W − 1, be the embedding induced by W . Then σk (X) has dimension kn + k − 1 and
♯(S(q, X)) = 1 for a general q ∈ σk (X).
We will apply Theorem 1.1 to the Segre-Veronese embeddings of a multiprojective space in
arbitrary characteristic (see Example 6.3), where we do not have an analogue of [26, Theorem 2.5],
hence the analysis of the previous section does not apply. Notice that Theorem 1.1 does not apply
to Segre embeddings, because it requires L very ample (or at least birationally very ample) and
h0 (X, R) > 1.
The authors wish to thank C. Bocci, for his help to understand the case of almost unbalanced
tensors (see Example 4.10 below) and the anonimous referee for pointing out an improvement in
the formula after Corollary 5.4.
2. General notation
In this section we work with projective spaces Pr defined over an algebraically closed field K.
Let X ⊂ Pr be an integral and non-degenerate n-dimensional variety. We will denote with Xreg
the set of regular points in X. For any x ∈ Xreg we will denote with Tx X the tangent space to X
at x.
For any integer k ≥ 2 we will denote with σk (X) the k-th secant variety of X which is an
irreducible projective variety. If k(n + 1) − 1 ≤ r, then k(n + 1) − 1 is the expected dimension of
σk (X) and X is k-defective if dim σk (X) < k(n + 1) − 1.
We will often use the fact that if X is not k-defective, then, for all k ′ ≤ k, X is not k ′ -defective.
Definition 2.1. For q ∈ Pr general, the X-rank rX (q) of a point q ∈ Pr is the minimal cardinality
of a finite set A ⊂ X such that q ∈ hAi.
We define SX (q) as the collection of subsets A ⊂ X of cardinality rX (q) such that q ∈ hAi.
When SX (q) is finite, we will write αX (q) for the cardinality of SX (q).
Remark 2.2. Since X is irreducible, then the secant varieties σk (X) are irreducible. Moreover a
general point of σk (X) has X-rank k.
By definition, the set SX (q) is finite for q ∈ σk (X) general if and only if dim σk (X) = k(n +
1) − 1 ≤ r. In this case, we will write simply αk for the cardinality of SX (q), where q ∈ σk (X) is
generic. Such an αk is often called the k-th secant degree of X.
The following result holds (with the same elementary proof) in arbitrary characteristic.
Proposition 2.3. Assume dim σk+1 (X) = (k+1)(n+1)−1 and αk > 1. Fix a general q ∈ σk+1 (X),
any A ∈ S(q) and any a ∈ A. There are at least αk elements B1 . . . , Bαk ∈ S(q) (one of them
being A) such that a ∈ Bi for all i = 1, . . . , αk .
Proof. Since dim σk+1 (X) = (k + 1)(n + 1) − 1, we have dim σk (X) = k(n + 1) − 1 and hence αk
is a well-defined finite positive integer. Fix a general q ∈ σk+1 (X) and take any A ∈ S(q) and any
a ∈ A. Set E := A \ {a}. Since q is general in σk+1 (X) and dim σk+1 (X) = (k + 1)(n + 1) − 1,
a dimensional count gives that for each D ∈ S(q) the set D is a general element of the (k + 1)symmetric product of X. Thus E is a general subset of X with cardinality k. Since ♯(A) = rX (q),
we have q ∈
/ hA′ i for any A′ ( A; in particular q ∈
/ E. Thus h{a, q}i is a line, which meets hEi
in a unique point q ′ (because hEi + ha, qi = hAi) and q ′ ∈
/ hE ′ i for any E ′ ( E. For a general
′
q ∈ hAi, q is a general element of hEi. Since E is general, q ′ is general in σk (X). Thus rX (q ′ ) = k,
E ∈ S(q ′ ) and ♯(S(q ′ )) = αk > 1. Fix any F ∈ S(q ′ ). We have q ∈ h{F, a}i. Since rX (q) = k + 1,
we have a ∈
/ F and F ∪ {a} ∈ S(q). Now we have exactly αk such F ’s, say F1 , . . . , Fαk , the Bi of
the statement are Bi = {Fi , a} for i = 1, . . . , αk concluding the proof.
3. Weak defectivity and identifiability
In this section, as well as in the next two sections, we assume that the algebraically closed base
field K has characteristic 0.
4
CONTACT LOCI AND IDENTIFIABILITY
For a subset A of cardinality k contained in the regular locus Xreg we set:
MA := h∪x∈A Tx Xi.
By Terracini’s lemma ([8, Corollary 1.11]), for a general choice of A the space MA is the tangent
space to σk (X) at a general point u ∈ hAi. Thus dim MA = min{k(n + 1) − 1, dimhXi} for A
general, unless X is k defective.
Definition 3.1. The tangential k-contact locus Γk (A) is the closure in X of the union of all the
irreducible components which contain at least one point of A, of the locus of points of Xreg where
MA is tangent to X.
We will write tk := dim Γk (A) for a general choice of A.
We say that X is weakly k-defective if tk > 0.
We will write just Γk instead of Γk (A), when A is general.
With the terminology just introduced we have the following remark:
Remark 3.2. [25, Remark 3.1(iii) at p. 159] If σk+1 (X) ( Pr has the expected dimension
(k + 1) dim X + (k + 1) − 1 < r and X is weakly k-defective, then X is also weakly (k + 1)-defective.
We will need the following result of [27].
Proposition 3.3. For a general choice of A ⊂ Xreg , the algebraic set Γk := Γk (A) is equidimensional and either it is irreducible or it has exactly k irreducible components, each of them containing
a different point of A.
Proof. It is Proposition 3.9 of [27]. (Notice that our σk (X) is called the (k − 1)-th secant variety
in [27], so that all the k’s there should be here replaced by k + 1, in order to apply the results of
[27] to our terminology).
Definition 3.4. Following [26] and [27], we say that Γk is of the type I (resp. type II ) if Γk is
irreducible (resp. it is not irreducible) for a general choice of A.
The fundamental technical results of the present paper is the following lemma. It shows how
one can control the growth of the dimension of contact loci. Mainly, one must take care of the case
of contact loci of type II, which are not irreducible.
Lemma 3.5. Let X ⊂ Pr be an irreducible projective variety such that dim σk+1 (X) = (k + 1)(n +
1) − 1 < r and assume that X is weakly k-defective. Then X is weakly (k + 1)-defective. Let B
(resp. A) be a general subset of k (resp. k + 1) points of X. Call Γk (resp. Γk+1 ) the tangential
contact locus of X at the points of B (resp. A). Then either:
• dim Γk < dim Γk+1 ; or
• both Γk and Γk+1 have type II, each of the irreducible components of Γk and Γk+1 is a
linear space and these components are linearly independent, i.e.
dimhΓk+1 i = dimhΓk i + dim Γk + 1 = (k + 1)(dim Γk + 1) − 1.
Proof. Since dim σk+1 (X) = (k + 1)(n + 1) − 1, we have dim σk (X) = k(n + 1) − 1. By Remark
3.2, X is weakly (k + 1)-defective. By [25, first two lines after Definition 1.2], X is not i-defective
for i = 1, . . . , k + 1.
By [27, Proposition 3.9] we know that both Γk and Γk+1 are equidimensional, that A is in
the smooth locus of Γk+1 , that B is in the smooth locus of Γk and that either Γk (resp. Γk+1 )
is irreducible or it has exactly k (resp. k + 1) irreducible components, each of them containing
exactly one point of B (resp. A).
Since B is general, we may assume that B ⊂ A. Thus Γk ⊆ Γk+1 . Since Γk is a proper subvariety
of X and Γk+1 contains B plus a general point of X, then Γk 6= Γk+1 . Thus if either Γk or Γk+1
is irreducible, then tk < tk+1 (where tk is defined in Definition 3.1).
Now assume that both Γk and Γk+1 are reducible, hence of type II, and that tk = tk+1 =: t.
Let B = {P1 , . . . , Pk } and A = {Pk+1 } ∪ B. Call Γi the component of Γk+1 passing through Pi ,
so that also Γk = Γ1 ∪ · · · ∪ Γk . Call Πi the linear span of Γi . Of course, by the generality of A,
all the Πi ’s have the same dimension. Assume that dim Πi > dim Γi and let us show that we get a
contradiction.
CONTACT LOCI AND IDENTIFIABILITY
5
For i = 2, . . . , k + 1 call Li the span of Π1 ∪ · · · ∪ Πi , which is also the span of Γ1 ∪ · · · ∪ Γi .
By [26, Proposition 2.5] we know that, for all i, dim(Li ) ≥ it + i − 1, since X is not i-defective.
Moreover by [27, Proposition 3.9] we know that dim(Lk ) = kt + k − 1 and dim(Lk+1 ) = (k + 1)t + k.
If dim(Lk−1 ) > (k − 1)t + k − 2, then by Grassmann formula, the intersection of Lk−1 and Πk has
dimension greater than (k − 1)t + k − 2 + dim(Πk ) − kt − k + 1 = dim(Πk ) − t − 1, so that, by
Grassmann formula again,
dim(Lk+1 ) < dim(Lk ) + dim(Πk+1 ) − dim(Πi ) + t + 1 = (k + 1)t + k,
a contradiction. Thus dim(Lk−1 ) = (k − 1)t + k − 2. Arguing by induction, we obtain that
dim(Li ) = it + i − 1 for i = 2, . . . , k + 1. In particular dim(Π1 ∩ Π2 ) = 2 dim(Π1 ) − 2t − 1 ≥ 0.
By the generality of A again, dim(Πi ∩ Πj ) = dim(Π1 ∩ Π2 ) for all i, j. Thus Π3 intersects L2 in
dimension at least dim Π1 ∩ Π2 ). It follows that L3 has dimension
dim(L3 ) ≤ dim(L2 ) + dim(Π3 ) − 2 dim(Π1 ) + 2t + 1 = (2t + 1) − dim(Πi ) + 2t + 1 < 3t + 2,
which yields the claimed contradiction.
4. Remarks on the secant degree of contact loci
We collect in this sections some remarks on the general tangential contact loci Γk of X, which
clarify the relations between the k-th secant degree of X and the k-th secant degree of a general
Γk . To do that, we need to distinguish the case where Γk is of type I (i.e. irreducible) from the
case where Γk is of type II.
Notation 4.1. For the rest of the section, let X ⊂ Pr be an integral and non-degenerate ndimensional variety and let k be a positive integer. Assume that X is weakly k-defective with
dim σk+n−1 (X) = (k + n − 1)(n + 1) − 1 < r.
Call Γk the tangential k-contact locus of X at a general set A = {P1 , . . . , Pk } of k points of X,
and call Πk the linear span of Γk .
The following technical result is the analogue of [26, Theorem 2.4] for the tangential contact
locus Γk . It is implicitly contained in part (iii) of [27, Proposition 3.9], but we prefer to make it
explicit here.
Lemma 4.2. For a general Q ∈ hAi ⊂ Πk , all subsets B ⊂ X of cardinality k such that Q ∈ hBi
are contained in the span of Γk .
Proof. Notice that, by the generality of A, Q is also a general point in the secant variety σk (X).
Since Q is general, the tangent space to σk (X) at Q is also the span of the tangent spaces to X
at the points of B, by the Terracini’s Lemma. The claim now follows from [27, proposition 3.9
(iii)].
Corollary 4.3. The k-secant degree of a projective variety X is equal to the k-secant degree of the
intersection of X with the span of its tangential k-contact locus.
When the general contact locus of X is of type I, we can give a more explicit statement.
Proposition 4.4. Assume that Γk is irreducible (i.e. type I). Assume the existence of a component
W of Πk ∩ X, different from Γk , such that Πk is tangent to X along W . Then for all a, b with
a > 0 and a + b = k, the join of the a-th secant variety of W with the b-th secant variety of Γk
cannot cover Πk .
Proof. Assume on the contrary that the join covers Πk . Since σk (X) has the expected dimension
and it is equal to the union of the spaces Πk ’s, when the k-th points P1 , . . . , Pk vary on X, then
for dimensional reasons there cannot be a subvariety Y ( X which contains all the respective
components W ’s. Thus, for P1 , . . . , Pk general, a general point Q of W is a general point of X.
Hence the contact variety of Q, P2 , . . . , Pk has the same properties of Γk . In particular Πk is
tangent to X along an irreducible variety of the same dimension of Γk , which contains P2 , . . . , Pk
and Q. Since Q ∈
/ Γk , this is excluded by the generality of P1 , . . . , Pk .
Corollary 4.5. Assume that Γk is irreducible (i.e. type I). Then the secant degree of X is equal
to the secant degree of the general contact locus Γk .
6
CONTACT LOCI AND IDENTIFIABILITY
The situation becomes completely different when the general Γk is of type II.
Remark 4.6. Assume that dim σk+n (X) = (k + n)(n + 1) − 1 < r, so that X is also weakly
(k + 1)-defective. Assume that a general tangential contact locus Γk+1 is formed by k + 1 linearly
independent linear spaces. Then also the general tangential contact locus Γk is formed by k linearly
independent linear spaces.
Indeed, with no loss of generality, as in the proof of Lemma 3.5, we may assume Γk ⊂ Γk+1 .
Thus Γk is of type II, and all its components are contained in one component of Γk+1 . But clearly
if one takes irreducible varieties Γi , i = 1, . . . , k, of the same dimension tk , such that each Γi spans
a linear space Πi of dimension > tk , and the linear spaces Πi ’s are linearly independent, then the
span of the union of the Γi ’s cannot have dimension kd + k − 1.
We point out the following example.
Example 4.7. ([18, Proposition 1.7]) Let X be the Grassmanniann G := G(P2 , P7 ) and let k = 3.
Then X is weakly 3-defective and, for a general choice of P1 , P2 , P3 ∈ X, Γ3 is the union of three
disjoint linear spaces L1 , L2 , L3 of dimension 3, such that Li contains Pi . The span Π3 of Γ3 is
also tangent to X along a linear space W = P5 , which misses the Pi ’s and intersects each Li along
a line.
One computes soon that the join of Li , Lj , W cannot fill Π3 , for obvious dimensional reasons.
It follows that the secant degree of X equals the secant degree of Γ3 , which is 1. Hence the generic
3-identifiability of G holds.
We can derive, as in the previous example a general statement for the case where the tangential
contact variety Γk is of type II.
Proposition 4.8. Assume that the general k-tangential contact locus Γk is of type II, i.e. it is the
union of components Γ1 , . . . , Γk . Assume that Πk is tangent to X along Γk ∪ Γ, where Γ 6⊂ Γk . If
hΓ1 ∪ · · · ∪ Γ̂I ∪ · · · ∪ Γk ∪ Γi =
6 Πk , where with Γ̂I we indicate any non empty subset of {Γ1 , . . . , Γk },
then the k-th secant degree of X is equal to the k-th secant degree of Γk .
Corollary 4.9. In the assumptions of the previous proposition, assume the Γi ’s are linear subspaces
of Pr , and linearly independent. Then X is j-identifiable for any j ≤ k.
Proof. Clearly if X is k-identifiable then it is also j-identifiable for every j ≤ k, hence we need
simply to show that X is k-identifiable. This last follows since the union of linearly independent
linear spaces is identifiable.
On the other hand, there are cases (where X is an almost unbalanced products of projective
spaces) in which Γk has type II and the previous proposition does not apply, moreover the k-th
secant degree of Γk is different from the k-th secant degree of X.
A first example is the following.
Example 4.10. Consider the Segre variety X of P2 ×P2 ×P5 and let k = 5. One computes that the
general tangential contact locus Γ5 consist of 5 linearly independent linear subspaces Γ1 ∪ · · · ∪ Γ5 ,
where each Γi is a P4 which passes through one of the points Pi .
The span Π5 of Γ5 is a P24 which is tangent to X along an extra Γ = P4 . Γ is linearly independent
from any subset of 4 among the Γi ’s. Thus the 5-th secant degree of Γ5 is 1, but the secant degree
of X equals the secant degree of Γ5 ∪ Γ, hence it is 6.
A similar situation holds for the almost unbalanced cases defined in Section 8 of [20].
5. Applications to concrete examples
We show below how Lemma 3.5 applies, thanks to the observations contained in Section 4,
to conclude the identifiability of some relevant cases of Segre products, Veronese varieties, SegreVeronese varieties and moment varieties.
Non-trivial Segre products X ⊂ Pr , X image of Pa1 × · · · × Pas , contain, for each j = 1, . . . , s,
linear spaces L1 , . . . , Lk , of dimension aj passing through k general points, which are linearly
independent, provided that k(aj + 1) is smaller than (r + 1). We do not know examples where
these spaces are contact loci of weakly defective Segre products. In any case, the previous lemma,
together with the analysis of weakly defective varieties contained in [26] and [29], yield the following.
CONTACT LOCI AND IDENTIFIABILITY
7
Theorem 5.1. Let X ⊂ Pr be an integral and non-degenerate n-dimensional variety and let k be a
positive integer. Assume that X is not uniruled by lines and that dim σk+n−1 (X) = (k + n − 1)(n+
1) − 1 < r (in particular X is not (k + n − 1)-defective). Then X is not weakly j-defective (hence
it is j-identifiable) for every j ≤ k.
Proof. If X is not weakly j-defective, then the claim follows by [26]. Assume that X is weakly
j-defective. Hence X is also weakly (j + i)-defective for i = 1, . . . , n − 1. Since X is not uniruled,
then Γj+i cannot be a union of linear spaces of positive dimension, so by Lemma 3.5 we have:
1 ≤ dim Γj < · · · < dim(Γj+n−1 ).
It follows that dim(Γj+n−1 ) ≥ n, a contradiction.
Theorem 5.1 is applicable to any embedding X ⊂ Pr with OX (1) ∼
= L⊗2 ⊗ M with L an ample
line bundle and M a nef line bundle. In particular it is applicable to any Veronese embedding of Pn ,
since in this case dim σi (X) is known for all i by a famous theorem of Alexander and Hirschowitz
([10]). However, the result that we get from Theorem 5.1 for symmetric tensors is weaker than the
results of [36, 31], where all the non-identifiable Veronese varieties are classified.
Let us turn to the study of Segre
Q varieties. Fix integer s ≥ 2, ni > 0, 1 ≤ i ≤ s. Set
n := n1 + · · · + ns and r := −1 + si=1 (ni + 1) and let X ⊂ Pr be the (n-dimensional) Segre
embedding of the multiprojective space Pn1 × · · · × Pns . Since X is uniruled by lines, we cannot
apply Theorem 5.1 directly. On the other hand, the structure of linear spaces contained in Segre
embeddings of multiprojective spaces is well known.
Remark 5.2. Let πi : X → Pni denote the projection of the Segre variety X into its i-th factor.
For any i = 1, . . . , s let OX (ǫi ) be the line bundle πi∗ (OPni (1)). Each OX (ǫi ) is a spanned line
bundle and OX (1) = ⊗si=1 OX (1). Let L ⊂ X be any linear space of positive dimension. Since
OL (1) ∼
= ⊕si=1 OX (ǫi )|L , there is j ∈ {1, . . . , s} such that OX (ǫi )|L ∼
= OL for all i 6= j, i.e. L is
contained in a fiber of πj . Fix an integer k ≥ 1 such that dim σk+1 (X) = (k + 1)(n + 1) − 1 < r
and assume that X is weakly (k + 1)-defective, with general tangential (k + 1)-contact locus Γk+1
equal to a union of linear spaces, say L1 , . . . , Lk , Lk+1 , of dimension ǫk > 0. Thus for each
i = 1, . . . , k + 1 there is a unique integer j(i) ∈ {1, . . . , s} such that πj(i) (Li ) is a point. We get a
function φk+1 : {1, . . . , k + 1} → {1, . . . , s} defined by the formula φk+1 (i) := j(i). The function
φk+1 cannot depend on the choice of k + 1 general points. Then, moving independently the points,
by monodromy we get that this function φk+1 must be invariant for the action of the full symmetric
group on {1, . . . , k + 1}. Thus j(i) = j(1) for all i.
Using the previous remark, we can prove the following.
Theorem 5.3. Let X be the Segre embedding of Pn1 ×· · ·×Pnk in Pr as above. Assume that for each
i ∈ {1, . . . , s} there is j ∈ {1, . . . , s} \ {i} such that nj = ni . Assume also that dim σk+n−1 (X) =
(k + n − 1)(n + 1) − 1 < r. Then X is not weakly j-defective (hence it is j-identifiable) for every
j ≤ k.
Proof. Assume that X is weakly j-defective. Thus it is also weakly (j + i)-defective, for i =
0, . . . , n − 1.
Each general contact locus Γj+i has dimension tj+i ≤ n − 1 and so there cannot be a strictly
increasing sequence of integer tj , . . . , tj+n−1 . By Lemma 3.5 there is an integer h ∈ {j, . . . , j +n−1}
such that Γh is a union L1 ∪ · · · ∪ Lh of h independent linear spaces of dimension th . Thus, by
Remark 5.2, there is an integer m ∈ {1, . . . , s} such that, for all i = 1, . . . , h, πm (Li ) is a linear
space of dimension dim Li , while πx (Li ) is a point for all x ∈ {1, . . . , s} \ {m}.
Notice that the general contact locus Γh is invariant for linear automorphisms of Pr preserving
X, in the sense that for any automorphism φ : Pr → Pr with φ(X) = X the set φ(Γh ) may be
taken as a general contact locus. Thus πx (h(Li )) is a point for each x ∈ {1, . . . , s} \ {m}, while
πm (h(Li )) is a linear space of dimension dim Πi . By assumption there is z ∈ {1, . . . , s} \ {j}
with nz = nj . Hence there is an automorphism ψ of X exchanging the m-th factor and the zth factor. This automorphism extends to a linear automorphism φ : Pr → Pr with φ(X) = X,
because the embedding X ֒→ Pr is induced by the complete linear system |OX (1, . . . , 1)| and
ψ ∗ (OX (1, . . . , 1)) ∼
= OX (1, . . . , 1). We would have that πz (Li ) is a linear space of dimension
dim Li , which yields a contradiction.
8
CONTACT LOCI AND IDENTIFIABILITY
The condition that for each i ∈ {1, . . . , s} there is j ∈ {1, . . . , s} \ {i} such that nj = ni restricts
the range of application of Theorem 5.3. However, the theorem applies to the products of many
copies of a given Pq , i.e. to the case of cubic tensors (provided that one knows the non-defectivity
of the corresponding Segre embedding).
In particular, in the case ni = 1 for all i, as an immediate corollary of [23] and Theorem 5.3, we
obtain the following result.
Corollary 5.4. Fix an integer s ≥ 5 and an integer k > 0 such that (k + s − 1)(s + 1) ≤ 2s − 1.
s
Let X ⊂ P2 −1 be the Segre embedding of (P1 )s . Then X is not weakly k-defective, hence it is
k-identifiable.
Thus, we obtain that binary tensors, i.e. tensors of type 2 × · · · × 2 (s times), are k-identifiable
for
2s − 1
− s + 1.
k≤
s+1
This bound is very near to k0 = ⌈2s /(s + 1)⌉ is the maximal integer for which k-identifiability can
hold.
As an immediate corollary of [9, Theorem 3.1] we get the following result.
Corollary 5.5. Take X = (Pm )s ⊂ Pr , m > 1, r + 1 = (m + 1)s . Set a := ⌊(m + 1)s /(ms + 1)⌋
and let ǫ be the only integer such that 0 ≤ ǫ ≤ m and a ≡ ǫ (mod m + 1). If either ǫ > 0 or
(k + ms)(sm + 1) ≤ (m + 1)s − 1, set a′ := a − ǫ. If ǫ = 0 and (ms + 1)a = (m + 1)s set a′ := a − 1.
For any positive integer k + ms − 1 ≤ a′ X is not weakly k-defective and hence it is k-identifiable.
Take X as in Corollary 5.5. Note that if k ≥ a + 2, then σa+2 (X) has expected dimension > r
and hence identifiability fails for σa+2 (X). In the special case s = 3, in Corollary 3.10 it is sufficient
to assume 3m(k + 3m − 1) < (m + 1)3 , because the secant varieties of (Pm )3 are not defective, by
[39].
Complete results on the defectivity of Segre products with more than 3 factors, except for the
case of many copies of P1 and for (Pm )3 ([39]) and an almost complete for [9], are not available.
As far as we know, the best procedure to find theoretically the non defectivity of Segre products
is the inductive method described in [5].
For Segre-Veronese varieties, which correspond to rank 1 partially symmetric tensors, we can
join Theorem 5.1 and Theorem 5.3 to get:
Theorem
5.6.
s ≥ 2, ni > 0, di > 0, 1 ≤ i ≤ s. Set n := n1 + · · · + ns and r =
Fix integer
Q di +n
r
i
.
Let
X
⊂
P
be
the (n-dimensional) Segre-Veronese embedding of the multiprojective
−1+
di
space Pn1 × · · ·× Pns , via the linear system Ld11 ⊗ · · ·⊗ Lds s , where each Li is the pull-back of OPni (1)
in the projection onto the i-th factor.
Assume that dim σk+n−1 (X) = (k + n − 1)(n + 1) − 1 < r. Assume also that either:
• di > 1 for all i; or
• for all i with di = 1 there exists j 6= i such that dj = 1 and ni = nj .
Then X is not weakly j-defective (hence it is j-identifiable) for every j ≦ k.
Proof. Notice that X is not uniruled by lines, unless di = 1 for some i. When some di is 1, one
can still apply the trick introduced in the proof of Theorem 5.3.
Example 5.7. C. Araujo, A. Massarenti and R. Rischter proved that if h ≤ a⌊log2 (d−1)⌋ , then
σh+1 (X) has the expected dimension (see [13, Theorem 1.1] for a more precise result).
For the case s = 2 there is a very strong conjecture and several results supporting it ([1, 2, 4]).
When s = 2 H. Abo and M. C. Brambilla give a list of 9 classes of secant varieties, which are known
to be defective and they conjecture that they are the only defective ones for s = 2 ([4, Conjecture
5.5]. See [8, Table 2] for the list of known (in 2010) non-defective cases when s = 2.
H. Abo and M. C. Brambilla also gave many examples of defective secant varieties when s = 3, 4
([3]). See [4, §5] for a general conjecture for the non-defectivity of Segre-Veronese varieties. If
di ≥ 2 for all i, then X is not uniruled by lines, so we may apply Theorem 5.1 to obtain the non
weak defectivity.
CONTACT LOCI AND IDENTIFIABILITY
9
Example 5.8. C. Améndola, J.-C. Faugère, K. Ranestad and B. Sturmfels ([11, 12]) studied the
Gaussian moment variety Gn,d ⊂ PN , N = n+d
− 1, whose points are the vectors of all moments
n
of degree ≤ d of an n-dimensional Gaussian distribution. We have dim Gn,d = n(n + 3)/2. They
proved that all the secant varieties of G1,d are non-defective, but σ2 (Gn,d ) is defective for all n ≥ 3
([12, Theorem 13 and Remark 15]). To apply Theorem 5.1 to G1,d we need prove that G1,d is not
uniruled by lines. If so, Theorem 5.1 may be applied to σk (G1,d ) for all k with 2(k +1) < d. We may
assume d ≥ 4. The variety G1,d ⊂ Pd is described in [11] and [12]; it is an arithmetically Cohen
Macaulay surface of degree d2 ([11, Proposition 1 and the last part of its proof and Corollary 2])
and its singular locus is a line L ⊂ G1,d ([12, Lemma 4]). Let ℓL : Pd \ L → Pd−2 the note the
linear projection of from L and let π : G1,d \ L → Pd−2 be the restriction of ℓL . Assume that
G1,d is ruled by line and call ∆ the one-dimensional family of lines giving the ruling. Since the
Grassmannian are complete, ∆ coves G1,d and hence the closure of the image of π would be an
integral and non-degenerate curve Y ⊂ Pd−2 such that a general L ∈ ∆ is mapped to a general
point of Y . The surface G1,d is covered by a family of degree d rational normal curves of Pd , each
of them tangent to L at a different point. Hence Y must be the rational normal curve of degree
d − 2 and X would
be contained in a degree d − 2 3-dimensional cone T ⊂ Pd with vertex L. Since
d
deg(G1,d ) = 2 > 3(d − 2), Bezout theorem gives that every cubic hypersurface containing G1,d
contains T , contradicting the fact that G1,d is cut out by cubics ([11, Proposition 1]).
For Segre products, the most general result that we can obtain with our techniques is resumed in
the following theorem. It can be applied to a wide class of Segre products, not covered by Theorem
5.3.
P
Theorem 5.9. Let X be the Segre embedding of a product Pn1 × · · · × Pnq in Pr , with n = ni =
dim(X). Let k be a positive integer. Assume that dim σk+n−1 (X) = (k + n − 1)(n + 1) − 1 < r (in
particular X is not (k + n − 1)-defective). Then X is j-identifiable for every j ≤ k.
Proof. Assume that X is weakly k-defective. Then, as in the proof of Theorem 5.1, there exists
some j, k ≤ j ≤ k + n − 2, such that the general tangential contact loci Γj and Γj+1 have the
same dimension. Thus, by Lemma 3.5 and its proof, the contact loci Γj and Γj+1 are formed by
a union of linearly independent linear spaces Li , say of dimension s. By Remark 5.2, the Li ’s are
contained in some component of the product, which, as in the proof of Theorem 5.6, is the same
for all i’s. Say for simplicity that the Li ’s sits in the first component. Call Y the Segre embedding
of Pn2 × · · · × Pnq and call Qi ∈ Y the point to which Li maps in the projection X → Y . The
Qi ’s are general points of Y . Since X is not j + 1 defective, the span M of the Li ’s has dimension
(j + 1)(s + 1) − 1. If M meets X only in the union of the Li ’s, which are linearly independent
subspaces, then by [19, S
Example 2.2] and Lemma 4.2 X is (j + 1)-identifiable. Otherwise there is
a point P ∈ (M ∩ X) \ Li . The point P cannot belong to the same fiber over Y of some Li , for
otherwise M contains the span of Li and P , and its dimension cannot be (j +1)(s+1)−1. Thus the
projection of M ∩ X to Y contains a point Q different from the Qi ’s. It follows that for a general
choice of j points Q1 , . . . , Qj in Y the span of the Qi ’s meets Y in a point Q different from the Qi ’s.
By the trisecant lemma ([14, p. 109], [24, Example 1.8], [25, Proposition 2.6]), this meansP
that the
codimension of Y in its span is at most j − 1. Since the codimension of Y is Π(ni + 1) − ni − 1
(where the product and the sum range from 2 to q), and moreover (j + 1)(n + 1) < Πqi=1 (a1 + 1),
after a short computation we get that:
a1 >
Πqi=2 (ai
+ 1) −
q
X
ai ,
i=2
i.e. Y is unbalanced in the sense of [20, Section 8]. Since we also have j ≥ a1 , it follows by [5,
Theorem 4.4] that either σj+1 (X) covers Pr , or X is (j + 1)-defective. In both cases we have a
contradiction.
We should remark that the inductive method of [5] is extended to the study of weak defectivity
in [20]. The results contained in [20] are, in several cases, improved by the results that one can get
via Theorem 5.9.
From the computational point of view, let us point out the following:
10
CONTACT LOCI AND IDENTIFIABILITY
Remark 5.10. For specific examples of Segre products X, the stacked hessian method introduced
in [30] can provide a positive answer to the k-identifiability of X. The stacked hessian method
requires the computations of the derivatives of the span of k tangent spaces at general points, when
the points move.
In order to prove the non (k + n)-defectivity of X, by the Terracini’s lemma, one has just to
compute the dimension of the span of k + n general tangent spaces (see e.g. [32, Section 6]). When
n is not too big with respect to k, this second method, together with Theorem 5.9, can provide
the k-identifiability of X with a faster procedure.
6. Positive characteristic
In this section we work over an algebraically closed field K of arbitrary characteristic.
Definition 6.1. Let X ⊂ Pr be an integral and non-degenerate variety. Set n := dim(X). X is
said to be very strange if for a general codimension n linear subspace M ⊂ Pr the set X ∩ M is
not in linearly general position in M , i.e. there is a hyperplane of M containing at least r − n + 2
points of M ∩ X ([42]).
We recall that X ⊂ Pr is not very strange if a general 0-dimensional section of X is in uniform
position ([14, Ch. 2],[42, Introduction]). Hence X ⊂ Pr is not very strange if either char(K) = 0
or X is non-singular in codimension 1 and r ≥ n + 3 ([42, Corollaries 1.6 and 2.2 and Theorem
0.1]). Thus none of the examples of the last section is very strange. X is not very strange if a
general curve section of X is not strange ([42, Lemma 1.1]) and in particular X is not very strange
if char(K) > deg(X).
Lemma 6.2. Let Y ⊂ Pr be an integral and non-degenerate variety, which is not very strange. Set
n := dim(Y ). Fix a general S ⊂ Y with ♯(S) ≤ r − n − 1. Then S is the set-theoretic base locus of
the linear system on Y induced by H 0 (Pr , IS (1)).
Proof. Let N ⊂ Pr be a general linear space of dimension n − r − 1. By Bertini’s theorem the
scheme N ∩ Y is a finite set of deg(Y ) points. Since Y is not very strange, the set N ∩ Y is in
linearly general position in N . Hence for any E ⊂ N ∩ Y with ♯(E) = ♯(S) ≤ n − r − 1, the
restriction of U := H 0 (Pr , IE (1)) to N ∩ Y has E as its set-theoretic base locus. Since N is a
linear space the restriction of U to Y has base locus contained in N ∩ Y . Thus E is the base locus
of the restriction of U to T . Since Y is integral and non-degenerate, a general subset A ⊂ Y with
cardinality at most r − n spans a general subspace of Pr with dimension ♯(A) − 1. Hence E is a
general subset of Y with cardinality ♯(S).
Proof of Theorem 1.1: Fix a general q ∈ σk (X). If dim σk (X) < kn + k − 1, then a dimensional
count gives that S(q, X) is infinite.
Hence to prove Theorem 1.1 it is sufficient to find a contradiction to the existence of A, B ∈
S(q, X) with A 6= B. Set E := B \ A ∪ B and Z := A ∪ E. We have h1 (Pr , IZ (1)) > 0, because
q ∈ hAi ∩ hBi and q ∈
/ hA ∩ Bi ([15, Lemma 1]), i.e. Z does not impose ♯(Z) independent conditions
to W . We may also assume that A and B are general in their irreducible component of the
constructible set S(q, X). Since q is general, we may assume that A (resp. B) is a general subset of
X with cardinality s, but of course A ∪ B is far from being general (in general). We identify X with
M , so that L and R are line bundles on X and OX (1) ∼
= L⊗R. Since A is general and ♯(A) ≤ dim V ,
the set of all f ∈ V vanishing at all points of A is a linear subspace of V with dimension dim(V )−s.
Hence h1 (Pr , IA (1)) = 0. Since A is general, by Lemma 6.2 there is a hypersurface T ∈ |V | with
A ⊂ T and E ∩ T = ∅. Let f ∈ V \ {0} be an equation of T . The multiplication by f induces
an isomorphism between H 0 (X, R) and the linear subspace W (−T ) := H 0 (X, IT ⊗ L ⊗ R) ∩ W
of all g ∈ W vanishing on T . Moreover, for any finite subset S ⊂ X \ T this isomorphism sends
H 0 (X, IS ⊗ R) onto W (−T − S) = {h ∈ W : h|T ∪S ≡ 0}. Since ♯(E) ≤ ♯(B) ≤ h0 (R) and
B is general in X, we have h0 (X, IE ⊗ R) = h0 (X, R) − ♯(E), i.e. E impose s independent
conditions to H 0 (X, R). Since T ∩ E = ∅, we get dim(W (−T − E)) = dim(W (−T )) − ♯(E). Hence
H 0 (Pr , IZ (1)) = dim W − ♯(Z), i.e. h1 (Pr , IZ (1)) = 0, a contradiction.
We give an example of application of Theorem 1.1.
CONTACT LOCI AND IDENTIFIABILITY
11
r
ns
n1
Example
Take X = P × · · · × P and positive integers di > 0 and let X ⊂ P , r =
Qs 6.3.
ni +di
−1+ i=1 ni . Write di = bi +ci with bi > 0, ci ≥ 0 and cj > 0 for some j. Set n := n1 +· · ·+ns .
If we fix any integer k > 0 such that
s
s
Y
n i + b i Y n i + ci
k ≤ min{−2 − n +
,
},
ni
ni
i=1
i=1
then Theorem 1.1 applies and proves the identifiability of a general q ∈ Pr of X-rank k.
References
[1] H. Abo, On non-defectivity of certain Segre-Veronese varieties. J. Symb. Comput. 45 (2010), no. 12, 1254–1269.
[2] H. Abo, and M. C. Brambilla, Secant varieties of Segre-Veronese varieties Pm × Pn embedded by O(1, 2).
Experim. Math. 18 (2009), no. 3, 369–384.
[3] H. Abo, and M. C. Brambilla, New examples of defective secant varieties of Segre-Veronese varieties. Collect.
Math. 63 (2012), no. 3, 287–297.
[4] H. Abo, and M. C. Brambilla, On the dimensions of secant varieties of Segre-Veronese varieties. Ann. Mat.
Pura Appl. 192 (2013), no. 1, 61–92.
[5] H. Abo, G. Ottaviani, and C. Peterson, Induction for secant varieties of Segre varieties. Trans. Amer. Math.
Soc. 361 (2009), no. 2, 767–792.
[6] H. Abo, G. Ottaviani, and C. Peterson, Non-defectivity of Grassmannians of planes. J. Alg. Geom. 21 (2012),
1–20.
[7] H. Abo, and N. Vannieuwenhoven, Most secant varieties of tangential varieties to Veronese varieties are nondefective. Trans. Amer. Math. Soc. 370 (2018), no. 1, 393–420.
[8] B. Ådlandsvik, Joins and higher secant varieties. Math. Scand. 62 (1987), 213-222.
[9] T. Aladpoosh, and H. Haghighi, On the dimension of higher secant varieties of Segre varieties Pn × · · · × Pn . J.
Pure Appl. Algebra 215 (2011), no. 5, 1040–1052.
[10] J. Alexander, and A. Hirschowitz, Polynomial interpolation in several variables. J. Alg. Geom. 4 (1995), no. 2,
201–222.
[11] C. Améndola, J.-C. Faugère, and B. Sturmfels, Moment varieties of Gaussian mixtures. J. Alg. Stat. 7 (2016),
no. 1, 14–28.
[12] C. Améndola, K. Ranestad, and B. Sturmfels, Algebraic Identifiability of Gaussian Mixtures. IMRN to appear.
Preprint arXiv:1612.01129 (2016).
[13] C. Araujo, A. Massarenti, and R. Rischter, On nonsecant defectivity of Segre-Veronese varieties. Preprint
arXiv:1611.01674 (2016).
[14] E. Arbarello, M. Cornalba, P. Griffiths, and J. Harris, Geometry of algebraic curves, Vol. I. Springer, Berlin Heidelberg - New York, 1985.
[15] E. Ballico and A. Bernardi, Decomposition of homogeneous polynomials with low rank, Math. Z. 271 (2012),
1141–1149.
[16] E. Ballico, A. Bernardi, M.V. Catalisano, and L. Chiantini, Grassman secants, identifiability, and linear systems
of tensors. Linear Algebra Appl. 438 (2013), 121–135.
[17] K. Baur, J. Draisma, and W. de Graaf, Secant dimensions of minimal orbits: computations and conjectures.
Experim. Math. 16 (2007), no. 2, 239–250.
[18] A. Bernardi, and D. Vanzo, A new class of non-identifiable skew symmetric tensors. Preprint arXiv:1606.04158
(2016).
[19] C. Bocci, and L. Chiantini, On the identifiability of binary Segre products. J. Algebraic Geometry 22 (2013),
1–11.
[20] C. Bocci, L. Chiantini, and G. Ottaviani, Refined methods for the identifiability of tensors. Ann. Mat. Pura
Appl. 193 (2014), no. 6, 1691–1702.
[21] A. Boralevi, A note on secants of Grassmannians. Rend. Istit. Mat. Univ. Trieste 45 (2013), 67–72.
[22] M. V. Catalisano, A. V. Geramita, and A. Gimigliano, Secant Varieties of Grassmann Varieties. Proc. Amer.
Math. Soc. 183 (2004), 633–642.
[23] M. V. Catalisano, A. V. Geramita, and A. Gimigliano, Secant varieties of P1 × · · · × P1 (n-times) are NOT
defective for n ≥ 5. J. Alg. Geom. 20 (2011), 295–327.
[24] L. Chiantini, and C. Ciliberto, A few remarks on the lifting problem. Astérisque 218 (1993), 95–109.
[25] L. Chiantini, and C. Ciliberto, Weakly defective varieties. Trans. Amer. Math. Soc. 454 (2002), no. 1, 151–178.
[26] L. Chiantini, and C. Ciliberto, On the concept of k-secant order of a variety. J. London Math. Soc. 73 (2006),
no. 2, 436–454.
[27] L. Chiantini, and C. Ciliberto, On the dimension of secant varieties. J. Europ. Math. Soc. 73 (2006), no. 2,
436–454.
[28] L. Chiantini, M. Mella, and G. Ottaviani, One example of general unidentifiable tensors. J. Algebr. Stat. 5
(2014), no. 1, 64–71.
[29] L. Chiantini, and G. Ottaviani, On generic identifiability of 3-tensors of small rank. SIAM J. Matrix Anal.
Applic. 33 (2012), 1018–1037.
[30] L. Chiantini, G. Ottaviani, and N.Vanniuwenhoven, An algorithm for generic and low-rank specific identifiability
of complex tensors. SIAM J. Matrix Anal. Applic. 35 (2014), 1265–1287.
12
CONTACT LOCI AND IDENTIFIABILITY
[31] L. Chiantini, G. Ottaviani, and N.Vanniuwenhoven, On identifiability of symmetric tensors of subgeneric rank.
Trans. Amer. Math. Soc. 369 (2017), no. 2, 4021–4042.
[32] L. Chiantini, G. Ottaviani, and N.Vanniuwenhoven, Effective criteria for specific identifiability of tensors and
forms. SIAM J. Matrix Anal. Applic. 38 (2017), 656–681.
[33] I. Domanov, and L. De Lathauwer, On the uniqueness of the canonical polyadic decomposition of third-order
tensors–part I: Basic results and unique- ness of one factor matrix. SIAM J. Matrix Anal. Appl. 34 (2013),
855–875.
[34] I. Domanov, and L. De Lathauwer, On the uniqueness of the canonical polyadic decomposition of third-order
tensors–part II: Uniqueness of the overall decomposition. SIAM J. Matrix Anal. Appl. 34 (2013), 876–903.
[35] I. Domanov, and L. De Lathauwer, Generic uniqueness conditions for the canonical polyadic decomposition and
INDSCAL. SIAM J. Matrix Anal. Applic. 36 (2015), no. 4, 1567-1589.
[36] F. Galuppi, and M. Mella, Identifiability of homogeneous polynomials and Cremona Transformations. Preprint
arXiv:1606.06895 (2016).
[37] J. B. Kruskal, Three-way arrays: rank and uniqueness of trilinear decompositions, with application to arithmetic
complexity and statistics. Linear Algebra Appl. 18 (1977), no. 2, 95–138.
[38] J.M. Landsberg, Tensors: Geometry and Applications Graduate Studies in Mathematics, Amer. Math. Soc.
Providence, 128 (2012).
[39] T. Lickteig, Typical tensorial rank. Linear Algebra Appl. 69 (1985), 95–120.
[40] A. Massarenti, and R. Rischter, Non-secant defectivity via osculating projections. Ann. Scuola Norm. Sup. to
appear. Preprint arXiv:1610.09332v1 (2016).
[41] M. Mella, Singularities of linear systems and the Waring problem. Trans. Amer. Math. Soc. 358 (2006), no. 12,
5523–5538.
[42] J. Rathmann, The uniform position principle for curves in characteristic p. Math. Ann. 276 (1987), no. 4,
565–579.
[43] N. D. Sidiropoulos, and R. Bro, On the uniqueness of multilinear decomposition of N-way arrays. J. Chemometrics 14 (2000), 229–239.
(Edoardo Ballico) Dipartimento di Matematica, Univ. Trento, Italy
E-mail address: [email protected]
(Alessandra Bernardi) Dipartimento di Matematica, Univ. Trento, Italy
E-mail address: [email protected]
(Luca Chiantini) Dipartimento di Ingegneria dell’Informazione e Scienze Matematiche, Univ. Siena,
Italy
E-mail address: [email protected]
| 0math.AC
|
A High-Performance HOG Extractor on FPGA
Vinh Ngo
Arnau Casadevall
Marc Codina
Department of Microelectronics and
Electronics Systems
Spain
[email protected]
Department of Microelectronics and
Electronics Systems
Spain
[email protected]
Department of Microelectronics and
Electronics Systems
Spain
[email protected]
David Castells-Rufas
Jordi Carrabina
Department of Microelectronics and
Electronics Systems
Spain
[email protected]
Department of Microelectronics and
Electronics Systems
Spain
[email protected]
ABSTRACT
Pedestrian detection is one of the key problems in emerging selfdriving car industry. And HOG algorithm has proven to provide
good accuracy for pedestrian detection. There are plenty of research
works have been done in accelerating HOG algorithm on FPGA
because of its low-power and high-throughput characteristics. In
this paper, we present a high-performance HOG architecture for
pedestrian detection on a low-cost FPGA platform. It achieves a
maximum throughput of 526 FPS with 640x480 input images,
which is 3.25 times faster than the state of the art design. The
accelerator is integrated with SVM-based prediction in realizing a
pedestrian detection system. And the power consumption of the
whole system is comparable with the best existing implementations.
KEYWORDS
Histogram of gradients, HOG extractor, FPGA HOG accelerator
1 INTRODUCTION
Pedestrian detection is a safety critical application on
autonomous cars. There are two main approaches to implement
pedestrian detection systems. On one hand, the detection algorithm
relies on all input image pixels. This approach uses deep learning
method and it requires costly computing platforms with not only
many processing cores but also large memory bandwidth and
capacity. On the other hand, only extracted features from the image
are input to the detection algorithm. This approach using HOG
(Histogram of Gradients) [1] has proven to have good accuracy in
detection [2]. While requiring less memory capacity, it is still a
computing-intensive algorithm, which needs a low latency and
high-throughput platform. FPGA, therefore, comes as suitable
solution thanks to its capability in parallel processing. More
importantly, FPGAs potentially have better energy efficiency in
comparison with alternative platforms such as CPU and GPU.
In this paper, we design and implement a hog feature extractor
on a low-cost FPGA device, targeting at high throughput and low
power consumption. This work is based on our previous work in
[3]. There are several improvements to help achieve a highperformance design. First, the fixed-point number is used to
represent values other than the integer number, which apparently
increases the feature’s accuracy with the cost of computational
complexity. Secondly, a pipeline for normalizing cell features to
take advantages of hardware’s capability in pipeline and parallel
execution. The output HOG normalized features are transferred to
the HPS (Hard Processor System) for prediction process. Third,
instead of buffering input images before extracting, which costs
memory, input pixels are processed directly from the sensor by a
pipeline. And finally, we optimize the pipeline design so as to
achieve the highest throughput. The HOG extractor design can
work at a maximum clock frequency of 162 MHz and provide a
throughput of 526 FPS, the highest throughput in the state of the
art. The design is then integrated into a heterogeneous system with
SVM-based prediction software. The energy efficiency is
comparable to the most efficient implementations.
The paper is outlined as follows. An overview of the original
HOG algorithm is described in section 2. Section 3 discusses
related works regarding FPGA implementations of real-time HOG
extractor. Section 4 presents our architectural design in detail. The
experimental results and discussions are shown in section 5.
Finally, the conclusions are presented in section 6.
2 HOG OVERVIEW
The HOG algorithm consists of two main steps: gradient
computation and histogram generation.
HIP3ES, January 2018, Manchester, United Kingdom
To compute the gradient of a pixel (x,y), first, we need to
calculate the intensity difference of its two pairs of neighbor pixels
in horizontal and vertical directions following the Equation (1) and
(2) respectively.
( , ) = ( + 1, ) − ( − 1, ) (1)
( , ) = ( , + 1) − ( , − 1) (2)
Then, the magnitude and the orientation of the gradient
at pixel(x,y) are computed by Equation (3) and (4).
| ( , )| =
( , ) +
( , )=
( , )
( , )
( , )
(3)
(4)
Figure 1: Histogram is generated cell by cell
Having the gradients, the histogram is generated cell by cell.
Each cell has a size of 8x8 pixels. Therefore, a cell consists of 64
pairs of magnitude and orientation gradient values. Depending on
the associated orientations, the magnitude gradients are
accumulated to the corresponding bins. A cell histogram with nine
bins is illustrated in Figure 1. Figure 2 describes in detail how the
orientation of the gradient is quantized into a range of 9 bins using
the scale from 0 to 180o.
Figure 2. Dividing into 9 bins from 0 to 180o
The magnitude G, in this example, should be accumulated to
bin 2 because its orientation is approximately 30o. For more
accuracy, G will be accumulated fairly between adjacent bins
depending on its exact orientation.
3 RELATED WORKS
To our knowledge, the works in [4], [5] presented the first
implementations of HOG extractors on FPGAs. In [4], the HOG
extractor is shown to have a good latency of only 312 µs. However,
V. Ngo et al.
this design does not include the normalization module and it
simplifies the computational process by using integer numbers. In
[6], the authors proposed to process the pixel data at twice the pixel
frequency and normalize the block histograms using L1-norm so
that the available resources are efficiently used and can address
parallel computing of multiple scales. With an input image of
1920x1080, the design achieves high speed with a latency of only
150 µs. But it is not clarified in the paper what this latency is about.
Similarly, the design used some kinds of frame buffer before HOG
processing module, which costs memory. Energy consumption of a
HOG-based detection system on FPGA is first reported in [2]. In
this work, the authors try to reduce the bit-width of the fixed-point
representation to boost the performance. With a 640x480 frame size
and a 13-bit fixed-point representation, the energy efficiency of the
HOG extractor module is 0.54J/Frame. Anyway, the design
leverages a costly hardware system with four FPGA devices and
each device has 16 64-bit memory channels. The memory space for
those 4 FPGA devices is 128 GB.
Another approach is presented in [7], in which the authors
investigate the cell size and number of histogram bins that provide
better performance. In this implementation, all the process of the
detection system is integrated into an FPGA device. With a
negligible loss in accuracy, the best set of parameters provides a
frame rate of 42.7fps and high energy-efficiency of only
0.451J/Frame. A detailed description of HOG implementation on
FPGA is presented in [8], which achieves a high processing speed
at 40fps, with 1920x1080 input image size. Interestingly, in [9],
HOG algorithm is analyzed on a heterogeneous system, including
CPU, GPU, and FPGA. Based on multiple configuration
experiments, the authors concluded that FPGA is best suited for
histogram extraction and classification tasks in the whole detection
flow because it produces a good trade-off between power and
speed.
Recently, our previous work is published in [3], which
simplifies the computing by using integer numbers. We achieved
high throughput in HOG extracting process by buffering the input
image. Besides, a look-up table is used to store the results of the
square root and arctan computations. This approach heavily
consumes on-chip memory. A low-complexity implementation of
HOG-based pedestrian detection is presented recently in [10].
Instead of the original HOG, the authors proposed the use of
histogram of significant gradients, and the hardware is, therefore,
less complex. In addition, hardware resource usage is optimized by
reducing the number of bits representing the intermediate values
during computation processes. Besides, the authors avoid using
complex representation numbers as well as DSP operations by precalculated values and simplification techniques.
4 IMPLEMENTATION
We implement the whole system in Terasic’s DE1-SOC board.
The system block diagram is shown in Figure 3. It includes the
hardware components such as the image sensor, the HOG pipeline,
the Hard Processor System (HPS), and other supporting modules.
A High-Performance HOG Accelerator on FPGA
HIP3ES, January 2018, Manchester, United Kingdom
four cells and L2 normalization [1] is chosen for the sake of
accuracy and simplicity.
Figure 5 describes our hardware line buffers that allow the
HOG module to compute the luminance difference Gx and Gy
between neighbor pixels in vertical and horizontal directions. This
design supports processing pixels on every clock cycle, which
means that the performance of the design can be boosted if input
pixels come at every clock cycle.
Figure 3. System diagram
Images from the sensor, after being filtered by the Bayern
Pattern, are transferred directly to both the HOG pipeline module
and the pixel FIFO. The pixel FIFO is necessary for later showing
the original image on the VGA. A custom Avalon master interface
is created to get pixels from this FIFO and write to the 1GB external
SDRAM controlled by the HPS.
The image sensor is configured through an I2C interface for
some key parameters such as image size, pixel clock. The Bayer
pattern filter module takes raw input pixels and calculates the three
colors pixel values. After that, the grayscale pixel value is
generated to provide the HOG extractor and the HPS for real-time
visualization.
The HOG extractor module is a long pipeline that generates the
normalized hog features. Our best implementation in throughput
used a 155 stages pipeline. The features are then written to the HPS
memory by a DMA (Direct Memory Access). A Python code
running on HPS will read these features out for predicting the
present of pedestrians. The detailed architecture inside the HOG
pipeline is presented in Figure 4.
Figure 4. HOG extractor block diagram
First of all, luminance differences Gx and Gy (Eq. 1,2) are
calculated by the DELTAXY module. These are 9 bit signed
integers. We used the vector translate function in CORDIC IP to
compute the magnitude and the orientation gradients. Both of them
are fixed-point numbers. To achieve 2 digits after the decimal point
accuracy, we choose to represent the orientation gradient by 13
fractional bits. Thus, the number of fractional bits for the magnitude
gradient is six, according to the configuring requirement of
CORDIC IP. Depending on the orientation gradient, the magnitude
gradient of each pixel will vote to appropriate bins. The
AGGREGATE module adds 64 histogram values of 64 pixels in a
cell bin by bin to output the final cell features. Finally, cell features
are block-wise contrast normalized. In this design, each block has
Figure 5. Pixel line buffers
The depth of each buffer corresponds to the row size of the
input image, in our case 640. The luminance differences, Gx and
Gy, at pixel P_11 are calculated using P_21 and P_01 for the
vertical direction, and P_10 and P_12 for horizontal direction as in
Equation (5) and (6).
(1,1) =
(1,1) =
−
−
(5)
(6)
Following the original HOG algorithm in [1], the final HOG
feature is extracted from every cell of 8x8 pixel size. And the
orientation is divided into 9 bins from 0 to 180º. In our case, with
the 640x480 image size, the final HOG feature is a vector of
80x60x9 dimension. The HOG module processes in a pipeline
approach every 8 continuous pixels in a row of a cell. And it
generates a partial hog vector with 9 bins aggregating 8 magnitude
gradient values. These partial hog vectors are put in a line buffer as
shown in Figure 6. Only 80 partial cell hogs are needed to be stored
so as to minimize the memory usage without stalling the pipeline.
In order to generate the full HOG feature for a cell, it is necessary
to aggregate 8 partial cell hogs from 8 different rows. The
cell_hog_valid signal will be active only if all the partial cell hogs
are fully collected.
Figure 6. Partial cell hog line buffer
HIP3ES, January 2018, Manchester, United Kingdom
The normalization of the cell histogram is done following the
equation in (7). In the equation, v is the cell hog features in the
block, and ǁvǁ2 is the L2-normalization of all the cell hog features
in the block. A small constant, Ɛ, is added to avoid dividing by zero.
As illustrated in Figure 4, each bin of the normalized hog
feature is represented by 32 bits. This is not the final HOG feature
value and it is represented in floating-point format. We only do the
conversion from the fixed-point format to the floating-point format
for the final step. All the intermediate results are calculated by
integer and fixed-point numbers depending on the specific task.
=
(7)
ǁ ǁ
Ɛ
We used the ModelSim simulator and a C golden model of the
HOG to verify the HOG extractor design. At the system level, we
build a heterogeneous system as in Figure 7. An Ubuntu
distribution runs on the HPS. We used Qsys (Quartus-II) to create
the system-on-chip FPGA architecture.
V. Ngo et al.
We have trained and tested several models with different
configurations using the INRIA Person Dataset [12] to achieve a
maximum yield in terms of accuracy. We have tested our model not
only with INRIA test dataset but also images from the camera
sensor to have a more generalized model.
5 RESULTS AND DISCUSSIONS
The hardware design of the HOG extractor is compared stepwise with a reference C model which uses floating point operations.
And the output results of the two implementations, which is the
final normalized hog features, have an average difference of only 2
units at the second digit after the decimal point, which corresponds
to 2% accurate because the range of a normalized feature value is
from 0 to 1.
Table 1 reports the key compilation results for the HOG
extractor and the heterogeneous system. For the HOG extractor, we
provide two versions. The normal version works at 49 MHz and
costs fewer hardware resources. The optimized one targets high
throughput applications. It can work at 162 MHz clock frequency.
Running at this frequency, the design only requires 2% more in
ALMs and 9 more DSP blocks. And the number of registers is
increased by nearly 20%. The on-chip memory usage is almost the
same between the two versions. On the other hand, the
heterogeneous system, as illustrated in Figure 7, is a system on chip
design which includes the HOG extractor module as a hardware
accelerator.
Table 1: Compilation report for Cyclone V device
Figure 7. HOG feature extraction system
The normalized HOG feature is written to the external DDR3
SDRAM memory by a DMA through the f2h_axi_slave bridge. A
custom Avalon bus master is created in Qsys to send image pixels
also to the DDR3 memory. These two memory locations are set to
be dedicated for FPGA. These transfer methods provide good
system performance because data are transmitted in parallel with
the HPS’s CPU execution.
In the opposite direction, the pixels in the memory and
detection results are sent to the VGA controller to visualize in real
time.
To extract all the data (pixel image and hog features) we used
a Python C/C++ API to write an extension module for binding the
communicating between the HPS memory and the Python
interpreter. That is useful to compute, in a high-level manner, the
detection and classification tasks using Machine Learning
techniques.
Our Python script is capable of reading both the frame and
HOG vector provided by the board by reading in C their respective
memory registers and, then, binding them with Python C/C++ API
as Numpy [11] arrays.
Design
Block
memory
Kbits
Logic
(in
ALMs)
DSP Registers Fmax
blocks
(MHz)
HOG extractor
324
(8%)
7,922
(25%)
65
(75%)
14,787
49
Optimized
HOG extractor
326
(8%)
8,610
(27%)
74
(85%)
17,697
162
12,138
(38%)
65
(75%)
21,715
69
Heterogeneous 437 (11%)
system
Other than the clock frequency, a design’s output throughput
also depends on the input throughput. Our design supports up to
one input pixel every clock cycle. It means that if the input pixel
clock is 162 MHz, our design throughput reaches 526 FPS as shown
in Table 2. It is worth noting that the figures in this table are based
on the HOG extractor design, not the heterogeneous system.
According to Table 2, our design speeds up 3.25x over the
equivalent design recently published [10]. In order to compare
designs that have different input frame size, we use pixels per
second unit. In this scale, our design also supports the highest
performance. Our implementation achieves 22% higher than the
work in [6]. Another interesting measurement is the number of
pixels per clock period. This number reflects the throughput of the
A High-Performance HOG Accelerator on FPGA
HIP3ES, January 2018, Manchester, United Kingdom
design without taking into account the clock frequency. Our design
achieves nearly 1 pixel on every clock, and it is the highest
throughput according to Table 2.
Regarding FPGA resource, our design is optimized for memory
usage and therefore consumes least memory resource among the
designs in Table 2. The reason for this is that our pipeline works on
every input pixel and there is not any buffer for input frames. For
other resources, implementation in [6] is the most efficient. It
targeted at low resource utilization by simplifying some
computational operations. In the voting part, magnitudes are voted
to only one unique bin without interpolation. Furthermore, all the
calculations use integer numbers. On the other hand, the
implementation in [13] is quite equivalent to ours. It consumes less
LUTs, DSPs and registers thanks to a Look-up Table for storing in
arctan values.
In terms of energy efficiency, we measure for the entire
detection system. Following Table 3, our design consumes
0.82J/Frame which is quite higher than results from [2] and [7].
Those are the best designs in the state-of-the-art in terms of energy
efficiency. The interesting point is that our design is better in power
consumption. Thus, the reason for lower energy efficiency is
because of the system detection rate. Our system only supports 11
fps although the HOG extractor can work at multiple times higher
speed. We believe that our system throughput will improve
significantly if we implement the classification task on FPGA, and
therefore the energy consumption is reduced accordingly. We
observed that sliding detection window, which is the most timeconsuming task in the classification process, can be done in parallel
in hardware but not in HPS’s software. And we still have a large
room to implement a classifier in hardware since the latency of the
hardware part is only 6.16 ms, which is 14% of the design in [2].
Last but not least, Table 3 shows that our design’s hardware
resource is significantly small in comparison to the others.
Table 2: Comparison of performance between different FPGA implementations
Design Frame size
FPGA
Max
FPS
frequency
(MHz)
Pixels per Second
(FPS*Frame Size)
Pixels per clock period
(Pixels per
Second/Frequency)
FPGA resources
Memory
(Kb)
LUTs DSPs Registers
[4]
800x600
Spartan 3
63
30
14,400,000
0.229
1080
42,435
-
-
[6]
1920x1080
Virtex 7
266
64
132,710,400
0.5
936
3,924
12
3,642
[10]
640x480
Cyclone IV
117.8
162
49,766,400
0.422
-
-
-
-
125
60
124,416,000
0.995
432
7,226
26
12,462
162
526
161,587,200
0.997
326
8,610
74
17,697
[13]
Ours
1920x1080 ZynQ 7000
640x480
Cyclone V
Table 3: Comparison of energy efficiency
[2]
Frame size
FPGA
Freq.
(MHz)
Latency
640x480
Virtex 6
150
44 ms
37
0.54
68.2
Virtex 7
266
-
19
0.451
Cyclone V
50
6.16 ms
9
0.82
[7] 1920x1080
Us
640x480
Power Energy FPS Memory
(W) (J/Frame)
(Kb)
# of
LUTs
DSPs
# of
FFs
13,738
184,953
190
208,666
42.7
4,079
30,360
364
48,576
11
437
12,138
65
21,715
6 CONCLUSION
A high-performance HOG feature extractor is implemented on
a low-cost FPGA device. Fixed-point representation is employed
for achieving approximately 2% different in comparison with the
floating-point golden model. The HOG extractor design, which
supports 526 FPS, can be a well-fitted IP in high-performance
pedestrian detection systems.
The energy consumption of the whole detection system is
0.82J/Frame, which is among the good existing solutions.
However, there is still room for future improvements to increase
the detection throughput of our system design and lower down the
energy consumption accordingly. This is can be done by
implementing a classifier on chip. Thus, the classification task can
slide the detection window through the HOG frame in parallel.
Furthermore, the classification task also can be started early and
pipelined together with the hog feature calculating process.
ACKNOWLEDGMENTS
HIP3ES, January 2018, Manchester, United Kingdom
This work was supported by Spanish projects TEC2014-59679C2-2.
REFERENCES
[1] N. Dalal and W. Triggs, “Histograms of Oriented Gradients for Human
Detection,” 2005 IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit.
CVPR05, vol. 1, no. 3, pp. 886–893, 2004.
[2] X. Ma, W. A. Najjar, and A. K. Roy-Chowdhury, “Evaluation and acceleration
of high-throughput fixed-point object detection on FPGAS,” IEEE Trans.
Circuits Syst. Video Technol., vol. 25, no. 6, pp. 1051–1062, 2015.
[3] V. Ngo, A. Casadevall, M. Codina, D. Castells-Rufas, and J. Carrabina, “A
pipeline hog feature extraction for real-time pedestrian detection on FPGA,” in
2017 IEEE East-West Design & Test Symposium (EWDTS), 2017, pp. 1–6.
[4] S. Bauer, U. Brunsmann, and S. Schlotterbeck-macht, “FPGA Implementation of
a HOG-based Pedestrian Recognition System,” in MPC-Workshop, 2009, no.
July.
[5] R. Kadota, H. Sugano, M. Hiromoto, H. Ochi, R. Miyamoto, and Y. Nakamura,
“Hardware architecture for HOG feature extraction,” IIH-MSP 2009 - 2009 5th
Int. Conf. Intell. Inf. Hiding Multimed. Signal Process., no. 3, pp. 1330–1333,
2009.
[6] M. Hahnle, F. Saxen, M. Hisung, U. Brunsmann, and K. Doll, “FPGA-Based realtime pedestrian detection on high-resolution images,” IEEE Comput. Soc. Conf.
V. Ngo et al.
Comput. Vis. Pattern Recognit. Work., pp. 629–635, 2013.
[7] A. Khan, M. Umar, K. Khan, M. Bilal, and C. Kyung, “Hardware Architecture
and Optimization of Sliding Window Based Pedestrian Detection on FPGA for
High Resolution Images by Varying Local Features,” pp. 142–148.
[8] J. Rettkowski, A. Boutros, and D. Gohringer, “Real-time pedestrian detection on
a xilinx zynq using the HOG algorithm,” 2015 Int. Conf. ReConFigurable
Comput. FPGAs, ReConFig 2015, 2016.
[9] C. Blair, N. M. Robertson, and D. Hume, “Characterising a Heterogeneous
System for Person Detection in Video using Histograms of Oriented Gradients:
Power vs. Speed vs. Accuracy,” IEEE J. Emerg. Sel. Top. Circuits Syst., vol. 3,
no. 2, pp. 236–247, 2013.
[10] M. Bilal, A. Khan, M. U. K. Khan, and C.-M. Kyung, “A Low-Complexity
Pedestrian Detection Framework for Smart VideoSurveillance Systems,” IEEE
Trans. CIRCUITS Syst. VIDEO Technol., vol. 27, no. 10, pp. 2260–2273, Oct.
2017.
[11] S. Van Der Walt, S. C. Colbert, and G. Varoquaux, “The NumPy array: a structure
for efficient numerical computation,” Feb. 2011.
[12] “INRIA
Person
dataset.”
[Online].
Available:
http://pascal.inrialpes.fr/data/human/. [Accessed: 16-Dec-2017].
[13] M. Hemmati, M. Biglari-Abhari, S. Berber, and S. Niar, “HOG Feature Extractor
Hardware Accelerator for Real-Time Pedestrian Detection,” in 2014 17th
Euromicro Conference on Digital System Design, 2014, pp. 543–550.
| 1cs.CV
|
Jolie Static Type Checker: a prototype
Daniel de Carvalho1 , Manuel Mazzara1 , Bogdan Mingela1
Larisa Safina1 , Alexander Tchitchigin2 , Nikolay Troshkov1
1
Innopolis University, Russia
Typeable.io LLC, Russia
arXiv:1702.07146v5 [cs.SE] 18 Oct 2017
2
Abstract. Static verification of a program source code correctness is an important
element of software reliability. Formal verification of software programs involves proving that a program satisfies a formal specification of its behavior. Many languages
use both static and dynamic type checking. With such approach, the static type
checker verifies everything possible at compile time, and dynamic checks the remaining. The current state of the Jolie programming language includes a dynamic type
system. Consequently, it allows avoidable run-time errors. A static type system for
the language has been formally defined on paper but lacks an implementation yet.
In this paper, we describe a prototype of Jolie Static Type Checker (JSTC), which
employs a technique based on a SMT solver. We describe the theory behind and the
implementation, and the process of static analysis.
1
Introduction
The microservice architecture is a style inspired by service-oriented computing that promises
to change the way in which software is perceived, conceived and designed [20]. The trend
of migrating monolithic architectures into microservices to reap benefits of scalability is
growing fast today [10,11]. Jolie [23] is the only language natively supporting microservice
architectures [8] and, currently, has dynamic type checking only.
Static type checking is generally desirable for programming languages improving software
quality, lowering the number of bugs and preventing avoidable errors. The idea is to allow
compilers to identify as many issues as possible before actually run the program, and therefore avoid a vast number of trivial bugs, catching them at a very early stage. Despite the
fact that, in the general case interesting properties of programs are undecidable [27], static
type checking, within its limits, is an effective and well established technique of program
verification. If a compiler can prove that a program is well-typed, then it does not need to
perform dynamic safety checks, allowing the resulting compiled binary to run faster.
A static type system for Jolie has been exhaustively and formally defined only on paper
[25], but still lacks an implementation. The obstacles of programming in a language without
a static type analyzer have been witnessed by Jolie developers, especially by newcomers.
However, implementing such system is a non trivial task due to technical challenges both
of general nature and specific to the language. In this paper, we introduce and describe
the Jolie Static Type Checker (JSTC), building on top of the previous work on the Jolie
programming language [23]. Our approach follows the formal derivation rules as defined in
[25]. The project is built as a Java implementation of source code processing and verification
via Z3 SMT solver [9] and it has to be intended as our community contribution to the Jolie
programming language [4].
Section 2 recalls the basic of Jolie and section 3 discusses related work. The description
of the static type-checking and the system architecture can be found in Section 4, while
Section 6 draws conclusive remarks and discusses open issues.
2
Background
Microservices [12] is an architectural style evolved from Service-Oriented Architectures [15].
According to this approach, applications are composed by small independent building blocks
2
Daniel de Carvalho et al.
that communicate via message passing. These composing parts are indeed called microservices. This paradigm has seen a dramatic growth in popularity in recent years [24]. Microservices are not limited to a specific technology. Systems can be built using a wide range
of technologies and still fit the approach. In this paper, however, we support the idea that
a paradigm-based language would bring benefit to development in terms of simplicity and
development cost.
Jolie is the first programming language constructed above the paradigm of microservices:
each component is autonomous service that can be deployed separately and operated by running in parallel processes. Jolie comprises formally-specified semantics, inspired by process
calculi such as CCS [21] and the π-calculus [22]. As for practical side, Jolie was inspired by
standards for Service-Oriented Computing such as WS-BPEL [2] and the attempts of formalizing it [18]. The composition of both theoretical and practical aspects allows Jolie to be
the preferred candidate for the application of modern research methodologies, e.g. runtime
adaptation, process-aware web applications, or correctness-by-construction of concurrent
software.
The basic abstraction unit of Jolie is the microservice [12]. It is based on a recursive
model where every microservice can be easily reused and composed for obtaining, in turn,
other microservices. Such approach allows distributed architecture and guarantees simple
management of all components, which reduces maintenance and development effort. Microservices communicate and work together by sending messages to each other. In Jolie,
messages are represented in tree structure. A variable in Jolie is a path in a data tree and
the type of a data tree is a tree itself. Equality of types must therefore be handled with
that in mind. Variables are not declared, wherefore the manipulation of the program state
must be inferred. Communications are type checked at runtime, when messages are sent or
received. Type checking of incoming messages is especially relevant, since it could moderate
the consequences of errors.
The Jolie language is constructed in three layers: The behavioural layer operates with
the internal actions of a process and the communication it performs seen from the process
point of view, the service layer deals with the underlying architectural instructions and the
network layer deals with connecting communicating services.
Other workflow languages are capable of expressing orchestration of (micro)services the
same way Jolie can do, for example WS-BPEL [2]. WS-BPEL allows developers to describe
workflows of services and other communication aspects (such as ports and interfaces), and it
has been also shown how dynamic workflow reconfiguration can be expressed [17]. However,
WS-BPEL has been designed for high-level orchestration, while programming the internal
logic of a single micro-service requires fine-grained procedural constructs. Here it is were
Jolie works better.
3
Related work
The implementation of a static type checker for Jolie is part of a broader attempt to enhance
the language for practical use. Previous work on the type system has been done, however
focusing mostly on dynamic type checking. Safina extended the dynamic type system as
described in [29], where type choices have been added in order to move computation from a
process-driven to a data-driven approach.
The idea to integrate dynamic and static type checking with the introduction of refinement types, verified via SMT solver, has been explored in [32]. The integration of the two
approaches allows a scenario where the static verification of internal services and the dynamic verification of (potentially malicious) external services cooperates in order to reduce
testing effort and enhancing security.
The idea of using SMT Solvers for static analysis, in particular in combination with
other techniques, has been successfully adopted before for other programming languages, for
example LiquidHaskell and F*. LiquidHaskell [14]3 is a notable example of implementation
of Liquid Types (Logically Qualified Data Types) [28]. It is a static verification technique
3
Online demo at http://goto.ucsd.edu/~rjhala/liquid/haskell/demo/
Jolie Static Type Checker: a prototype
3
combining automated deduction (SMT solvers), model checking (Predicate Abstraction), and
type systems (Hindley-Milner inference). Liquid Types have been implemented for several
other programming languages. The original paper presented an OCaml implementation. F*
[1] instead an ML-like functional programming language specifically designed for program
verification.The F* type-checker uses a combination of SMT solving and manual proofs to
guarantee correctness
Another direction in developing static type checking for Jolie is creating the verified
type checker4 by means of proof assistant instead of SMT solver [3]. Proof assistant is a
software tool needed to assist with the development of formal proofs by human-machine
collaboration and helps to ascertain the correctness of them. The type checker is expressed
as well-typed program with dependent types in Agda [26]. If the types are well formed, all
required invariants and properties are described and expressed in the types of the program
meaning that the program is correct. This work is currently in progress and evolves in
parallel with ours.
4
Static type-checking implementation
This paper builds on top of Julie Meinicke Nielsen’s work [25] at the Technical University
of Denmark implementing the type system of the Jolie language. The thesis represents
the theoretical foundation for the type checking of the core fragment of the language, which
excludes recursive types, arrays, subtyping of basic types, faults and deployment instructions
such as architectural primitives. The work of Nielsen presents the first attempt at formalizing
a static type checker for the core fragment of Jolie, and the typing rules expressed there are
the core theory behind our static checker.
In Nielsen’s work typing rules are represented in the style of type theory where type rules
are inference rules describing how a type system assigns a type to a syntactic construct of
the language [7]. The rules are then applied by the type system to determine if a program
is well typed or not. The main typing rules will be presented in the following of this paper.
The implementation of JSTC consists of two system components. Firstly, a Java program
accepts the source code of a Jolie program, builds an abstract syntax tree (AST), visits it
and produces a set of logical assertions written in SMT Lib [5] language. At the second
phase, the generated assertions are feed into Z3 solver. The basic idea is to implement, for
each Jolie node5 , methods containing statements expressed in the SMT Lib syntax. These
statements can then be processed via a solver. In Figure 1 the overall process is pictorially
represented and details are described in section 4.4.
The concept of SMT solvers is closely related to logical theorems. Logic, especially in the
field of proof theory, considers theorems as statements of a formal language. Existence of such
logical expressions allows to formulate a set of axioms and inference rules to formalize the
typing rules for each of Jolie syntax nodes and then perform the validation of the nodes using
constructed theorems. Consequently, the Jolie typing rules are the specific cases of logical
theorems, that are used in the project. The concept is implied from software verification
fundamentals [6].
Since Jolie program may contain complex expressions with function calls, it is also necessary to consider data structures representing a match between names and expressions, in
order to be able to avoid inconsistency and redundancy, that are likely to cause conflicts during type-checking. The project implementation considers using a stack during the recursive
checking of the nodes as illustrated in section 4.4.
The decision of using an SMT-solver, instead of more lightweight techniques, was made
in order to allow a future straightforward integration of Refinement Types into the type
checker, objective on which our team is already working [29,32]. Furthermore, relying on a
solid existing technology allowed us to prototype and release a proof of concept of the type
checker in a shorter period of time.
4
5
https://github.com/ak3n/jolie
Any syntax unit is considered a node. It can be a logical or arithmetic expression, an assignment;
a condition; a loop etc. Those nodes comprise the abstract syntax tree.
4
Daniel de Carvalho et al.
Fig. 1: Process of Type Checking in the Jolie Verifier
4.1
Jolie verifier
The Java program reuses an existing structure of a Visitor pattern that was used in a previous
project for formatting Jolie source code 6 . It accepts processed Jolie program source code
in the form of AST and performs traversing. For each kind of node the system creates one
or more logical formulas written using SMT-LIB [5] syntax, which are then stored into a
file on disk. At the current implementation state the theorems are collected in a single
data element.The verifier targets assignments, conditions, and other cases of variables usage
where type consistency can be violated.
4.2
SMT Solver
Z3 carries out the main functionality of program verification. Z3 is an SMT solver from
Microsoft Research [9]. It is targeted at solving problems that arise in software verification
and software analysis. Given a set of formulas that was previously created by the verifier
in Java, Z3 processes it and returns whether this set is satisfiable or not. In case of any
contradiction in the set, the solver will signal that the overall theorem is not satisfiable,
therefore alerting that the input program is not consistent in terms of types usage.
4.3
Typing rules
Our objective is to accurately translate Jolie typing rules into SMT statements, therefore
allowing static type checking 7 . The foreground activity so far is producing the set of statements for the construct of the behavioural layer of Jolie. The layer describes the internal
actions of a process and the communications it performs seen from the process’ point of view.
The layer is chosen for the first phase of the development because of being the foundation of
the syntactical structures of Jolie. Also there is a similarity of the layer with common programming languages in a sense of the abstraction level. So these facts make the behavioural
level to be the first entry in the world of Jolie language capabilities.
All statements at the behavioural layer of Jolie are called behaviours. We write Γ `B
B . Γ 0 to indicate a behaviour B, typed with respect to an environment Γ , which updates
Γ to Γ 0 during type checking [25].
There are some core rules presented and described below:
T-Nil. The typing rule for a nil behaviour is an axiom. In the conclusion the typing
environment is not changed, since the nil statement doesn’t affect the typing environment.
6
7
https://github.com/nickaleks/jolie
Please note that, at the moment, not all the rules in [25] have been implemented.
Jolie Static Type Checker: a prototype
5
Γ `B 0 . Γ
T-If-Then-Else. The rule for typing an if statement is standard: An if statement is
typable if its condition has type bool, and if the type checking of its branches perform the
same updates to the environment. We require the branches to perform the same updates
because we do not know which branch will be taken. The else part may also be omitted
and B2 may be replaced by an empty behaviour. The conditional typing statement is the
following:
Γ ` e : bool Γ `B B1 . Γ 0 Γ `B B2 . Γ 0
Γ `B if (e) B1 else B2 . Γ 0
T-While. The rule for typing a while statement is standard: A while statement is typable
if its condition has type bool, and if type checking its body has no influence on the typing
environment.
Γ ` e : bool Γ `B B . Γ
Γ `B while(e)B . Γ
Above, it is required that the body of the while loop does not change the typing of
variables because we do not know whether the body will be executed at all, and for how
many times. We also require that expression e is type checked against type bool.
T-Seq. A sequence statement typed with respect to an environment is typable if its first
component is typable with respect to the environment and its second component is typable
with respect to the update of the environment performed by the first component. The update
of the environment performed by the sequence statement is the update performed by the
second component with respect to the update performed by the first component.
Γ `B B1 . Γ 0 Γ 0 `B B2 . Γ 00
Γ `B B1 ; B2 . Γ 00
Thus, fundamental typing rules of the behavioral layer of Jolie programming language
are presented and explained for further topic revelation.
4.4
Typing rules to SMT translation
Here we will illustrate an example of the conditional rule translation in order to understand
the procedure in detail.
The typing rule of the if statement does not contradict intuition. The statement is
typeable when its condition expression is boolean, and the execution of both its branches
brings the same updates to the environment. This means that the set of matches between
expressions and variables with their types remains the same with no difference from a branch
choice. This is necessary since it is not possible to predict what branch will be executed at
runtime 8 .
The full implementation is available on github.9 Below we show the Java fragment that
builds the corresponding SMT statement.
public void visit(IfStatement n) {
for (Pair<OLSyntaxNode, OLSyntaxNode> statement : n.children()) {
OLSyntaxNode condition = statement.key();
OLSyntaxNode body = statement.value();
check(condition);
TermReference conditionTerm = usedTerms.pop();
writer.writeLine("(assert (hasType " + conditionTerm.id + " bool))");
if (body != null) {body.accept(this);}
}
if (n.elseProcess() != null){n.elseProcess().accept(this);}
}
8
9
The else part may also be omitted and B2 may be replaced by an empty behavior.
https://github.com/innopolis-jolie-smt-typechecker/jolie
6
Daniel de Carvalho et al.
The code structure represents basic steps to achieve a record with corresponding SMT
statements of the block as a result. Firstly, a condition of the if statement is separated from
the body. Then the condition is sent to be checked using the same visitor class. Eventually
after the last ’recursion’ step the condition is put in the stack of terms, which contains any
terms (expressions, variables etc.) processed during the checking. So the term corresponding
to the condition is expected to be on top of the stack. Then an assertion that says the
condition term is boolean is written. Afterwards the body is processed using one of the other
overloads of the visitor. These steps can be repeated in case of existence of nested conditional
statements. In the end of the method the else branch body of the very first if is processed
if it is present. There is also an important note is that the conditional statement does not
impose any other direct type restrictions besides the condition term that is confirmed by
the mentioned typing rule. Other implemented nodes can be seen in the source mentioned
above.
The Jolie verifier takes some input for processing. Let us consider a simple piece of Jolie
code with a conditional statement.
a = 2;
b = 3;
if ( a > b ) {
println@Console( a + b )()
} else{
println@Console( "Hello!" )()
}
In the case everything works, none of the typing rules is violated. Z3 agrees with the
opinion and results in ’sat’, that means the program state is satisfiable. We list here the
SMT statements representing the condition processing:
(declare-const $$__term_id_10 Term)
(assert (hasType $$__term_id_10 bool))
(assert (hasType $$__term_id_10 bool))
The first assertion is made based on an expression type determination: the expression
a > b is boolean. The second one is imposed by the typing rule: the condition expression
must be boolean. In this case there is no contradiction between these two assertions.
If the condition would be replaced with some other type expression the typing rule may
be violated. The corresponding example case with a replacement of a > b is shown below:
a = 2;
b = 3;
if ( 5 ) {
println@Console( a + b )()
} else{
println@Console( "Hello!" )()
}
And the constructed SMT statements for the condition expression are given here:
(declare-const $$__term_id_10 Term)
(assert (hasType $$__term_id_10 int))
(assert (hasType $$__term_id_10 bool))
Now the contradiction between the assertions is notable. The parser decided the expression to be an integer, which is correct. But the restriction on a condition type from the
typing rule simply contradict with the actual type. Consequently Z3 results in ’unsat’. This
means that the program state representing the assertion unsatisfiable and incorrect in terms
of the considered static type checking analysis.
Jolie Static Type Checker: a prototype
5
7
Evaluation
The question of how to prove correctness of verification tools has always been widely discussed. How can we be sure that the output of such tool is correct? It can be poorly written,
or the hardware could malfunction. However, in most cases we tend to trust verification
tools, and in our project we have to make sure that this tool is as trustworthy as any other.
The general solution is testing. Verification of the written code correctness was continuously
performed during the development process. The Jolie Team created a collection of examples
of Jolie programs.10 The verification results of some of these are presented in this section.
5.1
An unsatisfiable model
The general purpose of the type checker is to find inconsistency in types usage. The program
listed below is the most basic example of a program with inconsistent types. The variable
myInt is assigned an integer first, and then a string. The current design of the type checker
disallows this behavior.
main
{
myInt = 15;
myInt = "fifteen"
}
The resulting set of SMT theorems is listed below.
(declare-const myInt Term)
(declare-const $$__term_id_4 Term)
(assert (hasType $$__term_id_4 int))
(assert (sameType myInt $$__term_id_4))
(assert (hasType myInt int))
(declare-const $$__term_id_10 Term)
(assert (hasType $$__term_id_10 string))
(assert (sameType myInt $$__term_id_10))
(assert (hasType myInt string))
The model is unsatisfiable. Assertions on the lines 4-6 restrict type of myInt to integer,
whereas assertions on the lines 9-11 ensure that the same variable should be of type string.
These assertions cannot be evaluated to be true in the same model, considering the initial
theorems of the type checker model. Therefore, the overall model is going to be unsatisfiable
upon calling the check-sat Z3 command.
5.2
A satisfiable model
In case everything in the program code is correct in terms of type consistency, the type
checker should evaluate the resulting SMT model as satisfiable. It also should ignore cases
that have not being processed properly yet, without giving any false positives on any inconsistency. The program listed below has, in fact, an inconsistency in types usage. The
line 8 reassigns a variable to be of type integer, whereas at the line 5 the same variable
was introduced as a variable of type string. However, as long as this assignment include a
statement with a dynamic key, the type checker ignores it. The reason for this is inability
to determine which variable this variable path will point to at the moment of execution.
10
https://github.com/jolie/examples
8
Daniel de Carvalho et al.
main
{
key = "cat";
animals.cat = "I am a cat";
animals.(key) = 13
}
The resulting set of SMT theorems is listed below.
(declare-const key Term)
(declare-const $$__term_id_4 Term)
(assert (hasType $$__term_id_4 string))
(assert (sameType key $$__term_id_4))
(assert (hasType key string))
(declare-const $$__term_id_5 Term)
(declare-const animals Term)
(declare-const animals.cat Term)
(declare-const $$__term_id_10 Term)
(assert (hasType $$__term_id_10 string))
(assert (sameType animals.cat $$__term_id_10))
(assert (hasType animals.cat string))
(declare-const $$__term_id_11 Term)
(declare-const animals.DYNAMIC_PATH_$$__term_id_14 Term)
(declare-const $$__term_id_19 Term)
(assert (hasType $$__term_id_19 int))
(assert (sameType animals.DYNAMIC_PATH_$$__term_id_14 $$__term_id_19))
(assert (hasType animals.DYNAMIC_PATH_$$__term_id_14 int))
The workaround for the dynamic keys issue can be seen at Lines 20 and 21. Any time
a variable is used in order to construct another variable path, the whole Term is getting a
unique identifier marked with the DYNAMIC PATH substring. This way it will not interfere
with any other theorem, thus not affecting the overall model satisfiability.
6
Conclusions and future works
Jolie is the first programming language specifically oriented to the microservice architecture.
It has been shown how software attributes such as extensibility, modifiability and consistency
can significantly benefit from a migration into the microservice paradigm [10,11]. Projects
run by our team demonstrated the efficacy of the paradigm and of the Jolie programming
language in the field of ambient intelligence and smart buildings [31,30]. Social networks
implementation would also benefit from a reorganization of the software architecture [19].
Local projects, and beyond that a number of projects worldwide involving the use of Jolie,
would immensely benefit from a fully stable implementation of the Jolie Static Type Checker.
Static type checking allows compilers to identify certain programming mistakes (that
violate types) at compile time, i.e. before actually running the program. Therefore a vast
number of trivial bugs can be caught and fixed at a very early stage of the software lifecycle. In this paper we described JSTC, a static type checker for the Jolie programming
language which natively supports microservices. A static type system for the language has
been exhaustively and formally defined on paper, but so far still lacked an implementation.
We introduced our ongoing work on a static type checker and presented some details of the
implementation. The type checker prototype, at the moment, consists of a set of rules for the
type system expressed in SMT Lib language. The actual implementation covers operations
such as assignments, logical statements, conditions, literals and comparisons.
Jolie Static Type Checker: a prototype
9
JSTC is already able to validate programs, as it has been shown in this paper. However,
it works with certain assumptions. The main assumption is that programs do not contain
implicit type casts. The Jolie language allows implicit type casts, however, their behavior
is very complex. Handling such situations is an open issue left for future development and
future versions. Two other major issues have not been addressed.
Variable types can be changed at runtime. This strictly depends on the approach that has
been chosen. Generally, static typing guarantees that a variable has a type that cannot be
changed after declaration or assignment. However, Jolie allows this operation. We need to
determine which behavior we expect from the type checker, thus deciding how to process
type changes.
Implicit type casts in Jolie are ambiguous. This is a major problem, and further research
is required in order to find a solution. While Jolie allows implicit type casts, sometimes the
result of a cast is not obvious. For example, casting a negative Integer to Boolean will result
in a False. This is an unexpected behavior when compared to other programming languages.
There may be a solid rationale for this, however, we need to investigate all cases and make
sure that the type checker works accordingly to the Jolie actual behavior, and not to the
expected one.
JSTC future releases will need to be validated in real-life applications. The plan is to
use the Jolie programming language and the type checker as a basis for the development
of future research projects, the same way was done in [31] and [30]. Potential application
scenarios are cognitive architecture [33], automotive systems [13] and smart houses [16].
References
1. F*. https://www.fstar-lang.org/.
2. WS-BPEL OASIS Web Services Business Process Execution Language. accessed April 2016.
http://docs.oasis-open.org/wsbpel/2.0/wsbpel-specification-draft.html.
3. Evgenii Akentev, Alexander Tchitchigin, Larisa Safina, and Manuel Mazzara. Verified typechecker for jolie. https://arxiv.org/pdf/1703.05186.pdf.
4. Alexey Bandura, Nikita Kurilenko, Manuel Mazzara, Victor Rivera, Larisa Safina, and Alexander Tchitchigin. Jolie community on the rise. In 9th IEEE International Conference on ServiceOriented Computing and Applications, SOCA, 2016.
5. Clark Barrett, Aaron Stump, and Cesare Tinelli. The SMT-LIB Standard. Version 2.0 , 2010.
6. Hoare C.A.R. An axiomatic basis for computer programming. Communications of the ACM,
12:576–583, 1969.
7. Luca Cardelli. Type systems. ACM Comput. Surv., 28(1):263–264, 1996.
8. Manuel Mazzara Fabrizio Montesi Claudio Guidi, Ivan Lanese. Microservices: a language-based
approach. In Present and Ulterior Software Engineering. Springer, 2017.
9. Leonardo de Moura and Nikolaj Bjrner. Z3: An efficient smt solver. In Tools and Algorithms
for the Construction and Analysis of Systems, 14th International Conference, TACAS 2008,
Held as Part of the Joint European Conferences on Theory and Practice of Software, ETAPS
2008, Budapest, Hungary, March 29-April 6, 2008. Proceedings, volume 4963 of Lecture Notes
in Computer Science, pages 337–340. Springer, 2008.
10. N. Dragoni, I. Lanese, S. T. Larsen, M. Mazzara, R. Mustafin, and L. Safina. Microservices: How
to make your application scale. In A.P. Ershov Informatics Conference (the PSI Conference
Series, 11th edition). Springer, 2017.
11. Nicola Dragoni, Schahram Dustdar, Stephan T. Larse, and Manuel Mazzara. Microservices:
Migration of a mission critical system. https://arxiv.org/abs/1704.04173.
12. Nicola Dragoni, Saverio Giallorenzo, Alberto Lluch-Lafuente, Manuel Mazzara, Fabrizio Montesi, Ruslan Mustafin, and Larisa Safina. Microservices: yesterday, today, and tomorrow. In
Bertrand Meyer and Manuel Mazzara, editors, Present and Ulterior Software Engineering.
Springer, 2017.
13. Rainer Gmehlich, Katrin Grau, Felix Loesch, Alexei Iliasov, Michael Jackson, and Manuel Mazzara. Towards a formalism-based toolkit for automotive applications. 2013 1st FME Workshop
on Formal Methods in Software Engineering (FormaliSE), pages 36–42.
14. Ranjit Jhala. Liquid Haskell. http://goto.ucsd.edu/~rjhala/liquid/haskell/blog/about/.
10
Daniel de Carvalho et al.
15. M.C. MacKenzie et al. Reference model for service oriented architecture 1.0. OASIS Standard,
12, 2006.
16. Ilaria Baroni Marco Nalin and Manuel Mazzara. A holistic infrastructure to support elderlies’
independent living. Encyclopedia of E-Health and Telemedicine, IGI Global, 2016.
17. M. Mazzara, F. Abouzaid, N. Dragoni, and A. Bhattacharyya. Design, modelling and analysis
of a workflow reconfiguration. In Proceedings of the International Workshop on Petri Nets and
Software Engineering, Newcastle upon Tyne, UK, June 20-21, 2011, pages 10–24, 2011.
18. Manuel Mazzara. Towards Abstractions for Web Services Composition. Ph.D. thesis, University
of Bologna, 2006.
19. Manuel Mazzara, Luca Biselli, Pier Paolo Greco, Nicola Dragoni, Antonio Marraffa, Nafees
Qamar, and Simona de Nicola. Social networks and collective intelligence: a return to the
agora. IGI Global, 2013.
20. Manuel Mazzara, Ruslan Mustafin, Larisa Safina, and Ivan Lanese. Towards microservices and
beyond: An incoming paradigm shift in distributed computing. https://arxiv.org/pdf/1610.
01778.pdf.
21. Robin Milner. Communication and concurrency. Prentice Hall International (UK) Ltd., 1995.
22. Robin Milner. Communicating and Mobile Systems: The &Pgr;-calculus. Cambridge University
Press, New York, NY, USA, 1999.
23. Fabrizio Montesi, Claudio Guidi, and Gianluigi Zavattaro. Service-Oriented Programming with
Jolie. In Web Services Foundations, pages 81–107. Springer, 2014.
24. S. Newman. Building microservices. O’Reilly Media, Inc., 2015.
25. Julie Meinicke Nielsen. A Type System for the Jolie Language. Master’s thesis, Technical
University of Denmark, 2013.
26. Chalmers University of Technology. Accessed December 2016. Agda. http://wiki.portal.
chalmers.se/agda/pmwiki.php.
27. Henry Gordon Rice. Classes of recursively enumerable sets and their decision problems. Trans.
Amer. Math. Soc., 74:358–366, 1953.
28. Patrick M. Rondon, Ming Kawaguci, and Ranjit Jhala. Liquid types. SIGPLAN Not., 43(6):159–
169, June 2008.
29. Larisa Safina, Manuel Mazzara, Fabrizio Montesi, and Victor Rivera. Data-driven workflows
for microservices (genericity in jolie). In Proc. of The 30th IEEE International Conference on
Advanced Information Networking and Applications (AINA), 2016.
30. Dilshat Salikhov, Kevin Khanda, Kamill Gusmanov, Manuel Mazzara, and Nikolaos Mavridis.
Jolie good buildings: Internet of things for smart building infrastructure supporting concurrent
apps utilizing distributed microservices. In Proceedings of the 1st International conference on
Convergent Cognitive Information Technologies, pages 48–53, 2016.
31. Dilshat Salikhov, Kevin Khanda, Kamill Gusmanov, Manuel Mazzara, and Nikolaos Mavridis.
Microservice-based iot for smart buildings. In Proceedings of the 31st International Conference
on Advanced Information Networking and Applications Workshops (WAINA), 2017.
32. Alexander Tchitchigin, Larisa Safina, Manuel Mazzara, Mohamed Elwakil, Fabrizio Montesi,
and Victor Rivera. Refinement types in jolie. In Spring/Summer Young Researchers Colloquium
on Software Engineering, SYRCoSE, 2016.
33. Jordi Vallverdú, Max Talanov, Salvatore Distefano, Manuel Mazzara, Alexander Tchitchigin,
and Ildar Nurgaliev. A cognitive architecture for the implementation of emotions in computing
systems. Biologically Inspired Cognitive Architectures, 15(Supplement C):34 – 40, 2016.
| 6cs.PL
|
1
Unmixing urban hyperspectral imagery with a
Gaussian mixture model on endmember variability
arXiv:1801.08513v1 [cs.CV] 25 Jan 2018
Yuan Zhou, Student Member, IEEE, Erin B. Wetherley, and Paul D. Gader, Fellow, IEEE
Abstract—Spectral unmixing given a library of endmember
spectra can be achieved by multiple endmember spectral mixture
analysis (MESMA), which tries to find the optimal combination
of endmember spectra for each pixel by iteratively examining
each endmember combination. However, as library size grows,
computational complexity increases which often necessitates a
laborious and heuristic library reduction method. In this paper,
we model a pixel as a linear combination of endmembers sampled from probability distributions of Gaussian mixture models
(GMM). The parameters of the GMM distributions are estimated
using spectral libraries. Abundances are estimated based on the
distribution parameters. The advantage of this algorithm is that
the model size grows very slowly as a function of the library size.
To validate this method, we used data collected by the
AVIRIS sensor over the Santa Barbara region: two 16 m spatial
resolution and two 4 m spatial resolution images. 64 validated
regions of interest (ROI) (180 m by 180 m) were used to assess
estimate accuracy. Ground truth was obtained using 1 m images
leading to the following 6 classes: turfgrass, non-photosynthetic
vegetation (NPV), paved, roof, soil, and tree. Spectral libraries
were built by manually identifying and extracting pure spectra
from both resolution images, resulting in 3,287 spectra at 16
m and 15,426 spectra at 4 m. We then unmixed ROIs of each
resolution using the following unmixing algorithms: the setbased algorithms MESMA and AAM, and the distribution-based
algorithms GMM, NCM, and BCM. The original libraries were
used for the distribution-based algorithms whereas set-based
methods required a sophisticated reduction method, resulting in
reduced libraries of 61 spectra at 16 m and 95 spectra at 4 m. The
results show that GMM performs best among the distributionbased methods, producing comparable accuracy to MESMA, and
may be more robust across datasets.
Index Terms—spectral unmixing, endmember variability,
Gaussian mixture model, MESMA, hyperspectral image analysis
I. I NTRODUCTION
YPERSPECTRAL images have important applications
in astronomy, agriculture, geoscience, surveillance (such
as object identification), material identification, and detecting
processes [1]. Because limited photons enter the sensor when
collecting narrow bandwidth channels from a high altitude,
the spatial resolution of hyperspectral image is usually very
coarse, i.e. a pixel may correspond to a region with a diameter
H
Y. Zhou and P. D. Gader are with the Department of Computer and
Information Science and Engineering, University of Florida, Gainesville, FL,
USA. E-mail: yuan,[email protected].
E. B. Wetherley is with the Department of Geography, University
of California Santa Barbara, Santa Barbara, CA, USA. E-mail: [email protected].
of several meters. Hence, multiple materials may exist in
this region and contribute to the measured pixel spectrum,
also known as a mixed pixel [2]. One important problem in
hyperspectral imagery is to decompose mixed pixels to identify
the constituting materials (endmember) and their proportions
(abundance) that form the pixel spectrum.
The most common model that relates endmembers and
abundances to a pixel is the linear mixing model (LMM),
which assumes that the reflectance measured within each pixel
is a unique linear combination of the reflectances of each subpixel endmember, weighted by its abundance, plus some noise
[3]. The intuition behind this model is that the fractional area
of a material determines its representation in the measured
signal. However, when unmixing a hyperspectral image with
LMM, we usually encounter an additional problem that spectral reflectance for identical materials are often different. For
example, asphalt spectra can vary significantly based on age,
shadowing, and composite materials [4]. This is sometimes
called endmember variability [5], [6].
Several factors can contribute to endmember variability,
including both extrinsic factors and intrinsic factors. The
most significant extrinsic factor is illumination. When solar
incidence and emergence angles are different for a surface, the
observed signal will be different [7]. Material angle matters as
well, for example roofs can be present at a variety of angles
relative to incoming solar radiation, producing different spectral signatures for one material. Atmospheric condition can be
another extrinsic factor affecting reflectance, however this is
usually corrected during image processing. Measurement scale
represents an important intrinsic factor. Objects or materials
that may be considered “pure” may in reality be composed
of materials at smaller scales with varying reflectances [8].
For example, a tree canopy can be considered a single, pure
endmember, however this ignores the spectral variety of tree
leaves, bark, branches, and substrate that composes a single
tree pixel [9]. Similarly, soils are composed of particles with
different shapes, sizes, and chemical composition [6]. The
larger scale we use to define an endmember, the larger intrinsic
variability we may expect from its spectra. For example,
trees and turfgrass can be defined as individual endmembers,
however if we we wish to define a class of green vegetation
comprised of both turfgrass and tree, its variability will not be
less than the component endmember.
Considering endmember variability, we can generalize the
LMM to the following equation:
0000–0000/00$00.00 c 2018 IEEE
2
yn =
M
X
mnj αnj + nn , n = 1, . . . , N
(1)
j=1
where yn ∈ RB is the spectrum of the nth pixel in the
image, B is the number of bands, N is the number of pixels,
M is the number of endmembers. mnj ∈ RB is the jth
endmember for the nth pixel. αnj ∈ R is the abundance that
usually satisfies
the positivity and sum-to-one constraints, i.e.
P
αnj ≥ 0, j αnj = 1. Finally, we have some additive noise
nn .
When it comes to unmixing in terms of (1), we are referring to retrieving {mnj , αnj } from {yn }, or {αnj } from
{yn } and a library of endmember spectra. The former is
sometimes called unsupervised unmixing, and because it is
undetermined this can be a difficult problem. Studies that have
worked to solve unsupervised unmixing usually require several
assumptions, such as spatial smoothness of the abundances
and the existence of contiguous pure pixels [10], [11], [12].
The latter is called supervised unmixing and depends on a
library of known endmember spectra. If the library is small
enough to easily enumerate all possible spectral combinations,
the task can be trivial. However, applying this scheme on
larger libraries becomes computationally inefficient. This is
the problem we are addressing in this study.
Previous studies that have worked to solve this problem
have used methods that can be categorized as set-based or
distribution-based [5]. Set-based methods treat the endmember
library as an unordered set and try to pick the best combination
of endmembers to model each pixel. A widely used setbased method is multiple endmember spectral mixture analysis
(MESMA) [13]. The general idea of MESMA is to test
every endmember combination and select the one with the
smallest error within set thresholds that limit pixel complexity. There are many variations to MESMA. In multipleendmember linear spectral unmixing model (MELSUM), the
solution for abundances is obtained from directly solving
the linear equations and discarding the negative values [14].
In automatic Monte Carlo unmixing (AutoMCU), pixels are
unmixed using multiple sets of random combinations, with the
mean fractional values assigned as abundances [15], [16]. In
alternate angle minimization (AAM), projection is iteratively
used to find the spectrum index of one endmember given
the other endmembers fixed. Besides MESMA variants, there
is sparse unmixing that used the full spectral library with a
sparsity constraint on the abundances forcing them having only
a few nonzero elements [17].
Contrary to set-based methods, distribution-based methods
assume that the endmembers for each pixel are sampled from
probability distributions, hence the linear combinations of
these endmembers (pixels) also follow some distribution. It
works by modeling the spectral library as statistical distributions, extracting parameters to describe these distributions, and
unmixing the pixels based on the distribution parameters. The
most widely used distribution is Gaussian, and its application
for spectral unmixing is known as the normal compositional
model (NCM) [18], [19], [10], [20], [21], [22]. The popularity
of NCM comes from the fact that a linear combination of
Gaussian random variables is also a Gaussian random variable
whose mean and covariance matrix are linear combinations
from the endmember means and covariance matrices. Hence,
the resulting probability density function of the pixels has a
simple analytical form. Fitting the actual pixel values to the
pixel distribution, the abundances can be solved by several
techniques, such as expectation maximization [20], sampling
methods [18], [19], [10], and particle swarm optimization [22].
Following this philosophy, some have worked to extend the
idea to distributions beyond Gaussian. In [23], the authors
propose Beta distributions to model the spectral library. The
benefit is that Beta distributions have a domain in the range
0 – 1, so are more suitable for the reflectance range, and the
actual library may have a skewed mode in the distribution. In
[24], [12], the idea is further extended to use Gaussian mixture
models (GMM) for distributions. The rationale comes from
the observation that library endmembers may have multiple
modes, whose shape cannot be represented by a simple Gaussian or Beta distribution. Since GMM is more flexible, it can
approximate any distribution found in the library.
A. Our contribution
Many unmixing studies are not well evaluated in presence of
ground truth. Commonly used hyperspectral datasets include
Pavia University, Indian Pines, Cuprite, Mississippi Gulfport,
etc., which are not validated with ground truth endmembers
and abundances. Hence, the primary method for evaluating
their results include:
1) Compare the estimated endmembers with spectra in the
USGS spectra library (e.g. in Cuprite dataset) [25], [26].
2) Compare the estimated abundances with assumed segmentation maps of pure materials (e.g. in Indian Pines,
Pavia University, Gulfport datasets) [21], [27].
3) Calculate the reconstruction error of estimated endmembers and abundances and assume that a lower reconstruction error implies a better result [10], [11].
Each of these methods can be problematic. First, different conditions (sensor, atmosphere, light source) during data
collection will affect measured reflectances, making library
comparison less ideal. Second, high spatial resolution hyperspectral images are primarily composed of pure pixels,
and segmentation like abundance maps do not necessarily
indicate good unmixing capability for mixed pixels. Third,
reconstruction error is more related to model complexity than
unmixing accuracy since small reconstruction error could be
achieved by overfitting [28].
Moreover, these datasets are not comprehensive with respect
to spatial scales, scene diversity and generalization. For example, the Pavia University and Gulfport datasets have about 1 m
spatial resolution in which most are pure pixels. Also, they are
focused on only a few urban sites, which contain mostly manmade materials with segmentation like abundance distribution.
Developing unmixing algorithms on them will have a bias
on forcing smooth and sparse abundance maps. Hence, it is
unknown if the algorithms validated on these datasets can be
applied to datasets with generalized scenarios.
3
In this work, we introduce a supervised unmixing algorithm
based on modeling endmember variability by GMM distributions, and compare several set-based and distribution-based
algorithms with a highly validated, comprehensive dataset of
128 images with different spatial scales. The algorithm was
first introduced in [12], [24] and we modified it for this
application. The dataset was developed in [29] but only used
for evaluating MESMA. It contains two types of images, one
with about 16 m pixel size, the other with about 4 m pixel size.
It covers a wide range of landcover, including various kinds
of road, roof, vegetation, and soil. Validation abundances were
obtained by classifying high resolution images corresponding
to the hyperspectral images. Unlike MESMA, which requires
a small and well-curated spectral library, the GMM algorithm
uses the original source library without modification to unmix
fractions using inferred parameters from the library.
II. DATASET
We used two low-resolution images (16 m) and two highresolution images (4 m) in this study. The low-resolution
images were collected by the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) [1] over Santa Barbara, CA, on
August 29, 2014. The spatial resolutions are 15.6 m/pixel
and 15.8 m/pixel. The spectral range measures wavelengths
from 380 – 2500 nm with 224 bands of approximately 10 nm
bandwidth. High-resolution images were collected by AVIRISNext Generation with 3.9 m/pixel and 3.6 m/pixel spatial
resolutions. The spectral resolution is also higher, recording
432 bands of about 5 - 6 nm bandwidth across a similar
spectral range as the 16 m dataset. We spectrally resampled the
AVIRIS-Next Generation imagery to 224 bands to produce an
image with identical spectral parameters to the 16 m AVIRIS
image. We also removed certain bands from analysis due
to atmospheric interference, reducing the number of bands
to 164. Initial image processing was conducted by the Jet
Propulsion Laboratory, with additional processing in the lab
to reduce the effects of elevation change on pixel location.
The study area includes the cities of Santa Barbara and
Goleta as well as the land between them, near the California coast. Urban composition is typical of the southwestern
United States, including man-made materials such as asphalt,
concrete, metal, gravel, and brick, as well as vegetation in
the forms of turfgrass, various tree species, and large areas of
undeveloped land covered in senesced vegetation [30].
A. Validation Polygons
We produced 64 polygons that represented the variety of
landcover within the study area. Each polygon was 180 m by
180 m in size, or 11-12 pixels wide in the 16 m images (46
or 50 pixels wide in the 4 m images). Validation polygons
were randomly distributed across the area with a minimum
distance 400 m. If a polygon contained large areas of open
water or an undetermined material, it was discarded and a new
polygon randomly generated. Cover was determined within
each polygon using a 1 m NAIP high-resolution image. We
used a combination of image segmentation, using ECognition,
and manual adjustments to classify the cover within each
Table I
N UMBER OF SPECTRA FOR EACH ENDMEMBER CLASS IN THE LIBRARIES
Turfgrass
NPV
Paved
Roof
Soil
Tree
Total
Full
537
884
299
435
262
870
3287
16 m
Reduced
10
14
6
17
3
11
61
Full
1468
3465
2902
2941
1442
3208
15426
4m
Reduced
5
7
17
16
5
45
95
polygon as turf, tree, paved, roof, soil, or non-photosynthetic
vegetation (NPV). Cover was further confirmed by visually
inspecting August 2014 Google Earth imagery. Fig 1 displays
all polygons as they appear in the 16 m images.
Fig. 2 shows a scatter plot of the 64 ground truth abundances
when the 6 endmember classes are merged to 3 categories
of vegetation, impervious, and non-vegetated pervious. Most
polygons are dominated by a mixture of impervious and
vegetation materials. To improve the representation of less
common mixtures in the scene, we added 5 polygons with
high proportions of soil.
B. Library Building
We produced 240 polygons across the 4 m scene to extract
pure spectra and build the full spectral libraries. The polygons
were intended to capture class material variability as much
as possible, and so included multiple roof types, asphalt,
concrete, trees, turfgrass, soil, and NPV, as well as less
common materials like rubber, solar panels, tennis courts, and
plastic tarps. These materials were then grouped into one of
our 6 endmember classes: turfgrass, NPV, paved, roof, soil,
and tree.
The same polygons were used to extract spectra from the
16 m imagery, with necessary modifications as described in
[29]. Together, we produced a library of 16 m spectra and a
library of 4 m spectra. After removing duplicate spectra, the
final 16 m library was comprised of 3,287 spectra and the 4
m library contained 15,426 spectra.
These full spectral libraries were used to train the parameters of distribution-based algorithms. However, they were too
large to be used by MESMA, and required reduction. We
performed reduction in two steps. First, iterative endmember
selection (IES) [31] was used to automatically select a subset
of spectra that represented the larger library. This is achieved
iteratively, by gradually selecting the most representative
spectra and evaluating their representativeness using a kappa
coefficient. IES reduced the 16 m and 4 m library sizes to 226
and 187, respectively. Libraries were further reduced using
iterative classification reduction (ICR), which uses MESMA
as a classifier to quickly identify and remove spectra that tend
to map materials incorrectly [29]. This reduced the libraries to
a final size of 61 for 16 m images and 95 for 4 m images. The
spectra for each endmember class for all the cases are plotted
in Fig 3, and their numbers are shown in Table I.
4
Figure 1. Validation polygons on the site (a) and all 4 m ROI images (b). The two 16 m images are mosaicked by geographic coordinates.
each has numerous spectra in the library. A pixel can be
assumed to be generated by randomly picking one spectrum
for each endmember, and linearly mixing them based on
some abundances. In this way, if we use a probability density
function to represent the spectral distribution, the actual endmembers can be assumed to be sampled from this distribution.
Suppose the jth endmember for the nth pixel is sampled from
a distribution modeled by GMM
p (mnj |Θ) =
Kj
X
πjk N mnj |µjk , Σjk ,
k=1
Figure 2. Scatter plot of ground truth total abundances in terms of 3
categories, green vegetation (turfgrass and tree), pervious (NPV and soil),
and impervious (paved and roof). Most of them lie on the plane, which
corresponds with the selection of ROIs where almost all the pixels fall into
the 6 endmember classes.
III. M ETHOD
A. The Gaussian Mixture Model for Unmixing
Here we briefly introduce the GMM based unmixing [12],
which is a generative model that models a distribution on the
input space [32]. Suppose we have M endmember classes,
where Θ := πjk , µjk , Σjk : j = 1, . . . , M, k = 1, . . . , Kj
are the GMM parameters. Allowing GMM to represent the
library, we can get multiple Gaussian components for each
endmember. Take the dataset in Section II as an example,
which can be viewed from two perspectives. Fig. 4 shows
the pixels from all the validation ROIs, library endmembers,
and corresponding Gaussian components when projected to 2
dimensions. The method for estimating GMM parameters will
be discussed later, however we can see that the ellipses formed
by these parameters surround validation pixels at multiple
positions on the edge of the pixel cloud. The pixels can be
viewed as picking points within ellipses and combining these
points linearly. Fig. 5 shows the Gaussian components from
5
Figure 3. Original and reduced spectral libraries. The numbers of spectra in each category are shown in Table I.
the wavelength-reflectance perspective, where the centers of
Gaussian components and their variation patterns are shown
as curves. Compared to MESMA, which evaluates every
spectrum in the library, GMM tries to combine every center of
Gaussian components, but allows the center to move according
to its corresponding variation pattern.
Following
the
distribution
assumption,
if
{mnj : j = 1, . . . , M } are independent and the noise is
also sampled
from a Gaussian p (nn ) = N (nn |0, D), then
P
yn =
j mnj αnj + nn implies that the pixel follows a
distribution
X
p (yn |αn , Θ, D) =
πk N (yn |µnk , Σnk ) ,
If we assume each pixel is independently sampled, the
probability density function of all the pixels is the product
as
p (Y|A, Θ, D) =
N
Y
p (yn |αn , Θ, D) ,
n=1
T
where A := [α1 , . . . , αN ] ∈ RN ×M . Given Y, Θ, D,
the abundances A can be estimated by maximum likelihood estimation (MLE). Specifically, we want to maximize
p (Y|A, Θ, D), or minimize − log p (Y|A, Θ, D), which becomes the following optimization problem by combining the
above equations
k∈K
where K := {1, . . . , K1 } × {1, . . . , K2 } × · · · × {1, . . . , KM }
is the Cartesian product of the M index sets, k =
(k1 , . . . , kM ) ∈ K, πk ∈ R, µnk ∈ RB , Σnk ∈ RB×B are
defined by
E (A) = −
N
X
n=1
log
X
πk N (yn |µnk , Σnk ) ,
k∈K
s.t. αnj ≥ 0,
M
X
αnj = 1, ∀n.
j=1
πk =
M
Y
j=1
πjkj , µnk =
M
X
j=1
αnj µjkj , Σnk =
M
X
j=1
2
αnj
Σjkj +D.
The objective function can be minimized by a generalized
expectation maximization (EM) algorithm, which alternates
6
Figure 4. Scatter plot of GMM components on the pixels and library spectra. The projection is determined by performing PCA on all the spectra in the
library. The pixels of 64 images for each scale are combined and denoted by gray dots. The colored dots show the spectra in the library for each endmember
class. The ellipses denote the Gaussian components.
Figure 5. Wavelength-reflectance plot of GMM components on the library spectra. The spectra are put into 2-dimensional bins of wavelength-reflectance to
form a histogram shown as gray scale background images. The center of each Gaussian component is shown as solid curve. The center plus (minus) twice the
square root of the largest eigenvalue with its corresponding eigenvector is shown as a dashed curve, which indicates the major variation pattern of a Gaussian
component. The prior probabilities are shown in the legends.
7
between an E step and an M step [33]. The E step calculates the posterior probability of the latent variable given the
observed data and old parameters. The M step increases the
expected value of the complete data log-likelihood. In our case,
the E step calculates
πk N (yn |µnk , Σnk )
.
k∈K πk N (yn |µnk , Σnk )
γnk = P
(2)
The M step tries to minimize
EM = −
N X
X
γnk {log πk + log N (yn |µnk , Σnk )} .
LKj out of all the candidates. To avoid many components, we
tried Kj = 1, 2, 3, 4.
This approach can serve as an ideal modelQ
selection. However, the number of combinations |K| =
j Kj can still
be very large, especially in real datasets where the libraries
contain many spectra. Hence, we use a threshold TCV IC to
further reduce Kj manually. Let L0j be the maximum CVIC
for the jth endmember; we pick the smallest Kj such that
LKj − L0j ≤ TCV IC L0j . Hence when TCV IC = 0, we have
the ideal CVIC-based model selection. As TCV IC increases,
we can have a reduced number of components.
n=1 k∈K
It does not have a closed form solution for A. But we can use
gradient descent to minimize EM , where the derivative can be
calculated by
C. Projection
Analyzing the computation in Section
III-A, we see that
the time complexity is O |K| N B 3 [12]. In addition to the
number of combinations, the number of bands is also crucial
X
X
∂EM
T
T
=−
Λk Rk − 2A ◦
Ψk Sk ,
(3) to execution time. We can reduce the time cost by reducing
∂A
the dimensionality of the data.
k∈K
k∈K
We use PCA to reduce the dimensionality, which gives a
2
where Rk ∈ RM ×B , Sk ∈ RM ×B are defined by
center c ∈ RB and a projection matrix E ∈ RB×d such that
T
all the spectra are processed as ET (y − c). Note that (1) still
Rk = µ1k1 , µ2k2 , . . . , µM kM ,
holds if both the pixel spectra and endmember spectra are
T
projected in this way. Hence the estimated abundances for the
Sk = [vec (Σ1k1 ) , vec (Σ2k2 ) , . . . , vec (ΣM kM )] ,
projected spectra are the original abundances for the data. Also
2
and Λk ∈ RN ×B , Ψk ∈ RN ×B denote
note that if an endmember follows a GMM distribution, the
T
projected endmember also follows a GMM distribution. So we
Λk = [λ1k , λ2k , . . . , λN k ] ,
can directly estimate the GMM parameters from the projected
T
Ψk = [vec (Ψ1k ) , vec (Ψ2k ) , . . . , vec (ΨN k )] ,
library spectra.
As for the data input to find this projection, there are two
B×1
B×B
where λnk ∈ R
and Ψnk ∈ R
are
possibilities. One is to use all the pixel data. This works
if the image is big enough, such that all the endmembers
λnk = γnk Σ−1
nk (yn − µnk ) ,
have sufficient presence. However, if the image contains fewer
1
1
T
−T
−T
−T pixels (e.g. in the 16 m dataset) with limited endmembers, the
Ψnk = γnk Σnk (yn − µnk ) (yn − µnk ) Σnk − γnk Σnk .
2
2
directions determined by PCA will present the variation within
Given an initial A, we can update γnk and A alternately until the image, which may not distinguish different endmembers.
convergence, which leads to a local minimum of the objective The other method is to use the spectral library, i.e. combine
function. This algorithm requires several clarifications and we all the spectra in the library and perform PCA on them. We
adopted this method for our dataset. Specifically, we selected
will explain them in the following subsections.
an equal number of spectra for each endmember class in the
library and concatenated them. This ensures that the relative
B. Determining the GMM Parameters
sizes of endmember classes do not affect the direction, and
we have a library of endmember spectra
Suppose
also ensures that the mean lies in the center.
Yj ∈ RNj ×B : j = 1, . . . , M , with which we can estimate
the GMM parameters Θ. The difficulty comes from estimating
the number of components Kj for each endmember, as once D. The Algorithm
The implementation of the algorithm is described in Algowe know Kj , πjk , µjk , Σjk can be estimated by the standard EM algorithm. Estimating this Kj is sometimes called rithm 1. In step 1, the spectra in the library are concatenated
model selection and has several approaches [34]. We will use to form an input to PCA. We project the data to 10 dimensions
cross-validation-based information criterion (CVIC) [35] as in step 2. Step 3 is elaborated in Section III-B. Step 4
involves initialization of A, which utilizes the information
our metric to select Kj .
To be
Given a candidate Kj , we can evaluate CVIC in the follow- of multiple means from the Gaussian components.
−1
Rk yn , project
ing way. Let Yj be the spectra for the jth endmember in the specific, we set αnk ← Rk RTk + IM
library, we can divide them into V = 5 subsets with equal size. αnk onto the simplex space, and initialize αn ← αnk̂ where
For each subset Yjv , the remaining spectra are input to a MLE k̂ = arg mink kyn − RTk αnk k2 . Step 5 is the main body, in
with Kj Gaussian components, and the trained parameters are which the M step is the most complicated part. Because of
used to evaluate the log-likelihood of Yjv . Then the sum of the constraints on αn , we use projected gradient descent here.
all these log-likelihood values is calculated as LKj , which is The projection function can be found in [36], [37]. The step
our CVIC. Finally, the optimal Kj is the one that maximizes size τ can be set adaptively by using a small initial value
8
Algorithm 1 Spectral unmixing with GMM
T
Input: Y = [y1 , ..., yN ] , {Lj : j = 1, . . . , M }, TCV IC .
1) Determine the projection matrix by PCA.
2) Project the pixels and the spectra in the library to a low
dimensional subspace.
3) Estimate the numbers of components {Kj } using CVIC
and estimate the GMM parameters Θ using standard
EM.
4) Initialize A by choosing the αnk that minimizes the
reconstruction error.
5) Alternately update the E step and M step until convergence.
• E step: update γnk by (2).
∂EM
M
where ∂E
• M step: update A by φ A − τ ∂A
∂A is
defined in (3), τ is some step size, φ is the projection
function to the simplex space.
Output: A.
and gradually increasing it by a multiplier of 10 as long as
the objective function keeps decreasing. Since the covariance
matrix from endmember variability is usually much larger
than the noise covariance, the latter can be negligible in the
computation and we use D = 0.0012 IB .
IV. R ESULTS
A. Setup
We ran GMM on the 16 m and 4 m images after training
it using the same resolution spectral libraries. Since GMM
takes spectral libraries to estimate the Gaussian components,
we used the same components on all 64 images. For reproducibility, we ran model selection of GMM 15 times, selected
the most frequent combination, and applied it to the dataset.
For comparison, we ran 2 set-based methods and 4
distribution-based methods:
1) MESMA [13]. It was implemented in IDL by the original authors and provided as an extension Viper Tool to
ENVI. We used the same parameters as in [29], i.e. maximum RMSE 2.5%, threshold RMSE 0.7%, abundances
constrained between 0 and 1, maximum shade threshold
80%, and a maximum of three endmembers plus shade
for each pixel. Also, it will not allow multiple spectra
from one endmember class in the mixture. The obtained
fractions were normalized to give the final abundances.
2) Alternate angle minimization (AAM) [38]. Its code
was implemented in Matlab and downloaded from Rob
Heylen’s website. It tries every subset of endmembers,
iteratively updates the spectrum index of each endmember such that the reconstruction error is minimized
given the rest selected spectra, and hence finds the best
combination and abundances. Since it uses projection to
find the combination, theoretically it should work faster.
It is different from MESMA in several ways. First, it
may not find the global minimum because of its alternate
optimization. Second, it may find a pixel mixed by many
endmembers instead of maximum three. Finally, it does
not include a shade endmember to adjust for brightness
differences between endmember and measured spectra.
3) Gaussian mixture model (GMM). We used TCV IC =
0.05 for the 16 m dataset, which produced 216 combinations from the library. Because the 4 m dataset was about
20 times larger, and the number of spectra in the library
was 3 times larger, we used a larger TCV IC = 0.2 (18
combinations) such that the whole process could still
run in a few hours.
4) GMM-1. We set the number of combinations to be 1,
i.e. one component for each endmember, which makes
it to be NCM theoretically. However, it has the same
implementation of GMM hence reflects the difference
introduced by bringing multiple components.
5) Normal compositional model with sampling optimization (NCM Sampling). There are many variations of
NCM, with different optimization approaches [18], [19],
[10], [20], [21], [22]. We chose the sampling strategy
in [39] which does not assume statistical independence
between different bands.
6) Beta compositional model (BCM) [23]. It is available
from Alina Zare’s website. Assuming the independence
of bands, it uses Beta distribution to model each band
and unmixes the pixels.
Excluding MESMA, which was implemented in IDL, all
methods were implemented in Matlab. MESMA was run on
a PC with Intel Core i7-2760QM CPU and 8 GB memory.
The other methods were run on a PC with Intel Core i7-3820
CPU and 64 GB memory. For distribution-based methods,
the original libraries were input to train the parameters for
unmixing while set-based methods used the reduced libraries.
We used two metrics to measure the differences between the
estimated and ground truth fractions: mean absolute difference
(MAD) and correlation coefficient (R). They were calculated
for each endmember class based on the 64 pairs of values. To
visualize the values, we used a variation of the Bland-Altman
plot where the x-axis is the ground truth value and the yaxis is the estimated minus the ground truth value [40]. When
comparing different algorithms for unmixing quality, we will
mainly resort to MAD as correlation coefficient itself is not
sufficient (slope and intercept are needed to accompany R).
B. Accuracy and Efficiency
16 m Case. Table II shows the MAD and correlation
coefficient for the 16 m images. Original errors for 6 classes
implies that GMM and AAM have the best accuracy, followed
by MESMA. The difference comes from the paved, roof and
tree classes, where GMM outperformed MESMA. In general,
MESMA, AAM, GMM and GMM-1 had similar accuracy.
Comparing all the distribution-based methods with same input,
GMM has the best performance overall, with fewest errors for
NPV, paved and roof.
Merged errors show that GMM and AAM retain their
higher accuracy, with MESMA falling further behind due to
poor impervious accuracy. Since merged errors are differences
between summed up quantities of similar materials, such
as paved and roof, if the errors are enlarged compared to
9
individual errors, it means that both of the similar materials
are overestimated or underestimated, i.e. the algorithm tends
for confuse even dissimilar materials, such as pervious and
impervious.
Fig. 6 compares the estimated total abundances to validated
abundances for each material. We can see that NCM Sampling
and BCM tend to ignore paved or roof when they have
presence. The difference statistics between the estimated and
the ground truth are plotted in the Bland-Altman plots in
Fig. 7, where the set-based methods and GMM appear to be
better than the others.
4 m Case. The error statistics of 4 m data are shown in
Table III. MESMA and GMM are the most accurate, with
AAM is not as good as MESMA. One possible reason for the
high accuracy of MESMA is that MESMA inherently takes
shade into account while AAM only combines input library
spectra. Hence, the slightly better performance of GMM over
AAM is more significant since they both ignore shade.
Fig. 8 and Fig. 9 show the scatter plots and Bland-Altman
plots for this data. Similar to the statistics in Table III, GMM
has an advantage over the other distribution-based methods
on paved and roof. Compared to AAM, GMM presents stable
results to some extremely bad outliers for soil in AAM.
Efficiency. Since they were run on different machines with
different implementation (MESMA and NCM Sampling have
multiple threads), the time costs are not for comparison, but for
reference. In general, all the algorithms run in a few hours. The
fastest algorithm is GMM-1, which is our implemented GMM
with only one combination. NCM Sampling turns out to be the
slowest algorithm. It is expected since sampling algorithms are
usually slower than deterministic algorithms. The times costs
on the 4 m dataset are usually more than 10 times slower than
those on the 16 m dataset. This is because the image size of
the former is 16-19 times larger than the latter and the library
size is also larger. The least gap comes from GMM because
of a changed TCV IC leading to a significantly less number
of combinations. In [38] the authors show that with the same
implementation AAM is much faster than MESMA, but here
the result is converse. One reason is that the parameters of
MESMA force it to pick at most 3 endmembers for a pixel
instead of all the combinations. Also, multi-threading and
implementation techniques impact the real world time costs
significantly.
C. Extend to Semi-realistic Images
We extended the experiments to semi-realistic images to
check if the algorithm implementation or library reduction was
overfitted to this particular dataset. The method was to test the
algorithms on another batch of synthetic images generated by
the library spectra. Since all the algorithms assumed that a
pixel was a linear combination of endmember spectra from
the library, the creation of this synthetic dataset would use
this assumption.
We created this dataset following the literature that emphasizes realistic simulation [41], [42]. For each image in the
original dataset, we randomly sampled spectra from the full
library while the number of spectra for each endmember class
is equal to the number in the reduced library. Then we used
AAM to unmix the image with these sampled spectra. The
obtained abundances were sorted to keep the largest three
while the other were set to 0, and rescaled such that their
summation was one. This is to conform with the assumption
of MESMA. The endmembers and abundances were combined
according to the LMM to generate pixels. In this way, we can
have a dataset where the endmembers are randomly picked
from the full library, and the spatial distribution of abundances
looks similar to the original one. Fig. 10 shows all the 4 m
synthetic images generated in this way. Comparing it with
Fig. 1, we can see its similarity. But inherently, the synthetic
images follow exactly the LMM with at most 3 endmembers
for each pixel and they are randomly picked from the full
library.
We ran all the algorithms on this simulated dataset. Table IV
shows the unmixing results on this dataset. We can see that
GMM and NCM Sampling turn out to be the best methods.
The superior performance of NCM contrasts sharply to its
worst result in Table II. Since we evaluate the difference
between total abundances for a material, it is possible that the
relatively large pixel abundance error is mitigated by averaging
them. This is more possible for sampling algorithms because
statistically they tend to sample values around the correct ones.
Compared to the negligible difference between GMM and
set-based methods in Section IV-B, the advantage of GMM is
more obvious. Since the endmembers were randomly sampled
from the big library, set-based methods were less capable to
unmix the pixels using a reduced library that was derived
based on the another dataset. It is possible that a different
reduced library based on this simulated dataset may lead to
better results for set-based methods. However, that means, setbased methods may not be as robust as GMM across datasets.
V. D ISCUSSION AND C ONCLUSION
We have proposed an unmixing algorithm based on endmember variability modeled by GMM distributions. We validated the algorithm on a dataset consisting of 128 images
across 2 scales, with ground truth abundances obtained by
inspecting high resolution images. The results show that
with large libraries, GMM achieves comparable accuracy to
MESMA without the need for guided manual library reduction.
We will discuss several issues regarding the dataset, algorithm
and results in this Section.
The dataset was well developed with various scenes, but
the ground truth has intrinsic errors coming from UTM
coordinates. It happens when the universal coordinates are
used in the hyperspectral images and the other high-resolution
images for region correspondence. Because these airborne
images are spatially calibrated from its unstable collection
process, the coordinates derived from the map information
may not accurately reflect the real coordinates. Therefore, the
region for calculating the total abundances may have a small
shift compared to the region in the hyperspectral image. This
is more likely in the 16 m data, which may explain larger
overall errors for all the algorithms, and could be mitigated
by registering the two images to find exact correspondence in
the future [43].
10
Merged
Individual
Table II
C OMPARISON OF ERROR AND CORRELATION COEFFICIENT FOR THE 16 M IMAGES
a
b
Set-based
MESMA
AAM
0.029 / 0.693
0.029 / 0.703
0.069 / 0.830
0.069 / 0.819
0.093 / 0.588
0.093 / 0.685
0.087 / 0.240
0.079 / 0.241
0.069 / 0.773
0.080 / 0.768
0.098 / 0.835
0.065 / 0.855
0.074 / 0.660
0.069 / 0.678a
0.088 / 0.909
0.057 / 0.898
0.082 / 0.836
0.084 / 0.870
0.102 / 0.762
0.087 / 0.818
0.091 / 0.836
0.076 / 0.862
MAD / R2
Turfgrass
NPV
Paved
Roof
Soil
Tree
Average
GV
Pervious
Impervious
Average
NPV
Tree
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
0
0
1
0
1
0
0
1
0
0.5
0
0
1
0
0
0
0
1
0
BCM
Soil
1
0
NCM
Sampling
Roof
1
0
GMM-1
Paved
1
0
GMM
BCM
0.042 / 0.610
0.073 / 0.750
0.096 / 0.538
0.125 / 0.000
0.067 / 0.733
0.184 / 0.534
0.098 / 0.527
0.175 / 0.701
0.074 / 0.886
0.175 / 0.722
0.141 / 0.770
1
0
AAM
Distribution-based
GMM-1
NCM Sampling
0.041 / 0.629
0.193 / 0.117
0.071 / 0.813
0.132 / 0.381
0.093 / 0.768
0.196 / 0.058
0.105 / 0.032
0.112 / 0.362
0.067 / 0.781
0.071 / 0.000
0.071 / 0.849
0.107 / 0.632
0.075 / 0.645
0.135 / 0.258
0.073 / 0.913
0.156 / 0.515
0.075 / 0.880
0.145 / 0.515
0.088 / 0.819
0.199 / 0.181
0.079 / 0.871
0.167 / 0.404
the entries in red fonts denote the best two results in each category.
the running time on all the images for the 6 methods is 18, 27, 83, 0.3, 220, 101 minutes respectively.
Turfgrass
MESMA
GMM
0.045 / 0.632
0.064 / 0.805
0.081 / 0.691
0.078 / 0.279
0.071 / 0.834
0.076 / 0.798
0.069 / 0.673
0.070 / 0.872
0.072 / 0.880
0.084 / 0.845
0.075 / 0.866
0
0
0.5
1
Figure 6. Scatter plots of 64 abundance values in 16 m for ground truth (x-axis) and estimated (y-axis).
In applying GMM on the dataset, there are several implementation details that affected the results. First, the projection
affected the results of GMM. Hence, we used a carefully
determined projection in Algorithm 1. Second, the number of
combinations impacted the performance. If some training data
were present, we could gradually decrease TCV IC until the
error stopped improving during reasonable time. Otherwise,
we suggest gradually increasing TCV IC from 0 (without
running the whole algorithm) until the number of combinations
reduces to approximately 100, which is a reasonable time cost
multiplier of GMM-1. Third, we didn’t use a spatial prior in
the original GMM paper [12]. We found that the prior made
it work better on some images, while worse on some other
images when applied to this dataset. In total, it didn’t improve
the results much. This has two possible reasons: (i) the pixel
size is big enough such that smoothness and sparsity are not
obvious on the abundance maps; (ii) the dataset contains a
variety of scenes in which some of them violate this property.
We also have some remarks on the results. For the real
dataset, GMM is slightly better than MESMA for the 16
m data while slightly inferior for the 4 m data, so they are
close in accuracy. However, MESMA used a library reduction
11
Turfgrass
MESMA
0.5
0
0
0
0
0
0
-0.5
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-1
0
0.2
0.4
-0.5
0
0.5
1
0.2
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
-0.5
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-0.5
0
0.2
0.4
0.5
1
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
-0.5
0
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-0.5
0
0.2
0.4
0.5
1
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
-0.5
0
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-0.5
0
0.2
0.4
0.5
1
0.5
1
0.5
1
0.5
0
0
0
0
0
0
-0.5
0
0.5
-1
0
0.5
1
-0.5
0
0.5
1
-1
0
0.2
0.4
0.5
1
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
-0.5
0
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-0.5
0
0.2
0.4
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
-0.5
0
0.5
-0.5
1
-0.5
0
0.5
-0.5
0.5
-0.5
0
0.5
-0.5
0
-0.5
0
0.5
-0.5
BCM
Tree
1
0
NCM
Sampling
Soil
0.5
-0.2
GMM-1
Roof
0.5
0
GMM
Paved
0.5
-0.2
AAM
NPV
0.2
-0.5
0
0.5
1
Figure 7. Bland-Altman plots of 64 abundance values in 16 m for ground truth and estimated. The x-axis is the ground truth. The y-axis is the difference
between estimated and ground truth. The solid line in each plot is the mean of these differences while the dashed lines show the mean plus (minus) twice the
standard deviation of these differences.
method that relied on manual guidance and user knowledge of
the study area, while GMM used the original library without
refinement. Hence, GMM may be more applicable to datasets
without ground truth. Second, the results of MESMA were
slightly different from those reported in [29], which used this
same imagery and validation dataset. In [29], the average R2
for individual and merge cases are 0.642 and 0.867 for 16
m, 0.811 and 0.923 for 4 m. Comparing them to Table II
and Table III, they are very similar. Note that we used
Matlab to extract the polygon ROIs directly from the original
images while in [29] the images were resampled to a uniform
spatial resolution of 18 m and corrected for locational errors
using Delaunay triangulation. Additionally, that study used
178 spectral bands in contrast to 164 bands used here.
Future work could include developing a dataset with more
images and more scale variation. Also, validation abundances
can be more accurately obtained by registration of hyperspectral and high resolution images. Further work can also be done
to improve the efficiency of GMM using covariance matrices
with a simple form.
ACKNOWLEDGMENTS
The authors wish to thank Alina Zare and Sheng Zou for
providing the NCM sampling code. We wish to acknowledge
funding from the NASA Earth and Space Science Fellowship Program and the Belgian Science Policy Office in the
framework of the STEREO III Program— Project UrbanEARS
(SR/00/307). We also thank the Jet Propulsion Laboratory
for providing radiometrically calibrated, orthorectified AVIRIS
imagery.
R EFERENCES
[1] G. Vane, R. O. Green, T. G. Chrien, H. T. Enmark, E. G. Hansen,
and W. M. Porter, “The airborne visible/infrared imaging spectrometer
(AVIRIS),” Remote Sensing of Environment, vol. 44, no. 2-3, pp. 127–
143, 1993.
[2] C. Small, “Estimation of urban vegetation abundance by spectral mixture
analysis,” International journal of remote sensing, vol. 22, no. 7, pp.
1305–1334, 2001.
[3] J. Settle and N. Drake, “Linear mixing and the estimation of ground
cover proportions,” International Journal of Remote Sensing, vol. 14,
no. 6, pp. 1159–1177, 1993.
[4] M. Herold and D. Roberts, “Spectral characteristics of asphalt road
aging and deterioration: implications for remote-sensing applications,”
Applied Optics, vol. 44, no. 20, pp. 4327–4334, 2005.
[5] A. Zare and K. Ho, “Endmember variability in hyperspectral analysis:
Addressing spectral variability during spectral unmixing,” IEEE Signal
Processing Magazine, vol. 31, no. 1, pp. 95–104, 2014.
[6] B. Somers, G. P. Asner, L. Tits, and P. Coppin, “Endmember variability
in spectral mixture analysis: A review,” Remote Sensing of Environment,
vol. 115, no. 7, pp. 1603–1616, 2011.
[7] G. Meister, A. Rothkirch, H. Spitzer, and J. Bienlein, “BRDF field
studies for remote sensing of urban areas,” Remote Sensing Reviews,
vol. 19, no. 1-4, pp. 37–57, 2000.
[8] E. A. Wentz, D. A. Quattrochi, M. Netzband, and S. W. Myint,
“Synthesizing urban remote sensing through application, scale, data and
case studies,” Geocarto International, vol. 27, no. 5, pp. 425–442, 2012.
12
Merged
Individual
Table III
C OMPARISON OF ERROR AND CORRELATION COEFFICIENT FOR THE 4 M IMAGES
a
b
Set-based
MESMA
AAM
0.028 / 0.740
0.033 / 0.712
0.044 / 0.885
0.059 / 0.874
0.057 / 0.859
0.100 / 0.822
0.060 / 0.574
0.078 / 0.287
0.040 / 0.941
0.065 / 0.793
0.049 / 0.910
0.055 / 0.918
0.046 / 0.818a 0.065 / 0.735
0.039 / 0.943
0.045 / 0.931
0.053 / 0.935
0.091 / 0.858
0.054 / 0.943
0.072 / 0.817
0.049 / 0.941
0.069 / 0.868
MAD / R2
Turfgrass
NPV
Paved
Roof
Soil
Tree
Average
GV
Pervious
Impervious
Average
NPV
Tree
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
0
0
1
0
1
0
0
1
0
0.5
0
0
1
0
0
0
0
1
0
BCM
Soil
1
0
NCM
Sampling
Roof
1
0
GMM-1
Paved
1
0
GMM
BCM
0.049 / 0.710
0.074 / 0.822
0.089 / 0.607
0.105 / 0.181
0.094 / 0.734
0.102 / 0.682
0.086 / 0.623
0.120 / 0.823
0.078 / 0.897
0.151 / 0.791
0.116 / 0.837
1
0
AAM
Distribution-based
GMM-1
NCM Sampling
0.027 / 0.811
0.033 / 0.754
0.052 / 0.867
0.068 / 0.872
0.075 / 0.794
0.165 / 0.847
0.100 / 0.250
0.199 / 0.655
0.084 / 0.897
0.065 / 0.867
0.039 / 0.937
0.057 / 0.926
0.063 / 0.759
0.098 / 0.820
0.042 / 0.942
0.058 / 0.932
0.082 / 0.906
0.087 / 0.886
0.075 / 0.911
0.092 / 0.895
0.067 / 0.919
0.079 / 0.904
the entries in red fonts denote the best two results in each category.
the running time on all the images for the 6 methods is 165, 499, 213, 8, 3153, 1347 minutes respectively.
Turfgrass
MESMA
GMM
0.036 / 0.811
0.052 / 0.879
0.061 / 0.834
0.069 / 0.458
0.078 / 0.904
0.043 / 0.934
0.056 / 0.803
0.043 / 0.943
0.075 / 0.916
0.067 / 0.921
0.062 / 0.927
0
0
0.5
1
Figure 8. Scatter plots of 64 abundance values in 4 m for ground truth (x-axis) and estimated (y-axis).
[9] D. A. Roberts, S. L. Ustin, S. Ogunjemiyo, J. Greenberg, S. Z.
Dobrowski, J. Chen, and T. M. Hinckley, “Spectral and structural
measures of northwest forest vegetation at leaf to landscape scales,”
Ecosystems, vol. 7, no. 5, pp. 545–562, 2004.
[10] A. Halimi, N. Dobigeon, and J.-Y. Tourneret, “Unsupervised unmixing
of hyperspectral images accounting for endmember variability,” IEEE
Trans. on Image Processing, vol. 24, no. 12, pp. 4904–4917, 2015.
[11] L. Drumetz, M.-A. Veganzones, S. Henrot, R. Phlypo, J. Chanussot,
and C. Jutten, “Blind hyperspectral unmixing using an extended linear
mixing model to address spectral variability,” IEEE Transactions on
Image Processing, vol. 25, no. 8, pp. 3890–3905, 2016.
[12] Y. Zhou, A. Rangarajan, and P. D. Gader, “A Gaussian mixture model
representation of endmember variability in hyperspectral unmixing,”
IEEE Transactions on Image Processing, vol. To Appear, 2018.
[13] D. A. Roberts, M. Gardner, R. Church, S. Ustin, G. Scheer, and
R. Green, “Mapping chaparral in the Santa Monica Mountains using
multiple endmember spectral mixture models,” Remote Sensing of
Environment, vol. 65, no. 3, pp. 267–279, 1998.
[14] J.-P. Combe, S. Le Mouelic, C. Sotin, A. Gendrin, J. Mustard, L. Le Deit,
P. Launeau, J.-P. Bibring, B. Gondet, Y. Langevin, et al., “Analysis
of OMEGA/Mars express data hyperspectral data using a multipleendmember linear spectral unmixing model (MELSUM): Methodology
and first results,” Planetary and Space Science, vol. 56, no. 7, pp. 951–
975, 2008.
[15] G. P. Asner and D. B. Lobell, “A biogeophysical approach for
automated SWIR unmixing of soils and vegetation,” Remote sensing
of environment, vol. 74, no. 1, pp. 99–112, 2000.
[16] G. P. Asner and K. B. Heidebrecht, “Spectral unmixing of vegetation,
soil and dry carbon cover in arid regions: comparing multispectral and
hyperspectral observations,” International Journal of Remote Sensing,
13
Turfgrass
MESMA
0.5
0
0
0
0
0
0
-0.5
0.5
-0.2
0
0.5
1
-0.5
0
0.5
1
-0.5
0
0.2
0.4
-0.5
0
0.5
1
0.2
0.5
0.5
0.5
1
0.2
0
0
0
0
0
0
-0.5
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-1
0
0.2
0.4
0.5
1
0.5
0.5
0.5
0.5
0.2
0
0
0
0
0
0
-0.5
0
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-0.5
0
0.2
0.4
0.5
1
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
-0.5
0
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-0.5
0
0.2
0.4
0.5
1
0.5
1
0.4
0.5
0.5
0
0
0
0.2
0
0
-0.5
0
0.5
-1
0
0.5
1
0
0
0.5
1
-0.5
0
0.2
0.4
0.5
1
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
-0.5
0
0.5
-0.5
0
0.5
1
-0.5
0
0.5
1
-0.5
0
Figure 9. Bland-Altman plots of 64 abundance values in 4 m for ground truth and estimated.
Figure 10. Simulated 4 m ROI images. They are very similar to the real images in Fig. 1.
0.2
0.4
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
0
0.5
1
-0.5
0
0.2
-0.2
1
-0.5
0
0.2
-0.2
0.5
-0.2
0
0.2
-0.2
0
-0.2
0
0.2
-0.2
BCM
Tree
0.5
0
NCM
Sampling
Soil
0.5
-0.2
GMM-1
Roof
0.2
0
GMM
Paved
0.5
-0.2
AAM
NPV
0.2
-0.5
0
0.5
1
14
4m
16 m
Table IV
C OMPARISON OF ERROR AND CORRELATION COEFFICIENT FOR THE SYNTHETIC IMAGES
a
MAD / R2
Turfgrass
NPV
Paved
Roof
Soil
Tree
Average
Turfgrass
NPV
Paved
Roof
Soil
Tree
Average
Set-based
MESMA
AAM
0.044 / 0.791
0.050 / 0.837
0.039 / 0.948
0.027 / 0.940
0.074 / 0.512
0.115 / 0.563
0.063 / 0.441
0.063 / 0.569
0.038 / 0.643
0.033 / 0.723
0.090 / 0.882
0.058 / 0.822
0.058 / 0.703
0.058 / 0.742
0.032 / 0.838
0.019 / 0.912
0.040 / 0.955
0.016 / 0.965
0.062 / 0.836
0.121 / 0.819
0.048 / 0.478
0.051 / 0.537
0.043 / 0.661
0.042 / 0.845
0.050 / 0.906
0.035 / 0.911
0.046 / 0.779
0.047 / 0.832
GMM
0.021 / 0.913
0.023 / 0.963
0.078 / 0.745
0.054 / 0.742
0.028 / 0.738
0.051 / 0.880
0.042 / 0.830
0.021 / 0.921
0.012 / 0.990
0.070 / 0.773
0.046 / 0.472
0.026 / 0.840
0.055 / 0.864
0.038 / 0.810a
Distribution-based
GMM-1
NCM Sampling
0.025 / 0.916
0.016 / 0.954
0.026 / 0.956
0.018 / 0.978
0.116 / 0.598
0.038 / 0.927
0.118 / 0.259
0.027 / 0.894
0.028 / 0.781
0.016 / 0.956
0.060 / 0.923
0.029 / 0.960
0.062 / 0.739
0.024 / 0.945
0.015 / 0.921
0.014 / 0.919
0.013 / 0.991
0.019 / 0.995
0.103 / 0.742
0.041 / 0.890
0.087 / 0.254
0.035 / 0.617
0.028 / 0.813
0.036 / 0.953
0.045 / 0.880
0.033 / 0.951
0.049 / 0.767
0.030 / 0.887
BCM
0.042 / 0.827
0.041 / 0.848
0.101 / 0.307
0.129 / 0.170
0.067 / 0.604
0.179 / 0.669
0.093 / 0.571
0.019 / 0.896
0.027 / 0.955
0.053 / 0.723
0.088 / 0.222
0.041 / 0.766
0.099 / 0.748
0.054 / 0.718
the entries in red fonts denote the best two results in each category.
vol. 23, no. 19, pp. 3939–3958, 2002.
[17] A. Castrodad, Z. Xing, J. B. Greer, E. Bosch, L. Carin, and G. Sapiro,
“Learning discriminative sparse representations for modeling, source
separation, and mapping of hyperspectral imagery,” IEEE Transactions
on Geoscience and Remote Sensing, vol. 49, no. 11, pp. 4263–4281,
2011.
[18] O. Eches, N. Dobigeon, C. Mailhes, and J.-Y. Tourneret, “Bayesian
estimation of linear mixtures using the normal compositional model:
Application to hyperspectral imagery,” IEEE Trans. on Image Processing, vol. 19, no. 6, pp. 1403–1413, 2010.
[19] O. Eches, N. Dobigeon, and J.-Y. Tourneret, “Estimating the number
of endmembers in hyperspectral images using the normal compositional
model and a hierarchical Bayesian algorithm,” IEEE Journal of Selected
Topics in Signal Processing, vol. 4, no. 3, pp. 582–591, 2010.
[20] D. Stein, “Application of the normal compositional model to the analysis
of hyperspectral imagery,” in IEEE Workshop on Advances in Techniques
for Analysis of Remotely Sensed Data, 2003, pp. 44–51.
[21] A. Zare and P. D. Gader, “PCE: Piecewise convex endmember detection,” IEEE Trans. on Geoscience and Remote Sensing, vol. 48, no. 6,
pp. 2620–2632, 2010.
[22] B. Zhang, L. Zhuang, L. Gao, W. Luo, Q. Ran, and Q. Du, “PSO-EM:
A hyperspectral unmixing algorithm based on normal compositional
model,” IEEE Trans. on Geoscience and Remote Sensing, vol. 52, no.
12, pp. 7782–7792, 2014.
[23] X. Du, A. Zare, P. D. Gader, and D. Dranishnikov, “Spatial and
spectral unmixing using the Beta compositional model,” IEEE Journal
of Selected Topics in Applied Earth Observations and Remote Sensing,
vol. 7, no. 6, pp. 1994–2003, 2014.
[24] Y. Zhou, A. Rangarajan, and P. D. Gader, “A Gaussian mixture model
representation of endmember variability for spectral unmixing,” in 2016
8th Workshop on Hyperspectral Image and Signal Processing: Evolution
in Remote Sensing (WHISPERS), Aug 2016, pp. 1–5.
[25] J. M. Nascimento and J. M. Bioucas Dias, “Vertex component analysis:
A fast algorithm to unmix hyperspectral data,” IEEE Trans. on
Geoscience and Remote Sensing, vol. 43, no. 4, pp. 898–910, 2005.
[26] X. Lu, H. Wu, Y. Yuan, P. Yan, and X. Li, “Manifold regularized sparse
NMF for hyperspectral unmixing,” IEEE Trans. on Geoscience and
Remote Sensing, vol. 51, no. 5, pp. 2815–2826, 2013.
[27] A. Zare, P. D. Gader, O. Bchir, and H. Frigui, “Piecewise convex
multiple-model endmember detection and spectral unmixing,” IEEE
Trans. on Geoscience and Remote Sensing, vol. 51, no. 5, pp. 2853–
2862, 2013.
[28] K. P. Murphy, Machine learning: a probabilistic perspective, MIT press,
2012.
[29] E. B. Wetherley, D. A. Roberts, and J. P. McFadden, “Mapping
spectrally similar urban materials at sub-pixel scales,” Remote Sensing
of Environment, vol. 195, pp. 170–183, 2017.
[30] D. A. Roberts, D. A. Quattrochi, G. C. Hulley, S. J. Hook, and
R. O. Green, “Synergies between VSWIR and TIR data for the
urban environment: An evaluation of the potential for the hyperspectral
infrared imager (HyspIRI) decadal survey mission,” Remote Sensing of
Environment, vol. 117, pp. 83–101, 2012.
[31] A. N. Schaaf, P. E. Dennison, G. K. Fryer, K. L. Roth, and D. A.
Roberts, “Mapping plant functional types at multiple spatial resolutions
using imaging spectrometer data,” GIScience & Remote Sensing, vol.
48, no. 3, pp. 324–344, 2011.
[32] C. M. Bishop, Pattern recognition and machine learning, springer New
York, 2006.
[33] X.-L. Meng and D. B. Rubin, “Maximum likelihood estimation via the
ECM algorithm: A general framework,” Biometrika, vol. 80, no. 2, pp.
267–278, 1993.
[34] G. J. McLachlan and S. Rathnayake, “On the number of components
in a Gaussian mixture model,” Wiley Interdisciplinary Reviews: Data
Mining and Knowledge Discovery, vol. 4, no. 5, pp. 341–355, 2014.
[35] P. Smyth, “Model selection for probabilistic clustering using crossvalidated likelihood,” Statistics and Computing, vol. 10, no. 1, pp. 63–
72, 2000.
[36] J. Duchi, S. Shalev-Shwartz, Y. Singer, and T. Chandra, “Efficient
projections onto the l1-ball for learning in high dimensions,” in Proc. of
the 25th International Conference on Machine learning (ICML). ACM,
2008, pp. 272–279.
[37] Y. Zhou, A. Rangarajan, and P. D. Gader, “A spatial compositional model
for linear unmixing and endmember uncertainty estimation,” IEEE
Trans. on Image Processing, vol. 25, no. 12, pp. 5987–6002, 2016.
[38] R. Heylen, A. Zare, P. Gader, and P. Scheunders, “Hyperspectral unmixing with endmember variability via alternating angle minimization,”
IEEE Transactions on Geoscience and Remote Sensing, vol. 54, no. 8,
pp. 4983–4993, 2016.
[39] A. Zare, P. Gader, and G. Casella, “Sampling piecewise convex unmixing
and endmember extraction,” IEEE Transactions on Geoscience and
Remote Sensing, vol. 51, no. 3, pp. 1655–1665, 2013.
[40] J. M. Bland and D. Altman, “Statistical methods for assessing agreement
between two methods of clinical measurement,” The lancet, vol. 327,
no. 8476, pp. 307–310, 1986.
[41] L. Gao, Q. Du, B. Zhang, W. Yang, and Y. Wu, “A comparative study
on linear regression-based noise estimation for hyperspectral imagery,”
IEEE Journal of Selected Topics in Applied Earth Observations and
Remote Sensing, vol. 6, no. 2, pp. 488–498, 2013.
[42] Z. Hao, M. Berman, Y. Guo, G. Stone, and I. Johnstone, “Semi-realistic
simulations of natural hyperspectral scenes,” in IEEE International
Geoscience and Remote Sensing Symposium (IGARSS). IEEE, 2015, pp.
1004–1007.
[43] Y. Zhou, A. Rangarajan, and P. D. Gader, “Nonrigid registration of
hyperspectral and color images with vastly different spatial and spectral
resolutions for spectral unmixing and pansharpening,” in 2017 IEEE
Conference on Computer Vision and Pattern Recognition Workshops
(CVPRW), July 2017, pp. 1571–1579.
15
Yuan Zhou received the B.E degree in Software Engineering (2008), the M.E. degree in Computer Application Technology (2011), both from
Huazhong University of Science and Technology,
Wuhan, Hubei, China. Then he worked in Shanghai
UIH as a software engineer for two years. Since
2013, he has been a Ph.D. student in the Department
of CISE, University of Florida, Gainesville, FL,
USA. His research interests include image processing, computer vision and machine learning.
Erin Wetherley is a doctoral candidate in the
Geography Department at the University of California, Santa Barbara. She specializes in characterizing
urban environments using hyperspectral imagery,
thermal imagery, and sub-pixel analyses. Prior to
her doctoral work, Erin earned a bachelors degree in
Environmental Studies from Brown University, and
worked for several years as a GIS and database manager at a Washington, D.C. non-profit organization.
Paul Gader (M’86–SM’09–F’11) received the
Ph.D. degree in mathematics for image-processingrelated research from the University of Florida,
Gainesville, FL, USA, in 1986. He was a Senior Research Scientist with Honeywell, a Research
Engineer and a Manager with the Environmental
Research Institute of Michigan, Ann Arbor, MI,
USA, and a Faculty Member with the University
of Wisconsin, Oshkosh, WI, USA, the University of
Missouri, Columbia, MO, USA, and the University
of Florida, FL, USA, where he is currently a Professor of Computer and Information Science and Engineering. He performed
his first research in image processing in 1984 working on algorithms for
the detection of bridges in forward-looking infrared imagery as a Summer
Student Fellow at Eglin Air Force Base. He has since worked on a wide
variety of theoretical and applied research problems including fast computing
with linear algebra, mathematical morphology, fuzzy sets, Bayesian methods, handwriting recognition, automatic target recognition, biomedical image
analysis, landmine detection, human geography, and hyperspectral and light
detection, and ranging image analysis projects. He has authored/co-authored
hundreds of refereed journal and conference papers.
| 1cs.CV
|
Inertial Hegselmann-Krause Systems
arXiv:1502.03332v3 [cs.SY] 9 Mar 2016
Bernard Chazelle
∗
Chu Wang
†
March 10, 2016
Abstract
We derive an energy bound for inertial Hegselmann-Krause (HK) systems,
which we define as a variant of the classic HK model in which the agents can
change their weights arbitrarily at each step. We use the bound to prove the
convergence of HK systems with closed-minded agents, which settles a conjecture
of long standing. This paper also introduces anchored HK systems and show their
equivalence to the symmetric heterogeneous model.
1
Introduction
The Hegselmann-Krause model of multiagent consensus has emerged as a “model organism” for opinion dynamics [9]. In an HK system, a collection of n agents, each one
represented by a point in Rd , evolves by applying the following rule at discrete times:
move each agent to the mass center of all the agents within unit distance. It has been
shown that the system always freezes eventually [6, 10, 13, 18, 19]. While the model
has been the subject of numerous studies [2, 3, 5, 11, 12, 14] and much is known about
its convergence rate, its heterogeneous variant remains a mystery [24, 25, 26, 27, 21].
In that model, each agent can choose its own radius of confidence. In the HK model
with closed-minded agents, all of the agents have radius either 1 or 0. While extensive
simulations have pointed to the convergence of that system [14, 15, 17, 21], a proof has
remained elusive. This open question has been described by a leading researcher as
one of the outstanding gaps in our understanding of opinion dynamics [8]. This paper
resolves this issue by settling the conjecture in the affirmative: HK systems with closedminded agents always converge. Our proof entails making the problem a special case
∗
Department of Computer Science, Princeton University, [email protected] . This
author was supported in part by NSF grants CCF-0963825 and CCF-1420112
†
Program in Applied and Computational Mathematics, Princeton University, chuw@math.
princeton.edu
1
of a much broader class of dynamical systems, the inertial HK systems (more on which
below).
The relaxation time of the original HK model has been shown to be nO(n) in any fixed
dimension [6], a bound later improved to a polynomial bound in both n and d [1]. For
the particular case d = 1, a bound of O(n5 ) was established in [16], which was lowered to
O(n4) in [20] and then to O(n3 ) in [1]. The model can be generalized in various ways, its
ultimate expression being the grand unified model of influence systems [7], in which each
agent gets to pick its neighbors by following its own distinct, arbitrary criteria. Oddly,
even the most seemingly innocuous modifications of the original HK model have stumped
researchers in the field. This is the case of HK systems with closed-minded agents, where
any agent’s radius of confidence is either 0 or 1. To prove that these systems always
converge, we introduce the more general inertial HK systems and establish a bound on
their kinetic 2-energy. We also introduce the anchored variant of HK systems and prove
that it is equivalent to the symmetric heterogeneous model. This fairly surprising result
sheds new light on the convergence properties of these systems.
1.1
Inertial HK systems
Instead of being required to move to the mass center of its neighbors at each step, each
agent of an inertial HK system may move toward it by any fraction of length; setting
this fraction to zero makes the agent closed-minded, which means that it remains frozen
in place. Formally, the system consists of n agents represented by points x1 (t), . . . , xn (t)
in Rd at time t = 0, 1, 2, etc. Two agents i and j are said to be neighbors if they are
within unit distance: kxi (t) −xj (t)k2 ≤ 1. When the time t is understood, the neighbors
of i form a set we denote by Ni ; these sets form an undirected communication network
Gt with a self-loop at each of the n nodes. The dynamics of the system is specified by
xi (t + 1) = (1 − λ)xi (t) +
λ X
xj (t),
|Ni | j∈N
(1)
i
where λ ∈ [0, 1] is called the inertia. Not only λ need not have the same value for all
the agents, but it can be reset to a different value with each application of (1). In this
way, we can select any agent to be closed-minded by setting their inertia to 0. We can
also retrieve the original HK model by turning all the inertias to 1. In its full generality,
however, an inertial HK system is not deterministic. We tackle the issue of convergence
by turning our attention to their kinetic s-energy. The concept was introduced in [6]
as a generating function for studying averaging processes in dynamic networks. It is
defined as follows:
n
X X
kxi (t + 1) − xi (t)ks2 .
K(s) =
t≥0
i=1
We provide an upper bound for the case s = 2.
2
Theorem 1.1. The kinetic s-energy of an n-agent inertial HK-system whose inertias
are uniformly bounded from above by λ0 satisfies K(2) ≤ λ0 n2 /4.
We use this result to establish the convergence of HK systems with closed-minded
agents. Note that the convergence is asymptotic. This is even true for n = 2 with a
single closed-minded agent. Indeed, if the mobile agent is initialized close enough to the
closed-minded one, it will eventually converge to it by halving its distance at each step.
The network Gt becomes fixed in this case. In general, it changes with time, however.
Interestingly, fixed-point attraction does not automatically imply the convergence of the
communication network, so we address this issue separately.
Theorem 1.2. An HK system with any number of closed-minded agents converges
asymptotically to a fixed-point configuration. The communication network converges for
all initial conditions if d = 1 and for all initial conditions outside a set of measure zero
if d > 1.
The specific meaning of this last clause is that, in dimension two and higher, as long
as we perturb the closed-minded agents by an arbitrarily small amount at the beginning,
the communication network Gt will settle to a fixed graph in finite time almost surely.
The perturbation is likely an unnecessary artifact of the proof and it would be nice to
settle this point. The main open problem, however, is to derive an effective upper bound
on the relaxation time.
1.2
Anchored HK systems
An anchored HK system consists of n agents, each one represented by a vector zk =
(xk (t), yk ). The vector is a combination of a mobile part xk (t) ∈ Rd and a static part
′
yk ∈ Rd ; the dimensions d and d′ are the same for all the agents. Two agents i and j
are neighbors if and only if kzi (t) − zj (t)k2 ≤ r, where r is a fixed positive constant. At
each step, the mobile part of an agent moves to the mass center of all its neighbors while
its anchored part remains fixed. (Note that the averaging is done one coordinate at a
time, so the static coordinates affect only the neighborhood relationships and do not
participate in the averaging itself.) Anchored HK systems capture a notion of partial
closed-mindedness: agents are closed-minded in some coordinates but open-minded in
others. Both mobile and anchored parts, on the other hand, affect the communication
network.
By contrast, a symmetric heterogeneous HK system consists of n agents, each one
represented by a vector xk (t). For each pair of agents (i, j), a threshold rij specifies that
agents i and j are neighbors at time t whenever kxi (t) − xj (t)k2 ≤ rij . It is required that
rij = rji and rii ≥ 0 (the latter to create self-loops). Note that rij = 0 means that i and
j are neighbors only when their positions coincide, while rij < 0 implies that i and j
3
are never joined together. Surprisingly, anchored and symmetric heterogeneous systems
are conjugate: in other words, there exists a bijection between them that respects their
dynamics and establishes the equivalence of the two systems. Specifically, we prove the
following:
′
Theorem 1.3 Given any anchored HK system zk (t) = (xk (t), yk ) in Rd × Rd , there
exists a conjugate symmetric heterogeneous HK system x′k (t) in Rd . Conversely, a symmetric heterogeneous HK system of n agents in Rd is conjugate to an anchored HK
system zk (t) = (xk (t), yk ) with agents in Rd × Rn−1 . In both cases, the conjugacy is
formed by the trivial correspondence: xk (t) = x′k (t) for any k and t. Both anchored and
symmetric heterogeneous HK systems converge asymptotically to a fixed configuration.
If there is no pair of agents (i, j) such that kyi − yj k2 = r in an anchored HK system
or such that rij = 0 in a symmetric heterogeneous HK system, then the communication
network converges to a fixed graph.
While the convergence of symmetric heterogeneous HK systems can be inferred directly from known results, the convergence of the communication networks requires
special treatment, however. An interesting corollary of these results is the convergence
of HK systems embedded within a social network [22, 23, 28]. Imagine that the existence of an edge between two agents i, j is a function not only of their relative distance
but also of a predetermined, fixed relationship. By setting rij < 0, we can enforce the
absence of an edge. In this way we can restrict the HK action to the edges of a fixed,
arbitrary social network, and still assert convergence.
2
Inertial HK Systems
The purpose of this section is to prove Theorem 1.1. To do that, we assign each agent i
a certain amount of “money” Ci (0) at the beginning (t = 0) and specify a protocol for
spending and exchanging it with other agents as time progresses. If we knew ahead of
time the total contribution of agent i to the kinetic 2-energy, we could simply set Ci (0)
to that amount and let the agent “pay” for its contribution from its own pocket. This
information is not available, however, so we take an initial guess and set up an exchange
protocol so that no agent runs out of money. By giving money to their neighbors in a
judicious manner, we show how each agent remains in a position to pay for its share
of the 2-energy at each step. The proof is algorithmic: it is a message-passing protocol
that simulates the update of a distributed Lyapunov function. Our initial guess is
Ci (0) =
n
X
j=1
o
n
min kxi (0) − xj (0)k22 , 1 .
4
To specify the exchange protocol, we first simplify the notation as follows:
∆i = xi (t + 1) − xi (t)
dij = xi (t) − xj (t)
′
dij = xi (t + 1) − xj (t + 1).
The two rules below are applied to every agent i at any time step t ≥ 0:
• For every neighbor of j at time t (which includes i itself), agent i spends k∆i +∆j k22
units of money and gives to agent j an amount equal to 2(dij − ∆j )T ∆j .
• If agent j becomes a new neighbor of i at time t + 1 or, conversely, ceases to be
one, then agent i spends |kd′ij k22 − 1|.
Let Ci (t) be the amount of money held by agent i at time t, and let Niin (resp. Niout )
denote the set of agents that are neighbors of i at time t + 1 (resp. t) but not at time t
(resp. t + 1). Using the symmetry of the neighbor relation, we express the cash flow at
time t by
X
X
Ci (t + 1) − Ci (t) = 2
(dji − ∆i )T ∆i − 2
(dij − ∆j )T ∆j
j∈Ni
−
X
j∈Ni
j∈Ni
k∆i + ∆j k22 −
X
j∈Niin ∪Niout
|kd′ij k22 − 1|.
Since (djiP
− ∆i )T ∆i − (dij − ∆j )T ∆j = dTij (∆i − ∆j ) − 2dTij ∆i + k∆j k22 − k∆i k22 and,
by (1), λ j∈Ni dij = −|Ni |∆i , we have
Ci (t + 1) − Ci (t)
o
Xn
T
2
T
2dij (∆i − ∆j ) + k∆i − ∆j k2 − 4dij ∆i − 4|Ni |k∆i k22 −
=
j∈Ni
=
Xn
j∈Ni
o
2dTij (∆i − ∆j ) + k∆i − ∆j k22 + 4 λ−1 − 1 |Ni |k∆i k22 −
X
|kd′ij k22 − 1|
X
|kd′ij k22 − 1|.
j∈Niin ∪Niout
j∈Niin ∪Niout
Note that λ = 0 implies that ∆i = 0, so it is understood that (λ−1 − 1)|Ni |k∆i k22 = 0
in the identity above. Since d′ij = dij + ∆i − ∆j , the first summand in the last equality
above is equal to kd′ij k22 − kdij k22 ; therefore
Ci (t + 1) − Ci (t)
o
Xn
kd′ij k22 − kdij k22 −
=
j∈Ni
=
n
X
j=1
min
kd′ij k22 , 1
−
X
j∈Niin ∪Niout
n
X
j=1
|kd′ij k22 − 1| + 4 λ−1 − 1 |Ni |k∆i k22
min kdij k22 , 1 + 4 λ−1 − 1 |Ni |k∆i k22 .
5
Since |Ni | > 0 and λ ≤ λ0 , it follows that
Ci (t) ≥
n
X
j=1
min
kdij k22 , 1
+4
λ−1
0
−1
t−1
X
k=0
kxi (k + 1) − xi (k)k22 .
Being its own neighbor, agent i spends at least 4k∆i k22 money at each step. Summing up
over all the agents, this amounts to 4K(2). This shows that the initial
injection of money
−1
allows the system to spend 4K(2) and still be left with 4 λ0 − 1 K(2). Theorem 1.1
follows from the fact that the initial injection of money is at most n2 .
3
HK Systems with Closed-Minded Agents
This section proves Theorem 1.2. The bound on the kinetic 2-energy shows that the
system eventually slows down to a crawl but it falls short of proving convergence. Indeed,
an agent moving along a circle by 1/t at time t contributes finitely to the kinetic 2energy yet travels an infinite distance. We prove that HK systems with closed-minded
agents always converge asymptotically. We treat the one-dimensional separately for two
reasons: the proof is entirely self-contained and the convergence of the communication
network does not require perturbation. In dimension two and higher, we prove that the
agents always converge to a fixed position: the system has a fixed-point attractor. We
show how a tiny random perturbation ensures that the network eventually settles on a
fixed graph.
3.1
The one-dimensional case
We begin with the one-dimensional case, which is particularly simple. By Lemma 1.1,
we can choose a small enough ε > 0 and an integer tε large enough so that no agent
moves by a distance of more than ε at any time t ≥ tε . Fix t > tε and let xi (resp. Ni )
denote the position (resp. neighbors) of agent i at time t; we use primes and double
primes to indicate the equivalent quantities for time t + 1 and t + 2. The symmetric
difference between Ni and Ni′ , if nonempty, is the disjoint union of a set Li of agents
located at xi − 1 ± O(ε) at times t and t + 1 and a set Ri at locations xi + 1 ± O(ε). For
each subset, we distinguish between the agents of Ni not in Ni′ and vice-versa, which
out
gives the disjoint partitions Li = Lin
and Ri = Riin ∪ Riout . The locations x′i and
i ∪ Li
′′
xi of agent i at times t + 1 and t + 2 are given by
(
P
P
|Ni |x′i = ( j∈Ni ∩N ′ xj ) + ( j∈Lout∪Rout xj )
i
i
P
P i
|Ni′ |x′′i = ( j∈Ni ∩N ′ x′j ) + ( j∈Lin ∪Rin x′j ).
i
i
6
i
All x′k and x′′k are of the form xk ± O(ε), so subtracting the two identities shows that
out
(|Ni′ | − |Ni |)xi = (|Lin
i | − |Li |)(xi − 1)
+ (|Riin | − |Riout |)(xi + 1) ± O(εn).
Since the dynamics is translation-invariant, we can assume that xi = 0. Setting ε small
enough, the integrality of the set cardinalities implies that the net flow of neighbors on
the left of agent i is the same as it is on the right:
in
out
in
|Lout
i | − |Li | = |Ri | − |Ri |.
(2)
Among all the agents undergoing a change of neighbors between times t and t + 1, pick
the one that ends up the furthest to the right at time t + 1, choosing the one of largest
index i to break ties. We distinguish between two cases:
1. x′i ≥ xi : No agent of Riout can be closed-minded; nor can it be mobile since, ranks
being preserved, it would provide an agent undergoing a change of neighbors and
landing to the right of i at time t + 1, in contradiction with the definition of i. It
follows that Riout is empty, which in turn implies that Lin
i is not, since by our choice
of i not all four terms in (2) can be zero. Since agent i is not moving left, neither
is any agent j of Lin
i . Its set Nj of neighbors changes between time t and t + 1 and
Rjout is empty. To see why the latter is true, we first note that Nj cannot lose any
closed-minded agent to the right. Also, since any mobile agent in Rjout is to the
left of i at time t, it stays to the left of it by conservation of ranks; hence the agent
remains a neighbor of j, a contradiction. The argument so far uses the rightmost
status of agent i only to assert that Riout is empty. This means we are back to
square one and we can proceed inductively, eventually reaching a contradiction.
2. x′i < xi : The key observation is that our previous argument never uses time
directionality, so we can exchange the role of t and t + 1, which implies that now
x′i > xi . Note that the superscripts in and out must be swapped. While we chose
i as the mobile agent landing furthest to the right, by symmetry we must now
choose the one starting the furthest to the right: of course, since mobile agents
can never cross this make no difference.
We conclude that each agent is now endowed with a fixed set of neighbors, so the
dynamics is specified by the powers of a fixed stochastic matrix with positive diagonal,
which are well known to converge. The system is attracted to a fixed point at an
exponential rate, but of course we have no a priori bound on the time it takes to fall
into that basin of attraction. The communication network converges.
3.2
The higher-dimensional case
Generalizing the previous argument to higher dimension fails on several counts, the most
serious one being the loss of any left-right ordering. We follow a different tack, which
7
begins with a distinction between two types of agents. An agent is trapped at time t
if there exists a path in the current communication graph leading to a closed-minded
agent; it is said to be free otherwise. There exists a time to after which the agents fall
into two categories: some of them are never trapped past t0 and are called eternally free;
the others are chronically trapped (ie, trapped an infinite number of times). As we did
before, we pick a parameter ε > 0 (to be specified below) and tε > to large enough so
that no agent moves by a distance of more than ε at any time t ≥ tε . If two agents ever
get to share the same position, their fates become completely tangled since they can
never again get separated. Since such merges occur fewer than n times, we can make
tε big enough, if necessary, so that all merges are in the past. To summarize, past tε ,
the mobile agents move by increments less than ε, no merging occurs, and the system
consists only of eternally free and chronically trapped agents.
At any time, the state system is represented by a n-by-d matrix whose i-th row
encode the position of agent i in Rd . The matrix consists of two parts: x for the mobile
agents and y for the closed-minded ones. A transition of the system is a linear map of
the form x 7→ Ax + By, where each row of the nonnegative matrix (A | B) sums up to 1.
Lemma 3.1. Past tε , no agent can move while free.
Proof. Fix t ≥ tε and consider a connected component C of the graph induced by the free
agents. If z denotes its position matrix at time t and k its number of rows, then z ′ = Cz,
where primes refer to time t + 1 and C is a k-by-k stochastic matrix for a random walk
in the undirected graph C. Because the graph is connected, the eigenvalue 1 of C is
simple, so the null space of I − C, and hence of (I − C)T (I − C), is spanned by 1. By
Courant-Fischer, therefore, any vector u normal to 1 satisfies k(I − C)uk2 ≥ σkuk2 ,
where σ is the smallest positive singular value of I − C. If we define z̄ = z − k1 11T z, it
immediately follows that
√
σkz̄k2 ≤ k(I − C)z̄k2 = k(I − C)zk2 = kz − z ′ k2 ≤ ε n.
√
Setting ε < 12 σ/ n ensures that any two of the k agents are within unit distance. It
follows that C is the complete graph and C = k1 11T . Since the agents can no longer
merge, the only option left is for all k of them to be already merged at time t, hence
unable to move.
The lemma implies that eternally free agents can never move again past tε . Indeed,
it shows that an eternally free agent can only move if it is joined to a trapped one,
which, by definition, it cannot be. Since eternal freedom keeps the agents from playing
any role after time tε , we might as well assume that all the mobile agents in the system
are chronically trapped. This means that, at all instants, either an agent is trapped (ie,
joined to a closed-minded agent via a path) or it is isolated, meaning that the other
agents are either merged with it or at distance greater than one. An agent cannot move
while isolated.
8
The position matrix z of the k trapped agents at time t ≥ tε satisfies the relation
z ′ = T z +Uy, where primes denote time t+1 and the k-by-n matrix (T | U) has each row
summing up to 1. Being trapped implies that U is not the null matrix. In fact, viewed
as a Markov chain, the trapped agents correspond to transient states, which means that
T k tends to the null matrix as k goes to infinity. This shows that T cannot have 1 as
an eigenvalue; therefore I − T is nonsingular. Let µ be a uniform upper bound on the
singular values of all the (so-called fundamental) matrices
(I − T )−1 ; since their number
√
is finite, so is µ. Since z ′ = T z + Uy and kz ′ − zk2 ≤ ε n, the matrix z is very close to
(I − T )−1 Uy; specifically,
√
(3)
kz − (I − T )−1 Uyk2 = k(I − T )−1 (z − z ′ )k2 ≤ µkz − z ′ k2 ≤ µε n.
A matrix of the form (I − T )−1 Uy is called an anchor. Since the set of all possible
anchors (for given y) is finite, the minimum (Frobenius-norm) distance r between any
two distinct anchors is strictly positive. The value of r does not depend on ε, so
√ we can
always lower the value of the latter, if necessary, to ensure that r > (1 + 2µ)ε n.
By (3) and Lemma 3.1, we know that, at any time √
t past tε , any mobile agent is
either stuck in place (if free) or at distance at most µε n away from an anchor. As
a result, no agent can
change anchors since this would necessitate a one-step leap
√ ever √
of at least r − 2µε n > ε n for the positional matrix, hence the displacement of an
agent by a distance of at least ε, which has been ruled out. Since the argument holds
for any ε small enough, each mobile agent is thus constrained to converge toward its
chosen anchor. This concludes the proof that all agents converge to a fixed point in Rd .
The convergence is asymptotic and no bound can be inferred directly from our analysis.
The result does not imply that the communication network should also converge
to a fixed graph. The lack of convergence points to a situation where the agents are
still moving in increasingly small increments, yet edges of the network keep switching
forever. This can only occur if at least one pair of anchor points are at distance 1: by
anchor point, we mean the points formed by any row of an anchor matrix or of y. The
key observation is that all the anchor points are convex combinations of the rows of y,
so an interdistance of 1 is expressed by an equality of the form kv T yk2 = 1. There are
only a finite set of such equalities to consider and each one denotes an algebraic surface
of codimension 1. Any random perturbation of the closed-minded agents will result in
the convergence of the communication network almost surely. This completes the proof
of Theorem 1.2.
4
Anchored and Symmetric Heterogeneous HK Systems
This section proves Theorem 1.3. We begin with a proof of the conjugacy between the
two types of HK systems.
9
4.1
The bijection relation
To express an anchored HK system z(t) = (xk (t), yk ) as a symmetric heterogeneous one
is straightforward. We have the equivalence
kzi (t) − zj (t)k22 ≤ r 2 ⇔ kxi (t) − xj (t)k22 ≤ r 2 − kyi − yj k22 .
(4)
p
We define rij = r 2 − kyi − yj k22 if the right hand side of (4) is non-negative, and
rij = −1 otherwise. Then the system xk (t) together with thresholds rij forms a symmetric heterogeneous HK system. Notice that the equivalence (4) ensures that the communication graphs of the given anchored HK system and its corresponding symmetric
heterogeneous HK counterpart are identical.
For the other direction, we need to lift the given symmetric heterogeneous HK system
to an anchored HK version. We need the following lemma, whose proof can be found in
the Appendix.
Lemma 4.1. For any n-by-n symmetric matrix R = (rij ) with no negative terms in the
diagonal, there exist r > 0 and vectors yk ∈ Rn−1 (1 ≤ k ≤ n), such that
q
kyi − yj k2 = r 2 − rij2 sign (rij ),
(5)
for any i 6= j; here sign (x) = 1 if x ≥ 0 and −1 otherwise.
Given a symmetric heterogeneous HK system xk (t), we choose the anchors yk by
appealing to Lemma 4.1. For any rij ≥ 0, it then follows that
kxi (t) − xj (t)k22 ≤ rij2 ⇔ kxi (t) − xj (t)k22 + kyi − yj k22 ≤ r 2 ,
(6)
and for any rij < 0, and we always have
kxi (t) − xj (t)k22 + kyi − yj k22 > r 2 ,
(7)
for any i 6= j, which prevents any edge between i and j. This means that the dynamics
of the symmetric heterogenous HK system coincides precisely with that of the mobile
part of the lifted anchored system.
Remark: Lemma 4.1 asserts that, given (n − 1)n/2 lengths dij (i 6= j) of the form
(r 2 − rij2 sign(rij ))1/2 , we can find n points yk ∈ Rn−1 such that the pairwise distance
kyi − yj k2 = dij . Notice that, if dij itself is arbitrary, this is not always possible. For
example, in the case n = 3, the problem is equivalent to finding a triangle in R2 with
each side length given. The problem is solvable if and only if the three lengths satisfy
the triangle inequality. In our case, however, there is an extra parameter r that we can
use. Intuitively, if we choose a large r such that all the |rij | are relatively small, then
the problem of finding yk is equivalent to finding an almost regular polytope, each edge
of which is roughly of the same length r.
10
4.2
Proof of convergence
The fixed-point attraction of symmetric heterogeneous HK systems can be inferred directly from known results about infinite products of type-symmetric stochastic matrices [6, 10, 13, 18]. The same holds of anchored systems. In both cases, given any ε > 0
and any initial condition, the n agents will eventually reach a ball of radius ε that they
will never leave; we call this ε-convergence. We study the conditions for this to imply
that the corresponding communication networks themselves converge to a fixed graph.
It suffices to consider the case of a symmetric heterogeneous HK system. Consider a
connected component C of the graph and let z and z ′ = Cz denote the corresponding
position matrices at time t and t + 1, where C is the corresponding k-by-k stochastic
matrix associated with C. As we did in the proof of Lemma 3.1, we define σ to be a
uniform lower bound on any positive singular value of I − C for any such matrix C.
Setting
σ
ε = √ min rij
2 n rij >0
implies that
√
1
nǫ
1
1
′
≤ min rij ,
kz̄k2 ≤ k(I − C)z̄k2 = kz − z k2 ≤
σ
σ
σ
2 rij >0
where z̄ = z − k1 11T z is the projection of z onto the orthogonal space of 1. It follows
that, for any pair (i, j) in C such that rij > 0, there will be an edge between i and j.
With the assumption rij 6= 0, the communication graph is now fixed and convergence
proceeds at an exponential rate from that point on. The bijection result of the previous
section shows that the condition rij = 0 corresponds to kyi − yj k2 = r in the case of
anchored systems. This concludes the proof of Theorem 1.3.
References
[1] Bhattacharyya, A., Braverman, M., Chazelle, B., Nguyen, H.L. On the convergence
of the Hegselmann-Krause system, Proc. 4th ITCS (2013), 61–66.
[2] Blondel, V.D., Hendrickx, J.M., Tsitsiklis, J.N. On the 2R conjecture for multiagent systems, Proc. European Control Conference 2007 (ECC 2007), July 2007,
874-881, Kos (Greece), 2007.
[3] Blondel, V.D., Hendrickx, J.M., Tsitsiklis, J.N. On Krause’s multi-agent consensus
model with state-dependent connectivity, IEEE Transactions on Automatic Control
54, 11 (2009), 2586–2597.
11
[4] Bullo, F., Cortés, J., Martinez, S., Distributed Control of Robotic Networks, Applied Mathematics Series, Princeton University Press, 2009.
[5] Castellano, C., Fortunato, S., Loreto, V. Statistical physics of social dynamics, Rev.
Mod. Phys. 81 (2009), 591–646.
[6] Chazelle, B. The total s-energy of a multiagent system, SIAM J. Control Optim.
49 (2011), 1680–1706.
[7] Chazelle, B. Diffusive influence systems, SIAM J. Comput. 44 (2015), 1403–1442
[8] Fagnani, F. Lecture, Foundations of Computational Mathematics Conference
(FoCM’11), Budapest, Hungary, 2011.
[9] Hegselmann, R., Krause, U. Opinion dynamics and bounded confidence models,
analysis, and simulation, J. Artificial Societies and Social Simulation 5, 3 (2002).
[10] Hendrickx, J.M., Blondel, V.D. Convergence of different linear and non-linear Vicsek models, Proc. 17th International Symposium on Mathematical Theory of Networks and Systems (MTNS2006), Kyoto (Japan), July 2006, 1229–1240.
[11] Krause, U. A discrete nonlinear and non-autonomous model of consensus formation,
Communications in Difference Equations (2000), 227–236.
[12] Kurz, S., Rambau, J. On the Hegselmann-Krause conjecture in opinion dynamics,
Journal of Difference Equations and Applications (2009).
[13] Lorenz, J. A stabilization theorem for dynamics of continuous opinions, Physica A:
Statistical Mechanics and its Applications 355 (2005), 217–223.
[14] Lorenz, J. Continuous opinion dynamics under bounded confidence: a survey, International Journal of Modern Physics C 16 (2007), 1819–1838.
[15] Lorenz, J. Bounds of confidence: Meet, discuss and find consensus!, Complexity, 4
(2010), 43–52.
[16] Martinez, S., Bullo, F., Cortés, J., Frazzoli, E. On synchronous robotic networks
Part ii: Time complexity of rendezvous and deployment algorithms, IEEE Transactions on Automatic Control 52 (2007), 2214–2226.
[17] Mirtabatabaei A., Bullo, F. Opinion dynamics in heterogeneous networks: convergence conjectures and theorems, SIAM Journal on Control and Optimization, 50
(5):2763–2785, 2012.
[18] Moreau, L. Stability of multiagent systems with time-dependent communication
links, IEEE Transactions on Automatic Control 50 (2005), 169–182.
12
[19] Nedić, A., Touri, B., Multi-dimensional Hegselmann-Krause dynamics, Proceedings of the 51st IEEE Conference on Decision and Control (CDC), Maui, Hawaii,
December 9-13, 2012, pp. 68-73.
[20] Touri, B., Nedić, A. Discrete-time opinion dynamics, Proc. 45th IEEE Asilomar
Conference on Signals, Systems, and Computers (2011), 1172–1176.
[21] Guiyuan Fu, Weidong Zhang. Opinion Dynamics of Modified Hegselmann-Krause
Model with Group-based Bounded Confidence, Proc. 19th World Congress of the
International Federation of Automatic Control, Vol. 14, 2014.
[22] Das, Abhimanyu, Sreenivas Gollapudi, and Kamesh Munagala. ”Modeling opinion
dynamics in social networks.” Proceedings of the 7th ACM international conference
on Web search and data mining. ACM, 2014.
[23] Christakis, Nicholas A., and James H. Fowler. ”The collective dynamics of smoking
in a large social network.” New England journal of medicine 358.21 (2008): 22492258.
[24] Olfati-Saber, Reza, and Richard M. Murray. ”Consensus problems in networks of
agents with switching topology and time-delays.” Automatic Control, IEEE Transactions on 49.9 (2004): 1520-1533.
[25] Lorenz, Jan. ”Heterogeneous bounds of confidence: meet, discuss and find consensus!.” Complexity 15.4 (2010): 43-52.
[26] Mirtabatabaei, Anahita, and Francesco Bullo. ”On opinion dynamics in heterogeneous networks.” American Control Conference (ACC), 2011. IEEE, 2011.
[27] Etesami, Seyed Rasoul, and Tamer Basar. ”Game-theoretic analysis of the
Hegselmann-Krause model for opinion dynamics in finite dimensions.” Automatic
Control, IEEE Transactions on 60.7 (2015): 1886-1897.
[28] Weisbuch, Gerard. ”Bounded confidence and social networks.” The European Physical Journal B-Condensed Matter and Complex Systems 38.2 (2004): 339-343.
13
Appendix
Our proof of Lemma 4.1 relies on two technical facts. For convenience, we use bold
letters to denote vectors; for example, uk denotes the k-th coordinate of vector u.
Fact A. There exist n + 1 vectors u(k) ∈ Rn (0 ≤ k ≤ n) such that ku(i) − u(j) k2 = 1
√
(k)
(k)
(0 ≤ i < j ≤ n), ui = 0 for i > k ≥ 0 and all uk exceed 1/ 2 and decrease as k
grows.
√
Proof. Proceeding by induction, we write u(0) = 0, u(1) = e1 and u(2) = 12 e1 + 23 e2 ,
where ei is the unit vector in the i-th dimension. Assume we already
constructed u(k)
√
(k)
(k)
(0 ≤ k ≤ m < n) such that ui = 0 for i > k and uk > 1/ 2. Then we can write
u(k) as
k
X
(k)
(k)
u =
ui ei , k = 1, 2, . . . , m.
i=1
We define
u
(m+1)
=
m−1
X
(m)
ui ei
+
i=1
(m)
√
(m)
um
−
1
(m)
2um
em +
s
1−
Since um > 1/ 2, we have
(m+1)
um+1
>
s
1−
1
em+1 .
(m) 2
4 um
1
1
√
=√ .
2
4(1/ 2)
2
For k = 0, 1, . . . , m,
(m)
(k)
(m)
ku(m+1) − u(k) k22 = ku(m) − u(k) k22 + u(k)
m /um = (1 − δkm ) + um /um = 1.
Notice that, for 0 ≤ k < n,
(k+1) 2
uk+1
−
(k) 2
uk
=
1−
1
(k) 2
4 uk
which proves the monotonicity claim.
−
(k) 2
uk
(k)
= − uk −
1
(k)
2uk
2
≤ 0,
Fact B. For any integer n > 0, there is a positive number γ depending on n such
that, for any tij satisfying |1 − tij | ≤ γ and tij = tji (0 ≤ i < j ≤ n), there exist vectors
y (k) ∈ Rn (0 ≤ k ≤ n) such that ky (i) − y (j) k2 = tij , for 0 ≤ i < j ≤ n.
Proof. We make repeated use of the matrix infinity norm. Recall that if M is a p-by-q
matrix, its infinity norm is defined as the maximum absolute row sum of M:
kMk∞ := max
1≤i≤p
14
q
X
j=1
|mij |.
As one would expect of a matrix norm, the infinity norm is submultiplicative:
kMNk∞ ≤ kMk∞ kNk∞ ,
for any p-by-q matrix M and q-by-r matrix N. We define a constant
α = 5n + max kCk−1 k∞ ,
1≤k≤n
where Ck is the k-by-k matrix whose i-th row consists of the first k elements of the
vector u(i) in Fact A. Note that Ck is lower-triangular and invertible. Let γ = α−4n .
The intuition of the proof is that the vectors y (k) we are seeking should be close to the
vectors u(k) . We build the desired vectors by induction. Let y (0) = 0 and y (1) = t01 e1 .
Then it is obvious that ky (0) − y (1) k2 = t01 and y (0) and y (1) are close to the vectors
from Fact A:
ky (0) − u(0) k∞ = 0 < γ,
ky (1) − u(1) k∞ = |t01 − 1| ≤ γ ≤ α4 γ.
(j)
Suppose y (0) , y (1) , . . . , y (k−1) have been specified such that yi = 0 for i > j,
ky (i) − u(i) k∞ ≤ α4i γ
(0 ≤ i ≤ k − 1),
(8)
and
ky (i) − y (j) k2 = tij
(0 ≤ i < j ≤ k − 1).
(k)
We need to show is that there exists a vector y (k) such that yi
= 0 for i > k,
ky (k) − u(k) k∞ ≤ α4k γ
(9)
and
ky (k) − y (i) k2 = tik
This last relation is equivalent to
i
X
j=1
(k)
yj
−
(i) 2
yj
+
k
X
(0 ≤ i ≤ k − 1).
(k) 2
yj
j=i+1
= t2ik
(10)
(0 ≤ i ≤ k − 1).
(11)
By subtracting the equations for 1 ≤ i ≤ k − 1 from the one for i = 0, we get a linear
(k) (k)
(k)
system for ŷ := (y1 , y2 , . . . , yk−1)T :
Aŷ = b.
(i)
Here the (k − 1) × (k − 1) matrix A is a lower triangular matrix where Aij = yj (i ≥ j)
and b is a (k − 1) dimensional column vector where
i
X (i) 2
1 2
bi =
t0k − t2ik +
yj
.
2
j=1
15
We derive similar relations from Fact A:
i
X
(k)
uj
j=1
−
(i) 2
uj
+
k
X
(k) 2
uj
j=i+1
(k)
=1
(k)
(0 ≤ i ≤ k − 1),
(12)
(k)
which implies a linear system for û := (u1 , u2 , . . . , uk−1)T :
C û = d,
P
(i) 2
where C is shorthand for Ck−1 and di = 12 ij=1 uj . We already observed that C is
√
(i)
nonsingular; we note that, by (8) and ui > 1/ 2, the same is true of A. Next, we derive
upper bounds on the length of the vector b and its distance from d. By |1 − tij | ≤ γ
and γ < 1/2,
|t20k − t2ik | = |t0k + tik ||t0k − tik | ≤ (2 + 2γ) · 2γ < 6γ.
(13)
(i)
By our induction hypothesis (8), the fact that |yj | ≤ 1 + γ, and the definition of γ, we
have
(i) 2
(i) 2
(i)
(i)
(i)
(i)
yj
− uj
= |yj + uj ||yj − uj | ≤ (2 + α4i γ) · α4i γ < 3α4(k−1) γ.
(14)
Thus, by (13, 14),
kb − dk∞ ≤ 3(1 + nα4(k−1) /2)γ.
By inequality (13) and the fact that γ is small enough, we have
1
1
max |t20k − t2ik | + max ky (i) k22 < (6γ + (1 + γ)2 ) < 1.
kbk∞ ≤
1≤i≤k
2 1≤i≤k
2
(15)
(16)
We also claim that
kA−1 − C −1 k∞ ≤ 2nα4k−2 γ.
(17)
kC −1 (A − C)k∞ ≤ kC −1 k∞ kA − Ck∞ < nα4k−3 γ.
(18)
Here is why. First, notice that (8) implies kA − Ck∞ ≤ nα4(k−1) γ. Then based on the
definition of α, we have kC −1 k∞ < α, and hence
The right hand side of the above inequality is smaller than 1/2 based on the definition
of γ, which allows us to expand the matrix inverse [I + C −1 (A − C)]−1 as
[I + C
−1
(A − C)]
−1
=I+
∞
X
i=0
(−1)i [C −1 (A − C)]i ,
from which it follows that
k[I + C −1 (A − C)]−1 k∞ ≤ 2.
16
(19)
Notice that
A−1 − C −1 = [I + C −1 (A − C)]−1 C −1 (C − A)C −1 ,
then inequality (17) directly follows from inequalities (18) and (19). By (15, 16, 17) and
the fact that kC −1 k∞ < α, finally we have
kŷ − ûk∞ =
=
≤
≤
This shows that
kA−1 b − C −1 dk∞
k(A−1 − C −1 )b + C −1 (b − d)k∞
k(A−1 − C −1 )k∞ kbk∞ + kC −1 k∞ k(b − d)k∞
2nα4k−2 γ + 3(1 + nα4(k−1) /2)αγ < α4k−1 γ.
(k)
(k)
|yj − uj | ≤ α4k−1 γ
(1 ≤ j ≤ k − 1).
In turn, this implies that
(k) 2
(k) 2
(k)
(k)
(k)
(k)
yj
− uj
= |yj + uj ||yj − uj | < (2 + α4k−1 γ)α4k−1 γ < 3α4k−1γ.
(20)
(21)
It suffices now to set the remaining (nonzero) coordinate of y (k) yet to be specified,
(k)
which is yk . Recall that it must satisfy
k
X
j=1
(k) 2
yj
= t20,k
and, by our construction, this single equality suffices to imply all of (10). This implies a
(k)
unique setting of (positive) yk , so we need only be concerned with (9) and the positivity
(k) 2
of yk . Since |1 − t20k | = |1 − t0k ||1 + t0k | ≤ γ(2 + γ) < 3γ, inequality (12) for i = 0,
combined with (14), establishes that
(k) 2
yk
(k) 2
− uk
≤
k−1
X
i=1
(k) 2
yi
(k) 2
− ui
+ |1 − t20k | ≤ 3(1 + nα4k−1 )γ.
√
Since
> 1/ 2, it follows that
1
(k) 2
> − 3(1 + nα4k−1 )γ > 0.
yk
2
Furthermore,
(k) 2
(k) 2
√
− uk
yk
(k)
(k)
≤
3
2(1 + nα4k−1 )γ < α4k γ.
|yk − uk | =
(k)
(k)
yk + uk
(k)
uk
(22)
In conjunction with (20), this establishes (9), and completes the inductive construction.
It should be noted that Fact B can also be proven via the implicit function theorem
and a perturbation argument based on Fact A. The benefit of the proof given above is
to provide an explicit construction.
17
Lemma 4.1 For any n-by-n symmetric matrix R = (rij ) with no negative terms in
the diagonal, there exist r > 0 and vectors yk ∈ Rn−1 (1 ≤ k ≤ n), such that
q
kyi − yj k2 = r 2 − rij2 sign (rij ),
(23)
for any i 6= j; here sign (x) = 1 if x ≥ 0 and −1 otherwise.
Proof. Choose a sufficiently large r such that
max |rij | < γr,
i,j
where γ is the small positive constant from Fact B. We set tij to
q
1 − rij2 sign(rij )/r 2
and easily verify that |1 − tij | ≤ γ. Fact B guarantees the existence of vectors zk ∈ Rn−1
(1 ≤ k ≤ n) such that kzi − zj k2 = tij . Setting yk = rzk satisfies the requirements.
18
| 3cs.SY
|
TUE-AF-PO2.10
1
STEAM: A Hierarchical Co-Simulation Framework
for Superconducting Accelerator Magnet Circuits
arXiv:1801.08957v1 [physics.acc-ph] 26 Jan 2018
L. Bortot, B. Auchmann, I. Cortes Garcia, A.M. Fernandez Navarro, M. Maciejewski, M. Mentink,
M. Prioli, E. Ravaioli, S. Schöps, and A.P. Verweij
Abstract—Simulating the transient effects occurring in superconducting accelerator magnet circuits requires including the
mutual electro-thermo-dynamic interaction among the circuit
elements, such as power converters, magnets, and protection
systems. Nevertheless, the numerical analysis is traditionally done
separately for each element in the circuit, leading to possible
non-consistent results. We present STEAM, a hierarchical cosimulation framework featuring the waveform relaxation method.
The framework simulates a complex system as a composition
of simpler, independent models that exchange information. The
convergence of the coupling algorithm ensures the consistency of
the solution. The modularity of the framework allows integrating
models developed with both proprietary and in-house tools. The
framework implements a user-customizable hierarchical algorithm to schedule how models participate to the co-simulation,
for the purpose of using computational resources efficiently. As a
case study, a quench scenario is co-simulated for the inner triplet
circuit for the High Luminosity upgrade of the LHC at CERN.
Index Terms—Superconducting accelerator magnet; cosimulation; field-circuit coupling; finite element analysis; quench;
circuit modelling; CLIQ; Large Hadron Collider.
I. I NTRODUCTION
Circuits consisting of superconducting accelerator magnets
are complex systems that integrate components and technologies belonging to heterogeneous fields of engineering.
Each component is coupled with the others, showing mutual
interactions. Due to the physical size of the circuit of up to
several kilometres, the number of components, and the variety
of possible dynamic effects, the simulation of such a system
is intrinsically a multi-physics, multi-scale, and multi-rate
problem. In particular, this holds in case of a quench: quench
heaters (QH) [1] or the recently developed Coupling-Loss
Induced Quench system (CLIQ) [2] cause transient effects that
propagate through the circuit [3], interacting with the magnets,
the power converters and the protection electronics.
Simulations are crucial for understanding the transient phenomena occurring within superconducting accelerator magnet
L. Bortot, M. Prioli, A.M. Fernandez Navarro, M. Mentink and A.P. Verweij
are with CERN, Switzerland (e-mail: [email protected]).
B.Auchmann is with CERN, Switzerland, and with Paul Scherrer Institute,
5232 Villigen PSI, Switzerland.
M. Maciejewski is with CERN, Switzerland, and with Institute of Automatic
Control, Technical University of Łódź, 18/22 Stefanowskiego St., Poland.
I.C. Garcia and S. Schöps are with Technische Universität Darmstadt,
Karolinenpl. 5, 64289 Darmstadt, Germany.
E. Ravaioli is with the Lawrence Berkeley National Laboratory, Berkeley,
CA 94720 USA.
The authors I. Cortes Garcia and S. Schöps have been supported by the
Excellence Initiative of the German Federal and State Governments and the
Graduate School of CE at TU Darmstadt.
Figure 1: MQXF cross-section (left), and magnetic flux density
field at nominal current (right).
circuits. Numerical methods are widely used in the analysis
of both the magnets and the quench protection systems,
bringing insights on the quench behaviour and contributing
to prevent potentially irreversible consequences. Nevertheless,
the currently available high-performance tools cannot capture
within one model all the phenomena occurring in an accelerator magnet circuit. Therefore, the system is traditionally
decomposed in component-specific models, refined by expert
know-how. As a consequence, consistent results are achieved
only if all the models’ mutual influences are correctly taken
into account, with a proper two-way coupling strategy.
These considerations led to the development of STEAM,
a co-simulation framework [4]–[7] implemented in Java. The
key features are a communication bus and a user-customizable
hierarchical algorithm. The former exchanges information
between multiple models, the latter schedules how the models
participate to the co-simulation, solving them only when
needed for the accuracy of solution. The coupling of the
models occurs via a dedicated algorithm implementing the
waveform relaxation method [8]. The convergence of the
coupling algorithm ensures the consistency of the solution.
The framework architecture is expandable and can support
both proprietary and in-house tools.
In this paper we introduce the core algorithms and the
architecture of the framework. Then a case-study illustrates
the decomposition of the system, the choice of the solvers,
and the hierarchical organization of the models. The case-study
assumes a quench occurring in one of the Nb3 Sn quadrupole
magnets (MQXF) [9] (see Fig. 1) belonging to the future inner
triplet circuit for the High Luminosity upgrade of the LHC at
CERN [10]. The relevance of achieving consistent simulations
is discussed in the results section.
c 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.
TUE-AF-PO2.10
2
II. F RAMEWORK I MPLEMENTATION
The hierarchical co-simulation approach turns the analysis
of a circuit of accelerator magnets into a coupled problem [11].
The complexity of the original system is represented through
a composition of simpler models. The waveform relaxation
method [7] is applied to resolve the mutual dependencies
between the models, with the following algorithm. i) The
overall simulation time is divided into windows. ii) Within
a window, the models are solved and the relevant waveforms
are exchanged, following a Gauss-Seidel scheme [12]. iii) The
previous operation is repeated for the same window until the
waveforms belonging to two consecutive iterations differ by
less than a prescribed tolerance, then the algorithm moves to
the next window. The convergence of the algorithm for every
window ensures the consistency of the overall solution [13].
A. Architecture
The STEAM framework is developed on a three-layer,
scalable and expandable structure (see Fig. 2) [14]:
1) The top layer contains the hierarchical co-simulation algorithm implementing the waveform relaxation method. The
functionalities of the layer are to manage the execution of
the models over the simulation time windows, to check the
waveform relaxation convergence and to provide the necessary
input/output interfaces.
2) The middle layer exchanges information between the
models that participate to the co-simulation. The layer is
composed by a communication bus which expects a specific
communication protocol. The bus can handle both time- and
space-dependent signals. For the latter ones, the MpCCI [15]
mesh-based interpolation tool is in use [16].
3) The bottom layer implements a modular structure, composed by blocks called tool adapters. Each adapter exchanges
signals between the communication bus and the models belonging to a tool via a suitable Application Programming
Interface (API), which is tool-dependent. This ensures that
different models developed within the same tool rely on the
same tool adapter. The architecture is scalable and expandable:
new simulation tools can be interfaced with the framework by
developing the dedicated tool adapters.
B. Hierarchy
The transient phenomena characterizing a circuit of accelerator magnets might occur at different times, with different
durations. If these phenomena are distributed among separated
multiphysical models, then possibly not all the models are
necessary during the full co-simulation timespan. This observation justifies the introduction of a hierarchical state-machine
algorithm for the models’ management. Referring to Fig. 3, the
user defines both the states and the transitions of the system.
The simulation time is separated into states, and the subset of
active models are determined for each state. A transition can be
defined by a fixed time, or a conditional expression dependent
on the waveforms exchanged among the models. The benefit
is twofold: the state machine input explicitly determines the
causality relations existing between the models, and the overall
computational cost is reduced.
Figure 2: Framework architecture. A generic coupled problem
is decomposed in n models developed with k different tools.
Figure 3: Hierarchical organization of the n models representing a time-dependent coupled problem, in m states.
III. C ASE S TUDY
As an example, STEAM is used for analysing a quench
scenario of the future inner triplet circuit for the High Luminosity LHC. The circuit powering scheme requires three
power supply units (see Fig. 4): a main one (PSM) on the
outer current loop and two trim units (PST1, PST2) located in
nested current loops. Moreover, the system contains nonlinear
bypass components such as flywheel diodes and crowbars,
and six newly designed Nb3 Sn magnets Q1a , Q1b , Q2a , Q2b ,
Q3a , Q3b . The magnets are protected by a combination of
QHs and CLIQ units, with the latter ones connected over
single and multiple magnets [17]. Furthermore, the circuit
protection strategy requires the simultaneous intervention of
the protection systems for all the magnets, in case of a
quench. As a consequence, the mutual dependencies between
the system components justify the co-simulation approach for
the numerical analysis of the scenario.
In the case study, the current in the circuit is rampedup to nominal conditions, with IPSM = 16.47 kA, and
IPST1 = IPST2 = −2 kA. Subsequently, one of the high-field
turns of Q2a is assumed to quench. The longitudinal propagation of the quench within the conductor and the steadily
increasing temperature result in a growing resistive voltage. At
a certain moment the quench is detected: the power converters
are then switched off, the crowbars are activated and the
TUE-AF-PO2.10
3
Figure 4: Inner triplet circuit. the system decomposition and the choice of the tools are highlighted.
simulated with a dedicated 1D adiabatic model, due to the
local nature of the quench initiation and propagation. Since
the heat propagation between adjacent turns is neglected,
the model provides conservative results; 2) The quadrupole
magnets are represented by individual models simulating the
magnetothermal transient induced by the action of the protection systems; 3) The network (see Fig. 4) provides the
electrical coupling between all the components in the circuit.
Each component is included in the network using an equivalent
lumped-parameters representation. An additional a Java code
simulates the quench detection signals associated to each
magnet (Sec. IV).
B. Choice of tools
Figure 5: Transient phenomena occurring in a generic superconducting accelerator circuit, represented in a time-scale
reference frame.
quench protection systems of the magnets are triggered. In
order to give generality to the method, a malfunction in the
QHs of Q2a is assumed.
A. System decomposition
The inner triplet system decomposition is tailored to capture
only the transient phenomena which are relevant for the given
quench scenario (see Fig. 5). In particular, the study focuses on
simulating the initial quench propagation, the magnetothermal
dynamics of the magnets and the electrical behaviour of
the network. Nevertheless, the system decomposition can be
refined with more dedicated models, to include other devices
and physical phenomena. As an example, one can include
the thermal behaviour of the bypass diodes, the action of
the digital controllers of the power converters [18], or the
mechanical response of the magnets [16].
The system is partitioned in three sub-units (see Fig. 4):
the magnet turn where the initial quench occurs, the magnets and the remaining network. 1) The quenching turn is
As a consequence of the system decomposition, each model
covers only a subset of transient phenomena. At this point,
the most suitable simulation tools are determined for each
subset. 1) A SPICE [19] tool is used to calculate the currents
and voltages of the inner triplet equivalent network. The
proprietary distribution PSpicer [20] has been used, although
the freeware version LTspice [21] is also supported; 2) The
finite-element (FE) proprietary tool COMSOLr [22] is used
to calculate the quench initiation and propagation, and the
consequent resistive voltage; 3) Q2a is represented with a
COMSOLr magnetothermal model, which implementation
details are described in [23]. The FEM method is chosen for
providing a detailed analysis of the quenching magnet. The 2D
representation is justified due the large length/diameter ratio
of the coils, and the uniform energy deposition along the coils
by means of CLIQ; 4) The other magnets are modelled using
LEDET [2], [24]. In particular, the magnetothermal formulation is implemented using equivalent networks of lumped
elements which solve faster in comparison to the FE model.
C. Hierarchical organization of the models
The models are organized in a hierarchical structure, reflecting the different system’s states during the co-simulation
(see Fig. 6). At t0 the circuit is powered: only the network
solver is required to ramp up the circuit to nominal operation
conditions. At tquench the quench is introduced and the 1D
TUE-AF-PO2.10
4
(a)
(b)
Figure 6: Hierarchical organization of the models, as function
of the simulation time.
(c)
40
tquench
tdetection
tdischarge
T (K)
30
20
10
0
0
0.1
0.2
0.3
0.4
0.5
x (m)
(d)
0.2
VR
tquench
tdetection
tdischarge
VR (V)
0.15
0.1
0.05
0
-50
-40
-30
-20
-10
0
10
t (ms)
Figure 7: Top: temperature profile along the quenching turn,
during the quench propagation. Bottom: Evolution of the
associated resistive voltage.
model is activated. The resistive voltage grows until it is
detected, at tdischarge . At this point the 1D model is disabled.
The dynamics of the circuit is determined by the action of the
quench protection systems on the quadrupole magnets. For this
reason, the dedicated 2D MQXF models are enabled and kept
active until tend , when the energy in the circuit is completely
discharged.
IV. R ESULTS
The results focus on the dissipation of the energy stored in
the circuit after tquench . Firstly, the magnetothermal transient
occurring in Q2a is discussed in detail. Next, two magnetnetwork coupling strategies are applied, the former based on
an equivalent thermal representation and the latter on the fieldcircuit coupling technique. Lastly, the results given by the two
coupling strategies are compared.
The magnet Q2a is assumed to quench at tquench (Fig. 7,
top), developing a resistive voltage contribution VR (Fig. 7,
bottom). The normal zone propagates along the turn until the
quench detection threshold (VR > 100 mV) is reached, at
Figure 8: (a) Voltage-to-ground distribution, 5 ms after
tdischarge . (b) Inter-filament and inter-strand coupling losses,
5 ms after tdischarge . (c) Ohmic losses, 25 ms after tdischarge .
(d) Temperature distribution, 500 ms after tdischarge .
tdetection . Subsequently, a validation criterion is applied to VR ,
requiring the resistive voltage signal to exceed a threshold
of 100 mV for 10 ms. Once the quench is validated, the
protection systems are activated at tdischarge .
The 1000 V / 40 mF CLIQ unit is discharged in parallel
to the Q2a coil poles. This induces a voltage redistribution
(Fig. 8a), turning into an imbalance in the poles currents which
oscillate due to the resonance between the inductive and capacitive behaviour of the magnet and CLIQ unit, respectively.
The oscillation of the magnetic field determines the dissipation
within the coil of the energy stored in the CLIQ unit, as
inter-filament and inter-strand (Fig. 8b) coupling currents [25],
[26]. The energy deposition determines the quench of a bigger
volume of the coil, which becomes resistive and contributes
both to the discharge of the magnet through the dissipation
of Joule losses (Fig. 8c), and to a temperature increase in the
coil (Fig. 8d). The computational time of the co-simulation is
about 5h, on a standard workstation.
The magnets’ internal dynamics imposes the equivalent
impedances seen by the network. The network, in turn, determines the currents driving the magnets’ dynamics. Hence,
a reliable quench protection simulation requires a consistent
TUE-AF-PO2.10
(a)
5
and a lower Joule integral (MIITs). This is reflected both in
Fig. 9b) where the hot-spot temperature decrease by 42 K,
and in Fig. 9c), where the peak voltage to ground decrease
by 15 V. A summary of the comparison is provided in Tab. I,
showing an overall reduction in the estimation of the magnet
stress parameters.
20
I (kA)
15
10
5
I Q2a CSTh
I Q2a CSMTh
ICLIQ CSTh
ICLIQ CSMTh
Table I: Comparison of results
0
Quantity
100
MIITs
THotspot
VPeakGnd
101
102
103
t (ms)
(b)
Unit
CSTh
CSMTh
∆
∆%
MA2 s
K
V
27
253
935
24
211
920
-3
-42
-15
-9
-16
-2
300
250
Thot−spot CSTh
Thot−spot CSMTh
V. C ONCLUSIONS AND OUTLOOK
T (K)
200
150
100
50
0
100
101
102
103
102
103
t (ms)
(c)
1
VPeak CSTh
VPeak CSMTh
V (kV)
0.5
0
-0.5
-1
100
101
t (ms)
Figure 9: Q2a magnet. (a) Currents in the coil and in the CLIQ
unit; (b) Hot-spot temperature; (c) Maximum and minimum
voltage to ground.
two-way coupling between the magnets and the circuit. To
prove this, a comparison is provided. In the co-simulation
CSTh , each magnet is represented in the network as a linear
inductor in series with a time varying resistor. Such a simplification takes into account the thermal response of the magnet
but neglects both the saturation and the dynamic phenomena
in the superconducting coils. In the co-simulation CSMTh , the
dynamics of each magnet is consistently represented using the
field-circuit coupling technique [7], [27], [28].
The currents for Q2a and the related CLIQ unit are shown
in Fig. 9a), for both the co-simulations. The coil dynamics
greatly reduces the magnet’s differential inductance, determining steeper initial derivatives in the currents of CSMTh . This,
in turn, causes a higher deposition of dynamic losses, a more
homogeneous spread of the quench, a faster current decay
The role of simulations is twofold. On the one hand,
numerical methods provide support to the design of both
circuit components and quench protections systems. On the
other hand, they bring insights to the transient phenomena
occurring in superconducting accelerator magnet circuits, even
for quantities that cannot be measured. This holds true for the
high-performance, low-margin inner triplet circuit for the LHC
High Luminosity upgrade. The design requires accurate simulations, due to the mutual electro-thermal coupling occurring
among the magnets, the quench protection systems and the
rest of the network. Simulating such a complex system in a
single tool with high accuracy is currently not feasible.
For this reason, we propose STEAM, a Java-developed
framework which allows to decompose the original system into
simpler, independent models solved consistently, as a coupled
problem. The hierarchical algorithm ensures an efficient use
of computational resources, enforcing the models to be solved
only for the simulation time span where they contribute to
the accuracy of the solution. The framework architecture is
scalable and expandable, giving the possibility to add further
proprietary and in-house solvers in the future. A quench
protection scenario in the inner triplet circuit is used as a case
study, to illustrate the co-simulation approach. Results remark
the importance of taking into account the mutual influences
between the sub-units in the circuit, in a consistent way. The
analysis of the magnets’ dynamics is improved, leading to less
conservative results: welcomed margins are highlighted in the
baseline design of the high-performance MQXF magnets.
STEAM is actively supporting the simulation needs of the
most demanding accelerator projects [29]–[32]. The framework is ready to be extended to new modules and tools, to
cover more simulation needs: examples are the QHs dynamics,
the mechanical response of the magnet coils, the 3D quench
propagation, and the fluid dynamics of the coolant.
VI. ACKNOWLEDGEMENT
The authors would like to thank K. Król, and J.C. Garnier
from CERN for the continuous help provided in developing
the STEAM Framework, and S. Yammine from CERN for the
support in the development of the inner triplet PSpice model.
TUE-AF-PO2.10
6
R EFERENCES
[1] F. Rodriguez-Mateos and F. Sonnemann, “Quench heater studies for the
LHC magnets,” in Particle Accelerator Conference, 2001. PAC 2001.
Proceedings of the 2001, vol. 5. IEEE, 2001, pp. 3451–3453.
[2] E. Ravaioli, “CLIQ. A new quench protection technology for superconducting magnets,” Ph.D. dissertation, Universiteit Twente, 2015.
[3] E. Ravaioli, K. Dahlerup-Petersen, F. Formenti, V. Montabonnet, M. Pojer, R. Schmidt, A. Siemko, M. S. Camillocci, J. Steckert, H. Thiesen
et al., “Impact of the voltage transients after a fast power abort on
the quench detection system in the LHC main dipole chain,” IEEE
Transactions on Applied Superconductivity, vol. 22, no. 3, pp. 9 002 504–
9 002 504, 2012.
[4] F. Henrotte, E. Lange, and K. Hameyer, “An efficient field-circuit
coupling method by a dynamic lumped parameter reduction of the FE
model,” in Power Electronics and Motion Control Conference, 2008.
EPE-PEMC 2008. 13th. IEEE, 2008, pp. 2393–2399.
[5] E. Lange, F. Henrotte, and K. Hameyer, “An efficient field-circuit
coupling based on a temporary linearization of FE electrical machine
models,” IEEE Transactions on Magnetics, vol. 45, no. 3, pp. 1258–
1261, 2009.
[6] P. Zhou, D. Lin, W. Fu, B. Ionescu, and Z. Cendes, “A general cosimulation approach for coupled field-circuit problems,” IEEE transactions
on magnetics, vol. 42, no. 4, pp. 1051–1054, 2006.
[7] S. Schops, H. De Gersem, and A. Bartel, “A cosimulation framework
for multirate time integration of field/circuit coupled problems,” IEEE
Trans. Magn., vol. 46, no. 8, pp. 3233–3236, 2010.
[8] J. K. White and A. Sangiovanni-Vincentelli, “Waveform relaxation,” in
Relaxation Techniques for the Simulation of VLSI Circuits. Springer,
1987, pp. 79–100.
[9] P. Ferracin, G. Ambrosio, M. Anerella, A. Ballarino, H. Bajas, M. Bajko,
B. Bordini, R. Bossert, D. Cheng, D. Dietderich et al., “Development of
MQXF: The Nb3 Sn Low-β Quadrupole for the HiLumi LHC,” IEEE
Transactions on Applied Superconductivity, vol. 26, no. 4, pp. 1–7, 2016.
[10] H. L. Collaboration et al., “HL-LHC Preliminary Design Report,” FP7
High Luminosity Large Hadron Collider Design Study, CERN-ACC2014, vol. 300, 2014.
[11] K. Park and C. A. Felippa, “Partitioned analysis of coupled systems,”
Computational Methods for Transient Analysis, vol. 1, pp. 157–219,
1983.
[12] K. Burrage, Parallel and sequential methods for ordinary differential
equations. Clarendon Press, Oxford, 1995.
[13] A. Bartel, M. Brunk, M. Günther, and S. Schöps, “Dynamic iteration
for coupled problems of electric circuits and distributed devices,” SIAM
Journal on Scientific Computing, vol. 35, no. 2, pp. B315–B335, 2013.
[14] M. Maciejewski, B. Auchmann, L. Bortot, I. Cortes Garcia, A. M. Fernandez Navarro, M. Prioli, S. Schöps, and A. P. Verweij, “Architecture of
a Hierarchical Co-Simulation Framework for the Simulation of Transient
Effects in Superconducting Accelerator Circuits,” Under preparation.
[15] R. Ahrem, M. Hackenberg, P. Post, R. Redler, and J. Roggenbuck,
“MpCCI-Mesh Based Parallel Code Coupling Interface,” Institute for
Algorithms and Scientific Computing (SCAI), GMD, http://www. mpcci.
org, 2000.
[16] M. Maciejewski, B. Auchmann, L. Bortot, I. Cortes Garcia, A. M.
Fernandez Navarro, M. Prioli, S. Schöps, and A. P. Verweij, “Coupling
of Magneto-Thermal and Mechanical Superconducting Magnet Models
by Means of Mesh-Based Interpolation„” IEEE Transactions on Applied
Superconductivity, vol. 28, no. 3, 2018, accepted for publication.
[17] E. Ravaioli, G. Ambrosio, B. Auchmann, P. Ferracin, M. Maciejewski,
F. Rodriguez-Mateos, G. Sabbi, E. Todesco, and A. P. Verweij, “Quench
protection system optimization for the High Luminosity LHC Nb3 Sn
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
quadrupoles,” IEEE Transactions on Applied Superconductivity, vol. 27,
no. 4, pp. 1–7, 2017.
M. Maciejewski, I. C. Garcia, S. Schöps, B. Auchmann, L. Bortot,
M. Prioli, and A. Verweij, “Application of the Waveform Relaxation
Technique to the Co-Simulation of Power Converter Controller and
Electrical Circuit Models,” in 22nd International Conference on Methods
and Models in Automation and Robotics (MMAR 2017), Miedzyzdroje,
Poland, accepted for publication, 2017.
A. Vladimirescu, The SPICE book. John Wiley & Sons, Inc., 1994.
OrCAD, PSpice 17.2, [CD-ROM]. San Jose, CA: Cadence Design
Systems Inc., 2017.
M. Engelhardt, “Ltspice iv,” Linear Technology Corporation, Sept,
2011. [Online]. Available: http://www.linear.com/designtools/software/
#LTspice.[Accessed:Nov.01,2017].
COMSOL, COMSOL Multiphysics, v. 5.2, [CD-ROM]. Stockholm, Sweden: COMSOL AB, 2017.
L. Bortot, B. Auchmann, I. Cortes Garcia, A. M. Fernandez Navarro,
M. Maciejewski, M. Prioli, S. Schöps, and A. P. Verweij, “A 2-D FiniteElement Model for Electro-Thermal Transients in Accelerator Magnets,”
IEEE Transactions on Mangetics, vol. 54, no. 3, 2018, accepted for
publication.
E. Ravaioli, B. Auchmann, M. Maciejewski, H. ten Kate, and A. Verweij,
“Lumped-element dynamic electro-thermal model of a superconducting
magnet,” Cryogenics, vol. 80, pp. 346–356, 2016.
M. N. Wilson, Superconducting magnets. Clarendon Press, Oxford,
1983.
A. P. Verweij, “Electrodynamics of superconducting cables in accelerator
magnets,” Ph.D. dissertation, Universiteit Twente, 1995.
I. C. Garcia, S. Schops, M. Maciejewski, L. Bortot, M. Prioli, B. Auchmann, and A. P. Verweij, “Optimized field/circuit coupling for the
simulation of quenches in superconducting magnets,” IEEE Journal on
Multiscale and Multiphysics Computational Techniques, vol. 2, pp. 97–
104, May 2017.
L. Bortot, M. Maciejewski, M. Prioli, A. M. F. Navarro, J. B. Ghini,
B. Auchmann, and A. P. Verweij, “A consistent simulation of electrothermal transients in accelerator circuits,” IEEE Transactions on Applied
Superconductivity, vol. 27, no. 4, pp. 1–5, 2017.
M. Valette, L. Bortot, A. Fernandez Navarro, B. Lindstrom, R. Schmidt,
A. Verweij, and D. Wollmann, “Effect of Quench Heater and CLIQ Firing on the Circulating HL-LHC Beam,” in 8th Int. Particle Accelerator
Conf.(IPAC’17), Copenhagen, Denmark, 14 - 19 May, 2017. JACOW,
Geneva, Switzerland, 2017, pp. 2101–2104.
A. M. Fernandez Navarro, M. Maciejewski, A. P. Verweij, L. Bortot, M. Mentink, M. Prioli, B. Auchmann, S. Izquierdo Bermudez,
E. Ravaioli, and S. Yammine, “Simulation of a Quench Event in the
Reconfigured LHC Main Dipole Circuit Including the 11 T Magnets
for the High-Luminosity Upgrade,” IEEE Transactions on Applied
Superconductivity, vol. 28, no. 4, 2018, under review.
M. Prioli, T. Salmi, B. Auchmann, L. Bortot, M. Maciejewski,
M. Mentink, E. Ravaioli, A. Stenvall, and A. Verweij, “Strategies
to reduce the voltage to ground in the FCC main dipole circuits,”
in Contribution submitted to the FCC Week 2017 Conference ,
Berlin, Germany, 29 May - 2 June, 2017. [Online]. Available:
https://indico.cern.ch/event/556692/
M. Mentink, T. M. Salmi, M. Prioli, J. Zhao, and C. Lorin, “Magnet
quench protection of the FCC-hh 16 T block-type dipole magnet
by means of quench absorption coils,” in Contribution submitted to
the 25th International Conference on Magnets Technology (MT25),
August 26 - September 1, 2017, Amsterdam, The Netherlands. [Online].
Available: https://indico.cern.ch/event/445667/contributions/2563167/
| 5cs.CE
|
SINGLE STREAM PARALLELIZATION OF GENERALIZED LSTM-LIKE RNNS ON A GPU
Kyuyeon Hwang and Wonyong Sung
arXiv:1503.02852v1 [cs.NE] 10 Mar 2015
Department of Electrical and Computer Engineering
Seoul National University
Seoul 151-744, South Korea
Email: [email protected], [email protected]
ABSTRACT
Recurrent neural networks (RNNs) have shown outstanding performance on processing sequence data. However, they suffer from long
training time, which demands parallel implementations of the training procedure. Parallelization of the training algorithms for RNNs
are very challenging because internal recurrent paths form dependencies between two different time frames. In this paper, we first
propose a generalized graph-based RNN structure that covers the
most popular long short-term memory (LSTM) network. Then, we
present a parallelization approach that automatically explores parallelisms of arbitrary RNNs by analyzing the graph structure. The
experimental results show that the proposed approach shows great
speed-up even with a single training stream, and further accelerates
the training when combined with multiple parallel training streams.
Index Terms— Recurrent neural network (RNN), long shortterm memory (LSTM), generalization, parallelization, graphics processing unit (GPU)
1. INTRODUCTION
Deep neural networks have shown quite impressive performances
in several pattern recognition applications [1, 2]. Among the deep
neural networks, the feed-forward networks are suitable for processing input data with a fixed length, and they are usually used for image and phoneme recognition. On the other hand, recurrent neural
networks (RNNs) employ feedback inside, and they are suitable for
processing input data whose dimension is not fixed or limited. For
example, automatic speech recognition (ASR) systems can perform
better with an RNN-based language modeling [3].
Since RNNs contain feed-back loops inside, the past input can
be memorized and affect the current output. If RNNs are properly
trained, it is possible to compress the input history effectively and
yield good results even when there are considerable time delays between the input and output. Especially, the long short-term memory
(LSTM) RNN is known to solve the problems with long time lag
very successfully [4].
However, the LSTM RNN employs a very complex component
known to be the memory block. It demands much effort even for
slight modification of the structure because of the difficulty in deriving the corresponding training equation. Thus, it is needed to develop a generalized RNN structure that can be modified easily while
representing LSTM networks perfectly. Previously, a generalized
This work was supported in part by the Brain Korea 21 Plus Project
and the National Research Foundation of Korea (NRF) grants funded by the
Ministry of Education, Science and Technology (MEST), Republic of Korea
(No. 2012R1A2A2A06047297).
LSTM-like RNN structure with real-time recurrent learning (RTRL)
[5] was proposed in [6] with special gated connections. However, we
propose a much more general structure by introducing multiplicative
layers and delayed connections. Also, we derive a backpropagation through time (BPTT) [7] based training algorithm for our RNN
structure, which is generally more flexible than the RTRL-based one.
RNNs also demand very long training time, thus implementation
with GPUs or multiprocessors is needed. However, parallelization
of the network is difficult due to dependency induced by the internal feedback loops. The conventional approach uses independent
multiple training streams that employs plural copies of the network
[8]. However, this inter-stream parallelism demands huge memory,
which is a serious bottleneck for GPU based implementations.
In this paper, we propose a parallelization approach as well as
the generalized RNN structure. For this purpose, we first develop
training algorithms for the generalized RNNs. The training equations of conventional LSTM can be perfectly represented with the
generalized equations. Then, the parallelization approach exposes
single-stream parallelization (intra-stream parallelism) that does not
increase the size of mini-batches as the conventional multi-stream
parallelization (inter-stream parallelism). Experimental results show
that further speed-up can be achieved by combining the two parallelism.
This paper is organized as follows. The generalized LSTM-like
RNN structure is proposed and its training equations are derived in
Section 2. In Section 3, the intra-stream parallelism of the generalized RNNs is explored and combined with the conventional interstream parallelism. In Section 4, experimental results of the proposed approach on a GPU are presented, followed by concluding
remarks in Section 5.
2. GENERALIZATION
To apply our parallelization approach to various types of RNNs,
we first introduce a generalized RNN structure that can represent
complex RNNs using simple basic blocks. This generalization fully
covers advanced LSTM network structures with forget gates and
peephole connections, and their BPTT-based training algorithm.
Also, with the generalized RNN, one can easily design a new RNN
structure quite easily since every equation and the parallelization
approach remain the same.
2.1. Generalized RNN structure
The proposed generalized RNN structure is basically a directed
graph, which consists of a set of nodes and edges. Each node represents a layer and each edge makes a connection between two layers.
Softmax output
There are two types of connections: delayed or not. A delayed
connection makes a fixed amount of delay on the signal, and is used
to construct a recurrent loop. More specifically, the connection m
propagates the activation of the source layer k at the frame t − dm
to the destination layer at the frame t as
zm (t) = Wm yk (t − dm ),
(1)
where zm is the output of the connection m, Wm is the corresponding weight matrix, yk is the activation of the source layer k, and dm
is the amount of delay at the connection m. The value of dm is 0 for
non-delayed connections and larger than 0 for delayed connections.
In an additive layer, the inputs are summed up and the activation
function is applied on it:
X
sk (t) =
zm (t)
(2)
m∈Ak
yk (t) = fk (sk (t)),
(3)
where sk is the state (input), Ak is the set of the indices of the anterior connections, yk is the activation, and fk (·) is the activation
function of the layer k. In addition to the normal additive layers,
multiplicative layers are employed to represent gate units of LSTM
networks. A multiplicative layer performs element-wise multiplication of input vectors (or matrices for batched computation) as follows:
Y
sk,i (t) =
zm,i (t),
(4)
m∈Ak
where the subscript i represents the index of elements in a vector.
For generality, we introduce an aggregation function gk (·) as
sk (t) = gk ({zm (t)|m ∈ Ak }),
(5)
where gk (·) is a vector addition function for an additive layer or an
element-wise multiplication function for a multiplicative layer, or it
can be other nonlinear functions to add further nonlinearity to the
network.
In the previous approach on the generalized LSTMs [6], the gate
units are implemented with gated connections. However, the gated
connection has two input layers, so cannot be regarded as an edge
of a familiar directed graph structure, where each edge has one input
and one output.
In our approach, by introducing the multiplicative layers, LSTM
gates can be regarded as normal nodes in a graph structure, which allows general graph algorithms to be directly applied in Section 3. As
an example, Figure 1 shows a generalized representation of a singlelayer LSTM network with forget gates and peephole connections.
1
1
1
SCC
LSTM layer
Input
Fig. 1. Generalized representation of an LSTM network with forget gates and peephole connections. Thick arrows represent connections with full weight matrices. On the other hand, connections
with the thin arrows have identity weight matrices. The numbers on
the dashed lines indicate the corresponding delay amounts. A nonsingleton strongly connected component (SCC) is drawn, of which
nodes will be grouped into a single recurrent node to make the network acyclic.
These two variables will be back-propagated at the backward pass.
If the layer k is an output layer, δk,j (t) should be initialized by comparing the output with a desired output dk,j (t) according to the error
criterion defined by E(t) and the activation function of the output
layer. Using the minimum cross-entropy criterion with the softmax
activation function,
δk,j (t) = dk,j (t) − yk,j (t).
(9)
If the layer k is not an output layer,
δk,j (t) = −
X X ∂E total (t0 , t1 ) ∂zn,i (t + dn ) ∂yk,j (t)
∂zn,i (t + dn )
∂yk,j (t) ∂sk,j (t)
n∈P i∈I
k
2.2. Training
n
(10)
In this section, BPTT [7] based training equations for the generalized
RNN are derived. The objective is to minimize the following total
error from t0 + 1 to t1 :
X
E total (t0 , t1 ) =
E(t),
(6)
t0 <t≤t1
where E(t) is the error at frame t. For convenience, we define two
derivative variables as
δk,i (t) = −
∂E total (t0 , t1 )
∂sk,i (t)
(7)
m,i (t) = −
∂E total (t0 , t1 )
.
∂zm,i (t)
(8)
=
X X
n,i (t +
dn )Wn,ij fk0 (sk,j (t)),
(11)
n∈Pk i∈In
where Pk is the set of posterior connection indices of the layer k
and In is the set of element indices of the vector zn . Also, m,j (t)
becomes
∂E total (t0 , t1 ) ∂sk,j (t)
∂sk,j (t)
∂zm,j (t)
∂
= δk,j (t)
gk ({zn (t)|n ∈ Ak }),
∂zm,j (t)
m,j (t) = −
(12)
(13)
where k is the index of the destination layer of the connection m.
To truncate errors at t = t00 , we backpropagate the two derivative
variables while t > t00 where t00 ≤ t0 using (11) and (13). After
the backward pass, the truncated error gradient of the connection
m ∈ Pk can be acquired by
∂E total (t0 , t1 )
≈
∂Wm,ij
X
t00 <t≤t1
=−
∂E total (t0 , t1 ) ∂zm,i (t)
∂zm,i (t)
∂Wm,ij
(14)
m,i (t)yk,j (t − dm ).
(15)
X
LSTM layer
t00 <t≤t1
In matrix form, (11) can be represented as
X
δ k (t) =
WnT n (t + dn ) ◦ fk0 (sk (t)),
Recurrent
node
(16)
n∈Pk
where ◦ denotes element-wise vector multiplication. If the layer k is
an additive layer, then (13) becomes
m (t) = δ k (t).
(17)
Fig. 2. Feed-forward representation of the LSTM network that is
depicted in Figure 1.
Otherwise for the multiplicative layer k,
m (t) = δ k (t) ◦
◦
Y
Input
zn (t),
(18)
n∈Ak ,n6=m
Q
where element-wise multiplications are performed with . The error gradient matrix for the connection m ∈ Pk is computed by
X
∇Wm = −
m (t)ykT (t − dm ).
(19)
t00 <t≤t1
The error gradients can be used for the first order optimization methods such as stochastic gradient descent.
3. PARALLELIZATION
Parallelization of RNN computation is quite challenging due to dependencies between two consecutive frames. The state of an RNN
of the frame k cannot be determined until the computation for the
frame k − 1 is finished. In this section, we first develop a parallelization method for the forward and the backward pass with a single
stream (intra-stream parallelism), and then extend the approach to a
multi-stream case (inter-stream parallelism).
3.1. Intra-stream parallelism
The key concept of separating sequential parts from the parallel parts
of an RNN is to determine loops in the RNN and group each loop
into a single special node called a recurrent node. Then, the remaining structure becomes a directed acyclic graph (DAG), which can be
easily parallelized as in a mini-batch based feed-forward neural network computation. Only the internal computations of the recurrent
nodes are performed sequentially.
More specifically, strongly connected components (SCCs) are
found to determine which nodes should be grouped into a recurrent
node. An SCC is a subgraph that is strongly connected, that is, there
are one or more paths between every pair of two vertices inside the
subgraph. An SCC analysis finds a set of SCCs that form a partition
of the vertex set of the original graph. For SCCs that are singletons
and do not contain a self-loop, the original nodes inside the SCCs
remain unchanged. Otherwise, the nodes in each SCC are grouped
into a single recurrent node. Then, the final graph becomes a DAG
and be ready for parallel computation. An example of an LSTM network is shown in Figure 2. One of the famous algorithms for finding
SCCs is the Tarjan’s strongly connected component algorithm [9].
Tarjan’s algorithm also provides a reverse topological sort of the resulting DAG, which is useful to determine the activation order.
Once an RNN is represented as a DAG, the forward computation
becomes very similar to that of feedforward networks. As in the case
of feedforward networks, computations of nodes and edges are performed in a topological order of the DAG. These operations can be
done in parallel over several frames since the network is represented
as a DAG and there are no dependencies between different frames
except the isolated recurrent nodes.
Recurrent nodes are subgraphs of the original RNN and should
be computed sequentially. The computation of a recurrent node from
frame t0 to t1 in the forward pass requires t1 −t0 +1 sequential steps.
In each step of the forward pass, delayed connections are computed
first. Then the remaining part excluding the delayed connections
becomes a DAG and can be computed in a topological order. The
computation of a backward pass can be performed similarly with
reversed topological orders.
The sequential computations of recurrent nodes are quite expensive and often become a bottleneck of the overall performance. To
speed up these sequential parts, we need to employ the multi-stream
parallelization.
3.2. Inter-stream parallelism
Inter-stream parallelism can be explored in the multi-stream mode
where an RNN processes N streams with independent contexts. This
is equivalent to running N independent copies of the RNN. Therefore, the multi-stream mode greatly increases parallelism and the
overall execution speed. Recently, this approach was successfully
applied to speed up language model training with an Elman network
on a GPU [8].
For training an RNN in the multi-stream mode, the input and target streams are usually given by connecting randomly ordered training sequences. Since the lengths of the training sequences are very
long, we apply the efficient version of truncated BPTT(h), denoted
as BPTT(h; h0 ) proposed in [10]. BPTT(h; h0 ) is similar to the or-
·104
0.8
Elman / baseline
Elman / proposed
LSTM / baseline
LSTM / proposed
Processing speed (GFLOPS)
Training speed (words/sec)
1
0.6
0.4
4,000
2,000
1024 / baseline
1024 / proposed
2048 / baseline
2048 / proposed
4096 / baseline
4096 / proposed
Tesla K40
0.2
0
100
0
100
101
102
103
101
102
103
Number of streams
Number of streams
Fig. 3. Comparison of language model training speeds with Elman
and LSTM networks. The LSTM employs forget gates and peephole
connections. The sizes of the input layer, hidden or LSTM layer,
and output layer is 38,000, 512, and 20,000 respectively. The minibatch size is fixed to 1,024, so the error propagates from 1,024/N to
2,048/N − 1 previous steps where N is the number of streams.
dinary truncated BPTT(h) in that the network is unrolled h times.
However, in the forward pass of BPTT(h; h0 ), h0 time steps are computed at once. Also, the error gradients for the recent h0 output errors
are obtained by one iteration. These error gradients are summed up
over the N training streams. Therefore, output errors of total N × h0
frames affect the error gradients when updating weights after backward passes. We call the set of these frames as a mini-batch throughout the paper, as it is equivalent to a mini-batch in stochastic gradient
descent methods of feedforward neural networks.
Increasing N also speeds up the training. However, we cannot
make N very large since the size of a mini-batch, N × h0 , is limited
by the physical memory size of a GPU. Moreover, increasing the
size of a mini-batch results in infrequent update of the weights and
may slow down the convergence [11]. Also, the parameter h0 cannot
be easily modified since the training speed is approximately proportional to the ratio of h0 to h. For simplicity, let us assume h = 2h0
to fix the training speed. In this case, error propagates through h0 to
2h0 − 1 previous time steps in backward pass. Therefore h0 should
be set sufficiently large to solve long time lag problems.
4. EXPERIMENTAL RESULTS
Nvidia Tesla K40 GPU is used for the following experiments. For
all experiments, BPTT(2h; h) is used for simplicity. Since the training algorithm for the generalized RNN structure is mathematically
equivalent to that of Elman or LSTM networks, results with performance measures such as accuracy or the mean squared error (MSE)
are not reported.
To evaluate the proposed parallelization approach, we evaluate
the language model training speed with the multi-stream mode as in
[8]. The RNN architecture is an Elman network with 38,000 input,
512 hidden, and 20,000 output units. The mini-batch size is fixed
to 1,024 to use the same amount of GPU memory. Hence, with N
streams, h = 1,024/N and the error propagates from 1,024/N to
2,048/N − 1 previous time steps. For comparison, an LSTM version of the network with forget gates and peephole connections are
also evaluated. Note that the LSTM network has no self recurrent
Fig. 4. Comparison of GPU processing power utilizations when
training LSTM networks with the three different sizes of LSTM layers: 1,024, 2,048, and 4,096. The input and output layers have the
same size as the LSTM layer. Also, the theoretical peak performance
of Tesla K40 GPU is shown. The mini-batch size is fixed to 1,024.
connection from the output of the LSTM layer to the input of that.
The training speeds are compared in Figure 3 with varying number of streams. Since the baseline approaches does not exploit intrastream parallelism, they show poor training speeds when the number
of streams are small. On the other hand, the proposed approach employs intra-stream parallelism and shows over 10 times of speed-up
over the baseline approach when a single stream is used. Also, with
the proposed approach, we can obtain almost the maximum speed
only with 64 streams. This is a nice advantage since using less number of streams allows RNNs to learn longer time lags when the size
of mini-batch is limited, as discussed in Section 3.2.
To analyze scalability and GPU efficiency with various size of
networks, we perform another experiment with LSTM networks with
forget gates and peephole connections. All layers of each network
have the same size, which is 1,024, 2,048, or 4,096. To examine the
GPU utilizations, we present the number of single-precision floating point operations per second (FLOPS) in Figure 4 along with the
theoretical peak performance of Tesla K40 GPU. Note that only the
operations for parameters and error gradients are counted. Compared to the previous experiment where the input and output layers
are very large, this example is much closer to the deep RNN architectures in terms of the ratio of the sequential computations (inside
the recurrent nodes) to the parallel computations. As shown in the
figure, the GPU utilization gets higher as the layer size or the number of streams increases. Also, the intra-stream parallelism further
accelerates the training especially with the small number of streams.
5. CONCLUDING REMARKS
We introduced a generalized structure for RNNs which covers
LSTM networks with forget gates and peephole connections. This
generalized structure is represented as a directed graph where nodes
and edges correspond to layers and connections, respectively. Due
to the graph representation, we can automatically find loops inside
RNNs using the Tarjan’s strongly connected component algorithm
and explore intra-stream parallelism. The proposed intra-stream parallelism is combined with inter-stream parallelism in multi-stream
mode for further acceleration. The experiments show that exploiting
these two parallelisms greatly speeds up the training task on a GPU.
6. REFERENCES
[1] Geoffrey E Hinton and Ruslan R Salakhutdinov, “Reducing
the dimensionality of data with neural networks,” Science, vol.
313, no. 5786, pp. 504–507, 2006.
[2] Geoffrey Hinton, Li Deng, Dong Yu, George E Dahl, Abdelrahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent
Vanhoucke, Patrick Nguyen, Tara N Sainath, et al., “Deep neural networks for acoustic modeling in speech recognition: The
shared views of four research groups,” Signal Processing Magazine, IEEE, vol. 29, no. 6, pp. 82–97, 2012.
[3] Tomas Mikolov, Stefan Kombrink, Lukas Burget, JH Cernocky, and Sanjeev Khudanpur, “Extensions of recurrent neural network language model,” in Acoustics, Speech and Signal
Processing (ICASSP), 2011 IEEE International Conference on.
IEEE, 2011, pp. 5528–5531.
[4] Felix A Gers, Jürgen Schmidhuber, and Fred Cummins,
“Learning to forget: Continual prediction with LSTM,” Neural
computation, vol. 12, no. 10, pp. 2451–2471, 2000.
[5] Ronald J Williams and David Zipser, “A learning algorithm for
continually running fully recurrent neural networks,” Neural
computation, vol. 1, no. 2, pp. 270–280, 1989.
[6] Derek Monner and James A Reggia, “A generalized LSTMlike training algorithm for second-order recurrent neural networks,” Neural Networks, vol. 25, pp. 70–83, 2012.
[7] Paul J Werbos, “Backpropagation through time: what it does
and how to do it,” Proceedings of the IEEE, vol. 78, no. 10, pp.
1550–1560, 1990.
[8] Xie Chen, Yongqiang Wang, Xunying Liu, Mark JF Gales, and
Philip C Woodland, “Efficient GPU-based training of recurrent neural network language models using spliced sentence
bunch,” in INTERSPEECH, 2014.
[9] Robert Tarjan, “Depth-first search and linear graph algorithms,” SIAM journal on computing, vol. 1, no. 2, pp. 146–
160, 1972.
[10] Ronald J Williams and Jing Peng, “An efficient gradient-based
algorithm for on-line training of recurrent network trajectories,” Neural Computation, vol. 2, no. 4, pp. 490–501, 1990.
[11] Richard H Byrd, Gillian M Chin, Jorge Nocedal, and Yuchen
Wu, “Sample size selection in optimization methods for machine learning,” Mathematical programming, vol. 134, no. 1,
pp. 127–155, 2012.
| 9cs.NE
|
DID: Distributed Incremental Block Coordinate Descent for
Nonnegative Matrix Factorization
Tianxiang Gao, Chris Chu
arXiv:1802.08938v1 [cs.LG] 25 Feb 2018
Department of Electrical and Computer Engineering, Iowa State University, Ames, IA, 50011, USA
{gaotx, cnchu}@iastate.edu
Abstract
Nonnegative matrix factorization (NMF) has attracted much
attention in the last decade as a dimension reduction method
in many applications. Due to the explosion in the size of data,
naturally the samples are collected and stored distributively in
local computational nodes. Thus, there is a growing need to
develop algorithms in a distributed memory architecture. We
propose a novel distributed algorithm, called distributed incremental block coordinate descent (DID), to solve the problem. By adapting the block coordinate descent framework,
closed-form update rules are obtained in DID. Moreover, DID
performs updates incrementally based on the most recently
updated residual matrix. As a result, only one communication
step per iteration is required. The correctness, efficiency, and
scalability of the proposed algorithm are verified in a series
of numerical experiments.
1 Introduction
Nonnegative
matrix
factorization
(NMF)
(Lee and Seung 1999) extracts the latent factors in a
low dimensional subspace. The popularity of NMF
is due to its ability to learn parts-based representation by the use of nonnegative constraints. Numerous successes have been found in document clustering
(Xu and Gong 2004;
Lu, Hong, and Wang 2017),
computer
vision
(Lee and Seung 1999),
signal
processing
(Gao, Olofsson, and Lu 2016;
Lu, Hong, and Wang 2017), etc.
Suppose a collection of N samples with M nonnegative measurements is denoted in matrix form X ∈ RM×N
,
+
where each column is a sample. The purpose of NMF is to
approximate X by a product of two nonnegative matrices
B ∈ RM×K
and C ∈ RK×N
with a desired low dimen+
+
sion K, where K ≪ min{M, N }. The columns of matrix
B can be considered as a basis in the low dimension subspace, while the columns of matrix C are the coordinates.
NMF can be formulated as an optimization problem in (1):
minimize
f (B, C) =
subject to
B, C ≥ 0,
B,C
1
2
kX − BCkF
2
(1a)
(1b)
Copyright © 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
where “≥ 0” means element-wise nonnegative, and k·kF is
the Frobenius norm. The problem (1) is nonconvex with respect to variables B and C. Finding the global minimum is
NP-hard (Vavasis 2009). Thus, a practical algorithm usually
converges to a local minimum.
Many
algorithms
have
been
proposed
to
solve NMF such as multiplicative updates (MU)
(Lee and Seung 2001), hierarchical alternating least
square
(HALS)
(Cichocki, Zdunek, and Amari 2007;
Li and Zhang 2009), alternating direction multiplier
method (ADMM) (Zhang 2010), and alternating nonnegative least square (ANLS) (Kim and Park 2011). Amongst
those algorithms, ANLS has the largest reduction of objective value per iteration since it exactly solves nonnegative
least square (NNLS) subproblems using a block principal
pivoting (BPP) method (Kim and Park 2011). Unfortunately, the computation of each iteration is costly. The
algorithm HALS, on the other hand, solves subproblems
inexactly with cheaper computation and has achieved
faster convergence in terms of time (Kim and Park 2011;
Gillis and Glineur 2012). Instead of iteratively solving the
subproblems, ADMM obtains closed-form solutions by
using auxiliary variables. A drawback of ADMM is that it is
sensitive to the choice of the tuning parameters, even to the
point where poor parameter selection can lead to algorithm
divergence (Sun and Fevotte 2014).
Most of the proposed algorithms are intended for centralized implementation, assuming that the whole data matrix can be loaded into the RAM of a single computer
node. In the era of massive data sets, however, this assumption is often not satisfied, since the number of samples is too large to be stored in a single node. As a result, there is a growing need to develop algorithms in distributed system. Thus, in this paper, we assume the number of samples is so large that the data matrix is collected
and stored distributively. Such applications can be found
in e-commerce (e.g., Amazon), digital content streaming
(e.g., Netflix) (Koren, Bell, and Volinsky 2009) and technology (e.g., Facebook, Google) (Tan, Cao, and Fong 2016),
where they have hundreds of millions of users.
Many distributed algorithms have been published recently. The distributed MU (Liu et al. 2010;
Yin, Gao, and Zhang 2014) has been proposed as
the first distributed algorithm to solve NMF. How-
ever, MU suffers from slow and ill convergence in
some
cases
(Lin 2007).
Kannan, Ballard, and Park
(Kannan, Ballard, and Park 2016) proposed high performance ANLS (HPC-ANLS) using 2D-grid partition of
a data matrix such that each node only stores a submatrix
of the data matrix. Nevertheless, six communication steps
per iteration are required to obtain intermediate variables
so as to solve the subproblems. Thus, the communication
overhead is significant. Moreover, the computation is costly
as they use ANLS framework. The most recent work is
distributed HALS (D-HALS) (Zdunek and Fonal 2017).
However, they assume the factors B and C can be stored
in the shared memory of the computer nodes, which may
not be the case if N is large. Boyd et al. (Boyd et al. 2011)
suggested that ADMM has the potential to solve NMF
distributively. Du et al. (Du et al. 2014) demonstrated this
idea in an algorithm called Maxios. Similar to HPC-ANLS,
the communication overhead is expensive, since every
latent factor or auxiliary variable has to be gathered and
broadcasted over all computational nodes. As a result,
eight communication steps per iteration are necessary. In
addition, Maxios only works for sparse matrices since they
assume the whole data matrix is stored in every computer
node.
In this paper, we propose a distributed algorithm based
on block coordinate descent framework. The main contributions of this paper are listed below.
• We propose a novel distributed algorithm, called distributed incremental block coordinate descent (DID). By
splitting the columns of the data matrix, DID is capable
of updating the coordinate matrix C in parallel. Leveraging the most recent residual matrix, the basis matrix B is
updated distributively and incrementally. Thus, only one
communication step is needed in each iteration.
• A scalable and easy implementation of DID is derived using Message Passing Interface (MPI). Our implementation does not require a master processor to synchronize.
• Experimental results showcase the correctness, efficiency,
and scalability of our novel method.
The paper is organized as follows. In Section 2, the previous works are briefly reviewed. Section 3 introduces a distributed ADMM for comparison purpose. The novel algorithm DID is detailed in Section 4. In Section 5, the algorithms are evaluated and compared. Finally, the conclusions
are drawn in Section 6.
2 Previous Works
In this section we briefly introduce three standard algorithms
to solve NMF problem (1), i.e., ANLS, HALS, and ADMM,
and discuss the parallelism of their distributed versions.
Notations. Given a nonnegative matrix X ∈ RM×N
with
+
M rows and N columns, we use xri ∈ R1×N
to
denote
its
+
i-th row, xj ∈ RM×1
to
denote
the
j-th
column,
and
x
ij ∈
+
R+ to denote the entry in the i-th row and j-th column. In
×1
addition, we use xrT
∈ RN
and xTj ∈ R1×M
to denote
+
+
i
the transpose of i-th row and j-th column, respectively.
2.1 ANLS
The optimization problem (1) is biconvex, i.e., if either
factor is fixed, updating another is in fact reduced to a
nonnegative least square (NNLS) problem. Thus, ANLS
(Kim and Park 2011) minimizes the NNLS subproblems
with respect to B and C, alternately. The procedure is given
by
2
C := argminC≥0 kX − BCkF
B := argminB≥0 kX −
2
BCkF
(2a)
.
(2b)
The optimal solution of a NNLS subproblem can be
achieved using BPP method.
A naive distributed ANLS is to parallel Cminimization step in a column-by-column manner and
B-minimization step in a row-by-row manner. HPC-ANLS
(Kannan, Ballard, and Park 2016) divides the matrix X
into 2D-grid blocks, the matrix B into Pr row blocks, and
the matrix C into Pc column blocks so that the memory
NK
requirement of each node is O( PMN
+ MK
Pr + Pc ), where
r Pc
Pr is the number of rows processor and Pc is the number of
columns processor such that P = Pc Pr is the total number
of processors. To really perform updates, the intermediate
variables CC T , XC T , B T B, and B T X are computed and
broadcasted using totally six communication steps. Each of
them has a cost of log P · (α + β · N K), where α is latency,
and β is inverse bandwidth in a distributed memory network
model (Chan et al. 2007). The analysis is summarized in
Table 1.
2.2 HALS
Since the optimal solution to the subproblem is not required when updating one factor, a comparable method,
called HALS, which achieves an approximate solution is
proposed by (Cichocki, Zdunek, and Amari 2007). The algorithm HALS successively updates each column of B and
row of C with an optimal solution in a closed form.
The objective function in (1) can be expressed with respect to the k-th column of B and k-th row of C as follows
2
X − BC
F
= X−
PK
r
i=1 bi ci
= X−
r
i6=k bi ci
P
− bk crk
2
F
,
P
Let A , X − i6=k bi cri and fix all the variables except bk
or crk . We have subproblems in bk and crk
min kA − bk crk kF ,
2
(3a)
min kA − bk crk k2F
(3b)
bk ≥0
crk ≥0
By setting the derivative with respect to bk or crk to zero and
projecting the result to the nonnegative region, the optimal
solution of bk and crk can be easily written in a closed form
−1
bk := (crk crT
(AcrT
(4a)
k )
k ) +
T −1 T
r
(4b)
ck := (bk bk ) (A bk ) +
where [z]+ is max{0, z}. Therefore, we have K inner-loop
iterations to update every pair of bk and crk . With cheaper
computational cost, HALS was confirmed to have faster convergence in terms of time.
Zdunek and Fonal in 2017 proposed a distributed version
of HALS, called DHALS. They also divide the data matrix
X into 2D-grid blocks. Comparing with HPC-ANLS, the
resulting algorithm DHALS only requires two communication steps. However, they assume matrices B and C can be
loaded into the shared memory of a single node. Therefore,
DHALS is not applicable in our scenario where we assume
N is so big that even the latent factors are stored distributively. See the detailed analysis in Table 1.
2.3
ADMM
The algorithm ADMM (Zhang 2010) solves the NMF problem by alternately optimizing the Lagrangian function with
respect to different variables. Specifically, the NMF (1) is
reformulated as
1
minimize
kX − W Hk2F
(5a)
B,C,W,H
2
subject to
B = W, C = H
(5b)
B, C ≥ 0,
(5c)
where W ∈ RM×K and H ∈ RK×N are auxiliary variables without nonnegative constraints. The augmented Lagrangian function is given by
1
2
kX − W HkF + hΦ, B − W i
2
ρ
ρ
2
2
+ hΨ, C − Hi + kB − W kF + kC − HkF (6)
2
2
L(B, C, W, H; Φ, Ψ)ρ =
where Φ ∈ RM×K and Ψ ∈ RK×N are Lagrangian multipliers, h·, ·i is the matrix inner product, and ρ > 0 is the
penalty parameter for equality constraints. By minimizing L
with respect to W , H, B, C, Φ, and Ψ one at a time while
fixing the rest, we obtain the update rules as follows
be broadcasted to all other computational nodes. As a consequence, Maxios requires theoretically eight communication
steps per iteration and only works for sparse matrices. Table
1 summarizes the analysis.
3 Distributed ADMM
This section derives a distributed ADMM (DADMM)
for comparison purpose. DADMM is inspired by
another centralized version in (Boyd et al. 2011;
Hajinezhad et al. 2016), where the update rules can be
easily carried out in parallel, and is stable when ρ is small.
As the objective function in (1) is separable in columns,
we divide matrices X and C into column blocks of P parts
P
X1
1
2
2
kX − BCkF =
kXi − BCi k2 ,
2
2
i=1
M×Ni
i
where Xi ∈ R+
and Ci ∈ RK×N
are column blocks
+
PP
of X and C such that i=1 Ni = N . Using a set of auxiliary
variables Yi ∈ RM×Ni , the NMF (1) can be reformulated as
P
X
1
minimize
Yi ,B,C
i=1
2
kXi − Yi k2F
Yi = BCi ,
B, C ≥ 0.
subject to
(9a)
for i = 1, 2, · · · , P (9b)
(9c)
The associated augmented Lagrangian function is given by
L(Yi , B, C; Λi )ρ =
P
X
1
i=1
+
P
X
2
kXi − Yi k2F
hΛi , Yi − BCi i +
P
X
ρ
i=1
i=1
M×K
W := (XH T + Φ + ρB)(HH T + ρIK )−1
(7a)
where Λi ∈ R
sulting ADMM is
H := (W T W + ρIK )−1 (W T X + Ψ + ρC)
B := [W − Φ/ρ]+
(7b)
(7c)
Yi := argmin
C := [H − Ψ/ρ]+
(7d)
Φ := Φ + ρ(B − W )
Ψ := Ψ + ρ(C − H)
(7e)
(7f)
where IK ∈ RK×K is the identity matrix. The auxiliary
variables W and H facilitate the minimization steps for B
and C. When ρ is small, however, the update rules for W and
H result in unstable convergence (Sun and Fevotte 2014).
When ρ is large, ADMM suffers from a slow convergence.
Hence, the selection of ρ is significant in practice.
Analogous to HPC-ANLS, the update of W and B can be
parallelized in a column-by-column manner, while the update of H and C in a row-by-row manner. Thus, Maxios
(Du et al. 2014) divides matrix W and B in column blocks,
and matrix H and C in row blocks. However, the communication overhead is expensive since one factor update depends on the others. Thus, once a factor is updated, it has to
(8)
Yi
2
2
kYi − BCi kF , (10)
are the Lagrangian multipliers. The re-
1
ρ
2
2
kXi − Yi k2 + kΛi /ρ + Yi − BCi kF
2
2
(11a)
2
Ci := argmin kΛi /ρ + Yi − BCi k2
(11b)
B := argmin kΛ/ρ + Y − BCk2F
(11c)
Λi := argmaxhΛi , Yi − BCi i
(11d)
Ci ≥0
B≥0
Λi
where Λ , [Λ1 Λ2 · · · ΛP ] and Y , [Y1 Y2 · · · YP ].
Clearly, the Yi update has a closed-form solution by taking
the derivate and setting it to zero, i.e.,
Yi :=
1
(Xi − Λi + ρBCi )
1+ρ
(12)
Moreover, the updates for Yi , Ci , and Λi can be carried out
in parallel. Meanwhile, B needs a central processor to update since the step (11c) requires the whole matrices Y , C,
Algorithm
HPC-ANLS
D-HALS
Maxios
DADMM
DBCD
DID
Runtime
BPP
O (M N K(1/Pc + 1/Pr ))
O K 3 + M N K/P
BPP
O (M N K/P )
O (M N K/P )
Memory per processor
O (M N/(Pc Pr ) + M K/Pr + N K/Pc )
O (M N/(Pc Pr ) + M K + N K)
O (M N )
O (M N/P + M K)
O (M N/P + M K)
O (M N/P + M K)
Communication time
3(α + βN K) log Pr + 3(α + βM K) log Pc
(α + βN K) log Pr + (α + βM K) log Pc
4(2α + β(N + M )K) log P
(α + βM K) log P
K(α + βM K) log P
(α + βM K) log P
Communication volume
O (M KPc + N KPr )
O (M KPc + N KPr )
O ((M + N )KP )
O (M KP )
O (M KP )
O (M KP )
Table 1: Analysis of distributed algorithms per iteration on runtime, memory storage, and communication time and volume.
and Λ. If we use the solver BPP, however, we do not really
need to gather those matrices, because the solver BPP in fact
does not explicitly need Y , C, and Λ. Instead, it requires two
intermediate variables W , CC T and H , (Λ/ρ + Y )C T ,
which can be computed as follows:
W , CC T =
P
X
Ci CiT ,
(13a)
4.1 Distributed Block Coordinate Descent
i=1
H , (Λ/ρ + Y )C T =
P
X
(Λi /ρ + Yi )CiT .
(13b)
i=1
It is no doubt that those intermediate variables can be calculated distributively. Let Ui = Λi /ρ, which is called scaled
dual variable. Using the scaled dual variable, we can express
DADMM in a more efficient and compact way. A simple
MPI implementation of algorithm DADMM on each computational node is summarized in Algorithm 1.
Algorithm 1: DADMM for each computational node
Input: Xi , Ci , B
Initialize P processors, along with Yi , B, Ci , Xi
repeat
1
Ui := Ui + (Yi − BCi )
1
2
Yi := 1+ρ
(Xi − ρUi + ρBCi )
3
4
5
2
Ci := argminCi ≥0 kUi + Yi − BCi k2
(W, H) := Allreduce(Ci CiT , (Ui + Yi )CiT )
B := BPP(W, H)
until stopping criteria satisfied;
At line 4 in Algorithm 1, theoretically we need a master
processor to gather Ci CiT and (Ui +Yi )CiT from every local
processor and then broadcast the updated value of CC T and
(U + Y )C T back. As a result, the master processor needs
a storage of O (M KP ). However, we use a collaborative
operation called Allreduce (Chan et al. 2007). Leveraging it,
the master processor is discarded and the storage of each
processor is reduced to O (M K).
4
expensive as it is required to find optimal solutions of subproblems to ensure convergence. In this section, we will propose another distributed algorithm that adapts block coordinate descent framework and achieves approximate solutions
at each iteration. Moreover, leveraging the current residual
matrix facilitates the update for matrix B so that columns of
B can be updated incrementally.
Distributed Incremental Block Coordinate
Descent
The popularity of ADMM is due to its ability of carrying out
subproblems in parallel such as DADMM in Algorithm 1.
However, the computation of ADMM is costly since it generally involves introducing new auxiliary variables and updating dual variables. The computational cost is even more
We firstly introduce a naive parallel and distributed algorithm, which is inspired by HALS, called distributed block
coordinate descent (DBCD). Since the objective function in
(1) is separable, the matrix X is partitioned by columns, then
each processor is able to update columns of C in parallel,
and prepare messages concurrently to update matrix B.
Analogous to DADMM, the objective function in (1) can
be expanded as follows
2
X − BC
F
=
PN
j=1
2
xj − Bcj
=
PN
j=1
xj −
PK
2
k=1 bk ckj
By coordinate descent framework, we only consider one element at a time. To update cij , we fix the rest of variables as
constant, then the objective function becomes
X
XN
2
xj −
bk ckj − bi cij .
(14)
j=1
k6=i
Taking the partial derivative of the objective function (14)
with respect to cij and setting it to zero, we have
X
= 0.
(15)
bTi bi cij − xj −
bk ckj
k6=i
The optimal solution of cij can be easily derived in a closed
form as follows
"
#
P
bTi (xj − k6=i bk ckj )
cij :=
(16a)
bTi bi
+
T
b (xj − Bcj + bi cij )
= i
(16b)
bTi bi
+
bTi (xj − Bcj )
= cij +
(16c)
bTi bi
+
Based on the equation (16c), the j-th column of C is required so as to update cij . Thus, updating a column cj has to
be sequential. However, the update can be executed in parallel for all j’s. Therefore, the columns of matrix C can be
updated independently, while each component in a column
cj is optimized in sequence.
The complexity of updating each cij is O (M K).
Thus, the entire complexity of updating matrix C is
O M N K 2 /P . This complexity can be reduced by bringing xj − Bcj outside the loop and redefining as ej ,
xj − Bcj . The improved update rule is
ej := ej + bi cij
T
b ej
cij := iT
bi bi +
(17a)
(17b)
ej := ej − bi cij
(17c)
By doing so, the complexity is reduced to O (M N K/P ).
The analogous derivation can be carried out to update the
i-th column of matrix B, i.e., bi . By taking partial derivative
of the objective function (14) with respect to bi and setting
it to zero, we have equation
N
X
j=1
X
bi cij − xj −
k6=i
bk ckj
cij = 0
Solving this linear equation gives us a closed-form
optimal solution of bi
" PN
#
j=1 (xj − Bcj + bi cij )cij
bi :=
PN 2
j=1 cij
+
#
"
PN
j=1 (xj − Bcj )cij
= bi +
PN 2
j=1 cij
+
(X − BC)crT
i
= bi +
cri crT
i
+
(18)
to the
(19a)
zj ,
c2ij
4.2 Incremental Update for bi
(19c)
(20a)
(20b)
The vector yj and scaler zj can be computed in parallel. After receiving messages including yj ’s and zj ’s from other
processors, a master processor updates the column bi as a
PN
scaled summation of yj with scaler z , j=1 zj , that is,
bi := [y/z]+
PN
Algorithm 2: DBCD for each computational node
Input: xj , cj , B
repeat
// Update C
ej := xj − Bcj
for all i ∈ {1, 2, · · · , K} do
Update cij using equations (17)
end
// Update B
for all i ∈ {1, 2, · · · , K} do
ej = ej + bi cij
yj = ej cij
zj = c2ij
(y, z) = Allreduce(yj , zj )
bi := [y/z]+
ej = ej − bi cij
end
until stopping criteria satisfied;
(19b)
Unfortunately, there is no way to update bi in parallel since the equation (19c) involves the whole matrices X and C. That is the reason why sequential algorithms can be easily implemented in the shared memory but cannot directly be applied in distributed memory. Thus, other works (Kannan, Ballard, and Park 2016;
Zdunek and Fonal 2017; Du et al. 2014) either use gather
operations to collect messages from local processors or assume small size of the latent factors.
By analyzing the equation (19a), we discover the potential
parallelism. We define a vector yj and a scaler zj as follows
yj , (xj − Bcj + bi cij )cij = (ej + bi cij )cij
identify vectors yj ’s and scalars zj ’s to update matrix B, and
their computation can be executed concurrently among computational nodes. A MPI implementation of this algorithm
for each processor is summarized in Algorithm 2.
(21)
where y , j=1 yj . Thus, the update for matrix B can be
executed in parallel but indirectly. The complexity of updating bi is O (M N/P ) as we reserve error vector ej and concurrently compute yj and zj . The complexity of updating
entire matrix B is O (M N K/P ).
By partitioning the data matrix X by columns, the update
for matrix C can be carried out in parallel. In addition, we
The complexity of algorithm DBCD is O (M N K/P ) per
iteration, which is perfectly parallelizing a sequential block
coordinate descent algorithm. However, the performance of
DBCD could be deficient due to the delay in network. In
principle, DBCD sends totally KP messages to a master
processor per iteration, which is even more if we implement
DBCD using Allreduce. Any delay of a message could cause
a diminished performance. In contrast, the algorithm DID
has a novel way to update matrix B incrementally using only
a single message from each processor per iteration.
To successfully update matrix B, the bottleneck is to iteratively compute yj and zj for associated bi since once bi
is updated, the yj and zj have to be recomputed due to the
change occurred in matrix B from equation (19b). Nevertheless, we discovered this change can be represented as several
arithmetic operations. Thus, we in fact do not need to communicate every time in order to update each bi .
Suppose that after t-th iteration, the i-th column of matrix
B is given, i.e., bti , and want to update it to bt+1
. Let E =
i
X − BC, which is the most current residual matrix after t-th
iteration. From equation (19c), we have
EcrT
t
i
bt+1
:=
b
+
(22)
i
i
cri crT
i
+
Once we update bti to bt+1
, we need to update bi in matrix B
i
so as to get new E to update the next column of B, i.e., bi+1 .
However, we do not really need to recalculate E. Instead, we
can update the value by
cri
E := E + bti cri − bt+1
i
(23)
We define and compute a variable δbi as
δbi ,
bt+1
i
−
bti .
(24)
Using the vector δbi , we have a compact form to update E
E := E − δbi cri
(25)
The updated E is substituted into the update rule of bi+1 in
equation (22), and using bti+1 we obtain
(E − δbi cri )crT
i+1
t+1
t
bi+1 := bi+1 +
(26a)
cri+1 crT
i+1
+
EcrT
cri crT
i+1
i+1
t
(26b)
= bi+1 + r rT − r rT δbi
ci+1 ci+1
ci+1 ci+1
+
In the equation (26b), the first two terms is the same as general update rule for matrix B in DBCD, where Eci+1 can
be computed distributively in each computational node. On
the other hand, the last term allows us to update the column
bi+1 still in a closed form but without any communication
step. Therefore, the update for matrix B can be carried out
incrementally and the general update rule is given by
"
#
P
r rT
EcrT
k<i (ci ck )δbk
t+1
t
i
bi := bi + r rT −
(27)
ci ci
cri crT
i
+
Comparing to the messages used in DBCD, i.e., (yj , zj ), we
need to compute the coefficients for the extra term, that is,
cri crT
k for all k < i. Thus, a message communicated among
processors contains two parts: the weighted current residual
matrix Wj , and a lower triangular matrix Vj maintaining the
inner product of matrix C. The matrices Wj and Vj are defined as below
#
"
|
|
···
|
(28)
Wj , ej c1j ej c2j · · · ej cKj
|
|
···
|
2
c1j
0
0
···
0
c2j c1j
c22j
0
···
0
Vj , .
(29)
.
.
..
..
..
..
.
0
cKj c1j
cKj c2j
cKj c3j
···
c2Kj
Using variables Wj and Vj , the update rule to columns of
matrix B becomes
#
"
X
bi := bi + wi /vii −
(vik /vii )δbk
(30)
k<i
+
where wi is the i-th column of matrix W , vij is the i-th
component of j-th column of matrix V , and matrices W and
V are the summations of matrices Wj and Vj , respectively,
PN
PN
i.e., W , j=1 Wj and V , j=1 Vj .
For each processor, they store a column of X, a column of
C, and the matrix B. They execute the same algorithm and a
MPI implementation of this incremental algorithm for each
computational node is summarized in Algorithm 3. Clearly,
the entire computation is unchanged and the volume of message stays the same as DBCD, but the number of communication is reduced to once per iteration.
Algorithm 3: DID for each computational node
Input: xj , cj , B
repeat
// Update C
ej := xj − Bcj
for all i ∈ {1, 2, · · · , K} do
Update cij using equations (17)
end
Compute Wj and Vj from equations (28) and (29).
(W, V ) := Allreduce(Wj , Vj )
// Update B
for all i ∈ {1,
2, · · · , K} doP
bt+1
:= bti + wi /vii − k<i (vik /vii )δbk +
i
δbi := bt+1
− bti
i
end
until stopping criteria satisfied;
5
Experiments
We conduct a series of numerical experiments to compare
the proposed algorithm DID with HALS, ALS, ADMM,
BCD, DBCD, DADMM, and HPC-ANLS. The algorithm
BCD is the sequential version of DBCD. Due to the
ill convergence of ADMM and Maxios in (Zhang 2010;
Du et al. 2014), we derive DADMM in Section 3 and set
ρ = 1 as default. Since we assume M and K are much
smaller than N , HPC-ANLS only has column partition of
the matrix X, i.e., Pc = P and Pr = 1.
We use a cluster1 that consists of 48 SuperMicro servers
each with 16 cores, 64 GB of memory, GigE and QDR
(40Gbit) InfiniBand interconnects. The algorithms are implemented in C code. The linear algebra operations use GNU
Scientific Library (GSL) v2.42 (Gough 2009). The Message
Passing Interface (MPI) implementation OpenMPI v2.1.03
(Gabriel et al. 2004) is used for communication. Note that
we do not use multi-cores in each server. Instead, we use a
single core per node as we want to achieve consistent communication overhead between cores.
Synthetic datasets are generated with number of samples
N = 105 , 106 , 107 and 108 . Due to the storage limits of the
computer system we use, we set the dimension M = 5 and
low rank K = 3, and utilize P = 16 number of computational nodes in the cluster. The random numbers in the synthetic datasets are generated by the Matlab command rand
that are uniformly distributed in the interval [0, 1].
We also perform experimental comparisons on four realworld datasets. The MNIST dataset4 of handwritten digits
has 70,000 samples of 28x28 image. The 20News dataset5 is
a collection of 18,821 documents across 20 different newsgroups with totally 8,165 keywords. The UMist dataset6
1
http://www.hpc.iastate.edu/
http://www.gnu.org/software/gsl/
3
https://www.open-mpi.org/
4
http://yann.lecun.com/exdb/mnist/
5
http://qwone.com/~jason/20Newsgroups/
6
https://cs.nyu.edu/~roweis/data.html
2
0.25
0.12
4000
HALS
DBCD
0.1
ANLS
Time (seconds)
×10 -3
E r/E 0
BCD
15
0.06
3000
DADMM
0.15
×10 -3
0
r
E /E
HPC-ANLS
ADMM
0.08
Computation
Communication
3500
0.2
0.1
10
DID
15
10
2500
2000
1500
0.04
5
5
1000
0.05
0.02
1500
2000
2500
3000
3500
150
200
250
300
500
0
0
1000
2000
3000
4000
5000
6000
7000
0
0
100
200
300
400
Time (seconds)
500
600
700
800
900
1000
0
DBCD
Time (seconds)
(a) Sequential
DID
HPC-ANLS
DADMM
(c) Computation v.s. communication
(b) Distributed
Figure 1: Convergence behaviors of different algorithms with respect to time consumption of communication and computation
on the dataset with N = 108 samples.
N
105
106
107
108
MNIST
20News
UMist
YaleB
HALS
1281
225
596
339
495
302
677
1001
ANLS
141
238
1120
163
197
169
1001
352
ADMM
170
115
1191
97
199
169
953
224
Number of iterations
BCD HPC-ANLS
549
141
396
238
654
1120
302
163
492
197
231
169
622
1001
765
352
DADMM
170
115
1191
97
199
169
953
224
DBCD
549
396
654
302
492
231
622
765
DID
549
396
654
302
492
231
622
765
HALS
16.88
36.86
587.47
3779.11
705.32
2550.02
314.72
223.58
ANLS
56.59
630.43
29234.61
43197.12
395.61
745.28
657.14
201.22
ADMM
45.88
476.83
31798.51
27590.16
610.65
714.61
836.76
149.35
Time (seconds)
BCD HPC-ANLS
10.61
4.31
95.24
50.47
909.76
2372.47
8808.92
10172.55
942.68
31.84
2681.49
131.12
422.11
492.72
236.13
50.69
DADMM
3.46
37.06
2563.47
5742.37
46.17
172.69
471.01
40.61
DBCD
1.42
14.04
126.01
785.57
170.65
651.52
92.49
44.08
DID
1.17
8.61
106.60
610.09
133.50
559.70
82.34
36.45
Table 2: Performance comparison for algorithms on synthetic and real datasets with P = 16 number of computing nodes.
contains 575 images of 20 people with the size of 112x92.
The YaleB dataset7 includes 2,414 images of 38 individuals
with the size of 32x32. The MNIST and 20News datasets are
sparse, while UMist and YaleB are dense.
The algorithms HALS, (D)BCD, and DID could fail if
kbi k or kcri k is close to zero. This could appear if B or C is
badly scaled. That means the entries of E = X − BC are
strictly negative. We avoid this issue by using well scaled
initial points for the synthetic datasets and K-means method
to generate the initial values for the real datasets. All the
algorithms are provided with the same initial values.
When an iterative algorithm is executed in practice, a
stopping criteria is required. In our experiments, the stopping criteria is met if the following condition is satisfied
Et
2
F
≤ ǫ E0
2
F
,
(31)
where E t is the residual matrix after t-th iteration. Throughout the experiments, we set ǫ = 10−6 as default. In addition,
we combine the stopping criterion with a limit on time of
24 hours and a maximum iteration of 1000 for real datasets.
The experimental results are summarized in the Table 2.
Correctness In principle, the algorithms HALS, (D)BCD,
and DID have the same update rules for the latent factors
B and C. The difference is the update order. The algorithm
DID has the exact same number of iterations as BCD and
DBCD, which demonstrates the correctness of DID.
7
http://www.cad.zju.edu.cn/home/dengcai/Data/FaceData.html
Efficiency As presented in Table 2, DID always converges
faster than the other algorithms in term of time. HALS and
BCD usually use a similar number of iterations to reach the
stopping criteria. ANLS and ADMM use much fewer iterations to converge. Thanks to auxiliary variables, ADMM
usually converges faster than ANLS. Figure 1(a) shows that
comparing with HALS, BCD actually reduces the objective
value a lot at the beginning but takes longer to finally converge. Such phenomenon can also be observed in the comparison between ANLS and ADMM. In Figure 1(b), DID
is faster than DBCD. The reason is shown in Figure 1(c)
that DID involves much less communication overhead than
DBCD. Based on the result in Table 2, DID is about 1015% faster than DBCD by incrementally updating matrix B.
(HPC-)ANLS works better in MNIST and 20News datasets
because these datasets are very sparse.
Scalability As presented in Table 2, the runtime of DID
scales linearly as the number of samples increases, which is
much better than the others. It can usually speed up a factor
of at least 10 to BCD using 16 nodes. (D)ADMM is also
linearly scalable, which is slightly better than (HPC-)ANLS.
Due to the costly computation, (D)ADMM is not preferred
to solve NMF problems.
6
Conclusion
In this paper, we proposed a novel distributed algorithm DID
to solve NMF in a distributed memory architecture. Assume
the number of samples N to be huge, DID divides the matrices X and C into column blocks so that updating the matrix
C is perfectly distributed. Using the variables δb, the ma-
trix B can be updated distributively and incrementally. As a
result, only a single communication step per iteration is required. The algorithm is implemented in C code with OpenMPI. The numerical experiments demonstrated that DID has
faster convergence than the other algorithms. As the update only requires basic matrix operations, DID achieves
linear scalability, which is observed in the experimental results. In the future work, DID will be applied to the cases
where updating matrix B is also carried out in parallel. Using the techniques introduced by (Hsieh and Dhillon 2011)
and (Gillis and Glineur 2012), DID has the possibility to be
accelerated. How to better treat sparse datasets is also a potential research direction.
References
[Boyd et al. 2011] Boyd, S.; Parikh, N.; Chu, E.; Peleato, B.;
and Eckstein, J. 2011. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine Learning 1–122.
[Chan et al. 2007] Chan, E.; Heimlich, M.; Purkayastha, A.;
and Van De Geijn, R. 2007. Collective communication:
theory, practice, and experience. Concurrency and Computation: Practice and Experience 1749–1783.
[Cichocki, Zdunek, and Amari 2007] Cichocki, A.; Zdunek,
R.; and Amari, S.-i. 2007. Hierarchical als algorithms for
nonnegative matrix and 3d tensor factorization. In International Conference on Independent Component Analysis and
Signal Separation, 169–176. Springer.
[Du et al. 2014] Du, S. S.; Liu, Y.; Chen, B.; and Li, L. 2014.
Maxios: Large scale nonnegative matrix factorization for
collaborative filtering. In Proceedings of the NIPS 2014
Workshop on Distributed Matrix Computations.
[Gabriel et al. 2004] Gabriel, E.; Fagg, G. E.; Bosilca, G.;
Angskun, T.; Dongarra, J. J.; Squyres, J. M.; Sahay, V.;
Kambadur, P.; Barrett, B.; Lumsdaine, A.; Castain, R. H.;
Daniel, D. J.; Graham, R. L.; and Woodall, T. S. 2004. Open
MPI: Goals, concept, and design of a next generation MPI
implementation. In Proceedings, 11th European PVM/MPI
Users’ Group Meeting, 97–104.
[Gao, Olofsson, and Lu 2016] Gao, T.; Olofsson, S.; and Lu,
S. 2016. Minimum-volume-regularized weighted symmetric nonnegative matrix factorization for clustering. In 2016
IEEE Global Conference on Signal and Information Processing (GlobalSIP), 247–251. IEEE.
[Gillis and Glineur 2012] Gillis, N., and Glineur, F. 2012.
Accelerated multiplicative updates and hierarchical als algorithms for nonnegative matrix factorization. Neural computation 1085–1105.
[Gough 2009] Gough, B. 2009. GNU scientific library reference manual. Network Theory Ltd.
[Hajinezhad et al. 2016] Hajinezhad, D.; Chang, T.-H.;
Wang, X.; Shi, Q.; and Hong, M. 2016. Nonnegative matrix
factorization using admm: Algorithm and convergence
analysis. In 2016 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 4742–4746.
IEEE.
[Hsieh and Dhillon 2011] Hsieh, C.-J., and Dhillon, I. S.
2011. Fast coordinate descent methods with variable selection for non-negative matrix factorization. In Proceedings of
the 17th ACM SIGKDD international conference on Knowledge discovery and data mining, 1064–1072. ACM.
[Kannan, Ballard, and Park 2016] Kannan, R.; Ballard, G.;
and Park, H. 2016. A high-performance parallel algorithm
for nonnegative matrix factorization. In Proceedings of the
21st ACM SIGPLAN Symposium on Principles and Practice
of Parallel Programming, 9. ACM.
[Kim and Park 2011] Kim, J., and Park, H. 2011. Fast nonnegative matrix factorization: An active-set-like method and
comparisons. SIAM Journal on Scientific Computing 3261–
3281.
[Koren, Bell, and Volinsky 2009] Koren, Y.; Bell, R.; and
Volinsky, C. 2009. Matrix factorization techniques for recommender systems. Computer.
[Lee and Seung 1999] Lee, D. D., and Seung, H. S. 1999.
Learning the parts of objects by non-negative matrix factorization. Nature 788–791.
[Lee and Seung 2001] Lee, D. D., and Seung, H. S. 2001.
Algorithms for non-negative matrix factorization. In Advances in neural information processing systems, 556–562.
[Li and Zhang 2009] Li, L., and Zhang, Y.-J. 2009. Fastnmf: highly efficient monotonic fixed-point nonnegative matrix factorization algorithm with good applicability. Journal
of Electronic Imaging 033004–033004.
[Lin 2007] Lin, C.-J. 2007. On the convergence of multiplicative update algorithms for nonnegative matrix factorization. IEEE Transactions on Neural Networks 1589–1596.
[Liu et al. 2010] Liu, C.; Yang, H.-c.; Fan, J.; He, L.-W.; and
Wang, Y.-M. 2010. Distributed nonnegative matrix factorization for web-scale dyadic data analysis on mapreduce. In
Proceedings of the 19th international conference on World
wide web, 681–690. ACM.
[Lu, Hong, and Wang 2017] Lu, S.; Hong, M.; and Wang, Z.
2017. A nonconvex splitting method for symmetric nonnegative matrix factorization: Convergence analysis and optimality. IEEE Transactions on Signal Processing.
[Sun and Fevotte 2014] Sun, D. L., and Fevotte, C. 2014. Alternating direction method of multipliers for non-negative
matrix factorization with the beta-divergence. In Acoustics,
Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on, 6201–6205. IEEE.
[Tan, Cao, and Fong 2016] Tan, W.; Cao, L.; and Fong, L.
2016. Faster and cheaper: Parallelizing large-scale matrix
factorization on gpus. In Proceedings of the 25th ACM International Symposium on High-Performance Parallel and
Distributed Computing, 219–230. ACM.
[Vavasis 2009] Vavasis, S. A. 2009. On the complexity of
nonnegative matrix factorization. SIAM Journal on Optimization 1364–1377.
[Xu and Gong 2004] Xu, W., and Gong, Y. 2004. Document
clustering by concept factorization. In Proceedings of the
27th annual international ACM SIGIR conference on Research and development in information retrieval, 202–209.
ACM.
[Yin, Gao, and Zhang 2014] Yin, J.; Gao, L.; and Zhang,
Z. M. 2014. Scalable nonnegative matrix factorization
with block-wise updates. In Joint European Conference on
Machine Learning and Knowledge Discovery in Databases,
337–352. Springer.
[Zdunek and Fonal 2017] Zdunek, R., and Fonal, K. 2017.
Distributed nonnegative matrix factorization with hals algorithm on mapreduce. In International Conference on Algorithms and Architectures for Parallel Processing, 211–222.
Springer.
[Zhang 2010] Zhang, Y. 2010. An alternating direction algorithm for nonnegative matrix factorization. preprint.
| 2cs.AI
|
arXiv:1304.7992v3 [q-bio.MN] 23 Nov 2013
Fast Reconstruction of Compact Context-Specific Metabolic
Network Models
Nikos Vlassis∗
∗
Maria Pires Pacheco†
Thomas Sauter†
Luxembourg Centre for Systems Biomedicine, University of Luxembourg
†
Life Sciences Research Unit, University of Luxembourg
November 26, 2013
Abstract
Systemic approaches to the study of a biological cell or tissue rely increasingly on the
use of context-specific metabolic network models. The reconstruction of such a model
from high-throughput data can routinely involve large numbers of tests under different
conditions and extensive parameter tuning, which calls for fast algorithms. We present
fastcore, a generic algorithm for reconstructing context-specific metabolic network models from global genome-wide metabolic network models such as Recon X. fastcore takes
as input a core set of reactions that are known to be active in the context of interest (e.g.,
cell or tissue), and it searches for a flux consistent subnetwork of the global network that
contains all reactions from the core set and a minimal set of additional reactions. Our key
observation is that a minimal consistent reconstruction can be defined via a set of sparse
modes of the global network, and fastcore iteratively computes such a set via a series
of linear programs. Experiments on liver data demonstrate speedups of several orders of
magnitude, and significantly more compact reconstructions, over a rival method. Given
its simplicity and its excellent performance, fastcore can form the backbone of many
future metabolic network reconstruction algorithms.
1
Introduction
Cell metabolism is known to play a key role in the pathogenesis of various diseases [11] such
as Parkinson’s disease [31] and cancer [19]. The study of human metabolism has been greatly
advanced by the development of computational models of metabolism, such as Recon 1 [13],
the Edinburgh human metabolic network [18], and Recon 2 [39]. These are genome-scale
metabolic network models that have been reconstructed by combining various sources of
‘omics’ and literature data, and they involve a large set of biochemical reactions that can be
active in different contexts, e.g., different cell types or tissues [38].
To maximize the predictive power of a metabolic model when conditioning on a specific
context, for instance the energy metabolism of a neuron or the metabolism of liver, recent
efforts go into the development of context-specific metabolic models [3, 9, 22, 8, 25, 2]. These
are network models that are derived from global models like Recon 1, but they only contain
1
a subset of reactions, namely, those reactions that are active in the given context. Such
context-specific metabolic models are known to exhibit superior explanatory and predictive
power than their global counterparts [22, 15, 5].
Most algorithms for context-specific metabolic network reconstruction (see Section 2.5
for a short overview) first identify a relevant subset of reactions according to some ‘omics’
information (typically expression data and bibliomics), and then search for a subnetwork
of the global network that satisfies some mathematical requirements and contains all (or
most of) these reactions [3, 35, 22, 7, 20, 2]. The mathematical requirements are typically
imposed via flux balance analysis, which characterizes the steady-state distribution of fluxes
in a metabolic network via linear constraints that are derived from the stoichiometry of the
network and physical conservation laws [34, 36, 32, 16, 14]. The search problem may target
the optimization of a specific functionality of the model (e.g., biomass production) or some
other objective [4], and it may involve repeated tests under different conditions and parameter
tuning [3, 15, 29, 41]. The latter calls for fast algorithms.
We present fastcore, a generic algorithm for context-specific metabolic network reconstruction. fastcore takes as input a core set of reactions that are supported by strong
evidence to be active in the context of interest. Then it searches for a flux consistent subnetwork of the global network that contains all reactions from the core set and a minimal set
of additional reactions. Flux consistency implies that each reaction of the network is active
(i.e., has nonzero flux) in at least one feasible flux distribution [34, 1]. An attractive feature
of fastcore is its generality: As it only relies on a preselected set of reactions and a simple
mathematical objective (flux consistency), it can be applied in different contexts and it allows
the integration of different pieces of evidence (‘multi-omics’) into a single model.
Computing a minimal consistent reconstruction from a subset of reactions of a global
network is, however, an NP-hard problem [1], and hence some approximation is in order. Our
key observation is that a minimal consistent reconstruction can be defined via a set of sparse
modes of the global network, and fastcore is designed to compute a minimal such set.
Every iteration of the algorithm computes a new sparse mode via two linear programs that
aim at maximizing the support of the mode inside the core set while minimizing that quantity
outside the core set. fastcore’s search strategy is in marked contrast to related approaches,
in which the search for a minimal consistent reconstruction involves, for instance, incremental
network pruning [22]. fastcore is simple, devoid of free parameters, and its performance is
excellent in practice: As we demonstrate on experiments with liver data, fastcore is several
orders of magnitude faster, and produces much more compact reconstructions, than the main
competing algorithm MBA [22].
2
2.1
Methods
Background
A metabolic network of m metabolites and n reactions is represented by an m × n stoichiometric matrix S, where each entry Sij contains the stoichiometric coefficient of metabolite i
in reaction j. A flux vector v ∈ Rn is a tuple of reaction rates, v = (v1 , . . . , vn ), where vi
2
B
v2
v1
2
v3
A
v4
C
D
v6
v5
Figure 1: A metabolic network with one blocked reaction (A↔B). Note that A appears with
stoichiometric coefficient 2 in the boundary reaction →2A.
is the rate of reaction i in the network. Reactions are grouped into reversible ones (R) and
irreversible ones (I). For a reaction i ∈ I it holds that vi ≥ 0; this and other imposed flux
bounds, e.g., lower and upper bounds per reaction, are collectively denoted by B (which defines a convex set). A flux vector is called feasible or a mode if it satisfies a set of steady-state
mass-balance constraints that can be compactly expressed as:
v ∈ B.
Sv = 0,
(1)
An elementary mode is a feasible flux vector v 6= 0 with minimal support, that is, there is no
other feasible flux vector w 6= 0 with supp(w) ⊂ supp(v), where supp(v) = j ∈ {1, 2, . . . , n} :
vj 6= 0 is the support (i.e., the set of nonzero entries) of v [34, 16]. A reaction i is called
blocked if it cannot be active under any mode, that is, there exists no mode v ∈ Rn such that
vi 6= 0 (in practice |vi | ≥ ε, for some small positive threshold ε). A metabolic network model
that contains no blocked reactions is called (flux) consistent [34, 1].
2.2
Network consistency testing
Given a metabolic network model with stoichiometric matrix S, a problem of interest is to
test whether the network is consistent or not. Additionally, if the network is inconsistent, it
would be desirable to have a method that detects all blocked reactions.
It has been suggested that network consistency can be detected by a single linear program
(LP) [1]. The idea is to first convert each reversible reaction into two irreversible reactions
(and define a reversible flux as the difference of two irreversible fluxes), and then test if the
minimum feasible flux on the new set J of irreversible-only reactions is strictly positive (in
practice, at least ε). This is equivalent to testing if the following LP is feasible:
max
v,z
s.t.
z
z≥ε
z∈R
vi ≥ z
∀i ∈ J
Sv = 0
v ∈ B.
(LP-2)
This test of consistency, however, can produce spurious solutions. In Figure 1 we show a
toy metabolic network comprising four metabolites (A,B,C,D) and six reactions annotated
with corresponding fluxes v1 , . . . , v6 . Fluxes are bounded as 0 ≤ vi ≤ 3 for i =
6 2, and
3
|v2 | ≤ 3. All stoichiometric coefficients are equal to one, except for the reaction →2A.
The only reversible reaction is A↔B, which is a dead-end reaction and therefore blocked,
whereas all other reactions are irreversible and unblocked. After converting A↔B to a pair
of irreversible reactions, LP-2 achieves optimal value z ∗ = 1.5, which implies (wrongly) that
the network is consistent. The test here fails because the two irreversible copies of A↔B have
equal flux at the solution, thereby nullifying the actual net flux of A↔B.
A straightforward solution to the problem would involve iterating through all reactions,
computing the maximum and minimum feasible flux of each reaction via an LP that satisfies
the constraints in (1). Reactions with minimum and maximum flux zero would then be
blocked. This is the idea behind the FVA (Flux Variability Analysis) algorithm and the
reduceModel function of the COBRA toolbox [26, 33]. However, iterating through all reactions
can be inefficient. A faster variant is fastFVA [17], which achieves acceleration over FVA
via LP warm-starts. Another fast algorithm is CMC (CheckModelConsistency) [22], which
involves a series of LPs, where each LP maximizes the sum of fluxes over a subset J of
reactions:
X
max
vj
v
j∈J
(LP-3)
s.t. Sv = 0
v ∈ B.
The set J is initialized by J = R ∪ I (all reactions in the network), and it is updated after
each run of LP-3 so that it contains the reactions whose consistency has not been established
yet. When J cannot be reduced any further, we can reverse the signs of the columns of S
corresponding to the reversible reactions in J and resume the iterations. Eventually, all
remaining reactions may have to be tested one by one for consistency, as in FVA. Such an
iterative scheme is complete, in the sense that it will always report consistency if the network
is consistent, and if not, it will reveal the set of blocked reactions. However, as we will
clarify in the next section, LP-3 is not optimizing the ‘correct’ function, which may result in
unnecessarily many iterations. For example, when applied to the network of Figure 1, LP-3
will pick up the elementary mode that corresponds to the pathway A→C→D (because this
pathway achieves maximum sum of fluxes v1 + v4 + v5 + v6 = 1.5 + 3 + 3 + 3), and it will set
v3 = 0. To establish the consistency of the reaction A→D, an additional run of LP-3 would be
needed, where the set J would only involve the reactions A↔B and A→D. Hence, an iterative
algorithm like CMC that relies on LP-3 would need two iterations to detect the consistent
part of this network. However, one LP suffices to detect the consistent subnetwork in this
example, as we explain in the next section. In more general problems involving larger and
more realistic networks, CMC may involve unnecessarily many iterations, as we demonstrate
in the experiments.
2.3
Fast consistency testing
In most problems of interest there will be no single mode that renders the whole network
consistent, and an iterative algorithm like the one described in the previous section must be
used. For performance reasons it would therefore be desirable to be able to establish the
consistency of as many reactions as possible in each iteration of the algorithm.
4
Since consistency implies nonzero fluxes, it is sufficient to optimize a function that just
‘pushes’ all fluxes away from zero. Formally, this amounts to searching for modes v whose
cardinality—denoted by card(v) and defined as card(v) = #supp(v), i.e., the number of
nonzero entries of v—is as large as possible. Directly maximizing card(v) is, however, not
straightforward, for the following reasons: First, the card function is quasiconcave only for
v ∈ Rn+ (the nonnegative orthant), and it is nonconvex for general v ∈ Rn [6]. Second, even if
we restrict attention to nonnegative fluxes in each iteration (which we can do without loss of
generality by flipping the signs of the corresponding columns of S), it is not obvious how to
efficiently maximize the quasiconcave card(v). Third, in practice consistency implies fluxes
that are ε-distant from zero, in which case some adaptation of the card function is in order.
Here we propose an approach to approximately maximize card(v) over a nonnegative flux
subspace indexed by a set of reactions J . First note that the cardinality function can be
expressed as
X
card(v) =
θ(vi ) ,
(4)
i∈J
where θ : R → {0, 1} is a step function:
(
θ(vi ) =
0 if vi = 0
1 if vi > 0 .
(5)
The key idea is to approximate the function θ by a concave function that is the minimum of
a linear function and a constant function:
θ(vi ) ≈ min{
vi
, 1} ,
ε
(6)
where ε is the flux threshold. The problem of approximately maximizing card(v) can then be
cast as an LP: We introduce an auxiliary variable zi ∈ R+ for each flux variable vi , for i ∈ J ,
P
and take epigraphs [6], in which case maximizing card(v) = i∈J θ(vi ) can be expressed as
X
max
zi
v,z
s.t.
i∈J
zi ≤ θ(vi )
∀i ∈ J , zi ∈ R+
vi ≥ 0
∀i ∈ J
Sv = 0
v ∈ B.
Using (6) and assuming constant ε, this simplifies to
X
max
zi
v,z
s.t.
i∈J
zi ∈ [0, ε]
∀i ∈ J , zi ∈ R+
vi ≥ z i
∀i ∈ J
Sv = 0
v ∈ B.
(LP-7)
Note that LP-7 tries to maximize the number of feasible fluxes in J whose value is at least ε
(contrast this with LP-2).
5
Returning to the network of Figure 1, if J comprises all network reactions, then note that
the flux vector [v1 , v2 , v3 , v4 , v5 , v6 ] = [ε, 0, ε, ε, ε, 2ε] is an optimal solution of LP-7. Hence, a
single run of the latter can detect all unblocked reactions of that network. More generally,
a single run of LP-7 on an arbitrary subset J of a given network will typically detect all
unblocked irreversible reactions of J . The intuition is that LP-7 prefers flux ‘splitting’ over
flux ‘concentrating’ in order to maximize the number of participating reactions in the solution,
which, in the case of irreversible reactions, corresponds to flux cardinality maximization.
By construction, the above approximation of the cardinality function applies only to
nonnegative fluxes. In order to deal with reversible reactions that can also take negative fluxes,
we can embed LP-7 in an iterative algorithm (as in the previous section), in which reversible
reactions are first considered for positive flux via LP-7, and then they are considered for
negative flux. The latter is possible by flipping the signs of the columns of the stoichiometric
matrix that correspond to the reversible reactions under testing, in which case the fluxes of the
transformed model are again all nonnegative, and the above approximation of the cardinality
function can be used. This gives rise to an algorithm for detecting the consistent part of a
network that we call fastcc (for fast consistency check). Since fastcc is just a variant of
fastcore, we defer its detailed description until the next section.
Independently to this work, a similar approach to network consistency testing was recently
proposed, called OnePrune [12]. OnePrune first converts each reversible reaction into two
irreversible reactions, forming an augmented set J of irreversible-only reactions (as in LP-2
above), and then it employs an LP that coincides with LP-7 for the above choice of J and
ε = 1. However, such an approach is prone to the same drawback as LP-2, namely, that the
two irreversible copies of a blocked reaction can carry equal positive flux at the solution of LP7 due to the presence of cycles introduced by the transformation. The authors acknowledge
this problem but they do not fully resolve it. In our case, we avoid this problem by working
with the original reactions and a series of LPs with appropriate sign flips of the stoichiometric
matrix, thereby guaranteeing the completeness of the algorithm.
2.4
Context-specific network reconstruction
The reconstruction problem involves computing a minimal consistent network from a global
network and a ‘core’ set of reactions that are known to be active in a given context. Formally,
given (i) a consistent global network {N , SN } with reaction set N = {1, 2, . . . , n} and stoichiometric matrix SN , and (ii) a set C ⊂ N , the problem is to find the smallest set A ⊆ N
such that C ⊆ A and the subnetwork {A, SA } induced by the reaction set A is consistent.
(By SA we denote the submatrix of SN that contains only the columns indexed by A.) This
problem is known to be NP-complete [1], suggesting that a practical solution should entail
some approximation. (We note that Acuña et al. [1] prove NP-completeness of this problem
by noting that a special case involves C being the empty set, in which case the problem comes
down to finding the smallest elementary mode of the global network, which, as the authors
show, is NP-complete. However, this leaves open the case of a nonempty core set C, since a
solution to the minimal reconstruction problem need not constitute an elementary mode. We
conjecture that the problem remains NP-hard when C is nonempty, but we are not pursuing
6
this question here.)
Our approach hinges on the observation that a consistent induced subnetwork of the global
network can be defined via a set of modes of the latter:
Theorem 1. Let V be a set of modes of the global network {N , SN }, and let A = ∪v∈V supp(v)
be the union of the supports of these modes. The induced subnetwork {A, SA } is consistent.
Proof. For each v ∈ V, let vA be the ‘truncated’ v after dropping all dimensions not indexed
by A. Clearly, SA vA = 0, therefore each vA is a mode in the reduced model {A, SA }. By
construction of A, each reaction in A is in the support of some v ∈ V, and hence also in the
support of some mode vA of the reduced model.
This simple result allows one to cast the reconstruction problem as a search problem over
sets of modes of the global network:
min card(A)
V
[
s.t. A =
supp(v)
(NLP-8)
v∈V
C⊆A
∀v ∈ V : SN v = 0, v ∈ B .
Note that this optimization problem involves searching for a set V of modes of {N , SN }, such
that the union of the support of these modes (the set A) is a minimal-cardinality set that
contains the core set C. In order to practically make use of this theorem, one has to define
a search strategy over modes. Next we discuss two possibilities. The first gives rise to an
exact algorithm, but this algorithm does not scale to large networks. The second is a scalable
greedy approach that gives rise to fastcore.
Exact reconstruction via mixed integer linear programming
Note that, without loss of generality, in NLP-8 we can restrict the search for V over all
elementary modes of the global network, since the union of their supports covers the whole
set N . As we show next, if all elementary modes are available, NLP-8 can be cast as a mixed
integer linear program (MILP) and solved exactly. This MILP is defined as follows. Let r
be the number of elementary modes, and {m1 , . . . , mr } be a set of length-n binary vectors,
where each vector mj captures the support of elementary mode j (so, its ith entry is 1 if
reaction i is included in elementary mode j, and 0 otherwise). Also, let c = (c1 , . . . , cn ) be
a length-n binary vector with ci = 1 if reaction i is included in the core set C, and ci = 0
otherwise. The decision variables of the MILP are a length-n binary vector x = (x1 , . . . , xn )
and a length-r real vector y = (y1 , . . . , yr ). At an optimal solution of the MILP, the set A is
defined as A = {i ∈ N : x∗i = 1}.
Theorem 2. When all elementary modes are available, the following MILP-9 solves NLP-8
7
exactly.
min
x,y
s.t.
X
xi
i
1X
mj yj
r
j
X
c≤
mj yj
x≥
(MILP-9)
j
y ∈ [0, 1]
x ∈ {0, 1} .
Proof. By definition, x∗i = 1 implies that reaction i will be included in the reconstruction
P
A, hence the objective minimizes the cardinality of A. The sum j mj yj∗ is a vector whose
support is the union of the supports of all selected elementary modes at the solution, where
P
an elementary mode j is selected when yj∗ > 0. The first constraint x ≥ 1r j mj yj therefore
imposes that the set A must contain the union of the supports of the selected elementary
P
modes at the solution. (The factor 1r ensures that 1r j mj yj ≤ 1). Since superfluous reactions
P
are removed by the minimization of
i xi in the objective, the above implies that A is
precisely the union of the supports of the selected elementary modes at the solution. The
P
second constraint c ≤ j mj yj imposes that the core set must be included in the union of
the supports of the selected elementary modes at the solution, and hence the core set must
be included in A. Therefore, all constraints of NLP-8 are satisfied at the optimal solution
of MILP-9, and since the two programs minimize the same objective, an optimal solution of
MILP-9 must be an optimal solution of NLP-8.
Note, however, that MILP-9 does not scale to large networks, for the following reasons:
First, it requires computing all elementary modes of the global network, which can be a very
large number [16]. Second, the binary decision variables xi index all reactions of the global
network, and therefore MILP-9 needs to search over a binary hypercube of dimension n, which
can be prohibitive for large n. Nonetheless, it is reassuring to know that an exact solution to
the context-specific network reconstruction problem is possible, albeit with high complexity.
Next we describe fastcore, an approximate greedy algorithm that scales much better to
large networks, and we compare it to MILP-9 in the Results section.
Greedy approximation and the fastcore algorithm
An alternative search strategy for computing V in NLP-8 is a greedy approach, reminiscent
of greedy heuristics for the related set covering problem [10]. This is the idea behind the
proposed fastcore algorithm: We build up the set V in a greedy fashion, by computing
in each iteration a new mode of the global network. Further, as a means to approximately
minimize card(A), each added mode is constrained to have sparse support outside C. This is
implemented via L1 -norm minimization, which is a standard approach to computing sparse
solutions to (convex) optimization problems [6, 23].
The overall fastcore algorithm is shown in Algorithm 1. The algorithm maintains a set
J ⊆ C that is initialized with the irreversible reactions in C, and a ‘penalty’ set P = (N \C)\A
8
Algorithm 1 The fastcore algorithm
Input: A consistent metabolic network model {N , SN } and a reaction set C ⊂ N .
Output: A consistent induced subnetwork {A, SA } of {N , SN } such that C ⊆ A.
1: function fastcore( N , C )
2:
J ← C ∩ I, P ← N \ C
3:
f lipped ← F alse, singleton ← F alse
4:
A ← FindSparseMode( J , P, singleton )
5:
J ←C\A
6:
while J 6= ∅ do
7:
P ←P \A
8:
A ← A ∪ FindSparseMode( J , P, singleton )
9:
if J ∩ A =
6 ∅ then
10:
J ← J \ A, f lipped ← F alse
11:
else
12:
if f lipped then
13:
f lipped ← F alse, singleton ← T rue
14:
else
15:
f lipped ← T rue
16:
if singleton then
17:
J˜ ← J (1) (the first element of J )
18:
else
19:
J˜ ← J
20:
end if
21:
for each i ∈ J˜ \ I do
22:
flip the sign of the i’th column of SN and
23:
swap the upper and lower bounds of vi
24:
end for
25:
end if
26:
end if
27:
end while
28:
return A
29: end function
that contains all reactions outside C that have not been added yet to the set A. Each iteration
adds to the set A the support of a mode that is dense in J (i.e., contains as many nonzero
fluxes in J as possible) and sparse in P (i.e., contains as many zero fluxes in P as possible),
computed by the function FindSparseMode (Algorithm 2). This function first applies
an LP-7 to compute an active subset K of J , and then it applies the following L1 -norm
minimization LP constrained by the set K:
X
min
zi
v,z
s.t.
i∈P
vi ∈ [−zi , zi ]
∀i ∈ P, zi ∈ R+
vi ≥ ε
∀i ∈ K
SN v = 0
v ∈ B.
9
(LP-10)
Algorithm 2 The FindSparseMode function
Input: A set J ⊆ C, a penalty set P ⊆ N \ C, and the singleton flag.
Output: The support of a mode that is dense in J and sparse in P.
1: function FindSparseMode( J , P, singleton )
2:
if J = ∅ then
3:
return ∅
4:
end if
5:
if singleton then
6:
v ∗ ← LP-7 on set J (1)
7:
else
8:
v ∗ ← LP-7 on set J
9:
end if
10:
K ← {i ∈ J : vi∗ ≥ ε}
11:
if K = ∅ then
12:
return ∅
13:
end if
14:
v ∗ ← LP-10 on sets K, P
15:
return {i ∈ N : |vi∗ | ≥ ε}
16: end function
P
The LP-10 minimizes i∈P |vi |, the L1 norm of fluxes in the penalty set P (expressed via
epigraphs), subject to a minimum flux constraint on the set K. However, some care is needed
to preempt false negative solutions arising from the minimization of L1 norm in LP-10. For
example, suppose in the network of Figure 1 that the global network comprises all reactions
except A↔B, and C = J = K = {6} and P = {1, 3, 4, 5}. In this case, LP-10 could settle
to a solution [v1 , v3 , v4 , v5 , v6 ] = [ 2ε , ε, 0, 0, ε]. The flux v1 , being below ε, would be treated as
zero by FindSparseMode, in which case the reaction →2A would be erroneously excluded
from the reconstruction. A simple way to avoid this is to use a scaled version of ε (we used
105 ε) in the second constraint of LP-10, with an equal scaling of all flux bounds in B.
The fastcore algorithm first goes through the I ∩ C reactions (step 2), and then through
the R ∩ C ones (and eventually through each individual reversible reaction in the core set;
when singleton = T rue). The f lipped variable ensures that a reversible reaction is tested in
both the forward and negative direction. The algorithm terminates when all reactions in C
have been added to A, which is guaranteed since in the main loop the set J never expands
(step 10) and the global network is consistent. Note that fastcore has no free parameters
besides the flux threshold ε.
The fastcc algorithm for detecting the consistent part of an input network (see previous section) can be viewed as a variant of fastcore(N , N ) in which the steps 10–14 of
FindSparseMode are omitted (and there is no P set). It is easy to verify that fastcc is
complete, in the sense that it will always report consistency if the network is consistent, and
if not, it will reveal the set of blocked reactions.
10
Table 1: Summary of the main characteristics of GIMME [3], MBA [22], iMAT [43], mCADRE
[41], INIT [2], and fastcore (this paper) reconstruction algorithms.
Optimization
Computational cost
Function required
Omics required
Code available
2.5
GIMME
LP
low
yes
yes
yes
MBA
MILP
high
no
optional
yes
iMAT
MILP
high
no
yes
yes
mCADRE
MILP
high
yes
yes
yes
INIT
MILP
high
yes
yes
no
fastcore
LP
low
no
no
yes
Related work
Several algorithms have been published in the last years for extracting condition-specific
models from generic genome-wide models like Recon 1. Among them, mCADRE [41], INIT
[2], iMAT [43], MBA [22] and GIMME [3] are the most commonly used (see Table 1 for an
overview). Here we provide a short outline of the different algorithms, and refer to [4] for
a more extensive overview. For GIMME, iMAT, and MBA, we briefly discuss some notable
differences to fastcore.
GIMME [3] takes as input microarray data and a biological function to optimize for, such
as biomass production. GIMME starts by removing reactions with associated expression levels
below a user-defined threshold, and then it optimizes for the specified biological function using
linear programming. In case the pruning steps compromise the input biological function,
GIMME reintroduces some previously removed reactions that are in minimal disagreement
with the expression data. Since GIMME has not been designed to include all core reactions
in the solution (as fastcore does), the reconstructions obtained by GIMME and fastcore
can differ significantly: Running the createTissueSpecific function of the COBRA toolbox on
a set of liver core reactions (see Section 3) treating them as expressed reactions (and adding
a biomass reaction [41] and a sink reaction for glycogen to be used as optimization function),
only about 50% of the core reactions of the GIMME model were consistent at the solution. A
fairer comparison would require adapting fastcore to explicitly deal with omics data, which
is outside the scope of the current work.
iMAT [43] was originally designed for the integration of transcriptomic data. iMAT optimizes for the consistency between the experimental data and the activity state of the model
reactions. iMAT tries to include modes composed of reactions associated to genes with high
expression value, and therefore a threshold needs to be chosen to segregate between low,
medium, and highly expressed genes. The computational demands of iMAT are high due to
the repeated use of mixed integer linear programming. As with GIMME, direct comparison of
iMAT to fastcore is problematic. Nevertheless, we applied iMAT (own implementation) on
the liver problem (see Section 3), by setting the liver core reactions to RH (reaction high) and
all non-core reactions to RL (reaction low). iMAT determined 549 core reactions as active,
while 182 and 338 reactions were classified as undetermined and inactive, respectively. This
11
means that about 50% of the core reactions were lost during iMAT model building. As with
GIMME, this demonstrates the difficulty of directly comparing fastcore to algorithms that
optimize different objectives.
mCADRE [41] is similar to MBA, except that the pruning order is not random, but it
depends on the tissue-specific expression evidence and weighted connectivity to other reactions
of the network. Reactions that are associated to genes that are never tagged as expressed and
which are not connected to reactions associated to highly expressed genes are first evaluated
in the pruning step. Reactions are effectively removed if the removal does not impair core
reactions and metabolic functions to carry a flux (mCADRE removes core reactions if the
core/non-core reaction ratio is below a user-given threshold). mCADRE uses mixed integer
linear programming and therefore it does not scale up to large networks (but it is in general
faster than MBA).
INIT [2] uses data retrieved from public databases in order to assess the presence of a
certain reaction-respective metabolites in the cell type of interest. INIT uses mixed integer
linear programming to build a model in which all reactions can carry a flux. Contrary to
other algorithms, INIT does not rely on the assumption of a steady state, but it allows small
net accumulation of all metabolites of the model.
The closest algorithm to fastcore is the MBA algorithm of Jerby et al. [22]. MBA takes
as input two core sets of reactions, and it searches for a consistent network that contains all
reactions from the first set, a maximum number of reactions from the second set (for a given
tradeoff), and a minimal number of reactions from the global network. (fastcore can be
easily adapted to work with multiple core sets, by introducing a set of weights that reflect
the confidence of each reaction to be active in the given context, and adding appropriate
regularization terms in the objective functions of LP-7 and LP-10 that capture the given
tradeoff. We will address this variant in future work.) Both fastcore and MBA involve
a search for a minimal consistent subnetwork, however the search strategy of fastcore is
very different to MBA: Whereas fastcore iteratively expands the active set A starting with
A = ∅, MBA starts with A = N and iteratively prunes the set A by checking whether
the removal of each individual reaction (selected in random order) compromises network
consistency. As the pruning order affects the output model, this step of MBA is repeated
multiple times. MBA builds a final model by adding one by one non-core reactions with the
highest presence rate over all pruning runs, and it stops when a consistent final model is
obtained. Due to the multiple pruning runs, MBA has very high computational demands.
Consistency testing in MBA is carried out with the CMC algorithm that is based on LP-3,
as explained earlier. Hence, fastcore’s search strategy differs to MBA in two key aspects:
First, consistency testing in fastcore involves the maximization of flux cardinality (LP-7)
instead of sum of fluxes (LP-3), which results in fewer LP iterations. Second, the search
for compact solutions in fastcore involves L1 -norm minimization instead of pruning. The
advantage of the former is that it can be encoded by a single LP, resulting in significant
overall speedups (see Section 3.2).
12
Figure 2: Flowchart of the overall pipeline for generating consistent context-specific models.
3
Results
Generic metabolic reconstructions like Recon 2 are inconsistent models as they contain reactions that are not able to carry nonzero flux due to gaps in the network (see next section).
The first step towards obtaining a consistent context-specific reconstruction is therefore to
extract the consistent part of a global generic model. This can be achieved by fastcc or
other similar methods (see Section 2.2). The consistent global model serves then as input for
the context-specific reconstruction with fastcore. In Figure 2 we show a flowchart of the
overall pipeline.
We report results on two sets of problems, the first involving consistency verification of
an input model, and the second involving the reconstruction of a context-specific model from
an input model and a core set of reactions. The fastcore algorithm was implemented in
the COBRA toolbox [33], using Matlab 2013a and the IBM CPLEX solver (version 12.5.0.0).
Test runs were performed on a standard 1.8 GHz Intel Core i7 laptop with 4 GB RAM running
Mac OS X 10.7.5. In all experiments we used flux threshold ε =1e-4. The software is available
from bio.uni.lu/systems_biology/software/
3.1
Consistency testing
In the first set of experiments we applied fastcc, the consistency testing variant of fastcore, for consistency verification of four input models, and compared it against the FastFVA
algorithm of Gudmundsson and Thiele [17], and an own implementation (based on fastcc
but with LP-3 replacing LP-7) of the CMC algorithm of Jerby et al. [22]. We also tested
the FVA algorithm of the reduceModel function of the COBRA toolbox [33], and the MIRAGE algorithm of Vitkin and Shlomi [40], but we do not include them in the results as they
performed worse than the reported ones. The input models were the following:
• c-Yeast (#N = 1204), the consistent part of a yeast model [42].
13
Table 2: Comparing fastcc to fastFVA [17] and CMC [22] on four input models.
fastFVA
CMC
fastcc
c-Yeast
# LPs time∗
2408
3
18
0.5
7
0.1
c-Ecoli
# LPs time
3436
3
25
1
2
0.2
c-Recon1
# LPs time
4938
9
49
2
9
0.4
c-Recon2
# LPs time
11668
207
42
11
19
5
∗ in seconds
• c-Ecoli (#N = 1718), the consistent part of an E. coli model [29]. (Here we set to 1000
the upper bounds of all fluxes that were fixed to zero, and we multiplied all bounds by
1000 to avoid numerical issues.)
• c-Recon1 (#N = 2469), the consistent part of Recon 1 [13]. (Recon 1 was found to
contain 1273 blocked reactions.)
• c-Recon2 (#N = 5834), the consistent part of Recon 2 [39]. (Recon 2 was found to
contain 1606 blocked reactions.)
The results are shown in Table 2. fastcc is faster and it uses much fewer LPs than the
other two algorithms. We note that fastFVA is based on an optimized Matlab/C++ implementation with LP warm-starts, while fastcc is based on standard Matlab. These results
confirm the appropriateness of flux cardinality (LP-7) as a metric for network consistency
testing, in agreement with the theoretical analysis and the discussions above.
3.2
Reconstruction of a liver model
In the second set of experiments, we used the fastcore algorithm to reconstruct a liver
specific metabolic network model from the consistent part of Recon 1 (c-Recon1, #N = 2469),
and we compared against an own implementation of the MBA algorithm of Jerby et al. [22].
We applied the two algorithms in two settings. The first setting involves the liver specific
input reaction set of Jerby et al. [22], which is based on 779 ‘high’ core and 290 ‘medium’
core reactions (the latter set is supported by weaker biological evidence than the former). To
allow a comparison with fastcore, we defined a single core set as the union of the high and
medium core reaction sets, and we applied the two algorithms on this core set. The second
setting uses the ‘strict’ liver model of Jerby et al. [22], which contains 1083 high core reactions
and no medium core reactions, and therefore allows a direct comparison with fastcore.
The results for the two settings are shown in Table 3. We note that for MBA, the reported
number of LPs and the runtime refer to a single pruning iteration of the algorithm, whereas
the size of each reconstruction refers to the final model after 1000 pruning iterations. In
both settings, fastcore is several orders of magnitude faster than MBA, achieving a full
reconstruction of a liver specific model in about one second, using a much smaller number
of LPs. As MBA employs a greedy pruning strategy for optimization, the number of LPs
14
Table 3: Comparing fastcore to MBA [22] on liver model reconstruction from c-Recon1.
MBA
fastcore
liver core set (#C = 1069)
#A
IR∗ #LPs time‡
1826 1573 72279 7383
1746 1546
20
1
strict liver core set (#C = 1083)
#A
IR #LPs time
1888 1630 71546 6730
1818 1627
20
1
∗ number of intracellular reactions
‡ the reported time (in seconds), as well as the number of LPs, refer to a single pruning step of MBA, whereas
#A and IR refer to the full MBA.
that it uses and its total runtime can be very high, as also indicated by Wang et al. [41] who
reported runtime of a single pruning pass of MBA in the order of 10 hours on a 2.34 GHz
CPU computer.
The reconstructed models by fastcore are also more compact than those obtained by
MBA, with a difference of 70-80 non-core reactions. For the standard liver model, 1687
out of the 1746 reactions (96%) of the fastcore reconstruction appear also in the MBA
reconstruction, whereas for the strict liver model the common reactions are 1739 out of 1818
(95%). The two algorithms turned out to use alternative transporters to connect the core
reactions: In the standard liver model, 46 out of 59 reactions that are present exclusively
in the fastcore reconstruction are transporter reactions or other reactions which are not
associated to a specific gene and thus are not sufficiently supported in the core set, whereas
in MBA the corresponding numbers are 116 out of 139 reactions. Note that both MBA
and fastcore try to minimize the number of added non-core reactions in order to obtain a
compact consistent model. The above difference in the number of added non-core reactions
between MBA and fastcore is the result of the different optimization approaches taken by
the two algorithms, and no biological relevance should be attributed to each reconstruction
other than the one implied by the makeup of the core set. From this point of view, fastcore
performs in general better than MBA, as it tends to add fewer unnecessary reactions.
We also compared the solutions of fastcore to those of MILP-9, using core sets that are
randomly generated from a consistent subset of E. coli core [30]. This is a small model with
#N = 53 and 414 elementary modes (unfortunately, the dependence of the MILP-9 solver to
the number of elementary modes did not allow testing larger models). In Figure 3 we show
the size of the reconstructed models (mean values) obtained with the exact MILP solver vs.
fastcore, as a function of the size of the core set. fastcore is capable of obtaining very
good approximations to the optimal solutions, which improve with the size of the core set.
To evaluate fastcore’s performance in correctly identifying liver reactions, we performed
repeated random sub-sampling validation in which fastcore was used to reconstruct the liver
metabolism based on a reduced, randomly selected ‘subcore’ set of 80% of the original core
reactions. As in [22], we wanted to test whether fastcore is able to recover a significant
number of the 20% left-out core reactions. To test for the enrichment of the left-out core
reactions in the reconstructed model, we used a hypergeometric test, in which the total
15
Figure 3: Comparing fastcore to an exact MILP solver on a small E. coli model [30].
Shown are mean values of sizes of reconstructed models (over 50 repetitions for each core set;
standard deviations were small and are omitted to avoid clutter) as a function of the size of
the core set. fastcore computes near-optimal reconstructions, which improve with the size
of the core set.
population is defined by all non-subcore reactions in the global network, the number of draws
is defined as the number of non-subcore reactions included in the reconstruction, and the leftout core reactions are the ‘successes’. Under the null-hypothesis that there is no enrichment
for the left-out core reactions when reconstructing the liver model based on the subcore set, we
can compute a p-value for including at least the number of observed left-out core reactions in
the reconstruction. We repeated this random sub-sampling procedure 500 times and computed
the corresponding p-values. The median of these p-values was 0.0025, indicating the ability
of fastcore to capture liver-specific reactions that were included in the original core set.
As argued above, the reconstructions obtained by fastcore need not optimize for cellular
functions other than the ones implied by the composition of the input core set, and it is
an interesting research question how to modify fastcore so that it can explicitly capture
functional requirements in its reconstructions. Nevertheless, it is of interest to test whether
the current version of fastcore can produce reconstructions that are functionally relevant,
perhaps for slight variations of the core set. To this end, as in [22], we checked whether
the (standard) liver model reconstructed by fastcore can perform gluconeogenesis from
glucogenic amino acids, glycerol, and lactate (altogether 21 metabolites). If not yet included,
transporters from the extracellular medium to the cytosol were added to the model (glycerol,
glutamate, glycine, glutamine, and serine). This was necessary as the transport reactions were
not sufficiently supported in the core set. This ‘extended’ liver model was able to convert 17/21
metabolites (vs 12/21 metabolites of the non-extended model). The extended liver model was
then used to simulate the liver disorders hyperammonemia and hyperglutamenia, which affect
16
Figure 4: Mean urea/glutamine ratio in the extended liver model obtained by fastcore.
Healthy (normal homozygote), partial (heterozygote) and full knock-out cases. See text for
details.
the capacity to metabolize dietary amino acids into urea [22]. Loss of function mutations
of three enzyme-coding genes, argininosuccinate synthetase (ASS), argininosuccinate lyase
(ASL), and ornithine transcarbamylase (OTC) were identified in patients suffering from these
disorders. The rates of the reactions controlled by the three genes were fixed to 500, 250, or
zero, to mimic the healthy homozygote (no mutation), heterozygote (loss of one allele), and the
complete loss of function, respectively. To allow for a comparison with the experimental study
of Lee et al. [24] where labeled 15N-glutamine was administrated to patients suffering from
inborn errors affecting the three genes, we explicitly shut down the influx of other potential
nitrogen sources in the liver model, thereby simulating only the uptake and metabolism
of glutamine. By allowing the influx of only one nitrogen source, the fate of the latter
could be determined exactly in the model. The ratio of urea secretion level over glutamine
absorption was computed by sampling over the feasible space [32]. In accordance with the
wet lab observations [24], the severity of the disorders, characterized by the mean urea over
glutamine ratio, increased with the level of loss of function of the three genes ASS, ASL,
and OTC (see Figure 4). Null patients showed no native production of urea. Overall, the
ratios predicted by the fastcore model faithfully match the experimentally observed ones
[24]. (The corresponding ratios reported by Jerby et al. when using the MBA algorithm [22]
matched less well the experimental observations, probably because of the cross-feeding of
nitrogen to urea from multiple nitrogen sources. By running the above procedure on the
MBA model, we noticed that both models attained comparable urea / glutamine flux ratios.)
To summarize, the above experiments demonstrate that, by an informed choice of the core
set and influx bounds, fastcore can indeed give rise to functionally relevant models.
17
3.3
Reconstruction of a murine macrophage model
We also used the fastcore algorithm to build a cell-type specific murine macrophage model
from the consistent part of Recon1bio (comprising #N = 2474 reactions). Recon1bio (#N =
3745) is a modified Recon 1 model that contains three extra reactions (biomass, NADPOX,
and a sink reaction to balance the glycogenin self-glucosylation reaction) [5]. We used a
core set comprising 300 (out of 382) proteomics derived Raw264.7 macrophage reactions, as
described by Bordbar et al. [5]. (The remaining 82 reactions could not be added to the core set
as they are situated in an inconsistent region of Recon 1 and therefore carry a permanent zero
net flux.) For their macrophage reconstruction, Bordbar et al. used, among other methods,
GIMMEp—a variant of the GIMME algorithm [3] that is similar to the MBA algorithm—and
they obtained a network model containing 1026 intracellular reactions. Our main interest was
to investigate whether fastcore can obtain a functional network that is at least as compact
as the one obtained with GIMMEp. fastcore generated (in about one second and using 11
LPs) a consistent network model of 953 reactions, 831 of which are intracellular reactions.
This is a much more compact model than the one obtained with GIMMEp.
4
Discussion
fastcore is a generic algorithm for context-specific metabolic network reconstruction from
genome-wide metabolic models, and it was motivated by requirements of fast computation
and compactness of the output model.
The key advantage of having a fast reconstruction algorithm is that it permits the execution of multiple runs in order to optimize for extra parameters or test different core sets
extracted from the input data [15, 41]. For example, when working with gene expression
data, the definition of the core set may depend on the threshold used to segregate between
high expression genes (core reactions) and low expression genes (non-core reactions) [3]. As
the choice of threshold is rather arbitrary, a practical approach could involve evaluating the
robustness of the output model as a function of the chosen threshold. fastcore can perform this analysis in a few minutes, whereas for the same problem other algorithms would
need hours or days. (Algorithms like GIMME or GIMMEp that require manual curation and
assembly of subnetworks, would also fail in this kind of task.) Another example where fast
computation is imperative is cross-validation. In the current study (see Section 3) we ran a
random sub-sampling validation procedure 500 times, an operation that took a few minutes
with fastcore but that would barely be manageable with other reconstruction algorithms.
Other examples where fast computation is important are time-course experiments or experiments involving different patients or conditions [21]. There fastcore could more easily
identify differential models over time and/or input conditions.
Compactness is a key concept in various research areas of biology, such as the minimal
genome [28, 27]. Notwithstanding, the requirement of model compactness seems to be in
disagreement with the observation that biological systems are fairly redundant and this redundancy serves a specific purpose, namely, the fast adaptation to changes in the environment.
Alternative pathways that perform similar functions are known to be expressed in different
18
environmental conditions, allowing for instance to metabolize another type of sugar when
glucose is not available [37]. At any rate, the pursuit of compactness in metabolic network
reconstruction need not be in conflict with the notion of redundancy. Alternative pathways
will be included in a reconstructed model as long as ‘redundant’ reactions that are supported
by biological evidence are included in the core set.
Acknowledgments
We would like to thank Ines Thiele, Ronan Fleming, Nils Christian, Evangelos Symeonidis,
Nathan Price, and Rudi Balling for their feedback.
References
[1] V. Acuña, F. Chierichetti, V. Lacroix, A. Marchetti-Spaccamela, M. F. Sagot, and L. Stougie.
Modes and cuts in metabolic networks: Complexity and algorithms. Biosystems, 95(1):51–60,
2009.
[2] R. Agren, S. Bordel, A. Mardinoglu, N. Pornputtapong, I. Nookaew, and J. Nielsen. Reconstruction of genome-scale active metabolic networks for 69 human cell types and 16 cancer types using
INIT. PLoS Computational Biology, 8(5):e1002518, 2012.
[3] S. A. Becker and B. Ø. Palsson. Context-specific metabolic networks are consistent with experiments. PLoS computational biology, 4(5):e1000082, 2008.
[4] A. S. Blazier and J. A. Papin. Integration of expression data in genome-scale metabolic network
reconstructions. Frontiers in Physiology, 3:299, 2012.
[5] A. Bordbar, M. L. Mo, E. S. Nakayasu, A. C. Schrimpe-Rutledge, Y. M. Kim, T. O. Metz, M. B.
Jones, B. C. Frank, R. D. Smith, S. N. Peterson, et al. Model-driven multi-omic data analysis
elucidates metabolic immunomodulators of macrophage activation. Molecular Systems Biology,
8(1), 2012.
[6] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, Cambridge,
UK, 2004.
[7] S. Chandrasekaran and N. D. Price. Probabilistic integrative modeling of genome-scale metabolic
and regulatory networks in Escherichia coli and Mycobacterium tuberculosis. Proceedings of the
National Academy of Sciences, 107(41):17845–17850, 2010.
[8] R. L. Chang, L. Xie, L. Xie, P. E. Bourne, and B. Ø. Palsson. Drug off-target effects predicted
using structural analysis in the context of a metabolic network model. PLoS Computational
Biology, 6(9):e1000938, 2010.
[9] N. Christian, P. May, S. Kempa, T. Handorf, and O. Ebenhöh. An integrative approach towards
completing genome-scale metabolic networks. Mol. BioSyst., 5(12):1889–1903, 2009.
[10] V. Chvátal. A greedy heuristic for the set-covering problem. Mathematics of operations research,
4(3):233–235, 1979.
[11] R. J. DeBerardinis and C. B. Thompson. Cellular metabolism and disease: what do metabolic
outliers teach us? Cell, 148(6):1132–1144, 2012.
19
[12] J. M. Dreyfuss, J. D. Zucker, H. M. Hood, L. R. Ocasio, M. S. Sachs, and J. E. Galagan.
Reconstruction and validation of a genome-scale metabolic model for the filamentous fungus
Neurospora crassa using FARM. PLoS Comput Biol, 9(7):e1003126, 07 2013.
[13] N. C. Duarte, S. A. Becker, N. Jamshidi, I. Thiele, M. L. Mo, T. D. Vo, R. Srivas, and B. Ø.
Palsson. Global reconstruction of the human metabolic network based on genomic and bibliomic
data. Proceedings of the National Academy of Sciences, 104(6):1777–1782, 2007.
[14] R. M. T. Fleming, C. Maes, Y. Ye, M. A. Saunders, and B. Ø. Palsson. A variational principle for
computing nonequilibrium fluxes and potentials in genome-scale biochemical networks. Journal
of Theoretical Biology, 292:71–77, 2012.
[15] O. Folger, L. Jerby, C. Frezza, E. Gottlieb, E. Ruppin, and T. Shlomi. Predicting selective drug
targets in cancer through metabolic networks. Molecular systems biology, 7:501, 2011.
[16] J. Gagneur and S. Klamt. Computation of elementary modes: a unifying framework and the new
binary approach. BMC bioinformatics, 5(1):175, 2004.
[17] S. Gudmundsson and I. Thiele. Computationally efficient flux variability analysis. BMC bioinformatics, 11(1):489, 2010.
[18] T. Hao, H. W. Ma, X. M. Zhao, and I. Goryanin. Compartmentalization of the Edinburgh human
metabolic network. BMC bioinformatics, 11(1):393, 2010.
[19] K. Hiller and C. M. Metallo. Profiling metabolic networks to study cancer metabolism. Current
Opinion in Biotechnology, 24:60–68, 2013.
[20] P. A. Jensen and J. A. Papin. Functional integration of a metabolic network model and expression
data without arbitrary thresholding. Bioinformatics, 27(4):541–547, 2011.
[21] L. Jerby and E. Ruppin. Predicting drug targets and biomarkers of cancer via genome-scale
metabolic modeling. Clinical Cancer Research, 18(20):5572–5584, 2012.
[22] L. Jerby, T. Shlomi, and E. Ruppin. Computational reconstruction of tissue-specific metabolic
models: Application to human liver metabolism. Molecular Systems Biology, 6:401, 2010.
[23] A. A. Julius, M. Imielinski, and G. J. Pappas. Metabolic networks analysis using convex optimization. In 47th IEEE Conference on Decision and Control, pages 762–767, 2008.
[24] Brendan Lee, Hong Yu, Farook Jahoor, William O’Brien, Arthur L. Beaudet, and Peter Reeds.
In vivo urea cycle flux distinguishes and correlates with phenotypic severity in disorders of the
urea cycle. Proceedings of the National Academy of Sciences, 97(14):8021–8026, 2000.
[25] N. E. Lewis, G. Schramm, A. Bordbar, J. Schellenberger, M. P. Andersen, J. K. Cheng, N. Patel,
A. Yee, R. A. Lewis, R. Eils, et al. Large-scale in silico modeling of metabolic interactions between
cell types in the human brain. Nature biotechnology, 28(12):1279–1285, 2010.
[26] R. Mahadevan and C. H. Schilling. The effects of alternate optimal solutions in constraint-based
genome-scale metabolic models. Metabolic engineering, 5(4):264, 2003.
[27] J. Maniloff. The minimal cell genome: “on being the right size”. Proceedings of the National
Academy of Sciences, 93(19):10004, 1996.
[28] H. J. Morowitz. The completeness of molecular biology. Israel journal of medical sciences,
20(9):750, 1984.
20
[29] J. D. Orth, T. M. Conrad, J. Na, J. A. Lerman, H. Nam, A. M. Feist, and B. Ø. Palsson. A
comprehensive genome-scale reconstruction of Escherichia coli metabolism. Molecular systems
biology, 7(1), 2011.
[30] J. D. Orth, R. M. T. Fleming, and B. Ø. Palsson. Reconstruction and use of microbial metabolic
networks: the core Escherichia coli metabolic model as an educational guide. In A. Böck, R. Curtiss III, J. B. Kaper, P. D. Karp, F. C. Neidhardt, T. Nystrm, J. M. Slauch, C. L. Squires, and
D. Ussery, editors, Escherichia coli and Salmonella: Cellular and Molecular Biology. ASM Press,
Washington, DC, 2010.
[31] M. Pourfar, M. Niethammer, and D. Eidelberg. Metabolic networks in Parkinson’s disease. In
G. Grimaldi and M. Manto, editors, Mechanisms and Emerging Therapies in Tremor Disorders,
Contemporary Clinical Neuroscience, pages 403–415. Springer New York, 2013.
[32] N. D. Price, J. L. Reed, and B. Ø. Palsson. Genome-scale models of microbial cells: evaluating
the consequences of constraints. Nature Reviews Microbiology, 2(11):886–897, 2004.
[33] J. Schellenberger, R. Que, R. M. T. Fleming, I. Thiele, J. D. Orth, A. M. Feist, D. C. Zielinski,
A. Bordbar, N. E. Lewis, S. Rahmanian, J. Kang, D. R. Hyduke, and B. Ø Palsson. Quantitative
prediction of cellular metabolism with constraint-based models: the COBRA Toolbox v2.0. Nat
Protoc, 6(9):1290–1307, Sep 2011.
[34] S. Schuster and C. Hilgetag. On elementary flux modes in biochemical reaction systems at steady
state. Journal of Biological Systems, 2(2):165–182, 1994.
[35] T. Shlomi, M. N. Cabili, M.J. Herrgård, B. Ø. Palsson, and E. Ruppin. Network-based prediction
of human tissue-specific metabolism. Nature biotechnology, 26(9):1003–1010, 2008.
[36] G. Stephanopoulos, A. A. Aristidou, and J. Nielsen. Metabolic engineering: principles and methodologies. Academic Press, 1998.
[37] J. Suckow, P. Markiewicz, L. G. Kleina, J. Miller, B. Kisters-Woike, B. Müller-Hill, et al. Genetic
studies of the Lac repressor. XV: 4000 single amino acid substitutions and analysis of the resulting
phenotypes on the basis of the protein structure. Journal of molecular biology, 261(4):509, 1996.
[38] I. Thiele and B. Ø. Palsson. A protocol for generating a high-quality genome-scale metabolic
reconstruction. Nature protocols, 5(1):93–121, 2010.
[39] I. Thiele, N. Swainston, R. M. T. Fleming, A. Hoppe, S. Sahoo, M. K. Aurich, H. Haraldsdottir,
M. L. Mo, O. Rolfsson, M. D. Stobbe, et al. A community-driven global reconstruction of human
metabolism. Nature biotechnology (doi:10.1038/nbt.2488), 2013. doi:10.1038/nbt.2488.
[40] E. Vitkin and T. Shlomi. MIRAGE: a functional genomics-based approach for metabolic network
model reconstruction and its application to cyanobacteria networks. Genome biology, 13(11):R111,
2012.
[41] Y. Wang, J. A. Eddy, and N. D. Price. Reconstruction of genome-scale metabolic models for 126
human tissues using mCADRE. BMC Systems Biology, 6(1):153, 2012.
[42] A. R. Zomorrodi and C. D. Maranas. Improving the iMM904 S.cerevisiae metabolic model using
essentiality and synthetic lethality data. BMC systems biology, 4(1):178, 2010.
[43] Hadas Zur, Eytan Ruppin, and Tomer Shlomi. imat: an integrative metabolic analysis tool.
Bioinformatics, 26(24):3140–3142, 2010.
21
| 5cs.CE
|
Furthering Baseline Core Lucid Standard
Specification in the Context of the History of Lucid,
Intensional Programming, and Context-Aware
Computing
Paquet, Joey
Mokhov, Serguei A.
arXiv:1107.0940v4 [cs.PL] 21 Oct 2013
SIGLUCID
Abstract
This work is multifold. We review the historical literature on the Lucid programming
language, its dialects, intensional logic, intensional programming, the implementing systems,
and context-oriented and context-aware computing and so on that provide a contextual
framework for the converging Core Lucid standard programming model. We are designing
a standard specification of a baseline Lucid virtual machine for generic execution of Lucid
programs. The resulting Core Lucid language would inherit the properties of generalization
attempts of GIPL (1999–2013) and TransLucid (2008–2013) for all future and recent Lucidimplementing systems to follow. We also maintain this work across local research group in
order to foster deeper collaboration, maintain a list of recent and historical bibliography and
a reference manual and reading list for students. We form a (for now informal) SIGLUCID
group to keep track of this standard and historical records with eventual long-term goal
through iterative revisions for this work to become a book or an encyclopedia of the referenced
topics, and perhaps, an RFC. We first begin small with this initial set of notes.
Contents
1 Introduction
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Proposed Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 SIGLUCID . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2
3
3
3
2 Historical Perspective, Context, Dialects, and Applications
2.1 Lucid Dialects . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 Incomplete Brief History and The Family . . . . . . . .
2.2 List of Tools and Implementing Systems . . . . . . . . . . . . .
2.3 Application Domains . . . . . . . . . . . . . . . . . . . . . . . .
2.4 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
3
4
5
6
6
7
.
.
.
.
.
.
.
19
20
20
20
21
23
24
24
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Core Lucid Standard Specification Design
3.1 SIGLUCID Meetings . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1 SECASA 2010 Meeting at SERA 2010, Montreal, Canada . .
3.1.2 SECASA 2009 Meeting at COMPSAC 2009, Seattle, USA . .
3.1.3 SECASA 2008 Meeting at COMPSAC 2008, Turku, Finland
3.1.4 PLC 2005 Meeting at WORLDCOMP 2005, Las Vegas, USA
3.1.5 ISLIP 1999 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.6 ISLIP 1995 . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
3.2
3.3
3.4
TransLucid . . . . . . . . . . . . . . . . . . . . .
GIPSY . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Hybrid Interaction with Other Languages
3.3.2 Introduction to the GIPSY Type System
The Core Lucid Standard . . . . . . . . . . . . .
3.4.1 Syntax . . . . . . . . . . . . . . . . . . . .
3.4.2 Semantics . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
SIGLUCID
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
25
25
25
28
32
32
32
4 Conclusion
32
5 Future Work
33
6 Acknowledgments
33
List of Figures
1
2
3
GIPC Framework with the Preprocessor . . . . . . . . . . . . . . . . . . . . . . .
Example of Eductive Evaluation of Objective Lucid Progran . . . . . . . . . . . .
GIPSY Type System. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
27
30
Listings
1
Example of a hybrid GIPSY program. . . . . . . . . . . . . . . . . . . . . . . . .
29
List of Tables
1
1
Matching data types between Lucid and Java. . . . . . . . . . . . . . . . . . . . .
31
Introduction
This work gears toward a generalization on a number of previous results by various authors in
terms of context specification in the Lucid programming language and the Core Lucid dealing
with data types and have a virtual machine standard agreed to by SIGLUCID. Aside from
various data types (primarily to address the hybrid computing paradigms uniformly of Lucid
integrating with imperative dialects), the context definition should also be hierarchical for certain
application domains to allow for context nesting. The notion of context is central to Lucid as an
explicit meaning component that is specified as a first class-value. Traditional Lucid’s context
specification was assuming tags and the corresponding values were simple – i.e. a collection of
dimension names and the value pairs would denote a point in the context space. Then, the
notion of point was not sufficient for some Lucid dialects that needed higher-order contextual
notions, such as context sets to denote a context area or field instead of a point, as it was done
in Lucx. Another way to traverse a more complex notion of the context definition was done
in iHTML and related tools where nesting of the tags would denote the nesting of contextual
expressions forming a sort of contextual tree, where the actual tag values were at the leaves of the
tree. Then, a similar need arose in Forensic Lucid and MARFL to specify higher-order contexts
representing evidence and witness stories or configuration details, but allowing evaluation at
any level of the context tree rather than just the leaves. Thus, this work aims at unifying and
2
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
standardizing various context specifications under one uniform intermediate form that all Lucid
dialects can adhere to thereby making the community speak the same language and potentially
bring interoperability between various Lucid implementations and incarnations across University
groups working in the intensional programming domain.
1.1
Motivation
Higher-order context specification is needed for nested-level context that traditionally decomposes a higher-order value into its components, and equivalently from the components get to
the parent component. This is partitioned in any nested markup-like language, e.g. iHTML,
any XML-based definitions and descriptions of data and databases, configuration management
of a software system components, as well as domain-specific applications such as contextual
specification of a cyberforensic case where evidential statement is comprised of observation sequences representing encoded stories told by evidence and witnesses, which in turn decompose
into observations, and then into properties and duration components; which all-in-all comprise
a context of evaluation of a cyberforensic case. Thus, the need for higher-order contexts is
apparent as a fundamental pillar supporting higher-order intensional logic (HOIL).
Types other than the context also should be exposed to the programmer when needed and
allow for a wider range of data types and type systems to allow hybrid dialect interaction easier
as well as compiler optimization and run-time system parallelizations.
1.2
Proposed Solution
For the context specification, we propose to extend the notion of context to be a bi-directional
tree with the operators from GIPL, Lucx, iHTML and MARFL to query, switch, and traverse
the depth of the context hierarchy. The language that encompasses the new specification on the
syntax and semantic level is proposed to be called Core Lucid or Standard Lucid or Nominal
Lucid.
The type specification and code segments are augmented as presented in possible specification
from the SIGLUCID meetings and others in Section 3.1 and Section 3.3.1.
1.3
SIGLUCID
SIGLUCID: Special Interest Group on Lucid, Ubiquity, Context, Intensionality, and multiDimensionality. SIGLUCID is a working group of researchers in Lucid, intensional programming,
intensional logic, context-aware and context-oriented computing and the related application
domains (see Figure 2.3.
SIGLUCID currently is a loose affiliation of researchers, collaborators, and supporters in intensional logic, intensional programming, context-aware computing, etc. across Canada, Australia,
and other places.
Should you wish to be a part and contribute, contact the people listed at the title page. This
is a running draft to fill in the missing information as it becomes possible.
2
Historical Perspective, Context, Dialects, and Applications
The history of Lucid, multi-dimensional intensional programming and logic, context-orientation,
parallel, concurrent, and distributed eductive evaluation aspects can be traced through different
Lucid dialects, outlined in Section 2.1.
3
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
2.1
SIGLUCID
Lucid Dialects
Here we enumerate the Lucid dialects that came to be from either practical implementations
and/or theoretical frameworks to study the intensionality properties, context, and mathematical
and intensional logic foundations. We plan to make the list into a table or other presentation
means with the status of each language and the related citations.
• Lucid
• GIPL
• TransLucid
• Lucx
• GLU
• GLU#
• Indexical Lucid
• Tensor Lucid
• Partial Lucid
• JLucid
• Objective Lucid
• Onyx
• Forensic Lucid
• JOOIP
• MARFL
• IHTML
• IHTML2
• iPerl
• ISE
• vmake
• Lustre
• pLucid
• Luthid
4
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
2.1.1
SIGLUCID
Incomplete Brief History and The Family
From 1974 to Lucid Today (taken from [64], incomplete, to be updated:
1. Lucid as a Pipelined Dataflow Language through 1974-1977. Lucid was introduced by
Anchroft and Wadge in [7, 8]. Features:
• A purely declarative language for natural expression of iterative algorithms.
• Goals: semantics and verification of correctness of programming languages (for details
see [7, 8]).
• Operators as pipelined streams: one for initial element, and then all for the successor
ones.
2. Intensions, Indexical Lucid, GRanular Lucid (GLU, [45, 46]), circa 1996. More details on
these two dialects are provided further in the chapter as they directly relate to the theme
of this thesis. Features:
• Random access to streams in Indexical Lucid.
• First working hybrid intensional-imperative paradigm (C/Fortran and Indexical Lucid) in the form of GLU.
• Eduction or demand-driven execution (in GLU).
3. Partial Lucid, Tensor Lucid, 1999 [95].
• Partial Lucid is an intermediate experimental language used for demonstrative purposes in presenting the semantics of Lucid in [95].
• Tensor Lucid dialect was developed by Joey Paquet for plasma physics computations to illustrate advantages and expressiveness of Lucid over an equivalent solution
written in Fortran.
4. GIPL, 1999 [95].
• All Lucid dialects can be translated into this basic form of Lucid, GIPL through a set
of translation rules. (GIPL is in the foundation of the execution semantics of GIPSY
and its GIPC and GEE because its AST is the only type of AST GEE understands
when executing a GIPSY program).
5. RLucid, 1999, [34]
• A Lucid dialect for reactive real-time intensional programming.
6. JLucid, Objective Lucid, 2003 - 2005
• These dialects introduce a notion of hybrid and object-oriented programming in the
GIPSY with Java and Indexical Lucid and GIPL, and are discussed great detail in
the follow up chapters of this thesis.
7. Lucx [149], 2003 - 2005
• Kaiyu Wan introduces a notion of contexts as first-class values in Lucid, thereby
making Lucx the true intensional language.
5
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
8. Onyx [39], April 2004.
• Peter Grogono makes an experimental derivative of Lucid – Onyx to investigate on
lazy evaluation of arrays.
9. GLU# [93], 2004
• GLU# is an evolution of GLU where Lucid is embedded into C++.
2.2
List of Tools and Implementing Systems
1. GIPSY [102]
2. GLU
3. TransLucid
4. pLucid
5. libintense
2.3
Application Domains
1. Context-Aware Computing
2. Scientific Computing
3. Distributed and Parallel Evaluation
4. Ubiquitous and Mobile Computing
5. Wiki
6. Forensic Computing
7. Multimedia and Configuration Management
8. Program Verification
9. Software Engineering
10. Aspect-Oriented Programming
11. Web OS
12. Reactive Computing
13. Pervasive Computing
14. Autonomic Computing
15. Modeling and Simulation
16. Model Checking
6
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
2.4
SIGLUCID
Related Work
There is a vast amount of related and past work done. Over time we will provide brief historical
description of each or a group of works clustered by a specific theme either in this section or
relevant other sections. For now, however, we begin by citing them first, so anyone looking for
the references can look them up in a jiffy and make their choice accordingly. This is ideal for
graduate students and researchers starting in the subjects or looking for what’s been done that
they can benefit from.
Most recent on top:
• 2013
– Serguei A. Mokhov. Intensional Cyberforensics. PhD thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada,
September 2013
– Sleiman Rabah, Serguei A. Mokhov, and Joey Paquet. An interactive graph-based
automation assistant: A case study to manage the GIPSY’s distributed multi-tier
run-time system. In Proceedings of the ACM Research in Adaptive and Convergent
Systems (RACS 2013), pages 387–394, New York, NY, USA, October 2011–2013.
ACM. ISBN 978-1-4503-2348-2. Pre-print: http://arxiv.org/abs/1212.4123
• 2012
– Yi Ji, Serguei A. Mokhov, and Joey Paquet. Unifying and refactoring DMF to support
concurrent Jini and JMS DMS in GIPSY. In Bipin C. Desai, Sudhir P. Mudur,
and Emil I. Vassev, editors, Proceedings of the Fifth International C* Conference on
Computer Science and Software Engineering (C3S2E’12), pages 36–44, New York,
NY, USA, June 2010–2013. ACM. ISBN 978-1-4503-1084-0. doi: 10.1145/2347583.
2347588. Online e-print http://arxiv.org/abs/1012.2860
• 2011
– John Plaice. Cartesian programming. Technical Report UNSW-CSE-TR-1101, University of Grenoble, France, January 2011. Habilitation Thesis, online at ftp://ftp.
cse.unsw.edu.au/pub/doc/papers/UNSW/1101.pdf
– Yi Ji. Scalability evaluation of the GIPSY runtime system. Master’s thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal,
Canada, March 2011
– Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Reasoning about a simulated
printer case investigation with Forensic Lucid. In Pavel Gladyshev and Marcus K.
Rogers, editors, Proceedings of ICDF2C’11, number 0088 in LNICST, pages 282–296.
Springer, October 2011. ISBN 978-3-642-35514-1. doi: 10.1007/978-3-642-35515-8\
23. Submitted in 2011, appeared in 2012; online at http://arxiv.org/abs/0906.
5181
– Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. On the need for data flow
graph visualization of Forensic Lucid programs and forensic evidence, and their evaluation by GIPSY. In Proceedings of the Ninth Annual International Conference on
Privacy, Security and Trust (PST), 2011, pages 120–123. IEEE Computer Society,
July 2011. ISBN 978-1-4577-0582-3. doi: 10.1109/PST.2011.5971973. Short paper;
full version online at http://arxiv.org/abs/1009.5423
7
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
• 2010
– Serguei A. Mokhov and Joey Paquet. A type system for higher-order intensional logic
support for variable bindings in hybrid intensional-imperative programs in GIPSY. In
Tokuro Matsuo, Naohiro Ishii, and Roger Lee, editors, 9th IEEE/ACIS International
Conference on Computer and Information Science, IEEE/ACIS ICIS 2010, pages
921–928. IEEE Computer Society, May 2010. ISBN 978-0-7695-4147-1. doi: 10.1109/
ICIS.2010.156. Presented at SERA 2010; online at http://arxiv.org/abs/0906.
3919
– Serguei A. Mokhov and Joey Paquet. Using the General Intensional Programming
System (GIPSY) for evaluation of higher-order intensional logic (HOIL) expressions.
In Proceedings of SERA 2010, pages 101–109. IEEE Computer Society, May 2010.
ISBN 978-0-7695-4075-7. doi: 10.1109/SERA.2010.23. Online at http://arxiv.org/
abs/0906.3911
– Aihua Wu, Joey Paquet, and Serguei A. Mokhov. Object-oriented intensional programming: Intensional Java/Lucid classes. In Proceedings of SERA 2010, pages 158–
167. IEEE Computer Society, 2010. ISBN 978-0-7695-4075-7. doi: 10.1109/SERA.
2010.29. Online at: http://arxiv.org/abs/0909.0764
– Bin Han, Serguei A. Mokhov, and Joey Paquet. Advances in the design and implementation of a multi-tier architecture in the GIPSY environment with Java. In Proceedings
of SERA 2010, pages 259–266. IEEE Computer Society, 2010. ISBN 978-0-7695-40757. doi: 10.1109/SERA.2010.40. Online at http://arxiv.org/abs/0906.4837
– Bin Han. Towards a multi-tier runtime system for GIPSY. Master’s thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal,
Canada, 2010
– Yi Ji, Serguei A. Mokhov, and Joey Paquet. Unifying and refactoring DMF to support
concurrent Jini and JMS DMS in GIPSY. In Bipin C. Desai, Sudhir P. Mudur,
and Emil I. Vassev, editors, Proceedings of the Fifth International C* Conference on
Computer Science and Software Engineering (C3S2E’12), pages 36–44, New York,
NY, USA, June 2010–2013. ACM. ISBN 978-1-4503-1084-0. doi: 10.1145/2347583.
2347588. Online e-print http://arxiv.org/abs/1012.2860
– Serguei A. Mokhov. Hybrid Intensional Computing in GIPSY: JLucid, Objective
Lucid and GICF. LAP - Lambert Academic Publishing, March 2010. ISBN 978-38383-1198-2
– Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Towards automatic deduction
and event reconstruction using Forensic Lucid and probabilities to encode the IDS evidence. In S. Jha, R. Sommer, and C. Kreibich, editors, Proceedings of RAID’10, LNCS
6307, pages 508–509. Springer, September 2010. doi: 10.1007/978-3-642-15512-3\ 36
– Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. The need to support of
data flow graph visualization of Forensic Lucid programs, forensic evidence, and their
evaluation by GIPSY. [online], September 2010. Poster at VizSec’10; online at http:
//arxiv.org/abs/1009.5423
– Serguei A. Mokhov, Emil Vassev, Joey Paquet, and Mourad Debbabi. Towards a selfforensics property in the ASSL toolset. In Proceedings of C3S2E’10, pages 108–113.
ACM, May 2010. ISBN 978-1-60558-901-5. doi: 10.1145/1822327.1822342
8
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
• 2009
– Joey Paquet. Distributed eductive execution of hybrid intensional programs. In Proceedings of the 33rd Annual IEEE International Computer Software and Applications
Conference (COMPSAC’09), pages 218–224, Seattle, Washington, USA, July 2009.
IEEE Computer Society. ISBN 978-0-7695-3726-9
– Serguei A. Mokhov, Joey Paquet, and Xin Tong. A type system for hybrid intensionalimperative programming support in GIPSY. In Proceedings of C3S2E’09, pages 101–
107, New York, NY, USA, May 2009. ACM. ISBN 978-1-60558-401-0. doi: 10.1145/
1557626.1557642
– Ai Hua Wu. OO-IP Hybrid Language Design and a Framework Approach to the GIPC.
PhD thesis, Department of Computer Science and Software Engineering, Concordia
University, Montreal, Canada, 2009
– Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Towards automated deduction in blackmail case analysis with Forensic Lucid. In Joseph S. Gauthier, editor, Proceedings of the Huntsville Simulation Conference (HSC’09), pages 326–333. SCS, October 2009. ISBN 978-1-61738-587-2. Online at http://arxiv.org/abs/0906.0049
– Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Reasoning about a simulated
printer case investigation with Forensic Lucid. In Joseph S. Gauthier, editor, Proceedings of the Huntsville Simulation Conference (HSC’09), page 45. SCS, October
2009. ISBN 978-1-61738-587-2. Abstract, fully online at http://arxiv.org/abs/
0906.5181
– Serguei A. Mokhov and Emil Vassev. Self-forensics through case studies of small to
medium software systems. In Proceedings of IMF’09, pages 128–141. IEEE Computer
Society, September 2009. ISBN 978-0-7695-3807-5. doi: 10.1109/IMF.2009.19
– Serguei A. Mokhov. The role of self-forensics modeling for vehicle crash investigations
and event reconstruction simulation. In Joseph S. Gauthier, editor, Proceedings of
the Huntsville Simulation Conference (HSC’09), pages 342–349. SCS, October 2009.
ISBN 978-1-61738-587-2. Online at http://arxiv.org/abs/0905.2449
– Serguei A. Mokhov. Enhancing the formal cyberforensic approach with observation
modeling with credibility factors and mathematical theory of evidence. [online], also
in ;login: vol. 34, no. 6, p. 101, December 2009. Presented at WIPS at USENIX
Security’09, http://www.usenix.org/events/sec09/wips.html
– Serguei A. Mokhov. Towards improving validation, verification, crash investigations,
and event reconstruction of flight-critical systems with self-forensics. [online], June
2009. A white paper submitted in response to NASA’s RFI NNH09ZEA001L, http://
arxiv.org/abs/0906.1845, mentioned in http://ntrs.nasa.gov/archive/nasa/
casi.ntrs.nasa.gov/20100025593_2010028056.pdf
– Manuel Peralta, Supratik Mukhopadhyay, and Ramesh Bharadwaj. Automatic synthesis and deployment of intensional kahn process networks. In Dominik Ślȩzak, Tai
hoon Kim, Stephen S. Yau, Osvaldo Gervasi, and Byeong-Ho Kang, editors, Grid and
Distributed Computing, volume 63 of Communications in Computer and Information
Science, pages 73–87. Springer Berlin Heidelberg, 2009. ISBN 978-3-642-10548-7. doi:
10.1007/978-3-642-10549-4\ 10
• 2008
9
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Khaled M. Ben Hamed. Multidimensional Programs on Distributed Parallel Computers: Analysis and Implementation. PhD thesis, Computer Science, the University of
New Brunswick, February 2008
– John Plaice, Blanca Mancilla, and Gabriel Ditu. From Lucid to TransLucid: Iteration,
dataflow, intensional and Cartesian programming. Mathematics in Computer Science,
2(1):37–61, 2008. ISSN 1661-8270. doi: 10.1007/s11786-008-0043-9
– Mehmet A. Orgun, Chuchang Liu, and Abhaya C. Nayak. Knowledge representation,
reasoning and integration using temporal logic with clocks. Mathematics in Computer
Science, 2(1):143–163, 2008. ISSN 1661-8270. doi: 10.1007/s11786-008-0048-4
– Dominic A. Orchard and Steve Matthews. Integrating lucid’s declarative dataflow
paradigm into object-orientation. Mathematics in Computer Science, 2(1):103–122,
2008. ISSN 1661-8270. doi: 10.1007/s11786-008-0046-6
– Blanca Mancilla and John Plaice. Possible worlds versioning. Mathematics in Computer Science, 2(1):63–83, 2008. ISSN 1661-8270. doi: 10.1007/s11786-008-0044-8
– Angelos Charalambidis, Athanasios Grivas, Nikolaos S. Papaspyrou, and Panos Rondogiannis. Efficient intensional implementation for lazy functional languages. Mathematics in Computer Science, 2(1):123–141, 2008. ISSN 1661-8270. doi: 10.1007/
s11786-008-0047-5
– Amir Hossein Pourteymour. Comparative study of Demand Migration Framework
implementation using JMS and Jini. Master’s thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada, September
2008
– Xin Tong. Design and implementation of context calculus in the GIPSY. Master’s
thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada, April 2008
– Serguei A. Mokhov. Towards syntax and semantics of hierarchical contexts in multimedia processing applications using MARFL. In Proceedings of the 32nd Annual IEEE International Computer Software and Applications Conference (COMPSAC), pages 1288–1294, Turku, Finland, July 2008. IEEE Computer Society. doi:
10.1109/COMPSAC.2008.206
– Joey Paquet, Serguei A. Mokhov, and Xin Tong. Design and implementation of
context calculus in the GIPSY environment. In Proceedings of the 32nd Annual
IEEE International Computer Software and Applications Conference (COMPSAC),
pages 1278–1283, Turku, Finland, July 2008. IEEE Computer Society. doi: 10.1109/
COMPSAC.2008.200
– John Plaice, Blanca Mancilla, Gabriel Ditu, and William W. Wadge. Sequential
demand-driven evaluation of eager TransLucid. In Proceedings of the 32nd Annual
IEEE International Computer Software and Applications Conference (COMPSAC),
pages 1266–1271, Turku, Finland, July 2008. IEEE Computer Society. doi: 10.1109/
COMPSAC.2008.191
– Toby Rahilly and John Plaice. A multithreaded implementation for TransLucid. In
Proceedings of the 32nd Annual IEEE International Computer Software and Applications Conference (COMPSAC), pages 1272–1277, Turku, Finland, July 2008. IEEE
Computer Society. doi: 10.1109/COMPSAC.2008.191
10
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Amir Hossein Pourteymour, Emil Vassev, and Joey Paquet. Design and implementation of demand migration systems in GIPSY. In Proceedings of PDPTA 2009. CSREA
Press, June 2008
– Emil Vassev and Joey Paquet. Towards autonomic GIPSY. In Proceedings of the
Fifth IEEE Workshop on Engineering of Autonomic and Autonomous Systems (EASE
2008), pages 25–34. IEEE Computer Society, 2008. ISBN 978-0-7695-3140-3. doi:
10.1109/EASe.2008.9
– Serguei A. Mokhov and Joey Paquet. Formally specifying and proving operational
aspects of Forensic Lucid in Isabelle. Technical Report 2008-1-Ait Mohamed, Department of Electrical and Computer Engineering, Concordia University, Montreal,
Canada, August 2008. In Theorem Proving in Higher Order Logics (TPHOLs2008):
Emerging Trends Proceedings. Online at: http://users.encs.concordia.ca/~tphols08/
TPHOLs2008/ET/76-98.pdf and http://arxiv.org/abs/0904.3789
– Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Formally specifying operational semantics and language constructs of Forensic Lucid. In Oliver Göbel, Sandra
Frings, Detlef Günther, Jens Nedon, and Dirk Schadt, editors, Proceedings of the
IT Incident Management and IT Forensics (IMF’08), LNI140, pages 197–216. GI,
September 2008. ISBN 978-3-88579-234-5. Online at http://subs.emis.de/LNI/
Proceedings/Proceedings140/gi-proc-140-014.pdf
– Serguei A. Mokhov. Encoding forensic multimedia evidence from MARF applications
as Forensic Lucid expressions. In Tarek Sobh, Khaled Elleithy, and Ausif Mahmood,
editors, Novel Algorithms and Techniques in Telecommunications and Networking,
proceedings of CISSE’08, pages 413–416, University of Bridgeport, CT, USA, December 2008. Springer. ISBN 978-90-481-3661-2. doi: 10.1007/978-90-481-3662-9\ 71.
Printed in January 2010
– Serguei A. Mokhov. Towards security hardening of scientific distributed demanddriven and pipelined computing systems. In Proceedings of the 7th International
Symposium on Parallel and Distributed Computing (ISPDC’08), pages 375–382. IEEE
Computer Society, July 2008. ISBN 978-0-7695-3472-5. doi: 10.1109/ISPDC.2008.52
• 2007
– Amir Hossein Pourteymour, Emil Vassev, and Joey Paquet. Towards a new demanddriven message-oriented middleware in GIPSY. In Proceedings of PDPTA 2007, pages
91–97. PDPTA, CSREA Press, June 2007
– Xin Tong, Joey Paquet, and Serguei A. Mokhov. Complete context calculus design
and implementation in GIPSY. [online], 2007–2008. http://arxiv.org/abs/1002.
4392
– Gabriel Ditu. The Programming Language TransLucid. PhD thesis, University of
New South Wales, Australia, 2007
• 2006
– Kaiyu Wan. Lucx: Lucid Enriched with Context. PhD thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada,
2006
• 2005
11
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Kaiyu Wan, Vasu Alagar, and Joey Paquet. Real time reactive programming in
lucid enriched with contexts. In Zhiming Liu and Keijiro Araki, editors, Theoretical
Aspects of Computing - ICTAC 2004, volume 3407 of Lecture Notes in Computer
Science, pages 387–402. Springer Berlin Heidelberg, 2005. ISBN 978-3-540-25304-4.
doi: 10.1007/978-3-540-31862-0\ 28
– Kaiyu Wan, Vasu Alagar, and Joey Paquet. Lucx: Lucid enriched with context. In
Proceedings of the 2005 International Conference on Programming Languages and
Compilers (PLC 2005), pages 48–14. CSREA Press, June 2005
– Serguei Mokhov and Joey Paquet. General imperative compiler framework within
the GIPSY. In Proceedings of the 2005 International Conference on Programming
Languages and Compilers (PLC 2005), pages 36–42. CSREA Press, June 2005
– Serguei A. Mokhov. Towards hybrid intensional programming with JLucid, Objective
Lucid, and General Imperative Compiler Framework in the GIPSY. Master’s thesis,
Department of Computer Science and Software Engineering, Concordia University,
Montreal, Canada, October 2005. ISBN 0494102934; online at http://arxiv.org/
abs/0907.2640
– Serguei Mokhov and Joey Paquet. Objective Lucid – first step in object-oriented
intensional programming in the GIPSY. In Proceedings of the 2005 International
Conference on Programming Languages and Compilers (PLC 2005), pages 22–28.
CSREA Press, June 2005
– Peter Grogono, Serguei Mokhov, and Joey Paquet. Towards JLucid, Lucid with
embedded Java functions in the GIPSY. In Proceedings of the 2005 International
Conference on Programming Languages and Compilers (PLC 2005), pages 15–21.
CSREA Press, June 2005
– Joey Paquet and Ai Hua Wu. GIPSY – a platform for the investigation on intensional
programming languages. In Proceedings of the 2005 International Conference on
Programming Languages and Compilers (PLC 2005), pages 8–14. CSREA Press, June
2005
– Emil Vassev and Joey Paquet. A general architecture for demand migration in a
demand-driven execution engine in a heterogeneous and distributed environment.
In Proceedings of the 3rd Annual Communication Networks and Services Research
Conference (CNSR 2005), pages 176–182. IEEE Computer Society, May 2005. doi:
10.1109/CNSR.2005.9
– Ai Hua Wu and Joey Paquet. Object-oriented intensional programming in the GIPSY:
Preliminary investigations. In Proceedings of the 2005 International Conference on
Programming Languages and Compilers (PLC 2005), pages 43–47. CSREA Press,
June 2005
– Emil Vassev and Joey Paquet. A generic framework for migrating demands in the
GIPSY’s demand-driven execution engine. In Proceedings of the 2005 International
Conference on Programming Languages and Compilers (PLC 2005), pages 29–35.
CSREA Press, June 2005
– Emil Iordanov Vassev. General architecture for demand migration in the GIPSY
demand-driven execution engine. Master’s thesis, Department of Computer Science
and Software Engineering, Concordia University, Montreal, Canada, June 2005. ISBN
0494102969
12
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Kaiyu Wan, Vasu Alagar, and Joey Paquet. A context theory for intensional programming. In Workshop on Context Representation and Reasoning (CRR05), July
2005
– Melvin C. Fitting. FOIL axiomatized. [online], August 2005. http://comet.lehman.
cuny.edu/fitting/bookspapers/pdf/papers/FOILAxioms.pdf
• 2004
– Paul Swoboda. A Formalisation and Implementation of Distributed Intensional Programming. PhD thesis, The University of New South Wales, Sydney, Australia, 2004
– Bo Lu. Developing the Distributed Component of a Framework for Processing Intensional Programming Languages. PhD thesis, Department of Computer Science and
Software Engineering, Concordia University, Montreal, Canada, March 2004
– Paul Swoboda and John Plaice. A new approach to distributed context-aware computing. In A. Ferscha, H. Hoertner, and G. Kotsis, editors, Advances in Pervasive
Computing. Austrian Computer Society, 2004. ISBN 3-85403-176-9
– Paul Swoboda and John Plaice. An active functional intensional database. In
F. Galindo, editor, Advances in Pervasive Computing, pages 56–65. Springer, 2004.
LNCS 3180
– Peter Grogono. Intensional programming in Onyx. Technical report, Department
of Computer Science and Software Engineering, Concordia University, Montreal,
Canada, April 2004
– Joey Paquet, Aihua Wu, and Peter Grogono. Towards a framework for the General
Intensional Programming Compiler in the GIPSY. In Proceedings of the 19th Annual
ACM Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA 2004), pages 164–165, New York, NY, USA, October 2004. ACM.
doi: 10.1145/1028664.1028731
– Lei Tao. Warehouse and garbage collection in the GIPSY environment. Master’s
thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada, 2004
– Yimin Ding. Automated translation between graphical and textual representations
of intensional programs in the GIPSY. Master’s thesis, Department of Computer
Science and Software Engineering, Concordia University, Montreal, Canada, June
2004. http://newton.cs.concordia.ca/~paquet/filetransfer/publications/
theses/DingYiminMSc2004.pdf
– Nikolaos S. Papaspyrou and Ioannis T. Kassios. GLU# embedded in C++: a marriage
between multidimensional and object-oriented programming. Softw., Pract. Exper.,
34(7):609–630, 2004. ISSN 0038-0644. doi: 10.1002/spe.582
– Vasu S. Alagar, Joey Paquet, and Kaiyu Wan. Intensional programming for agent
communication. In Jo ao Leite, Andrea Omicini, Paolo Torroni, and Pinar Yolum,
editors, Declarative Agent Languages and Technologies II, volume 3476 of Lecture
Notes in Computer Science, pages 239–255. Springer Berlin Heidelberg, 2005. ISBN
978-3-540-26172-8. doi: 10.1007/11493402\ 14
• 2003
13
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Bo Lu, Peter Grogono, and Joey Paquet. Distributed execution of multidimensional
programming languages. In Proceedings of the 15th IASTED International Conference on Parallel and Distributed Computing and Systems (PDCS 2003), volume 1,
pages 284–289. International Association of Science and Technology for Development,
November 2003
– Ai Hua Wu, Joey Paquet, and Peter Grogono. Design of a compiler framework in
the GIPSY system. In Proceedings of the 15th IASTED International Conference on
Parallel and Distributed Computing and Systems (PDCS 2003), volume 1, pages 320–
328. International Association of Science and Technology for Development, November
2003
– William W. Wadge. Hammings problem example. [online], December 2003. http:
//i.csc.uvic.ca/home/hei/lup/contents.html
– Anand Ranganathan and Roy H. Campbell. A middleware for context-aware agents
in ubiquitous computing environments. In Markus Endler and Douglas Schmidt,
editors, Proceedings of Middleware 2003, volume 2672 of Lecture Notes in Computer
Science, pages 143–161. Springer Berlin Heidelberg, 2003. ISBN 978-3-540-40317-3.
doi: 10.1007/3-540-44892-6\ 8
– Simon Gay and Rajagopal Nagarajan. Intensional and extensional semantics of
dataflow programs. Formal Aspects of Computing, 15(4):299–318, 2003. ISSN 09345043. doi: 10.1007/s00165-003-0018-1
• 2002
– Peter Grogono. GIPC increments. Technical report, Department of Computer Science
and Software Engineering, Concordia University, Montreal, Canada, April 2002
– Ai Hua Wu. Semantic checking and translation in the GIPSY. Master’s thesis,
Department of Computer Science and Software Engineering, Concordia University,
Montreal, Canada, 2002
– Chun Lei Ren. General intensional programming compiler (GIPC) in the GIPSY.
Master’s thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada, 2002
• 2000
– Joey Paquet and Peter Kropf. The GIPSY architecture. In Proceedings of Distributed
Computing on the Web, Quebec City, Canada, 2000
• 1999
– Peter Kropf and John Plaice. Intensional objects. In International symposium on
Languages for Intensional Programming, pages 37–45, Athens, Greece, June 1999.
Demokrits Institute
– Panagiotis Rondogiannis. Adding multidimensionality to procedural programming
languages. Software: Practice and Experience, 29(13):1201–1221, 1999. ISSN 1097024X. doi: 10.1002/(SICI)1097-024X(199911)29:13h1201::AID-SPE278i3.0.CO;2-0
– Joey Paquet. Scientific Intensional Programming. PhD thesis, Department of Computer Science, Laval University, Sainte-Foy, Canada, 1999
14
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Manolis Gergatsoulis and Panos Rondogiannis, editors. Proceedings of ISLIP’99,
volume Intensional Programming II, June 1999. World Scientific. ISBN: 981-02-40953
– William W. Wadge. Intensional logic in context. In Gergatsoulis and Rondogiannis
[37], pages 1–13. Tutorial
– Thom Frühwirth. Constraint solving with constraint handling rules. In Gergatsoulis
and Rondogiannis [37], pages 14–30. Tutorial
– William W. Wadge and M. C. Schraefel. Putting the hyper back in hypertext. In
Gergatsoulis and Rondogiannis [37], pages 31–39
– M. C. Schraefel, Blanca Mancilla, and John Plaice. Intensional hypertext. In Gergatsoulis and Rondogiannis [37], pages 40–54
– Yu Zhang and Kang Zhang. Associative query for multi-version web documents. In
Gergatsoulis and Rondogiannis [37], pages 55–64
– Jiannong Cao, Alvin Chan, and Kang Zhang. Programming dynamically reconfigurable web server groups using the DyGOP model. In Gergatsoulis and Rondogiannis
[37], pages 65–77
– Alessandra Raffaetà and Thom Frühwirth. Two semantics for temporal annotated
constraint logic. In Gergatsoulis and Rondogiannis [37], pages 78–92
– Michael Fisher and Tony Kakoudakis. Flexible agent grouping in executable temporal
logic. In Gergatsoulis and Rondogiannis [37], pages 93–105
– Costas D. Koutras and Christos Nomikos. On the computational complexity of stratified negation in linear-time temporal logic programming. In Gergatsoulis and Rondogiannis [37], pages 106–117
– Manolis Gergatsoulis. Extensions of the branching-time logic programming language
Cactus. In Gergatsoulis and Rondogiannis [37], pages 118–132
– Themis Panayiotopoulos. Temporal reasoning with TRL. In Gergatsoulis and Rondogiannis [37], pages 133–148
– Loı̈c Besnard, Patricia Bourani, Thierry Gautier, Nicolas Halbwachs, Simin NadjmTehrani, and Annie Ressouche. Design of a multi-formalism application and distribution in a data-flow context: An example. In Gergatsoulis and Rondogiannis [37],
pages 149–167
– Jean-Raymond Gagné and John Plaice. Demand-driven real-time computing. In
Gergatsoulis and Rondogiannis [37], pages 168–181. ISBN: 981-02-4095-3
– Weiqiang Lin and Mehmet A. Orgun. Applied hidden periodicity analysis for mining
descrete-valued time series databases. In Gergatsoulis and Rondogiannis [37], pages
182–196
– Ion Androutsopoulos. Temporal meaning representation in a natural language frontend. In Gergatsoulis and Rondogiannis [37], pages 197–213
– Win Maung, Chit Swe, and Mehmet A. Orgun. Statistical queries on historical relational databases. In Gergatsoulis and Rondogiannis [37], pages 214–228
– Themis Panayiyotopoulos and L. C. Baxevanaki. Statistical queries on historical
relational databases. In Gergatsoulis and Rondogiannis [37], pages 229–243
15
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Weichang Du. Toward an intensional model for programming large scale distributed
systems. In Gergatsoulis and Rondogiannis [37], pages 244–258
– Joey Paquet and John Plaice. The semantics of dimensions as values. In Gergatsoulis
and Rondogiannis [37], pages 259–273
– Panos Rondogiannis. Adding multidimensionality to procedural programming languages. In Gergatsoulis and Rondogiannis [37], pages 274–291
– John Plaice and Peter G. Kropf. Intensional communities. In Gergatsoulis and
Rondogiannis [37], pages 292–295
– Paul Swoboda and William W. Wadge. Vmake and ISE general tools for the intensionalization of software systems. In Gergatsoulis and Rondogiannis [37], pages
310–320. ISBN: 981-02-4095-3
• 1998
– William W. Wadge, G. Brown, M. C. Schraefel, and T. Yildirim. Intensional HTML.
In 4th International Workshop PODDP’98, March 1998
• 1997
– Q. Zhao. Implementation of an object-oriented intensional programming system. Master’s thesis, Department of Computer Science, University of New Brunswick, Canada,
1997
– Raganswamy Jagannathan, Chris Dodd, and Iskender Agi. GLU: A high-level system
for granular data-parallel programming. In Concurrency: Practice and Experience,
volume 1, pages 63–83, 1997
– Mehmet A. Orgun and Weichang Du. Multi-dimensional logic programming: Theoretical foundations. Theoretical Computer Science, 185(2):319–345, 1997. ISSN
0304-3975. doi: 10.1016/S0304-3975(97)00048-0
• 1996
– Chris Dodd. Intensional Programming I, chapter Rank analysis in the GLU compiler,
pages 76–82. Volume Intensional Programming I of Orgun and Ashcroft [87], May
1996. ISBN: 981-02-2400-1
– Raganswamy Jagannathan and Chris Dodd. GLU programmer’s guide. Technical
report, SRI International, Menlo Park, California, 1996
• 1995
– Joey Paquet. Relational databases as multidimensional dataflows. Master’s thesis,
Departement d’Informatique, Université Laval, Québec, Canada, 1995
– Edward A. Ashcroft, Anthony A. Faustini, Rangaswamy Jagannathan, and William W.
Wadge. Multidimensional Programming. Oxford University Press, London, February
1995. ISBN: 978-0195075977
– Mehmet A. Orgun and Edward A. Ashcroft, editors. Proceedings of ISLIP’95, volume
Intensional Programming I, May 1995. World Scientific. ISBN: 981-02-2400-1
– Paul Caspi and Pouzet. A functional extension to Lustre. In Orgun and Ashcroft
[87], pages 15–29. Invited Contribution
16
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Edward A. Ashcroft. Multidimensional program verification: Reasoning about programs that deal with multidimensional objects. In Orgun and Ashcroft [87], pages
30–41. Invited Contribution
– David Abramson and Rok Sosič. Relative debugging using multiple program versions.
In Orgun and Ashcroft [87], pages 42–55. Invited Contribution
– William W. Wadge. Possible WOOrlds. In Orgun and Ashcroft [87], pages 56–62.
Invited Contribution
– R. Jagannathan. Intensional and extensional graphical models for GLU programming.
In Orgun and Ashcroft [87], pages 63–75
– Jiannong Cao, Lichucha Fernando, and Kang Zhang. Programming distributed systems based on graphs. In Orgun and Ashcroft [87], pages 83–95
– Yaowei Liu and John Staples. Building logic constructs into procedural programming
languages. In Orgun and Ashcroft [87], pages 96–109
– Lorenzo Verdoscia. ALFA fine grain dataflow machine. In Orgun and Ashcroft [87],
pages 110–134
– Iskender Agi. GLU for multidimensional signal processing. In Orgun and Ashcroft
[87]. URL citeseer.ist.psu.edu/agi95glu.html. ISBN: 981-02-2400-1
– John Plaice. Particle in-cell simulation with Lucid. In Orgun and Ashcroft [87], pages
149–161
– Satoshi Yamane. Real-time object-oriented specification and verification. In Orgun
and Ashcroft [87], pages 162–185
– Wanli Ma and Mehmet A. Orgun. Verifying MULTRAN programs with temporal
logic. In Orgun and Ashcroft [87], pages 186–206
– William W. Wadge and Alan Yoder. The Possible-World Wide Web. In Orgun and
Ashcroft [87], pages 207–213
– Joey Paquet and John Plaice. The intensional relation. In Orgun and Ashcroft [87],
pages 214–227
– Panos Rondogiannis and William W. Wadge. Extending the intensionalization algorithm to a broader class of higher-order programs. In Orgun and Ashcroft [87], pages
228–233
– Padmanabhan Krishnan. An asynchronous calculus based on absence of actions. In
Orgun and Ashcroft [87], pages 234–248
– Richard Buckland. Choice as a first class citizen. In Orgun and Ashcroft [87], pages
249–259
– Seiki Akama. A meta-level approach to modal logic programming. In Orgun and
Ashcroft [87], pages 260–272
– Tu Van Le. Fuzzy temporal Prolog. In Orgun and Ashcroft [87], pages 273–280
– Chuchang Liu and Mehmet A. Orgun. Knowledge-based simulation with Chronolog.
In Orgun and Ashcroft [87], pages 281–295
• 1994
– Panagiotis Rondogiannis. Higher-Order Functional Languages and Intensional Logic.
PhD thesis, Department of Computer Science, University of Victoria, Victoria, Canada,
1994
17
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Joey Paquet and John Plaice. On the design of an indexical query language. In
Proceedings of the Seventh International Symposium on Lucid and Intensional Programming, pages 28–36, 1994
– S. Tao. TLucid and Intensional Attribute Grammars. PhD thesis, Department of
Computer Science, Victoria University, Canada, 1994
– Weichang Du. Object-oriented implementation of intensional language. In Proceedings
of the 7th International Symposium on Lucid and Intensional Programming, pages
37–45, Menlo Park, California, USA, September 1994. SRI International
• 1993
– John Plaice and William W. Wadge. A new approach to version control. IEEE
Transactions on Software, 19(3):268–276, March 1993
– John Plaice, Ridha Khedri, and Ren’e Lalement. From abstract time to real time.
In In Proceedings of the Sixth International Symposium on Lucid and Intensional
Programming, pages 83–93, 1993
– Anthony A. Faustini and R. Jagannathan. Multidimensional problem solving in Lucid.
Technical Report SRI-CSL-93-03, SRI International, 1993
• 1991
– Weichang Du. Indexical Parallel Programming. PhD thesis, Department of Computer
Science, Victoria University, Canada, 1991
– B. Freeman-Benson. Lobjcid: Objects in Lucid. In Proceedings of the 1991 Symposium
on Lucid and Intensional Programming, pages 80–87, Menlo Park, California, USA,
April 1991. SRI International
• 1990
– Weichang Du and William W. Wadge. The eductive implementation of a threedimensional spreadsheet. Software Practice and Experience, 20(11):1097–1114, November 1990. ISSN 0038-0644
– Weichang Du and William W. Wadge. A 3D spreadsheet based on intensional logic.
IEEE Software, 7(3):78–89, June 1990. doi: 10.1109/52.55232
• 1989
– Anthony A. Faustini and E. B. Lewis. Towards a Real-Time Dataflow Language.
IEEE Computer Society, Los Alamitos, CA, USA, 1989. ISBN 0-8186-0819-6
• 1988
– J. van Benthem. A Manual of Intensional Logic. CSLI Publications, Stanford and
The University of Chicago Press, 1988. ISBN 0-937073-29-6
• 1987
– Sally C. Johnson. A strategy for automatically generating programs in the Lucid
programming language (NASA technical memorandum). Technical report, NASA,
Scientific and Technical Information Office, 1987. ASIN: B000711R3Q
18
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
– Antony A. Faustini and William W. Wadge. An eductive interpreter for the language
Lucid. SIGPLAN Not., 22(7):86–91, 1987. ISSN 0362-1340. doi: 10.1145/960114.
29659
• 1985
– William W. Wadge and Edward A. Ashcroft. Lucid, the Dataflow Programming
Language. Academic Press, London, 1985
• 1982
– Anthony A. Faustini. The Equivalence of a Denotational and an Operational Semantics of Pure Dataflow. PhD thesis, University of Warwick, Computer Science
Department, Coventry, United Kingdom, 1982
– Edward A. Ashcroft and William W. Wadge. R for semantics. ACM Transactions
on Programming Languages and Systems, 4(2):283–294, April 1982
• 1981
– C. B. Ostrum. The Luthid 1.0 Manual. Department of Computer Science, University
of Waterloo, Ontario, Canada, 1981
• 1977
– Edward A. Ashcroft and William W. Wadge. Lucid, a nonprocedural language with
iteration. Communications of the ACM, 20(7):519–526, July 1977. ISSN 0001-0782.
doi: 10.1145/359636.359715
– Edward A. Ashcroft and William W. Wadge. Erratum: Lucid – a formal system for
writing and proving programs. SIAM J. Comput., 6(1):200, 1977
• 1976
– Edward A. Ashcroft and William W. Wadge. Lucid – a formal system for writing
and proving programs. SIAM J. Comput., 5(3), 1976
Wikipedia and other Wiki entries:
• http://en.wikipedia.org/wiki/Lucid_(programming_language)
• http://en.wikipedia.org/wiki/Category:Intensional_programming_languages
• http://en.wikipedia.org/wiki/Intensional_logic
• http://www.haskell.org/haskellwiki/Lucid
3
Core Lucid Standard Specification Design
The Core Lucid standard design and specification is an ongoing process influenced by the two
core proposals: GIPL and TransLucid developed in 1999 and 2008 respectively.
19
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
3.1
SIGLUCID
SIGLUCID Meetings
Here’s the brief summary of the SIGLUCID meetings at various workshops and conferences,
attendees, and works contributing to the collaboration and developing the Core Lucid standard.
3.1.1
SECASA 2010 Meeting at SERA 2010, Montreal, Canada
Works
The following works were presented:
1. [43]
2. [75]
3. [155]
4. [74]
T ODO
Attendees
1. Serguei A. Mokhov
2. Joey Paquet
3. Emil Vassev
4. Bin Han
T ODO
3.1.2
SECASA 2009 Meeting at COMPSAC 2009, Seattle, USA
Works
The following works were presented:
1. [96]
T ODO
20
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
Attendees
1. Joey Paquet
2. John Plaice
T ODO
3.1.3
SECASA 2008 Meeting at COMPSAC 2008, Turku, Finland
The first discussion about standardizing the types, evaluation, and overview of the current
candidates for the Lucid Core from different research groups, such as GIPL, TransLucid. The
needs of various in-progress Lucid dialects were discussed to be accommodated in the core, such
as MARFL, Forensic Lucid. Below are the points from the meeting minutes.
Works
The following works were presented:
1. [104]
2. [112]
3. [118]
4. [67]
Attendees
1. Weichang Du
2. Blanca Mancilla
3. Serguei A. Mokhov
4. Joey Paquet
5. John Plaice
6. Toby Rahilly
7. William W. Wadge
Notes
1. Constants appearing in the expressions:
type<string> (const type -- int8<42> != int16<42>)
[| string |]
12
true
false
21
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
2. header – default types for integers, etc.
3. #parens – see the sections on the GIPSY type system and a hybrid program example
Section 3.3.1 and Section 3.3.2
4. Proposal of type (’type’ is a keyword).
• type<float 32>
• uchar<> uchar<>
• shorthand syntactical sugar:
[||| 1.2 |||] – 64 bit IEEE float
{{{ 1.2 }}} – 32 bit float
5. special<...> – correspond to exceptions and error situations for handling later on
6. Joey: Dimensions syntactically are allowed to taken on default values other than always
implicit default of zero.
• special<undecl> special¡arith¿ = special¡undecl¿+
– lose details, e.g. where it happened in the code or even within the imperative code?
• if(isspecial<undecl> E)
then ...
7. Bill has done something like that with someone in pLucid, with well defined semantics,
etc.
8. Bill complained about eagerness:
{E1 : E2 , ..., En : En2 }
eager: only lefts (multithreaded [118]), else all —, but right-hand-sides are lazy.
9. Q: How to stop people from producing recursive/infinite contexts?
10. Bill: risky: \# a@{a:P, E:Q} != P ??? If E does not terminate or special – can’t prove
it’s constant.
11. Toby: threading, sequential scheduling
12. Audience concluded: GIPL context-eager, TransLucid dimension-eager (LHS)
13. Variables
variable x
dimension d
type of x is context-dependent.
Future:
id<x>
dimension<d>
expr<E>
22
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
14. Dynamic ranks analysis (Tony Faustini and Bill in the ’80)
(X, C)?
? = eduction evaluation engine
W?(X, {})
->42
->{d1,....,dn}?
v1=C(d1)
vn=C{dn}
v’1=C(d1’)
v’m=C{dm’}
W?(X, {d1:v1, ....})
->42
-> {d’2,...,d’n}?
x, W, C, Cs, Ci # 3
W’ = W U {(X, Cs) |-> {3}?}
Ci – current internal context
15. Toby: optimization: demand grouping demands as early as possible, lifting up
16. optimization for constant vs. run-time dimensions thus
dimension d
is an optimization hint.
17. Binary representation (portable) ??? a-la Java byte-code
3.1.4
PLC 2005 Meeting at WORLDCOMP 2005, Las Vegas, USA
Works
The following works were presented:
1. [40]
2. [63]
3. [62]
4. [136]
5. [153]
6. [101]
7. [149]
T ODO
23
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
Attendees
1. Weichang Du
2. Serguei A. Mokhov
3. Joey Paquet
4. Emil Vassev
5. William W. Wadge
6. Kaiyu Wan
7. Aihua Wu
T ODO
3.1.5
ISLIP 1999
Works
The following works were presented:
1. [37]
T ODO
Attendees
1. William W. Wadge
2. John Plaice
3. Joey Paquet
4. ...
T ODO
3.1.6
ISLIP 1995
Works
The following works were presented:
1. [87]
T ODO
24
SIGLUCID
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
Attendees
1. William W. Wadge
2. John Plaice
3. Joey Paquet
4. ...
T ODO
3.2
TransLucid
T ODO
3.3
3.3.1
GIPSY
Hybrid Interaction with Other Languages
GIPC Preprocessor The Preprocessor [64, 62] is something that is invoked first by the
GIPC (see Figure 1) on incoming GIPSY program’s source code stream. The Preprocessor’s
role is to do preliminary program analysis, processing, and splitting the source GIPSY program
into “chunks”, each written in a different language and identified by a language tag. In a very
general view, a GIPSY program is a hybrid program consisting of different languages in one or
more source file; then, there has to be an interface between all these code segments. Thus, the
Preprocessor after some initial parsing (using its own preprocessor syntax) and producing the
initial parse tree, constructs a preliminary dictionary of symbols used throughout the program.
This is the basis for type matching and semantic analysis applied later on. This is also where
the first step of type assignment occurs, especially on the boundary between typed and typeless
parts of the program, e.g. Java and a specific Lucid dialect. The Preprocessor then splits the
code segments of the GIPSY program into chunks preparing them to be fed to the respective
concrete compilers for those chunks. The chunks are represented through the CodeSegment class
that the GIPC collects.
GIPSY Program Segments There are four baseline types of segments defined to be used
in a GIPSY program. These are:
• #funcdecl program segment declares function prototypes written as imperative language
functions defined later or externally from this program to be used by the intensional language part. The syntactical form of these prototypes is particular to GIPSY programs and
need not resemble the actual function prototype declaration they describe in their particular programming language. They serve as a basis for static and dynamic type assignment
and checking within the GIPSY type system with regards to procedural functions called
by other parts of the GIPSY program, e.g. the Lucid code segments.
25
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
Figure 1: GIPC Framework with the Preprocessor
• #typedecl segment lists all user-defined data types that can potentially be used by the
intensional part; usually objects. These are the types that do not explicitly appear in the
matching table in Table 1 describing the basic data types allowed in GIPSY programs.
• #<IMPERATIVELANG> segment declares that this is a code segment written in whatever
IMPERATIVELANG may be, for example #JAVA for Java, #CPP for C++, #FORTRAN for
Fortran, #PERL for Perl, #PYTHON for Python, etc.
• #<INTENSIONALLANG> segment declares that this is a code segment written in whatever
INTENSIONALLANG may be, for example #GIPL, #LUCX, #JOOIP, #INDEXICALLUCID,
26
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
Figure 2: Example of Eductive Evaluation of Objective Lucid Progran
27
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
#JLUCID, #OBJECTIVELUCID, #TENSORLUCID, #ONYX [39], #FORENSICLUCID [73], and #TRANSLUCID,
etc. as specified by the available GIPSY implementations and stubs. An example of a hybrid program is presented in Listing 1. The preamble of the program with the type and
function declaration segments are the main source of type information that is used at
compile time to annotate the nodes in the tree to help both static and semantic analysis.
3.3.2
Introduction to the GIPSY Type System
The introduction of JLucid, Objective Lucid, and GICF [64, 63, 62, 40] prompted the development of the GIPSY Type System as implicitly understood by the Lucid language and its
incarnation within the GIPSY to handle types in a more general manner as a glue between the
imperative and intensional languages within the system. Further evolution of Lucx introducing contexts as first-class values and JOOIP highlighted the need of the further development
of the type system to accommodate the more general properties of the intensional and hybrid
languages.
Matching Lucid and Java Data Types Here we present a case of interaction between
Lucid and Java. Allowing Lucid to call Java methods brings a set of issues related to the
data types, especially when it comes to type checks between Lucid and Java parts of a hybrid
program. This is pertinent when Lucid variables or expressions are used as parameters to Java
methods and when a Java method returns a result to be assigned to a Lucid variable or used
in an intensional expression. The sets of types in both cases are not exactly the same. The
basic set of Lucid data types as defined by Grogono [38] is int, bool, double, string, and
dimension. Lucid’s int is of the same size as Java’s long. GIPSY and Java double, boolean,
and String are roughly the same. Lucid string and Java String are simply mapped internally
through StringBuffer; thus, one can think of the Lucid string as a reference when evaluated
in the intensional program. Based on this fact, the lengths of a Lucid string and Java String
are the same. Java String is also an object in Java; however, at this point, a Lucid program
has no direct access to any String’s properties (though internally we do and we may expose
it later to the programmers). We also distinguish the float data type for single-precision
floating point operations. The dimension index type is said to be an integer or string (as far
as its dimension tag values are concerned), but might be of other types eventually, as discussed
in [133]. Therefore, we perform data type matching as presented in Table 1. Additionally, we
allow void Java return type which will always be matched to a Boolean expression true in
Lucid as an expression has to always evaluate to something. As for now our types mapping
and restrictions are as per Table 1. This is the mapping table for the Java-to-IPL-to-Java type
adapter. Such a table would exist for mapping between any imperative-to-intensional language
and back, e.g. the C++-to-IPL-to-C++ type adapter.
Overview of the Design and Implementation of the Type System. While the main language of GIPSY, Lucid, is polymorphic and does not have explicit types, co-existing with other
languages necessitates definition of GIPSY types and their mapping to a particular language
being embedded. Figure 3 presents the detailed design of the GIPSY Type System.
Each class is prefixed with GIPSY to avoid possible confusion with similar definitions in the
java.lang package. The GIPSYVoid type always evaluates to the Boolean true, as described
earlier in Section 3.3.2. The other types wrap around the corresponding Java object wrapper
28
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
/∗ ∗
∗ Language−mix GIPSY program .
∗ @author S e r g u e i Mokhov
∗/
#t y p e d e c l
myclass ;
#f u n c d e c l
m y c l a s s f o o ( int , double ) ;
f l o a t bar ( int , i n t ) : ” f t p : / / newton . c s . c o n c o r d i a . ca / c o o l . c l a s s ” : baz ;
int f 1 ( ) ;
#JAVA
m y c l a s s f o o ( i n t a , double b )
{
return new m y c l a s s (new I n t e g e r ( ( i n t ) ( b + a ) ) ) ;
}
class myclass
{
public m y c l a s s ( I n t e g e r a )
{
System . out . p r i n t l n ( a ) ;
}
}
#CPP
#i n c l u d e <i o s t r e a m >
i n t f 1 ( void )
{
c o u t << ” h e l l o ” ;
return 0 ;
}
#OBJECTIVELUCID
A + bar (B, C)
where
A = f o o (B, C) . i n t V a l u e ( ) ;
B = f1 () ;
C = 2.0;
end ;
/∗
∗ i n t h e o r y we c o u l d w r i t e more than one i n t e n s i o n a l chunk ,
∗ t h e n t h o s e chunks would e v a l u a t e as s e p a r a t e p o s s i b l y
∗ t o t a l l y i n d e p e n d e n t e x p r e s s i o n s i n p a r a l l e l t h a t happened
∗ t o u s e t h e same s e t o f i m p e r a t i v e f u n c t i o n s .
∗/
// EOF
Listing 1: Example of a hybrid GIPSY program.
29
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
Figure 3: GIPSY Type System.
30
SIGLUCID
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
Table 1: Matching data types between Lucid and Java.
Return Types of Java Methods
Types of Lucid Expressions
Internal GIPSY Types
int, byte, long
float
double
boolean
char
String
Method
Method
[]
Object
Object
void
int, dimension
float
double
bool
char
string, dimension
function
operator
[]
class
URL
bool::true
GIPSYInteger
GIPSYFloat
GIPSYDouble
GIPSYBoolean
GIPSYCharacter
GIPSYString
GIPSYFunction
GIPSYOperator
GIPSYArray
GIPSYObject
GIPSYEmbed
GIPSYVoid
Parameter Types Used in Lucid
Corresponding Java Types
Internal GIPSY Types
string
float
double
int
dimension
bool
class
URL
[]
operator
function
String
float
double
int
int, String
boolean
Object
Object
[]
Method
Method
GIPSYString
GIPSYFloat
GIPSYDouble
GIPSYInteger
Dimension
GIPSYBoolean
GIPSYObject
GIPSYEmbed
GIPSYArray
GIPSYOperator
GIPSYFunction
classes for the primitive types, such as Long, Float, etc. Every class keeps a lexeme (a lexical representation) of the corresponding type in a GIPSY program and overrides toString() to show
the lexeme and the contained value. These types are extensively used by the Preprocessor,
imperative and intensional (for constants) compilers, the SequentialThreadGenerator, and
SemanticAnalyzer for the general type of GIPSY program processing, and by the GEE’s
Executor.
The other special types that have been created are either experimental or do not correspond
to a wrapper of a primitive type. GIPSYIdentifier type case corresponds to a declaration of
some sort of an identifier in a GIPSY program to be put into the dictionary, be it a variable or a
function name with the reference to their definition. Constants and conditionals may be anonymous and thereby not have a corresponding identifier. GIPSYEmbed is another special type that
encapsulates embedded code via the URL parameter and later is exploded into multiple types
corresponding to procedural demands (Java or any other language methods or functions) [64, 40].
GIPSYFunction and its descendant GIPSYOperator correspond to the function types for regular operators and user-defined functions. A GIPSYFunction can either encapsulate an ordinary
Lucid function (which is immutable as in functional programming) or a procedure (e.g. a Java
method), which may often be mutable (i.e. with side effects). These four types (identifier,
embed, function, and operator) are not directly exposed to a GIPSY programmer and at this
point are managed internally. By the latter we mean we have not reached the stage when we
31
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
can provide them for explicit use by programmers; however, the semantics of is still defined and
specified at the requirements, design, and implementation levels. GIPSYContext and Dimension
are a new addition to the type system implementation since [64]. They represent context-asfirst-class-values in the context calculus defined by Wan in [147] and refined and implemented by
Tong [132]. The rest of the type system is exposed to the GIPSY programmer in the preamble
of a GIPSY program, i.e., the #funcdecl and #typedecl segments, which result in the embryo
of the dictionary for linking, semantic analysis, and execution. Once imperative compilers of
procedural demands return, the type data structures (return and parameter types) declared in
the preamble are matched against what was discovered by the compilers and if the match is
successful, the link is made. By capturing the types such as identifier, embed, function, operator
and context, dimension, the GIPSY type system lays down fundamentals the higher-order intensional logic (HOIL) support that combines functional programming, intensional logic, context
calculus, and in some instances hybrid paradigm support, and the corresponding types. We
describe various properties of the concrete GIPSY types and their more detailed specification in
Appendix ?? and Appendix ??. There we detail the inner workings of each type in more detail as
well describe some of the properties through the notions of existential, union, intersection, and
linear types. They have been excluded from the main body of the article due to size limitations.
3.4
The Core Lucid Standard
The Core Lucid standard specification, syntax, semantics, translation rules, type system, and
verifications are to be placed in this section upon consensus of the SIGLUCID members.
T ODO
3.4.1
Syntax
T ODO
3.4.2
Semantics
T ODO
4
Conclusion
We have layed out the first foundational notions of a practical Lucid standard at the 1st SECASA
in 2008 in Turku, Finland, associated with COMPSAC 2008. Since then two (3) more SECASA’s
happened: in 2009 in Seattle, 2010 in Montreal, and another one is planned in 2011. Prior that
we are producing this first set of notes from the meeting and related work.
32
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
5
SIGLUCID
Future Work
We plan on further meet and refine these notes and the standards and further accrete the
related work. Our eventual goal after the standard draft is complete publish it along with a
comprehensive survey of the recent related work as well as historical review.
6
Acknowledgments
This research work was funded by the respective grants, faculties and departments of the authors.
Acknowledgment also goes to the former and present colleagues collaborated with us in the past.
References
[1] David Abramson and Rok Sosič. Relative debugging using multiple program versions. In
Orgun and Ashcroft [87], pages 42–55. Invited Contribution.
[2] Iskender Agi. GLU for multidimensional signal processing. In Orgun and Ashcroft [87].
URL citeseer.ist.psu.edu/agi95glu.html. ISBN: 981-02-2400-1.
[3] Seiki Akama. A meta-level approach to modal logic programming. In Orgun and Ashcroft
[87], pages 260–272.
[4] Vasu S. Alagar, Joey Paquet, and Kaiyu Wan. Intensional programming for agent communication. In Jo ao Leite, Andrea Omicini, Paolo Torroni, and Pinar Yolum, editors,
Declarative Agent Languages and Technologies II, volume 3476 of Lecture Notes in Computer Science, pages 239–255. Springer Berlin Heidelberg, 2005. ISBN 978-3-540-26172-8.
doi: 10.1007/11493402\ 14.
[5] Ion Androutsopoulos. Temporal meaning representation in a natural language front-end.
In Gergatsoulis and Rondogiannis [37], pages 197–213.
[6] Edward A. Ashcroft. Multidimensional program verification: Reasoning about programs
that deal with multidimensional objects. In Orgun and Ashcroft [87], pages 30–41. Invited
Contribution.
[7] Edward A. Ashcroft and William W. Wadge. Lucid – a formal system for writing and
proving programs. SIAM J. Comput., 5(3), 1976.
[8] Edward A. Ashcroft and William W. Wadge. Erratum: Lucid – a formal system for writing
and proving programs. SIAM J. Comput., 6(1):200, 1977.
[9] Edward A. Ashcroft and William W. Wadge. Lucid, a nonprocedural language with iteration. Communications of the ACM, 20(7):519–526, July 1977. ISSN 0001-0782. doi:
10.1145/359636.359715.
[10] Edward A. Ashcroft and William W. Wadge. R for semantics. ACM Transactions on
Programming Languages and Systems, 4(2):283–294, April 1982.
[11] Edward A. Ashcroft, Anthony A. Faustini, Rangaswamy Jagannathan, and William W.
Wadge. Multidimensional Programming. Oxford University Press, London, February 1995.
ISBN: 978-0195075977.
33
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[12] Loı̈c Besnard, Patricia Bourani, Thierry Gautier, Nicolas Halbwachs, Simin NadjmTehrani, and Annie Ressouche. Design of a multi-formalism application and distribution in
a data-flow context: An example. In Gergatsoulis and Rondogiannis [37], pages 149–167.
[13] Richard Buckland. Choice as a first class citizen. In Orgun and Ashcroft [87], pages
249–259.
[14] Jiannong Cao, Lichucha Fernando, and Kang Zhang. Programming distributed systems
based on graphs. In Orgun and Ashcroft [87], pages 83–95.
[15] Jiannong Cao, Alvin Chan, and Kang Zhang. Programming dynamically reconfigurable
web server groups using the DyGOP model. In Gergatsoulis and Rondogiannis [37], pages
65–77.
[16] Paul Caspi and Pouzet. A functional extension to Lustre. In Orgun and Ashcroft [87],
pages 15–29. Invited Contribution.
[17] Angelos Charalambidis, Athanasios Grivas, Nikolaos S. Papaspyrou, and Panos Rondogiannis. Efficient intensional implementation for lazy functional languages. Mathematics in
Computer Science, 2(1):123–141, 2008. ISSN 1661-8270. doi: 10.1007/s11786-008-0047-5.
[18] Yimin Ding.
Automated translation between graphical and textual representations of intensional programs in the GIPSY. Master’s thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada,
June 2004. http://newton.cs.concordia.ca/~paquet/filetransfer/publications/
theses/DingYiminMSc2004.pdf.
[19] Gabriel Ditu. The Programming Language TransLucid. PhD thesis, University of New
South Wales, Australia, 2007.
[20] Chris Dodd. Intensional Programming I, chapter Rank analysis in the GLU compiler,
pages 76–82. Volume Intensional Programming I of Orgun and Ashcroft [87], May 1996.
ISBN: 981-02-2400-1.
[21] Weichang Du. Indexical Parallel Programming. PhD thesis, Department of Computer
Science, Victoria University, Canada, 1991.
[22] Weichang Du. Object-oriented implementation of intensional language. In Proceedings
of the 7th International Symposium on Lucid and Intensional Programming, pages 37–45,
Menlo Park, California, USA, September 1994. SRI International.
[23] Weichang Du. Toward an intensional model for programming large scale distributed systems. In Gergatsoulis and Rondogiannis [37], pages 244–258.
[24] Weichang Du and William W. Wadge. A 3D spreadsheet based on intensional logic. IEEE
Software, 7(3):78–89, June 1990. doi: 10.1109/52.55232.
[25] Weichang Du and William W. Wadge. The eductive implementation of a three-dimensional
spreadsheet. Software Practice and Experience, 20(11):1097–1114, November 1990. ISSN
0038-0644.
[26] Anthony A. Faustini. The Equivalence of a Denotational and an Operational Semantics
of Pure Dataflow. PhD thesis, University of Warwick, Computer Science Department,
Coventry, United Kingdom, 1982.
34
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[27] Anthony A. Faustini and R. Jagannathan. Multidimensional problem solving in Lucid.
Technical Report SRI-CSL-93-03, SRI International, 1993.
[28] Anthony A. Faustini and E. B. Lewis. Towards a Real-Time Dataflow Language. IEEE
Computer Society, Los Alamitos, CA, USA, 1989. ISBN 0-8186-0819-6.
[29] Antony A. Faustini and William W. Wadge. An eductive interpreter for the language
Lucid. SIGPLAN Not., 22(7):86–91, 1987. ISSN 0362-1340. doi: 10.1145/960114.29659.
[30] Michael Fisher and Tony Kakoudakis. Flexible agent grouping in executable temporal
logic. In Gergatsoulis and Rondogiannis [37], pages 93–105.
[31] Melvin C. Fitting. FOIL axiomatized. [online], August 2005. http://comet.lehman.
cuny.edu/fitting/bookspapers/pdf/papers/FOILAxioms.pdf.
[32] B. Freeman-Benson. Lobjcid: Objects in Lucid. In Proceedings of the 1991 Symposium
on Lucid and Intensional Programming, pages 80–87, Menlo Park, California, USA, April
1991. SRI International.
[33] Thom Frühwirth. Constraint solving with constraint handling rules. In Gergatsoulis and
Rondogiannis [37], pages 14–30. Tutorial.
[34] Jean-Raymond Gagné and John Plaice. Demand-driven real-time computing. In Gergatsoulis and Rondogiannis [37], pages 168–181. ISBN: 981-02-4095-3.
[35] Simon Gay and Rajagopal Nagarajan. Intensional and extensional semantics of dataflow
programs. Formal Aspects of Computing, 15(4):299–318, 2003. ISSN 0934-5043. doi:
10.1007/s00165-003-0018-1.
[36] Manolis Gergatsoulis. Extensions of the branching-time logic programming language Cactus. In Gergatsoulis and Rondogiannis [37], pages 118–132.
[37] Manolis Gergatsoulis and Panos Rondogiannis, editors. Proceedings of ISLIP’99, volume
Intensional Programming II, June 1999. World Scientific. ISBN: 981-02-4095-3.
[38] Peter Grogono. GIPC increments. Technical report, Department of Computer Science and
Software Engineering, Concordia University, Montreal, Canada, April 2002.
[39] Peter Grogono. Intensional programming in Onyx. Technical report, Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada, April
2004.
[40] Peter Grogono, Serguei Mokhov, and Joey Paquet. Towards JLucid, Lucid with embedded
Java functions in the GIPSY. In Proceedings of the 2005 International Conference on
Programming Languages and Compilers (PLC 2005), pages 15–21. CSREA Press, June
2005.
[41] Khaled M. Ben Hamed. Multidimensional Programs on Distributed Parallel Computers:
Analysis and Implementation. PhD thesis, Computer Science, the University of New
Brunswick, February 2008.
[42] Bin Han. Towards a multi-tier runtime system for GIPSY. Master’s thesis, Department
of Computer Science and Software Engineering, Concordia University, Montreal, Canada,
2010.
35
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[43] Bin Han, Serguei A. Mokhov, and Joey Paquet. Advances in the design and implementation of a multi-tier architecture in the GIPSY environment with Java. In Proceedings of
SERA 2010, pages 259–266. IEEE Computer Society, 2010. ISBN 978-0-7695-4075-7. doi:
10.1109/SERA.2010.40. Online at http://arxiv.org/abs/0906.4837.
[44] R. Jagannathan. Intensional and extensional graphical models for GLU programming. In
Orgun and Ashcroft [87], pages 63–75.
[45] Raganswamy Jagannathan and Chris Dodd. GLU programmer’s guide. Technical report,
SRI International, Menlo Park, California, 1996.
[46] Raganswamy Jagannathan, Chris Dodd, and Iskender Agi. GLU: A high-level system for
granular data-parallel programming. In Concurrency: Practice and Experience, volume 1,
pages 63–83, 1997.
[47] Yi Ji. Scalability evaluation of the GIPSY runtime system. Master’s thesis, Department
of Computer Science and Software Engineering, Concordia University, Montreal, Canada,
March 2011.
[48] Yi Ji, Serguei A. Mokhov, and Joey Paquet. Unifying and refactoring DMF to support
concurrent Jini and JMS DMS in GIPSY. In Bipin C. Desai, Sudhir P. Mudur, and
Emil I. Vassev, editors, Proceedings of the Fifth International C* Conference on Computer
Science and Software Engineering (C3S2E’12), pages 36–44, New York, NY, USA, June
2010–2013. ACM. ISBN 978-1-4503-1084-0. doi: 10.1145/2347583.2347588. Online e-print
http://arxiv.org/abs/1012.2860.
[49] Sally C. Johnson. A strategy for automatically generating programs in the Lucid programming language (NASA technical memorandum). Technical report, NASA, Scientific
and Technical Information Office, 1987. ASIN: B000711R3Q.
[50] Costas D. Koutras and Christos Nomikos. On the computational complexity of stratified
negation in linear-time temporal logic programming. In Gergatsoulis and Rondogiannis
[37], pages 106–117.
[51] Padmanabhan Krishnan. An asynchronous calculus based on absence of actions. In Orgun
and Ashcroft [87], pages 234–248.
[52] Peter Kropf and John Plaice. Intensional objects. In International symposium on Languages for Intensional Programming, pages 37–45, Athens, Greece, June 1999. Demokrits
Institute.
[53] Tu Van Le. Fuzzy temporal Prolog. In Orgun and Ashcroft [87], pages 273–280.
[54] Weiqiang Lin and Mehmet A. Orgun. Applied hidden periodicity analysis for mining
descrete-valued time series databases. In Gergatsoulis and Rondogiannis [37], pages 182–
196.
[55] Chuchang Liu and Mehmet A. Orgun. Knowledge-based simulation with Chronolog. In
Orgun and Ashcroft [87], pages 281–295.
[56] Yaowei Liu and John Staples. Building logic constructs into procedural programming
languages. In Orgun and Ashcroft [87], pages 96–109.
36
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[57] Bo Lu. Developing the Distributed Component of a Framework for Processing Intensional
Programming Languages. PhD thesis, Department of Computer Science and Software
Engineering, Concordia University, Montreal, Canada, March 2004.
[58] Bo Lu, Peter Grogono, and Joey Paquet. Distributed execution of multidimensional programming languages. In Proceedings of the 15th IASTED International Conference on
Parallel and Distributed Computing and Systems (PDCS 2003), volume 1, pages 284–289.
International Association of Science and Technology for Development, November 2003.
[59] Wanli Ma and Mehmet A. Orgun. Verifying MULTRAN programs with temporal logic.
In Orgun and Ashcroft [87], pages 186–206.
[60] Blanca Mancilla and John Plaice. Possible worlds versioning. Mathematics in Computer
Science, 2(1):63–83, 2008. ISSN 1661-8270. doi: 10.1007/s11786-008-0044-8.
[61] Win Maung, Chit Swe, and Mehmet A. Orgun. Statistical queries on historical relational
databases. In Gergatsoulis and Rondogiannis [37], pages 214–228.
[62] Serguei Mokhov and Joey Paquet. General imperative compiler framework within the
GIPSY. In Proceedings of the 2005 International Conference on Programming Languages
and Compilers (PLC 2005), pages 36–42. CSREA Press, June 2005.
[63] Serguei Mokhov and Joey Paquet. Objective Lucid – first step in object-oriented intensional programming in the GIPSY. In Proceedings of the 2005 International Conference
on Programming Languages and Compilers (PLC 2005), pages 22–28. CSREA Press, June
2005.
[64] Serguei A. Mokhov. Towards hybrid intensional programming with JLucid, Objective
Lucid, and General Imperative Compiler Framework in the GIPSY. Master’s thesis, Department of Computer Science and Software Engineering, Concordia University, Montreal,
Canada, October 2005. ISBN 0494102934; online at http://arxiv.org/abs/0907.2640.
[65] Serguei A. Mokhov. Towards security hardening of scientific distributed demand-driven
and pipelined computing systems. In Proceedings of the 7th International Symposium on
Parallel and Distributed Computing (ISPDC’08), pages 375–382. IEEE Computer Society,
July 2008. ISBN 978-0-7695-3472-5. doi: 10.1109/ISPDC.2008.52.
[66] Serguei A. Mokhov. Encoding forensic multimedia evidence from MARF applications as
Forensic Lucid expressions. In Tarek Sobh, Khaled Elleithy, and Ausif Mahmood, editors,
Novel Algorithms and Techniques in Telecommunications and Networking, proceedings of
CISSE’08, pages 413–416, University of Bridgeport, CT, USA, December 2008. Springer.
ISBN 978-90-481-3661-2. doi: 10.1007/978-90-481-3662-9\ 71. Printed in January 2010.
[67] Serguei A. Mokhov. Towards syntax and semantics of hierarchical contexts in multimedia
processing applications using MARFL. In Proceedings of the 32nd Annual IEEE International Computer Software and Applications Conference (COMPSAC), pages 1288–1294,
Turku, Finland, July 2008. IEEE Computer Society. doi: 10.1109/COMPSAC.2008.206.
[68] Serguei A. Mokhov. Enhancing the formal cyberforensic approach with observation modeling with credibility factors and mathematical theory of evidence. [online], also in ;login:
vol. 34, no. 6, p. 101, December 2009. Presented at WIPS at USENIX Security’09,
http://www.usenix.org/events/sec09/wips.html.
37
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[69] Serguei A. Mokhov. The role of self-forensics modeling for vehicle crash investigations
and event reconstruction simulation. In Joseph S. Gauthier, editor, Proceedings of the
Huntsville Simulation Conference (HSC’09), pages 342–349. SCS, October 2009. ISBN
978-1-61738-587-2. Online at http://arxiv.org/abs/0905.2449.
[70] Serguei A. Mokhov. Towards improving validation, verification, crash investigations, and
event reconstruction of flight-critical systems with self-forensics. [online], June 2009. A
white paper submitted in response to NASA’s RFI NNH09ZEA001L, http://arxiv.org/
abs/0906.1845, mentioned in http://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.
gov/20100025593_2010028056.pdf.
[71] Serguei A. Mokhov. Hybrid Intensional Computing in GIPSY: JLucid, Objective Lucid
and GICF. LAP - Lambert Academic Publishing, March 2010. ISBN 978-3-8383-1198-2.
[72] Serguei A. Mokhov. Intensional Cyberforensics. PhD thesis, Department of Computer
Science and Software Engineering, Concordia University, Montreal, Canada, September
2013.
[73] Serguei A. Mokhov and Joey Paquet. Formally specifying and proving operational aspects
of Forensic Lucid in Isabelle. Technical Report 2008-1-Ait Mohamed, Department of Electrical and Computer Engineering, Concordia University, Montreal, Canada, August 2008.
In Theorem Proving in Higher Order Logics (TPHOLs2008): Emerging Trends Proceedings. Online at: http://users.encs.concordia.ca/~tphols08/TPHOLs2008/ET/76-98.
pdf and http://arxiv.org/abs/0904.3789.
[74] Serguei A. Mokhov and Joey Paquet. Using the General Intensional Programming System
(GIPSY) for evaluation of higher-order intensional logic (HOIL) expressions. In Proceedings of SERA 2010, pages 101–109. IEEE Computer Society, May 2010. ISBN 978-0-76954075-7. doi: 10.1109/SERA.2010.23. Online at http://arxiv.org/abs/0906.3911.
[75] Serguei A. Mokhov and Joey Paquet. A type system for higher-order intensional logic
support for variable bindings in hybrid intensional-imperative programs in GIPSY. In
Tokuro Matsuo, Naohiro Ishii, and Roger Lee, editors, 9th IEEE/ACIS International
Conference on Computer and Information Science, IEEE/ACIS ICIS 2010, pages 921–928.
IEEE Computer Society, May 2010. ISBN 978-0-7695-4147-1. doi: 10.1109/ICIS.2010.156.
Presented at SERA 2010; online at http://arxiv.org/abs/0906.3919.
[76] Serguei A. Mokhov and Emil Vassev. Self-forensics through case studies of small to medium
software systems. In Proceedings of IMF’09, pages 128–141. IEEE Computer Society,
September 2009. ISBN 978-0-7695-3807-5. doi: 10.1109/IMF.2009.19.
[77] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Formally specifying operational semantics and language constructs of Forensic Lucid. In Oliver Göbel, Sandra
Frings, Detlef Günther, Jens Nedon, and Dirk Schadt, editors, Proceedings of the IT Incident Management and IT Forensics (IMF’08), LNI140, pages 197–216. GI, September 2008. ISBN 978-3-88579-234-5. Online at http://subs.emis.de/LNI/Proceedings/
Proceedings140/gi-proc-140-014.pdf.
[78] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Towards automated deduction in
blackmail case analysis with Forensic Lucid. In Joseph S. Gauthier, editor, Proceedings of
the Huntsville Simulation Conference (HSC’09), pages 326–333. SCS, October 2009. ISBN
978-1-61738-587-2. Online at http://arxiv.org/abs/0906.0049.
38
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[79] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Reasoning about a simulated
printer case investigation with Forensic Lucid. In Joseph S. Gauthier, editor, Proceedings
of the Huntsville Simulation Conference (HSC’09), page 45. SCS, October 2009. ISBN
978-1-61738-587-2. Abstract, fully online at http://arxiv.org/abs/0906.5181.
[80] Serguei A. Mokhov, Joey Paquet, and Xin Tong. A type system for hybrid intensionalimperative programming support in GIPSY. In Proceedings of C3S2E’09, pages 101–107,
New York, NY, USA, May 2009. ACM. ISBN 978-1-60558-401-0. doi: 10.1145/1557626.
1557642.
[81] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. The need to support of data flow
graph visualization of Forensic Lucid programs, forensic evidence, and their evaluation by
GIPSY. [online], September 2010. Poster at VizSec’10; online at http://arxiv.org/abs/
1009.5423.
[82] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Towards automatic deduction and
event reconstruction using Forensic Lucid and probabilities to encode the IDS evidence. In
S. Jha, R. Sommer, and C. Kreibich, editors, Proceedings of RAID’10, LNCS 6307, pages
508–509. Springer, September 2010. doi: 10.1007/978-3-642-15512-3\ 36.
[83] Serguei A. Mokhov, Emil Vassev, Joey Paquet, and Mourad Debbabi. Towards a selfforensics property in the ASSL toolset. In Proceedings of C3S2E’10, pages 108–113. ACM,
May 2010. ISBN 978-1-60558-901-5. doi: 10.1145/1822327.1822342.
[84] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. On the need for data flow graph
visualization of Forensic Lucid programs and forensic evidence, and their evaluation by
GIPSY. In Proceedings of the Ninth Annual International Conference on Privacy, Security
and Trust (PST), 2011, pages 120–123. IEEE Computer Society, July 2011. ISBN 978-14577-0582-3. doi: 10.1109/PST.2011.5971973. Short paper; full version online at http:
//arxiv.org/abs/1009.5423.
[85] Serguei A. Mokhov, Joey Paquet, and Mourad Debbabi. Reasoning about a simulated
printer case investigation with Forensic Lucid. In Pavel Gladyshev and Marcus K. Rogers,
editors, Proceedings of ICDF2C’11, number 0088 in LNICST, pages 282–296. Springer,
October 2011. ISBN 978-3-642-35514-1. doi: 10.1007/978-3-642-35515-8\ 23. Submitted
in 2011, appeared in 2012; online at http://arxiv.org/abs/0906.5181.
[86] Dominic A. Orchard and Steve Matthews. Integrating lucid’s declarative dataflow
paradigm into object-orientation. Mathematics in Computer Science, 2(1):103–122, 2008.
ISSN 1661-8270. doi: 10.1007/s11786-008-0046-6.
[87] Mehmet A. Orgun and Edward A. Ashcroft, editors. Proceedings of ISLIP’95, volume
Intensional Programming I, May 1995. World Scientific. ISBN: 981-02-2400-1.
[88] Mehmet A. Orgun and Weichang Du. Multi-dimensional logic programming: Theoretical
foundations. Theoretical Computer Science, 185(2):319–345, 1997. ISSN 0304-3975. doi:
10.1016/S0304-3975(97)00048-0.
[89] Mehmet A. Orgun, Chuchang Liu, and Abhaya C. Nayak. Knowledge representation,
reasoning and integration using temporal logic with clocks. Mathematics in Computer
Science, 2(1):143–163, 2008. ISSN 1661-8270. doi: 10.1007/s11786-008-0048-4.
39
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[90] C. B. Ostrum. The Luthid 1.0 Manual. Department of Computer Science, University of
Waterloo, Ontario, Canada, 1981.
[91] Themis Panayiotopoulos. Temporal reasoning with TRL. In Gergatsoulis and Rondogiannis [37], pages 133–148.
[92] Themis Panayiyotopoulos and L. C. Baxevanaki. Statistical queries on historical relational
databases. In Gergatsoulis and Rondogiannis [37], pages 229–243.
[93] Nikolaos S. Papaspyrou and Ioannis T. Kassios. GLU# embedded in C++: a marriage
between multidimensional and object-oriented programming. Softw., Pract. Exper., 34(7):
609–630, 2004. ISSN 0038-0644. doi: 10.1002/spe.582.
[94] Joey Paquet. Relational databases as multidimensional dataflows. Master’s thesis, Departement d’Informatique, Université Laval, Québec, Canada, 1995.
[95] Joey Paquet. Scientific Intensional Programming. PhD thesis, Department of Computer
Science, Laval University, Sainte-Foy, Canada, 1999.
[96] Joey Paquet. Distributed eductive execution of hybrid intensional programs. In Proceedings
of the 33rd Annual IEEE International Computer Software and Applications Conference
(COMPSAC’09), pages 218–224, Seattle, Washington, USA, July 2009. IEEE Computer
Society. ISBN 978-0-7695-3726-9.
[97] Joey Paquet and Peter Kropf. The GIPSY architecture. In Proceedings of Distributed
Computing on the Web, Quebec City, Canada, 2000.
[98] Joey Paquet and John Plaice. On the design of an indexical query language. In Proceedings
of the Seventh International Symposium on Lucid and Intensional Programming, pages
28–36, 1994.
[99] Joey Paquet and John Plaice. The intensional relation. In Orgun and Ashcroft [87], pages
214–227.
[100] Joey Paquet and John Plaice. The semantics of dimensions as values. In Gergatsoulis and
Rondogiannis [37], pages 259–273.
[101] Joey Paquet and Ai Hua Wu. GIPSY – a platform for the investigation on intensional
programming languages. In Proceedings of the 2005 International Conference on Programming Languages and Compilers (PLC 2005), pages 8–14. CSREA Press, June 2005.
[102] Joey Paquet, Serguei A. Mokhov, Emil I. Vassev, Xin Tong, Yi Ji, Amir H. Pourteymour, Kaiyu Wan, Aihua Wu, Sleiman Rabah, Bin Han, Bo Lu, Lei Tao, Yimin
Ding, Chun Lei Ren, and The GIPSY Research and Development Group. The General Intensional Programming System (GIPSY) project. Department of Computer Science and Software Engineering, Concordia University, Montreal, Canada, 2002–2013.
http://newton.cs.concordia.ca/~gipsy/, last viewed April 2012.
[103] Joey Paquet, Aihua Wu, and Peter Grogono. Towards a framework for the General
Intensional Programming Compiler in the GIPSY. In Proceedings of the 19th Annual
ACM Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA 2004), pages 164–165, New York, NY, USA, October 2004. ACM. doi:
10.1145/1028664.1028731.
40
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[104] Joey Paquet, Serguei A. Mokhov, and Xin Tong. Design and implementation of context
calculus in the GIPSY environment. In Proceedings of the 32nd Annual IEEE International
Computer Software and Applications Conference (COMPSAC), pages 1278–1283, Turku,
Finland, July 2008. IEEE Computer Society. doi: 10.1109/COMPSAC.2008.200.
[105] Manuel Peralta, Supratik Mukhopadhyay, and Ramesh Bharadwaj. Automatic synthesis and deployment of intensional kahn process networks. In Dominik Ślȩzak, Tai hoon
Kim, Stephen S. Yau, Osvaldo Gervasi, and Byeong-Ho Kang, editors, Grid and Distributed Computing, volume 63 of Communications in Computer and Information Science, pages 73–87. Springer Berlin Heidelberg, 2009. ISBN 978-3-642-10548-7. doi:
10.1007/978-3-642-10549-4\ 10.
[106] John Plaice. Particle in-cell simulation with Lucid. In Orgun and Ashcroft [87], pages
149–161.
[107] John Plaice. Cartesian programming. Technical Report UNSW-CSE-TR-1101, University
of Grenoble, France, January 2011. Habilitation Thesis, online at ftp://ftp.cse.unsw.
edu.au/pub/doc/papers/UNSW/1101.pdf.
[108] John Plaice and Peter G. Kropf. Intensional communities. In Gergatsoulis and Rondogiannis [37], pages 292–295.
[109] John Plaice and William W. Wadge. A new approach to version control. IEEE Transactions on Software, 19(3):268–276, March 1993.
[110] John Plaice, Ridha Khedri, and Ren’e Lalement. From abstract time to real time. In In
Proceedings of the Sixth International Symposium on Lucid and Intensional Programming,
pages 83–93, 1993.
[111] John Plaice, Blanca Mancilla, and Gabriel Ditu. From Lucid to TransLucid: Iteration,
dataflow, intensional and Cartesian programming. Mathematics in Computer Science, 2
(1):37–61, 2008. ISSN 1661-8270. doi: 10.1007/s11786-008-0043-9.
[112] John Plaice, Blanca Mancilla, Gabriel Ditu, and William W. Wadge. Sequential demanddriven evaluation of eager TransLucid. In Proceedings of the 32nd Annual IEEE International Computer Software and Applications Conference (COMPSAC), pages 1266–1271,
Turku, Finland, July 2008. IEEE Computer Society. doi: 10.1109/COMPSAC.2008.191.
[113] Amir Hossein Pourteymour. Comparative study of Demand Migration Framework implementation using JMS and Jini. Master’s thesis, Department of Computer Science and
Software Engineering, Concordia University, Montreal, Canada, September 2008.
[114] Amir Hossein Pourteymour, Emil Vassev, and Joey Paquet. Towards a new demanddriven message-oriented middleware in GIPSY. In Proceedings of PDPTA 2007, pages
91–97. PDPTA, CSREA Press, June 2007.
[115] Amir Hossein Pourteymour, Emil Vassev, and Joey Paquet. Design and implementation
of demand migration systems in GIPSY. In Proceedings of PDPTA 2009. CSREA Press,
June 2008.
[116] Sleiman Rabah, Serguei A. Mokhov, and Joey Paquet. An interactive graph-based automation assistant: A case study to manage the GIPSY’s distributed multi-tier run-time
41
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
system. In Proceedings of the ACM Research in Adaptive and Convergent Systems (RACS
2013), pages 387–394, New York, NY, USA, October 2011–2013. ACM. ISBN 978-1-45032348-2. Pre-print: http://arxiv.org/abs/1212.4123.
[117] Alessandra Raffaetà and Thom Frühwirth. Two semantics for temporal annotated constraint logic. In Gergatsoulis and Rondogiannis [37], pages 78–92.
[118] Toby Rahilly and John Plaice. A multithreaded implementation for TransLucid. In Proceedings of the 32nd Annual IEEE International Computer Software and Applications
Conference (COMPSAC), pages 1272–1277, Turku, Finland, July 2008. IEEE Computer
Society. doi: 10.1109/COMPSAC.2008.191.
[119] Anand Ranganathan and Roy H. Campbell. A middleware for context-aware agents
in ubiquitous computing environments. In Markus Endler and Douglas Schmidt, editors, Proceedings of Middleware 2003, volume 2672 of Lecture Notes in Computer Science, pages 143–161. Springer Berlin Heidelberg, 2003. ISBN 978-3-540-40317-3. doi:
10.1007/3-540-44892-6\ 8.
[120] Chun Lei Ren. General intensional programming compiler (GIPC) in the GIPSY. Master’s
thesis, Department of Computer Science and Software Engineering, Concordia University,
Montreal, Canada, 2002.
[121] Panagiotis Rondogiannis. Higher-Order Functional Languages and Intensional Logic. PhD
thesis, Department of Computer Science, University of Victoria, Victoria, Canada, 1994.
[122] Panagiotis Rondogiannis. Adding multidimensionality to procedural programming languages. Software: Practice and Experience, 29(13):1201–1221, 1999. ISSN 1097-024X.
doi: 10.1002/(SICI)1097-024X(199911)29:13h1201::AID-SPE278i3.0.CO;2-0.
[123] Panos Rondogiannis. Adding multidimensionality to procedural programming languages.
In Gergatsoulis and Rondogiannis [37], pages 274–291.
[124] Panos Rondogiannis and William W. Wadge. Extending the intensionalization algorithm
to a broader class of higher-order programs. In Orgun and Ashcroft [87], pages 228–233.
[125] M. C. Schraefel, Blanca Mancilla, and John Plaice. Intensional hypertext. In Gergatsoulis
and Rondogiannis [37], pages 40–54.
[126] Paul Swoboda. A Formalisation and Implementation of Distributed Intensional Programming. PhD thesis, The University of New South Wales, Sydney, Australia, 2004.
[127] Paul Swoboda and John Plaice. An active functional intensional database. In F. Galindo,
editor, Advances in Pervasive Computing, pages 56–65. Springer, 2004. LNCS 3180.
[128] Paul Swoboda and John Plaice. A new approach to distributed context-aware computing.
In A. Ferscha, H. Hoertner, and G. Kotsis, editors, Advances in Pervasive Computing.
Austrian Computer Society, 2004. ISBN 3-85403-176-9.
[129] Paul Swoboda and William W. Wadge. Vmake and ISE general tools for the intensionalization of software systems. In Gergatsoulis and Rondogiannis [37], pages 310–320. ISBN:
981-02-4095-3.
42
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[130] Lei Tao. Warehouse and garbage collection in the GIPSY environment. Master’s thesis, Department of Computer Science and Software Engineering, Concordia University,
Montreal, Canada, 2004.
[131] S. Tao. TLucid and Intensional Attribute Grammars. PhD thesis, Department of Computer
Science, Victoria University, Canada, 1994.
[132] Xin Tong. Design and implementation of context calculus in the GIPSY. Master’s thesis, Department of Computer Science and Software Engineering, Concordia University,
Montreal, Canada, April 2008.
[133] Xin Tong, Joey Paquet, and Serguei A. Mokhov. Complete context calculus design and
implementation in GIPSY. [online], 2007–2008. http://arxiv.org/abs/1002.4392.
[134] J. van Benthem. A Manual of Intensional Logic. CSLI Publications, Stanford and The
University of Chicago Press, 1988. ISBN 0-937073-29-6.
[135] Emil Vassev and Joey Paquet. A general architecture for demand migration in a demanddriven execution engine in a heterogeneous and distributed environment. In Proceedings
of the 3rd Annual Communication Networks and Services Research Conference (CNSR
2005), pages 176–182. IEEE Computer Society, May 2005. doi: 10.1109/CNSR.2005.9.
[136] Emil Vassev and Joey Paquet. A generic framework for migrating demands in the GIPSY’s
demand-driven execution engine. In Proceedings of the 2005 International Conference on
Programming Languages and Compilers (PLC 2005), pages 29–35. CSREA Press, June
2005.
[137] Emil Vassev and Joey Paquet. Towards autonomic GIPSY. In Proceedings of the Fifth
IEEE Workshop on Engineering of Autonomic and Autonomous Systems (EASE 2008),
pages 25–34. IEEE Computer Society, 2008. ISBN 978-0-7695-3140-3. doi: 10.1109/EASe.
2008.9.
[138] Emil Iordanov Vassev. General architecture for demand migration in the GIPSY demanddriven execution engine. Master’s thesis, Department of Computer Science and Software
Engineering, Concordia University, Montreal, Canada, June 2005. ISBN 0494102969.
[139] Lorenzo Verdoscia. ALFA fine grain dataflow machine. In Orgun and Ashcroft [87], pages
110–134.
[140] William W. Wadge. Possible WOOrlds. In Orgun and Ashcroft [87], pages 56–62. Invited
Contribution.
[141] William W. Wadge. Intensional logic in context. In Gergatsoulis and Rondogiannis [37],
pages 1–13. Tutorial.
[142] William W. Wadge. Hammings problem example. [online], December 2003. http://i.
csc.uvic.ca/home/hei/lup/contents.html.
[143] William W. Wadge and Edward A. Ashcroft. Lucid, the Dataflow Programming Language.
Academic Press, London, 1985.
[144] William W. Wadge and M. C. Schraefel. Putting the hyper back in hypertext. In Gergatsoulis and Rondogiannis [37], pages 31–39.
43
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
SIGLUCID
[145] William W. Wadge and Alan Yoder. The Possible-World Wide Web. In Orgun and
Ashcroft [87], pages 207–213.
[146] William W. Wadge, G. Brown, M. C. Schraefel, and T. Yildirim. Intensional HTML. In
4th International Workshop PODDP’98, March 1998.
[147] Kaiyu Wan. Lucx: Lucid Enriched with Context. PhD thesis, Department of Computer
Science and Software Engineering, Concordia University, Montreal, Canada, 2006.
[148] Kaiyu Wan, Vasu Alagar, and Joey Paquet. A context theory for intensional programming.
In Workshop on Context Representation and Reasoning (CRR05), July 2005.
[149] Kaiyu Wan, Vasu Alagar, and Joey Paquet. Lucx: Lucid enriched with context. In Proceedings of the 2005 International Conference on Programming Languages and Compilers
(PLC 2005), pages 48–14. CSREA Press, June 2005.
[150] Kaiyu Wan, Vasu Alagar, and Joey Paquet. Real time reactive programming in lucid enriched with contexts. In Zhiming Liu and Keijiro Araki, editors, Theoretical Aspects of Computing - ICTAC 2004, volume 3407 of Lecture Notes in Computer Science, pages 387–402. Springer Berlin Heidelberg, 2005. ISBN 978-3-540-25304-4. doi:
10.1007/978-3-540-31862-0\ 28.
[151] Ai Hua Wu. Semantic checking and translation in the GIPSY. Master’s thesis, Department
of Computer Science and Software Engineering, Concordia University, Montreal, Canada,
2002.
[152] Ai Hua Wu. OO-IP Hybrid Language Design and a Framework Approach to the GIPC. PhD
thesis, Department of Computer Science and Software Engineering, Concordia University,
Montreal, Canada, 2009.
[153] Ai Hua Wu and Joey Paquet. Object-oriented intensional programming in the GIPSY:
Preliminary investigations. In Proceedings of the 2005 International Conference on Programming Languages and Compilers (PLC 2005), pages 43–47. CSREA Press, June 2005.
[154] Ai Hua Wu, Joey Paquet, and Peter Grogono. Design of a compiler framework in the
GIPSY system. In Proceedings of the 15th IASTED International Conference on Parallel and Distributed Computing and Systems (PDCS 2003), volume 1, pages 320–328.
International Association of Science and Technology for Development, November 2003.
[155] Aihua Wu, Joey Paquet, and Serguei A. Mokhov. Object-oriented intensional programming: Intensional Java/Lucid classes. In Proceedings of SERA 2010, pages 158–167. IEEE
Computer Society, 2010. ISBN 978-0-7695-4075-7. doi: 10.1109/SERA.2010.29. Online
at: http://arxiv.org/abs/0909.0764.
[156] Satoshi Yamane. Real-time object-oriented specification and verification. In Orgun and
Ashcroft [87], pages 162–185.
[157] Yu Zhang and Kang Zhang. Associative query for multi-version web documents. In
Gergatsoulis and Rondogiannis [37], pages 55–64.
[158] Q. Zhao. Implementation of an object-oriented intensional programming system. Master’s
thesis, Department of Computer Science, University of New Brunswick, Canada, 1997.
44
Index
API
[], 31
bool, 28, 31
bool::true, 31
boolean, 28, 31
byte, 31
char, 31
CodeSegment, 25
Dimension, 31, 32
dimension, 28, 31
double, 28, 31
Executor, 31
Float, 31
float, 28, 31
GIPC, 25
GIPSYArray, 31
GIPSYBoolean, 31
GIPSYCharacter, 31
GIPSYContext, 32
GIPSYDouble, 31
GIPSYEmbed, 31
GIPSYFloat, 31
GIPSYFunction, 31
GIPSYIdentifier, 31
GIPSYInteger, 31
GIPSYObject, 31
GIPSYOperator, 31
GIPSYString, 31
GIPSYVoid, 28, 31
int, 28, 31
java.lang, 28
Long, 31
long, 28, 31
Method, 31
Object, 31
Preprocessor, 25, 31
SemanticAnalyzer, 31
SequentialThreadGenerator, 31
special, 22
String, 28, 31
string, 28, 31
StringBuffer, 28
toString(), 31
true, 28
type, 22
void, 28, 31
AST, 5
C, 5
C++, 6, 26, 28
data types
matching Lucid and Java, 28
Forensic Lucid, 2, 4, 21
Fortran, 5, 26
Frameworks
GEE, 5
GICF, 28
GIPC, 5
GIPSY Type System, 28
GEE, 5
GICF, 28
GIPC, 5
Preprocessor, 25
GIPL, 1, 3–5, 19, 21, 22
GIPSY, 5, 6, 25, 28
Type System, 28
Types, 28
GIPSY Program
Segments, 25
GIPSY Type System, 28
GLU, 4–6
GLU#, 4, 6
Indexical Lucid, 4, 5
Java, 5, 25, 26, 28
JLucid, 4, 5, 28
JOOIP, 4, 28
Lucid, 1–6, 28
Family, 5
History, 5
Pipelined Dataflows, 5
Lucx, 2–5, 28
Lustre, 4
MARFL, 2–4, 21
Onyx, 4, 6
45
Core Lucid Standard Specification: SIGLUCID’s Contextual Design
Partial Lucid, 4, 5
Perl, 26
Preprocessor, 25
GIPC, 25
Python, 26
Segments
#<IMPERATIVELANG>, 26
#<INTENSIONALLANG>, 26
#CPP, 26
#FORENSICLUCID, 28
#FORTRAN, 26
#GIPL, 26
#INDEXICALLUCID, 26
#JAVA, 26
#JLUCID, 28
#JOOIP, 26
#LUCX, 26
#OBJECTIVELUCID, 28
#ONYX, 28
#PERL, 26
#PYTHON, 26
#TENSORLUCID, 28
#TRANSLUCID, 28
#funcdecl, 25, 32
#typedecl, 26, 32
Tensor Lucid, 4, 5, 27, 28
Tools
libintense, 6
TransLucid, 1, 4, 6, 19, 21, 22, 25
Types, 28
46
SIGLUCID
| 6cs.PL
|
arXiv:cs/0307053v1 [cs.CE] 23 Jul 2003
Preprint typeset in JHEP style - HYPER VERSION
Hamevol1.0: a C++ code for differential equations based
on Runge-Kutta algorithm. An application to matter
enhanced neutrino oscillation
P. Aliania,b , V. Antonellib,c , M. Picariellob,c , Emilio Torrente-Lujand,e
a
b
c
d
e
Dept. de Physique, Université Libre de Bruxelles, Bruxelles, Belgium
Dip. di Fisica, Università degli Studi di Milano, Milano, Italy
I.N.F.N., Sezione di Milano, Milano, Italy
Departamento de Fisica, Grupo de Fisica Teorica, Universidad de Murcia, Murcia, Spain
CERN-TH, CH-1211 Geneve 23, Switzerland
Abstract: We present a C++ implementation of a fifth order semi-implicit Runge-Kutta algorithm for solving Ordinary Differential Equations. This algorithm can be used for studying
many different problems and in particular it can be applied for computing the evolution of any
system whose Hamiltonian is known. We consider in particular the problem of calculating the
neutrino oscillation probabilities in presence of matter interactions. The time performance and
the accuracy of this implementation is competitive with respect to the other analytical and numerical techniques used in literature. The algorithm design and the salient features of the code
are presented and discussed and some explicit examples of code application are given.
e-mail addresses: [email protected]; [email protected]; [email protected];
[email protected]
Submitted to Comput. Phys. Commun.
Contents
1. Program Summary
1
2. Introduction. The mathematical problem.
2
3. The Physical motivations.
3
4. Code Structure
4.1 Algorithmics
4.2 The distribution
4.3 The RK algorithm. Main Subroutines
4.4 Sample Program and Inputs
6
6
7
8
10
5. Conclusions
10
6. Acknowledgments
11
7. Appendices
7.1 Using Hamevol
7.2 Example of Hamiltonian definition
7.3 Sample outputs
12
12
14
15
1. Program Summary
• Title of the program: Hamevol
• Version number: 1.0
• Avaible at: http://wwwteor.mi.infn.it/ antonell/programs/RKutta , and mirrors
• Programming language: C++
• Platform: any platform supporting a C++ compiler (examples: Linux, Unix, Windows)
• Tested on:Pentium PC, AMD PC
• Memory requirements for execution: Standard application: 40 Kbytes
• No. of bytes in distributed program, including test data, etc: 235.000
• Keywords: Numerical algorithms, Differential equations, Hamiltonian evolution, Oscillation,
• Nature of physical problem: Numerical solution of Hamiltonian differential equations. Application to the numerical calculation of the oscillation probability for a quantum system
(like, for instance, neutrinos of any kind propagating in a medium)
• Method of solution: Algorithm based on fifth order semi-implicit Runge-Kutta method
• Typical running time: ≃ 30 seconds for every single point on a Penthium IV PC
2. Introduction. The mathematical problem.
The use of numerical algorithms and suitable computational techniques has often been very
useful to find the solution of difficult problems which are of interest for mathematics and other
applied science. This is particularly true in our days, when the relationships between information
technology and other sciences are becoming closer and closer.
In this paper we discuss the adaptation of well known numerical techniques to a general
class of problems which are described by ordinary differential equations and we present some
examples taken from physics.
The numerical code and algorithm we are presenting in this work is based on the implementation of the Runge-Kutta method and it can find significant applications in the study of
different physical systems. In fact the evolution of every system can be fully described once
we are able to solve the differential equations that drive this evolution. The typical example
is the case in which one wants to study a linear quantum system that is described by a vector
X ≡ (Xi , i = 1..N ), where Xi are the elements of an appropriate basis decribing the system.
The system of linear differential equations we are interested in can be put in the simple form:
i
dX
= H(t)X,
dt
(2.1)
where H is a matrix determining the evolution of the system.
In the language of physics H is the Hamiltonian of the system and Equation (2.1) is the
corresponding Schroedinger equation. It is clear, however, that the system of differential equations (2.1) is very general and they can describe also problems of different nature in fields that
are completely different from physics. Depending on the kind of problems one has to solve,
the requirements which are fundamental for the solutions can be different. This can suggest
the choice of a particular algorithm in order to fulfill these requirements in the best possible
way. For instance, as we are going to discuss later, in the physical problems we are intersted in,
the delicate point is efficiency more than accuracy and this justifies the choiche of a particular
version of adaptive Runge-Kutta method that, if properly adapted to our purpose, enable us to
obtain satisfactory results.
The solution of equation (2.1) can be written in terms of the fundamental system of solutions,
or equivalently called evolution operator U (t, t0 ), defined by the expression:
U (t, t0 ) = exp [−iH (t − t0 )],
(2.2)
where t0 is the initial time at which we know the state of the system. The simple formula given
above is valid for a time independent Hamiltonian. In the case in which the Hamiltonian of the
system is changing in the time (like, for instance, in presence of matter effects), formula (2.2)
must be replaced by path-ordered exponential
Z t
U (t, t0 ) = P exp −i
dτ H[τ ] .
(2.3)
t0
The code we are presenting here can be used to solve in an iterative way the system of
differential equations appearing in Equation (2.1). This is particularly important in the case of
Hamiltonians which are explicitly time dependent or which contain terms fastly oscillating in
time.
The structure of this paper is the following. We start discussing in section 3 some concrete
examples, taken from physics, of situations in which the application of the algorithm presented
here is particularly suited. In the next section we present the algoritmic structure and salient
aspects of the code we developed. This section includes a presentation of the algorithm, all the
essential informations about the distribution, the main subroutines and functions and about the
way of working of a sample program. In section 5 we draw our final conclusions. An example of
a sample program, with some specific Hamiltonian set up and the relative outputs are presented
in the Appendices.
3. The Physical motivations.
A very interesting example is given by the study of neutrino physics, which has been one of the
central topics of elementary particle physics in the last years. A detailed discussion about the
main properties of neutrinos and the relevance of their study for our knowledge of the intimate
structure of matter is beyond the scope of this work. Therefore we refer the interested reader
to the many reviews one can find in literature [1]. Here we just recall that during the last years
the answer has been given to the central question of neutrino physics, which puzzled physicists
for more than seventy years, that is to discriminate whether this particle is massive or massless.
We know by now that neutrinos are massive and oscillating particles and the proof of this
has been given by the important results obtained mainly in the last decade (and especially in
the very last years) by the experiments looking for oscillation signals of neutrinos from different
sources: solar [2, 3], reactor [4] and atmospheric [5] neutrinos.
The great relevance of these results is confirmed by the fact that this is up to now the strongest
indication of oscillation we have in the leptonic sector and it is impossible to accomodate it in the
usual “minimal version” of the Standard Model (the theory describing very well the electroweak
interactions of elementary particles). One can say that neutrino oscillations is a hint for physical
phenomena beyond to what is presently known.
All these experimental evidences in favors of the oscillation hypothesis have proved that the
flavor eigenstates of neutrinos, that is the ones entering in the weak processes, are in fact quantum superspositions of different mass eigenstates (at least three different mass eigenstates are
needed to explain the full set of experimental data, if one doesn’t take into account the controversial results of the LSND experiment). During the evolution the composition of this quantum
system can change giving rise to the oscillation phenomenon detected in the experiments.
Any study of neutrino oscillation is necessarily based on the calculation of the so called
neutrino survival (or transition) probability. This is the probability that a neutrino emitted by
a source with a certain flavor, for instance an electronic neutrino emitted in the solar fusion
processes, remains with the same flavour (or is converted into a different flavor neutrino) before
reaching the detector.
Hence the basic quantity to compute is the survival probability in matter for a neutrino of
a certain flavor:
P (νi → νi ; t, Eν ) = |hνi (t0 ) |U (t, t0 ) |νi (t0 )i |2
(3.1)
where t0 is the inital time at which the neutrino is assumed to be in the flavour eigenstate νi
(with i = e, µ, τ ) and U (t, t0 ) is the evolution operator given by Equation (2.3).
In absence of matter the basis in which the Hamiltonian is diagonal is simply the mass basis,
whose eigenstates να are connected to the neutrino flavor eigenstates νi by means of the mixing
matrix U :
X
νi =
Uiα να , i = e, µ, τ and α = 1, 2, 3 .
(3.2)
α
Using the same compact notation of Equation (2.1), we denote the set of the three neutrino
mass eigenstates with the vector ν, where ν ≡ (να , α = 1, 2, 3). This notation can be simply
extended to the case in which one has more than three neutrinos 1 . The Schroedinger equation
describing the evolution of the system in vacuum under the relativistic approximation and in
the hypothesis of equal momentum for different mass eigenstates is:
i
where
H 0 = DiagEα =
p
dν
= H 0ν
dt
p2 + m2 ≃ Eν +
(3.3)
m21 /2Eν
m22 /2Eν
m23 /2Eν
(3.4)
The last experimental data (both for atmospheric and solar neutrinos) have proved that to
describe neutrino evolution one has to take into account also the modification of the oscillation
pattern due to the very important effects of interaction with matter. This gives rise to the well
known MSW effect [6]. The problem of calculating neutrino oscillation probability in presence
of matter effects has been faced by many authors with different approaches, both numerical and
analytical, in the case of two neutrino flavors [7, 8, 9]. Exact solutions to the three neutrino
MSW equations were derived [10, 11] for simple matter densities. Numerical algorithms for
direct computation of the solar neutrino survival probability with all three active neutrinos have
been presented in [12, 13, 14] .
The purpose of our code Hamevol is to calculate the electron-neutrino survival probability
for a given neutrino energy, a given density profile, given neutrino masses and mixing matrix.
This set of (three) oscillation probabilities will be used subsequently by other (Fortran) programs
to calculate expected signals from diverse neutrino experiments. Due to the fact that both the
mixing matrix and the density function are considered input parameters of Hamevol , all kinds
of neutrino oscillation problems may be tackled, including those relevant to anti-neutrinos only.
In presence of standard matter with arbitrary electron number density, the propagation is
usually well described by the following system of differential equations:
1
This is the case if one introduces also sterile neutrinos in the analysis
dν
(3.5)
= H 0 + ρ(t)U V U † ν,
dt
where V is a matrix with V11 as only non-zero element, ρ(t), essentially a forward scattering
amplitude, is proportional to the electron number density of the medium Ne (t)
√
ρ(t) = ± 2GF Ne (t)
(3.6)
i
and U is the mixing matrix connecting the neutrino flavor eigenstates with the mass eigenstates.
In the case of three neutrino generations, adopting the Particle Data Group [15] convention for
the mixing matrix, one gets:
νe
c1 c3
s1 c3
s3
ν1
(3.7)
νµ = −s1 c2 − c1 s3 s2 c1 c2 − s1 s3 s2 c3 s2 × ν2
ντ
s1 s2 − c1 s3 c2 −c1 s2 − s1 s3 c2 c3 c2
ν3
where ci ≡ cos θi and si ≡ sin θi and the three mixing angles θ1 = θ12 , θ2 = θ23 , and θ3 = θ13
roughly measure mixing between mass eigenstates (1-2), (2-3), and (1-3) respectively. We have
neglected the CP violating phase, which is irrelevant in this problem.
The plus/minus sign in formula of Equation (3.6) is for neutrinos/antineutrinos respectively.
The time dependence of the electron density Ne is a crucial factor involved in solving the
evolution equation.
The difference of the eigenvalues of H 0 ( the inverse of the usually defined oscillation length)
is typically considered to be, in the solar case, of the order
m2i − m2j
10−4 − 10−5 eV 2
≈
≈ 10−10 − 10−11 eV
2E
1 M eV
In [10] an analytical solution of the problem has been found for the case in which the electron
number density is parametrized (data taken from [16]) , for sufficiently far distances from the
solar core as:
Ne (r) = N0 exp(−λr) ; λ ≃ 10.6
r
r0
(3.8)
with r, r0 the distance from the center and solar radius respectively.
An important peculiarity of neutrino propagation in matter with respect to vacuum is that
for a certain value of the parameters a resonance appears at a certain point along the neutrino
trajectory. This resonance takes place if the following condition is satisfied:
ρ(tres ) =
∆m2 cos 2θ
,
E
(3.9)
where θ is the mixing angle between the 2 neutrino flavors taking part to the oscillation phenomenon. Apart from this approximate results, exact solutions have appeared for particular
forms of the function ρ: for linear densities, in terms of Weber-Hermite functions ([17, 18]); for
functions of the form ρ(t) = C(1 + tanh(λt)) in [19], and for exponentially decaying densities
ρ(t) = ce−λt in [7, 8].
The parametrization of Equation (3.8) is the one used also in the sample program we present
as an example in subsection 4.4.
In any case our numerical alghorithm enables us to find a solution of the problem for every
expression one chooses for the electron density number.
The capability of solving the system of Equations (3.5) and, therefore, of computing the
neutrino survival (or transition) probabilities as a function of the mixing parameters is an
essential ingredient in every analysis of neutrino data. The theoretical expected signal in every
experiment is obtained by convoluting neutrino fluxes, oscillation probabilities, neutrino cross
sections and detector energy response functions. The comparison of these expected signal as a
function of the mixing parameters with the experimental results is then usually performed by
means of a χ2 statistical analysis. The outcome of this kind of analyses is typically the production
of exclusion plots selecting the regions of the mixing parameter plan which are in agreement
with the data at a certain confidence level. The algorithm we are presenting in this work enables
us to numerically solve the neutrino evolution equations for all the oscillation parameter space,
without the need to introduce the approximation which are required in different approaches
based on the use of semi-analytical expressions in portions of the parameter space. For a more
detailed description of the full procedure we adopted for a phenomenological analysis of solar
and reactor neutrino data we refer the interested reader to [20].
The numerical alghoritm we are presenting here finds also other relevant applications. Some
significant examples are:
a)the study of the neutrino evolution inside stochastic media and neutrino propagation in solar
magnetic field [21]. In these cases one has to solve systems of differential equations where,
respectively, (2N )2 and (2N ) equations appear (N is the number of neutrino species);
b) solution of the renormalization group equation for the Minimal Supersymmetric Standard
Model in a Supergravity scenario. Here the number of differential equations in the systems that
have to be solved simultaneously (with a complicated mixture of initial and boundary conditions)
is typically between thirty and one hundred. [22]
4. Code Structure
4.1 Algorithmics
There are different ways of solving a possibly stiff set of equations and the advantages and
drawbacks of each of them must be evaluated keeping in mind the kind of problem one wants to
study. In our case we aim for efficiency rather than precision (a relative precision of 10−3 −10−5 in
conversion probabilities for example would be sufficient in a typical application in particle physics
propagation problems). Therefore we opted for an adaptive Runge-Kutta (RK) algorithm.
The Runge-Kutta method [23] is particularly suitable for solving differential equations,
starting from the knowledge of the function at the a fixed intial point X and advancing the
solution from X to X + h, by using the evaluation of the function at intermediate points inside
the interval h . By properly combining these evaluations one can reduce the error in the final
output. The method is conventionally denoted of order n if its error term is O hn+1
An essential characteristic of a good Ordinary Differential Equations integrator is the capability of having an adaptive control over its progress and a mechanism for adapting its stepsize,
in such a way to obtain the required accuracy with the minimal possible computational effort.
This property is possessed by the algorithm we implemented. Although an implicit RK would
be advised for stiff equations, there are several alternatives, such as the semi-implicit fifth-order
RK routine we have chosen. This routine requires the determination of the function at five
different points in the interval between the chosen steps. They are
k1 = hf (xn , yn )
k2 = hf (xn + a2 h, yn + b21 k1 )
...
(4.1)
k6 = hf (xn + a6 h, yn + b61 k1 + . . . b65 k5 )
6
X
ci ki
yn+1 = yn +
i=1
where the coefficients ai , bij and ci must satisfy certain constraints in order to ensure stability
and convergence. This algorithm is well suited for adaptive stepping, due to the fact that among
the six evaluations in Eq. (4.1) there is an embedded fourth-order combination, which, although
redundant, gives us an estimate of the error at each evaluation thereby allowing us to adjust the
step size. Table (1) gives a list of ai , bij and ci as determined by Cash and Karp [23].
i
ai
bij
1
4
1
5
3
10
3
5
5
1
6
7
8
2
3
1
5
3
40
3
10
11
− 54
1631
55296
9
40
9
− 10
5
2
175
512
6
5
− 70
27
575
13824
35
27
44275
110592
253
4096
1
2
3
4
5
j=
ci
c∗i
37
378
2825
27648
0
0
250
621
125
594
18575
48384
13525
55296
277
14336
1
4
0
512
1771
Table 1: Cash-Karp coefficients for our RK routines, taken from Ref.[23].
4.2 The distribution
The distribution of the program is contained in the tarred gzipped file hamevol1.0.tar.gz. In
any Linux or Unix system, unpacking and untarring this distribution file will produce a local
directory called Hamevol1.0 containing the following ascii files:
hamevol-rungekutta.hpp
hamevol-util.hpp
hamevol-sample.hpp
hamevol-sample.cpp
Makefile
The file with the main routines.
Some auxiliary utilities.
The sample program header.
The sample program.
A simple compiling make example.
In the same Linux or Unix system the execution of the command
./make
or directly the explicit call to the C++ compiler
g++ -O3 -o hamevol.x hamevol-sample.cpp
should produce an executable file from our sample program dedicated to solve some particular neutrino propagation problem. To run the executable the following command should be
typed
./hamevol.x OPTION
where in our sample program OPTION=1,0 depending on whether full Sun+Earth or only
Sun propagation is demanded. As a consequence, some brief ouput will appear on the standard
ouput, mainly information about parameter settings and options. In addition, as a result of
the execution our sample program writes an output file runge.out with the neutrino conversion
probabilities along the neutrino trajectory.
In our distribution, we have well separated the code corresponding to the general routines
implementing the Runge-Kutta algorithm from those corresponding to a particular application
(the “sample” files) of interest to us and that are presented here: the computation of oscillation
neutrino probabilities. Within our sample programs, it is also well differentiated the driver code
which calls the RK routines from the part where a concrete hamiltonian is built. In the most
basic case, a general user should be able to use our program as a black box for his own purposes
simply plugging his own definition for the hamiltonian.
In the following we will first describe the main routines included in file hamevol-rungekutta.hpp
and then those contained in the sample programs.
4.3 The RK algorithm. Main Subroutines
The kern code is built up with five main subroutines. They correspond to procedures and methods well known in the literature. We have improved and adapted them for our purposes. The
following classes are located in file hamevol-rungekutta.hpp. Here it follows a brief description of
any of them together with its calling sequence. For brevity, arguments which are also referenced
somewhere else are omitted here.
• void runge(CNumber y[], CNumber dydx[], CNumber (*H)(...), int n, Number x,
Number h, CNumber yout[], void (*derivs)(...))
Given the value y[1,..,n] of the vector state describing a phyisical system made up with n
components and evolving according to Schroedinger equation and knowing the Hamiltonian
H of the system, the subroutine produces the advanced solution as the function at the
incremented variables yout[1..n].
• void odeint(CNumber ystart[], int nvar, Number x1, Number x2, Number eps,
Number h1, Number hmin, CNumber (*H)(...), void (*derivs)(...), void (*rkqs)(...))
This is a Runge-Kutta driver with adaptive stepsize control. It Integrates the starting
values ystart[1..nvar] from x1 to x2 with accuracy eps, storing the intermediate results
in global variables.
A value h1 should be set as a guessed first stepsize, hmin as the minimum allowed stepsize
(it can be zero). On output nok and nbad are the number of good and bad (but retried and
fixed) steps taken, and ystart is replaced by values at the end of the integration interval.
• void rkqs(CNumber y[], CNumber dydx[], int n, Number *x, Number htry, Number eps,
CNumber yscal[], Number *hdid, Number *hnext, CNumber (*H)(...), void (*derivs)(...)).
This routine rkqs is implemented in order to perform an adaptive 5th order Runge-Kutta
integration. The method enables to have a monitoring of local truncation error, in order
to ensure the required accuracy and adjust the stepsize. The inputs are the independent
variable vector y[1..n] and its derivative dydx[1..n] at the starting value of the independent variable x. Other inputs are the stepsize htry, the required accuracy eps, and the
vector yscal[1..n] against which the error is scaled. On output, y and x are replaced
by their new values, hdid is the stepsize that was actually accomplished, and hnext is the
estimated next stepsize.
• void rkck(CNumber y[], CNumber dydx[], int n, Number x, Number h, CNumber yout[],
CNumber yerr[], CNumber (*H)(...), void (*derivs)(...))
Used in adaptive size Runge-Kutta integration. Given values for the variables y[1..n]
and their derivatives dydx[1..n] known at x, advance solution over an interval h and
return the incremented variables as yout[1..n]. Also returns an estimate of the local
truncation error in yout using embedded fourth-order method.
The user supplies the routine H(x,i,j), which returns the element (i,j) of the hamiltonian of the evolution.
• void deriv(Number x, CNumber y[], CNumber dy[], int n, CNumber (*H)(...)).
The user-supplied routine derivs is used for calculating the right-hand side derivative.
The user supplies the routine derivs(x,y,dydx,n,H), which returns the derivatives dydx of
the many variable function y with respect to the vector x at the point x.
The following Auxiliary functions which will allocate the following data structures are included in the auxiliary distribution file hamevol-util.hpp.
CNumber *Cvector(long nh):
Number *vector(long nh):
int *ivector(long nh):
unsigned long *lvector(long nh):
unsigned char *cvector(long nh):
template <class Type> Type **matrix(...)
allocate a CNumber vector with subscript range v[1..nh]
a vector with subscript range v[1..nh].
a vector with subscript range v[1..nh].
a vector with subscript range v[1..nh].
a vector with subscript range v[1..nh].
allocate a Type matrix with a subscript range.
4.4 Sample Program and Inputs
The void main routine in hamevol-sample.cpp file takes the values of the neutrino wave functions
at initial starting points for physical initial conditions which are standard for solar neutrino
physics (νe (0) = 1, νµ (0) = 0, ντ (0) = 0) and calculates the final wave function and corresponding
probabilities at the target final points. The algorithm includes the following steps:
• It takes from the command line the user argument “1” (Sun) or “0” (Earth) propagation.
• It performs argument validation, set internal flags and writes to the standard output a list
of current values of diverse parameters.
• It declares the output file stream out ‘‘runge.dat’’, class ofstream included within
<fstream.h>.
• It declares the Cvector objects nu,dnu, respectively instances of the neutrino wave function
and their vector derivative. It performs diverse other initializations.
• Finally functions odeint and evolve are repeteadly called until the desired final point or
the maximum number of steps is reached.
Different parameters, for example the number of equations, or in this physical case the
number of neutrino species (N = 2, 3), have to be set in the header file hamevol-sample.hpp .
We run the RK algorithm to obtain the transition probabilites of neutrinos produced at the sun
center with a given energy and mixing angle as a function to its position along the trajectory
sun-earth. The user should provide routines for computing the electron densities at Sun and the
Earth along the neutrino trajectory. They appear in the matter part of the neutrino hamiltonian.
We include examples of the main program and other smaller routines as appendices.
5. Conclusions
A new code based on a semi-implicit fifth order adaptive Runge-Kutta algorithm has been
developed by us. It can be used as solver for many systems of differential equations, like, for
instance, the ones that usually describe the evolution of a system in physics and in other fields.
This algorith is particularly suited for the solution of differential equations in which the operator
driving the evolution of the system is changing in time.
Here we focus our attention to the application of this code to the study of physical problems,
like solving the Schroedinger equation for a system that is a quantum superposition of different
possible states. The explicit example we present is the study of the evolution and calculation
of transition probabilty for neutrinos emitted by a source and travelling in a medium. This
code has been already applied by us as a useful tool to obtain a check with respect to other
possible numerical algorithms (like the ones based on the evolution operator formalism) in
our phenomenological analysis of different neutrino oscillation experiments. This analysis has
confirmed the validity of neutrino oscillation hypothesys and enabled us to determine the allowed
region for mixing parameters, a topic of great relevance in Elementary Particle Physics.
In this paper we discuss the structure of the algorithm we developed and the main features
of our code. We also present a sample program and give some typical outputs, as a concrete
example of application of our algorithm.
6. Acknowledgments
We are really grateful to R. Ferrari for many useful discussions and suggestions and for the
continuos support given to our work. One of us (V. A.) would like to thank S. Petcov for very
useful discussions about the analytical study of the neutrino progragation in matter. One of us
(E.T) would like to thank the hospitality of CERN-TH division, the department of physics of
University of Milan, and financial support of INFN-CICYT grant. The computations presented
here have been all performed on the computer farm of the Theoretical group of Milano University.
7. Appendices
7.1 Using Hamevol
Here we present the main sample program, corresponding to the file hamevol-sample.cpp, where
we show explicitly the use of the main routines.
#include "hamevol.hpp"
#include <string.h>
/* Inizialize the vacuum values
/* #define parameters
(.......)*/
(.......)*/
Number Var;
CNumber **HH0;
// is the vacuum Hamiltonian in the flavour eingenstates
struct mix{
// contains the mixing matrix
Number mass[N + 1];
CNumber U[N + 1][N + 1];
} mixing;
/* nu are the wave functions, dnu their derivative */
CNumber *nu, *dnu;
int main(int arg, char** argv){
Number x1=0.;
Number x2;
/*
/*
Argument and parameter validation (............) */
Information output
(.............)*/
time_t ti, tf;
CNumber *onu;
nu = Cvector (N);
dnu = Cvector (N);
onu = Cvector (N);
ofstream out ("runge.dat"); /* save data in runge.dat */
srand (time (&ti));
time (&ti);
Var = VarI;
Number VarO = Var;
Number h = (VarF - VarI) / INIT_STEPS;
Number eps = Eps_Error;
vacuum_values ();
/* Main Routines */
odeint (nu, N, x1, x2, eps, dist, dist_min, H, deriv, rkqs);
evolute (nu, &out, VarO);
for (int nstp = 1; nstp <= MAX_STEPS; nstp++)
{ /* Take at most MAXSTP steps */
for (int i = 1; i <= N; i++)
onu[i] = nu[i];
vacuum_values ();
if ((VarO + h - VarF) * (VarF - VarI) > 0.0) /* Are we done? */
h = VarF - VarO;
Var = VarO + h;
odeint (nu, N, x1, x2, eps, dist, 0.0, H, deriv, rkqs);
if (distance (nu, onu, N) > Prob_Error)
{
//cout << "DECREASE: h=" << h << endl;
h *= DECREASE;
Number htemp = ((h < 0) ?
FMIN (h, -abs ((VarI - VarF) / MAX_STEPS)) :
FMAX (h, abs ((VarI - VarF) / MAX_STEPS)));
if (htemp != h)
{
evolute (nu, &out, Var);
VarO = Var;
}
h = htemp;
}
else
{
evolute (nu, &out, Var);
h *= INCREASE;
Number htemp = ((h < 0) ?
FMAX (h, -abs ((VarI - VarF) / MIN_STEPS)) :
FMIN (h, abs ((VarI - VarF) / MIN_STEPS)));
if (htemp != h)
h = htemp;
VarO = Var;
}
if ((Var - VarF) * (VarF - VarI) > 0.0)
{ /* Are we done? */
cout << "t=" << time (&tf) - ti << endl;
return 0; /* normal exit */
}
}
cout << "Too many steps in routine evolution_matter!" << endl;
return -1;
}
7.2 Example of Hamiltonian definition
Here is the hamiltonian we use in our sample program:
/************************************************************************/
/*
The Hamiltonian
*/
/************************************************************************/
inline CNumber V(int i, int j){
return (i == j == 1) ? 1. : 0.;
}
inline CNumber U(int i, int j){
return mixing.U[i][j];
}
CNumber H(Number r, int i, int j){
CNumber HH = 0;
return HH0[i][j] + V (i, j) * rho (r) * sqrt (2.) * Gf;
}
void vacuum_values(){
mixing.mass[1] = 1.e-2;
mixing.mass[2] = 1.e-1;
Number th12 = M_PI / 3.;
Number th13 = M_PI / 3.;
Number th23 = M_PI / 3.;
nu[1] = 1.;
nu[2] = 0.;
CNumber H0[N + 1][N + 1];
HH0 = matrix
Number sth12
Number cth12
Number sth13
Number cth13
Number sth23
Number cth23
/*
((CNumber) 1, N, 1, N);
= sin (th12);
= cos (th12);
= sin (th13);
= cos (th13);
= sin (th23);
= cos (th23);
Three neutrinos
(mixing.U)[1][1]
(mixing.U)[1][2]
(mixing.U)[1][3]
(mixing.U)[2][1]
=
=
=
=
*/
CNumber
CNumber
CNumber
CNumber
(cth12 * cth13);
(sth12 * cth13);
(sth13);
(-sth12 * cth23 - cth12 * sth23 * sth13);
(mixing.U)[2][2] = CNumber (cth12 * cth23 - sth12 * sth23 * sth13);
(mixing.U)[2][3] = CNumber (sth23 * cth13);
(mixing.U)[3][1] = CNumber (sth12 * sth23 - cth12 * cth23 * sth13);
(mixing.U)[3][2] = CNumber (-cth12 * sth23 - sth12 * cth23 * sth13);
(mixing.U)[3][3] = CNumber (cth23 * cth13);
break;
default:
cerr << "Number of neutrina (" << N << ") not implemented!" << endl;
exit (-1);
}
/*
cout << "Vacuum values:\n";
cout << "- Hamiltonian:\n";
Hamiltonian HH = H0();
cout << HH;
cout << "- Mixing matrix:\n";
cout << *(mixing.U);
*/
for (int i = 1; i <= N; i++)
for (int j = 1; j <= N; j++)
{
HH0[i][j] = 0;
/* H0 is the vacuum Hamiltonian in the mass eingenstates */
if ((i == 1) && (j == 1))
H0[i][i] = 1. / pow (10, Var); // This is OK for two neutrina
else
H0[i][j] = 0;
/* HH0 is the vacuum Hamiltonian in the flavour eingenstates */
for (int k = 1; k < N; k++)
for (int l = 1; l < N; l++)
HH0[i][j] += conj (U (k, i)) * H0[k][l] * U (l, j);
}
return;
}
The user should provide routines for computing the electron densities at Sun and the Earth
along the neutrino trajectory. They appear in the matter part of the neutrino hamiltonian.
7.3 Sample outputs
The following is the verbatim output of our program
hamevol.x 0
for the values of the parameters which appears in the first information lines (included by
default in the sample program).
./hamevol.x 0
Starting evolution in the Sun
Used parameters:
MAX_STEPS
100000
INIT_STEPS
DECREASE
0.1 INCREASE
5
VarI
-2.39794
VarF
Eps_Error
1e-08
Prob_Error
x2/VarI
0.00881916 x2/VarF
-2.3979
-2.4979
-2.5979
-2.6979
-2.7979
-2.8979
-2.9979
-3.0979
-3.1979
1
1
1
1
1
1
1
1
1
0.00382
0.00382
0.00481
0.00605
0.00762
0.00959
0.0121
0.0152
0.0191
4.62e-44
4.62e-44
4.62e-44
4.62e-44
4.62e-44
4.62e-44
4.62e-44
4.62e-44
4.62e-44
10000
-12.3979
0.01
8.81916e+07
4e-08
4e-08
-1.4e-08
-8.5e-08
1.5e-09
-1.7e-08
3.6e-08
-2.8e-08
7.8e-08
(......................)
References
[1] See, for instance :
M. C. Gonzalez-Garcia and Y. Nir, Rev. Mod. Phys. 75 (2003) 345; A. Y. Smirnov,
arXiv:hep-ph/0306075; S. M. Bilenky, C. Giunti and W. Grimus, Prog. Part. Nucl. Phys. 43, 1
(1999); P. Aliani, V. Antonelli, R. Ferrari, M. Picariello and E. Torrente-Lujan, AIP Conf. Proc.
655 (2003) 103 [arXiv:hep-ph/0211062]; P. Aliani, V. Antonelli, R. Ferrari, M. Picariello and
E. Torrente-Lujan, Les Rencontres de Physique de la Vallee d’Aoste: Results and Perspectives in
Particle Physics. (Vol. 28,Frascati Physics Series), pp. 151-163 [arXiv:hep-ph/0206308] ;
A. Strumia and F. Vissani, Int. J. Mod. Phys. A 17 (2002) 1755
[2] Q. R. Ahmad et al. [SNO Collaboration], Phys. Rev. Lett. 89 (2002) 011301 ; Q. R. Ahmad et al.
[SNO Collaboration], Phys. Rev. Lett. 89 (2002) 011302
[3] Y. Koshio, arXiv:hep-ex/0306002; S. Fukuda et al. [Super-Kamiokande Collaboration], Phys. Lett.
B 539 (2002) 179
[4] K. Eguchi et al. [KamLAND Collaboration], Phys. Rev. Lett. 90 (2003) 021802
[5] G. L. Fogli, E. Lisi and A. Marrone, Nucl. Instrum. Meth. A 503 (2003) 179.
[6] L. Wolfenstein, Phys. Rev. D 17, 2369 (1978); S.P. Mikheyev and A. Yu. Smirnov, Yad. Fiz. 42,
1441 (1985) [Sov. H. Nucl. Phys. 42, 913 (1985)]
[7] S. T. Petcov. Phys. Lett. B 200,3 373 (1988)
[8] U. Toshev. Phys. Lett. B 196 170 (1987)
[9] S. Parke, Phys. Rev. Lett. 57, 1275 (1986); P.I. Krastev and S.T. Petcov, Phys. Lett. B 207, 64
(1988); S.T. Petcov, Phys. Lett. B 406, 355 (1997); M. Bruggen, W.C. Haxton, and Y-Z Qian,
Phys. Rev. D 51, 4028 (1995); A.B. Balantekin, J.F. Beacom, and J.M. Fetter, Phys. Lett. B 427,
317 (1998); G. Fiorentini, M. Lissia, G. Mezzorani, M. Moretti, and D. Vignaud, Phys. Rev. D 49,
6298 (1994).
[10] E. Torrente Lujan, Phys. Rev. D 53, 4030 (1996) See also: E. Torrente-Lujan, Phys. Rev. D 60
(1999) 085003; V. Antonelli and E. Torrente Lujan, Phys. Rev. A 58 (1998) 1980
[11] T. Ohlsson and H. Snellman, hep-ph/9910546; P. Osland and T.T. Wu, hep-ph/9912540
[12] J.S. Kim, Y.S. Chae and J.D. Kim, Comp. Phys. Comm. 120, 41 (1999).
[13] J.S. Kim and C.W. Kim, hep-ph/9909428.
[14] J. S. Kim and K. Lee, Comput. Phys. Commun. 135, 176 (2001)
[15] K. Hagiwara et al. [Particle Data Group Collaboration], Phys. Rev. D 66, 010001 (2002)
[16] J.N. Bahcall, R. Ulrich. Rev. Mod. Phys. 60 (1988) 267
[17] S. T. Petcov. Phys. Lett. B Vol.191 (1987) p.299
[18] W.C. Haxton, Phys. Rev. D 35 (1987) 2352
[19] D. Notzold, MPI-PAE/Pth 08/87 (Munich 87)
[20] P. Aliani, V. Antonelli, M. Picariello and E. Torrente-Lujan, Nucl. Phys. B 634 (2002) 393;
P. Aliani, V. Antonelli, R. Ferrari, M. Picariello and E. Torrente-Lujan, Phys. Rev. D 67 (2003)
013006; P. Aliani, V. Antonelli, M. Picariello and E. Torrente-Lujan, arXiv:hep-ph/0212212;
P. Aliani, V. Antonelli, M. Picariello and E. Torrente-Lujan, JHEP 0302 (2003) 025; P. Aliani,
V. Antonelli, M. Picariello and E. Torrente-Lujan, New J. Phys. 5 (2003) 2
[21] E. Torrente-Lujan, Phys. Rev. D 59 (1999) 073001. E. Torrente-Lujan, JHEP 0304 (2003) 054
E. Torrente-Lujan, Phys. Lett. B 389 (1996) 557. V. B. Semikoz and E. Torrente-Lujan, Nucl.
Phys. B 556 (1999) 353. E. Torrente-Lujan, Phys. Rev. D 59 (1999) 093006.
[22] S. Khalil, C. Munoz and E. Torrente-Lujan; New J. Phys. 4 (2002) 27; D. G. Cerdeno, E. Gabrielli,
S. Khalil, C. Munoz, E. Torrente-Lujan and E. Torrente-Lujan, Nucl. Phys. B 603 (2001) 231;
E. Gabrielli, S. Khalil and E. Torrente-Lujan, Nucl. Phys. B 594 (2001) 3
[23] W. H. Press, et al.,”Numerical Recipes in C++”, Cambridge University Press, Cambridge, 2002
Variable
default
description
hamevol-rungekutta.hpp
1000000
1.0 × 10−10
0.9
-0.2
-0.25
1.89 × 10−4
MAXSTP
TINY
SAFETY
PGROW
PSHRNK
ERRCON
hamevol-sample.cpp
MAX-STEPS
MIN-STEPS
INIT-STEPS
DECREASE
INCREASE
1.0 × 105
10000
10000
0.1
5.0
RK algorithm internal parameter
id.
id.
id.
id.
id.
Steeper method
id.
id.
id.
id.
id.
hamevol-sample.hpp
fermi-MeV
m-eV
Gf
Na
RSun
REarth
Eps-Error
Prob-Error
N
dist
dist-min
EARTH
SUN
1.0/197.326
15
fermi-MeV × 10
106
1.66 × 10−23
6.022 × 1023
6.961 × 108 × m-eV
6.378 × 106 × m-eV
1.0 × 10−8
1.0 × 10−2
2
0.00001
0.0000001
0
1
conversion f → 1/M eV
conversion m → 1/eV
The Fermi constant in 1/eV 2
Avogadro number
Radius of the Sun ( 1/eV)
Radius of the Earth (1/eV)
number of equations
initial stepsize for Runge-Kutta
minimal stepsize for Runge-Kutta
Program option flag
Program option flag
Table 2: Here is a list of the most important switches and constants
Subroutine
Purpose
derivs
runge
odeint
Computes the derivatives dy/dx
Given the functions y and their derivatives dy/dx, it returns the advanced solution
RK driver with adaptive stepsize control. Integrates the starting value over an interval
with a required accuracy
Used to monitor accuracy and adjust stepsize during RK integration
Returns advanced solution over an interval together with the estimate of truncation error
rkqs
rkck
Table 3: The main subroutines and functions used in the code are reported together with a brief
explanation of their meaning.
| 5cs.CE
|
FRACTIONAL BROWNIAN MARKETS WITH TIME-VARYING VOLATILITY
AND HIGH-FREQUENCY DATA
arXiv:1707.06416v1 [math.ST] 20 Jul 2017
1
Ananya Lahiri1,† and Rituparna Sen2
Chennai Mathematical Institute, Chennai, India
2
Indian Statistical Institute, Chennai, India
Abstract. Diffusion processes driven by Fractional Brownian motion (FBM) have often been considered in modeling stock price dynamics in order to capture the long range dependence of stock price
observed in reality. Option prices for such models had been obtained by Necula (2002) under constant drift and volatility. We obtain option prices under time varying volatility model. The expression
depends on volatility and the Hurst parameter in a complicated manner. We derive a central limit
theorem for the quadratic variation as an estimator for volatility for both the cases, constant as well
as time varying volatility. That will help us to find estimators of the option prices and to find their
asymptotic distributions.
Keywords: asymptotic normality, fractional Black Scholes model, Malliavin calculus, option price,
volatility, Wick financing, Wick Ito Skorohod integration, Wiener chaos.
1. Introduction
It has been proposed to model stock prices as a diffusion driven by fractional Brownian motion
(fBm) in order to capture long range dependence of stock price in reality. See Cont (2005) for
evidence of long memory in finance and relation to fractional Brownian motion.
Cheridito (2003) has shown that the solution of the diffusion equation driven by fBm with suitably
time lag will lead to an arbitrage-free model. Guasoni (2006) has shown no arbitrage under transaction cost for fBm model. Elliott and Van der Hoek (2003), Biagini et al. (2004) have shown under
Wick Ito Skorohod notion of integration one can get arbitrage free market with fBm in some sense.
Option prices for such models are obtained by Necula (2002) under constant drift and volatility. One
of the aims of this paper is to obtain estimator for some functional of volatility which can be used
to price option under time varying volatility model.
For Brownian motion (Bm) setup, the estimation of one of the important functional of volatility
appeared in option price formula, called integrated volatility, is performed using sum of frequently
sampled squared data. For high frequency data with equal interval this estimator is essentially quadratic variation. FBm is long memory process for Hurst parameter H ∈ ( 12 , 1). It is well established
result that for pure fBm with H ∈ ( 12 , 43 ) quadratic variation is asymptotically normal. Using that
result we will show that in the diffusion driven by fBm with H < 34 with constant volatility, quadratic
variation is asymptotically normal. For similar model and low frequency data with fixed time gap
asymptotic normality for volatility estimator was obtained by Xiao et al.(2013). In our paper we
consider the estimator for high frequency data with time intervals decreasing to zero and show the
asymptotic normality for the estimator. Confidence intervals for volatility can now be translated to
confidence intervals for option prices as the expression for option price involve the quantity volatility.
For diffusion driven by fBm with H ∈ ( 21 , 34 ) and time varying bounded volatility case also we will
show the asymptotic normality of the estimator from high frequency data.
The objective of this paper is two fold. Firstly for the diffusion driven by fBm with time varying
volatility we will find the option price in terms of some functional of volatility and the Hurst parameter. Secondly we will show the asymptotic normality property for the estimator for such parametric
†Corresponding author, E-mail: [email protected].
1
2
function. The estimator requires the prior knowledge of Hurst parameter. Once the estimate of
functional of volatility is found one can apply the estimate to get the option price.
The rest of the paper is organized as follows. In section 2 we describe the diffusion model for
the stock price. In section 3 we introduce the proper notion of integration required to obtain an
arbitrage-free solution of the diffusion equation. In section 4 we present the option pricing results.
The central limit theorems for the proposed estimator are obtained in section 5. We conclude and
summarize the current and future research directions in section 6.
2. Model
The introduction of the fractional Black-Scholes model, where the Bm in the classical Black-Scholes
model is replaced by a fBm, have been motivated by empirical studies (see for example Mandelbrot
(1997), Shiryaev (1999)). The risk free asset equation is
dPt = rt Pt dt,
P0 = 1,
0≤t≤T
(2.1)
The risky asset equation is
dSt = µt St dt + σt St dBtH ,
S0 = S > 0,
0≤t≤T
(2.2)
where BtH is FBM with initial condition S0 = S > 0. Here H is Hurst parameter, for 0 < H < 1. µt
is real valued deterministic function of time t, called drift and σt2 is positive real valued deterministic
function of t, called volatility. BtH is a continuous and centered Gaussian process starting at 0 with
covariance and variance functions as follows: ∀H ∈ (0, 1), s, t > 0
BtH
1
E(BtH BsH ) = (t2H + s2H − | t − s |2H )
2
(2.3)
E(BtH )2 = t2H ,
(2.4)
H
Bt+s
− BsH
BtH
has homogeneous increments, i.e.,
has same law as
for all s, t > 0. Increments are
H
H
H
dependent and correlation between the increments Bt+h − Bt and Bs+h − BsH with s + h ≤ t and
t − s = nh is as follows:
1 2H
[(n + 1)2H + (n − 1)2H − 2n2H ]
(2.5)
ρH
n = h
2
Firstly we will find the European call option price for this model. Secondly we try to provide an
estimator for option price. In this process we see that it is enough to study the quadratic variation
of this process for given high frequency data 0 = t0 < t1 < · · · < tN = 1 with Stj , j = 0, · · · , N and
1
tj+1 − tj =
∀ j = 0, · · · , N − 1. We will propose suitable estimator with this data and see where
N
it converges and how that is useful for estimating option price. We observe that the analysis is based
on high frequency data, as sample size increases the time difference between two consecutive data
point decreases. We also note that through out our analysis we know H, we do not estimate H from
data.
3. Regarding the solution of the SDE, Wick Ito Skorohod Integral, H ∈ (0, 1)
In order to find the solution of the diffusion equation (2.2), we need to note that the whole analysis
depends on how we interpret the term dBtH . For H 6= 1/2, BtH is not a semimartingale.
There are different notions of integration with respect to BtH , out of which we choose Wick Ito
Skorohod (WIS) integral notion to solve equation (2.2), for H ∈ (0, 1), due to financial reason
outlined in, for example, see Elliot and Van der Hoek (2003) and Biagini et.al. (2004),(2008).
Rogers (1997) explains why other common notions of integral are inappropriate.
Following the WIS notion of integration the solution of the stochastic differential equation (2.2) is
Z t
Z
Z t
1
2
H
(M(σs χ[0,t] )) ds
(3.1)
St = S0 exp
σs dBs +
µs ds −
2 R
0
0
3
where M is an operator acting on s and depends on H; χ[0,t] is the indicator function. See Biagini et
Rt
al. (2004, 2008). We will discuss about the operator M in next subsection and meaning of 0 σs dBsH
in following subsection.
3.1. The integral operator M. Let us elaborate about the operator M which will be needed in
future sections, See Biagini et al. (2004). Let S(R) denote the Schwartz space of smooth rapidly
decreasing functions on R. M is defined on S(R) to L2 (R) as follows:
Z
3
d
Mf (x) = − CH (t − x)|t − x|H− 2 f (t)dt
(3.2)
dx
R
CH is constant.
1
with Fourier transform fˆ defined as
Mf (y) = |y| 2 −H fˆ(y), y ∈ R
fˆ(y) =
Z
e−ixy f (x)dx.
(3.3)
(3.4)
R
1
[Γ(2H + 1) sin(πH)] 2
, Γ(.) is gamma function and explicit expression
It turns out that CH =
[2Γ(H − 12 ) cos( 12 π(H − 21 ))]
for M is as follows:
Z
1
f (x − t) − f (x)
for H ∈ (0, ), Mf (x) = CH
dt
(3.5)
3
2
|t| 2 −H
R
1
for H = , Mf (x) = f (x)
2
Z
f (t)
1
dt
for H ∈ ( , 1), Mf (x) = CH
3
−H
2
R |t − x| 2
(3.6)
(3.7)
M extends S(R) to L2H where
1
L2H (R) = {f : R → R(deterministic); |y| 2 −H fˆ(y) ∈ L2 (R)}
= {f : R → R; Mf (x) ∈ L2 (R)}
= {f : R → R; kf kL2H < ∞}, wherekf kL2H = kMf kL2 }
We also have for f ∈ L2H (R)
hf, giL2H = hMf, MgiL2
(3.8)
d , ĝiL2 = hMf, giL2
d L2 = hMf
hf, MgiL2 = hfˆ, Mgi
(3.9)
and for f, g ∈ L2 (R) ∩ L2H (R)
3.2. Wiener Integral with respect to FBM, H ∈ (0, 1). Let f ∈ L2H (R), deterministic. Then
Mf ∈ L2 (R). The Wiener integral with respect to fractional Brownian motion is defined as
Z
Z
H
f (s)dBs = (Mf )(s)dBs
(3.10)
R
R
For detail see Appendix.
4. On the way to calculate European call option price
In this section we follow similar line of argument in that of Elliot and Van der Hoek (2003).
4
4.1. Risk-neutral measure. Let us write equation (2.2) with the notion if Wick product. Now
following Elliott and Van der Hoek (2003) we rewrite equation (2.2) as follows:
dSt = St ⋄ [µt + σt WtH ]dt
(4.1)
dBtH
where ⋄ is the Wick product for two processes and
=
. For meaning of Wick product and
dt
H
Wt see Appendix (8.2). From theorem of “Wick Ito integral” Biagini et al. (2008) or (Appendix
(8.2)) we note that St dBtH = St ⋄ WtH dt. We denote trading strategy or portfolio as θ(t, ω) = θ(t) =
(u(t), v(t)) = (ut , vt ) where u(t) and v(t) are the number of units of bond and stock respectively in
the portfolio at time t and the processes are adaptive. The value process is defined as
WtH
ztθ = ut Pt + vt St
(4.2)
Definition: The concept analogous to self-financing in the fractional Brownian setting is Wickfinancing. Elliott and Van der Hoek mention it as self financing but Wick financing strategy is not
usual buy and hold strategy. A portfolio is Wick-financing if
dztθ = ut dPt + vt dSt = ut dPt + vt St ⋄ [µt + σt WtH ]dt
(4.3)
dztθ = ut dPt + vt St ⋄ (µt + σt WtH )dt
= ut rt Pt dt + vt St ⋄ µt dt + σt vt St ⋄ WtH dt
= (zt − vt St )rt dt + µt vt St dt + σt vt St ⋄ WtH dt
µ t − rt
H
= zt rt dt + σt vt St ⋄
+ Wt dt
σt
From the Girsanov theorem in Elliott and Van der Hoek (2003) the translated process B̂tH as
Z t
µ s − rs
H
B̂t =
ds + BtH
(4.4)
σ
s
0
Z
1
dP̂
2
= exp(hω, φi− kφkL2 ) = exp( φ(s)dBs −
is fBm with respect to new measure P̂ defined on F by
dP
2
R
1
r(s)
−
µ(s)
−1
2
1
2
I[0,t] (s). For notation exp(hω, φi − 2 kφkL2 ) see Appendix
kφkL2 ) where φ(s) = M
2
σ(s)
(8.1). We note that φ(s) has to be in L2 R.
Now we can rewrite 4.3 as
dzt = rt zt dt + σt vt St ⋄ ŴtH dt
(4.5)
Z
t
dB̂tH
where, ŴtH =
. Multiplying both sides with exp(r̃t ) with r̃t =
rs ds and integrating, we get
dt
0
Z t
−r̃t
e zt − z0 =
e−r̃s σs vs Ss ⋄ ŴsH ds
(4.6)
0
and
Ê e−r̃T zT = z0
where Ê is expectation under measure P̂ . Thus there exists a risk-neutral measure.
We note that under risk neutral measure P̂ we have
dSt = rt St dt + σt St dB̂tH
which will be useful for calculating option price.
(4.7)
5
4.2. Complete Market. Let Ft = σ{BsH , 0 ≤ s ≤ t} be the filtration. The market is complete
if ∀ FT measurable bounded random variable F , ∃ z ∈ R and portfolio (ut , vt ) such that F = zT
almost surely P̂ , where zT is given by 4.2. We now proceed to verify this. By fractional Clark-Ocone
theorem in Elliott and Van der Hoek (2003) applied to F , we have,
Z T
h
i
−r̃
−r̃T
e F = Ê e T F +
(4.8)
ẼP̂ D̂t (e−r̃T F ) | Ft ⋄ ŴtH dt
0
Here ẼP̂ denotes the quasi-conditional expectation and D̂t is the fractional Hida Malliavin derivative
with respect to B̂tH . For detail see Elliot and van der Hoek (2003)and Biagini et. al. (2008). We
take z = Ê e−r̃T F . Now comparing equations (4.6) and (4.8) we get
h
i
ẼP̂ D̂t (F ) | Ft = er̃T −r̃t σt vt St
This is the condition for completeness of the market. Here we note that there is criticism about
this notion of completeness with Wick financing instead of self financing, see Bjrk and Hult (2005).
Fractional Black Scholes market has weak arbitrage but no strong arbitrage, see Biagini et al.(2008).
In the context of quasi conditional expectation we require following lemma which will be useful
for calculating option price in next section.
Z t
Z
1
H
Lemma 4.1. a) If gt = exp
σs dB̂s −
(M(σs χ[0,t] ))2 ds then for T > t, ẼP̂ [gT |Ft] = gt .
2 R
0
1,2
Z b) If F ∈ L (P̂ ) (similar to Definition A4 of Elliott and Van der Hoek (2003)), and Gt =
t
0
Ft dB̂tH , then for T > t, ẼP̂ [GT |Ft ] = Gt .
Proof. Proof can be done by direct calculation.
4.3. Price of European Call Option. We next will find the European call option price ẼP̂ [(ST −
K)+ |Ft ] for this model. When µt = µ and σt = σ, Necula (2008) obtains the price C at every
t ∈ (0, T ) of an European call option with strike price K and maturity T as
C(t, St ) = St Φ(d1 ) − Ke−r(T −t) Φ(d2 )
where
d1 =
2
S
log( Kt )+r(T −t)+ σ2 (T 2H −t2H )
σ T 2H −t2H
√
and
(4.10)
2
S
d2 =
(4.9)
log( Kt )+r(T −t)− σ2 (T 2H −t2H )
σ
√
T 2H −t2H
(4.11)
and Φ() is the cumulative probability of the standard normal distribution.
The confidence intervals for σ 2 obtained in section 5 can be translated to prediction intervals for
C as in Mykland (2000) or Avellaneda et al (1995).
Next for time varying µt and σt let us calculate option price. We need the following lemma.
Lemma 4.2. The price at every t ∈ [0, T ] of bounded FT measurable function F ∈ L2 (P̂ ) is given
by F (t) = exp(−r̃T + r̃t )ẼP̂ [F |Ft ].
Proof. Proof can be followed in similar line as in Theorem 4.1 from Necula (2008) and using part b)
of lemma 4.1.
For European call option price F will be F (ω) = (S(T, ω) − K)+ where K is the strike price.
Now following similar line of approach from Theorem 3.1 of Necula (2008) and using part a) of
lemma 4.1 we get
Z T
Z t
Z
Z
1
1
H
H
2
2
ẼP̂ exp
σs dB̂s |Ft = exp
σs dB̂s −
(M(σs χ[0,t] )) ds +
(M(σs χ[0,T ] )) ds
2 R
2 R
0
0
(4.12)
Equation (4.12) will be used for proving next theorem.
6
Theorem 4.1. The price at every T ∈ [0, T ] of an European call option with strike price K and
maturity T is given by St Φ(d1 ) − K exp(−r̃T + r̃t )Φ(d2 ) where
R
R
ln(St /K) + r̃T − r̃t + R (M(σs χ[0,T ] ))2 ds − R (M(σs χ[0,t] ))2 ds
qR
d1 =
R
2 ds −
(M(σ
χ
))
(M(σs χ[0,t] ))2 ds
s
[0,T
]
R
R
and
R
R
ln(St /K) + r̃T − r̃t − R (M(σs χ[0,T ] ))2 ds + R (M(σs χ[0,t] ))2 ds
qR
d2 =
R
(M(σs χ[0,T ] ))2 ds − R (M(σs χ[0,t] ))2 ds
R
.
Proof. Proof is in similar line as that of given in Necula (2008).
5. Estimation of volatility from discrete observations
Assume that the process is observed at discrete-time instants 0 = t0 < t1 < t2 < · · · < tN = T .
Thus the observation vector is S = (St1 , St2 , · · · , StN ). We note that this is high frequency data. In
1
> 0. In section 5.1 we present
particular, we assume tk = kh, k = 1, 2, · · · , N for a step size, h =
N
the results when σt is constant and in section 5.3 when σt is time varying. In section 5.2 and 5.4 we
present simulation studies.
5.1. Constant σ. Let us start with the estimator of σ 2 as
N −1
2
1 X
ˆ
2
σ =
log(S
)
−
log(S
)
t
t
k+1
k
Nh2H k=0
(5.1)
We shall prove a central limit theorem for σˆ2 . The main component of the proof is central limit
theorem for quadratic variations of fractional Brownian motion. We also need to bound the additional
terms that comes from the geometric nature of our process. Some of these arguments are similar to
those of Nourdin (2008, 2009). The main theorem for this section is given below.
Theorem 5.1. Assume that the stock price follows the diffusion model specified by equation (4.4)
with no drift and constant volatility σ. Also assume that H ∈ (0, 3/4). N → ∞ with the observation
interval Nh = T remaining constant. Without loss of generality we can assume T = 1. Then
√
2
N(σˆ2 − σ 2 ) =⇒ N (0, σH,2
)
(5.2)
2
where σH,2
is a constant that can be computed explicitly,
X
k
1
(1 − )[(k + 1)2H + (k − 1)2H − 2k 2H ]2 ).
= 2σ lim (1 + 2(1 − )(22H − 1)2 +
N →∞
N
N
k=2
N
2
σH,2
4
Proof. Under the condition of µ = 0 and σt = σ, the solution (3.1) of the stochastic differential
equation (4.2) simplifies to
1 2 2H
H
St = S0 exp σdBt − σ t
.
(5.3)
2
Putting this solution in the definition of σˆ2 in equation (5.1), we get,
N
−1
X
σ2
2
1
H
H
2H
2H
σ B(k+1)h − Bkh −
{(k + 1)h} − {kh}
σˆ2 =
Nh2H k=0
2
2
P
N −1
H
H
B
−
B
kh
(k+1)h
k=0
n
1
o
P
2
N
−1
H
H
2H
2H
=
(5.4)
σ
B(k+1)h − Bkh {(k + 1)h} − {kh}
Nh2H −σ k=0
2 PN −1
2
2H
− {kh}2H
+ σ4
k=0 {(k + 1)h}
7
It is already known, for example putting κ = 2 in equation (1.5) of Nourdin(2008), that if H ∈
(0, 3/4),
N −1
i
1 X h −2H
H
H 2
2
h
B(k+1)h
− Bkh
− 1 =⇒ N (0, σH,2
)
(5.5)
XN := √
N k=0
Combining (5.4) and (5.5), we have
√
N
−1
X
H
σ3
H
2
2
ˆ
B(k+1)h − Bkh
{(k + 1)h}2H − {kh}2H
N ( σ − σ ) = XN − √
N h2H k=0
N
−1
X
2
σ4
+ √
{(k + 1)h}2H − {kh}2H .
4 N h2H k=0
(5.6)
It is shown in lemma 5.1 that the second term converges to zero in L2 as N → ∞. In lemma 5.2 it
is shown that the third term converges to zero. The theorem now follows by applying Chebyshev’s
inequality to get convergence in L2 implies convergence in probability and Slutsky’s theorem to get
final asymptotic normality.
Lemma 5.1. Under the assumptions of theorem 5.1,
√
Proof. As h =
N
−1
X
H
L2
σ3
H
B(k+1)h − Bkh
{(k + 1)h}2H − {kh}2H −→ 0
N h2H k=0
(5.7)
1
l.h.s of (5.7) is
N
3
U1 = σ N
N
−1
X
2H− 21
k=0
H 2
Now we have E(B H
k+1 − B k ) =
N
N
H
(B H
k+1 − B k )((
N
N
k + 1 2H
k
) − ( )2H )
N
N
(5.8)
1
,
N 2H
H
H
H
E(B H
k+1 − B k )(B k+2 − B k+1 ) =
N
N
N
N
22H − 1
1
∼ 2H
2H
N
N
k − j − 1 2H
k − j 2H i
1 h k − j + 1 2H
|
| +|
| − 2|
| for |k − j| > 1.
N
N
N
N
2
N
N
N
Then second moment of (5.8) becomes
H
H
H
E(B H
k+1 − B k )(B j+1 − B j ) =
σ 6 N 4H−1
N
−1 N
−1
X
X
k=0 j=0
=
σ6
N
N
−1 N
−1
X
X
((
k + 1 2H
k
j + 1 2H
j
H
H
H
) − ( )2H ) (
) − ( )2H E(B H
k+1 − B k )(B j+1 − B j )
N
N
N
N
N
N
N
N
(2H(k + 1)2H−1 )(2H(j + 1)2H−1 ) × [
k=0 j=0
k=j,|k−j|=1
1
]
N 2H
N −1 N −1
σ6 X X
(2H(k + 1)2H−1 )(2H(j + 1)2H−1 )
+
N
j=0
k=0
|k−j|>1
k − j − 1 2H
k − j 2H i
1 h k − j + 1 2H
| +|
| − 2|
|
× |
2
N
N
N
8
≤
σ6
2
N 2H+1
4H [N
4H−1
+ 2N
4H−1
N −1 N −1
σ6 X X
(2H(k + 1)2H−1 )(2H(j + 1)2H−1 )
]+
N k=0 j=0
|k−j|>1
h 1
k − j 2H−2 i
× 2 2 2H(2H − 1)|
|
N
N
N −1 N −1
16σ 6 H 3 (2H − 1) X X
≤ 12σ 6 H 2 [N 2H−2 ] +
(k + 1)2H−1 (j + 1)2H−1 |k − j|2H−2
N 2H+1
j=0
k=0
|k−j|>1
= 12σ 6 H 2 [N 2H−2 ] +
16σ 6 H 3 (2H − 1)
T1
N 2H+1
Now,
T1 =
N
−1
X
k 2H−2
≤
=
k 2H−2
k
2H−2
≤
≤
k
2H−2
N
X
(2Hj 2H−1 )2
N
X
(2Hj 2H−1 )2
j=1
k=1
N
−1
X
(2H(k + j)2H−1 )2
j=k+1
k=1
N
−1
X
N
−k
X
j=1
k=1
N
−1
X
(2Hj 2H−1 )(2H(k + j)2H−1 )
j=1
k=1
N
−1
X
N
−k
X
k 2H−2 (4H 2 N 4H−1 )
k=1
2H−1
≤ N
4H 2 N 4H−1
= 4H 2N 6H−2
So, we get E(U12 ) < 12σ 6 H 2 N 2H−2 + 16σ 6 H 3 (2H − 1)N 4H−3 → 0 as N → ∞ if H < 43 . Hence the
result.
Lemma 5.2. Under the assumptions of theorem 5.1,
N
−1
X
2
σ4
√
{(k + 1)h}2H − {kh}2H → 0
4 N h2H k=0
(5.9)
N −1
1
σ 4 2H− 1 X k + 1 2H
k
2
Proof. Again putting h =
l.h.s of (5.9) is U2 = N
| − | |2H )2
(|
N
4
N
N
k=0
U2
N −1
1
3
σ4
σ 4 −2H− 1 X
2
4H 2 [(k + 1)2H−1 ]2 ≤ N −2H− 2 4H 2 N 4H−1 = σ 4 H 2 N 2H− 2
≤
N
4
4
k=0
So, U2 → 0 as n → ∞ if 2H −
3
2
< 0 that is H < 43 . Hence the result.
X
k
1
(1 − )[(k + 1)2H + (k − 1)2H − 2k 2H ]2 ).
= 2σ lim (1 + 2(1 − )(22H − 1)2 +
N →∞
N
N
N
We note that
2
σH,2
4
k=2
9
Table 1. The MEAN, VAR, MSE, ASYV of the estimators when σ 2 = 0.4
σ 2 =0.4
H
0.55
0.65
0.74
MEAN
0.399936
0.3986239
0.4016842
VAR 0.0002994642 0.00044069611 0.0007573279
MSE 0.0002994683 0.0004088546 0.0007601643
Table 2. The MEAN, VAR, MSE, ASYV of the estimators when σ 2 = 1.6
σ 2 =1.6
H
0.55
0.65
MEAN
1.601847
1.605567
VAR
0.00456971 0.008053313
MSE 0.004573123 0.008084309
Table 3. The MEAN, VAR, MSE, ASYV of the
0.74
1.607953
0.01498881
0.01505207
estimators when σ 2 = 6.4
σ 2 =6.4
H
0.55
0.65
0.74
MEAN 6.401082
6.450209 6.720847
VAR 0.08589091 0.1253994 0.4416412
MSE 0.08589208 0.1279204 0.5445839
5.2. Simulation Studies for fixed σ. In this section we present the simulation result for FBM
driven model and it’s estimator. We use somebm packages from R to simulate fractional Brownian
motion. We keep drift parameter µ = 0. We generate each sample paths with N = 1000 points and
replicate with replication number, say r =, 200 times to find the mean, variance and mean squared
1 3
error. We repeat the simulation for different values of σ 2 and for H ∈ ( , ). Simulation shows that
2 4
1
3
estimators are excellent when H > and H < for high frequency data with 1000 values in the
2
4
time interval 0 to 1.
5.3. Time varying σt . In this section we will prove
the result for long memory process only, i.e.
Z
1
H > . Our parameter of interest would be θ = (M(σs χ[0,1] ))2 ds.
2
R
To get the properties of time varying volatility estimator we need some Mathematical foundation.
Readers are referred to Appendix for background for time varying volatility estimator before starting
of this section. We denote
Z
Z
H
ηk :=
σs χ[ k , k+1 ] dBs =
fk (s)dBsH = I1 (fk )
R
N
N
R
where fk (s) = σs χ[ k , k+1 ] (s) and I1 is Wiener integral with respect to fBm BtH so ηk is same as the
N N
Wiener integral discussed before. Define
!
N
−1
X
√
−2H+1 2
h
ηk − θ̃N
(5.10)
XN = N
k=0
where θ̃N =
N
−1
X
h−2H+1 E(ηk2 ).
k=0
Theorem 5.2. Assume that the stock price follows the diffusion model specified by equation (4.4)
with no drift (µ = 0) and time varying volatility σt . Also assume that H ∈ (1/2, 3/4). N → ∞
10
with the observation interval Nh = T remaining constant. Without loss of generality we can assume
T = 1. Then,
2
XN =⇒ N (0, σH,2,∗
)
(5.11)
2
where σH,2,∗
can be computed explicitly given the form of σ(t) with the formula
2
σH,2,∗
= lim N
4H−1
N →∞
2
N
−1 N
−1
X
X
(
Proof. r.h.s. of 5.10 can be rewritten as XN =
√
Mfk (s)Mfk′ (s)ds)2
R
k=0 k ′ =0
.
Z
N 2H−1
N
N
−1
X
k=0
Z
!
ηk2 − θ̃N .
Let us introduce some notations.
Z
2
Let us denote E(ηk ) = θk . We get θk = (M(σs χ[ k , k+1 ] ))2 ds as expectation of Wiener integral
N
R
R
N
σs χ[ k , k+1 ] dBsH . Now
N
N
E(N
2H−1
N
−1
X
ηk2 )
=N
2H−1
k=0
N
−1
X
θk = N
2H−1
k=0
N
2H−1
N
−1 Z
X
k=0
R
N
−1 Z
X
k=0
R
(M(σs χ[ k , k+1 ] ))2 ds
N
N
(M(σs χ[ k , k+1 ] ))2 ds = θ̃N
N
N
If H > 1/2 and σ(s) < Σ ∀ s then
Z
R
Mfk (s)Mfj (s)ds ≤ Σ2 H(2H − 1)
|k − j|2H−2
for |k − j| > 1
N 2H
1
c for |k − j| = 1, c = (22H − 1) a constant
2N 2H
Σ2
for k = j
≤
N 2H
≤ Σ2
Using product formula for Wiener chaos integrals (8.7) we get
ηk2 = I12 (fk (s))
= I2 (fk ⊗0 fk ) + I0 (fk ⊗1 fk )
Z
= I2 (fk ⊗0 fk ) + (M(σs χ[ k , k+1 ] ))2 ds
R
N
N
and ηk2 − θk = I2 (fk ⊗0 fk )
R
We note that I0 (fk ⊗1 fk ) = hfk , fk iH = R (M(σs χ[ k , k+1 ] ))2 ds.
N N
Z
2
Then E(ηk ) = EI2 (fk ⊗0 fk ) + (M(σs χ[ k , k+1 ] ))2 ds. So, we get EI2 (fk ⊗0 fk ) = 0.
R
N
Let us now calculate the second moment of
N
N
−1
X
k=0
ηk2 .
11
E
N
−1
X
ηk2
k=0
!2
=
N
−1 N
−1
X
X
k=0
=
k ′ =0
N
−1 N
−1
X
X
k=0
k ′ =0
E ((I2 (fk ⊗0 fk ) + θk )(I2 (fk′ ⊗0 fk′ ) + θk ))
[E(I2 (fk ⊗0 fk )I2 (fk′ ⊗0 fk′ ))
+θk EI2 (fk′ ⊗0 fk′ ) + θk′ EI2 (fk ⊗0 fk ) + θk θk′ ]
= A1 + A2 + A3 + A4
Where A1 , A
!2 , A3 , A4 are respective terms in the summation. Now A2 = A3 = 0. We observe that
N
−1
X
Var
ηk2 = A1 .
k=0
A1 = 2
N
−1 N
−1
X
X
k ′ =0
k=0
= 2
N
−1 N
−1
X
X
k=0 k ′ =0
= 2
hfk ⊗0 fk , fk′ ⊗0 fk′ iH⊗2
hfk , fk′ i2H⊗1
N
−1 N
−1
X
X
(E(I1 (fk )I1 (fk′ ))2
k=0 k ′ =0
= 2
N
−1 N
−1
X
X
(E(ηk ηk′ ))2
k=0 k ′ =0
= 2
N
−1 N
−1
X
X
(
k=0 k ′ =0
A1 ≈ 2(N
1−4H
= 2(N
1−4H
= 2(N
= 2N
+
N
−1 N
−1
X
X
k=0 k ′ =0
k6=k ′
1−4H
1−4H
4
2
Z
Mfk (s)Mfk′ (s)ds)2
R
k ′ − k 2H−2 1
Σ H(2H − 1) |
|
N
N2
2
2
+ Σ H (2H − 1) N
4
2
2
+ Σ H (2H − 1) N
4
2
2
−4H
N
−1 N
−1
X
X
k=0 k ′ =0
k6=k ′
N
−1
X
−4H
+ 4Σ H (2H − 1) N
NX
−1+k
(N − 1)
Then N 4H−1 A1 < ∞ if H < 43 . So, we can see N 4H−1 E
N
−1
X
k=0
)
| k ′ − k |4H−4 )
k=−N +1 k ′ =k
k6=0
−4H
2
N
−1
X
| k |4H−4 )
k 4H−4 for H > 1/2
k=1
ηk2
!2
= N 4H−1 A1 .
12
Let us write XN in terms of multiple Wiener Ito integral.
!
N
−1
X
√
XN =
N N 2H−1
(I1 (fk ))2 − θ̃N
k=0
=
√
N 2H−1
N
N
−1
X
k=0
= N
2H− 21
I2 (
N
−1
X
k=0
= YN +
√
(I2 (fk ⊗0 fk ) + θk ) − θ̃N
−1
X
√ 2H−1 N
θk − θ̃N
(fk ⊗0 fk )) + N N
k=0
N(N 2H−1
N
−1
X
k=0
Now
√
N (N
2H−1
N
−1
X
k=0
!
θk − θ̃N )
θk − θ̃N ) = 0 for all N. Observe that E(YN2 ) = N 4H−1 A1 = SN . Define
YN
GN = √ .
SN
To prove asymptotic normality we will use the two theorems (8.1) and (8.2). Using the theorems
we want to show that kDGN k2H → 2 in L2 . For that matter we first show lim E[kDGN k2H ] = 2 and
N →∞
then lim E[kDGN k2H − 2]2 = 0. Now,
N →∞
[kDGN k2H − 2]2
= [kDGN k2H − E[kDGN k2H ] + E[kDGN k2H ] − 2]2
= [kDGN k2H − E[kDGN k2H ]2 + [E[kDGN k2H ] − 2]2 + 2[kDGN k2H − E[kDGN k2H ][E[kDGN k2H ] − 2]
= A + B + 2C
where A, B, C are respective terms. Now EC = 0, B → 0 as N → ∞. We will be interested in A for
further analysis.
Using (8.11) we get
N
−1
X
1
fj (t)I1 (fj ).
Dt YN = 2N 2H− 2
j=0
So
kDYN k2H
= 4N
4H−1
k=0 j=0
Now
E[kDYN k2H ]
N
−1 N
−1
X
X
4H−1
= 4N
I1 (fj )I1 (fk )hfj , fk iH .
N
−1 N
−1
X
X
k=0 j=0
(hfj , fk iH )2 = 4N 4H−1
A1
= 2SN
2
We note that E[kDYN k2H ] = 2E[YN2 ]. So E[kDGN k2H ] → 2 as N → ∞. Let us calculate the following
kDYN k2H − E[kDYN k2H ]
N
−1 N
−1
X
X
4H−1
[(I2 (fj ⊗0 fk ) + I0 (fk ⊗1 fk )hfj , fk iH − hfj , fk i2H ]
= 4N
k=0 j=0
= 4N
4H−1
N
−1 N
−1
X
X
k=0 j=0
[I2 (fj ⊗0 fk )hfj , fk iH ]
13
And then
E[kDYN k2H
−
E[kDYN k2H ]]2
= 16N
8H−2
N
−1 N
−1
X
X
E[
k=0 j=0
[I2 (fj ⊗0 fk )hfj , fk iH ]]2
≤ constant N 8H−6 → 0
For last part of the calculation see lemma 5.2 of Tudor (2013). as N → ∞ and H < 43 . Hence the
proof.
Theorem 5.3. Under the conditions of theorem 5.2,
√
2
)
N (σˆ2 − θ̃N ) =⇒ N (0, σH,2,∗
(5.12)
2
where σH,2,∗
is a constant that can be computed explicitly, given the form of σ(t).
Proof. Under the condition of µ = 0, the solution (3.1) of the stochastic differential equation (4.2)
simplifies to
Z t
Z
1
H
2
St = S0 exp
σs dBs −
(M(σs χ[0,t] )) ds .
(5.13)
2 R
0
Z
Z
k
X
2
˜
˜
Let us denote fk (s) =
fk (s) and δk = (M fk+1 (s)) − (M f˜k (s))2 . Putting the solution (5.13)
R
j=0
R
in the definition of σˆ2 in equation (5.1), we get,
2
N −1
1
1 X
ηk − δk
Nh2H
2
k=0
#
"N −1
N
−1
N
−1
X
X
X
1
δk2
= N 2H−1
ηk2 −
ηk δk +
4
k=0
k=0
k=0
σˆ2 =
(5.14)
Combining (5.14) and (5.11), we have
√
1
N(σˆ2 − θ̃N ) = XN − N 2H− 2
N
−1
X
k=0
N
−1
X
1
1
ηk δk + N 2H− 2
δk2 .
4
(5.15)
k=0
where XN is defined in theorem 5.2. It is shown in lemma 5.3 that the second term converges to zero
in L2 as N → ∞. In lemma 5.4 it is shown that the third term converges to zero. The theorem now
follows by applying Chebyshev and Slutsky with theorem 5.2 as before.
Lemma 5.3. Under the assumptions of theorem 5.2,
G=N
2H− 12
N
−1
X
k=0
Proof. Let us recall
Eηk ηk′ =
δk =
Z
R
Z
L2
ηk δk −→ 0
Mfk (s)Mfk′ (s)ds
R
[(M f˜k+1 (s))2 − (M f˜k (s))2 ]ds
(5.16)
14
For H >
1
2
δk = H(2H − 1)[
Z
= H(2H − 1)[
k+1
N
0
Z
Z
k+1
N
0
k+1
N
k
N
Z
k+1
N
σ(s)σ(t)|s − t|
+2
k
N
Z
k
N
0
Z
k+1
N
2H−2
dsdt −
Z
0
k
N
Z
k
N
0
σ(s)σ(t)|s − t|2H−2 dsdt]
]
k
N
So
Σ2
[1 + (k + 1)2H − (k)2H − 1]
N 2H
Σ2
≈
(k + 1)2H−1
N 2H
|δk | ≤
We look at the L2 norm of G.
2
E(G ) = N
4H−1
N
−1 N
−1
X
X
k=0
→ 0.
using estimates for δk and estimates of
Z
δk δk′
Z
Mfk (s)Mfk′ (s)ds
R
k ′ =0
Mfk (s)Mfk′ (s)ds.
R
Lemma 5.4. Under the assumptions of theorem 5.2, then
N
2H− 12
N
−1
X
k=0
Proof. Again using the estimates of δk we have
N
2H− 12
N
−1
X
k=0
δk2
≤ N
2H− 21
4
= ΣN
δk2 → 0
N
−1
X
k=0
−2H− 12
Σ2
(k + 1)2H−1
N 2H
(5.17)
2
N 4H−1
3
1
as well as H < .
2
4
5.4. Simulation studies (time varying volatility). InRthis section we did simulation studies to
see the difference between our actual parameter of interest R (M(σ(s)χ[0,1] ))2 ds and what we achieve
θ̃N for sample size N, for different Hurst parameter with different σ(s) function. We have chosen
different functions σ(t), necessarily bounded, on the interval [0, 1], calculate θ̃N and report the results.
We take N = 1000 and compute θ̃N and θ and note the difference.
Let us consider σ(t) a sub linear function of the form σ(t) = σtα , t ∈ (0, 1), 21 < H < 43 ,
σ > 0, 0 < α < 1. Next we consider functions of the form σ(t) = σ(tα + tβ ) for t ∈ (0, 1), 0 < α < 1
and β > 1, i.e. polynomial with positive fraction and integer powers. We note that θ̃N may not
converge as N → ∞ and it will indeed not converges. So the simulation result is only for N = 1000.
This proves the lemma for H >
For practical purpose the sub linear functions seems best.
6. Conclusions
In this paper we sketch the way to obtain the option price for fBm driven model with time varying
volatility. We identify the parameter of interest for calculating option price. Next we have proposed
estimator from high frequency data for parameter similar to so called ”integrated volatility”, in
case of constant volatility and time varying volatility model driven by fBm. We have shown that
estimators are asymptotically normally distributed for H < 34 . For time varying volatility model, the
15
Table 4. The comparison of θ̃N and θ
H
θ̃N
θ
H
θ̃N
θ
σ=0.4, α = 0.3
0.55
0.65
0.74
0.1000039 0.1000005 0.09992656
0.09962605 0.09868432 0.09770835
σ=6.4, α = 0.3
0.55
0.65
0.74
25.60027
25.60061
25.6002
25.50445
25.2632
25.01328
Table 5. The comparison of θ̃N and θ
H
θ̃N
θ
H
θ̃N
θ
σ=0.4,
0.55
0.1777476
0.1705813
σ=6.4,
0.55
45.50396
43.66903
α = 0.8, β = 2
0.65
0.74
0.1777324 0.1776835
0.1579971 0.1482963
α = 0.8, β = 2
0.65
0.74
45.50438
45.5041
40.44747 37.96394
estimator will not asymptotically unbiased for our parameter of interest. Through some simulation
study we showed how close of the parameter of interest can be achieved by the estimators under
consideration.
6.1. Future directions.
(1) In all these we assume H as a known quantity. The estimation for H also exists separately.
See Prakasa Rao (2010), Breton et al. (2009). Is there any way to combine?
(2) Following Zhang et al (2005) we want to develop method of inferring volatility when the
process is observed discretely with noise.
(3) Following Barndorff-Nielsen and Shephard (2004) we want to consider estimators for jump
process.
6.2. Comments.
• Why consider fBM driven models? Non-stationary time series will also take care of thick-tails
and long-range dependence in returns. But it is not easy to put them in an option pricing
framework.
• Why do we have confidence intervals for option prices, when looked at as solution of an
optimization problem? There is uncertainty in utility/preferences.
7. Acknowledgement
First author wants to acknowledge Department of Science and Technology, India, for financial
support to conduct this research work.
8. Appendix
8.1. Wiener integral . Ω := S ′ (R), dual of the space of Schwartz class functions S, is tempered
distributions with sigma algebra F . hω, f i is the random variable by action of ω ∈ Ω on f ∈ S(R).
1
2
Bochnor Minlos probability measure P on (Ω, F ) is such that E[exp ihω, f i] = e− 2 kf k , kf k2 =
16
Z
R
f 2 (x)dx. We also have expectation E[hω, f i] = 0 and variance E[hω, f i]2 = kf k2 under P . Let
I(0, t) is indicator function. Then B̃t (ω) = hω, I(0, t)i ∈ L2 (P ) and B̃t is Gaussian random variable
for each t. Using Kolmogorov’s continuity theorem B̃t has continuous version as Bt and Bt is standard
2
Brownian motion.
Z This duality can be extended for f ∈ L (R) approximating f by step functions
we get hω, f i =
f (t)dBt (ω).
R
H
H
Let M(0, t) = MI(0, t) ∈ L2 (R).
Z Then B̃t (ω) = hω, M(0, t)i. B̃t is Gaussian random variable
1
M(0, t)(x)M(0, s)(x)dx = (t2H + s2H − |t − s|2H ). Again take
with E B̃tH = 0 and E B̃tH B̃sH =
2
R
Z
continuous version of B̃tH as BtH . So we get BtH = (MI(0, t)(s)dBs .
R
Let f ∈ L2H (R), deterministic. Then Mf ∈ L2 (R). The Wiener integral with respect to fractional
Brownian motion is defined as
Z
Z
H
f (s)dBs = (Mf )(s)dBs
(8.1)
R
R
Take f ∈ S. Construct Bochnor MinlosZ probability measure P H on (Ω, F ) is such that E[exp ihω, f iH ] =
1
2
e− 2 kf kH , kf k2H = kf kL2H = kMf kL2 =
R
(Mf )2 (x)dx. We also have expectation E H [hω, f iH ] = 0
and variance E H [hω, f iH ]2 = kf k2H under P H . Let I(0, t) is indicator function. Then B̃tH (ω) =
hω, I(0, t)iH ∈ L2 (P H ) and B̃tH is Gaussian random variable for each t. Using Kolmogorov’s continuity theorem B̃tH has continuous version as BtH and ZBtH is fractional Brownian motion. This
duality can be extended for f ∈ L2H (R) we get hω, f iH =
f (t)dBtH (ω). So, we note that the term
R
Z t
H
σs dBs appears in solution of the SDE (3.1) is a Wiener integral with respect to Brownian motion.
0
The first two moments of Wiener integral with respect to fractional Brownian motion are as follows
Z
E( f (s)dBsH ) = 0
R
and
Z
E[ f (s)dBsH ]2 = kf kL2H = kMf kL2
R
Z
Z
H
E[ f (s)dBs
g(s)dBsH ] = hf, giL2H
R
R
8.2. Wick product and related topics . This section consists of background material required
for section 4.
Chaos expansion Theorem Let X
F ∈ L2 (P ). Then there exists a unique family cα , α ∈ I of
constants, cα ∈ R such that F (ω) =
cα Hα (ω), Hα is multi indexed Hermite polynomial of Bm
α∈I
X
(convergence in L2 (P )). Moreover, we have the isometry E(F 2 ) =
c2α α!.
α∈I
Let us define (S) as space of stochastic test functions and (S ∗ ) as space of stochastic
distributions.
X
(S) is collection of all F ∈ L2 (P ) such that it’s expansion is F (ω) =
cα Hα (ω) where kF k2k =
X
α
α∈I
α!c2α (2N)kα < ∞ for all integer k = 1, 2, · · · with (2N)kγ
m
Y
=
(2j)γj , γ = (γ1 , · · · , γm ) ∈ I.
j=1
17
(S ∗ ) is collection of all G ∈ L2 (P ) such that it’s expansion is G(ω) =
X
β!c2β (2N)−qβ
< ∞ for some integer q < ∞.
β
X
β∈I
cβ Hβ (ω) where kGk2q =
(S ∗ ) is dual X
of (S) with duality relation as follows:
X
If F (ω) =
aα Hα (ω) ∈ (S) and G(ω) =
bα Hα (ω) ∈ (S ∗ ) then action of G on F is
α∈I
α∈I
X
hG, F i(S ∗ ),(S) =
α!aα bα .
α∈I
If L2 (P ) ⊂ (S ∗ ) and (S) ⊂ L2 (P ) then action of G on F is hG, F i(S ∗ ),(S) = hG, F iL2(P ) = E(GF ).
X
X
If F (ω) =
aα Hα (ω) ∈ (S ∗ ) and G(ω) =
bβ Hβ (ω) ∈ (S ∗ ) then Wick product ⋄ is defined
α∈I
as (F ⋄ G)(ω) =
X
α,β∈I
β∈I
∗
aα bβ Hα+β (ω) ∈ (S ).
Fractional white noise
WtH
dBtH
=
is an element of (S ∗ ) see Elliott and Van der Hoek (2003),
dt
Biagini et al. (2004) for detail.
Wick Ito Skorohod integral with respect to BtH : If Y : R → (S ∗ ) is such that Yt ⋄ WtH is
integrable in (S ∗ ) then we define
Z
Z
H
Yt dBt =
Yt ⋄ WtH dt.
(8.2)
R
If f ∈ L2H (R) then
Z
f (s)dBsH
R
=
R
Z
R
f (s) ⋄
WsH ds
=
Z
(Mf )(s)dBs .
R
8.3. Background to deal with time varying volatility estimator . In this section we introduce
some notations and established results which will be needed for our future calculation. Our fractional
Brownian motion BtH is centered, continuous, mean zero Gaussian processes with covariance functions
H
as RB
= cov(BsH , BtH ) = 12 [t2H + s2H − |t − s|2H ]. Let E be the set of real valued step functions.
H
For φP= I[0, t], ψ = I[0, s] ∈ EPlet us define inner P
product hφ, ψiE = hI[0,s] , I[0,t] iE = RB
. For
H
H
H
H
φ = j aj I[0, tj ], set B (φ) = j aj Btj . Let ψ = j bj I[0, tj ]. So, E(B (φ)B (ψ)) = hφ, ψiE .
Next for φ ∈ H, there are φn ∈ E such that φn → φ in H then B H (φ) is the L2 limit of B H (φn ). So
we get hφ, ψiH = EB H (φ)B H (ψ). {B H (φ), φ ∈ H} is called isonormal Gaussian process.
Let Hn be n th Hermite polynomial satisfying
d
Hn (x) = Hn−1 (x), n ≥ 1.
dx
(8.3)
Take φ ∈ H such that kφkH = 1. Consider random variables Hn (B H (φ)) and take the closure of the
span of these random variables. This is the n th order Wiener chaos Wn .
In , the multiple stochastic (Wiener Ito) integral with respect to isonormal Gaussian process B H ,
1
is a map from H⊙n to Wn , H⊙n being symmetric tensor product of H. H⊙n has norm √ k.kH⊗n ,
n!
H⊗n is tensor product of H.
Then for f ∈ H⊙n , we also have In (f ) = In (f˜), f˜ is symmetrization of f .
For φ ∈ H, In (φ⊗n ) = n!Hn (I1 (φ)) = n!Hn (B H (φ)) is linear isometry between H⊙n and Wn .
Now for f ∈ H⊙n and g ∈ H⊙m we have followings:
E(In (f )Im (g)) = n!hf˜, g̃iH⊗n if m = n
E(In (f )Im (g)) = 0 if m 6= n
(8.4)
(8.5)
18
Let {ei , i ≥ 1} be an orthonormal basis of H, m, n ≥ 1, r = 0, · · · , n ∧ m. f ⊗r g ∈ H⊗(m+n−2r) is
contraction is defined as
∞
X
(8.6)
f ⊗r g =
hf, ei1 ⊗ · · · ⊗ eir iH⊗r hg, ei1 ⊗ · · · ⊗ eir iH⊗r .
i1,··· ,ir=1
This definition does not depend on the choice of orthonormal basis and hf, ei1 ⊗· · ·⊗eir iH⊗r ∈ H⊙(n−r) ,
˜ r g is symmetrization
hg, ei1 ⊗ · · · ⊗ eir iH⊗r ∈ H⊙(m−r) . f ⊗r g is not necessarily symmetric. Let f ⊗
of f ⊗r g. Then
m∧n
X nm
˜ r g).
In (f )Im (g) =
r!
I
(f ⊗
(8.7)
r r n+m−2r
r=0
Also for n = m = r we have
I0 (f ⊗r g) = hf, giH⊗r .
(8.8)
H
H 2
Let F be a functional of the isonormal Gaussian process B such that E(F (B ) ) < ∞ then
thereX
is unique sequence fn ∈ H⊙n and F can be written as sum of multiple stochastic integrals as
F =
In (fn ) with and I0 (f0 ) = E(F ) where the series converges in L2
n≥0
For φ1 , · · · , φn ∈ H, let F = g(B H (φ1 ), · · · , B H (φn )) with g smooth compactly supported. Then
Malliavin derivative D is H valued random variable defined as follows:
n
X
∂g H
DF =
(B (φ1 ), · · · , B H (φn ))φi .
(8.9)
∂x
i
i=1
If H is L2 (R) for some non atomic measure then DF can be identified as follows: DF = (Dt F )t∈R
n
X
∂g H
Dt F =
(B (φ1 ), · · · , B H (φn ))φi (t), t ∈ R
∂x
i
i=1
(8.10)
If F = In (f ), f ∈ H⊙n , for every t ∈ R, then
Dt F = Dt In (f ) = nIn−1 f (., t).
(8.11)
In−1 (f (., t)) means n − 1 multiple stochastic integral is taken with respect to first n − 1 variables
t1 , · · · , tn−1 of f (t1 , · · · , tn−1 , t), t is kept fixed. For Malliavin calculus details, see Nualart (1995),
Nourdin (2012). To prove asymptotic normality we will use the following two theorems [5.1] and
[5.2] taken from Tudor C.A. (2008):
Theorem 8.1. Let In (f ) be a multiple integral of order n ≥ 1 with respect to an isonormal process
M. Then
1
d(L(In (f )), N (0, 1)) ≤ cn [E(|DIn (f )|2H − n)2 ] 2
where D is the Malliavin derivative with respect to B H and H is the canonical Hilbert space associated
to B H . Here d can be any of the distances like Kolmogorov Smirnov distance, or total variation
distance etc. and depending upon d and the order n one will end up a constant cn . L(B H ) stands
for law of B H .
Theorem 8.2. Fix n ≥ 2 and let (Fk , k > 1), Fk = In (fk ) (with fk ∈ H⊙n , for every k ≥ 1) be
a sequence of square integrable random variables in the nth Wiener chaos of an isonormal process
B H such that E[Fk2 ]2 → 1 as k → ∞. Then the following are equivalent: (i) The sequence (Fk )k≥0
converges in distribution to the normal law N (0, 1). (ii) One has E[Fk4 ] → 3 as k → ∞. (iii) For
all 1 ≤ l ≤ n − 1 it holds that lim |fk ⊗l fk |H⊗2(n−l) = 0. (iv) |DFk |2H → n in L2 as k → ∞, where
k→∞
D is the Malliavin derivative with respect to B H .
The above theorems we will use to prove asymptotic normality for our proposed estimator.
19
References
[1] Avellaneda, M., Levy, A. and Pars, A. (1995), “Pricing and hedging derivative securities in markets with uncertain
volatilities” Applied Mathematical Finance, 2(2), 73-88.
[2] Barndorff-Nielsen, O. E. and Shephard, N. (2004), “Power and Bipower Variation with Stochastic Volatility and
Jumps”, Journal Of Financial Econometrics, 2 (1), 1-37.
[3] Biagini, F., ksendal, B., Sulem, A. and Wallner, N. (2004), “An introduction to white-noise theory and Malliavin
calculus for fractional Brownian motion”, Proc. Royal Society London, 460, 347-372.
[4] Biagini, F., Hu,Y., Oksendal,B. and Zhang, T. (2008), “Stochastic Calculus for Fractional Brownian Motion and
Applications , Springer.
[5] Bjrk, T. and Hult, H. (2005), “A note on Wick products and the fractional Black-Scholes model”, Finance and
Stochastics, 9(2), 197-209.
[6] Breton, J. et al. (2009), “Exact confidence intervals for the Hurst parameter of a fractional Brownian motion”,
Electronic Journal of Statistics 3, 416–425.
[7] Cheridito, P. (2003), “Arbitrage in fractional Brownian motion models” Finance and Stochastics 7 (4), 533-553.
[8] Cont, R. (2005), “Long range dependence in financial markets”, In J. Lvy-Vhel & E. Lutton (Eds.), Fractals in
engineering: New trends in theory and applications (pp. 159-180): Springer
[9] Elliott, R.J., van der Hoek, J. (2003), “A general White noise theory and applications to finance”, Mathematical
Finance 13, 301330.
[10] Guasoni, P. (2006), “No arbitrage under transaction costs, with fractional Brownian motion and beyond”, Mathematical Finance 16 (3), 569-582.
[11] Mandelbrot, B. B. (1997), “Fractals and scaling in finance, discontinuity, concentration, risk”, Springer.
[12] Mykland, P. A. (2000), “Conservative delta hedging”, Annals of Applied Probability, 10(2), 664-683.
[13] Necula,C. (2008), “Option Pricing in a Fractional Brownian Motion Environment”, Advances in Economic and
Financial Research - DOFIN Working Paper Series 2, Bucharest University of Economics, Center for Advanced
Research in Finance and Banking - CARFIB.
[14] Nourdin, I. (2008), “Asymptotic behavior of weighted quadratic and cubic variations of fractional brownian
motion”, Annals of probability, 36(6), 2159-2175.
[15] Nourdin, I. and Reveillac, A. (2009), “Asymptotic Behavior Of Weighted Quadratic Variations Of Fractional
Brownian Motion: The Critical Case H = 1/4”, Annals of Probability, 37(6), 2200-2230.
[16] Nourdin, I., (2012), “Selected aspects of fractional Brownian motion”, Springer.
[17] Nualart, D., (1995), “Malliavin calculus and related topics”, Springer.
[18] Prakasa Rao, B.L.S. (2010), “Statistical Inference for Fractional Diffusion Processes”, Wiley.
[19] Rogers, L.C.G. (1997), ”Arbitrage with fractional Brownian motion”, Mathematical Finance, 7, 95-105.
[20] Shiryaev, A. N. (1999), “Essentials of stochastic finance: facts, models, theory”, World Scientific.
[21] Tudor, C. A., (2008), “Analysis of variance for self-similar processes”, Springer.
[22] Xiao,W., Zhang, W. and Zhang, X. (2013), “Parameter identification for the discretely observed geometric fractional Brownian motion”, Journal of Statistical Computation and Simulation, 85(2), 269-283.
[23] Zhang, L. et al. (2005), “A tale of two scales: determining integrated volatility with noisy high frequency data”,
Journal of American Statistical Association , 100(472), 1394-1411.
| 10math.ST
|
Parallel Tempering for the planted clique problem
Angelini Maria Chiara1
arXiv:1802.05903v1 [cond-mat.dis-nn] 16 Feb 2018
1
Dipartimento di Fisica, Ed. Marconi, ”Sapienza” Università di Roma, P.le A. Moro 2, 00185 Roma Italy
The theoretical information threshold for the planted clique problem is 2 log2 (N ), however no
polynomial algorithm is known to recover a planted clique of size O(N 1/2−ǫ ), ǫ > 0. In this paper
we will apply a standard method for the analysis of disordered models, the Parallel-Tempering (PT)
algorithm, to the clique problem, showing numerically that its time-scaling in the hard region is
indeed polynomial for the analyzed sizes. We also apply PT to a different but connected model, the
Sparse Planted Independent Set problem. In this situation thresholds should be sharper and finite
size corrections should be less important. Also in this case PT shows a polynomial scaling in the
hard region for the recovery.
2
The planted clique problem is the following [1]: we extract a random graph with N nodes, each node is connected
to another one with probability p. Then, we plant a clique C, imposing K nodes among the N ones to be connected to
each other with probability q = 1. Given the resulting graph, we search for an algorithm able to identify the elements
of the planted clique. The problem can be studied for all the values of the probabilities p and q, but in the following
we will focus on the values p = 12 , q = 1. Given a random graph with probability p = 12 to have a link between two
nodes, one can easily show that the largest purely random clique is of size Kran ≃ 2 · log2 (N ) for N large [2]. As a
consequence, an exhaustive search for a clique of size K returns the planted clique if K > Kran . However, it has long
been conjectured that no polynomial-time algorithm can find cliques of size N 1/2−ǫ , ǫ > 0. In Ref. [3] this has been
proved for the class of sum-of-squares
algorithms. In Ref. [4], a message passing algorithm has been constructed that
q
fails unless K > KBP = Ne . In Ref. [5], an analogous gap between the threshold for exhaustive and polynomial
algorithms has been found in the sparse clique problem. In this case, it is shown how the problem undergoes two
phase transitions: the first one is a dynamical one, and below that threshold no local algorithm is able to find the
planted clique. The second transition is a static transition that identifies the threshold for exhaustive search. The
existence of such kind of gaps is common in recovery problems, however the clique problem is in a certain manner
special because for this problem the two thresholds scale differently with N . Some recent works prove hardness in
other problems assuming the hardness of planted clique (in the corresponding region) [6],[7],[8],[9]. For this reason,
there is a lot of current interest in the planted clique problem. In the first part of this paper we will show that also
in this dense limit, the clique problem undergoes a static and a dynamic phase transition, as demonstrated in [5] in
the sparse case. Then applying standard methods for the analysis of disordered models, and in particular a Parallel
Tempering (PT) algorithm, we will see how to find the planted clique down to Kran . The thermal algorithms are
not easy to be analyzed: in particular, the challenge is to understand how the time of convergence scales with N .
We show that data from PT are in very good agreement with a polynomial scaling. However, for the clique problem
an exhaustive search algorithm can find the planted clique of size K > 2 log2 (N ) in a time O(exp(c log2 (N ))): It is
sufficient to find a clique of size k0 = 2 log2 (N ), that takes a time kN0 and then to expand starting from that one.
Thus it is quite difficult to distinguish between a polynomial or a non-polynomial O(exp(c log2 (N ))) behavior. For
this reason, in the second part of this paper we move to a different but connected model: the planted Independent Set
(IS) model. Being this problem sparse, the thresholds are sharper: the exhaustive algorithm can find solutions in time
O(exp(N )). For this problem in the hard region we numerically show that for the analyzed sizes the PT algorithm
can find solutions in polynomial time. This gives good reasons to believe that also in the case of the planted clique
problem the PT algorithm finds solutions in polynomial time.
To be concrete, we construct a graph of N nodes with a planted clique C of size K. On each node there is a variable
vi , vi = 1 if node i ∈ C, vi = 0 if node i ∈
/ C. On each edge between nodes i and j we put a variable Aij . The edge
variables are 0 or 1 with the following probabilities:
(
1 if vi vj = 1
(1)
p(Aij = 1|{v}) = 1
otherwise
2
Given a realization of the graph we want to estimate the variables vi . We will call our estimation xi . We introduce a
Belief-Propagation (BP) algorithm that is essentially the one proposed in ref. [4]. Following the Bayes formula, the
posterior probability for xi given the graph is
P (xi |{Aij }) = P (Aij |{xi })P (xi ),
(2)
where the likelihood P (Aij |{xi }) is the one of eq. (1) and P (x) is the prior probability. The original problem has
a global constraint
on the size of the clique to be recovered (we are treating the case of known K): P ({x}) should
P
be zero if i xi 6= K. However we are using local algorithms, that cannot implement global constraints. Thus we
x
1−x
1− K
. The BP algorithm is a way to extract the marginal
choose a local prior on the single node: P (x) = K
N
N
probabilities for each node from Eq. (2). We introduce cavity messages ψi→j (xi ) proportional to the probability that
node i takes value xi , conditioned on the absence of edge (ij). Iterative equations on the messages read:
3
N −1
N −K 1
ψi→j (xi = 0) =
N
2
N −1 Y
K 1
ψi→j (xi = 1) =
[1 + (2Aij − 1)ψk→i (1)]
N 2
k\j
ψ
(x )
i
Cavity probabilities are obtained from the normalization of the cavity messages: ηi→j (xi ) = i→j
zi→j , with zi→j =
ψi→j (0) + ψi→j (1).
Once the iteration of the cavity messages has reached a fixed point, marginal probabilities are obtained as:
N
N −K 1
N
2
N Y
K 1
ψi (xi = 1) =
[1 + (2Aij − 1)ψk→i (1)]
N 2
ψi (xi = 0) =
(3)
(4)
k
ηi (xi ) = ψiz(xi i ) , with zi = ψi (0) + ψi (1).
i
hP
P
− N1
ij log(zij ) , with zij =
i log(zi ) −
The Bethe free energy associated to the reached solution is f =
zi
zi→j .
We then assume the elements of the clique to be the first K
p
elements with largest ηi (1). In the thermodynamic limit the recovery is possible if K > KBP = N/e [4]. Let us
emphasize the difference in the Fixed Point of BP when the algorithm finds and does not find the planted clique: If
the planted clique is recovered, the BP messages are completely polarized: ηi (1) = 0 (if i ∈
/ C) or ηi (1) = 1 (if i ∈ C).
If the planted clique is not recovered, messages that are not completely polarized: ηi (1) = ηi∗ (1) 6= 0, 1 ∀i. We will
call this solution the paramagnetic solution.
First of all, we study the stability of the planted solution. To do this, we initialize the BP messages near enough
to the planted solution and we look to the solution reached after iteration. The planted solution is reached well
below the threshold Kran = 2 · log2 (N ), until the threshold Ksp ≃ 1.3 · log2 (N ) (computed up to N = 10000): We
know that there exist other random cliques of size K ≤ Kran , however for Ksp < K < Kran the planted solution is
still locally stable under small perturbations: Speaking in terms of the free energy, the planted solution represents a
local minimum, well separated from the paramagnetic one. In the statistical physics language, Ksp corresponds to
the spinodal point for the existence of the planted solution. Below the spinodal of the planted solution, in general
the recovery in the non-planted ensemble is easy [11]. This should imply that for K < Ksp there should exist a
polynomial algorithm able to find a random clique of size K. However, the threshold for the polynomial algorithms
in the non-planted case is believed to be KKarp = log(N ) in the large N limit [10]. Whether the difference between
Ksp and KKarp is just due to finite size effects or has a deeper meaning should be better analyzed and will be the
subject of a subsequent work.
Then, we compare the Bethe free energy fplan and fran of the solutions found respectively from planted and random
initialization. We name as Ks the threshold at which fran = fplan ; it corresponds to a static phase transition, at
which the global minimum of the free energy changes from the planted to the paramagnetic solution. For large N , it
is known that Ks = 2 log2 (N ) [12]. Finite size corrections however are huge and for finite N , Ks (N ) is quite different
from its large N limit: Ks (10000) ≃ 1.6 log2 (N ). The values found for Ks (N ) are in good agreement with the finite
size estimate for the largest size kmax (N ) of a random clique in a random graph of size N . kmax (N ) can be obtained
by the following probabilistic argument: The expected number of cliques of size k in an E-R graph of size N with
(k)
2 . We define k
bonds present with probability p is given by E(N, k) = N
max (N ) as the largest integer k for which
k p
E(N, k) > 1. The largest naturally occurring clique is shown to have with high probability size kmax or kmax + 1 in
graphs with N large [13–15].
Summarizing:
• For K > KBP there is just one minimum of the free energy, that is the one corresponding to the planted clique,
found by randomly initialized BP. We will call this phase an easy phase for the recovery.
• For Ks < K < KBP there are two local minima of f : the global one corresponds to the planted clique, and it
is reached by BP with planted initialization; the other local minimum corresponds to a paramagnetic solution,
that is the one where BP stops if randomly initialized. If we know how to nucleate the planted solution, it is
still possible to find it. We will call this phase a hard phase.
4
• For Ksp < K < Ks the planted solution is still a local minimum of the free energy, that can be found if BP is
initialized around the planted solution. However, the global minimum of f is the paramagnetic one. This phase
corresponds to an impossible phase.
• For K < Ksp the number of random cliques is large and the planted solution is no more stable, it is no more a
minimum of the free energy: even starting near to the planted solution, the algorithm will flow to the unique
paramagnetic minimum.
A similar analysis was performed in ref. [16] for the so-called stochastic block model and in Ref. [5] for the problem
of finding a highly connected subset of vertices in a sparse graph. From the results
p of ref. [5], the dynamical threshold
associated with the failure of BP in the sparse case is shown to reduce to K = N/e in the dense planted √
clique case,
while the thresholds associated with the static and the spinodal phase transition are located at K ∗ = o( N ) in the
dense limit. In statistical physics such a situation is called a first order transition: there are two competing minima
of f ; varying the parameters of the problem, the global one changes from one to another. Such a situation is present
in other recovery problems, a well-known example is the compressed sensing [18].
We will now look to the statistical mechanical model constructed introducing a Hamiltonian associated to the Bayes
probability for the clique problem, following an approach similar to ref. [17]. In statistical physics, the appearance
of metastable minima is encountered in a large number of problems. In these cases, the more efficient algorithm
for the research of the true minimum is the so-called Parallel Tempering (PT) [19], that we will apply to the clique
problem. We define an Hamiltonian associated to the posterior probability of eq. (2) as: P ({x}|{A}) ≡ e−βH , where
the Hamiltonian has the form:
H({x}) = −
X
log(P (xi ))+
i
−
X
(1 + xi xj )
(1 − xi xj )
+ Aij log
(1 − Aij ) log
2
2
ij
The energy assumes the value H = ∞ if Aij = 0 and xi xj = 1, preventing from having configurations without links
between two elements of a clique. We have introduced an additional parameter, the inverse temperature β = T1 , that
takes the value β = 1 in the original problem. Given a realization of the graph {A}, we introduce n = 19 replicas of
the system with the same graph, each replica i is at a different inverse temperature: βi = 1 − i · 0.05, i ∈ [0, 18]. The
i = 0 replica is the original system. For each replica we perform a standard Metropolis Monte Carlo (MC) simulation.
After 5 MC steps for each replica, we try to flip the configuration of the i-th and (i + 1)-th replicas with probability
p = min 1, e(βi −βi+1 )(Ei −Ei+1 ) ,
(5)
where Ei is the
P actual energy value of the i-th replica. We then measure the magnetization M of the original replica
i = 0: M ≡ j xj . When M = K we stop our simulation, having identified the planted clique. If K < KBP , the
original system will be firstly attracted to the secondary minimum of the free energy and should overcome a barrier
to reach the true minimum. However at β < 1, this barrier lowers or eventually it disappears. Replicas at higher
temperature are free to explore a larger part of the phase space in less time. Flipping replicas thus permits to the
original system to reach the true minimum in a smaller time. Being the system fully connected, one could naively
think that the MC algorithm takes O(N 2 ) time: an iteration step is intended as the attempt to flip each of the N
variables, and the computation of the new energy is of order O(N ). However, the proposal to flip a spin is accepted
only O(K) times, because if xi = 0 and we propose to flip it, the flip is accepted only if Aij = 1 for all spins j
with xj = 1, while if xi = 1 it is always possible to flip it. The whole algorithm thus is O(K · N ). In Ref. [5] it is
stated that no local algorithm can find the planted solution for K < KBP . PT is not local because of the flipping
procedure between replicas. It finds the planted solution for K > Ks . As in all first-order phase transitions, the
time for convergence is diverging at the static transition point Ks . The time of convergence seems to diverge as
Nν
t(N, K) ∝ ((K−Ks )/
log2 (N ))a , as shown in Fig. 1. The value for the exponents are ν = 5.78(4), a = 3.64(12). In
the right part of fig. 1 we show the collapse of the data for the convergence time as a function of K for different
sizes N once the proper scaling variables are used. Summarizing, the time of convergence grows as a power law
with N and diverges with a power law at the static threshold Ks . The value for the critical exponents a and ν is
not optimal. In fact, the parameters of PT (number of replicas, spacing in temperature...) can be modified leading
probably to a changing in the exponents. However our purpose is not to optimize the PT but only to show that it
is indeed a polynomial algorithm. For the clique problem, an exhaustive search algorithm can find the planted clique
in a time O(exp(c log2 (N ))). Even if we showed that the collapse of the data with a polynomial scaling is very good,
one could always criticize that the analyzed sizes are too small to capture the difference between a polynomial and
5
1e+12
1000
a
b/(x-x0)
N=2000
N=3000
N=4000
N=5000
1e+11
1e+10
100
t
ν
t/(bN )
1e+09
1e+08
1e+07
1e+06
N=5000
N=4000
N=3000
N=2000
10
1
0.1
100000
10000
0.01
1.6 1.8
2
2.2 2.4 2.6 2.8
K/log2(N)
3
3.2 3.4 3.6
0
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
(K-Ks)/log2(N)
2
FIG. 1: Left: Time of convergence for PT changing the size K of the planted clique, and the size of the graph. Each
point is the average over 20 realizations of the graph. The time of convergence diverges as a power-law at Ks . Right:
Collapse of data for the convergence time once proper rescaled variables are used. The time of convergence grows as
a power law with N .
an exponential O(exp(c log2 (N ))) behavior. For this reason, now we move to a different but connected model: the
planted Independent Set model. Being this problem sparse, the thresholds are sharper. We will show that also for this
problem in the hard region the scaling of the PT time to find solutions is well-fitted by a power-law for the analyzed
sizes. This gives indications to believe that also in the case of the planted clique problem the PT algorithm really
finds solutions in polynomial time.
An Independent Set (IS) is a subset of vertices of a graph that are not connected. It is clear that a clique becomes
an Independent Set on the complementar graph. The IS problem can be safely defined on a sparse graph of degree d.
We define the density of the IS ρ = K
N . For d > 30, one can show that the paramagnetic solution of the BP equations
to find a random IS of size ρ is stable up to density ρl (d) well above the density ρmax (d) of the largest random IS [21].
This means that if we study the planted problem, we expect the existence of a region in the density of the planted IS
ρmax (d) < ρ(d) < ρl (d) where it will be hard to find the planted IS even if we are above the theoretical threshold for
the recovery: We are in a situation analogous to the case of the planted clique of size Ks < K < KBP . The thresholds
for the IS problem as a function of d has been computed in ref. [21].
To be concrete, we plant a IS I of size K on a graph of size N with average degree d in the following way: we
extract link Aij = 1 between nodes i, j ∈
/ I with probability cin , and links between i ∈ I, j ∈
/ I with probability
cout . We do not put links between i, j ∈ I. Imposing that all elements have average degree d (in this way a generic
d
algorithm cannot classify elements on the basis of their degree), we find cin = Nd(1−2ρ)
(1−ρ)2 , cout = N (1−ρ) . The thresholds,
in this case, are slightly different from the ones in ref. [21] that were for graphs of fixed degree, however the behavior
is the same. We restrict ourselves to the case of average degree d = 40. In this case, for fixed degree, following ref.
[21], the thresholds are ρl (40) = 0.138, ρmax (40) = 0.1273, ρs (40) = 0.1231. We write the BP equations following the
same reasoning of the ones for the clique problem, practically the equations are those in Eqs. (26),(27) of ref. [16]
for the general case of clustering. We numerically extract the threshold ρl as the limit for the convergence of the BP
equations to the planted solution once they are randomly initialized: ρl (40) = 0.135(1). Analogously to the clique
problem, we identify the static transition as the threshold at which the free energy of the planted solution, reached by
BP initialized near enough to the planted solution, is equal to the free energy of the paramagnetic solution, reached
by randomly initialized BP, finding ρs (40) = 0.1218. Please note that in the case of sparse IS the static threshold
ρs does not corresponds to the threshold ρmax for the maximal density of random IS, as in the clique problem. We
have that ρs < ρmax . Even if for ρs < ρ < ρmax there exist random ISs, in this region the planted IS dominates the
measure and can be reconstructed. In the limit of d → ∞, ρmax = ρs [21], as happens for the clique problem.
Having identified the important transitions, we run PT in the hard region for the recovery of the planted solution.
As for the clique problem, we write the Hamiltonian associated to the posterior Bayes probability that is essentially
the one in Eq. (8) of ref. [16]. A MC step takes time O(dN ). We introduce n replicas at different inverse temperatures
β: βi = 1 − i · 0.02, i ∈ [0, 18]. We run PT at two different densities: ρ = 0.14 that is in the easy phase, and ρ = 0.13
that is in the hard phase for the reconstruction of the planted solution. PT succeeds to find the planted solution, and
6
1e+07
1e+08
b
ax
ρ=0.13
ρ=0.14
1e+07
PT steps
PT steps
1e+06
100000
10000
b
ax
ρ=0.13
ρ=0.14
1e+06
100000
10000
1000
1000
1000
10000
N
1000
10000
N
FIG. 2: Average time of convergence (left) and 90th percentile time of convergence (right) for PT changing N for
d = 40 and planted IS of density ρ = 0.13 and ρ = 0.14. Times grow as a power-law with N . Each point is the
average over ∼ 103 − 104 realization of the graph.
the times are reported in Fig. 2 as a function of N . For both ρ = 0.13 and ρ = 0.14 we have tried to fit data both
with a polynomial f (x) = axb and an exponential function g(x) = c exp(dx). In both cases, the exponential fit has
to be excluded, while times are well fitted with a polynomial growth. The best-fit parameters for the exponent b are:
b(ρ = 0.14) = 2.5(1), b(ρ = 0.13) = 3.15(9). Also the time of convergence of the 90% of samples grows as a power-law
with N (see Fig. 2).
Concluding, we have applied a standard method for the analysis of disordered models in statistical physics, the
Parallel Tempering algorithm, to the planted clique problem. The performances are quite surprising: it succeeds in
finding the planted solution down to the information Theoretical threshold in a time numerically compatible with a
power-law in the size of the system. Moving to the planted IS sparse problem, that should be harder because thresholds
become sharper, the performances of PT still remains really good, succeeding in finding the correct solution in the
hard region in polynomial time. The PT algorithm is a non-local algorithm because replicas of the system at different
effective temperatures are flipped during the simulation. When looking to the associated statistical mechanical
problem, the hardness of the planted clique or planted IS problems relies on the presence of an extensive barrier in
the free energy landscape between the correct planted solution and a wrong paramagnetic solution. However, when
the temperature is added, this barrier can become smaller and eventually disappears. For this reason, in the PT
algorithm, replicas at higher temperature can explore rapidly a larger space of configurations preventing from being
trapped by the paramagnetic solution. This paper just shows numerical evidence that there could exist a polynomial
algorithm for the clique problem and related problems. A crucial point in the PT algorithm is the choice of the
number of replicas that should satisfy two important properties:
• The β associated to the last replica should be low enough to allow the system to explore the whole phase space,
without trapping barriers.
• The spacing in β should be not so large: in this way, the difference in the energy associated to two near replicas
could be small enough to permit the flipping with a non-null probability in Eq. (5).
For the analyzed sizes we have seen that n = 19 is a good number to have both these properties. One could criticize
that n could grow with N . It could be possible, however for the analyzed sizes n ≃ O(log(N )), that will, however,
lead to a polynomial scaling for the time of convergence of the algorithm. An analytical study of the performance of
thermic algorithms is quite difficult. However, we think it could be of crucial importance for the determination of the
real scaling of PT, and its application also for other optimization problems. The application of PT algorithm to the
largest clique (and largest IS) problem on a Random Regular Graph is currently under study.
I thank Andrea Montanari for introducing me to the planted clique problem and for the suggestion to look at the
sparse Independent Set problem. I thank also Scott Kirkpatrick, Raffaele Marino, Federico Ricci-Tersenghi for very
interesting discussions.
7
[1] M. Jerrum, Random Struct. Algorithms 3 (1992) p.347.
[2] Geoffrey R Grimmett and Colin JH McDiarmid,On coloring random graphs, Mathematical Proceedings of the Cambridge
Philosophical Society, vol. 77, Cambridge Univ. Press, 313–324 (1975).
[3] Barak, Boaz, et al. ”A nearly tight sum-of-squares lower bound for the planted clique problem.” Foundations of Computer
Science (FOCS), 2016 IEEE 57th Annual Symposium on. IEEE,
p 2016.
[4] Y. Deshpande, A. Montanari, ”Finding hidden cliques of size N/e in nearly linear time.” Foundations of Computational
Mathematics 15.4 (2015): 1069-1128.
[5] A. Montanari, J. Stat. Phys. 161, 273 (2015).
[6] Q. Berthet and P. Rigollet, arXiv preprint, arXiv:1304.0828 (2013).
[7] Bruce E. Hajek, Yihong Wu, and Jiaming Xu. Computational lower bounds for community detection on random graphs.
In COLT, 899–928, (2015).
[8] Z. Ma and Y. Wu, Ann. Stat. 43 (2015) p.1089.
[9] T.T. Cai, T. Liang and A. Rakhlin, Ann. Statist. 45, 1403 (2017).
[10] Karp, ”Probabilistic Analysis of Some Combinatorial Search Problems.”, in Algorithms and Complexity: New Directions
and Recent Results, Academic Press, NY 1976.
[11] Florent Krzakala and Lenka Zdeborova, PRL 102, 238701 (2009).
[12] Arias-Castro, Ery, and Nicolas Verzelen. ”Community detection in dense random networks.” The Annals of Statistics 42.3
(2014): 940-969.
[13] D. Matula, On the complete subgraphs of a random graph, Combinatory Mathematics and its. Applications (Chapel Hill,
1970) 356–369
[14] D. W. Matula. The largest clique size in a random graph. Technical report, Department of Computer Science, Southern
Methodist University, 1976
[15] B. Bollobas and P. Erdos, Cliques in random graphs, Math. Proc. Camb. Phil. Soc. 80, 419 (1976).
[16] Decelle, A., Krzakala, F., Moore, C., Zdeborová, L., Asymptotic analysis of the stochastic block model for modular
networks and its algorithmic applications. Phys. Rev. E 84(6), 066106 (2011).
[17] D. Hu, P. Ronhovde and Z. Nussinov, Philosophical Magazine 92, 406 (2012).
[18] F. Krzakala, M. Mézard, F. Sausset, Y. Sun, L. Zdeborová, Phys. Rev. X 2, 021005 (2012).
[19] K. Hukushima and K. Nemoto, J. Phys. Soc. Jpn. 65, 1604 (1996).
[20] A. Braunstein, R. Zecchina, Phys. Rev. Lett. 96, 030201 (2006).
[21] J. Barbier, F. Krzakala, L. Zdeborova, and P. Zhang, The hard-core model on random graphs revisited, J. Phys.: Conf.
Series 473, 012021 (2013)
| 8cs.DS
|
Compositional Invariant Generation via Linear
Recurrence Analysis
arXiv:1502.00138v1 [cs.PL] 31 Jan 2015
Azadeh Farzan and Zachary Kincaid
University of Toronto
Abstract. This paper presents a new method for automatically generating numerical invariants for imperative programs. Given a program, our
procedure computes a binary input/output relation on program states
which over-approximates the behaviour of the program. It is compositional in the sense that it operates by decomposing the program into
parts, computing an abstract meaning of each part, and then composing
the meanings. Our method for approximating loop behaviour is based on
first approximating the meaning of the loop body, extracting recurrence
relations from that approximation, and then using the closed forms to
approximate the loop. Our experiments demonstrate that on verification
tasks, our method is competitive with leading invariant generation and
verification tools.
1
Introduction
Compositional program analyses operate by decomposing a program into parts,
computing an abstract meaning of each part, and then composing the meanings. Compositional analyses have a number of desirable properties, including
scalability, parallelizability, and applicability to incomplete programs. However,
compositionality comes with a price: since each program fragment is analyzed
independently of its context, the analysis cannot benefit from contextual information. This paper presents a compositional method for numerical invariant
generation which, despite loss of contextual information, compares favourably
with leading (non-compositional) verification techniques.
The analysis proposed in this paper aims to compute a transition relation
which over-approximates the behaviour of a given program. The use of transition
relations in compositional analysis (e.g., [23,25,1,21,17,5]) stems from the fact
that they can be composed: for example, consider a program P = P1 ; P2 which
consists of two sub-programs P1 and P2 which are executed in sequence. A transition invariant JP K for P can be computed by computing transition invariants
JP1 K and JP2 K for the subprograms and then taking JP K to be the relational
composition: JP K = {(s, s′′ ) : ∃s′ .(s, s′ ) ∈ JP1 K ∧ (s′ , s′′ ) ∈ JP2 K}.
A crucial question is how to compute abstractions of loops (i.e., loop summaries [17]). Our analysis is based on a classical idea: find recurrence relations for
variables modified in the body of a loop, and then use the closed forms for these
recurrences as the abstraction of the loop. The focus of research on recurrence
analysis has mainly been on computing the exact behaviour of a (necessarily)
limited class loops, e.g. loops where the body is a sequence of affine assignments
(see Section 6 for a discussion of related literature). We shift the goal to computing over-approximate behaviour of arbitrary loops. The main novelty of our
approach is to make synergistic use of recurrence analysis and compositionality:
on one hand, recurrence analysis can be used to compute accurate transition
formulas for loops; on the other hand, transition formulas for loop bodies can be
mined for recurrence relations to enable recurrence analysis.
Compositionality enables using recurrence analysis for arbitrary loops in two
ways. First, the fact that the transition formula for a loop is computed from
a transition formula for its body makes the control structure of the loop irrelevant (e.g., whether it is a sequence of assignments or contains branching or
nested loops – its transition formula is just a formula). Second, having access
to a loop body formula when computing a loop summary opens the door to
using Satisfiability Modulo Theories (SMT) solvers to extract a broad range
semantic recurrences. In particular, our analysis is able to exploit approximate
recurrences (inequations over linear terms) to compute interesting loop invariants even for variables which do not satisfy recurrence equations in the classical
sense, thus extending the applicability of recurrence-based invariant generation
and overcoming a major barrier in its practical use.
In summary, this paper presents a compositional method for generating numerical invariants (polynomial inequalities of unbounded degree among integer
and rational variables) for programs. The main technical contributions are as
follows.
1. We give a method for computing abstractions of loops using summaries for
their bodies. This allows our analysis to apply to arbitrary code (with nested
loops, unstructured loops, and arbitrary branching). It also makes it possible
to use SMT solvers to extract semantic recurrence relations rather than
syntactic recurrences obtained by pattern-matching source code.
2. We identify a class recurrence (in)equations that can be efficiently extracted
from loop bodies using SMT solving technology and solved using simple
linear algebra.
3. We give a linearization algorithm which enables tractable (but necessarily
approximate) reasoning about non-linear formulas over rationals and integers
(Section 4).
4. We collect ideas from a diverse range of sources (including algebraic program
analysis [10], recurrence analysis [1,15,2], linearization [20], and symbolic
abstraction [26,22,19]), and synthesize them into a cohesive presentation
which can be used as a foundation for futher research on recurrence analysis.
We implemented linear recurrence analysis and used it to verify assertions for
a suite of benchmarks. Linear recurrence analysis is able to prove the correctness
of more benchmarks in this suite than any of the leading verification tools for
integer programs.
r := x // remainder
q := 0 // quotient
while(r >= y):
// subtract y from r
t := y
while(t != 0)
r := r - 1
t := t - 1
q := q + 1
assert(x = q*y + r)
v entry
r := x
v1
q := 0
v2
[r < y]
assert(x = q*y+r)
v exit
q := q + 1
[r >= y]
v3
v8
t := y
v4
[t != 0]
v7
[t = 0]
t := t - 1
v5
v6
r := r - 1
(a) Program text
(b) Flow graph
Fig. 1. An integer division program, computing a quotient and remainder. Statements
of the form [ψ] represent assumptions; i.e., statements which block if ψ does not hold.
2
Overview
We will adopt a simple intraprocedural model in which a program is represented
by a control flow automaton (CFA) where edges are labeled by program statements. Figure 1 depicts such a CFA for a program which computes the quotient
and remainder of division of a variable x by a variable y. We use this model for
the sake of simplicity and to help keep the presentation of our analysis short and
self-contained. We hope that the basic idea behind the extension to procedures
(implemented in the tool), using the analysis to compute procedure summaries
[29], is clear without formal explanation.
Our analysis, linear recurrence analysis (LRA), is presented in the algebraic
framework described in [10]. Suppose that we wish to prove that the assertion
assert(x = q*y + r) always succeeds. We begin by computing the set of paths
from v entry to v8 (the location corresponding to the assert statement in the
CFA). This set of paths is represented by a path expression for the vertex v8 ,
which is a regular expression over an alphabet of control flow edges. In principle,
this can be accomplished by Kleene’s well-known algorithm for converting a
finite automaton into a regular expression [14] (but more efficient algorithms
exist [30]). For example, the following is a path expression for v8 :
Inner loop
hv
entry
z
}|
{
∗
, v1 i·hv1 , v2 i· hv2 , v3 i·hv3 , v4 i· (hv4 , v5 i·hv5 , v6 i·hv6 , v4 i)∗ ·hv4 , v7 i·hv7 , v2 i ·hv2 , v8 i
|
{z
}
Outer loop
Once we have a path expression representing the paths to v8 , we compute an
over-approximation of the executions to v8 by evaluating the path expression in
some abstract domain. The main benefit of this algebraic framework is that an
analysis is defined simply by providing an interpretation for each of the regular
expression operators (sequencing, choice, and iteration, corresponding to the
control structures of structured programs), and then we may rely on a path
expression algorithm ([14,30]) to efficiently “lift” the analysis to programs with
arbitrary control flow.
Formally, a program analysis (in the framework of [10]) is defined by an interpretation, which consists of a semantic algebra and a semantic function. A
semantic algebra consists of a universe which defines the space of possible program meanings, and sequencing, choice, and iteration operators, which define
how to compose program meanings. A semantic function is a mapping from control flow edges to elements of the universe which defines the meaning of each
control flow edge. A path expression is evaluated by interpreting the individual
edges using the semantic function, and interpreting the regular expression operators using the corresponding operators of the semantic algebra (to compose the
interpretations of individual edges into interpretations of sets of program paths).
Keeping this overall algorithm in mind, we proceed to describe the interpretation which defines linear recurrence analysis.
LRA Universe. The semantic universe of LRA (i.e., the space of program
meanings) is the set of (not necessarily linear) arithmetic transition formulas.
If we let Var denote the set of program variables and Var′ the set of “primed”
copies of program variables, then a transition formula is an arithmetic formula
with free variables in Var ∪ Var′ . Such a formula represents an input/output
relation between program states.
LRA Semantic Function. The semantic function J·K is a function that maps
each edge of a control flow automaton to its interpretation as a transition formula. For example (again, considering Figure 1), we have
Jhv entry , v1 iK = r′ = x ∧ stable({q, t, x, y})
Jhv1 , v2 iK = q ′ = 0 ∧ stable({r, t, x, y})
Jhv2 , v3 iK = r > y ∧ stable({q, r, t, x, y})
V
′
where for X ⊆ Var, we have stable(X) ,
x∈X x = x ; we use this to factor
out equalities from the formulas and make them more legible. Boxes around
formulas have no meaning, and are used only to make it easier to distinguish
between equalities in formulas and the meta-language.
LRA Operators. The sequencing and choice operators of our analysis are
defined as follows:
ϕ ⊙ ψ = ∃x′′ .ϕ[x′′ /x′ ] ∧ ψ[x′′ /x]
Sequencing
ϕ⊕ψ =ϕ∨ψ
Choice
(where ϕ[x′′ /x′ ] denotes ϕ with each primed variable x′ replaced by its doubleprimed counterpart x′′ , and ψ[x′′ /x] similarly replaces unprimed variables with
double-primed variables).
The semantic function, sequencing, and choice operators are sufficient to
analyze loop-free code. For example, we may consider how LRA computes a
transition invariant for the body of the inner loop of Figure 1:
Jhv4 , v5 i · hv5 , v6 iK = Jhv4 , v5 iK ⊙ Jhv5 , v6 iK
= t > 0 ∧ r′ = r − 1 ∧ stable({q, t, x, y})
Jhv4 , v5 i · hv5 , v6 i · hv6 , v4 iK = Jhv4 , v5 i·hv5 , v6 iK ⊙ Jhv6 , v4 iK
= t > 0 ∧ r′ = r − 1 ∧ t′ = t − 1 ∧ stable({q, x, y})
The final step in describing our analysis is to provide a definition of the
iteration operator (⍟) of LRA. The idea behind the definition of the iteration
operator is to use an SMT solver to extract recurrence relations from the loop
body, and then use the closed form of these recurrences for the abstraction of
the loop. We explain this in detail in Section 3. Here, we illustrate how LRA
works on the running example to provide some intuition on the analysis.
After computing a formula ϕinner representing Recurrence Closed form
the body of the inner loop (as given above), we r′ = r − 1 r(k) = r(0) − k
apply the iteration operator ⍟ to compute a for- t′ = t − 1 t(k) = t(0) − k
mula representing any number of executions of
the inner loop. The iteration operator begins by extracting the recurrence equations shown to the right. It then computes closed forms for these recurrences,
also shown to the right (where x(k) denotes the value that the variable x takes
on the kth iteration of the loop). Note that this table omits “uninteresting” recurrences (such as q ′ = q + 0) which indicate that a variable does not change in
a loop. These closed forms are used to abstract the loop as follows:
′
′
ϕ⍟
inner = ∃k.k ≥ 0 ∧ r = r − k ∧ t = t − k ∧ stable({q, x, y})
= r′ = r + t′ − t ∧ t′ ≤ t ∧ stable({q, x, y})
We may use this summary ϕ⍟
inner for the inner loop to compute a transition
formula representing the body of the outer loop:
ϕouter = Jhv2 , v3 iK ⊙ Jhv3 , v4 iK ⊙ ϕ⍟
inner ⊙ Jhv4 , v7 iK ⊙ Jhv7 , v2 iK
= q ′ = q + 1 ∧ r′ = r + t′ − y ∧ t′ = 0 ∧ r ≥ y ∧ stable({x, y})
We then apply the iteration operator to Recurrence Closed form
compute a transition formula for the outer q ′ = q + 1 q (k) = q (0) + k
loop. The recurrences found for the outer loop r′ = r − y r(k) = r(0) − y (0) k
and their closed forms are shown to the right
(again, with “uninteresting” recurrences omitted). We note that our algorithm
extracts these recurrences from ϕouter using only semantic operations: the fact
that ϕouter is an abstraction of a looping computation is completely transparent to the analysis. Using the closed forms of the recurrences to the right, we
compute the following transition formula for the outer loop:
′
′
ϕ⍟
outer = ∃k.k ≥ 0 ∧ q = q + k ∧ r = r − ky ∧ stable({x, y})
= q ′ ≥ q ∧ r′ = r − (q ′ − q)y ∧ stable({x, y})
Finally, we compute a transition formula which approximates all executions
which end at v8 as follows:
ϕP = Jhv entry , v1 i · hv1 , v2 iK ⊙ ϕ⍟
outer ⊙ Jhv2 , v8 iK
= q ′ ≥ 0 ∧ r′ = x − q ′ y ∧ r ≤ y ∧ stable({x, y})
This formula is strong enough to imply that assertion x′ = q ′ ∗ y ′ + r′ holds
at v8 . This is particularly interesting because it requires proving a non-linear
transition invariant for the loop, which is out of scope for many state-of-the-art
program analyzers.
3
Abstracting Loops with Linear Recurrence Analysis
In this section, we describe the iteration operator of linear recurrence analysis.
Suppose that we have a formula ϕbody which approximates the behaviour of the
body of a loop. Our goal is to compute a formula ϕ⍟
body which represents the
effect of zero or more executions of the loop body. Our iteration operator works
by extracting recurrence relations from the formula ϕbody and then computing
closed forms for these relations. We present our iteration operator in three stages,
based on the types of recurrence relations being considered: simple recurrence
equations, stratified recurrence equations, and linear recurrence (in)equations.
Simple and stratified recurrences are classical classes of recurrence equations.
Linear recurrence (in)equations generalize the class of inequations presented in
[2] by using stratified recurrences to generate polynomial (rather than just linear)
inequations. The main conceptual contribution of this section is the idea to use
SMT solvers to extract recurrences (and other relevant information) from a loop
body formula.
In the remainder of this section, we fix a formula ϕbody representing the body
of a loop. We assume that ϕbody is expressed in linear (rational and integer)
arithmetic; our strategy for dealing with non-linear arithmetic is described in
Section 4. We
that ϕbody is satisfiable (if it is not, then we can take
V also assume
′
x
=
x,
which
represents zero iterations of the loop).
ϕ⍟
to
be
x∈Var
body
3.1
Simple recurrence equations
We start by defining simple recurrences and induction variables.
Definition 1. A simple recurrence for a formula ϕ is an equation of the form
x′ = x + c (for a constant c) such that ϕ |= x′ = x + c. If x′ = x + c is a simple
recurrence for ϕ, we say that x satisfies the recurrence x′ = x + c, and if there
is some c such that x satisfies the recurrence x′ = x + c, we say that x is an
induction variable.
Simple recurrences can be detected by first querying an SMT solver for a model
m of ϕbody , and then asking whether ϕbody implies x′ = x + Jx′ − xKm (where
Jx′ − xKm denotes the interpretation of the term x′ − x in the model m). This
implication holds iff x is an induction variable.
If x is an induction variable that satisfies the recurrence x′ = x + c, then the
closed form for x is x(k) = x(0) + kc (writing x(k) for the value that x obtains on
the kth iteration of the loop). To provide some early intuition on the iteration
operator to be developed in the remainder of this section, let us suppose that
we are only interested in simple recurrences. Then a possible definition for the
iteration operator is
V ′
′
ϕ⍟
body , ∃k ≥ 0. {x = x + kc : x = x + c ∈ SR(ϕbody )}
where SR(ϕbody ) is the set of simple recurrences satisfied by ϕbody .
The iteration operator defined above is sound (it over-approximates the behaviour of any number of iterations of the loop, since each variable is either
described exactly by a recurrence or is not constrained at all), but it is imprecise. The remainder of this section discusses more general recurrence equations
which can be used to compute more precise transition invariants for loops.
3.2
Stratified recurrences equations
Consider the loop shown to the right. We can see that while(x ≤ 10):
x := x + 1
x satisfies a simple recurrence equation x′ = x + 1, and
y := y + x
that y satisfies a (non-simple) recurrence equation y′ =
z := 2 * x
y + x + 1. A closed form for y’s recurrence is y (k) = y(0) +
Pk−1 (i)
′
(x
+1).
Since
x
satisfies
a
simple
recurrence
(x
=
x+1),
we have a closed
i=0
form for x(i) , so we may simplify this recurrence and remove the summation:
y(k) = y(0) +
k−1
k−1
X
X
k(k + 1)
i = y(0) + kx(0) +
(x(0) + i + 1) = y(0) + kx(0) + k +
.
2
i=0
i=0
Stratified recurrence equations generalize this idea: starting from simple recurrence equations, we solve more and more complicated recurrences using the
closed forms for simpler ones. As with the example above, stratified recurrences
have non-linear closed forms. Non-linear invariant generation is not the main
focus of our work, but it is sometimes a necessary intermediate step for proving linear invariants in a compositional setting: since our analysis cannot take
advantage of contextual information when analyzing a loop, we generate a nonlinear invariant and then, after the analysis has examined more context, simplify
it (using the linearization algorithm from Section 4).
Definition 2. Let ϕ be a formula. The stratified recurrence equations (and stratified induction variables) of ϕ are defined inductively as:
– A simple recurrence equation which is satisfied by ϕ is a stratified recurrence equation of ϕ (and a simple induction variable is a stratified induction
variable) at stratum 0.
– Let y denote a vector of the stratified induction variables of strata ≤ N . A
recurrence of the form x′ = x + cy (where c is a vector of constants) is a
stratified recurrence at stratum N + 1 (and if x satisfies such a recurrence,
it is a stratified induction variable at stratum N + 1).
We use siv(ϕ) to denote the set of all stratified induction variables of ϕ.
Let us now discuss how stratified recurrences are detected from a loop body
formula ϕbody . We begin by computing the affine hull aff(ϕbody ) of ϕbody (Algo-
Algorithm 1: Affine hull.
Input : Satisfiable formula ϕbody
Output: Affine hull of ϕbody
H ← ⊥; ψ ← ϕbody ;
while there
V exists a model m of ψ do
H ′ ← {x = JxKm : x ∈ Var ∪ Var′ };
H ← H ⊔= H ′ ;
/*Join in the domain of linear equalities*/
ψ ← ψ ∧ ¬H;
end
return H
rithm 1).1
Definition 3. The affine hull aff(ϕ) of a formula ϕ is the smallest affine set
which contains ϕ, represented
as (the set
of solutions to) a system of equations
Ax = b, where x = x1 · · · xn x′1 · · · x′n . Logically, aff(ϕ) is a system of equations which satisfies the following three properties: (1) ϕ |= aff(ϕ), (2) every
linear equation over Var ∪ Var′ which is implied by ϕ is also implied by aff(ϕ),
and (3) no equation in aff(ϕ) is implied by the others.
Our strategy for detecting stratified recurrences is based on the following
lemma. Combined with property (2) of aff(ϕbody ) above, this lemma implies
that any equation implied by ϕbody can be expressed as a linear combination of
the equations in aff(ϕbody ).
Lemma 1 ([28], Corollary 3.1d). Let A be a matrix, b be a column vector,
c be a row vector, and d be a constant. Assume that the system Ax = b has
a solution. Then Ax = b implies cx = d iff there is a row vector λ such that
λA = c and λb = d.
Let us write aff(ϕbody ) as Ax = b. Suppose that we have detected all recurrences of strata < N , and that we want to determine whether a variable xi
(0 ≤ i ≤ n) is an induction variable at stratum N . Then we ask whether there
exists λ, c, and d such that:
– λA = c and λb = d (i.e, cx = d is implied by aff(ϕbody ) and thus by ϕbody )
– ci = 1 and ci+n = −1 (the coefficients of xi and x′i are 1 and -1, respectively)
– For all j such that j 6= i + n and n ≤ j ≤ 2n, cj = 0 (except for x′i , all
coefficients of primed variables are 0).
– For all j such that j 6= i such that xj is not an induction variable of strata
< N and n ≤ j ≤ 2n, cj = 0 (except for xi and induction variables of strata
< N , all coefficients for unprimed variables are 0).
Thus, after computing the affine hull of ϕbody , determining whether a given
variable satisfies a stratified recurrence is simply a matter of solving a system of
linear equations (e.g., using Gaussian elimination).
1
This algorithm is a specialization of the one in [26] to the abstract domain of linear
equalities.
Closed forms for stratified recurrences. We first state a lemma:
Lemma 2. The closed form for a stratified induction variable of strata N is of
the form
(0)
x(k) = p0 (k) + p1 (k)y1 +· · · + pn (k)yn(0)
where each yi is a stratified induction variable of strata < N and each pi (k) ∈
Q[k] is a polynomial of one variable with rational coefficients.
Our algorithm for solving stratified recurrences is based on a constructive
proof for this lemma. We proceed by induction on strata. The base case is trivial.
Suppose that we have a recurrence at strata N (and all y1 , ..., yn are of strata
Pk−1
(i)
< N ): x′ = x+c1 y1 +· · ·+cn yn +b. Then we may write x(k) = x(0) + i=0 c1 y1 +
(i)
(i)
· · · + cn yn + b . By our induction hypothesis, each yj can be written as a linear
term with coefficients from Q[k]. It follows that there exists p0 , ..., pn ∈ Q[k] so
that
(i)
(0)
c1 y1 +· · · + cn yn(i) + b = p0 (i) + p1 (i)y1 +· · · + pn (i)yn(0)
Thus we have
x(k) = x(0) +
k−1
X
(0)
p0 (i) + p1 (i)y1 +· · · + pn (i)yn(0)
i=0
= x(0) +
k−1
X
i=0
(0)
p0 (i) + y1
k−1
X
i=0
p1 (i) +· · · + yn(0)
k−1
X
pn (i)
i=0
The closed form of a summation of a polynomial of degree m is a polynomial
of degree m + 1. We can find this polynomial via curve fitting (i.e., we compute
the first m + 1 terms of the summation and then solve the corresponding linear
system of equations for the coefficients of the polynomial).
3.3
Linear recurrence (in)equations
Recurrence equations (such as the simple and while(x ≥ 0 ∧ y ≥ 0):
stratified varieties) yield very accurate approxiif (*): x := x - 1
mations for some variables, but what about varielse: y := y - 1
ables which do not satisfy any recurrence equation? For example, consider that neither x nor y satisfy a recurrence equation in
the loop to the right. However, they do satisfy recurrence inequations: x−1 ≤ x′ ,
x′ ≤ x, y − 1 ≤ y′ , and y′ ≤ y. These inequations can be closed to yield
x(0) −k ≤ x(k) and x(k) ≤ x(0) , y (0) −k ≤ y (k) , and y (k) ≤ y (0) . In this section, we
discuss linear recurrence (in)equations, which allow us to compute good approximations for loops that cannot be completely described by recurrence equations.
Definition 4. A linear recurrence (in)equation of a formula ϕ is an (in)equation
which is implied by ϕ and which is of the form
cx′ ⊲⊳ cx + by + d
where ⊲⊳ ∈ {<, ≤, =}, x is any vector of variables, y is a vector of stratified
induction variables in ϕbody , c, b are constant vectors, and d is a constant.
Linear recurrence (in)equations generalize recurrence equations in two ways:
first, they allow for inequalities rather than equations. Second, they allow recurrences for linear terms, rather than just variables. For example, the linear
recurrence equation (x′ + y′ ) = (x + y) + 1 is satisfied by the body of the loop
above, which can be closed to yield (x(k) + y(k) ) = (x(0) + y(0) ) + k.
We now describe our method for detecting and solving linear recurrence
(in)equations. We begin by introducing a set of difference variables δx , one for
each variable x ∈
/ siv(ϕbody ) (variables which do belong to siv(ϕbody ) are already
precisely described by recurrence equations, so we need not approximate them).
We then compute (via Algorithm 2) the convex hull of the formula ψ defined as:
^
ψ , ∃X.ϕbody ∧ {δx = x′ − x : x ∈ Var \ siv(ϕbody )}
where X is Var′ ∪ (Var \ siv(ϕbody )).
Algorithm 2: Convex hull.
Input : Satisfiable formula ψ, set of variables X
Output: Convex hull of ∃X.ψ
P ← ⊥;
while there exists a model m of ψ do
Let Q be a cube of the DNF of ψ s.t. m |= Q;
Q ← project(Q, X) ;
/*Polyhedral projection*/
P ←P ⊔Q ;
/*Polyhedral join*/
ψ ← ψ ∧ ¬P ;
end
return P
Geometrically, the convex hull hull(ϕbody ) is the smallest convex polyhedron
which contains ϕbody . Logically, it is a set of (in)equations such that (1) every
(in)equation in hull(ϕbody ) is implied by ϕbody , and (2) any linear (in)equation
(over Var ∪ Var′ ) which is implied by ϕbody is also implied by hull(ϕbody ). For
example, hull(ϕbody ) for the loop above is:
0 ≤ δx ∧ δx ≤ 1 ∧ 0 ≤ δy ∧ δy ≤ 1 ∧ δx + δy = 1
We note that the only variables which appear in the (in)equations in hull(ϕbody )
are (stratified) induction variables and difference variables. Thus, we may write
any (in)equation in hull(ϕbody ) as cδ ⊲⊳ by +d (where δ is the vector of difference
variables, y is the vector of stratified induction variables, c and b are constant
vectors, and d is a constant). Recalling the definition of the difference variables,
we may rewrite such an inequation as c(x′ − x) ⊲⊳ by + d and then rewrite
again as cx′ ⊲⊳ cx + by + d, which matches the definition of linear recurrence
(in)equations given in Definition 4.
We may close such a linear recurrence (in)equation as follows:
k−1
X
by (i) + d
cx(k) ⊲⊳ cx(0) +
i=0
We can compute a closed form for the summation
preceding section.
3.4
Pk−1
i=0
by (i) + d as in the
Loop guards
A loop body typically contains crucial information about the execution of the
loop that cannot be captured by recurrence relations. For example, consider the
loop in Section 3.2. Supposing that the loop executes n times, we must have that
x(k) ≤ 10 for each k < n. Further, consider that the variable z is a function of
the simple induction variable x, and so z(k) can be described precisely in terms
of the pre-state variables (even though it does not itself satisfy any recurrence):
(
z(0)
if k = 0
(k)
z =
2(x(0) + k + 1) otherwise.
The question is: how can we recover this type of information from a loop body
formula?
We define the guard of a transition formula ϕ as follows:
guard(ϕ) , (∃Var.ϕ) ∧ (∃Var′ .ϕ)
If ϕ is a loop body formula, then guard(ϕ) is a formula which over-approximates
the effect of executing at least one execution of the loop. Intuitively, (∃Var.ϕ) as
a precondition that must hold before every iteration of the loop and (∃Var′ .ϕ)
as a post-condition of the loop that must hold after each iteration.
Consider again the example loop in Section 3.2, we have the following loop
body formula
ϕbody = x ≤ 10 ∧ x′ = x + 1 ∧ y ′ = y + x′ ∧ z ′ = 2x′
We compute guard(ϕbody ) as follows:
guard(ϕbody ) = (∃x, y, z.ϕbody ) ∧ (∃x’, y’, z’.ϕbody )
≡ (x ≤ 10) ∧ (x’ ≤ 11 ∧ z = 2x’) ,
and thereby recover the desired information about x and z.
Since loop body formulas may be large, it may be adventageous in practice
to simplify the guard formula by eliminating the quantifiers (as we did above).
A second option, which is more efficient but less precise, is to over-approximate
quantifier elimination. Two possibilities are to use Algorithm 2 to compute the
convex hull of guard(ϕbody ), or to use optimization modulo theories [19] to compute intervals for each pre- and post-state variable in ϕbody .
3.5
Bringing it all together
We close this section by describing how the pieces defined in this section fit into
the iteration operator of linear recurrence analysis. We let CR(ϕbody ) denote
the set of closed linear recurrence (in)equations (including simple and stratified
recurrence equations) satisfied by ϕbody . Each such (in)equation is of the form
cx(k) ⊲⊳ t, where the free variables of t are drawn from {x(0) : x ∈ Var} and a
distinguished variable k ∈
/ Var indicating the loop iteration. We define
V
ϕ+
{cx′ ⊲⊳ t[x(0) 7→ x] : cx′ ⊲⊳ t ∈ CR(ϕbody )}
body , ∃k.k ≥ 1 ∧
where t[x(0) 7→ x] denotes the term t with every variable of the form x(0) is
replaced by the corresponding variable x.
Finally, our iteration operator is defined as:
V
+
′
ϕ⍟
body , (ϕbody ∧ guard(ϕbody )) ∨
x∈Var x = x.
4
Linearization
The iteration operator presented in the previous section relies heavily on using
an SMT solver to extract information from loop body formulas. This strategy
requires that loop body formulas are expressed in a decidable theory which is
supported by SMT solvers (in particular, linear arithmetic). However, a program may contain non-linear instructions, and even if it does not, our iteration
operator may introduce non-linearity (consider Example 1, where the transition
′
′
formula for the outer loop ϕ⍟
outer contains the non-linear proposition r = x−q y).
Our solution to this problem is to linearize non-linear formulas before passing
them to the iteration operator.
Linearization is an operation that, given an (arbitrary) arithmetic formula
ϕ, computes a formula lin(ϕ) which over-approximates ϕ (i.e., ϕ ⇒ lin(ϕ)), but
which is expressed in linear arithmetic. There is generally no best approximation
of a non-linear formula as a linear formula, so our method is (necessarily) a
heuristic.
We explain our linearization algorithm informally using an example. Consider
the following non-linear formula (where w, x, y, z are integers):
ψ ,1≤w =x<y <5∧w∗y ≤z ≤x∗y
Our algorithm begins by normalizing ψ, separating it into a linear part and a set
of non-linear equations (introducing existentially quantified temporary variables
as necessary). For example, the result of normalizing ψ is:
1 ≤ w = x < y < 5∧ ≤ γ0 ≤ z ≤ γ1 ∧ γ0 = w ∗ y ∧ γ1 = x ∗ y
The left conjunct is a linear over-approximation of ψ, but it is very imprecise:
semantically equal (but syntactically distinct) non-linear terms become semantically unequal in the over-approximation, and all information about the magnitude of non-linear terms is lost. To increase precision of this approximation,
we use two strengthening steps.
1. We replace the non-linear operations with uninterpreted function symbols
and then compute the affine hull of the resulting formula to infer equalities
between non-linear terms. For our example ψ, the we discover that γ0 = γ1 .
2. We compute concrete and symbolic intervals for non-linear terms. Consider
γ0 = x ∗ y from our example ψ. We first compute concrete (x ∈ [1, 3] and
y ∈ [2, 4]) and symbolic (x ∈ [x, x] and y ∈ [y, y]) intervals for the operands
x and y, using symbolic optimization [19] to compute the concrete intervals.
We obtain a concrete interval for x ∗ y (x ∗ y ∈ [2, 12]) by multiplying the
concrete intervals of its operands. We obtain symbolic intervals for x ∗ y
(x ∗ y ∈ [y, 3y] and x ∗ y ∈ [2x, 4x]) by multiplying the concrete interval
for x by the symbolic interval for y and vice-versa. As a result of interval
computation, we discover: 2 ≤ γ1 ≤ 12 ∧ y ≤ γ1 ≤ 3y ∧ 2x ≤ γ1 ≤ 4x
Finally, we take lin(ψ) to be the initial coarse linear approximation of ψ conjoined
with the facts discovered by the two strengthening steps.
We expect linearization to have broad applications outside of the context in
which we presented it, particularly in program analysis, where over-approximation
can be tolerated but non-linear terms cannot. Finding improved linearization
heuristics is an interesting direction of future work.
5
Experiments
We wrote a tool which implements LRA and analyzes C code (using the CIL
[24] frontend).2 We use Z3 [9] to resolve SMT queries that result from applying
the iteration operator and checking assertion violations. Polyhedra operations
are passed to the New Polka library implemented in Apron [4]. The quantifier
elimination algorithm from [22] is used to compute loop guards.
We tested two different configurations of LRA: one which is fully compositional (LRA-Comp) and does not take advantage of contextual information,
and one (LRA) which uses an intraprocedural polyhedron analysis [8] to gain
some contextual information, but which is otherwise compositional. We compare
LRA’s performance against the state-of-the-art invariant generation and verification tools CPAChecker (overall winner of the 2015 Software Verification
Competition) and SeaHorn (winner of the loops category among tools which
are sound for verification).
To evaluate the precision of LRA we used it to verify the correctness of a
suite of 119 small loop benchmarks of varying difficulty. Our benchmark suite
was drawn from the loops category of the 2015 Software Verification Competition
(SVComp-15), as well as a set of non-linear benchmarks (Non-linear), such as the
one in Figure 1. The results for the 81 safe, integer-only benchmarks from these
suites are shown in Table 1. The suite also contains 38 unsafe benchmarks: LRA
and LRA-Comp have no false negatives on these benchmarks; CPAChecker
has 3 and SeaHorn has 2.
Our results demonstrate that LRA is an effective invariant generation algorithm. Even the fully compositional variant of LRA (LRA-Comp) is able to
prove safety for 80% of the benchmarks we considered). We also note that there
are 8 benchmarks for which LRA can prove safety but which CPAChecker and
SeaHorn cannot.
2
The tool and benchmarks are available at http://cs.toronto.edu/~ zkincaid/lra.
Benchmark suite # Bench
LRA
LRA-Comp CPAChecker SeaHorn
SVComp-15
74
65
60
37
65
Non-linear
7
6
5
1
3
Total
81
71 (88%) 65 (80%) 38 (47%) 68 (85%)
Running time across all benchmark suites
Mean
5.4s
3.0s
42.4s
37.7s
Median
0.8s
0.8s
1.6s
0.2s
Table 1. Experimental results.
6
Related work
There is a great deal of work on compositional invariant generation and acceleration which is related to the technique described in this paper. In this section,
we compare our technique to a sampling of this work.
Recurrence analysis. The idea of using closed forms of recurrence relations to
approximate loops has appeared in a number of other papers. Generally speaking, our work differs from previous work in two essential ways: first, we use an
SMT solver to extract semantic recurrences, rather than syntactic recurrences.
Second, we consider approximate recurrences (inequations over linear terms)
rather than exact recurrences (equations over variables). A survey of some of
this work follows.
Ammarguellat and Harrison present a method for detecting induction variables which is compositional in the sense that it uses closed forms for inner loops
in order to recognize nested recurrences [1]. Maps from variables to symbolic
terms (effectively a symbolic constant propagation domain) is used as the abstract domain. Kovács presents a technique for discovering invariant polynomial
equations based on solving recurrence relations [15]. The simple and stratified
recurrence equations considered in this paper are a strict subset of the recurrences considered in [15], but our algorithm for solving recurrences is simpler.
Kroening et al. [16] presents a technique for computing under -approximations
of loops which uses polynomial curve-fitting to directly compute closed forms
for recurrences rather than extracting recurrences and then solving them in a
separate step.
Ancourt et al. present a method for computing recurrence inequations for
while loops with affine bodies [2]. Like the method we present on Section 3.3,
their method is based on using difference variables and polyhedral projections.
Our method generalizes this work by (1) extending it to arbitary control flow,
with (possibly non-linear) formulas as bodies rather than affine transformations,
(2) integrating recurrence inequations with stratified induction variables, thereby
allowing enabling the computation of invariant polynomial inequations. Ancourt
et al. briefly discuss a method for computing invariant polynomial inequations,
but it is based on higher-order differences rather than stratified recurrence inequations. For example, in Figure 1, the analysis discussed in [2] would be able
to prove that r is decremented by a constant amount at every loop iteration,
but could not prove that the constant amount is exactly y.
Acceleration. Acceleration is a technique closely related to recurrence analysis
that was pioneered in infinite-state model checking [6,11,3], and which has recently found use in program analysis [12,18,13]. Given a set of reachable states
and an affine transformation describing the body of a loop, acceleration computes
an exact post-image which describes the set of reachable states after executing
any number of iterations of the loop (although there is recent work on abstract
acceleration uses computes over-approximate post-images [12,13]). In contrast,
our technique is approximate rather than exact, and computes loop summaries
rather than post-images. A result of these two features is that our analysis to
be applied to arbitrary loops, while acceleration is classically limited to simple
loops where the body consists of a sequence of assignment statements.
Compositional program analysis. Compositional program analysis has a
long history. Particular examples are interprocedural analyses based on summarization [29] and elimination-style dataflow analyses (a good overview of which
can be found in [27]). The following surveys recent work on compositional analysis for numerical invariants.
Kroening et al. [17] and Biallas et al. [5] present compositional analysis techniques based on predicate abstraction. In addition to predicate abstraction, there
are a few papers which use numerical abstract domains for compositional analysis. These include an algorithm for detecting affine equalities between program
variables [23], an algorithm for detecting polynomial equalities between program
variables [7], a disjunctive polyhedra analysis which uses widening to compute
loop summaries [25], and a method for automatically synthesizing transfer functions for template abstract domains using quantifier elimination [21]. Our abstract domain is the set of arbitrary arithmetic formula, which is more expressive
than these domains, but which (as usual) incurs a price in performance. It would
be interesting to apply abstractions to our formulas to improve the performance
of our analysis.
Linearization. Our linearization algorithm was inspired by Miné’s procedure
for approximating non-linear abstract transformers [20]. Miné’s procedure abstracts non-linear terms by linear terms with interval coefficients using the abstract value in the pre-state to derive intervals for variables. Our algorithm abstracts non-linear terms by sets of symbolic and concrete intervals, and applies
to the more general setting of approximating arbitrary formulas.
7
Conclusion
This paper presents a fully compositional algorithm for generating numerical
invariants of imperative programs. Our method for abstracting loops makes essential use of compositionality: we assume that we are given a formula which
approximates the body of a loop, and we use an SMT solver to extract recurrence relations and then use the closed forms of these recurrences to approximate
the loop. We have demonstrated experimentally that our method is competitive
with leading invariant generation and verification tools.
References
1. Z. Ammarguellat and W. L. Harrison, III. Automatic recognition of induction
variables and recurrence relations by abstract interpretation. PLDI, pages 283–
295, 1990.
2. C. Ancourt, F. Coelho, and F. Irigoin. A modular static analysis approach to affine
loop invariants detection. Electron. Notes Theor. Comput. Sci., 267(1):3–16, Oct.
2010.
3. S. Bardin, A. Finkel, J. Leroux, and P. Schnoebelen. Flat acceleration in symbolic
model checking. In ATVA, pages 474–488. 2005.
4. J. Bertrand and A. Miné. Apron: A library of numerical abstract domains for
static analysis. In CAV, pages 661–667, 2009.
5. S. Biallas, J. Brauer, A. King, and S. Kowalewski. Loop leaping with closures. In
SAS, pages 214–230, 2012.
6. B. Boigelot and P. Wolper. Symbolic verification with periodic sets. In CAV, pages
55–67. 1994.
7. M. A. Colón. Approximating the algebraic relational semantics of imperative programs. In SAS, pages 296–311. 2004.
8. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints among
variables of a program. In POPL, pages 84–96, 1978.
9. L. De Moura and N. Bjørner. Z3: an efficient SMT solver. TACAS, pages 337–340,
2008.
10. A. Farzan and Z. Kincaid. An algebraic framework for compositional program
analysis. CoRR, abs/1310.3481, 2013.
11. A. Finkel and J. Leroux. How to compose Presburger-accelerations: Applications
to broadcast protocols. In FST TCS, pages 145–156, 2002.
12. L. Gonnord and N. Halbwachs. Combining widening and acceleration in linear
relation analysis. In SAS, pages 144–160. 2006.
13. B. Jeannet, P. Schrammel, and S. Sankaranarayanan. Abstract acceleration of
general linear loops. In POPL, pages 529–540, 2014.
14. S. Kleene. Representation of events in nerve nets and finite automata. In C. Shannon and J. Mccarthy, editors, Automata Studies, pages 3–42. Princeton University
Press, Princeton, N.J., 1956.
15. L. Kovács. Reasoning algebraically about P-solvable loops. In TACAS, pages
249–264. 2008.
16. D. Kroening, M. Lewis, and G. Weissenbacher. Under-approximating loops in C
programs for fast counterexample detection. In CAV, pages 381–396. 2013.
17. D. Kroening, N. Sharygina, S. Tonetta, A. Tsitovich, and C. Wintersteiger. Loop
summarization using abstract transformers. In ATVA, pages 111–125. 2008.
18. J. Leroux and G. Sutre. Accelerated data-flow analysis. In SAS, pages 184–199,
2007.
19. Y. Li, A. Albarghouthi, Z. Kincaid, A. Gurfinkel, and M. Chechik. Symbolic
optimization with SMT solvers. In POPL, pages 607–618, 2014.
20. A. Miné. Symbolic methods to enhance the precision of numerical abstract domains. In VMCAI, pages 348–363, 2006.
21. D. Monniaux. Automatic modular abstractions for linear constraints. In POPL,
pages 140–151, 2009.
22. D. Monniaux. Quantifier elimination by lazy model enumeration. In CAV, pages
585–599, 2010.
23. M. Müller-Olm and H. Seidl. Precise interprocedural analysis through linear algebra. POPL, pages 330–341, 2004.
24. G. C. Necula, S. McPeak, S. P. Rahul, and W. Weimer. CIL: Intermediate language
and tools for analysis and transformation of C programs. In CC, pages 213–228,
2002.
25. C. Popeea and W.-N. Chin. Inferring disjunctive postconditions. ASIAN, pages
331–345, 2007.
26. T. W. Reps, S. Sagiv, and G. Yorsh. Symbolic implementation of the best transformer. In VMCAI, pages 252–266, 2004.
27. B. G. Ryder and M. C. Paull. Elimination algorithms for data flow analysis. ACM
Comput. Surv., 18(3):277–316, Sept. 1986.
28. A. Schrijver. Theory of Linear and Integer Programming. John Wiley & Sons, Inc.,
New York, NY, USA, 1986.
29. M. Sharir and A. Pnueli. Two approaches to interprocedural data flow analysis,
chapter 7, pages 189–234. Prentice-Hall, Englewood Cliffs, NJ, 1981.
30. R. E. Tarjan. Fast algorithms for solving path problems. J. ACM, 28(3):594–614,
July 1981.
| 6cs.PL
|
Representing Extended Finite State Machines for SDL by
A Novel Control Model of Discrete Event Systems
Peng Wang and Kai-Yuan Cai
Department of Automatic Control
Beijing University of Aeronautics and Astronautics
Beijing 100083, China
[email protected] [email protected]
Abstract
This paper discusses EFSM for SDL and
transforms EFSM into a novel control model of
discrete event systems. We firstly propose a control
model of discrete event systems, where the event set is
made up of several conflicting pairs and control is
implemented to select one event of the pair. Then we
transform EFSM for SDL to the control model to
clarify the control mechanism functioning in SDL flow
graphs. This work views the EFSM for SDL in the
perspective of supervisory control theory, and this
contributes to the field of software cybernetics, which
explores the theoretically justified interplay of software
and the control.
system (DES) by an automaton, and control is
implemented by another automaton called supervisor.
This control structure has been widely accepted and
followed up by many other advanced models in this
field, such as the partially observable DES, the
decentralized control approach and so forth. Nowadays
it has formed a theoretical framework, named by RW
Framework [9]. Considering the inherent connection of
automata and computer science, the supervisory
control theory has been introduced into the computer
science these years in order to design and analyze the
computer program more formally and safely. Some
approaches have be explored, and some topics have
been discussed with respect to the programming in
robotics [10], the software design for the power
transformer station [11] as well as developing software
by the polynomial dynamic system approach [12].
1. Introduction
Extended Finite State Machines (EFSMs) are
widely used in computer science and software
engineering [1, 2, 3], especially in the field of program
analysis and testing. This model typically simulates
flow graphs of software programs, and can be
conditionally specified to be the computing model in
Specification and Description Language (SDL) [3, 4].
As well known in computer science and engineering,
SDL is a language to specify the communication of
several processes, where each process is modeled by an
EFSM, and the whole communicating system is
modeled and computed by CEFSMs, namely
Communicating Extended Finite State Machines [5, 6].
This provides one of the backgrounds of our research
work.
Another background of this paper refers to the
supervisory control theory of discrete event systems,
firstly proposed by P. J. Ramadge and W. M. Wonham
in 1980s [7, 8]. This theory represents a discrete event
Fig. 1. Represent EFSM for SDL by control model of DES.
In this paper we propose a new control model of
discrete event systems on the basis of RW Framework
in order to analyze EFSM for the computing model in
SDL. This new model is different from that in RW
Framework because the new model does not partition
the event set by the controllable and uncontrollable
subsets. Rather, the event set in the new model is
composed of several conflicting pairs. Then the
supervisor, which is defined in a similar manner as that
in RW Framework, implements control by selecting
one event of the conflicting pair. By the new control
model, we represent the EFSM for SDL from the
perspective of supervisory control theory. The main
idea of our research work is illustrated in Fig. 1.
To bridge the gap between EFSM and supervisory
control theory, a topic has been introduced as
“embedded supervisory control of discrete event
systems” in [13]. That is, given a controlled object in
terms of a finite automaton and a supervisor in the
sense of the RW framework, an equivalent EFMS of a
special form can be obtained. In other words, a discrete
event system coupled with a supervisor can be
transformed to an EFMS in [13], and this also
contributes to bridging the same gap. However, we
study the topic in the opposite course, that is, we
analyze the given EFSM by decomposing it to a novel
control model of discrete event systems, and this
control model is different from the classic one in RW
Framework. Besides this, the EFSM discussed in this
paper is more generalized in some aspects than that
discussed in [13], and is more applicable to practical
problems because it is specified as the computing
model of SDL.
One point to be emphasized is, the work of this
paper is fundamentally inspired by the thoughts of
software cybernetics [14, 15], a new field which
studies the theoretically justified interplay between
software and the control. And the main result of this
paper also contributes to this fields concerning that
EFSM, a computing model in computer science, is
analyzed from the perspective of supervisory control
theory in this paper.
The rest of this paper is organized as follows.
Section 2 derives the new control model of discrete
event systems. Section 3 introduces EFSM and
specifies the general EFSM model to be a special one
for SDL. In Section 4 we transform the EFSM for SDL
to the new control models via two algorithms. Section
5 concludes this paper and prospects the following
research topics.
2. Control Model
Systems
of
Discrete
Event
This section proposes a new control model of
discrete event systems. The new model continues to
use the same supervisor/controller structure as that in
RW Framework. However, the two models have
different partitions on the event sets. That is, the classic
model in RW Framework partitions the event set as the
controllable and uncontrollable subsets while the new
model composes the event set by several conflicting
pairs. Based on these pairs supervisory control is
implemented by supervisor.
2.1 Discrete Event Systems
A discrete event system is modeled by a Mealy
finite deterministic automaton
G = (Q, Σ, Ζ, δ , λ , q0 )
where Q is the state set, Σ is the input event set, Ζ is
the output event set, δ : Q × Σ → Q is the transition
function, λ : Q × Σ → Ζ is the output function, q0 ∈ Q
is the initial state. In general, δ and λ are partial
functions on the domain, and λ (q,σ ) is defined if and
only if δ (q, σ ) is defined. Here the transition function
can be extended to δ : Q × Σ* → Q as below.
(1) δ (q, ε ) = q
(2) δ (q, sσ ) = δ (δ (q, s ),σ )
And similarly the output function can be extended to
λ : Q × Σ* → Ζ* as below.
(1) λ (q, ε ) = ε
(2) λ ( q , sσ ) = λ ( q , s )λ (δ ( q, s ),σ ) .
It is well known that languages can be used to
identify the behaviors of discrete event systems. And
as for the Mealy Automaton, it has both input
sequences and output sequences. Thus its behavior
should naturally be represented by the two sequences
combined together. Hence, Given a Mealy Automaton,
we define the input language by
Linput (G) = {s ∈ Σ * | δ (q0 , s ) is defined} ,
and accordingly define the output language by
Loutput (G ) = {t ∈ Ζ * | ∃s ∈ Σ ∗ s.t . λ (q0 , s) = t} .
However, it is noted that there exists the situation
that more than one input sequence are respondent to a
same output sequence. Therefore, it is also necessary to
reflect the inherent connection of the input and output
sequences in the defined language. So we further give
the following definition.
s
Σ
L(G ) = { ∈ ( )* | δ (q0 , s ) is defined and λ (q0 , s ) = t}
t
Ζ
And for ( s / t ) ∈ L(G ) , it is easily checked that s and
t belong to input and output languages respectively,
and s is definitely connected to t by λ (q0 , s ) = t .
And therefore this language can completely identify
the behavior of the Mealy Automaton. And we also
note that the regular expression can be applied here to
represent the language especially when the Mealy
Automaton is finite.
In this paper the input event set is especially
composed of several conflicting pairs. This can be
represented by Σ = I ∪ I . Consider I = {a, b, c} for
example, then we have I = {a , b , c} , and this yields the
event set Σ = {a, b, c, a , b , c} , where a and a , b and b ,
c and c are the given conflicting event pairs. Here we
note that this event set can also be represented by
Σ = I × {true, false} , or be interpreted by a relation
defined on the event set.
2.2 Supervisor
Let Σ = I ∪ I , define the control pattern γ by
γ ⊆ Σ and (∀σ ∈ Σ)(σ ∈ γ ⇒ σ ∉ γ )
Then all control patterns compose the set
Γ = {γ ∈ 2 Σ | (∀σ ∈ Σ)(σ ∈ γ ⇒ σ ∉ γ )} .
Event σ is said to be enabled by γ if σ ∈ γ ; or
disabled by γ if σ ∉ γ . And it is easily checked that
control pattern γ satisfies (∀σ ∈ Σ)(σ ∈ γ ⇒ σ ∉ γ ) .
Besides, it is also noted that this definition permits the
situation that σ ∉ γ ∧ σ ∉ γ .
The supervisor is a pair Φ = (S ,ψ ) , where
S = ( X , Σ, ξ , x0 ) is a deterministic automaton and
ψ : X → Γ is a state feedback map. Couple Φ to G
and this yields the supervised discrete event system by
(Φ / G ) = A( X × Q, Σ, Ζ, δˆ, λˆ, < x , q >)
0
0
where the transition function δˆ : X × Q × Σ → X × Q is
defined by
< ξ ( x, σ ), δ ( q, σ ) > if σ ∈ψ ( x ) and
ˆ
δ ( x, q , σ ) =
ξ ( x, σ ), δ (q , σ ) are defined
undefined
otherwise
and the output function λ̂ : X × Q × Σ → Ζ is defined by
λ (q, σ )
if δˆ( x, q, δ ) is defined
λˆ ( x, q, σ ) =
undefined otherwise
Let L(Φ / G) denote the language generated by
Φ / G . And one point to be emphasized is that, if
ξ ( x, σ ) is defined implies σ ∈ψ (x ) , then we have
δˆ( x, q, σ ) is defined ⇔
ξ ( x, σ ) is defined ∧ δ (q, σ ) is defined
Therefore the supervised system is equivalent to the
synchronized product [9] of S and G , namely
Φ /G = S ×G .
Thereby given a complete supervisor [9] in the
form of Φ = ( S ,ψ ) , we can always reduce it via the
automaton S . Specifically speaking, we firstly reduce
S by eliminating ξ ( x, σ ) for σ ∉ψ (x ) , then we
further eliminate the inaccessible states and then obtain
S mod ify . Thereby the supervisory control can be
equivalently achieved by G × S mod ify . Here we note
that automaton S mod ify corresponds to the supervisor
mentioned in [13].
3. Extended Finite State Machines for
Computing Model of SDL
In this section we firstly introduce the general
model of EFSM. Then the general model of EFSM is
specilized to be the computing model in SDL. An
example is given along with the formal explanation.
3.1 General Model of EFSM.
According to [1] we gives the definition of EFSM
as follows. An extended finite state machine is
structured by a five-tuple
EFSM = (Y , I , O,V , T )
where Y , I , O represent the state set, the input event
set and the output event set, respectively. V is the
variable set, composed of several variables, namely
V = {v1, v2 ""vn } . And each variable vi has its
G
domain, denoted by dom(vi ) . Let V denote the vector
composed by all the variables in V , namely
G
V =< v1, v2 "vn > . Then accordingly we have
G
dom(V ) = dom(v1) × dom(v2 )" × dom(vn ) . T is the
transition set. For a transition t ∈ T , t is denoted by
t = ( yt _ src , yt _ dest , it , ot , Pt , At )
where yt _ src ∈ Y and yt _ dest ∈ Y represent the source
state and destination state
respectively. it ∈ I and ot ∈ O
of the transition,
represent the input
G
event and the output event, respectively. Pt (V ) is the
predicate of the transition, which is defined on the
G
variable set V . At (V ) is the update function, which
updates the values of the variables.
By the definition as above, static structure of
EFSM is illustrated. And to make the system dynamic,
there are two other points to be mentioned:
(1) The initial condition for EFSM. The initial
condition includes the initial state y0 ∈ Y and the
initial values of the variables, namely
G
V0 =< v10 , v20 "vn 0 > .
(2) The dynamic rule for EFSM. This rule regulates
the dynamic behavior of the system. For the
transition t = ( yt _ src , yt _ dest , it , ot , Pt , At ) , the rule
is explained as follows: when event it is
imported to state yt _ src , if the current value of
G
G
V makes Pt (V ) true, then the transition is
enabled, the current state turns to yt _ dest , export
G
the event ot and the value of V is updated by
G
At (V ) ; otherwise the transition is disabled, the
G
current state and the value of V remain
unchanged.
Here we note that an EFSM can be nonG
deterministic if the current value of V enables more
than one transition with the one definite input
imported. However, this paper only discusses the
deterministic EFSM where only one transition is
enabled with the one input event at anytime of the
system evolvement.
Besides, it is also reasonable to use languages to
identify the behavior of the EFSM. And by the similar
manners as introduced in Mealy Automata, we can also
prefigure the input and output languages as well as the
combined language to depict the evolvement behavior
of the EFSM model. Here we do not give the formal
definition of languages for EFSMs, but corresponding
concepts about the languages can be well sensed in the
following section.
Fig. 2 EFSM for example.
Based on the aforementioned explanation, an
example of EFSM is presented in Fig. 2, where the
state set is Y = {I , II } , the input event set is I = {a}
and the output event set is O = {m, n} . The variable set
is given by V = {v} , and dom(v ) = {0,1,2"9} . The
predicates on variable v are given as Fig. 2 shows,
namely v > 7 , v ≤ 7 , v > 3 and v ≤ 3 . The update
functions on variable v consist of v := v + 1 and
v := v − 1 . The initial state of EFSM is I , and the
initial value of v is v0 = 0 . Then the EFSM can
evolve by the dynamic rule as mentioned above, and
this EFSM is definitely deterministic.
3.2 EFSM model for SDL.
In this paper, we focus on a special model of
EFSM, which functions as the computing model in
Specification and Description Language. And we note
that SDL specifies the predicate function as Fig. 3
shows.
false
G
Pt (V )
true
Fig. 3 Predicates in SDL flow graphs.
For an input event, the predicate decides where
the system goes by its Boolean values. That is, if the
current values of variables make the predicate true,
then the system runs to one state; otherwise the system
runs to another state.
G
¬Pt (V )
G
Pt (V )
Fig. 4. Predicates in EFSM for SDL.
Therefore the predicate in SDL can be interpreted
as a special case of the predicate in general EFSM, as
shown in Fig. 4. That is, for the computing model of
SDL, the EFSM is characterized by several transition
pairs, and each pair consists of two conflicting
G
transitions, of which one is marked by predicate Pt (V )
and the other one is marked by the contrary predicate
G
¬Pt (V ) . Then the transition pair is denoted by
t = ( yt _ src , yt _ dest , it , ot , Pt , At )
and
t = ( yt _ src , yt _ dest , it , ot , ¬Pt , At ) .
Therefore the EFSM for the computing model of
SDL is formularized by
G
EFSM SDL = (Y , I , O, V , T , y0 ,V0 )
where Y , I , O , V have the same meanings as the
counterparts in general EFSM. To incorporate the
initial conditions and the dynamic rules in EFSM SDL ,
G
let y0 and V0 =< v10 , v20 "" vn 0 > denote the initial
state and initial values of variables, and the transition
set of the EFSM for SDL is defined by
G
G
T : Y × I × P (V ) × {true, false} → Y × O × A(V ) , where
G
P (V ) × {true, false} denotes the conflicting predicates
in transition pairs.
Here we note that the EFSM for the computing
model of SDL must be deterministic because the SDL
flow graphs as well as the programming code cannot
run ambiguously, but be definitely explained to be
deterministic.
4. Representing EFSM for SDL by the
Control Model of DES
Based on the work of previous sections this
section discusses the connection between EFSM and
control models of DES. We transform EFSM for SDL
to the control model of discrete event systems.
Especially, the update function and the predicate
together act as a supervisor to control where program
flow goes. Here the update function acts as the
automaton of the supervisor, and the predicate in
EFSM plays the role of state feedback map. The basic
idea of this section is shown in Fig. 6.
G
G
T : Q × I × P(V ) × {true, false} → Q × O × A(V )
Fig. 5. SDL flow graph.
Then we transform the EFSM given in Fig. 2 to
the SDL flow graph in Fig. 5. And the programming
code can be further derived from of the SDL flow
graph as follows.
enum State {I, II}; static State s;
enum Input {a}; enum Output {m,n};
static Input i; static Output o; static int v;
void Initialization()
{ s=I; v=0; }
void Transition (Input i)
{ switch(s)
{ case I:
switch(i)
{ case a: if(v>7)
{s=II; o=n; v＝v+1;}
else
{s=I; o=m; v=v+1;}
break;
} break;
case II:
switch(i)
{ case a: if(v<=3)
{s=I; o=m; v＝v-1;}
else
{s=II; o=n; v=v-1;}
break;
} break;
}
}
ξ : X ×Σ → X
ψ :X →Γ
Φ =< S ,ψ >
Fig. 6. Extract the supervisor from EFSM for SDL.
This section represents the EFSM for SDL by the
control model proposed in Section 1. By this approach
we highlight the control functioning in the general
program flows, and study the inherent control
mechanism in program flow graphs. This work
contributes to the field of software cybernetics, which
explores the theoretically justified interplay of the
software and the control.
In the rest of this section we firstly specify the
problem formally, and the solution to the problem is
given via two algorithms.
G
Problem: Give an EFSM SDL = (Y , I , O, V , T , y0 ,V0 ) ,
transform this computing model to the control model
of DES proposed in Section 1, namely a DES coupled
with a supervisor.
4.1 Controlled DES
Firstly we extract a controlled DES from EFSM
of the computing model in SDL. Consider that the
output event set and the output function are defined in
EFSM, the controlled DES is extracted as a Mealy
Automaton.
G
Algorithm 1: Give EFSM SDL = (Y , I , O, V , T , y0 ,V0 ) .
Extract controlled DES G .
Step 1. Structure G by G = (Y , Σ, O, δ , λ , y0 ) , where
Σ := I × {true, false} ;
Step 2. Let i ∈ I and bin ∈{true, false} . Construct
transition function δ : Y × Σ → Y by
G
y dest
if T(y src , i , Pt (V ), bin)
G
δ ( y src , < i, bin >) =
=< y dest , o, At (V ) >
undefined otherwise
Construct the output function λ : Y × Σ → O by
G
o
if T(y src , i , Pt (V ), bin)
G
λ ( y src , < i , bin >) =
=< y dest , o, At (V ) >
undefined otherwise
Two points should be emphasized here. Firstly we
construct the input event set by Σ = I × {true, false} ,
and this yields the conflicting event pairs as introduced
in Section 1. Secondly, if the transition function
δ : Y × Σ → Y and the output function λ : Y × Σ → O
combine together and additionally consider
Σ = I × {true, false} , then it yields
Y × I × {true, false} → Y × O
And this corresponds to the transition of EFSM SDL by
G
G
eliminating
P (V )
and
A(V )
from
G
G
T : Y × I × P (V ) × {true, false} → Y × O × A(V ) .
Recalling the example given in Section 3, we
extract the Mealy Automaton to be the controlled DES
as shown in Fig. 7. And the conflicting event pair is
obtained by < a, true > and < a, false > .
G
construct automaton S by the update function A(V ) ;
(2) construct state feedback map ψ by the predicate
G
P (V ) .
G
Algorithm 2: Give EFSM SDL = (Y , I , O, V , T , y0 ,V0 ) .
Extract supervisor Φ = ( S ,ψ ) .
Step 1. Structure automaton S = ( X , Σ, ξ , x0 ) , where
G
X := dom(V ) × Y ; Σ := I × {true, false} ;
G
x0 :=< V0 , y0 > .
Step 2. Let bin ∈{true, false} . Construct transition
ξ : X × Σ → X as follows:
G
ξ (< V , y src >, < i , bin >)
,
if
consider
G
G
T ( ysrc , i, Pt (V ), bin) =< y dest , o, At (V ) >
is
G
G
defined
and
At (V ) ∈ dom(V )
,
then
G
G
ξ (< V , y src >, < i, bin >) =< At (V ), y dest >
;
G
otherwise ξ (< V , y src >, < i , bin >) is undefined.
Step 3. Let bin ∈ {true, false} . Let projVG ( x ) and
G
projy ( x ) denote the segments V and y
subject to state x . Namely,
G
x =< V , y >=< v1 , v2 " vn , y > ,
we
G
projVG ( x) = V =< v1 , v2 " vn >
given
have
and
projy ( x) = y .
Construct state feedback map ψ : X → Γ as
follows: < i,bin >∈ψ (x ) if and only if
G
T (proj y ( x ), i, Pt (V ), bin)
G
Pt (V ) |VG =proj G ( x ) = bin .
is
defined
and
V
Considering the classical logic system,
G
G
V = projVG ( x) can only make Pt (V ) to be true or false.
And note that only deterministic EFSMs are discussed,
it is easily checked that the state feedback map follows
the definition given in Section 1. That is,
< i,true >∈ψ (x) implies < i,false >∉ψ (x) .
Recalling the example presented in Section 3, we
construct the state set by X = dom(v) × {I , II } . The
initial state for automaton S is obtained by x0 = 0 I .
Fig. 7. Controlled DES extracted from EFSM.
4.2 Supervisor
Secondly we extract the supervisor from EFSM.
To obtain the supervisor, two steps should be taken: (1)
And the event set for automaton S is
Σ = {< a, true >, < a, false >} .
Apply Step 2 in Algorithm 2, and we construct the
transition function of automaton S . And automaton
S is illustrated in Fig. 8.
Fig. 8. Automaton of the supervisor.
Apply Step 3 in Algorithm 2, and we construct the
state feedback map by the predicates in the EFSM.
And the supervisor is finally obtained as shown below.
derived is I × {true, false} , we can actually extract the
first sub-element from it, and this makes the two
models comparable with respect to input events. And
thereby it is checked that the two models have the
same set of the input sequences as well as the
corresponding the output sequences, that is, the two
models follow the same disciplines during evolvement.
Consider the forgoing example, and we illustrate
the input and output sequences of the control model by
the diagram as below.
Fig. 11. The input and output sequences of
the control model.
Fig. 9. Supervisor extracted from EFSM.
According to Section 1 the supervisory control
implemented by Φ = ( S ,ψ ) can be equivalently
implemented by an automaton S mod ify . Then the
supervised system can be achieved by the synchronized
product of the controlled DES and the automaton,
namely G × S mod ify . Here we illustrate S mod ify in Fig.
10.
Here regular expressions can be derived to
represent the sequences as shown in Fig. 11. And we
extract the first sub-elements of the control model’s
events. Then < a, true > and < a, false > are
simplified as a . And if we mark state I and v = 2 ,
Then the input language is formulated by a 2 (a 14 )* ,
and the output language is given by m 2 (m 6 n 7 m )* .
Consequently we can formulate the combined
sequences by the language as below.
a
a
a a
( ) 2 (( ) 6 ( ) 7 )*
m
m n m
Consider the EFSM presented in Section 3, we
illustrate its input and output sequences by the diagram
in Fig. 12.
Fig. 10. Supervisor modified.
After introducing the algorithms we give
explanation as follows. It is noted that the evolvement
of state-transition systems can be represented by the
event sequences, and both of the EFSM and the control
model of DES have the input sequences and output
sequences, where each input sequence corresponds to
an output sequence definitely. So the question is
whether the algorithms keep the sequences fixed
between the two models. The answer is affirmative.
Although the input event set of the control model
Fig. 12. The input and output sequences of EFSM.
Then it is easily checked that the EFSM above has
the same input and output sequences as the derived
control model. Specifically speaking, from the diagram
in Fig. 12 we can exactly deduce the same input and
output languages as the derived control model.
Furthermore, if we also mark state I and v = 2 , then
the same regular expression can be formulated to
denote the combined sequence as the same as that of
the control model.
5. Concluding Remarks
This paper refers to two research topics, the
control model of DES and the EFSM for SDL, and
bridges the gap between the two topics via the given
algorithms. We firstly propose a novel control model
of DES, which is characterized by selecting one event
of the conflicting event pair. Then we focus on EFSM,
and specify the general EFSM to be the computing
model for SDL. Then the connection between these
two models is studied and the algorithms are given to
transform the EFSM for SDL to the control model of
DES.
The work of this paper highlights the control
mechanism functioning in flow graphs of SDL, and
contributes to the field of software cybernetics, which
explores the theoretically justified interplay between
software and the control. Based on the work of this
paper, some follow-up topics can be approached. And
one topic refers to Communicating Extended Finite
State Machines, which functions as the communicating
model in SDL. We can analyze CEFSM based on the
work of this paper and further propose a novel
approach to verify the communicating system modeled
by SDL from the perspective of supervisory control
theory.
References
[1] C. Bourhfir, E. Aboulhamid, F. Khendek, R. Dssouli,
“Test Cases Selection from SDL Specifications”, Computer
Networks, 2001, Vol. 35, pp.693-708.
[2] A. Petrenko, S. Boroday, R. Groz, “Confirming
Configurations in EFSM Testing”, IEEE Transaction on
Software Engineering, 2004, Vol. 30, No. 1, pp. 29-42.
[3] R.E. Miller, Y. Xue, “Bridging the Gap between Formal
Specification and Analysis of Communication Protocols”,
Proceedings of the 1996 IEEE Fifteenth Annual International
Phoenix Conference on Computers and Communications,
March 1996, pp. 225-231.
[4] “Introduction to SDL 88”. http://www.sdl-forum.org/
sdl88tutorial/index.html, 2002.
[5] C. Bourhfir, E. Aboulhamid, R. Dssouli, N. Ricob, “A
Test Case Generation Approach for Conformance Testing of
SDL Systems”, Computer Communications, 2001, Vol. 24,
pp. 319-333.
[6] J. J. Li, W. E. Wong, “Automation Test Generation from
Communicating Extended Finite State Machine (CEFSM)Based Model”, Proceedings of the fifth IEEE International
Symposium on Object-Oriented Real-Time Distributed
Computing, 2002, pp. 181-185.
[7] P. J. Ramadge and W. M. Wonham, “Supervisory Control
of a Class of Discrete Event Processes”, SIAM Journal of
Control and Optimization, 1987, Vol. 25, No. 1, pp. 206-230.
[8] W. M. Wonham and P. J. Ramadge, “On the Supremal
Controllable Sublanguage of a Given Language”, SIAM
Journal of Control and Optimization, 1987, Vol. 25, No. 3,
pp. 637-659.
[9] W.M. Wonham, “Supervisory Control of Discrete-Event
Systems”, Dept. of Electrical & Computer Engineering,
University of Toronto, ECE 1636F/1637S 2004-05, Revised
2004.07.01.
[10] E. Rutten and H. Marchand, “Task-Level Programming
for Control Systems Using Discrete Control Synthesis”,
February 2002, Research Report INRIA, No. 4389.
[11] H. Marchand and M. Samaan, “Incremental Design of a
Power Transformer Station Controller Using a Controller
Synthesis Methodology”, IEEE Transactions on Software
Engineering, August 2000, Vol.26, No.8, pp. 729-741.
[12] X. Y. Wang, Y. C. Li and K. Y. Cai, “On the
Polynomial Dynamic System Approach to Software
Development”, Science in China (series F), 2004, Vol. 46,
No. 4, pp. 437-457.
[13] Y. Yang, R. Gohari, “Embedded Supervisory Control of
Discrete-Event Systems”, IEEE International Conference on
Automation Science and Engineering, 2005, pp. 410-415.
[14] K. Y. Cai, J. W. Cangussu, R. A. DeCarlo, A. P.
Mathur, “An Overview of Software Cybernetics”,
Proceedings of the 11th Annual International Workshop on
Software Technology and Engineering Practice, IEEE
Computer Society Press, 2004, pp. 77-86.
[15] K. Y. Cai, “Optimal Software Testing and Adaptive
Software Testing in the Context of Software Cybernetics”,
Information and Software Technology, 2002, Vol. 44, pp.
841-855.
| 3cs.SY
|
FAST L EARNING R ATE OF D EEP L EARNING
Fast learning rate of deep learning via a kernel perspective
arXiv:1705.10182v1 [math.ST] 29 May 2017
Taiji Suzuki
TAIJI @ MIST. I . U - TOKYO . AC . JP
Department of Mathematical Informatics
The University of Tokyo
7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan,
PRESTO, Japan Science and Technology Agency,
Center for Advanced Integrated Intelligence Research, RIKEN
Abstract
We develop a new theoretical framework to analyze the generalization error of deep learning,
and derive a new fast learning rate for two representative algorithms: empirical risk minimization
and Bayesian deep learning. The series of theoretical analyses of deep learning has revealed its
high expressive power and universal approximation capability. Although these analyses are highly
nonparametric, existing generalization error analyses have been developed mainly in a fixed dimensional parametric model. To compensate this gap, we develop an infinite dimensional model that
is based on an integral form as performed in the analysis of the universal approximation capability.
This allows us to define a reproducing kernel Hilbert space corresponding to each layer. Our point
of view is to deal with the ordinary finite dimensional deep neural network as a finite approximation
of the infinite dimensional one. The approximation error is evaluated by the degree of freedom of
the reproducing kernel Hilbert space in each layer. To estimate a good finite dimensional model,
we consider both of empirical risk minimization and Bayesian deep learning. We derive its generalization error bound and it is shown that there appears bias-variance trade-off in terms of the
number of parameters of the finite dimensional approximation. We show that the optimal width of
the internal layers can
√ be determined through the degree of freedom and the convergence rate can
be faster than O(1/ n) rate which has been shown in the existing studies.
Keywords: Deep Learning, Fast Learning Rate, Kernel Method, Degree of Freedom, Generalization Error Bounds, Empirical Risk Minimizer, Bayesian Deep Learning
1. Introduction
Deep learning has been showing great success in several applications such as computer vision,
natural language processing, and many other area related to pattern recognition. Several highperformance methods have been developed and it has been revealed that deep learning possesses
great potential. Despite the development of practical methodologies, its theoretical understanding
is not satisfactory. Wide rage of researchers including theoreticians and practitioners are expecting
deeper understanding of deep learning.
Among theories of deep learning, a well developed topic is its expressive power. It has been
theoretically shown that deep neural network has exponentially large expressive power against the
number of layers. For example, Montufar et al. (2014) showed that the number of polyhedral regions created by deep neural network can exponentially grow as the number of layers increases.
Bianchini and Scarselli (2014) showed that the Betti numbers of the level set of a function created
1
T. S UZUKI
by deep neural network grows up exponentially against the number of layers. Other researches
also concluded similar facts using different notions such as tensor rank and extrinsic curvature
(Cohen et al., 2016; Cohen and Shashua, 2016; Poole et al., 2016).
Another important issue in neural network theories is its universal approximation capability. It
is well known that 3-layer neural networks have the ability, and thus the deep neural network also
does (Cybenko, 1989; Hornik, 1991; Sonoda and Murata, 2015). When we discuss the universal
approximation capability, the target function that is approximated is arbitrary and the theory is
highly nonparametric in its nature.
Once we knew the expressive power and universal approximation capability of deep neural network, the next theoretical question naturally arises in its generalization error. The generalization
ability is typically analyzed by evaluating the Rademacher complexity. Bartlett (1998) studied 3layer neural networks and characterized its Rademacher complexity using the norm of weights.
Koltchinskii and Panchenko (2002) studied deep neural network and derived its Rademacher complexity under norm constraints. More recently, Neyshabur et al. (2015) analyzed the Rademacher
complexity based on more generalized norm, and Sun et al. (2015) derived a generalization error
√
bound with a large margine assumption. As a whole, the studies listed above derived O(1/ n)
convergence of the generalization error where n is the sample size. One concern in this line of
√
convergence analyses is that the convergence of the generalization error is only O(1/ n) where
n is the sample size. Although this is minimax optimal, it is expected that we could show faster
convergence rate with some additional assumptions such as strong convexity of the loss function.
Actually, in a regular parametric model, we have O(1/n) convergence of the generalization error
(Hartigan et al., 1998). Moreover, the generalization error bound has been mainly given in finite dimensional models. As we have observed, the deep neural network possesses exponential expressive
power and universal approximation capability which are highly nonparametric characterizations.
This means that the theories are developed separately in the two regimes; finite dimensional parametric model and infinite dimensional nonparametric model. Therefore, theories that connect these
two regimes are expected to comprehensively understand statistical performance of deep learning.
In this paper, we consider both of empirical risk minimization and Bayesian deep learning and
analyze the generalization error using the terminology of kernel methods. Consequently, (i) we de√
rive a faster learning rate than O(1/ n) and (ii) we connect the finite dimensional regime and the infinite dimensional regime based on the theories of kernel methods. The empirical risk minimization
is a typical approach to learn the deep neural network model. It is usually performed by applying
stochastic gradient descent with the back-propagation technique (Widrow and Hoff, 1960; Amari,
1967; Rumelhart et al.). To avoid over-fitting, such techniques as regularization and dropout have
been employed (Srivastava et al., 2014). Although the practical techniques for the empirical risk
minimization have been extensively studied, there is still much room for improvement in its generalization error analysis. Bayesian deep learning has been recently gathering more attentions mainly
because it can deal with the estimation uncertainty in a natural way. Examples of Bayesian deep
learning researches include probabilistic backpropagation (Hernandez-Lobato and Adams, 2015),
Bayesian dark knowledge (Balan et al., 2015), weight uncertainty by Bayesian backpropagation
(Blundell et al., 2015), dropout as Bayesian approximation (Gal and Ghahramani, 2016). To analyze a sharper generalization error bound, we utilize the so-called local Rademacher complexity technique for the empirical risk minimization method (Mendelson, 2002; Bartlett et al., 2005;
Koltchinskii, 2006; Giné and Koltchinskii, 2006), and, as for the Bayesian method, we employ
the theoretical techniques developed to analyze nonparametric Bayes methods (Ghosal et al., 2000;
2
FAST L EARNING R ATE OF D EEP L EARNING
Table 1: Summary of derived bounds for the generalization error kfb − f o k2L2 (P (X)) where n is the
sample size, R is the norm of the weight in the internal layers, R̂∞ is an L∞ -norm bound
of the functions in the model, σ is the observation noise, dx is the dimension of the input,
mℓ is the width of the ℓ-th internal layer and Nℓ (λℓ ) for (λℓ > 0) is the degree of freedom
(Eq. (5)).
Error bound
PL
2
2 P
General setting
L ℓ=2 RL−ℓ+1 λℓ + σ +nR̂∞ L
ℓ=1 mℓ mℓ+1 log(n)
under an assumption that mℓ & Nℓ (λℓ ) log(Nℓ (λℓ )).
σ2 +R̂2∞ PL
∗ ∗
Finite dimensional model
ℓ=1 mℓ mℓ+1 log(n)
n
where m∗ℓ is the true width of the ℓ-th internal layer.
1
2
P
L−ℓ+1 n− 1+2sℓ log(n) + dx log(n)
Polynomial decay eigenvalue
L L
ℓ=2 (R ∨ 1)
n
where sℓ is the decay rate of the eigenvalue of the kernel
function on the ℓ-th layer.
van der Vaart and van Zanten, 2008, 2011). These analyses are quite advantageous to the typical
Rademacher complexity analysis because we can obtain convergence rate between O(1/n) and
√
√
O(1/ n) which is faster than that of the standard Rademacher complexity analysis O(1/ n).
As for the second contribution, we first introduce an integral form of deep neural network as
performed in the research of the universal approximation capability of 3-layer neural networks
(Sonoda and Murata, 2015). This allows us to have a nonparametric model of deep neural network as a natural extension of usual finite dimensional models. Afterward, we define a reproducing
kernel Hilbert space (RKHS) corresponding to each layer like in Bach (2017, 2015). By doing so,
we can borrow the terminology developed in the kernel method into the analysis of deep learning. In
particular, we define the degree of freedom of the RKHS as a measure of complexity of the RKHS
(Caponnetto and de Vito, 2007; Bach, 2015), and based on that, we evaluate how large a finite dimensional model should be to approximate the original infinite dimensional model with a specified
precision. These theoretical developments reveal that there appears bias-variance trade-off. That
is, there appears trade-off between the size of the finite dimensional model approximating the nonparametric model and the variance of the estimator. We will show that, by balancing the trade-off,
a fast convergence rate is derived. In particularly, the optimal learning rate of the kernel method
is reproduced from our deep learning analysis due to the fact that the kernel method can be seen
as a 3-layer neural network with an infinite dimensional internal layer. A remarkable property of
the derived generalization error bound is that the error is characterized by the complexities of the
RKHSs defined by the degree of freedom. Moreover, the notion of the degree of freedom gives a
practical implication about determination of the width of the internal layers.
The obtained generalization error bound is summarized in Table 11 .
1. a ∨ b indicates max{a, b}.
3
T. S UZUKI
2. Integral representation of deep neural network
Here we give our problem settings and the model that we consider in this paper. Suppose that n
input-output observations Dn = (xi , yi )ni=1 ⊂ Rdx × R are independently identically generated
from a regression model
yi = f o (xi ) + ξi (i = 1, . . . , n)
where (ξi )ni=1 is an i.i.d. sequence of Gaussian noises N (0, σ 2 ) with mean 0 and variance σ 2 , and
(xi )ni=1 is generated independently identically from a distribution P (X) with a compact support in
Rdx . The purpose of the deep learning problem we consider in this paper is to estimate f o from the
n observations Dn .
To analyze the generalization ability of deep learning, we specify a function class in which the
true function f o is included, and, by doing so, we characterize the “complexity” of the true function
in a correct way.
In order to give a better intuition, we first start from the simplest model, the 3-layer neural network. Let η be a nonlinear activation function such as ReLU (Nair and Hinton, 2010; Glorot et al.,
2011); η(x) = (max{xi , 0})di=1 for a d-dimensional vector x ∈ Rd . The 3-layer neural network
model is represented by
f (x) = W (2) η(W (1) x + b(1) ) + b(2)
where we denote by m2 the number of nodes in the internal layer, and W (2) ∈ R1×m2 , W (1) ∈
Rm2 ×dx , b(1) ∈ Rm2 and b(2) ∈ R. It is known that this model is universal approximator and it is
important to consider its integral form
Z
f (x) = h(w, b)η(w⊤ x + b)dwdb + b(2) .
(1)
where (w, b) ∈ Rdx × R is a hidden parameter, h : Rdx × R → R is a function version of the weight
matrix W (2) , and b(2) ∈ R is the bias term. This integral form appears in many places to analyze
the capacity of the neural network. In particular, through the ridgelet analysis, it is shown that
there exists the integral form corresponding to any f ∈ L1 (Rdx ) which has an integrable Fourier
transform for an appropriately chosen activation function η such as ReLU (Sonoda and Murata,
2015).
Motivated by the integral form of the 3-layer neural network, we consider a more general representation for deeper neural network. To do so, we define a feature space on the ℓ-th layer. The
feature space is a a probability space (Tℓ , Bℓ , Qℓ ) where Tℓ is a Polish space, Bℓ is its Borel algebra,
and Qℓ is a probability measure on (Tℓ , Bℓ ). This is introduced to represent a general (possibly)
continuous set of features as well as a discrete set of features. For example, if the ℓ-th internal layer
is endowed with a dℓ -dimensional finite feature space, then Tℓ = {1, . . . , dℓ }. On the other hand,
the integral form (1) corresponds to a continuous feature space T2 = {(w, b) ∈ Rdx × R} in the second layer. Now the input x is a dx -dimensional real vector, and thus we may set T1 = {1, . . . , dx }.
Since the output is one dimensional, the output layer is just a singleton TL+1 = {1}. Based on these
feature spaces, our integral form of the deep neural network is constructed by stacking the map on
the ℓ-th layer fℓo : L2 (Qℓ ) → L2 (Qℓ+1 ) given as
Z
(2a)
hoℓ (τ, w)η(g(w))dQℓ (w) + boℓ (τ ),
fℓo [g](τ ) =
Tℓ
4
FAST L EARNING R ATE OF D EEP L EARNING
where hoℓ (τ, w) corresponds to the weight of the feature w for the output τ and hoℓ ∈ L2 (Qℓ+1 × Qℓ )
and hoℓ (τ, ·) ∈ L2 (Qℓ+1 ) for all τ ∈ Tℓ+1 2 . Specifically, the first and the last layers are represented
as
f1o [x](τ ) =
dx
X
ho1 (τ, j)xj Q1 (j) + bo1 (τ ),
(2b)
hoL (w)η(g(w))dQL (w) + boL ,
(2c)
j=1
fLo [g](1)
=
Z
TL
where we wrote hoL (w) to indicate hoL (1, w) for simplicity because TL+1 = {1}. Then the true
function f o is given as
o
◦ · · · ◦ f1o (x).
f o (x) = fLo ◦ fL−1
(3)
Since, the shallow 3-layer neural network is a universal approximator, and so is our generalized
deep neural network model (3). It is known that deep neural network tends to give more efficient
representation of a function than the shallow network. Actually, Eldan and Shamir (2016) gave
an example of a function that the 3-layer neural network cannot approximate under a precision
unless its with is exponential in the input dimension but the 4-layer neural network can approximate
with polynomial order widths (see Safran and Shamir (2016) for other examples). In other words,
each layer of a deep neural network can be much “simpler” than one of a shallow network (more
rigorous definition of complexity of each layer will be given in the next section). Therefore, it is
quite important to consider the integral representation of a deep neural network rather than a 3-layer
network.
The integral representation is natural also from the practical point of view. Indeed, it is well
known that the deep neural network learns a simple pattern in the early layers and it gradually
extracts more complicated features as the layer is going up. The trained feature is usually continuous
one. For example, in computer vision tasks, the second layer typically extracts gradients toward
several degree angles (Krizhevsky et al., 2012). The angle is a continuous variable and thus the
feature space should be continuous to cover all angles. On the other hand, the real network discretize
the feature space because of limitation of computational resources. Our theory introduced in the next
section offers a measure to evaluate this discretization error.
3. Finite approximation of the integral form
The integral form is a convenient way to describe the true function. However, it is not useful to
estimate the function. When we estimate that, we need to discretize the integrals by finite sums due
to limitation of computational resources as we do in practice. In other word, we consider the usual
finite sum deep learning model as an approximation of the integral form. However, the discrete
approximation induces approximation error. Here we give an upper bound of the approximation
error. Naturally, there arises the notion of bias and variance trade-off, that is, as the complexity of
the finite model increases the “bias” (approximation error) decreases but the “variance” for finding
the best parameter in the model increases. Afterwards, we will bound the variance for estimating
2. Note that, for g ∈ L2 (Qℓ ), fℓ [g] is also square integrable with respect to L2 (Qℓ+1 ) if η is Lipschitz continuous
because h ∈ L2 (Qℓ+1 × Qℓ ).
5
T. S UZUKI
the finite approximation in Section 4.3. Combining these two notions, it is possible to quantify the
bias-variance trade-off and find the best strategy to minimize the entire generalization error.
The approximation error analysis of the deep neural network can be well executed by utilizing
notions of the kernel method. Here we construct RKHS for each layer in a way analogous to
Bach (2015, 2017) who studied shallow learning and the kernel quadrature rule. Let the output
of the ℓ-th layer be Fℓo (x, τ ) := (fℓo ◦ · · · ◦ f1o (x))(τ ). We define a reproducing kernel Hilbert
space (RKHS) corresponding to the ℓ-th layer (ℓ ≥ 2) by introducing its associated kernel function
kℓ : Rdx × Rdx → R. We define the positive definite kernel kℓ as
Z
o
o
′
(x′ , τ ))dQℓ (τ ).
(x, τ ))η(Fℓ−1
η(Fℓ−1
kℓ (x, x ) :=
Tℓ
It is easy to check that kℓ is actually symmetric and positive definite. It is known that there exists a
unique RKHS Hℓ corresponding the kernel kℓ (Aronszajn, 1950). Close investigation of the RKHS
for several examples for shallow network has been given in (Bach, 2017).
Under this setting, all arguments at the ℓ-th layer can be carried out through the theories of
kernel methods. Importantly, for g ∈ Hℓ , there exists h ∈ L2 (Qℓ ) such that
Z
o
(x, τ ))dQℓ (τ ).
h(τ )η(Fℓ−1
g(x) =
Tℓ
Moreover, the norms of g and h are connected as
kgkHℓ = khkL2 (Qℓ ) ,
(4)
(Bach, 2015, 2017). Therefore, the function
Z
o
(x, w))dQℓ (w),
hoℓ (τ, w)η(Fℓ−1
x 7→
Tℓ
representing the magnitude of a feature τ ∈ Tℓ+1 for the input x is included in the RKHS and its
RKHS norm is equivalent to that of the internal layer weight kho (τ, ·)kL2 (Qℓ ) because of Eq. (4).
To derive the approximation error, we need to evaluate the “complexity” of the RKHS. Basically,
the complexity of the ℓ-th layer RKHS Hℓ is controlled by the behavior of the eigenvalues of the
kernel. To formally state this notion, we introduce the integral operator associated with the kernel
kℓ defined as
Z
kℓ (·, x)g(x)dP (x),
Tℓ :g 7→
X
L2 (P (X)) → L2 (P (X)).
If the kernel function admits an orthogonal decomposition
kℓ (x, x′ ) =
∞
X
(ℓ) (ℓ)
(ℓ)
µj φj (x)φj (x′ ),
j=1
(ℓ)
in L2 (P (X) × P (X)) where (µj )∞
j=1 is the sequence of the eigenvalues ordered in decreasing orP
(ℓ) ∞
(ℓ)
der, and (φj )j=1 forms an orthonormal system in L2 (P (X)), then for g(x) = ∞
j=1 αj φj (x), the
6
FAST L EARNING R ATE OF D EEP L EARNING
integral operation is expressed as Tℓ g =
P∞
(ℓ) (ℓ)
j=1 αj µj φj
(see Steinwart and Christmann (2008);
(ℓ)
Steinwart and Scovel (2012) for more details). Therefore each eigenvalue µj plays a role like a “fil(ℓ)
ter” for each component φj . Here it is known that for all g ∈ Hℓ , there exists h̄ ∈ L2 (P (X)) such
that g = Tℓ h̄ and kgkHℓ = kh̄kL2 (P (X)) (Caponnetto and de Vito, 2007; Steinwart et al., 2009).
Combining this with Eq. (4), we have kgkHℓ = khkL2 (Qℓ ) = kh̄kL2 (P (X))
Based on the integral operator Tℓ , we define the degree of freedom Nℓ (λ) of the RKHS as
Nℓ (λ) = Tr[(Tℓ + λ)−1 Tℓ ]
for λ > 0. The degree of freedom can be represented as Nℓ (λ) =
(5)
P∞
(ℓ)
µj
j=1 µ(ℓ) +λ
j
by using the
eigenvalues of the kernel.
Now, we assume that the true function f o satisfies a norm condition as follows.
Assumption 1 For each ℓ, hoℓ and boℓ satisfy that
khoℓ (τ, ·)kL2 (Qℓ ) ≤ R (∀τ ∈ Tℓ ),
|boℓ (τ )| ≤ Rb (∀τ ∈ Tℓ ).
By Eq. (4), the first assumption khoℓ (τ, ·)kL2 (Qℓ ) ≤ R is interpreted as Fℓo (τ, ·) ∈ Hℓ and
kFℓo (τ, ·)kHℓ ≤ R. This means that the feature map Fℓo (τ, ·) in each internal layer is well regulated by the RKHS norm.
Moreover, we also assume that the activation function is scale invariant.
Assumption 2 We assume the following conditions on the activation function η.
• η is scale invariant: η(ax) = aη(x) for all a > 0 and x ∈ Rd (for arbitrary d).
• η is 1-Lipschitz continuous: |η(x) − η(x′ )| ≤ kx − x′ k for all x, x′ ∈ Rd .
The first assumption on the scale invariance is essential to derive tight error bounds. The second
one ensures that deviation in each layer does not affect the output so much. The most important
example of an activation function that satisfies these conditions is ReLU activation. Another one is
the identity map η(x) = x.
Finally we assume that the input distribution has a compact support.
Assumption 3 The support of P (X) is compact and it is bounded as
kxk∞ := max |xi | ≤ Dx (∀x ∈ supp(P (X))).
1≤i≤dx
We consider a finite dimensional approximation f ∗ given as follows: let mℓ be the number of
nodes in the ℓ-th internal layer (we set the dimensions of the output and input layers to mL+1 = 1
and m1 = dx ) and consider a model
fℓ∗ (g) = W (ℓ) η(g) + b(ℓ) (g ∈ Rmℓ , ℓ = 2, . . . , L),
f1∗ (x) = W (1) x + b(1) ,
∗
f ∗ (x) = fL∗ ◦ fL−1
◦ · · · ◦ f1∗ (x),
where W (ℓ) ∈ Rmℓ+1 ×mℓ and b(ℓ) ∈ Rmℓ+1 .
7
T. S UZUKI
Theorem 1 (Finite approximation error bound of the nonparametric model) For any 1 > δ >
0 and λℓ > 0, suppose that
mℓ ≥ 5Nℓ (λℓ ) log (32Nℓ (λℓ )/δ) (ℓ = 2, . . . , L),
then there exist W (ℓ) ∈ Rmℓ+1 ×mℓ and b(ℓ) ∈ Rmℓ+1 such that, by letting ĉδ =
kW (ℓ) k2F ≤ ĉδ R2 (ℓ = 1, . . . , L),
kb(ℓ) k ≤ Rb /(1 − δ) (ℓ = 1, . . . , L),
4
1−δ ,
(6a)
(6b)
and
kf o − f ∗ kL2 (P (X)) ≤
∗
kf k∞
q
L
X
p
2 ĉδL−ℓ RL−ℓ+1 λℓ ,
(7)
ℓ=2
L p
X
p
Rb
L
.
≤ ( ĉδ R) Dx +
( ĉδ R)L−ℓ 1−δ
(8)
ℓ=1
The proof is given in Appendix A. This theorem is proven by borrowing the theoretical technique recently developed for the kernel quadrature rule (Bach, 2015). We also employed some techniques analogous to the analysis of the low rank tensor estimation (Suzuki, 2015; Kanagawa et al.,
2016; Suzuki et al., 2016). Intuitively,
the degree of freedom Nℓ (λℓ ) is the intrinsic dimension of
√
the ℓ-th layer to achieve the λℓ approximation error. Indeed, we show √
in the proof that the ℓ-th
layer is approximated by the mℓ dimensional nodes with the precision λℓ under the condition
mℓ = Ω(Nℓ (λℓ ) log(Nℓ (λℓ ))). The error bound (7) indicates that the total approximation
error of
√
λ
of
each
layer
the whole network
is
basically
obtained
by
summing
up
the
approximation
error
ℓ
q
where the factor ĉδL−ℓ RL−ℓ+1 is a Lipschitz constant for error propagation.
We would like to emphasize that the approximation error bound (7) and the norm bounds (6)
of W (ℓ) and b(ℓ) are independent of the dimensions (mℓ )L
ℓ=1 of the internal layers. This is due to
the scale invariance property of the activation function. This is quite beneficial to derive a tight
generalization error bound. Indeed,
invariance, we only have a much looser bound
q without the scale √
P
L
L−ℓ
RL−ℓ+1 λℓ , and kW (ℓ) k2F ≤ mℓ+1 ĉδ R2 , kb(ℓ) k2 ≤
kf o − f ∗ kL2 (P (X)) ≤
ℓ=2 2 mℓ+1 ĉδ
mℓ+1 Rb2 which depend on the dimensions (mℓ )L
ℓ=1 and could be huge for small λℓ . This would
support the practical success of using the ReLU activation.
Let the norm bounds shown in Theorem 1 be
p
R̄ = ĉδ R, R̄b = Rb /(1 − δ).
Remind that Theorem 1 gives an upper bound of the infinity norm of f ∗ , that is, kf ∗ k∞ ≤ R̂∞
where
L
X
R̂∞ = R̄L Dx +
R̄L−ℓ R̄b .
ℓ=1
Let the set of finite dimensional functions with the norm constraint (6) be
F = {f (x) = (W (L) η(·)+b(L) )◦· · ·◦(W (1) x+b(1) ) | kW (ℓ) kF ≤ R̄, kb(ℓ) k ≤ R̄b (ℓ = 1, . . . , L)}.
Then, we can show that the infinity norm of F is also uniformly bounded as the following lemma.
8
FAST L EARNING R ATE OF D EEP L EARNING
Lemma 1 For all f ∈ F, it holds that kf k∞ ≤ R̂∞ .
The proof is given in Appendix A.2. Because of this, we can derive the generalization error bound
with respect to the population L2 -norm instead of the empirical L2 -norm. One can check that
kf k∞ ≤ R̂∞ for all f ∈ F by Lemma 1 or Lemma 3.
4. Generalization error bounds
In this section, we define the two estimators in the finite dimensional model introduced in the last
section: the empirical risk minimizer and the Bayes estimator. The generalization error bounds for
both of these estimators are derived. We also give some examples in which the generalization error
is analyzed in details.
4.1 Notations
Before we state the generalization error bounds, we prepare some notations. Let Ĝ = LR̄L−1 Dx +
P
L
L−ℓ , and define δ̂
3
1,n , δ̂2,n as
ℓ=1 R̄
δ̂1,n =
q
L
X
p
2 ĉδL−ℓ RL−ℓ+1 λℓ ,
ℓ=2
2
δ̂2,n
L
2X
mℓ mℓ+1 log+ 1 +
=
n
ℓ=1
√
√
4 2Ĝ max{R̄,R̄b } n
√PL
.
σ
ℓ=1 mℓ mℓ+1
Note that δ̂1,n is the finite approximation error given in Theorem 1. Roughly speaking, δ̂2,n corresponds to the amount of deviation of the estimators in the finite dimensional model.
4.2 Empirical risk minimization
In this section, we define the empirical risk minimizer and investigate its generalization error. Let
the empirical risk minimizer be fb:
fb := argmin
f ∈F
n
X
(yi − f (xi ))2 .
i=1
Note that there exists at least one minimizer because the parameter set corresponding to F is a
compact set and η is a continuous function. fb needs not necessarily be the exact minimizer but it
could be an approximated minimizer. We, however, assume fb is the exact minimizer for theoretical
simplicity. In practice, the empirical risk minimizer is obtained by using the back-propagation technique. The regularization for the norm of the weight matrices and the bias terms are implemented
by using the L2 -regularization and the drop-out techniques.
The generalization error of the empirical risk minimizer is bounded as in the following theorem.
Theorem 2 For any δ > 0 and λℓ > 0, suppose that
mℓ ≥ 5Nℓ (λℓ ) log (32Nℓ (λℓ )/δ) (ℓ = 2, . . . , L).
3. We define log+ (x) = max{1, log(x)}.
9
(9)
T. S UZUKI
Then, there exists universal constants C1 such that, for any r > 0 and r̃ > 1,
"
(
!
#)
√
2 + σ2)
(
R̂
n
∞
2
2
2
2
o
2
log+
kfb − f kL2 (P (X)) ≤C3 r̃δ̂1,n + (σ + R̂∞ )δ̂2,n +
+r
n
min{σ/R̂∞ , 1}
with probability 1 − exp −
2 (r̃−1)2
nδ̂1,n
11R̂2∞
− 2 exp(−r) for every r > 0 and r̃ ′ > 1.
The proof is given in Appendix C. This theorem can be shown by evaluating the covering number of the model F and applying the local Rademacher complexity technique (Mendelson, 2002;
Bartlett et al., 2005; Koltchinskii, 2006; Giné and Koltchinskii, 2006).
√
2
2
It is easily checked that the third term of the right side ( (R̂∞n+σ ) log+ min{σ/nR̂ ,1} + r ) is
∞
smaller than the first two terms, therefore the generalization error bound can be simply evaluated as
Based on a rough evaluation
2
2
kfb − f o k2L2 = Op (δ̂1,n
+ δ̂2,n
).
2
δ̂1,n
≃L
L
X
ℓ=1
2
λℓ , δ̂2,n
≃
L
X
mℓ mℓ+1
ℓ=1
n
log(n),
and the constraint mℓ & Nℓ (λℓ ) log(Nℓ (λℓ )), we can observe the bias-variance trade-off for the
generalization error bound because, as λℓ decreases, the required width of the internal layer mℓ
increases by the condition (9) and thus the deviation δ̂2,n in the finite dimensional model should increase. In other words, if we want to construct a finite dimensional model which well approximates
the true function, then a more complicated model is required and we should pay larger variance of
the estimator. A key notion for the bias-variance trade-off is the degree of freedom Nℓ (λℓ ) which
expresses the “complexity” of the RKHS Hℓ in each layer. The degree of freedom of a complicated
RKHS grows up faster than a simpler one as λ goes to 0. This is also informative in practice because, to determine the width of each layer, the degree of freedom gives a good guidance. That is,
if the degree of freedom is small compared with the sample size, then we may increase the width
of the layer. An estimate of the degree of freedom can be computed from the trained network by
computing the Gram matrix corresponding to the kernel induced from the trained network (where
the kernel is defined by the finite sum instead of the integral form) and using the eigenvalue of the
Gram matrix.
To obtain the best generalization error bound, (λℓ )L
ℓ=1 should be tuned to balance the biasvariance terms (and accordingly (mℓ )L
should
also
be
fine-tuned). The examples of the best
ℓ=2
achievable generalization error will be shown in Section 4.4.
4.3 Bayes estimator
In this section, we formulate a Bayes estimator and derive its generalization error. To define the
Bayes estimator, we just need to specify the prior distribution. Let Bd (C) be the ball in the Euclidean
space Rd with radius C > 0 (Bd (C) = {x ∈ Rd | kxk ≤ C}), and U(Bd (C)) be the uniform
distribution on the ball Bd (C). Since Theorem 1 ensures the norms of W (ℓ) and b(ℓ) are bounded
above by R̄ and R̄b , it is natural to employ a prior distribution that possesses its support on the set of
10
FAST L EARNING R ATE OF D EEP L EARNING
parameters with norms not greater than those norm bounds. Based on this observation, we employ
uniform distributions on balls with the radii indicated above as a prior distribution:
W (ℓ) ∼ U(Bmℓ+1 ×mℓ (R̄)), b(ℓ) ∼ U(Bmℓ+1 (R̄b )).
In practice, the Gaussian distribution is also employed instead of the uniform distribution. However,
the Gaussian distribution does not give good tail probability bound for the infinity norm of the deep
neural network model. That is crucial to develop the generalization error bound. For this reason,
we decided to analyze the uniform prior distribution.
The prior distribution on the parameters (W (ℓ) , b(ℓ) )L
ℓ=1 induces the distribution of the function f
in the space of continuous functions endowed with the Borel algebra corresponding to the L∞ (Rdx )norm. We denote by Π the induced distribution. Using the prior, the posterior distribution is defined
via the Bayes principle:
Π(df |Dn ) = R
Pn
(yi −f (xi ))2
)Π(df )
i=1
2σ2
.
Pn (yi −f ′ (xi ))2
′)
exp(− i=1
)Π(df
2
2σ
exp(−
Since the purpose of this paper is to give a theoretical analysis for the generalization error,
we do not pursue the computational issue of the Bayesian deep learning. See, for example,
Hernandez-Lobato and Adams (2015); Blundell et al. (2015) for practical algorithms.
The following theorem gives how fast the Bayes posterior contracts around the true function.
Theorem 3 Fix arbitrary δ > 0 and λℓ > 0 (ℓ = 1, . . . , L), and suppose that the condition (9) on
mℓ is satisfied. Then, for all r ≥ 1, the posterior tail probability can be bounded as
q
R̂2∞
o
EDn Π(f : kf − f kL2 (P (X)) ≥ (δ̂1,n + σ δ̂2,n )r max{12, 33 σ2 }|Dn )
"
#
2
2
r2
2 (r − 1)
≤ exp −nδ̂1,n
+ 12 exp −n(δ̂1,n + σ δ̂2,n )2 2 .
2
8σ
11R̂∞
The proof is given in Appendix B. The proof is accomplished by using the technique for
non-parametric Bayes methods (Ghosal et al., 2000; van der Vaart and van Zanten, 2008, 2011).
Roughly speaking this theorem indicates that the posterior distribution concentrates in the distance
δ̂1,n + σ δ̂2,n from the true function f o . The tail probability is sub-Gaussian and thus the posterior mass outside the distance δ̂1,n + σ δ̂2,n from the true function rapidly decrease. Here we again
observe that there appears bias-variance trade-off between δ̂1,n and δ̂2,n . This can be understood
essentially in the same way as the empirical risk minimization.
From the posterior contraction rate, we can derive the generalization error bound of the posterior
mean.
Corollary 1 Under the same setting as in Theorem 3, there exists a universal constant C1 such that
the generalization error of the posterior mean fb is bounded as
h
o
i
n
2
2
R̂∞ 2
σ
2
EDn kfb − f o k2L2 (P (X)) ≤ C1 max 12, 33σR̂2∞ 1 + q
(δ̂1,n + σ 2 δ̂2,n
)+ .
n
nδ̂2
1,n
11
T. S UZUKI
2 /R̂2 ≥ 1 (which is the regime of our interest),
Therefore, for sufficiently large n such that nδ̂1,n
∞
the generalization error is simply bounded as
2
2
2 2
kfb − f o k2L2 (P (X)) = Op max{1, R̂σ∞
}(
δ̂
+
σ
δ̂
)
.
2
1,n
2,n
4.4 Examples
Here, we give some examples of the generalization error bound. We have seen that both of the
empirical risk minimizer and the Bayes estimators have a simplified generalization error bound as
2
2
kfb − f o k2L2 (P (X)) = Op (δ̂1,n
+ δ̂2,n
) = Op
L
L
X
R̄L−ℓ+1 λℓ +
ℓ=2
L
X
mℓ mℓ+1
ℓ=1
n
!
log(n) ,
q
L
by supposing σ, R̂∞ and ĉL
δ R are in constant order. We evaluate the bound under the best choice
of mℓ balancing the bias-variance trade-off.
One way to balance the terms is to set λℓ so that
L
X
λℓ =
ℓ=2
L
X
mℓ mℓ+1
ℓ=1
n
where the log(n)-factor and L are dropped for simplicity. Since the inequality of arithmetic sum
P
P
m2ℓ
mℓ mℓ+1
≤ L+1
geometric mean gives L
ℓ=1 n , we may set mℓ to satisfy
ℓ=1
n
λℓ =
m2ℓ
(ℓ = 2, . . . , L).
n
(10)
Considering this relation and the constraint mℓ & Nℓ (λℓ ) log(Nℓ ) (Eq. (9)), we can estimate the
best width mℓ that minimizes the upper bound of the generalization error.
4.4.1 F INITE
DIMENSIONAL INTERNAL LAYER
If all RKHSs are finite dimensional, say m∗ℓ -dimensional. Then Nℓ (λ) ≤ m∗ℓ for all λ ≥ 0.
Therefore, by setting λℓ = 0 (∀ℓ), the generalization error bound is obtained as
kfb − f o k2L2 (P (X)) .
L
2 X
σ 2 + R̂∞
m∗ℓ m∗ℓ+1 log(n),
n
(11)
ℓ=1
where we omitted the factors depending only on log(R̄R̄b Ĝ). Note that, although there appears the
L∞ -norm bound R̂∞ , this convergence rate is independent of the Lipschitz constant R̄L−ℓ+1 and
R̄b up to log-order but is solely dependent on the number of parameters. Moreover, the convergence
rate is O(log(n)/n) in terms of the sample size n. This is much faster than the existing bounds that
√
utilize the Rademacher complexity because their bounds are O(1/ n). This result matches more
precise arguments for a finite dimensional 3-layer neural network based on asymptotic expansions
(Fukumizu, 1999; Watanabe, 2001) which also showed the generalization error of the 3-layer neural
network can be evaluated as O((m∗1 m∗2 + m∗2 m∗3 )/n).
12
FAST L EARNING R ATE OF D EEP L EARNING
4.4.2 P OLYNOMIAL
DECREASING RATE OF EIGENVALUES
(ℓ)
We assume that the eigenvalue µj decays in polynomial order as
(ℓ)
µ j ≤ aℓ j
− s1
ℓ
,
(12)
for a positive real 0 < sℓ < 1 and aℓ > 0. This is a standard assumption in the analysis of kernel
methods (Caponnetto and de Vito, 2007; Steinwart and Christmann, 2008), and it is known that this
assumption is equivalent to the usual covering number assumption (Steinwart et al., 2009). For
small sℓ , the decay rate is fast and it is easy to approximate the kernel by another one corresponding
to a finite dimensional subspace. Therefore small sℓ corresponds to a simple model and large sℓ
corresponds to a complicated model. In this setting, the degree of freedom is evaluated as
Nℓ (λℓ ) . (λℓ /aℓ )−sℓ .
(13)
This can be shown as follows: for any positive integer M , the degree of freedom can be bounded as
Nℓ (λℓ ) =
∞
X
j=1
(ℓ)
∞
X
µj
(ℓ)
µj
(ℓ)
µj + λℓ
≤M+
j=M +1
aℓ
≤M+
λℓ
λℓ
Z
∞
x−1/sℓ dx
M
≤ M + (aℓ /λℓ )(1 − 1/sℓ )−1 M 1−1/sℓ .
s
Letting M = (aℓ /λℓ )(1 − 1/sℓ )−1 ℓ to balance the first and the second term, we obatain the
evaluation (13). Hence, we can show that, according to Eq. (10),
2sℓ
1+2sℓ
λℓ = aℓ
n
1
− 1+2s
ℓ
gives the optimal rate, and we obtain the generalization error bound as
kfb − f o k2L2 (P (X)) . L
L
2sℓ
X
1
d2
2(L−ℓ+1) 1+2sℓ − 1+2sℓ
(R̄ ∨ 1)
aℓ
log (n) + x log(n),
n
n
(14)
ℓ=2
where we omitted the factors depending on sℓ , log(R̄R̄b Ĝ), σ 2 and R̂∞ . This indicates that the
complexity sℓ of the RKHS affects the convergence rate directly. As expected, if the RKHSs are
simple (that is, (sℓ )L
ℓ=2 are small), we obtain faster convergence.
4.4.3 O NE
INTERNAL LAYER : KERNEL METHOD
Finally, we consider a simple but important situation in which there is only one internal layer (L =
2). In this setting, we only need to adjust m2 because the dimensions of input and output are fixed
as m1 = dx and m3 = 1. We assume that the same condition (12) for ℓ = 2. Then, applying the
condition (10),
− 1
s2
1+s2
(dx + 1)m2
n
1+s2
λ2 =
⇒ λ2 ≃ a2
log(n)
n
dx + 1
gives the optimal convergence rate. Actually, we obtain
s2
1+s2
kfb − f o k2L2 (P (X)) . ((R̄ ∨ 1)2 a2
13
1
)(dx + 1) 1+s2 n
1
− 1+s
2
log(n).
T. S UZUKI
This convergence rate is equivalent to the minimax optimal convergence rate of the kernel ridge regression (Caponnetto and de Vito, 2007; Steinwart et al., 2009) (up to constant and log(n) factors).
It is known that the kernel method corresponds to the 3-layer neural network with an infinite dimensional internal layer. In that sense, our analysis includes that of kernel methods. In particular, the
finite dimensional approximation we performed in Section 3 can be seen as the kernel quadrature
rule (Bach, 2015). Thus, the analysis here ensures that the kernel quadrature rule can achieve the
optimal rate as a byproduct of the neural network analysis. In that sense, we can say that the deep
neural network is a method that constructs an optimal kernel in a layer-wise manner using a kernel
quadrature rule.
Some concrete examples have bee investigated in Bach (2017) for the three layer neural network.
However, the analysis does not assume that the loss function is strongly convex, and thus the local
Rademacher complexity analysis is not applied. Consequently, the convergence rate is slower than
√
O(1/ n).
1
− 1+s
) when L = 2, but we have already observed that the conver1
P
− 1+2s
ℓ : The sample complexity in
gence rate (14) of the deep neural network is basically L
ℓ=2 n
each layer is slow for deep neural network (there is a factor 2 before sℓ ). This is because we need
to estimate a matrix in each layer for deep neural network and the number of output grows up as the
sample size increases, and as a result, the number of parameters that should be estimated is much
larger than the 3-layer neural network in which the dimensions of the input and output are fixed.
The convergence rate is O(n
2
5. Relations to existing work
In this section, we describe the relation of our work to existing works
The sample complexity of deep neural network has been extensively studied by analyzing its
Rademacher complexity. For example, Bartlett (1998) characterized the generalization error of
a 3-layer neural network by the norm of the weight vectors instead of the number of parameters.
Koltchinskii and Panchenko (2002) studied more general deep neural network and derived its generalization error bound of deep neural network under a norm constraint. They showed the Rademacher
complexity of the deep neural network is bounded by the sum of those of single layer neural networks. This is similar to our generalization error bound obtained in Eqs.(14) and (11). More
recently, Neyshabur et al. (2015) analyzed the Rademacher complexity of the deep neural network
based on the norms of the weight matrix ({W (ℓ) }L
ℓ=1 in our paper). Sun et al. (2015) also derived
the Rademacher complexity and showed the generalization error under a large margin assumption.
As consequences of these studies, they derived the following type of inequalities:
n
CR2L
1X b
l(f (xi ), yi ) + √
EX,Y [l(fb(X), Y )] ≤
n
n
i=1
with high probability where l is a loss function, fb is the empirical risk minimizer with or without regularization and R is a norm bound of the internal layers (the definition of the norm differs
between papers). Basically, these studies considered the finite dimensional situation which was
studied in Section 4.4.1 as a special case of our analysis, and a connection to an infinite dimensional model has not been closely discussed. In particular, the bias-variance trade-off has not been
√
analyzed. Moreover, the generalization error is O(1/ n) which is much slower dependency on the
sample size than that of our rate O(1/n). This is a big difference from the existing analysis. To
14
FAST L EARNING R ATE OF D EEP L EARNING
improve the convergence rate to faster one for empirical risk minimization approaches, so called
local Rademacher complexity was important in our analysis. Moreover, we have observed that the
Bayesian analysis also gave faster convergence rate.
Analysis of the bias-variance trade-off in three layer neural network from the kernel point of
view has been investigated by Bach (2017). The analysis is given for several concrete examples.
However, the loss function is not assumed to be strongly convex, and thus the obtained rate is not
√
faster than O(1/ n).
Another important topic for the analysis of the generalization ability is VC-dimension analysis.
VC-dimension of the deep neural network has been studied by, for example, Bartlett et al. (1998);
Karpinski and Macintyre (1997); Goldberg and Jerrum (1995). However, VC-dimension is a notion
independent of the input distributions. On the other hand, the degree of freedom considered in our
paper depends on the input distribution and is more data specific. Hence, our analysis gives tighter
bound and could be practically more useful.
In our analysis, the kernel formulation of deep neural network model was the key ingredient
for the analysis. Some authors have proposed methods that utilize the representation in the internal
layers as a feature map into some RKHS as in our formulation. For example, Cho and Saul (2009);
Mairal et al. (2014); Mairal (2016) have proposed novel methods to construct a kernel via deep
learning. Their purpose is to suggest a new hierarchical method to construct a kernel, and their
proposed methods are different from the ordinary deep learning model. On the other hand, our
theoretical approach states that deep learning itself can be interpreted as a kind of kernel learning.
Moreover, their studies are not for theories but for methodologies. Hence, our analysis and their
studies are in complementary relationship, and our analysis would give theoretical support for their
methods.
6. Conclusion and future work
In this paper, we proposed to use the integral form of deep neural network for generalization error
analysis, and based on that, we derived the generalization error bound of the empirical risk minimizer and the Bayes estimator. The integral form enabled us to define an RKHS in each layer, and
import the theoretical techniques developed in kernel methods into the analysis of deep learning.
In particular, we defined the degree of freedom of each RKHS and showed that the approximation
error between a finite dimensional model and the integral form can be characterized by the degree
of freedom. In addition to the approximation error, we also derived the estimation error in the finite
dimensional model. We have observed that there appears bias-variance trade-off depending on the
size of the finite dimensional model.
Based on the analysis, we also derived generalization error bounds of some examples including
the situation where the eigenvalues of the kernel function decay in a polynomial order, the one
where the true model is finite dimensional, and the one where there is only one internal layer. We
have observed that the analysis of the 3-layer neural network reproduces the optimal learning rate
of the kernel method up to log(n)-order.
Our theoretical frame-work offered a clear description of the bias-variance trade-off for deep
learning. This was particularly useful to determine the optimal widths of the internal layers. We
believe this study opens up a new direction of a series of theoretical analyses of deep learning.
There remain several topics to be studied. One is characterization of the space of the deep neural
network described by the integral form. The integral form can approximate arbitrary function, but
15
T. S UZUKI
under the norm constraints as we have assumed, it is unclear how large the function class is. Solving
this issue is interesting future work.
Acknowledgment
This work was partially supported by MEXT kakenhi (25730013, 25120012, 26280009, 15H01678
and 15H05707), JST-PRESTO and JST-CREST (JPMJCR1304 and JPMJCR14D7).
Appendix A. Proof of Theorem 1
A.1 Approximation error bound for the finite dimensional model
To derive the approximation error bound, we utilize the following proposition that was proven by
Bach (2015).
Proposition 1 For λ > 0, there exists a probability density qℓ (τ ) with respect to the measure Qℓ
such that, for any δ ∈ (0, 1), i.i.d. sample v1 , . . . , vm from qℓ satisfies that
sup
kf kHℓ
inf
4
≤1 β∈Rm :kβk22 ≤ m
with probability 1 − δ, if
f−
m
X
2
βj qℓ (vj )−1/2 η(Fℓ−1 (·, vj ))
j=1
≤ 4λ,
L2 (P (X))
m ≥ 5Nℓ (λ) log(16Nℓ (λ)/δ).
By the scale invaliance of η, η(ax) = aη(x) (a > 0), we have the following proposition based
on Proposition 1.
Proposition 2 For λ > 0, and any 1/2 > δ > 0, if
m ≥ 5Nℓ (λ) log(16Nℓ (λ)/δ),
then there exist v1 , . . . vm ∈ Tℓ , w1 , . . . , wm > 0 such that
sup
inf
2
kf kHℓ ≤R β∈Rm :kβk22 ≤ 4R
m
and
f−
m
X
2
≤ 4λR2 ,
βj η(wj Fℓ−1 (·, vj ))
j=1
L2 (P (X))
m
1 X 2
wj ≤ (1 − 2δ)−1 .
m
j=1
R
1 Pm
−1
−1
Proof Notice that E[ m
=
= E[qℓ (v)−1 ] =
j=1 qℓ (vj ) ]
Tℓ qℓ (v) qℓ (v)dQℓ (v)
R
P
m
1
−1
≤ 1/(1 − 2δ)
j=1 qℓ (vj )
Tℓ 1dQℓ (v) = 1, thus an i.i.d. sequence {v1 , . . . , vm } satisfies m
with probability 2δ by the Markov’s inequality. Combining this with Proposition 1, the i.i.d. sequence {v1 , . . . , vm } and wj = qℓ (vj )−1/2 satisfies the condition in the statement with probability
m
1 − (δ + 1 − 2δ) = δ > 0. This ensures the existence of sequences {vj }m
j=1 and {wj }j=1 that
satisfy the assertion.
16
FAST L EARNING R ATE OF D EEP L EARNING
From now on, we define
c0 = 4, c1 = 4, cδ = (1 − 2δ)−1 .
The next theorem gives the proof of the approximation error bound in Theorem 1. The L∞ -norm
bound of f o is given later in Lemma 3. Substituting δ ← δ/2 into the statement in the following
Lemma 2 and letting ĉδ = c1 cδ/2 , we obtain Theorem 1.
Lemma 2 (Approximation error bound of the nonparametric model) For any 1/2 > δ > 0
and given λℓ > 0, let mℓ ≥ 5Nℓ (λℓ ) log(16Nℓ (λℓ )/δ). Then there exist W (ℓ) ∈ Rmℓ+1 ×mℓ and
b(ℓ) ∈ Rmℓ+1 (ℓ = 1, . . . , L) where mL+1 = 1 and m1 = dx such that
kW (ℓ) k2F ≤ c1 cδ R2 , kb(ℓ) k2 ≤
√
cδ Rb (ℓ = 1, . . . , L − 1),
kW (L) k2F ≤ c1 R2 , kb(L) k2 ≤ Rb ,
and
kf o − f ∗ kL2 (P (X)) ≤
L q
X
(c1 cδ )L−ℓ c0 RL−ℓ+1
ℓ=2
p
λℓ .
Proof
We construct the asserted finite dimensional network recursively from ℓ = L to ℓ = 1. Let
(ℓ)
(ℓ) mℓ
(ℓ) mℓ
ℓ
b
{vj }m
j=1 and {wj }j=1 be the sequences given in Proposition 2. Let Tℓ = {vj }j=1 . With
slight abuse of notation, we identify fℓ∗ : Rmℓ → Rmℓ+1 to a function fℓ∗ : Tbℓ → Tbℓ+1 in a
(ℓ+1)
) to express
canonical way. For a function F : Rdx × Tbℓ → R, we denote by fℓ∗ [F ](x, vi
P
(ℓ+1)
(ℓ)
(ℓ)
(ℓ)
(ℓ+1)
mℓ
∗
f [F (x, ·)](v
) =
W F (x, v ) + b for v
∈ Tbℓ+1 . When we write f ∗ [F ] for
ℓ
i
j=1
i,j
j
i
i
ℓ
× Tbℓ . We define
× Tℓ → R ((x, v) 7→ F (x, v)), we deal with F as a restriction of F on
F :
∗
∗
the output from the ℓ-th layer of the approximated network f as Fℓ (x, v) for v ∈ Tbℓ and x ∈ Rdx .
∗ ](x, v).
More precisely, it is recursively defined as Fℓ∗ (x, v) = fℓ∗ [Fℓ−1
We use an analogous notation for other networks such as fℓo . That is, Fℓo (x, v) = (fℓo ◦ · · · ◦
o ](x, v).
o
f1 (x))(v) for v ∈ Tℓ and x ∈ Rdx , and Fℓo (x, v) = fℓo [Fℓ−1
Step 1 (the last layer, ℓ = L).
We consider the following approximation of the L-th layer (the last layer): Remember that
mL+1 = 1 and thus the output from the L-th layer is just one dimensional. We denote by TL+1 =
{1} which is the index set of the output (which is just a singleton consisting of an element 1). As a
candidate of a good approximation to the true L-th layer, define
Rdx
Rdx
f˜L∗ [FL−1 ](x, 1) =
mL
X
√
1
(L)
(L)
(L)
mL βj η √
w FL−1 (x, vj ) + bL
mL j
(15)
j=1
by β (L) ∈ RmL and w(L) ∈ RmL satisfying kβ (L) k22 ≤ m1L c1 R2 and kw(L) k22 ≤ mL cδ . Here,
define that
√
⊤
(L)
W1,: = mL β (L) , b(L) = (boL (1)).
17
T. S UZUKI
Note that the model (15) can be rewritten as
f˜L∗ [FL−1 ](x, 1) =
mL
X
(L)
(L)
(L)
(L) √
W1,j η( mL −1 wj FL−1 (x, vj )) + b1 .
j=1
Because of Proposition 2 and Assumption 1, the norms of the weight W (L) and the bias b(L) are
bounded as
(L)
kW (L) kF = kW1,: k2 ≤
√
c1 R, kb(L) k2 = |bL | ≤ Rb .
(16)
By the Cauchy-Schwartz inequality and the Lipschitz continuity of η, we have that
′
|f˜L∗ [FL−1 ](x, 1) − f˜L∗ [FL−1
](x, 1)|
m
L
X (L) √ −1 (L)
√
(L)
(L)
(L) ′
≤|
W1,j (η( mL wj FL−1 (x, vj )) − η( mL −1 wj FL−1
(x, vj )))|
j=1
(L)
(L)
(L)
(L) √
′
L
(x, vj )))m
≤ kW1,: k2 mL −1 k(wj (FL−1 (x, vj ) − FL−1
j=1 k2
(L) √
(L)
(L)
′
L
≤ kW1,: k2 mL −1 kw(L) k2 k(FL−1 (x, vj ) − FL−1
(x, vj ))m
j=1 kmax
p
p
(L)
(L)
′
L
≤ c1 R2 cδ mL /mL k(FL−1 (x, vj ) − FL−1
(x, vj ))m
j=1 kmax
√
(L)
(L)
′
L
= c1 cδ Rk(FL−1 (x, vj ) − FL−1
(x, vj ))m
j=1 kmax ,
′
for FL−1 , FL−1
: TbL × Rdx → R. Moreover, Proposition 2 ensures that β (L) and w(L) can be taken
so that
o
o
kf˜L∗ [FL−1
](·, 1) − fLo [FL−1
](·, 1)k2L2 (P (X)) ≤ c0 λL R2 .
Hereinafter, we fix β (L) and w(L) so that this inequality and the norm bound (16) are satisfied.
Step 2 (internal layers for ℓ = 2, . . . , L−1). As for the ℓ-th internal layer, we consider the following
approximation:
(ℓ+1)
)=
f˜ℓ∗ [g](vi
mℓ
X
√
(ℓ+1)
(ℓ) √
(ℓ)
(ℓ)
),
mℓ βi,j η( mℓ −1 wj g(vj )) + boℓ (vi
j=1
(ℓ)
for g : Tbℓ → R with β (ℓ) ∈ Rmℓ+1 ×mℓ and w(ℓ) ∈ Rmℓ satisfying kβj,: k22 ≤ m1ℓ c1 R2 (∀j =
1, . . . , mℓ+1 ) and kw(ℓ) k22 ≤ mℓ cδ . Then, the Lipschitz continuity of f˜ℓ∗ can be shown as
(ℓ+1)
|f˜ℓ∗ [Fℓ−1 ](x, vi
≤
mℓ
X
√
j=1
(ℓ)
(ℓ+1)
′
) − f˜ℓ∗ [Fℓ−1
](x, vi
)|
√
√
(L)
(ℓ)
(ℓ)
(ℓ)
(ℓ) ′
(x, vj )))
mℓ βi,j (η( mℓ −1 wj Fℓ−1 (x, vj )) − η( mℓ −1 wj Fℓ−1
(ℓ)
(ℓ)
′
ℓ
(x, vj ))m
≤ kβi,: k2 kw(ℓ) k2 k(Fℓ−1 (x, vj ) − Fℓ−1
j=1 kmax
r
c1 √
(ℓ)
(ℓ)
′
ℓ
R cδ mℓ k(Fℓ−1 (x, vj ) − Fℓ−1
(x, vj ))m
≤
j=1 kmax
mℓ
18
FAST L EARNING R ATE OF D EEP L EARNING
=
√
(ℓ)
(ℓ)
′
L
(x, vj ))m
c1 cδ Rk(Fℓ−1 (x, vj ) − Fℓ−1
j=1 kmax ,
(ℓ+1)
∈ Tb(ℓ+1) . Proposition 2 asserts that there exit β (ℓ) and w(ℓ) that give an upper bound
for any vi
of the approximation error of the ℓ-th layer as
o
o
max kf˜ℓ∗ [Fℓ−1
](·, vjℓ+1 ) − fℓo [Fℓ−1
](·, vjℓ+1 )k2L2 (P (X)) ≤ c0 λℓ R2 .
j=1,...,mℓ
Finally, let
(ℓ)
Wij =
r
1
mℓ (ℓ) (ℓ+1) (ℓ)
(ℓ+1) o (ℓ+1)
(ℓ+1) o (ℓ+1) ⊤
b (vmℓ+1 )) ,
), . . . , wm
βij wi
, b =√
(w
bℓ (v1
ℓ+1 ℓ
mℓ+1
mℓ+1 1
then, by Assumption 1 and Proposition 2, the norms of these quantities can be bounded as
kW (ℓ) k2F
mℓ+1 mℓ
mℓ X X
(ℓ)2 (ℓ+1)2
=
βij wi
mℓ+1
≤
i=1 j=1
mℓ+1
mℓ X (ℓ+1)2 c1 R2
wi
mℓ+1
mℓ
i=1
≤ c1 cδ R2 ,
and
kb(ℓ) k22
mℓ+1
1 X (ℓ+1) 2 2
Rb ≤ cδ Rb2 .
w
≤
mℓ+1
j=1
Step 3 (the first layer, ℓ = 1).
For the first layer, let
(2)
f˜∗ (x, vi ) =
dx
X
(2)
(2)
ho1 (vi , j)Q1 (j)xj + bo1 (vi )
j=1
(2)
for vi
∈ Tb2 . By the definition of f o , it holds that
(2)
(2)
f˜∗ (x, vi ) = f o (x, vi ).
(2)
(2)
√1 (Q1 (j)w ho (v , j))i,j
1 i
i
m2
(2)
(2)
√1 (w bo (1), . . . , wm2 bo (m2 ))⊤ ∈ Rm2 .
Then, by
1
1
1
m2
Let
W (1)
=
∈
Rm2 ×dx
and
dx
m2 X
X
1 (2) 2 o (2) 2
w h1 (vi , j) Q1 (j)2
=
m2 i
i=1 j=1
!
m2
dx
X
X
1 (2) 2
(2)
≤
max
ho1 (vi , j)2 Q1 (j)2
w
1≤i≤m2
m2 i
i=1
j=1
19
=
Assumption 1 and Proposition 2, it
holds that
kW (1) k2F
b(1)
T. S UZUKI
≤cδ max
1≤i≤m2
and
kb(1) k22
dx
X
j=1
(2)
ho1 (vi , j)2 Q1 (j) ≤ cδ R2 ,
m2
1 X
(2) 2
wi Rb2 ≤ cδ Rb2 .
≤
m1
i=1
Step 4.
Finally, we combine the results we have obtained above. Note that
o
∗
kfLo ◦ fL−1
◦ · · · ◦ f1o − f˜L∗ ◦ f˜L−1
◦ · · · ◦ f˜1∗ kL2 (P (X))
=kf o ◦ f o ◦ · · · ◦ f o − f˜∗ ◦ f o ◦ · · · ◦ f o
L
L−1
1
L
L−1
1
..
.
∗
o
∗
o
+ f˜L∗ ◦ · · · ◦ f˜ℓ+1
◦ fℓo ◦ fℓ−1
◦ · · · ◦ f1o − f˜L∗ ◦ · · · ◦ f˜ℓ+1
◦ f˜ℓ∗ ◦ fℓ−1
◦ · · · ◦ f1o
..
.
∗
˜
+ f ◦ · · · f˜∗ ◦ f o − f˜∗ ◦ · · · f˜∗ ◦ f˜∗ kL (P (X))
L
≤
L
X
ℓ=1
2
1
L
2
1
2
o
∗
∗
o
◦ · · · ◦ f1o kL2 (P (X)) .
◦ f˜ℓ∗ ◦ fℓ−1
◦ · · · ◦ f1o − f˜L∗ ◦ · · · ◦ f˜ℓ+1
kf˜L∗ ◦ · · · ◦ f˜ℓ+1
◦ fℓo ◦ fℓ−1
Then combining the argument given above, we have
o
∗
o
∗
◦ · · · ◦ f1o kL2 (P (X))
◦ f˜ℓ∗ ◦ fℓ−1
◦ · · · ◦ f1o − f˜L∗ ◦ · · · ◦ f˜ℓ+1
◦ fℓo ◦ fℓ−1
kf˜L∗ ◦ · · · ◦ f˜ℓ+1
q
p
p
√
≤( c1 cδ R)L−ℓ ( c0 λℓ R) = (c1 cδ )L−ℓ c0 RL−ℓ+1 λℓ ,
for ℓ = 2, . . . , L. And the right hand side is 0 for ℓ = 1. This yields that
kf − f˜∗ kL2 (P (X)) ≤
o
L
X
R
L−ℓ+1
ℓ=2
q
(c1 cδ )L−ℓ c0
p
λℓ .
By substituting W (ℓ) and b(ℓ) for ℓ = 1, . . . , L defined above into the definition of f ∗ , then it is easy
to see that
f ∗ = f˜∗
as a function. Then, we obtain the assertion.
A.2 Bounding the L∞ -norm
The next lemma shows the L∞ -norm of the true function f o and that of f ∈ F. This gives the
L∞ -norm bound of every f ∈ F in Lemma 1 and thus that of f ∗ in Theorem 1 because f ∗ ∈ F.
20
FAST L EARNING R ATE OF D EEP L EARNING
Lemma 3 Under Assumptions 1, 2 and 3, the L∞ -norms of f o and that of f ∈ F are bounded as
o
L
kf k∞ ≤ R Dx +
L
X
RL−ℓ Rb ,
ℓ=1
L
X√
√
kf k∞ ≤ ( c1 cδ )L RL Dx +
( c1 cδ R)L−ℓ R̄b .
ℓ=1
Proof
Suppose that
o
kFℓ−1
(x, ·)kL2 (Qℓ ) ≤ G.
Then, Fℓo can be bounded inductively: for all τ ∈ Tℓ+1
|Fℓo (x, τ )|
=
≤
Z
o
h◦ℓ (τ, w)η(Fℓ−1
(x, w))dQℓ (w) + boℓ (τ )
Tℓ
o
(x, ·)kL2 (Qℓ ) + |boℓ (τ )|
kh◦ℓ (τ, ·)kL2 (Qℓ ) kFℓ−1
≤ RG + Rb ,
by Assumption 1. Similarly, as for ℓ = 1, it holds that, for all τ ∈ T2 and x ∈ Rdx ,
|f1o (x, τ )| = |
≤|
≤
dx
X
i=1
dx
X
h◦1 (τ, i)xi Q1 (i) + bo1 (τ )|
h◦1 (τ, i)xi Q1 (i)| + |bo1 (τ )|
i=1
kh◦1 (τ, ·)kL2 (Q1 ) kxkL2 (Q1 )
+ Rb
≤ RDx + Rb .
Applying the same argument recursively, we have
kf o k∞ ≤ RL Dx +
L
X
RL−ℓ Rb .
ℓ=1
We can bound the L∞ -norm of any f ∈ F through a similar argument. Note that W (ℓ) satisfies
√
√
kW (ℓ) kF ≤ c1 cδ R for ℓ = 1, . . . , L − 1, W (L) satisfies kW (L) kF ≤ c1 R, and b(ℓ) satisfies
√
kb(ℓ) k2 ≤ cδ Rb by its construction. Therefore, though a similar argument to the bound for f o , we
have that
"L−1
#
Y √
√
√
kf k∞ ≤ c1 R
cδ RDx
( c1 cδ R)
ℓ=2
+
L−2
X
ℓ=1
√
c1 R
"
L−1
Y
ℓ′ =ℓ+1
#
!
√
√
√
√
√
cδ Rb + c1 R cδ Rb + cδ Rb
( c1 cδ R)
21
T. S UZUKI
≤ (c1 cδ )L/2 RL Dx +
L
X
√
( c1 cδ R)L−ℓ R̄b .
ℓ=1
Appendix B. Bounding the posterior contraction rate
In this section, we prove Theorem 3. The proof is divided into two parts: posterior
pPn contraction
2
rate with respect to the in-sample error (i.e., the empirical L2 -norm kf kn =
i=1 f (xi ) /n)
and that with respect to the out-of-sample error (i.e., the population L2 -norm kf kL2 (P (X)) =
qR
f (X)2 dP (X)).
Here, let
ǫn = δ̂1,n + σ δ̂2,n , ǫ̃n = δ̂1,n + δ̂2,n .
B.1 In-sample error
Here we show the in-sample error bound. Let Xn = (x1 , . . . , xn ), Yn = (y1 , . . . , yn ) and Dn =
(Xn , Yn ). For given Xn , the probability distribution of Yn associated with a function f (i.e., yi =
f (xi ) + ǫi ) is denoted by Pn,f . The expectation of a function h of Yn with respect to Pn,f is denoted
by Pn,f (h). The density function of Pn,f with respect to Yn is denoted by pn,f .
For r̃ ≥ 1, let Ar̃ be the event such that
Z
pn,f (Yn )
Π(df ) ≥ exp(−nǫ̃2n r̃ 2 /σ 2 )Π(f : kf − f ∗ k∞ ≤ δ̂2,n r̃).
pn,f o (Yn )
The probability of this event is bounded by Lemma 4.
Using a test function φn defined later (here, a test function is a measurable function of Dn that
takes its value in [0, 1]), we decompose the expected posterior mass as
h
i
√
E Π(kf − f o kn ≥ 2ǫn r|Dn )
≤E [φn ] + P (Ar̃c )
+ E[(1 − φn )1Ar̃ Π(f ∈ F c |Dn )]
+ E[(1 − φn )1Ar̃ Π(f ∈ F : kf − f o k2n ≥ 2ǫr 2 |Dn )]
=:An + Bn + Cn + Dn ,
(17)
for ǫn > 0 where the expectation is taken with respect to Dn = (Xn , Yn ) distributed from the true
distribution. We give an upper bound of An , Bn , Cn and Dn in the following.
Step 1.
√
For arbitrary r ′ > 0, define Cr′ = {f ∈ F | r ′ ≤ nkf − f o kn /σ}. We construct a maximum
√
cardinality set Θr′ ⊂ Cr′ such that each f, f ′ ∈ Θr′ satisfies nkf − f ′ kn /σ ≥ r ′ /2. Here we
denote by D(ǫ, F, k · k) the ǫ-packing number of a normed space F attached with a norm k · k.
22
FAST L EARNING R ATE OF D EEP L EARNING
√
Then, the cardinality of Θr′ is equal to D(r ′ /2, Cr′ , nk · kn /σ). Then, following Lemma 13 of
van der Vaart and van Zanten (2011), one can construct a test φ̃r′ such that
√
√
1 ′2
1 ′2
Pn,f o φ̃r′ ≤ 9D(r ′ /2, Cr′ , nk · kn /σ)e− 8 r ≤ 9D(r ′ /2, F, nk · kn /σ)e− 8 r ,
1 ′2
sup Pn,f (1 − φ̃r′ ) ≤ e− 8 r ,
f ∈Cr ′
for any r ′ > 0. √
√
Substituting 2 nǫn r/σ into r ′ and denoting φn = φ̃√2√nǫn r/σ , we obtain
1
2 2 +log(D(r ′ /2,F ,
Pn,f o φn ≤ 9e− 4σ2 nǫn r
sup
f ∈C2√2√nǫn r
− 12 nǫ2n r 2
4σ
Pn,f (1 − φn ) ≤ e
√
nk·kn /σ))
.
(18)
(19)
√
Hence,
we just need to evaluate the (log-)packing number log(D(r ′ /2, F, nk · kn /σ)) where
√
r ′ = 2nǫn r/σ. It is known that the packing number is bounded from above by the internal
covering number4 , and the packing number of unit ball in d-dimensional Euclidean space and that
of the covering number is bounded as
D(ǫ, Bd (1), k · k) ≤ N (ǫ, Bd (1), k · k) ≤
4+ǫ
ǫ
d
.
Based on this we evaluate the packing number of F.
Let f, f ′ ∈ F be two functions corresponding to parameters (W (ℓ) , b(ℓ) )L
ℓ=1 and
′(ℓ)
′(ℓ)
L
(ℓ)
′(ℓ)
(ℓ)
′(ℓ)
(W , b )ℓ=1 . Notice that if kW − W kF ≤ ǫ and kb − b k ≤ ǫ, then
kf − f ′ k∞ ≤ LǫR̄L−1 Dx +
L
X
ǫR̄L−ℓ = ǫ(LR̄L−1 Dx +
ℓ=1
L
X
R̄L−ℓ ).
(20)
ℓ=1
Therefore, if ǫ ≤ δ/Ĝ where
Ĝ = (LR̄L−1 Dx +
L
X
R̄L−ℓ ),
ℓ=1
then kf − f ′ k∞ ≤ δ. Hence, the packing number of the function space F can be bounded by using
that of the parameter space as
√
√
√
log(D(r ′ /2, F, nk · kn /σ)) = log(D(r ′ /2, F, nk · kn /σ)) ≤ log(D(σr ′ /(2 n), F, k · k∞ ))
√
≤ log(N (σr ′ /(2 n), F, k · k∞ ))
≤
L
X
ℓ=1
L
X
√
√
log(N (σr /(2 nĜ), Bmℓ+1 ×mℓ (R̄), k · k)) +
log(N (σr ′ /(2 nĜ), Bmℓ (R̄b ), k · k))
′
ℓ=1
4. The ǫ-internal covering number of a (semi)-metric space (T, d) is the minimum cardinality of a finite set such that
every element in T is in distance ǫ from the finite set with respect to the metric d. We denote by N (ǫ, T, d) the
ǫ-internal covering number of (T, d).
23
T. S UZUKI
≤
L
X
ℓ=1
mℓ+1 mℓ log
σr ′
√
2 nĜR̄
σr ′
√
2 nĜR̄
4+
+
L
X
ℓ=1
mℓ log
σr ′
√
2 nĜR̄b
σr ′
√
2 nĜR̄b
4+
!
!
√
√
L
L
X
X
4 2ĜR̄
4 2ĜR̄b
=
mℓ+1 mℓ log 1 +
+
.
mℓ log 1 +
ǫn r
ǫn r
ℓ=1
(21)
ℓ=1
Therefore, by Eq. (18), we have that
"
!
!#
√
√
L
L
X
X
1
4
4
2
Ĝ
R̄
2
Ĝ
R̄
b
An ≤ 9 exp − 2 nǫ2n r 2 +
+
.
mℓ log 1 +
mℓ+1 mℓ log 1 +
4σ
ǫn r
ǫn r
ℓ=1
ℓ=1
Step 2. Here, we evaluate Bn . It can be evaluated by Lemma 4 as
2
2
Bn ≤ exp(−nǫ̃2n r̃ 2 /(8σ 2 )) + exp(−nδ̂1,n
(r̃ 2 − 1)2 /(11R̂∞
)).
Step 3. Since F is the support of the prior distribution, it is obvious that Cn = 0.
Step 4. Here, we evaluate Dn . Remind that Dn is defined as
i
h
√
Dn = EXn Pn,f o [Π(f ∈ F : kf − f o kn > 2ǫr|Yn )(1 − φn )1Ar̃ ] .
Define
Ξn (r̃) := − log(Π(f : kf − f ∗ k∞ ≤ δ̂2,n r̃))
for r̃ > 0. Then, Dn can be bounded as
(
"R
#)
√
1{f : kf − f o kn > 2ǫr}pn,f dΠ(f )
F
R
Dn = EXn Pn,f o
(1 − φn )1Ar̃
F pn,f dΠ(f )
√
(
"R
#)
pn,f
o
dΠ(f )
2ǫr} pn,f
o
F 1{f : kf − f kn >
R pn,f
= EXn Pn,f o
(1 − φn )1Ar̃
F pn,f o dΠ(f )
#)
(
"Z
2
2
2
pn,f /pn,f o dΠ(f ) exp(nǫ̃n r̃ /σ + Ξn (r̃))(1 − φn )1Ar̃
≤ EXn Pn,f o
√
= E Xn
≤ exp
(Z
f ∈F :kf −f o kn > 2ǫr
√
f ∈F :kf −f o kn > 2ǫr
Pn,f [(1 −
nǫ2n r 2
nǫ̃2n r̃ 2
+
Ξ
(r̃)
−
n
σ2
4σ 2
φn )1Ar̃ ] exp(nǫ̃2n r̃ 2 /σ 2
.
By using the relation (20), the prior mass Ξn (r̃) can be bounded as
Ξn (r̃) = − log(Π(f : kf − f ∗ k∞ ≤ δ̂2,n r̃))
24
)
+ Ξn (r̃))dΠ(f )
FAST L EARNING R ATE OF D EEP L EARNING
≤ − log(Π(f : kf − f ∗ k∞ ≤ δ̂2,n ))
≤−
−
≤
L
X
ℓ=1
L
X
ℓ=1
L
X
log(Π(W (ℓ) : kW (ℓ) − W ∗(ℓ) kF ≤ δ̂2,n /Ĝ))
log(Π(b(ℓ) : kb(ℓ) − b∗(ℓ) k2 ≤ δ̂2,n /Ĝ))
mℓ mℓ+1 log(R̄Ĝ/(δ̂2,n /2)) +
L
X
mℓ log(R̄b Ĝ/(δ̂2,n /2)).
(22)
ℓ=1
ℓ=1
Step 5. Finally, we combine the results obtained above.
h
i
√
E Π(kf − f o kn ≥ 2ǫn r|Yn )
"
!
!#
√
√
L
L
X
X
1
4 2ĜR̄b
4 2ĜR̄
2 2
≤9 exp − 2 nǫn r +
mℓ log 1 +
+
mℓ+1 mℓ log 1 +
4σ
ǫn r
ǫn r
ℓ=1
ℓ=1
2
2
+ exp(−nǫ̃2n r̃ 2 /(8σ 2 )) + exp(−nδ̂1,n
(r̃ 2 − 1)2 /(11R̂∞
))
nǫ2n r 2
n 2 2
ǫ̃
r̃
+
Ξ
(r̃)
−
.
+ exp
n
n
σ2
4σ 2
(23)
Now, let 1 ≤ r̃ ≤ r. Then, since ǫn ≥ δ̂2,n and r ≥ 1, we have that
!
!
(
!)
√
√
4 2ĜR′
4 2ĜR′
2ĜR′
, log 1 +
,
max log
≤ log 1 +
ǫn r
δ̂2,n
δ̂2,n
for all R′ > 0. Now, we set δ̂2,n to satisfy
!
!
√
√
L
L
2
X
X
nδ̂2,n
4 2ĜR̄
4 2ĜR̄b
≥
mℓ mℓ+1 log 1 +
+
(≥ Ξn (r̃)),
mℓ log 1 +
σ2
δ̂
δ̂
2,n
2,n
ℓ=1
ℓ=1
which can be satisfied by
2
δ̂2,n
√
√
4 2Ĝ max{R̄, R̄b } n
qP
=
.
mℓ mℓ+1 log+ 1 +
n
L
σ
m
m
ℓ=1
ℓ=1 ℓ ℓ+1
2σ 2
L
X
2 ≤ nǫ̃2 and Eq. (22), the RHS of Eq. (23) is upper bounded by
Then, by noticing nδ̂2,n
n
n 2 2 nǫ2n r 2
2 2
2
2
2
2
2
exp(−nǫ̃n r̃ /(8σ )) + exp(−nδ̂1,n (r̃ − 1) /(11R̂∞ )) + 10 exp 2 2 ǫ̃n r̃ −
.
σ
4σ 2
Here, by setting r 2 = 12r̃ 2 ≥ 12, then the RHS is further bounded as
2
2
exp(−nδ̂1,n
(r̃ 2 − 1)2 /(11R̂∞
)) + exp(−nǫ̃2n r̃ 2 /(8σ 2 )) + 10 exp(−nǫ2n r̃ 2 /σ 2 )
h
i
2
2
≤ exp −nδ̂1,n
(r̃ 2 − 1)2 /(11R̂∞
) + 11 exp(−nǫ2n r̃ 2 /(8σ 2 )).
25
(24)
T. S UZUKI
Lemma 4 Then, for any r̃ > 1, it holds that
PDn
Z
pn,f (Yn )
Π(df ) ≥ exp(−nǫ̃2n r̃ 2 /σ 2 )Π(f : kf − f ∗ k∞ ≤ δ̂2,n r̃)
pn,f o (Yn )
2
2
≥ 1 − exp(−nǫ̃2n r̃ 2 /(8σ 2 )) − exp(−nδ̂1,n
(r̃ 2 − 1)2 /(11R̂∞
)).
Proof
Note that Lemma 14 of van der Vaart and van Zanten (2011) showed that
PYn |Xn
Z
pn,f (Yn )
2 2
2
o
Π(df ) ≥ exp(−nǫ̃n r̃ /σ )Π(f : kf − f kn ≤ ǫ̃n r̃) ≥ 1−exp(−nǫ̃2n r̃ 2 /(8σ 2 )).
pn,f o (Yn )
where PYn |Xn represents the conditional distribution of Yn = (yi )ni=1 conditioned by Xn = (xi )ni=1 .
Therefore the proof is reduced to show kf − f o kn ≤ δ̂1,n r̃ + kf − f ∗ k∞ with high probability. Note
that
kf − f o kn ≤ kf − f ∗ kn + kf ∗ − f o kn ≤ kf − f ∗ k∞ + kf ∗ − f o kn .
Hence, we just need to show kf ∗ − f o kn ≤ r̃kf ∗ − f o kL2 (P (X)) ≤ r̃δ̂1,n with high probability.
This can be shown by Bernstein’s inequality:
P
√
1 + r̃ ′ kf ∗ − f o kL2 (P (X)) ≤ kf ∗ − f o kn ≤ exp −
nr̃ ′2 kf ∗ − f o k4L2 (P (X))
2(v + kf ∗ − f o k2∞ kf ∗ − f o k2L2 (P (X)) /3)
where v = EX [((f ∗ (X) − f o (X))2 − kf ∗ − f o k2L2 (P (X)) )2 ]. Now v ≤ EX [(f ∗ (X) − f o (X))4 ] ≤
kf ∗ − f o k2∞ kf ∗ − f o k2L2 (P (X)) = kf ∗ − f o k2∞ kf ∗ − f o k2L2 (P (X)) . This yields that
P
√
∗
o
∗
o
1 + r̃ ′ kf − f kL2 (P (X)) ≤ kf − f kn
"
3nr̃ ′2
≤ exp −
8
kf ∗ − f o kL2 (P (X))
kf ∗ − f o k∞
2 #
.
(25)
Since kf ∗ − f o k∞ ≤ 2R̂∞ and kf ∗ − f o kL2 (P (X)) ≤ δ̂1,n , the RHS is further bounded by
2
3nr̃ ′2 δ̂1,n
exp − 32R̂2
.
∞
2 r̃ ′2
3nδ̂1,n
, it holds that
Therefore, with probability 1 − exp − 32R̂2
∞
kf − f o kn ≤ kf − f ∗ k∞ +
√
1 + r̃ ′ kf ∗ − f o kL2 (P (X)) ≤ kf − f ∗ k∞ +
for all f such that kf k∞ < ∞. Thus by setting r̃ ′ so that r̃ =
26
√
√
1 + r̃ ′ δ̂1,n
1 + r̃ ′ , we obtain the assertion.
!
,
FAST L EARNING R ATE OF D EEP L EARNING
B.2 Out of sample error
Now, we are going to show the posterior contraction rate with respect to the out-of-sample predictive
error:
EDn Π(f : kf − f o kL2 (P (X)) ≥ ǫn r|Dn ) ,
(26)
for sufficiently large r ≥ 1.
To bound the posterior tail, we divide that into four parts:
I = EDn 1Ar̃c ,
i
h
√
II = EDn 1Ar̃ Π(f : 2kf − f o kn > ǫn r, kf k∞ ≤ R̂∞ | Dn ) ,
h
i
√
III = EDn 1Ar̃ Π(f : kf − f o kL2 (P (X)) > ǫn r ≥ 2kf − f o kn , kf k∞ ≤ R̂∞ | Dn ) ,
h
i
IV = EDn 1Ar̃ Π(f : kf k∞ > R̂∞ | Dn ) .
The term I and II are already evaluated in Section B.1, that is, I + II is bounded by the right
hand side of Eq. (17) which is what we have upper bounded in Section B.1.
The term III is bounded as follows. To bound this, we need to evaluate the difference between
the empirical norm kf − f o kn and the expected norm kf − f o kL2 (P (X)) , which can be done by
Bernstein’s inequality. Following the same argument to derive Eq. (25), it holds that
o
P kf − f kL2 (P (X)) ≥
√
o
2kf − f kn ≤ exp −
nkf − f o k2L2 (P (X))
2
11R̂∞
!
.
Therefore, we arrive at the following bound of III:
"
III ≤ EXn Pn,f o
≤
"Z
exp(nǫ̃2n r̃ 2 /σ 2
≤ exp
≤ exp
√
f ∈F :kf −f o kL2 (P (X)) >ǫn r≥ 2kf −f o kn
+ Ξn (r̃))
Z
f ∈F :kf −f o kL2 (P (X)) >ǫn r
nǫ̃2n r̃ 2
nǫ2n r 2
+
Ξ
(r̃)
−
n
2
σ2
11R̂∞
!
2nǫ̃2n r̃ 2
nǫ2n r 2
.
−
2
σ2
11R̂∞
!
#
pn,f /pn,f o dΠ(f ) exp(nǫ̃2n r̃ 2 /σ 2 + Ξn (r̃))1Ar̃
P (kf − f o kL2 (P (X)) ≥
√
#
2kf − f o kn )dΠ(f )
Finally, since all f ∈ F satisfies kf k∞ ≤ R̂∞ , IV = 0.
Combining the results we arrive at
h
i
2
2
EDn Π(f : kf − f o kL2 (P (X)) ≥ ǫn r|Dn ) ≤ exp −nδ̂1,n
(r̃ 2 − 1)2 /(11R̂∞
) + 12 exp −nǫ̃2n r̃ 2 /(8σ 2 ) ,
2 /σ 2 }r̃ 2 . This concludes the proof of Theorem 3.
for all r̃ ≥ 1 and r ≥ max{12, 33R̂∞
27
T. S UZUKI
Appendix C. Convergence rate for the empirical risk minimizer
Proposition 3 (Gaussian concentration inequality (Theorem 2.5.8 in Giné and Nickl (2015)))
Let (ξi )ni=1 be i.i.d. Gaussian sequence with mean 0 and variance σ 2 , and (xi )ni=1 ⊂ X be a given
set of input variables. Then,
Pn for1 a set F̃ of functions from X to R which is separable with respect to
L∞ -norm and supf ∈F̃
i=1 n ξi f (xi ) < ∞ almost surely, it holds that for every r > 0,
P
#
!
"
n
n
X
1
1X
ξi f (xi ) ≥ E sup
ξi f (xi ) + r ≤ exp[−nr 2 /2(σkF̃ kn )2 ]
sup
n
n
f ∈F̃
f ∈F̃
i=1
i=1
where kF̃k2n = supf ∈F̃
1
n
Pn
2
i=1 f (xi ) .
Here the probability is taken with respect to (ξi )ni=1 .
Remind that every f ∈ F satisfies kf kn ≤ kf k∞ ≤ R̂∞ . Hence kFkn ≤ R̂∞ . For an
observation (xi )ni=1 , let Gδ = {f − f ∗ | kf − f ∗ kn ≤ δ, f ∈ F}. It is obvious that Gδ is separable
with respect to L∞ -norm. Then, by the Gaussian concentration inequality, we have that
#
!
"
n
n
X
1
1X
P sup
ξi f (xi ) ≥ E sup
ξi f (xi ) + r ≤ exp[−nr 2 /2(σδ)2 ]
n
f ∈Gδ
f ∈Gδ n
i=1
i=1
√
√
for every r > 0. By applying this inequality for δj = 2j−1 σ/ n for j = 1, . . . , ⌈log 2 (R̂∞ n/σ)⌉
and using the uniform bound, we can show that, for every r > 0, with probability
√
⌈log2 (R̂∞ n/σ)⌉ exp[−nr 2 /2σ 2 ], it holds that any f ∈ Gδ uniformly satisfies
#
"
n
n
X
1
1X
ξi (f (xi ) − f ∗ (xi )) ≥ E sup
ξi f (xi ) + 2δr
n
f ∈G2δ n
i=1
i=1
√
where δ is any positive real satisfying δ ≥ σ/ n and f ∈ Gδ .
Lemma 5 There exists a universal constant C such that for any δ it holds that
v
#
!
"
u PL
n
u
4Ĝ max{R̄, R̄b }
1X
ℓ=1 mℓ mℓ+1
t
.
ξi f (xi ) ≤ Cσδ
log+ 1 +
E sup
n
δ
f ∈G2δ n
i=1
P
Proof Since f 7→ √1n ni=1 ξi f (xi ) is a sub-Gaussian process relative to the metric k · kn . By the
chaining argument (see, for example, Theorem 2.3.6 of Giné and Nickl (2015)), it holds that
"
#
n
√ σ Z 2δ p
1X
E sup
ξi f (xi ) ≤ 4 2 √
log(2N (ǫ, G2δ , k · kn ))dǫ.
n 0
f ∈G2δ n
i=1
PL
m m
R̄,R̄b }
, the
Since log N (ǫ, G2δ , k · kn ) ≤ log N (ǫ, F, k · k∞ ) ≤ 2 ℓ=1 n ℓ ℓ+1 log 1 + 4Ĝ max{
ǫ
right hand side is bounded by
v
!
PL
Z 2δ u
Z 2δ p
u
m
m
4
Ĝ
max{
R̄,
R̄
}
ℓ
ℓ+1
b
tlog(2) + 2 ℓ=1
log(2N (ǫ, F, k · kn ))dǫ ≤
dǫ
log 1 +
n
ǫ
0
0
28
FAST L EARNING R ATE OF D EEP L EARNING
v
u PL
u
mℓ mℓ+1
log+
≤ Cδt ℓ=1
n
!
4Ĝ max{R̄, R̄b }
1+
,
δ
where C is a universal constant. This gives the assertion.
Therefore, by substituting δ ←
lowing inequality holds:
kf −
f ∗ kn
∨σ
q PL
ℓ=1
mℓ mℓ+1
n
√
and r ← σr/ n, the fol-
n
1X
−
ξi (f (xi ) − f ∗ (xi ))
n
i=1
s
v
uP
PL
√
u
L
2
σ
mℓ mℓ+1
4 nĜ max{R̄, R̄b }
ℓ=1 mℓ mℓ+1 u
log+ 1 + qP
≤ Cσ kf − f ∗ kn ∨
t ℓ=1
n
n
L
σ
ℓ=1 mℓ mℓ+1
s
P
σ2 L
∗
ℓ=1 mℓ mℓ+1 √r
+ 2 kf − f kn ∨
σ
n
n
s
2
PL
2
σ
m
m
1
ℓ=1 ℓ ℓ+1
kf − f ∗ kn ∨
≤
4
n
!
!
PL
√
2
m
m
r
4
n
Ĝ
max{
R̄,
R̄
}
b
ℓ=1 ℓ ℓ+1
+4
,
log+ 1 +
+ 2C 2 σ 2
n
σ
n
√
uniformly for all f ∈ F with probability 1 − ⌈log2 (R̂∞ n/σ)⌉ exp[−r 2 /2]. Here let
PL
√
2
mℓ mℓ+1
r
4 nĜ max{R̄, R̄b }
Ψr,n := 2C 2 σ 2 ℓ=1
log+ 1 + qP
+ 4 .
n
n
L
σ
ℓ=1 mℓ mℓ+1
Remind that the empirical risk minimizer in the model F is denoted by fb:
fb := argmin
f ∈F
n
X
(yi − f (xi ))2 .
i=1
Since fb minimizes the empirical risk, it holds that
n
n
i=1
i=1
1X
1X
(yi − fb(xi ))2 ≤
(yi − f ∗ (xi ))2
n
n
n
2X
yi (f ∗ (xi ) − fb(xi )) + kfbk2n − kf ∗ k2n ≤ 0
⇒
n
i=1
n
2X
(ξi + f o (xi ))(f ∗ (xi ) − fb(xi )) + kfbk2n − kf ∗ k2n ≤ 0
⇒
n
i=1
29
T. S UZUKI
n
n
i=1
i=1
2X
2X o
⇒
ξi (f ∗ (xi ) − fb(xi )) +
f (xi )(f ∗ (xi ) − fb(xi )) + kfbk2n − kf ∗ k2n ≤ 0
n
n
n
2X
ξi (f ∗ (xi ) − fb(xi )) + kfb − f o k2n ≤ kf ∗ − f o k2n .
⇒
n
i=1
Therefore, we have
s
2
PL
2
σ
m
m
1 b
ℓ=1 ℓ ℓ+1
−
kf − f ∗ kn ∨
− Ψr,n + kfb − f o k2n ≤ kf ∗ − f o k2n .
4
n
Let us assume kfb − f ∗ k2n ≥
σ2
PL
ℓ=1 mℓ mℓ+1
n
. Then, by Eq. (27), we have
1
− kfb − f ∗ k2n − Ψr,n + kfb − f o k2n ≤ kf ∗ − f o k2n
4
1
1 b
⇒ − kf − f ∗ k2n − Ψr,n + kfb − f ∗ k2n − kf ∗ − f o k2n ≤ kf ∗ − f o k2n
4
2
1 b
∗ 2
∗
o 2
⇒ kf − f kn ≤ 2kf − f kn + Ψr,n .
4
σ2
Otherwise, we trivially have kfb − f ∗ k2n <
Combining the inequalities, it holds that
(27)
PL
(28)
mℓ mℓ+1
.
n
ℓ=1
kfb − f ∗ k2n ≤ 8kf ∗ − f o k2n + 4Ψr,n +
σ2
PL
ℓ=1 mℓ mℓ+1
n
.
(29)
Based on this inequality, we derive a bound for kfb − f ∗ kL2 (P (X)) instead of the empirical L2 -norm
kfb − f ∗ kn .
Proposition 4 (Talagrand’s concentration inequality (Talagrand, 1996; Bousquet, 2002)) Let
(xi )ni=1 be an i.i.d. sequence of input variables in X . Then, for a set F̃ of functions from X to R
which is separable with respect to L∞ -norm and kf k∞ ≤ R̃ for all f ∈ F̃, it holds that for every
r > 0,
"
# s 2 2
n
n
2
X
X
k
F̃
k
r
1
1
r R̃
L2 (P (X))
P sup
f (xi )2 − E[f 2 ] ≥ C E sup
f (xi )2 − E[f 2 ] +
+
f ∈F̃ n
n
n
f ∈F̃ n
i=1
i=1
≤ exp(−r)
where kF̃ 2 k2L2 (P (X)) = supf ∈F̃ E[f (X)4 ].
Let Gδ′ = {f − f ∗ | kf − f ∗ kL2 (P (X)) ≤ δ, f ∈ F}. By the bound kf k∞ ≤ R̂∞ for all f ∈ F
2 δ 2 . Hence,
(Lemma 3), kgk∞ ≤ 2R̂∞ for all g ∈ Gδ′ . Therefore, we have kGδ′2 k2L2 (P (X)) ≤ 4R̂∞
Talagrand’s concentration inequality yields that
"
# s
n
n
2
2
2
X
X
1
1
δ R̂∞ r r R̃
sup
f (xi )2 − E[f 2 ] ≥ C1 E sup
f (xi )2 − E[f 2 ] +
+
f ∈G ′ n
n
n
f ∈Gδ′ n
δ
i=1
i=1
(30)
with probability 1 − exp(−r) where C1 is a universal constant.
30
FAST L EARNING R ATE OF D EEP L EARNING
Lemma 6 There exists a universal constant C > 0 such that, for all δ > 0,
"
#
n
1X
2
2
E sup
f (xi ) − E[f ]
f ∈Gδ′ n i=1
v
"
!
u PL
u
m
m
4
Ĝ
max{
R̄,
R̄
}
ℓ ℓ+1
b
≤ C δR̂∞ t ℓ=1
log+ 1 +
n
δ
!#
PL
4
Ĝ
max{
R̄,
R̄
}
m
m
b
ℓ
ℓ+1
2
ℓ=1
.
log+ 1 +
∨ R̂∞
n
δ
Proof
Let (ǫi )ni=1 be i.i.d. Rademacher sequence. Then, by the standard argument of Rademacher
complexity, we have
"
#
"
#
n
n
X
1
1X
E sup
f (xi )2 − E[f 2 ] ≤ 2E sup
ǫi f (xi )2
f ∈Gδ′ n
f ∈Gδ′ n
i=1
i=1
(see, for example, Lemma 2.3.1 in van der Vaart and Wellner (1996)). Since kf k∞ ≤ 2R̂∞ for
all f ∈ Gδ′ , the contraction inequality (Ledoux and Talagrand, 1991, Theorem 4.12) gives an upper
bound of the RHS as
#
#
"
"
n
n
X
1
1X
ǫi f (xi )2 ≤ 4(2R̂∞ )E sup
ǫi f (xi ) .
2E sup
f ∈Gδ′ n
f ∈Gδ′ n
i=1
i=1
We further bound the RHS. By Theorem 3.1 in Giné and Koltchinskii (2006) or Lemma 2.3 of
Mendelson (2002) with the covering number bound (21), there exists a universal constant C ′ such
that
"
#
n
1X
E sup
ǫi f (xi )
f ∈Gδ′ n i=1
" v
!
u PL
u
m
m
4
Ĝ
max{
R̄,
R̄
}
ℓ
ℓ+1
b
≤ C ′ δt ℓ=1
log+ 1 +
n
δ
!#
PL
4Ĝ max{R̄, R̄b }
ℓ=1 mℓ mℓ+1
∨ R̂∞
.
log+ 1 +
n
δ
This concludes the proof.
Let Φn :=
PL
mℓ mℓ+1
n
√
4 nĜ max{R̄,R̄b }
√PL
R̂∞
ℓ=1 mℓ mℓ+1
. Then, applying the inequality (30) for
log+ 1 +
√
√
δ = 2j−1 R̂∞ / n for j = 1, . . . , ⌈log2 ( n)⌉, it is shown that there exists an event with probability
√
1 − ⌈log2 ( n)⌉ exp(−r) such that, uniformly for all f ∈ F, it holds that
s
n
2
2
X
p
1
R̂∞ r r R̂∞
2
(f (xi ) − f ∗ (xi ))2 − E[(f − f ∗ )2 ] ≤ C1 C(2δR̂∞ Φn ) ∨ (R̂∞
+
Φn ) + δ
n
n
n
ℓ=1
i=1
31
T. S UZUKI
δ2
R̂2 r
2
+ 2C12 (2C 2 + 1)R̂∞
Φn + (C12 + C1 ) ∞ ,
2
n
P
L
2
where δ is any positive real such that δ2 ≥ E[(f − f ∗ )2 ] and δ2 ≥ R̂∞
ℓ=1 mℓ mℓ+1 /n. The right
hand side can be further bounded by
≤
r
δ2
2
+ C2 R̂∞ Φn +
2
n
for an appropriately defined universal constant C2 . Applying this inequality for f = fb to Eq. (29)
gives that
r
1 b
2
+ 8kf ∗ − f o k2n + 4Ψr,n +
Φn +
kf − f ∗ k2L2 (P (X)) ≤ C2 R̂∞
2
n
2
σ 2 + R̂∞
n
!
L
X
mℓ mℓ+1 .
ℓ=1
Finally, by the Bernstein’s inequality (25), the term kf ∗ − f o k2n is bounded as
2
kf ∗ − f o k2n ≤ (1 + r̃ ′ )kf ∗ − f o k2L2 (P (X)) ≤ (1 + r̃ ′ )δ̂1,n
with probability 1 − exp −
2 r̃ ′2
3nδ̂1,n
32R̂2∞
for every r̃ ′ > 0.
Combining all inequalities, we obtain that
L
2 )X
r
2(σ 2 + R̂∞
∗ 2
2
′ 2
b
kf − f kL2 (P (X)) ≤ 2C2 R̂∞ Φn +
+ 16(1 + r̃ )δ̂1,n + 4Ψr,n +
mℓ mℓ+1 .
n
n
ℓ=1
This gives a bound for the distance between fb and f ∗ . However, what we want is a bound
on the distance from the true function f o to fb. This can be accomplished by noticing that
2 ,
kfb − f o k2L2 (P (X)) ≤ 2(kfb − f ∗ k2L2 (P (X)) + kf o − f ∗ k2L2 (P (X)) ) ≤ 2kfb − f ∗ k2L2 (P (X)) + 2δ̂1,n
and conclude that
L
2 )X
4(σ 2 + R̂∞
r
′ 2
o 2
2
b
+ (34 + 32r̃ )δ̂1,n + 8Ψr,n +
mℓ mℓ+1 .
kf − f kL2 (P (X)) ≤ 4C2 R̂∞ Φn +
n
n
ℓ=1
More concisely, letting
α(U ) := U
2
PL
ℓ=1 mℓ mℓ+1
n
√
Ĝ max{R̄,R̄b }
log+ 1 + 4 √nP
,
L
U
the right side is further upper bounded as
(
"
2 + σ2)
(
R̂
∞
o
2
kfb − f kL2 (P (X)) ≤C3 α(R̂∞ ) + α(σ) +
log+
n
with probability 1 − exp −
2 r̃ ′2
3nδ̂1,n
32R̂2∞
ℓ=1
mℓ mℓ+1
√
n
min{σ/R̂∞ , 1}
− 2 exp(−r) for every r > 0 and r̃ ′ > 0.
32
!
#
+ r + (1 + r̃
′
2
)δ̂1,n
)
FAST L EARNING R ATE OF D EEP L EARNING
References
S. Amari. A theory of adaptive pattern classifiers. IEEE Transactions on Electronic Computers,
(3):299–307, 1967.
N. Aronszajn. Theory of reproducing kernels. Transactions of the American Mathematical Society,
68:337–404, 1950.
F. Bach. On the equivalence between kernel quadrature rules and random feature expansions. arXiv
preprint arXiv:1502.06800, 2015.
F. Bach. Breaking the curse of dimensionality with convex neural networks. Journal of Machine
Learning Research, 18(19):1–53, 2017.
A. K. Balan, V. Rathod, K. P. Murphy, and M. Welling. Bayesian dark knowledge. In Advances in
Neural Information Processing Systems 28, pages 3438–3446, 2015.
P. Bartlett, O. Bousquet, and S. Mendelson. Local Rademacher complexities. The Annals of Statistics, 33:1487–1537, 2005.
P. L. Bartlett. The sample complexity of pattern classification with neural networks: the size of
the weights is more important than the size of the network. IEEE transactions on Information
Theory, 44(2):525–536, 1998.
P. L. Bartlett, V. Maiorov, and R. Meir. Almost linear VC-dimension bounds for piecewise polynomial networks. Neural Computation, 10(8):2159–2173, 1998.
M. Bianchini and F. Scarselli. On the complexity of neural network classifiers: A comparison
between shallow and deep architectures. IEEE transactions on neural networks and learning
systems, 25(8):1553–1565, 2014.
C. Blundell, J. Cornebise, K. Kavukcuoglu, and D. Wierstra. Weight uncertainty in neural network. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of
Proceedings of Machine Learning Research, pages 1613–1622, 2015.
O. Bousquet. A Bennett concentration inequality and its application to suprema of empirical process. Comptes Rendus de l’Académie des Sciences -Series I- Mathematics, 334:495–500, 2002.
A. Caponnetto and E. de Vito. Optimal rates for regularized least-squares algorithm. Foundations
of Computational Mathematics, 7(3):331–368, 2007.
Y. Cho and L. K. Saul. Kernel methods for deep learning. In Advances in Neural Information
Processing Systems 22, pages 342–350, 2009.
N. Cohen and A. Shashua. Convolutional rectifier networks as generalized tensor decompositions.
In Proceedings of the 33th International Conference on Machine Learning, volume 48 of JMLR
Workshop and Conference Proceedings, pages 955–963, 2016.
N. Cohen, O. Sharir, and A. Shashua. On the expressive power of deep learning: A tensor analysis.
In Proceedings of the 29th Annual Conference on Learning Theory, pages 698–728, 2016.
33
T. S UZUKI
G. Cybenko. Approximation by superpositions of a sigmoidal function. Mathematics of Control,
Signals, and Systems (MCSS), 2(4):303–314, 1989.
R. Eldan and O. Shamir. The power of depth for feedforward neural networks. In Proceedings of
the 29th Annual Conference on Learning Theory, pages 907–940, 2016.
K. Fukumizu. Generalization error of linear neural networks in unidentifiable cases. In International
Conference on Algorithmic Learning Theory, pages 51–62. Springer, 1999.
Y. Gal and Z. Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty
in deep learning. In Proceedings of the 33nd International Conference on Machine Learning,
volume 48 of JMLR Workshop and Conference Proceedings, pages 1050–1059, 2016.
S. Ghosal, J. K. Ghosh, and A. W. van der Vaart. Convergence rates of posterior distributions. The
Annals of Statistics, 28(2):500–531, 2000.
E. Giné and V. Koltchinskii. Concentration inequalities and asymptotic results for ratio type empirical processes. The Annals of Probability, 34(3):1143–1216, 2006.
E. Giné and R. Nickl. Mathematical Foundations of Infinite-Dimensional Statistical Models. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, 2015.
X. Glorot, A. Bordes, and Y. Bengio. Deep sparse rectifier neural networks. In Proceedings of the
14th International Conference on Artificial Intelligence and Statistics, volume 15 of Proceedings
of Machine Learning Research, pages 315–323, 2011.
P. W. Goldberg and M. R. Jerrum. Bounding the Vapnik-Chervonenkis dimension of concept classes
parameterized by real numbers. Machine Learning, 18(2-3):131–148, 1995.
J. Hartigan et al. The maximum likelihood prior. The Annals of Statistics, 26(6):2083–2103, 1998.
J. M. Hernandez-Lobato and R. Adams. Probabilistic backpropagation for scalable learning of
bayesian neural networks. In Proceedings of the 32nd International Conference on Machine
Learning, volume 37 of Proceedings of Machine Learning Research, pages 1861–1869, 2015.
K. Hornik. Approximation capabilities of multilayer feedforward networks. Neural networks, 4(2):
251–257, 1991.
H. Kanagawa, T. Suzuki, H. Kobayashi, N. Shimizu, and Y. Tagami. Gaussian process nonparametric tensor estimator and its minimax optimality. In Proceedings of the 29th International
Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research,
pages 1632–1641, 2016.
M. Karpinski and A. Macintyre. Polynomial bounds for VC dimension of sigmoidal and general
pfaffian neural networks. Journal of Computer and System Sciences, 54(1):169–176, 1997.
V. Koltchinskii. Local Rademacher complexities and oracle inequalities in risk minimization. The
Annals of Statistics, 34(6):2593–2656, 2006.
V. Koltchinskii and D. Panchenko. Empirical margin distributions and bounding the generalization
error of combined classifiers. The Annals of Statistics, 30(1):1–50, 2002.
34
FAST L EARNING R ATE OF D EEP L EARNING
A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional
neural networks. In Advances in Neural Information Processing Systems 25, pages 1097–1105,
2012.
M. Ledoux and M. Talagrand. Probability in Banach Spaces. Isoperimetry and Processes. Springer
Berlin Heidelberg, 1991.
J. Mairal. End-to-end kernel learning with supervised convolutional kernel networks. In Advances
in Neural Information Processing Systems 29, pages 1399–1407. 2016.
J. Mairal, P. Koniusz, Z. Harchaoui, and C. Schmid. Convolutional kernel networks. In Advances
in Neural Information Processing Systems 27, pages 2627–2635, 2014.
S. Mendelson. Improving the sample complexity using global data. IEEE Transactions on Information Theory, 48(7):1977–1991, 2002.
G. F. Montufar, R. Pascanu, K. Cho, and Y. Bengio. On the number of linear regions of deep neural
networks. In Advances in Neural Information Processing Systems 27, pages 2924–2932. 2014.
V. Nair and G. E. Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th International Conference on Machine Learning, pages 807–814, 2010.
B. Neyshabur, R. Tomioka, and N. Srebro. Norm-based capacity control in neural networks. In
Proceedings of the 28th Conference on Learning Theory, volume 40 of Proceedings of Machine
Learning Research, pages 1376–1401, 2015.
B. Poole, S. Lahiri, M. Raghu, J. Sohl-Dickstein, and S. Ganguli. Exponential expressivity in deep
neural networks through transient chaos. In Advances in Neural Information Processing Systems
29, pages 3360–3368. 2016.
D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning representations by back-propagating
errors. Cognitive modeling, 5(3):1.
I. Safran and O. Shamir. Depth separation in ReLU networks for approximating smooth non-linear
functions. arXiv preprint arXiv:1610.09887, 2016.
S. Sonoda and N. Murata. Neural network with unbounded activation functions is universal approximator. Applied and Computational Harmonic Analysis, 2015.
N. Srivastava, G. E. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple
way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):
1929–1958, 2014.
I. Steinwart and A. Christmann. Support Vector Machines. Springer, 2008.
I. Steinwart and C. Scovel. Mercer’s theorem on general domains: On the interaction between
measures, kernels, and rkhss. Constructive Approximation, 35(3):363–417, 2012.
I. Steinwart, D. Hush, and C. Scovel. Optimal rates for regularized least squares regression. In
Proceedings of the 22nd Annual Conference on Learning Theory, pages 79–93, 2009.
35
T. S UZUKI
S. Sun, W. Chen, L. Wang, and T.-Y. Liu. Large margin deep neural networks: theory and algorithms. arXiv preprint arXiv:1506.05232, 2015.
T. Suzuki. Convergence rate of Bayesian tensor estimator and its minimax optimality. In Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of
Machine Learning Research, pages 1273–1282, 2015.
T. Suzuki, H. Kanagawa, H. Kobayashi, N. Shimizu, and Y. Tagami. Minimax optimal alternating minimization for kernel nonparametric tensor learning. In Advances In Neural Information
Processing Systems 29, pages 3783–3791, 2016.
M. Talagrand. New concentration inequalities in product spaces. Inventiones Mathematicae, 126
(3):505–563, 1996.
A. W. van der Vaart and J. H. van Zanten. Rates of contraction of posterior distributions based on
Gaussian process priors. The Annals of Statistics, 36(3):1435–1463, 2008.
A. W. van der Vaart and J. H. van Zanten. Information rates of nonparametric Gaussian process
methods. Journal of Machine Learning Research, 12:2095–2119, 2011.
A. W. van der Vaart and J. A. Wellner. Weak Convergence and Empirical Processes: With Applications to Statistics. Springer, New York, 1996.
S. Watanabe. Learning efficiency of redundant neural networks in bayesian estimation. IEEE
Transactions on Neural Networks, 12(6):1475–1486, 2001.
B. Widrow and M. E. Hoff. Adaptive switching circuits. In IRE WESCON convention record, pages
96–104. Institute of Radio Engineers, 1960.
36
| 10math.ST
|
Object-Oriented Knowledge Extraction using
Universal Exploiters
Dmytro Terletskyi
Faculty of Computer Science and Cybernetics
Taras Shevchenko National University of Kyiv
Kyiv, Ukraine
[email protected]
Abstract — This paper contains analysis and extension of
exploiters-based knowledge extraction methods, which allow
generation of new knowledge, based on the basic ones. The main
achievement of the paper is useful features of some universal
exploiters proof, which allow extending set of basic classes and
set of basic relations by finite set of new classes of objects and
relations among them, which allow creating of complete lattice.
Proposed approach gives an opportunity to compute quantity of
new classes, which can be generated using it, and quantity of
different types, which each of obtained classes describes;
constructing of defined hierarchy of classes with determined
subsumption relation; avoidance of some problems of inheritance
and more efficient restoring of basic knowledge within the
database.
Keywords — knowledge extraction; object-oriented dynamic
networks; universal exploiters; lattice; semilattice; inheritance;
hierarchies of classes.
I.
INTRODUCTION
During recent years application of knowledge-based
systems has extremely increased, therefore variety of systems
and knowledge bases for different domains were developed. In
spite of this, the invention of efficient methods for knowledge
representation (KR), inference and extraction is still topical
issue.
Nowadays there are many knowledge representation
formalisms (KRFs), which are used for knowledge-based
systems (KBSs) development. Currently the most commonlyused approaches are semantic networks, ontologies, logical and
rule-based formalisms. However, the certain programming
paradigm, language and some stack of programming
technologies should be chosen for development of a KBS. This
choice is very important, because each programming paradigm
and language provides certain tools for system development
and determined mechanisms of interaction among its modules,
in particular interaction with database. Thus, chosen KRF and
programming technologies for its implementation, should be at
least compatible with respect to each other. Otherwise,
developed KBS will have complicated interaction between the
level of KRF and the level of its implementation. Consequently
it can decrease the efficiency of such system. Despite this,
chosen formalism should provide efficient representation of
hierarchically-structured knowledge about particular domain,
because concepts hierarchy makes KR more compact and
allows performing of reasoning over itself. Furthermore, the
hierarchy should be stored in the database in such way, that
KBS can be able to extract the knowledge efficiently and
represent them in terms of programming language, using which
the system was developed. However, the representation of
hierarchies is possible, only if chosen KRF and programming
language support mechanism of inheritance.
Currently, the most commonly used programming paradigm
is an object-oriented programming (OOP). All OOP-languages
and many KRFs support single inheritance. However, as it was
shown in [1]–[3], inheritance mechanism causes problem of
exceptions, redundancy and ambiguity, which usually arise
during construction of hierarchies and reasoning over them.
II.
KNOWLEDGE EXTRACTION
According to [4]–[7], knowledge extraction is defined as
creation or acquiring of knowledge from structured (e.g.
relational databases, object-oriented database models, UML,
XML and their fuzzy extensions, proposed in [8], [9]), semistructured (e.g. infoboxes) and unstructured (e.g. text,
documents, images) data sources. In addition, the extracted
knowledge should be represented in machine-processable
format that enables inference.
According to [6], there are two main paradigms of KE:
ontology-based and open domain. They also can be called as
close world knowledge extraction and open world knowledge
extraction. The idea of first approach is to use ontology as
vocabulary, which defines the types of concepts used in the
knowledge base. It means that knowledge base contains
defined number of types of entities and relationships. Thus,
only relations included in the vocabulary can be extracted from
the knowledge base.
In the second approach, knowledge-based system does not
have any vocabulary and pre-specified relationship types in the
knowledge base. It means that each entity or relation in
knowledge base can be considered as a candidate. Therefore,
any possible relation or assertion in the knowledge base can be
extracted.
However, Unbehauen, Hellmann, Auer, Stadler et al. in [4],
[5] argued about absence of clear definition of what extracted
knowledge is and paid attention to the fact, that mere usage
such KRFs as RDF/OWL can not sufficiently define the notion
of «knowledge». They have formulated two important
questions:
1. What is the result of data representation in terms of
RDF/OWL (triplification process)? Structured data or
represented knowledge?
2. When does structured data became knowledge?
Analyzing these questions, it is possible to conclude that result
of such knowledge extraction, first of all, will be structured
data, which then can be interpreted as some knowledge.
However, such interpretation can be performed only using
particular KRF, where notion of knowledge is defined in a
proper way. Therefore, any KRF can be considered as
interpreter of data, according to its own specifics and specifics
of particular domain, for representation of which the formalism
was developed.
One of attempts to solve earlier mentioned problems is such
KRF as object-oriented dynamic networks (OODN), which
was proposed in [10]. It provides representation of knowledge
in OOP-like style and is compatible with respect to many OOPlanguages. In addition, as it was demonstrated in [3], OODN
allow constructing of polyhierarchies and avoiding, in many
cases, problems of inheritance, which were mentioned above.
Moreover, OODN have fuzzy extension, proposed in [11],
[12], which provides representation of vague and imprecise
knowledge, using the same structure as for the crisp case. One
more feature of OODN is exploiters-based knowledge
extraction (KE) methods, which provide generating of finitely
defined set of new classes of objects and finitely set of new
relations among them, based on the set of basic classes and
relations among them. It allows calculation of quantity of new
classes, which can be extracted, and quantity of different types,
which each obtained class describes. Furthermore, according to
[13], the set of basic classes of any OODN, extended by
extracted classes, together with union exploiter, create upper
semilattice. Constructed upper semilattice forms a hierarchy of
classes, where each class satisfies subsumption relation defined
over the hierarchy that makes it possible to find more general
class for arbitrary pair of classes. Such approach allows
extracting of new knowledge from the basic ones and provides
an ability to reconstruct the knowledge base for increasing its
compactness.
III.
UNIVERSAL EXPLOITERS AND KNOWLEDGE
EXTRACTION
As it was shown in [12], some universal exploiters can be
efficiently used for KE. According to [12, Th. 1], all possible
applications of union exploiter, including all its possible
superpositions, to homogeneous classes of objects, which do
not have common properties and methods, always generate
finite quantity of new classes of objects, which can be precisely
calculated.
However, there are situations when homogeneous classes of
objects can have common properties and (or) methods. Before
we start to consider them, let us make clear what we mean by
type, subtype and subclass. As it was mentioned in [13],
inhomogeneous class of objects describes at least two different
types of objects within one class, where type is defined as
follows.
Definition 1. Type of objects t i of arbitrary inhomogeneous
class of objects T (Core(T ), pr1 (T ),..., prn (T )) , which
describes types t1 ,..., t n , is a homogeneous class of objects
ti (Core(T ), pri (ti )) , where i 1, n .
Consequently, each homogeneous class of objects describes
particular type of objects. The definition shows that type and
class of objects does not always mean the same, more
precisely, homogeneous class of objects is equivalent to type of
objects, however inhomogeneous class of objects is not
equivalent to type of objects, because it describes some set of
types. Now let us define notion of subtype.
Definition 2. Arbitrary type of objects
arbitrary type of objects
t1 is a subtype of
t2 , i.e. t1 t2 if and only if
Pt1 Pt2 F t1 F t2 ,
where P t1 , P t 2 are specifications of types
F t1 , F t2 are their signatures.
t1 , t2 and
This definition actually defines the notion of subclass for
the case of homogeneous classes, however it is not enough for
the inhomogeneous classes of objects. The notion of subclass
for inhomogeneous classes was introduced in [13], nevertheless
it is restricted and does not take into account some cases, when
classes of objects have common properties and methods. Let us
consider an example for clear understanding.
Suppose we have three homogeneous classes of objects
T1 p1 T1 ,..., pn T1 , f1 T1 ,..., f m T1 ,
T2 p1 T2 ,..., pk T2 , f1 T2 ,..., f w T2 ,
T3 p1 T3 ,..., pv T3 , f1 T3 ,..., f r T3 .
Let us assume that
p1 T1 p1 T2 p1 T3 ; p2 T1 p2 T2 p2 T3 ,
p3 T1 p3 T2 p3 T3 ; p4 T1 p4 T2 .
Using union exploiter, let us compute
T1 T2 , T1 T3 and
T1 T2 T3 , i.e.
T1 T2 T12 CoreT12 , pr1 t1 , pr2 t2
p1 T12 , p2 T12 , p3 T12 , p4 T12 ,
p1 t2 ,..., pk t2 , f1 t2 ,..., f w t2 ,
p5 t1 ,..., pn t1 , f1 t1 ,..., f m t1 ,
p1 t3 ,..., pv t3 , f1 t3 ,..., f r t3 .
p5 t2 ,..., pk t2 , f1 t2 ,..., f w t2 .
According to [13, Def. 12], T13 T123 and T12 T123 ,
T1 T3 T13 CoreT13 , pr1 t1 , pr2 t3
therefore all results, which were presented in [13] are correct.
That is why, let us extend the notion of subclass given in [13],
using Def. 1 and Def. 2.
p1 T13 , p2 T13 , p3 T13 ,
p4 t1 ,..., pn t1 , f1 t1 ,..., f m t1 ,
p4 t3 ,..., pv t3 , f1 t3 ,..., f r t3 .
T1 T2 T3 T123
CoreT123 , pr1 t1 , pr2 t 2 , pr3 t3
p1 T123 , p2 T123 , p3 T123 ,
p4 t1 ,..., pn t1 , f1 t1 ,..., f m t1 ,
p4 t2 ,..., pk t2 , f1 t2 ,..., f w t2 ,
p4 t3 ,..., pv t3 , f1 t3 ,..., f r t3 .
According to [13, Def. 12], T13 T123 , however T12 T123 .
Nevertheless, according to Def. 1 and Def. 2,
T1 CoreT123 , pr1 t1 ,
T2 CoreT123 , pr1 t1 .
T1 , which describes
is a subclass of arbitrary class of objects T2 ,
Definition 3. Arbitrary class of objects
1
1
t ,..., t
types
1
n,
which describes types
t12 ,..., tm2 , i.e. T1 T2 if and only if
ti1t 2j | ti1 t 2j , where i 1, n , j 1, m and n, m 1 .
T12 , T13 and T123
from Example 1, we can conclude that T12 T123 and
T13 T123 for both cases, when classes T1 , T2 and T3 have
Now, using this definition for classes
common properties and methods and when they do not have
them.
Let us consider homogeneous classes of objects T1 ,..., Tn ,
which describes types of objects t1 ,..., t n . Let us assume that
t , that t t1 , t t2 ,..., t tn . It means
that classes of objects T1 ,..., Tn have some common properties
there is such type
and (or) methods. It is clearly, that the application of union
exploiter to them will produce a set of new classes of objects.
Using this idea, let us formulate and prove the following
theorem.
Despite this, [13, Def. 12] is correct for the case when classes
of objects have no common properties and methods. Let us
assume that classes T1 , T2 and T3 do not have common
Theorem 1. For any
properties and methods, then we have
where T1 ,..., Tn are homogeneous classes, which describe
T1 T2 T12 pr1 t1 , pr2 t2
p1 t1 ,..., pn t1 , f1 t1 ,..., f m t1 ,
p1 t2 ,..., pk t2 , f1 t2 ,..., f w t2 .
T1 T3 T13 pr1 t1 , pr2 t3
p1 t1 ,..., pn t1 , f1 t1 ,..., f m t1 ,
p1 t3 ,..., pv t3 , f1 t3 ,..., f r t3 .
T1 T2 T3 T123
pr1 t1 , pr2 t 2 , pr3 t3
p1 t1 ,..., pn t1 , f1 t1 ,..., f m t1 ,
OODN O, C T1 ,..., Tn , R, E {}, M ,
types of objects t1 ,..., t n and there is a type
t , such that
t t1 , t t 2 ,..., t t n , all possible applications of union
exploiter, including all possible its superpositions, to classes of
objects from the set C and obtaining classes of objects using
union exploiter, always generate finite quantity of new classes
of objects, which can be precisely calculated by the following
formula:
q CE 2n n 1 ,
where
n C .
Proof: According to definition of union exploiter for classes
of objects [13, Def. 14], the result of union of two arbitrary
nonequivalent classes of objects T1 and T2 , which describe
type of objects
t1 and t2 respectively, is inhomogeneous class
of objects T , which describes both these types. If there is a
type t , such that t t1 , t t 2 ,..., t t n , then class T will
have the following structure
T Core(T ), pr1 (t1 ), pr2 (t2 ) .
from the set
k 2, n different classes
C . It is known that
k
C
k
n
2n .
However, we cannot create classes of objects, which describe
1 and 0 different types, applying union exploiter to the
0
n
C , i.e. we do not count C and
1
n.
C Therefore, we can conclude that
n
n
q CE C C C C 2 n 1 . ■
k
n
1
n
3.
T1 T2 T3 T1 T2 T3 ,
that T1 TJSL TJSL , where TJSL
it follows
T1 ... Tn .
C is a poset. For this we
T2 T1 T2 T2 and
show that is a relation of partial order under the set C .
Now we need to prove that
should define T1 , T2 C | T1
Taking into account that we have two types of classes, we need
to define three kinds of relation, i.e.
n 0
classes of objects from the set
T1 T2 T2 T1 ,
where T1 , T2 , T3 C . From the definition of
According to proof of [13, Th. 1], the number of all
possible unique classes of objects created from the basic set of
classes C T1 ,..., Tn using union exploiter can be
represented as a combination of
2.
0
n
k
n
k 0
1.
homogeneous
homogeneous,
2.
homogeneous
inhomogeneous,
3.
inhomogeneous
It was done in Def. 2 and Def. 3. Now let us prove reflexivity,
anti-symmetry and transitivity of these relations.
1.
Reflexivity: T1 T1
idempotency of ;
2.
Anti-symmetry:
n
k 2
inhomogeneous.
T1 T1 T1 follows from
Similarly to [13, Th. 2], we can formulate and prove the
following theorem.
T1 T2 T1 T2 T2 ,
Theorem 2. Set of classes of objects
T2 T1 T2 T1 T1
C T1 ,..., Tn , Tn1 ,..., T2n 1
of any OODN, extended according to Th. 1, with union
exploiter create the join-semilattice JSL (C , E {,1}) ,
where class TJSL T1 ... Tn is its greatest upper bound,
and from commutativity of
T1 T2 ;
3.
Transitivity:
T1 T2 T1 T2 T2 , T2 T3 T2 T3
i.e. 1 .
Proof: According to the definition of join-semilattice given
in [14], [15], it is a system JSL ( A, {,1}) , where A
is a poset, is a binary, idempotent, commutative and
associative operation and 1 is an unary operation, which are
defined over the set A . In addition a A , 1 satisfies
L1 : a 1 1 (identity law).
According to the theorem, carrier of join-semilattice is the
set of classes C , set of exploiters E contains binary operation
and unary operation 1 , which are defined over the set C .
Therefore JSL (C , E {,1}) , where
C T ,..., T , T
1
n
n 1
,..., T2n 1
.
From the [13, Def. 14] it follows, that mentioned properties
of are also true for , i.e.
1.
T1 T1 T1 ,
, we can conclude that
T3 T1 T2 T3 T1 T2 T3
T1 T3 T3 T1 T3 T3 T1 T3 .
Therefore,
JSL C T1 ,..., Tn , Tn1 ,..., T2n 1, E {,1}
is a join-semilattice, where class TJSL T1 ... Tn is its
greatest upper bound, i.e. 1 .
■
Now let us define intersection exploiter for classes of
objects, using Def. 1.
Definition 4. Intersection of two arbitrary nonequivalent
classes of objects T1 T2 , which describe types of objects
t11 ,..., tn1 and t12 ,..., tm2 , respectively, where n, m 1 , is
inhomogeneous class of objects T , which describes types of
3
3
objects t1 ,..., t w , where w 1, such that
t , t
3
k
1
i
t t t t ,
1
i
l
where
C T1 ,..., Tn , Tn1 ,..., T2n 1
t 2j | tk3 ti1 tk3 t 2j tl | tk3 tl
2
j
l
k 1, w , i 1, n , j 1, m .
Using this definition, let us formulate and prove the
following theorem.
Theorem 3. For any
OODN O, C T1 ,..., Tn , R, E {}, M ,
of any OODN, extended according to Th. 3, with exploiter
create the meet-semilattice MSL (C , E {,0}) , where
class TMSL T1 ... Tn is its least lower bound, i.e.
Proof: According to definition of meet-semilattice given in
[14], [15], it is a system MSL ( A, {,0}) , where A
is a poset, is binary, idempotent, commutative and
associative operation and 0 is unary operation, which are
defined over the set A . In addition, a A , 0 satisfies
L1 : a 0 0 (identity law).
where T1 ,..., Tn are homogeneous classes, which describe
types of objects t1 ,..., t n and there is a type t , such that
t t1 ... t tn t | t t
t t1 ... t t n ,
all possible applications of intersection exploiter, including all
possible its superpositions, to classes of objects from the set C
and obtaining classes of objects, using intersection exploiter,
always generate finite quantity of new classes of objects, which
can be precisely calculated by the following formula:
q CE 2n n 1 ,
where
n C .
Proof: According to Def. 4, the result of intersection of two
arbitrary nonequivalent classes of objects T1 and T2 is
inhomogeneous class of objects T that describes subtypes,
which are common for all types of class T1 and T2
simultaneously.
It is known that the number of all possible unique classes of
objects created from the basic set of classes C T1 ,..., Tn
using intersection exploiter can be represented as a
combination of
known that
k 2, n different classes from the set C . It is
k
C
k
n
2 .
However, intersection exploiter is a binary operation, that is
why we cannot count
According to the theorem, carrier of meet-semilattice is the
set of classes C , set of exploiters E contains binary operation
and unary operation 0 , which are defined over the set C .
Therefore, MSL (C , E {,0}) , where
C T1 ,..., Tn , Tn1 ,..., T2n 1 .
From the Def. 4 it follows, that all mentioned properties of
are also true for , i.e.
1.
T1 T1 T1 ,
2.
T1 T2 T2 T1 ,
3.
T1 T2 T3 T1 T2 T3 ,
Cn0 and Cn1 , therefore
n
n
k 0
k 2
q CE Cnk Cn1 Cn0 Cnk 2n n 1 . ■
Similarly to [13, Th. 2] and Th. 2, we can formulate and
prove the following theorem.
Theorem 4. Set of classes of objects
where T1 , T2 , T3 C . From the definition of
that, T1 TMSL TMSL , where TMSL
it follows
T1 ... Tn .
C is a poset. For this we
T1 , T2 C | T1 T2 T1 T2 T1 and
show that is a relation of partial order under the set C .
Now we need to prove that
should define
Taking into account that we have two types of classes, we need
to define three kinds of relation, i.e.
1.
homogeneous
homogeneous,
2.
homogeneous
inhomogeneous,
3.
inhomogeneous
n
n 0
0.
inhomogeneous.
It was done in Def. 2 and Def. 3. Now let us prove reflexivity,
anti-symmetry and transitivity of these relations.
1.
Reflexivity: T1 T1
idempotency of ;
2.
Anti-symmetry:
T1 T1 T1 follows from
T1 T2 T1 T2 T1 , and
T2 T1 T2 T1 T2 and from commutativity
of , we can conclude that T1 T2 ;
3.
i.e. 1 and class TMSL T1 ... Tn is its least lower bound,
Transitivity:
T2 T1 T1 T2 T2 , T3 T2 T2 T3
i.e.
0.
T3 T1 T2 T3 T1 T2 T3
Proof: According to definition of complete lattice given in
[14], [15], it is a system L ( A, {,,1,0}) , where A
T1 T3 T3 T1 T3 T3 T3 T1 .
is a poset and
Therefore,
L1 : (a b) c a (b c) (associative laws)
MSL C T1 ,...,Tn , Tn 1 ,...,T2n 1 , E {,0}
(a b) c a (b c)
is a meet-semilattice, where class TMSL T1 ... Tn is its
least lower bound, i.e.
, , 1 and 0 satisfy, for all a, b, c A :
L2 : a b b a (commutative laws)
■
0.
a b b a
Using Th. 1 and Th. 3, let us formulate and prove the
following theorem.
L3 : a a a
Theorem 5. For any OODN (O, C , R, E , M ) , where
aa a
C T1 ,..., Tn , Tn1 ,..., T2n 1 , E {,} , and T1 ,..., Tn
are homogeneous classes, which describe types of objects
t1 ,..., t n and there is a type t , such that
t t1 ... t tn t | t t
L4 : a (a b) a
L5 : a 0 a (identity laws)
a 1 a
all possible applications of union and intersection exploiters,
including all possible their superpositions, to classes of objects
from the set C and obtaining classes of objects, using these
exploiters respectively, always generate finite quantity of new
classes of objects, which can be precisely calculated by the
following formula:
q CE 2 n 1 2(n 1) ,
a 1 1
According to the theorem, carrier of the lattice is the set of
classes C , set of exploiters E contains two binary operations
, and two unary operations 1 and 0 , which are defined
over the set C . Therefore, L (C , E {,,1,0}) , where
C T1 ,..., Tn , Tn 1 ,..., T2n 1 , T2n ,..., T2n1 2 ( n 1)
Proof: Proof of the theorem follows from proofs of Th. 1
and Th. 3, i.e.
n
k 0
k 2
q CE 2 Cnk Cn1 Cn0 2 Cnk 2n 1 2(n 1) ,
where
a 0 0
n C .
n
n C .
■
Similarly to Th. 2 and Th. 4, we can formulate and prove
the following important theorem.
Theorem 6. Set of classes of objects
C T1 ,..., Tn , Tn 1 ,..., T2n 1 , T2n ,..., T2n1 2 ( n 1)
(absorption laws)
a ( a b) a
t t1 ... t t n ,
where
(idempotency laws)
of any OODN, extended according to Th. 5, with exploiters ,
create the complete lattice L (C , E {,,1,0}) ,
where class TJSM T1 ... Tn is the greatest upper bound,
Facts that
1.
(C , ) is a poset,
2.
and satisfy the laws L1 L3 ,
3.
TJSL T1 ... Tn is 1 of join-semilattice,
4.
TMSL T1 ... Tn is 0 of meet-semilattice,
were shown in the proves of Th. 2 and Th. 4.
From the [13, Def. 14] and Def. 4 it follows, that
T1 TMSL T1 , T1 TJSL T1 , T1 TMSL TMSL , and
T1 TJSL TJSL , where T1 C , TJSL T1 ... Tn ,
TMSL T1 ... Tn . Therefore, L (C , E {,,1,0})
is a complete lattice, where
C T1 ,..., Tn , Tn 1 ,..., T2n 1 , T2n ,..., T2n1 2 ( n 1) ,
v4 ( p4 (S )) 360) ,
and TJSL T1 ... Tn is its greatest upper bound, i.e. 1 and
p6 ( S ) – verification function, which defines property «all
TMSL T1 ... Tn is its least lower bound, i.e. 0 .
IV.
■
EXPLOITERS-BASED KNOWLEDGE EXTRACTION
Let us consider classes of objects, which describe such
types of convex polygons as square (S ) , rhombus (Rb) ,
parallelogram (P ) , and rectangle (Rt ) . Let us define for
them an OODN
Quadrangle (O, C , R, E , M ) .
For this purpose, we need to define set of classes of objects
C {S , Rb, P, Rt} and set of exploiters E {,} . Sets
sides are equal», i.e. vf6 ( S ) : p6 ( S ) {0,1} , where
p6 ( S ) (v1 ( p3 ( S )) v2 ( p3 ( S )) v3 ( p3 ( S ))
v4 ( p3 ( S ))) ,
p7 ( S ) – verification function, which defines property «all
angles are equal to 90 », i.e. vf7 ( S ) : p7 ( S ) {0,1} ,
o
where
p7 ( S ) (v1 ( p4 ( S )) v4 ( p4 ( S )) v3 ( p4 ( S ))
v4 ( p4 (S )) 90) ,
O and R will be undefined, because of the lack of
information. In addition, we do not define the set of modifiers
M , because it is not necessary within consideration of
exploiters-based KE. Suppose classes from the set C have
following structures
f1 ( S ) – method for perimeter computing, and
method for area computing;
f 2 (S ) –
Rb ( p1 ( Rb) (4, sides),
p2 ( Rb) (4, angles),
S ( p1 (S ) (4, sides),
p3 ( Rb) ((v1 ( p3 ( Rb)), cm), (v2 ( p3 ( Rb)), cm),
p2 (S ) (4, angles)
p3 ( S ) ((v1 ( p3 ( S )), cm), (v2 ( p3 ( S )), cm),
(v3 ( p3 ( Rb)), cm), (v4 ( p3 ( Rb)), cm)),
p4 ( Rb) ((v1 ( p4 ( Rb)),o ), (v2 ( p4 ( Rb)),o ),
(v3 ( p3 ( S )), cm), (v4 ( p3 ( S )), cm)),
p4 ( S ) (90o ,90o ,90o ,90o ),
(v3 ( p4 ( Rb)),o ), (v4 ( p4 ( Rb)),o )),
p5 ( S ) vf5 ( S ) 1,
p5 ( Rb) vf5 ( Rb) 1,
p6 ( S ) vf6 ( S ) 1,
p6 ( Rb) vf6 ( Rb) 1,
p7 ( S ) vf7 ( S ) 1,
f1 ( Rb) (v1 ( p3 ( Rb)) 4, cm),
f1 ( S ) (v1 ( p3 ( S )) 4, cm),
f 2 ( Rb) v1 ( p3 ( Rb))2 sin(v1 ( p4 ( Rb))),cm2 ,
f 2 (S ) v1 ( p3 (S ))2 , cm2 ,
p1 ( Rb) – quantity of sides, p2 ( Rb) – quantity of
angles, p3 ( Rb) – sizes of sides, p4 ( Rb) – measures of
internal angles, p5 ( Rb) – verification function, which defines
where
p1 (S ) – quantity of sides, p2 ( S ) – quantity of
angles, p3 ( S ) – sizes of sides, p4 ( S ) – measures of internal
angles, p5 ( S ) – verification function, which defines property
where
o
property «sum of internal angles is equal to 360 », i.e.
vf5 ( Rb) : p5 ( Rb) {0,1} ,
o
«sum of internal angles is equal to 360 », i.e.
vf5 ( S ) : p5 ( S ) {0,1} ,
where
p5 ( S ) (v1 ( p4 ( S )) v2 ( p4 ( S )) v3 ( p4 ( S ))
where
p5 ( Rb) (v1 ( p4 ( Rb)) v2 ( p4 ( Rb)) v3 ( p4 ( Rb))
v4 ( p4 ( Rb)) 360) ,
v4 ( p4 ( P))) ,
p6 ( Rb) – verification function, which defines property «all
sides are equal», i.e. vf6 ( Rb ) : p6 ( Rb ) {0,1} , where
p7 ( P) – verification function, which defines property
«opposite sides are equal», i.e. vf7 ( P) : p7 ( P) {0,1} ,
p6 ( Rb) (v1 ( p3 ( Rb)) v2 ( p3 ( Rb)) v3 ( p3 ( Rb))
where
v4 ( p3 ( Rb))) ,
p7 ( P) (v1 ( p3 ( P)) v3 ( p3 ( P))) (v2 ( p3 ( P))
f1 ( Rb) – method for perimeter computing, and
method for area computing;
f 2 ( Rb)
–
P ( p1 ( P) (4, sides),
v4 ( p3 ( P))) ,
f1 ( P) – method for perimeter computing, and f 2 ( P) –
method for area computing;
p2 ( P) (4, angles),
p3 ( P) ((v1 ( p3 ( P)), cm), (v2 ( p3 ( P)), cm),
Rt ( p1 ( Rt ) (4, sides),
p2 ( Rt ) (4, angles),
(v3 ( p3 ( R)), cm), (v4 ( p3 ( R)), cm)),
p3 ( Rt ) ((v1 ( p3 ( Rt )), cm), (v2 ( p3 ( Rt )), cm),
p4 ( P) ((v1 ( p4 ( P)), ), (v2 ( p4 ( P)), ),
o
o
o
(v3 ( p3 ( Rt )), cm), (v3 ( p3 ( Rt )), cm)),
o
(v3 ( p4 ( P)), ), (v4 ( p4 ( P)), )),
p4 ( Rt ) ((90,o ), (90,o ), (90,o ), (90,o )),
p5 ( P) vf5 ( P) 1,
p5 ( Rt ) vf5 ( Rt ) 1,
p6 ( P) vf6 ( P) 1,
p6 ( Rt ) vf6 ( Rt ) 1,
p7 ( P) vf7 ( P) 1,
f1 ( Rt ) (2 (v1 ( p3 ( Rt )) v2 ( p3 ( Rt ))),cm)),
f1 ( P) (2 (v1 ( p3 ( P)) v2 ( p3 ( P))),cm),
f 2 ( P) (v1 ( p3 ( P)) v2 ( p3 ( P))
sin(v1 ( p4 ( P))),cm ,
2
p1 ( P) – quantity of sides, p2 ( P) – quantity of
angles, p3 ( P ) – sizes of sides, p4 ( P) – measures of internal
angles, p5 ( P ) – verification function, which defines property
where
f 2 ( Rt ) v1 ( p3 ( Rt )) v2 ( p3 ( Rt )), cm2 ,
p1 ( Rt ) – quantity of sides, p2 ( Rt ) – quantity of
angles, p3 ( Rt ) – sizes of sides, p4 ( Rt ) – measures of
internal angles, p5 ( Rt ) – verification function, which defines
where
o
property «sum of internal angles is equal to 360 », i.e.
vf5 ( Rt ) : p5 ( Rt ) {0,1} ,
o
«sum of internal angles is equal to 360 », i.e.
vf5 ( P) : p5 ( P) {0,1} ,
where
p5 ( Rt ) (v1 ( p4 ( Rt )) v2 ( p4 ( Rt )) v3 ( p4 ( Rt ))
where
p5 ( P) (v1 ( p4 ( P)) v2 ( p4 ( P)) v3 ( p4 ( P))
v4 ( p4 ( P)) 360) ,
p6 ( P) – verification function, which defines property
«opposite sides are parallel», i.e. vf6 ( P) : p6 ( P) {0,1} ,
where
p6 ( P) (v1 ( p4 ( P)) v3 ( p4 ( P))) (v2 ( p4 ( P))
v4 ( p4 ( Rt )) 360) ,
p6 ( Rt ) – verification function, which defines property
«opposite sides are equal», i.e. vf6 ( Rt ) : p6 ( Rt ) {0,1} ,
where
p6 ( Rt ) (v1 ( p3 ( Rt )) v3 ( p3 ( Rt ))) (v2 ( p3 ( Rt ))
v4 ( p3 ( Rt ))) ,
f1 ( Rt ) – method for perimeter computing, and f 2 ( Rt ) –
method for area computing.
We have defined OODN for early mentioned types of
convex polygons. It is clear, that all elements of the set C
represent basic knowledge. Let us apply union and intersection
exploiters to them and obtain all possible new classes of
objects. According to [13, Def. 14],
S Rb SRb CoreSRb , pr1 (S ), pr2 ( Rb) ,
where
CoreSRb p1 SRb , p2 SRb , p3 SRb ,
p4 SRb , f1 SRb ,
p1 SRb – quantity of sides, p2 SRb – quantity
of angles, p3 SRb – verification function, which defines
Structure of the
equalities
p1 (S ) p1 ( Rb) , p2 (S ) p2 ( Rb) , p5 ( S ) p5 ( Rb) ,
p6 ( S ) p6 ( Rb ) , f1 (S ) f1 ( Rb) .
Indeed, according to [13, Def 4],
o
vf3 SRb : p3 SRb {0,1} ,
where
p3 SRb (v1 ( p4 (ti )) v2 ( p4 (ti )) v3 ( p4 (ti ))
v4 ( p4 (ti )) 360) , i {S , Rb} ,
p4 SRb – verification function, which defines property
«all sides are equal», i.e. vf4 SRb : p4 SRb {0,1} ,
(v1 ( p1 (S )),u1 ( p1 (S ))) (v1 ( p1 ( Rb)),u1 ( p1 ( Rb)))
(4, sides) ,
(v1 ( p2 (S )),u1 ( p2 (S ))) (v1 ( p2 ( Rb)),u1 ( p2 ( Rb)))
( 4, angles) .
Form the [13, Def. 5] it follows that p5 ( S ) p5 ( Rb) and
p6 ( S ) p6 ( Rb ) , i.e.
vf
S
5
v4 ( p3 (ti ))) , i {S , Rb} ,
f1 SRb – method for perimeter computing, which is
defined as follows f1 SRb (4 v1 ( p3 (ti )), cm) , where
i {S , Rb} .
Projections
structure
pr1 ( S ) and pr2 ( Rb) have the following
pr1 ( S ) ( p5 ( S ), p6 ( S ), p7 ( S ), f 2 ( S )) ,
pr2 ( Rb) ( p5 ( Rb), p6 ( Rb), f 2 ( Rb)) ,
where p5 ( S ) – sizes of sides, p6 ( S ) – measures of internal
angles, p7 ( S ) – verification function, which defines property
o
«all angles are equal to 90 »,
f 2 (S ) – method for area
computing, p5 ( Rb) – sizes of sides, p6 ( Rb) – measures of
internal angles, f 2 ( Rb) – method for area computing.
(S ) vf5Rb (S ) vf5S ( Rb) vf5Rb ( Rb) ,
that can be computed in the following way
vf5S (S ) v1 ( p4 (S )) v2 ( p4 (S )) v3 ( p4 (S ))
v4 ( p4 (S )) 360 ,
vf5Rb (S ) v1 ( p4 (S )) v2 ( p4 (S )) v3 ( p4 (S ))
v4 ( p4 (S )) 360 ,
where
p4 SRb (v1 ( p3 (ti )) v2 ( p3 (ti )) v3 ( p3 (ti ))
p1 (S ) p1 ( Rb) and
p2 (S ) p2 ( Rb) , i.e.
where
property «sum of internal angles is equal to 360 », i.e.
CoreSRb follows from the following
vf5S ( Rb) v1 ( p4 ( Rb)) v2 ( p4 ( Rb)) v3 ( p4 ( Rb))
v4 ( p4 ( Rb)) 360,
vf5Rb ( Rb) v1 ( p4 ( Rb)) v2 ( p4 ( Rb)) v3 ( p4 ( Rb))
vf
S
6
v4 ( p4 ( Rb)) 360 ;
(S ) vf6Rb (S ) vf6S ( Rb) vf6Rb ( Rb)
vf6S (S ) (v1 ( p3 (S )) v2 ( p3 (S )) v3 ( p3 (S ))
v4 ( p3 ( S ))) ,
vf6Rb (S ) (v1 ( p3 (S )) v2 ( p3 (S )) v3 ( p3 (S ))
v4 ( p3 ( S ))) ,
vf6S ( Rb) (v1 ( p3 ( Rb)) v2 ( p3 ( Rb)) v3 ( p3 ( Rb))
v4 ( p3 ( Rb))) ,
vf6Rb ( Rb) (v1 ( p3 ( Rb)) v2 ( p3 ( Rb)) v3 ( p3 ( Rb))
v4 ( p3 ( Rb))) .
As the result, in both cases we have (1 1) (1 1) , i.e.
1 1 1 .
From the [13, Def. 7] it follows that
f
S
1
f1 (S ) f1 ( Rb) , i.e.
( S ) f1Rb ( S ) f1S ( Rb) f1Rb ( Rb) ,
that can be calculated in the following way
f1S (S ) (v1 ( p3 (S )) 4, cm) ,
f1Rb (S ) (v1 ( p3 (S )) 4, cm) ,
f1S ( Rb) (v1 ( p3 ( Rb)) 4, cm) ,
Rb
1
f
Fig. 1. Complete lattice created by the set of classes and set of exploiters.
( Rb) (v1 ( p3 ( Rb)) 4, cm) ,
C S , Rb, P, Rt , SRb ,..., PRt , SRbP ,..., RbPRt ,
as the result we have
(v1 ( p3 (S )) 4, cm) (v1 ( p3 (S )) 4, cm)
SRbPRt , SRb ,..., PRt ,.., SRbP ,..., RbPRt ,
SRbPRt .
(v1 ( p3 ( Rb)) 4, cm) (v1 ( p3 ( Rb)) 4, cm) ,
According to Th. 6, the set C together with exploiters and
create the complete lattice L (C , E {,,1,0}) ,
i.e. 1 1 1 .
According to [13, Def. 14], the class of objects SRb is
the result of application of union exploiter to classes of objects
S and Rb . From the Def. 4, we can conclude, that the result
of application of intersection exploiter to these classes is equal
to the core of their union, i.e.
S Rb SRb CoreSRb .
In the result of all possible applications of union and
intersection exploiters we obtained such 6 classes, that each
class describes 2 different types of objects SRb , SP ,
SRt , RbP , RbRt , PRt such 4 classes, that each
class describes 3 different types of objects SRbP ,
SRbRt , SPRt , RbPRt and 1 class, that describes 4
different types of objects SRbPRt . In addition, we obtained
such 6 classes, that each class describes intersection of 2
different types of objects SRb , SP , SRt , RbP ,
RbRt , PRt , such 4 classes, that each class describes
intersection of 3 different types of objects SRbP , SRbRt ,
SPRt , RbPRt , and 1 class, that describes intersection f
4 different types of objects SRbPRt .
Using exploiters and , we have extended the set
by adding 22 new classes of objects, i.e.
C
SRbPRt is its greatest upper bound, i.e. 1 and
SRbPRt is its least lower bound, i.e. 0 . This lattice can be
where
graphically represented as it is shown on Fig. 1.
In addition, we define the set of relations R , by adding
new relations, namely 56 relations for classes S,..., Rt ,
96
32
SRb ,..., PRt and SRb ,..., PRt , 8 for
classes SRbP ,..., RbPRt and SRbP ,..., RbPRt .
for classes
Analyzing Fig. 1, we can see that obtained lattice defines
hierarchy of classes with determined subsumption relation .
It allows performing of subsumption reasoning for information
classifying and retrieving. Moreover, obtained hierarchy is
protected from ambiguity problem, because all classes, except
basic ones, are inhomogeneous.
Join-semilattice of the lattice L contains inhomogeneous
classes of objects, which define all possible sets of objects of
different types, which can be obtained from the basic classes of
objects S , Rb , P and Rt . Meet-semilattice of the lattice
L contains inhomogeneous classes of objects, which define
common subtypes for basic classes.
The greatest upper bound SRbPRt of the lattice L
gives an opportunity to represent and to store the knowledge in
the database in more efficient way by storing only one class
SRbPRt instead of four basic classes of objects. Moreover,
such storing requires less memory resources then storing of S ,
Rb , P and Rt , because instead of storing of 26 properties
and 8 methods, it is possible to store only 17 properties and
5 methods.
We can conclude that during KE using universal exploiters
we have obtained 22 new classes of objects, 96 new
relations among them, defined hierarchy of classes with
determined subsumption relation . Using obtained
knowledge it is possible to restore basic knowledge in database
more efficiently and perform subsumption reasoning within the
constructed hierarchy of classes.
V.
CONCLUSIONS AND OUTLOOK
Invention of KE techniques is very crucial for future
development of KRFs and area of KR in general. In this paper
the main attention was paid to consideration and extension of
KE method within such object-oriented KRF as object-oriented
dynamic networks. The main idea of proposed approach is
usage of universal exploiters, which allow generation of new
classes of objects and relations among them.
The main achievement of the paper is proof of useful
features of union and intersection exploiters, which allow
extending set of basic classes and create complete lattice.
Proposed approach has the following features:
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
ability to calculate before the generation:
o
quantity of new classes, which can be
generated, using proposed approach,
[12]
o
quantity of different types, which each of
obtained classes describes;
[13]
extension of the sets of basic classes and relations by
adding new classes of objects and relations among
them;
construction of defined hierarchy of classes with
determined subsumption relation , which allows
performing of subsumption reasoning for information
classifying and retrieving;
more efficient restoring of basic knowledge within the
database;
avoidance of inheritance problems, in particular
ambiguity problem in the case of multiple inheritance.
However, despite all noted advantages, proposed approach
requires further research, at least in the following directions:
using of useful properties of complete lattices;
adaptation to different kinds of knowledge sources;
extension to the case of fuzzy knowledge;
adaptation and usage in other known object-oriented
knowledge representation formalisms.
[14]
[15]
D. S. Touretzky, The mathematics of inheritance systems. Los Altos,
CA, USA: Morgan Kaufmann Publishers, 1986.
R. Al-Asady, Inheritance theory: an artificial intelligence approach.
Norwood, NJ, USA: Ablex Publishing Corporation, 1995.
D.
Terletskyi,
“Inheritance
in
object-oriented
knowledge
representation,” in Inform. and Software Techn., Commun. in Comput.
and Inform. Sci., vol. 538, G. Dregvaite and R. Damasevicius Eds., AG,
Switzerland: Springer, 2015, pp. 293–305.
S. Hellmann, J. Unbehauen et al., “Report on knowledge extraction from
structured sources,” Technical Report LOD2 D3.1.1, 2011
J. Unbehauen, S. Hellmann, S. Auer and C. Stadler, “Knowledge
extraction from structured sources,” in Search Computing: Broadening
Web Search, Lecture Notes in Comput. Sci., vol. 7538, S. Ceri and
M. Brambilla Eds. Berlin, Germany: Springer, 2012, pp. 34–52.
N. Takhirov, “Extracting knowledge for cultural heritage knowledge
base population,” Ph.D. dissertation, Dept. Comput. and Inform. Sci.,
Norwegian Univ. of Sci. and Technol., Trondheim, Norway, 2013.
J. Polpinij, “Ontology-based knowledge discovery from unstructured
and semi-structured text,” Ph.D. dissertation, School of Comput. Sci.
and Software Eng., Univ. of Wollongong, Wollongong, NSW, Australia,
2014.
Z. Ma, F. Zhang, L. Yan and J. Cheng, Fuzzy knowledge management
for the semantic web. Berlin, Germany: Springer, 2014.
L. Yan, Z. Ma and F. Zhang, Fuzzy XML data management. Berlin,
Germany: Springer, 2014.
D. Terletskyi and A. Provotar, “Object-oriented dynamic networks,” in
Computational Models for Bus. and Eng. Domains, Int. Book Series
Inform. Sci. & Computing, vol. 30, G. Setlak and K. Markov Eds.
Rzeszow, Poland: ITHEA, 2014, pp. 123–136.
D. A. Terletskyi and A. I. Provotar, “Fuzzy object-oriented dynamic
networks. I,” Cybern. and Syst. Anal., vol. 51, no. 1, pp. 34–40, Jan.
2015.
D. A. Terletskyi and A. I. Provotar, “Fuzzy object-oriented dynamic
networks. II,” Cybern. and Syst. Anal., vol. 52, no. 1, pp. 38–45, Jan.
2016.
D. Terletskyi, “Exploiters-based knowledge extraction in objectoriented knowledge representation,” in Proc. 24th Int. Workshop
Concurrency, Specification & Programming, Rzeszow, 2015, vol. 2, pp.
211–221.
G. Birkhoff, Lattice theory. 3rd revised ed. New York, NY, USA:
American Mathematical Society Colloquium Publications, 1967.
B. A. Davey and H. A. Priestley, Introduction to lattices and order. 2nd
ed. New York, NY, USA: Cambridge University Press, 2002.
| 2cs.AI
|
UAI 2009
MOSTAFAVI & MORRIS
419
Using the Gene Ontology Hierarchy when Predicting Gene Function
Sara Mostafavi
Quaid Morris
Department of Computer Science, Banting and Best Department of Medical Research and
University of Toronto
Departments of Computer Science and Molecular Genetics,
University of Toronto
Abstract
The problem of multilabel classification when
the labels are related through a hierarchical
categorization scheme occurs in many application domains such as computational biology. For example, this problem arises naturally when trying to automatically assign
gene function using a controlled vocabularies like Gene Ontology. However, most existing approaches for predicting gene functions
solve independent classification problems to
predict genes that are involved in a given
function category, independently of the rest.
Here, we propose two simple methods for incorporating information about the hierarchical nature of the categorization scheme. In
the first method, we use information about
a gene’s previous annotation to set an initial
prior on its label. In a second approach, we
extend a graph-based semi-supervised learning algorithm for predicting gene function in
a hierarchy. We show that we can efficiently
solve this problem by solving a linear system
of equations. We compare these approaches
with a previous label reconciliation-based approach. Results show that using the hierarchy information directly, compared to using
reconciliation methods, improves gene function prediction.
1
Introduction
We are interested in the problem of multilabel classification when the labels are related through a hierarchical categorization scheme and the input data is represented as a similarity metric between objects. This
problem arises naturally when trying to automatically
assign gene function, using controlled vocabularies like
Gene Ontology (GO) (Ashburner et al., 2000), En-
zyme Commission (EC) number (Bairoch, 2000), and
Structural Classification of Protein (SCOP) categories
(Murzin et al., 1995). The hierarchical multilabel classification problem, as we have posed it, arises in other
application domains including automatic image segmentation and automatic web-page annotation.
In the domain of gene and protein function prediction, the input data is most naturally represented using a network whose nodes represent genes (or proteins) and whose edges are positively weighted according to the biological evidence for shared function of
the connected genes. Such networks, which are called
functional association or function linkage networks,
can be derived from a variety of genomics and proteomics data sources, including gene expression and
genetic interaction data, by using an appropriate similarity metric. For example, Pearson correlation coefficient is often used to measure similarity between gene
expression profiles.
Although genes have multiple functions, existing approaches for predicting gene function typically solve a
binary classification problem to identify positive genes
for each function category independently (Pavlidis
et al., 2002; von Mering et al., 2003; Lanckriet et al.,
2004; Tsuda et al., 2005; Myers et al., 2005; Mostafavi
et al., 2008). However, hierarchical gene classification
schemes, such as GO, organize gene function categories
as a directed acyclic graph (DAG) in which categories
describing broader functions (e.g. eye development)
are ancestors of those describing more specific functions (e.g. eye photoreceptor cell differentiation). Annotations, i.e., assignments of genes to a given category, satisfy the “true path” rule: genes annotated to
a given category are also assigned to all of its ancestors; for example, an annotation in the category photoreceptor cell differentiation implies annotation in eye
development.
When making an annotation, curators place genes in
the most specific category supported by the available
data. Often genes are annotated in internal nodes of
420
MOSTAFAVI & MORRIS
the DAG because there is insufficient evidence to annotate genes in the most specific, i.e., leaf, categories.
For example, a mouse gene can be annotated as being
involved in development if mice with defective copies of
that gene die as embryos. Further investigations may
determine whether the gene functions in eye, heart, or
brain development, warranting a more specific annotation. These internal node annotations can provide
helpful hints when classifying genes in descendent categories, so long as the classification algorithm incorporates prior knowledge about the hierarchy.
Here, we introduce two new classification methods that
leverage DAG-based categorization hierarchies. Both
of our algorithms extend the Gaussian random fields
(GRF) algorithm (Zhou et al., 2004; Zhu et al., 2003).
Our interest in this algorithm stems from its success
at predicting gene function compared with other binary and hierarchy-based classification schemes (PenaCastillo et al., 2008; Mostafavi et al., 2008; Tsuda
et al., 2005). Intuitively, the GRF algorithm takes
as input a similarity network (here a functional linkage network) and a set of real-valued label biases and
assigns real-valued discriminant scores to each node;
these scores are assigned so that the linked nodes have
similar discriminant scores and the discriminant score
of each node is not too different from its initial label bias. Our first method, which we call Hierarchical label propagation (HLProp), replicates the similarity network for each category and then links the
nodes representing the same gene in parent and child
categories, thus ensuring that the discriminant scores
of a gene in related function categories also remain
close. By applying the GRF algorithm to this new
(much larger though sparsely-connected) network, we
can perform multilabel classification efficiently by solving a linear system of equations. We also describe a
second method, Hierarchical label bias (HLBias), that
uses the GO hierarchy to set label biases of genes with
annotations in internal category nodes. This second
approach builds on the previous work of (Eisner et al.,
2005) which used the structure of the GO hierarchy
to define positive and negative examples for a given
category of interest.
In Section 3, we briefly review the GeneMANIA version of the GRF algorithm (Mostafavi et al., 2008);
show the relationship between GRF label propagation
and Gaussian inference; and then use that relationship
to justify our two extensions: HLBias and HLProp. In
Sections 4 and 5, we compare these two methods, and
two simplifications of HLProp, a reconciliation method
called Isotonic Regression (Obozinski et al., 2008), and
unaugmented binary classification using GeneMANIA
GRF (Mostafavi et al., 2008). We evaluate performance in two settings, test and novel settings, using
UAI 2009
the data from the MouseFunc challenge (Pena-Castillo
et al., 2008). The test setting evaluates each classifier
in a cross-validation framework in which the GO annotations of test genes are completely hidden. The
novel setting evaluates the performance in the task of
predicting new annotations given the state of the GO
annotation database from the previous year. Many annotations in the updated GO database are refinements
of pre-existing internal node annotations.
2
Previous Work
To date, most classification algorithms that make use
of the GO hierarchy have built on top of binary classification schemes; standard structured output classification algorithms (Taskar et al., 2003) are difficult to
employ here due to the size the classification problem
and the non-trivial tree-width of the GO hierarchy.
Augmented binary classification schemes include cascaded classification (Kiritchenko et al., 2004) which
trains one binary classifier for each category to predict whether annotation in the child category is warranted given annotation in the parent category (or categories). Annotation predictions are made by querying classifiers in a cascade from the root down. An
alternative approach is to independently train binary
classifiers for each category and then to reconcile their
predictions so that the true path rule is enforced. For
example, (Barutcuoglu et al., 2008) reconciled predictions of binary SVMs for 100 GO categories (a subset
of about 2,000 well-annotated GO categories) by using
a Bayesian network to model the GO hierarchy over
the considered categories, (Obozinski et al., 2008) extended this approach to the entire GO hierarchy using
approximate inference.
Obozinski and colleagues (Obozinski et al., 2008)
compared ten reconciliation methods including reconciliation with a Bayesian networks (as done in
(Barutcuoglu et al., 2008)), Isotonic Regression, and a
cascade of classifiers approach (Cascade Logistics Regression) on the MouseFunc benchmark. Surprisingly,
many reconciliation methods did not significantly improve prediction accuracy, however, the performance
of Isotonic Regression was better than that of Cascade
Logistic Regression and other reconcilation methods.
As such, in this report, we restrict our comparisons to
Isotonic Regression. Other hierarchy-aware gene function prediction schemes that have only been applied
to tree-structured hierarchies, e.g. (Shahbaba & Neal,
2006), will not be considered here.
UAI 2009
3
MOSTAFAVI & MORRIS
Algorithm
We assume that we are given as input a weighted, undirected network represented as a positive, symmetric
affinity matrix W = W T , wij ≥ 0, where wij indicates the strength of the evidence of co-functionality
between genes i and j; a classification hierarchy over d
categories denoted by Hd×d , where hmc = 1 indicates
that category c is a child of category of m; and a label
bias matrix Y = [y~1 , ..., y~d ], where ~yc ∈ {+1, k, −1}n×1
consists of the label biases of all genes in the category
c. As we will show later, the parameter k is set to a
scalar between -1 and +1 and reflects our prior bias
on the mean of the labels of unlabeled genes. We consider a transductive setting, i.e., we assume that all
input vectors are available during training; this is a
natural assumption since the known gene complement
of genomes is relatively stable.
Below, we first describe how to predict gene function
using the GRF algorithm; we then show how to use the
GO hierarchy to derive more informative label biases.
Next, we show how to extend GRF when a classification hierarchy is available. Finally, we describe other
approaches for making gene function predictions with
a classification hierarchy.
3.1
Predicting a single gene function
To predict gene function, we use the Gaussian random
fields algorithm (Zhu et al., 2003; Zhou et al., 2004) to
assign a discriminant score fi ∈ [−1, 1] to each node
(protein) i in the network. These discriminant scores
can then be thresholded to classify the genes. Below,
we write the Gaussian random fields algorithm in the
following general form:
f~∗
=
arg min
f~
n
X
σi (yi − fi )2 +
i=1
n
X
wij (fi − fj )2
i,j=1
=
arg min (f~ − ~y ) Σ(f~ − ~y ) + f~Lf~
=
(Σ + L)−1 Σ~y
T
f~
(1)
where ~σ = [σ1 , ..., σn ]T are model parameters, Σ is
a diagonal matrix with Σii = σi , L = D − W is
the graph
P Laplacian and D is a diagonal matrix with
Dii = j wij . The above objective ensures that the
discriminant scores remain close to their initial labels
(first term in (1)) and that the discriminant scores of
linked genes (as indicated by wij > 0) are similar to
each other (second term in (1)).
To ensure that Σ + L is invertible, we can set σi > 0
and thus ensure that Σ + L is diagonally dominant.
However, to solve for f~, we only need to solve a linear
421
system of equations (Σ + L)f~ = Σ−1 ~y , which we can
do with various existing fast iterative solvers (Nocedal
& Wright, 2006). Here, we use the conjugate gradient
(CG) algorithm which is well-suited to this problem
because our coefficient matrix is very sparse and the
CG iterations only require us to take matrix-vector
products with the coefficient matrix.
The solution to (1) can also be interpreted as the maximum a posteriori (MAP) estimate of f~, where the observations ~y ∼ N (f~, Σ−1 ) with a prior on f~, namely
f~ ∼ N (~0, L−1 ) where N (~
µ, K) is the normal distribution with mean µ
~ and covariance matrix K (Shental
et al., 2008).
This relationship suggests that the label bias yi can
be viewed as a noisy estimate of a soft label fi with
the regularization parameter σi as the precision of the
estimate and the weights wij as inverse prior covariance between fi and fj . This interpretation suggests
that a node’s label bias should reflect our prior beliefs
about its label. We have previously shown (Mostafavi
et al., 2008) in unbalanced problems, we can achieve a
large gain in classification performance by setting the
+
−
label bias k of unlabeled nodes to be the k = nn+ −n
+n− ,
the mean of the labels of the labelled nodes. In the
following section, we describe how we use the GO hierarchy to set the label bias for nodes which we have
previously labelled as negative.
3.2
Hierarchical label bias
In Hierarchical label bias (HLBias), we use the GO
hierarchy directly to set the initial label bias of nonpositive genes. HLBias builds on the previous work of
(King et al., 2003) and (Eisner et al., 2005). In particular, King and colleagues (King et al., 2003) used
a gene’s annotations as a feature vector for predicting
additional annotations for the given gene. Eisner and
colleagues (Eisner et al., 2005), used the structure of
the GO hierarchy to define appropriate negative examples for predicting a given gene function: they used as
negatives all genes with initial annotation (i.e., before
calculating the transitive closure using the true path
rule) in neither descendant nor ancestral categories.
In our approach, we use a gene’s previous annotations
to estimate our prior bias that it will be annotated to
a given category of interest. To do so, when predicting category c, we first use as negatives all genes that
are annotated in any sibling category of c. We assign
this negative label because genes are rarely annotated
in more than one child of the same parent category.
For other genes i with an annotation in an ancestral
n+
+
category a of c, we set yic = 2 × nac
+ − 1 where na
a
is the number of positive examples in category a and
422
MOSTAFAVI & MORRIS
n+
ac is the number of positive examples in category a
that were also annotated in category c; this initial label bias is proportional to the probability of a gene
being annotated to category c given its annotation in
category a. For a gene i with multiple annotations, we
set yic to its mean value. Having set these label biases,
we then solve for discriminant values independently.
3.3
Given a matrix of label biases Yn×d and a hierarchical
classification scheme represented by Hd×d , we solve for
discriminant values for all d classes simultaneously. To
do so, we solve the following problem:
F ∗ = arg min
n X
d
X
fic
d X
n
X
c
σi (yic − fic )2 +
(2)
c
i
wij (fic − fjc )2 + λ
i,j
d
n X
X
i
hmc (fim − fic )2
c,m
where F = [f~1 , ..., f~d ] and hmc denotes the relationship between category m and c, and λ, σi ’s are the
regularization constants. Without the third term (by
setting λ = 0), equation (2) corresponds to solving d
independent binary classification problems. The third
term encourages the discriminant values of a gene in
two related function categories to be similar to each
other (see Figure 1 for an example).
In this work, we use the GO hierarchy to define H.
GO is a DAG (directed acyclic graph), however, as we
will discuss later, to ensure that problem (2) is convex,
we treat GO as an undirected graph. In particular,
hcm ∈ {0, +1} represent the parent child relationships
in GO: we set hcm = hmc = 1 if m is a parent of c in
the GO hierarchy and 0 otherwise. In addition, in our
experiments we set λ and each σi to a fixed value of
1 and so we drop these constants from our subsequent
equations.
3.3.1
Optimization
∗
We can solve for F by solving the following problem:
F∗
=
arg min trace(F T F − 2F T Y )
+
trace(F T LF ) + trace(F GF T )
yi,1
F
(3)
where G is the graphP
Laplacian of H, G = V −H, V =
d
diag(vcc ), and vcc = m hmc .
Differentiating equation (3) with respect to F , we get
the matrix equation (I + L)F + F G = Y . Equivalently, we can find F by solving a large sparse linear
system: A(vec(F )) = vec(Y ), where vec(Y ) is an operator that stacks the columns of Y atop of each other,
Gene Ontology
(GO) Categories
Sensory organ
development: y1
Sensory organ
development: y1
Eye
development: y2
fj,4
h2,4
Lens
development: y3
Hierarchical label propagation
UAI 2009
Retinal
development: y4
Eye
development: y2
Lens
development: y3
Retinal
development: y4
Figure 1: A graphical example of our model. On the
left figure, there are four identical networks over four
genes (nodes); the association between different genes
is shown in black edges. The color of the smaller nodes
attached to each gene represents the initial label of
the gene. If we wish to predict which genes are involved in eye development, we need to consider other
related categories (as shown on the right). In our modification, we introduce an edge between the same gene
(blue edges) in the different networks (in this figure, we
have only shown blue edges for one gene); these edges
will encourage the discriminant value of the same gene
(depicted as the color of the bigger nodes) in related
categories to be similar.
A(n×d)×(n×d) = (Id×d ⊗ (I + L) + G ⊗ In×n ), and ⊗
denotes the Kronecker matrix product. As an example, the matrix A that corresponds to the example in
Figure 1 can be represented as:
(2I + L)
−I
0
0
−I
(3I + L)
−I
−I
A=
0
−I
(2I + L)
0
0
−I
0
(2I + L)
In general, A can be represented as a block matrix with
diagonal blocks Aii = (I + L + vii I) and non-diagonal
blocks −hij I.
When H is symmetric, then A is also symmetric. Furthermore, since A is diagonally dominant with positive
diagonals, A is symmetric positive definite (SPD) and
thus invertible. However, with large d (number of gene
function categories) and n (number of genes), constructing A may be infeasible. Instead, we can solve
for the f~c ’s iteratively: given all f~c ’s for c 6= m, we can
solve for f~m by solvingPthe system of linear equations:
d
(I + L + vmm I)f~m = c hcm f~m + ~ym . In our setting,
problem (3) is convex and we can calculate F ∗ by iteratively updating f~m ’s; we have empirically observed
∗
that we need 10 or fewer iterations to solve each f~m
when there are approximately 50 GO categories that
are related to each other.
UAI 2009
3.4
MOSTAFAVI & MORRIS
Other approaches for using the GO
hierarchy
Here, we compare gene function prediction using
HLBias, HLProp, two heuristics approaches based on
HLProp (Down- and Up- propagation), and Isotonic
regression, a method for reconciling the predictions of
independent classifiers to satisfy the true-path rule.
Down- or Up-propagation. Both of our heuristics are
inspired by the observation that if the classification hierarchy were a polytree, then we would be able to solve
for F* using a belief propagation-like message passing
algorithm which passes a single message in each direction along parent and child category link and these
messages are used to calculate the label biases for the
message recipient. In Up-progation, these messages
are only passed from child categories to parent categories and in Down-propagation, messages are only
passed from parents to their children. In particular, in
Down-propagation, we first calculate f~r∗ , the vector of
discriminant scores for genes in the root category using
equation (1), then the root passes f~r∗ as a message to
each of its child categories c who set their label biases
to be ~yc + f~r∗ , where ~yc is the vector of initial label
biases for category c. We then apply this procedure
recursively from parent to child, summing together all
parent messages before calculating label biases whenever a node has multiple parents. Up-propagation is
similar though messages are passed from child nodes
to parent nodes.
Isotonic Regression (IR). Given a set of independent predictions for a gene i in d categories ~xi =
[x1 , ..., xd ], xc ∈ <, IR (Barlow et al., 1972) solves the
following problem:
arg min
zc
d
X
(xc − zc )2 qc
c
subject to zm ≥ zc ∀(m, c) ∈ H
where qc ’s are the parameters. We apply IR to the
discriminant values obtained using Gaussian random
fields algorithm to each function category independently. In our experiment, as done in (Obozinski et al.,
2008), we set qc = 1 for c = 1, ..., d and solve IR heuristically, using the generalized PAV algorithm (GPAV)
(Burdakov et al., 2006).
4
Methods
Regularization parameters In our experiments, we set
all of the precisions σi = 1, i ∈ {1, ..., n}, i.e. Σ = I
and we set the regularization constant in equation (2)
to λ = 1.
423
Benchmark data We use the mouse benchmark data
of (Pena-Castillo et al., 2008) which consists of ten
genomics and proteomics datasets. We first represent
each data source as a similarity network using Pearson
correlation coefficient (r ) and sparsify each network by
setting to zero any interaction that is not among the
top 50 highest r values for either gene. To ensure that
all interactions are positive (i.e. wij ≥ 0), we only
consider positive interactions (r > 0) and set to zero
all the negative interactions. We normalize each network: W̃ = D−1/2 W D−1/2 . We then combine all the
networks, by simply adding them together to obtain a
single functional linkage network and re-normalize this
composite network.
Gene Ontology To evaluate gene function prediction
we use GO function categories for M. musculus. We
use the true-path rule to associate each gene with all
of its functions, i.e. if a gene is annotated in a child
category, we consider it to be annotated in all of its
ancestor categories. In addition, we remove all annotations that were only supported with “inferred from
electronic annotation” (IEA) evidence code; IEA annotations are the only annotations made based on previous computational predictions which have not been
reviewed by a curator (Ashburner et al., 2000).
Evaluation As in (Pena-Castillo et al., 2008), we evaluate our predictions based on two sets of genes: (a)
test genes (i.e. cross-validation) and (b) novel genes.
For predicting test genes, we perform 3-fold crossvalidation on 2,634 GO biological process categories
which have between 3 and 300 annotations (GO association file download on September 1, 2007). Note
that when constructing our hierarchy from which we
derive the HLbias, only GO categories in this set are
included. For predicting novel genes, we train on the
2007 GO association file and then we evaluate our
performance in a “real-life” setting by comparing our
predictions to the updated GO association file (GO
association file download on September 1, 2008). In
particular, for predicting novel genes, we restrict our
evaluation to 903 GO categories that obtained three or
more new annotations since 2007. We report the performance on predicting test and novel genes in terms of
error as measured by 1-area under the ROC (Receiver
Operating Characteristic) curve (1-AUC). The AUC
under the ROC curve (Fawcett, 2006) corresponds to
the probability that a random positive instance will
be scored higher than a random negative instance. In
addition to AUC of ROC, we investigate the performance in predicting novel genes as measured by AUC
of precision recall curve.
424
MOSTAFAVI & MORRIS
Hierarchical label bias
(HLBias)
0.6
Error (1−AUC)
0.5
0.4
UAI 2009
HLBias
HLProp
Down−Propagation
Up−Propagation
IR
GRF
Hierarchical label propagation
(HLProp)
yi,1
initial labels bias
-1
?
+1
GO category 1
?
0.3
GO category 2
?
0.2
0.1
0
yi,2
500
1000
1500
number of GO categories
2000
2500
Figure 2: Cumulative performance (error) of various
methods in predicting the function of test genes in
2,634 GO categories (i.e. using 3-fold cross-validation
on 2008 GO annotation file).
5
Experimental Results
We first show the performance on test genes and then
focus our analysis in predicting novel genes.
5.1
Figure 3: An example illustrating the difference between HLBias and HLProp in assigning discriminant
scores to a test gene. The test gene is depicted by the
node whose initial label bias is a question mark. On
the right figure, the nodes in the same column depict
the same gene in different categories; only one of the
five blue edge representing the edges ha,c ’s is shown.
In HLBias, when predicting GO category 2, the two
neighbouring nodes of the test gene have a more positive label bias. In HLProp, the previous annotations
needs to be propagated through more edges to effect
the discriminant score of the test gene.
Predicting test genes
Figure 2 shows the cumulative distribution of the error
(1-AUC of ROC) of the Hierarchical label bias, HLProp, Down- and Up-propagation, Isotonic Regression
(IR) and Gaussian random fields (GRF) algorithms.
Table 1 summarizes the mean and median error of each
method.
Table 1: Mean and median error in predicting test
genes in 2,634 GO categories. The last column shows
the standard error in the estimate of the mean error.
Approach
mean median
SE
Hierarchical label bias
HLProp
Down-Propagation
Up-Propagation
IR
GRF
0.1035
0.1382
0.1282
0.1735
0.1745
0.1760
0.0881
0.1200
0.1109
0.1405
0.1438
0.1464
0.0016
0.0020
0.0019
0.0028
0.0028
0.0028
HLProp, HLBias, and Down-propagation considerably
reduce the error in gene function prediction. Specifically, despite being the simplest method, Hierarchical label bias achieves the best overall performance in
terms of 1-AUC of ROC curve (Table 1). Note that in
the cross-validation setting, for the Hierarchical label
bias method, the test genes are labeled as unknowns
and the initial label of other non-positive genes is set
according to their previous annotations in the GO hier-
archy (see Section 3.4). One explanation for the better
performance of HLBias compared to HLProp is that,
in using HLBias, genes that have an incomplete annotation in an ancestral category more directly influence
the discriminant scores of their linked genes. This is
because the initial label bias of a gene essentially needs
to be propagated through a minimum of two edges to
effect the label bias of a test gene; in HLProp the incomplete annotation information needs to be propagated through a minimum of three edges to effect the
discriminant score of a test gene (see Figure 3 for a
pictoral description). This explanation may also offer
an insight into why Down-Propagation performs better than Up-Propagation.
In addition, we observed that IR and Up-propagation
do not significantly improve the performance. This result is consistent with the observations in (Obozinski
et al., 2008) that most reconciliation methods often
perform similar to the baseline of independent predictions.
5.2
Predicting novel genes
Here we report the performance in predicting novel
gene function; in particular, to evaluate the performance on a given category, we use newly annotated
genes as positives and all other genes (excluding previously annotated genes, that is, those annotated in the
2007 GO file) as negatives. In predicting novel genes
UAI 2009
MOSTAFAVI & MORRIS
1
True Positive Rate (TPR)
Table 2: Mean and median error in predicting novel
genes in 903 GO categories. The last column shows
the standard error.
Approach
mean median
SE
HLBias
HLProp
IR
GRF
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
False Positive Rate (FPR)
Figure 4: The ROC curve for predicting the GO category “Osteoblast Differentiation” which had 26 annotations in the 2007 GO association file and 9 new
annotations in the 2008 version.
0.5
HLBias
HLProp
Down−Propagation
Up−Propagation
IR
GRF
0.45
0.4
0.35
Error (1−AUC)
425
0.3
Hierarchical label bias
HLProp
Down-Propagation
Up-Propagation
IR
GRF
0.1367
0.1404
0.1415
0.1767
0.1768
0.1772
0.1249
0.1262
0.1305
0.1671
0.1672
0.1673
0.0025
0.0032
0.0032
0.0031
0.0031
0.0031
one newly annotated gene had a previous annotation
in one of the sibling categories. The performance of
HLBias (mean error of 0.1406) was worse than HLProp
in these 13 categories (mean error of 0.1091) and similar to the baseline (mean error of 0.1469). In addition,
some of the newly annotated genes had no previous
annotations in ancestral categories with less than 300
annotations; as we discuss later, the performance of
HLBias was degraded in predicting these categories.
0.25
0.2
HLBias
HLProp
Down−Propagation
Up−Propagation
IR
GRF
0.25
0.15
0.1
0.2
0
0
100
200
300
400
500
600
700
number of GO categories
800
900
1000
Figure 5: Comparison of performance in terms of error
in predicting novel genes in 903 GO categories that
acquired three or more annotations in a span of one
year.
with the HLBias method, for a given GO category, we
adjust the initial label of all except the positive genes
(which have an initial label bias of +1 according to
the 2007 GO file) by using the incomplete annotation
information in GO.
Figure 4 shows a typical set of ROC curves; these
curves were generating for predicting GO category
“Osteoblast Differentiation”. Figure 5 shows the cumulative performance of each method in predicting
novel genes in 903 categories as measured by error (1AUC of ROC). Table 2 shows the mean and median
error. As shown in Figure 5, unlike the cross-validation
setting, the performance of HLProp at high percentiles
(e.g. greater than 75% percentile) is better that that
of HLBias. In particular, we found that our assumption about using genes that are annotated in siblings’
categories as negative example may not always hold.
For example, in 13 of the 903 GO categories, at least
Mean error (1−AUC)
0.05
0.15
0.1
0.05
0
[3−10]
[11−30]
[31−100]
number of annotations
[101−300]
Figure 6: Comparison of performance in terms of error
(1-AUC of ROC) in predicting novel genes. The performance is measured in predicting GO categories with
[3-10] (136 categories), [11-30] (350 categories), [31100] (314 categories), and [101-300] (101 categories)
positive annotations in the 2008 GO file.
To better understand the difference between the various methods, we measured the mean performance in
predicting GO categories at four different specificity
levels; those with 3-10, 11-30, 31-100, and 101-300 annotations in the 2008 GO file. As shown in Figure
6, HLBias performs better than the others in predicting GO categories with 3 to 10 and 101 to 300 annotations whereas the performance of HLProp, Downpropagation, and GO label bias is similar when predicting GO categories with 11 to 100 annotations.
Furthermore, using hierarchical information yields the
426
MOSTAFAVI & MORRIS
0
10
HLBias
HLProp
Down−Propagation
Up−Propagation
IR
GRF
−1
Average precision
10
−2
10
−3
10
−4
10
0
100
200
300
400
500
600
number of GO categories
700
800
900
Figure 7: Comparison of performance in terms of average under the precision-recall curve when predicting
novel genes in 903 GO categories that acquired three
or more annotations in a span of one year.
most improvement on GO categories with a very few
positive examples.
In addition to measuring performance in terms of 1AUC of ROC, we also investigated the performance
in predicting novel gene functions in terms of area
under the precision recall curve. As shown in Figure 7, the performance of HLProp in terms of average
precision, is better than all other methods. Interestingly, in contrast to the ROC measure, we observed
that on average, the performance of HLBias is not significantly different than the baseline approach (t-test
with α=0.05). However, the cumulative performance
of HLBias follows the same trend as measured in error or average precision (compare Figure 5 and 7);
HLBias has a lower precision at high percentile but
higher precision at lower percentiles. The lower performance of HLBias, as measured in terms of precision,
can be explained by our observation that genes with
new annotations in 338 of the 903 novel categories had
no previous annotation in the corresponding ancestral
categories and were therefore deemed negatives and
given a highly negative label bias: the performance
of HLBias in terms of precision at lower recalls (e.g.
recall of 10%) was specifically degraded in these categories.
6
Discussion
Here we have shown that by using the GO hierarchy
information directly, either by setting initial label biases using GO or using our formulation of hierarchical
Gaussian random fields (HLProp), we can significantly
improve gene function prediction. On the other hand,
our results are consistent with the previous report that
reconciliation methods may rarely improve the performance of independent classifiers (Obozinski et al.,
UAI 2009
2008). In contrast, in our setting, reconciliation of independent GO category results in a performance very
similar to the baseline of un-corrected classifications
obtained by GRF.
In order to be able to solve HLProp efficiently, we ignored the directionality of the GO hierarchy. To do so,
we set pmc = pcm if category c is a child of category
m. In contrast, the two heuristics variants (Up- and
Down-propagation) only propagate information about
discriminant scores in one direction. Our results indicate that propagating information down the hierarchy results in most gain whereas Up-Propagation
does not significantly effect the performance. This result is consistent with that of (Obozinski et al., 2008)
which found that reconciliation with a Bayesian network model of GO hierarchy where the arrows are directed from parents to the children classes performs
better than the opposite model where the arrows are
directed from children to the parents.
In its most general form, the GRF and the HLProp
algorithms contain a number of regularization parameters. We have had general success by setting all these
parameters to one though we have not rigorously investigated the effect of changing them because performing
cross-validation on genome scale datasets is computationally costly and sometimes infeasible. However, a
future area of improvements may be to find an optimal
setting for these parameters.
Acknowledgements
We would like to thank David Warde-Farley for helpful
comments. This work was supported by an NSERC
operating grant, a Genome Canada grant administered
by OGI and a CFI/MRI-LOF equipment grant to QM.
SM was partially supported by an OGS fellowship.
References
Ashburner, M., Ball, CA., Blake, JA., Botstein, D.,
Butler, H., & et. al. 2000. Gene Ontology: tool
for unification of biology. Nature Genetics, 25,
25–29.
Bairoch, A. 2000. The Enzyme database in 2000. Nucleic Acids Research, 28, 304–305.
Barlow, R.E., Bartholomew, D.J., Bremmer, J.M., &
Brunk, H.D. 1972. Statistical inference under order restrictions; the theory and application of isotonic regression. New York. ISBN 0-4-71-04970-0:
Wiley.
Barutcuoglu, Z., Schapire, R.E., & Troyanskaya, O.G.
2008. Hierarchical multi-label prediction of gene
function. Bioinformatics, 22(7), 830–836.
UAI 2009
MOSTAFAVI & MORRIS
Burdakov, O., Sysoeve, O., Grimvall, A., & Hussain,
M. 2006. Large-Scale Nonlinear Optimization.
In Nonconvex Optimization and Its Application.
Berlin: Springer-Verlag. Pages 25–33.
427
Pena-Castillo, L., Tasan, M., Myers, CL., Lee, H.,
Joshi, T., & et. al. 2008. A critical assessment
of Mus musculus gene function prediction using
integrated genomic evidence. Genome Biology,
9(Suppl 1), S2.
Eisner, R., Poulin, B., Szafron, D., Lu, P., & Greiner,
R. 2005. Improving Protein Function Prediction
Using the Hierarchical Structure of the Gene Ontology. CIBCB Conference, 354–363.
Shahbaba, B., & Neal, R. 2006. Gene function prediction with hierarchical models with hierarchybased prior. BMC Bioinformatics, 7, 448.
Fawcett, T. 2006. An introduction to ROC analysis.
Pattern Recognition Letters, 27, 861–874.
Shental, O., Siegel, P., Wolf, J., Bickson, D., & Dolve,
D. 2008. Gaussian Belief Propagation Solver for
Systems of Linear Equations. arXiv:0810.1736.
King, O., Foulger, R., Dwight, S., White, J., & Roth,
F. 2003. Predicting Gene Function From Patterns
of Annotations. Genome Research, 13.
Kiritchenko, S., Matwin, S., & Famili, A. Fazel. 2004.
Hierarchical Text Categorization as a Tool of Associating Genes with Gene Ontology Codes. European Workshop on Data Mining and Text Mining
for Bioinformatics at ECML, 26–30.
Lanckriet, G.R.G, Bie, T. De, N.Cristianini, Jordan,
M.I., & Noble, W.S. 2004. A statistical framework for genomic data fusion. Bioinformatics, 20,
2626–2635.
Mostafavi, S., Ray, D., Warde-Farley, D., Grouios, C.,
& Morris, Q. 2008. GeneMANIA: a real-time Multiple Association Network Integration Algorithm
for predicting gene function. Genome Biology,
9(Suppl 1), S4.
Murzin, A., Brenner, S., Hubbard, T., & Chothia, C.
1995. SCOP: a structural classification of proteins
database for the investigation of sequences and
structures. Journal of Molecular Biology, 247,
536–540.
Myers, C.L., Robson, D., Wible, A., Hibbs, M.,
Chiriac, C., Theesfeld, C.L., Dolinski, K., &
Troyanskaya, O.G. 2005. Discovery of biological
networks from diverse functional genomic data.
Genome Biology, 6, R114.
Nocedal, J., & Wright, S.J. 2006. Numerical Optimization. Springer.
Obozinski, G., Lanckriet, G., Grant, C., Jordan, M.I.,
& Noble, W.S. 2008. Consistent probabilistic outputs for protein function prediction. Genome Biology, 9(Suppl 1), S6.
Pavlidis, P., Weston, J., Cai, J., & Noble, W.S. 2002.
Learning gene functional classification from multiple data types. Journal of Computational Biology,
9, 401–411.
Taskar, B., Guestrin, C., & Koller, D. 2003. Max Margin Markov Networks. Advances in Neural Information Processing Systems (NIPS).
Tsuda, K., Shin, H.J., & Schoelkopf, B. 2005. Fast protein classification with multiple networks. Bioinformatics, 21(Suppl 2), ii59–ii65.
von Mering, C., Huynen, M., Jaeggi, D., Schmidt, S.,
Bork, P., & Snel, B. 2003. STRING: a database
of predicted functional associations between proteins. Nucleic Acids Research, 31(1), 258–261.
Zhou, D., Bousquet, O., Weston, J., & Scholkopf, B.
2004. Learning with local and global consistency.
Advances in Neural Information Processing Systems (NIPS), 16, 321–328.
Zhu, X., Ghahramani, Z., & Lafferty, J. 2003. SemiSupervised Learning Using Gaussian Fields and
Harmonic Functions. International Conference on
Machine Learning (ICML), 912–919.
| 5cs.CE
|
Finding Subcube Heavy Hitters in Analytics Data Streams
arXiv:1708.05159v2 [cs.DS] 20 Feb 2018
Branislav Kveton∗
Adobe Research
[email protected]
S. Muthukrishnan
Rutgers University
[email protected]
Hoa T. Vu†
University of Massachusetts
[email protected]
Yikun Xian
Rutgers University
[email protected]
Abstract
Modern data streams typically have high dimensionality. For example, digital analytics streams consist of user
online activities (e.g., web browsing activity, commercial site activity, apps and social behavior, and response to
ads). An important problem is to find frequent joint values (heavy hitters) of subsets of dimensions.
Formally, the data stream consists of d-dimensional items and a k-dimensional subcube T is a subset of k distinct
coordinates. Given a theshold γ, a subcube heavy hitter query Query(T, v) outputs YES if fT (v) ≥ γ and NO if
fT (v) < γ/4 where fT is the ratio of the number of stream items whose coordinates T have joint values v. The all
subcube heavy hitters query AllQuery(T ) outputs all joint values v that return YES to Query(T, v). The problem
is to answer these queries correctly for all T and v.
We present a simple one-pass sampling algorithm to solve the subcube heavy hitters problem in Õ(kd/γ) space.
Õ(·) suppresses polylogarithmic factors. This is optimal up to polylogarithmic factors based on the lower bound of
Liberty et al. [13] In the worst case, this bound becomes Θ(d2 /γ) which is prohibitive for large d.
Our main contribution is to circumvent this quadratic bottleneck via a model-based approach. In particular,
we assume that the dimensions are related to each other via the Naive Bayes model. We present a new two-pass,
Õ(d/γ)-space algorithm for our problem, and a fast algorithm for answering AllQuery(T ) in Õ((k/γ)2 ) time.
We demonstrate the effectiveness of our approach on a synthetic dataset as well as real datasets from Adobe
and Yandex. Our work shows the potential of model-based approach to data streams.
∗ The
† Part
authors are listed alphabetically.
of this work was done at Adobe Research.
1
1
Introduction
We study the problem of finding heavy hitters in high dimensional data streams. Most companies see transactions
with items sold, time, store location, price, etc. that arrive over time. Modern online companies see streams of
user web activities that typically have components of user information including ID (e.g. cookies), hardware (e.g.,
device), software (e.g., browser, OS), and contents such as web properties, apps. Activity streams also include events
(e.g., impressions, views, clicks, purchases) and event attributes (e.g., product id, price, geolocation, time). Even
classical IP traffic streams have many dimensions including source and destination IP addresses, port numbers and
other features of an IP connection such as application type. Furthermore, in applications such as Natural Language
Processing, streams of documents can be thought of as streams of a large number of bigrams or multi-grams over
word combinations [8]. As these examples show, analytics data streams with 100’s and 1000’s of dimensions arise in
many applications. Motivated by this, we study the problem of finding heavy hitters on data streams focusing on d,
the number of dimensions, as a parameter. Given d one sees in practice, d2 in space usage is prohibitive, for solving
the heavy hitters problem on such streams.
Formally, let us start with a one-dimensional stream of items x1 , . . . xm where each xi ∈ [n] := {1, 2, . . . , n}.
We can look at the count c(v) = |{i : xi = v}| or the frequency ratio f (v) = c(v)/m. A heavy hitter value v is one
with c(v) ≥ γm or equivalently f (v) ≥ γ, for some constant γ. The standard data stream model is that we maintain
data structures of size polylog(m, n) and determine if v is a heavy hitter with probability of success at least 3/4,
that is, if f (v) ≥ γ output YES and output NO if f (v) < γ/4 for all v.1 We note that if γ/4 ≤ f (v) < γ, then either
answer is acceptable.
Detecting heavy hitters on data streams is a fundamental problem that arises in guises such as finding elephant
flows and network attacks in networking, finding hot trends in databases, finding frequent patterns in data mining,
finding largest coefficients in signal analysis, and so on. Therefore, the heavy hitters problem has been studied
extensively in theory, databases, networking and signal processing literature. See [4] for an early survey and [19]
for a recent survey.
Subcube heavy hitter problems Our focus is on modern data streams such as in analytics cases, with d dimensions, for large d. The data stream consists of d-dimensional items x1 , · · · , xm . In particular,
xi = (xi,1 , . . . , xi,d ) and each xi,j ∈ [n] .
A k-dimensional subcube T is a subset of k distinct coordinates {T1 , · · · , Tk } ⊆ [d]. We refer to the joint values of
the coordinates T of xi as xi,T .
The number of items whose coordinates T have joint values v is denoted by cT (v), i.e., cT (v) = |{i : xi,T = v}| .
Finally, we use XT to denote the random variable of the joint values of the coordinates T of a random item. We have
the following relationship
cT (v)
.
fT (v) := Pr [XT = v] =
m
For a single coordinate i, we slightly abuse the notation by using fi and f{i} interchangeably. For example, fTi (v)
is the same as f{Ti } (v). Similarly, Xi is the same as X{i} .
We are now ready to define our problems. They take k, γ as parameters and the stream as the input and build
data structures to answer:
• Subcube Heavy Hitter: Query(T, v), where |T | = k, and v ∈ [n]k , returns an estimate if fT (v) ≥ γ. Specifically,
output YES if fT (v) ≥ γ and NO if fT (v) < γ/4. If γ/4 ≤ fT (v) < γ, then either output is acceptable. The
required success probability for all k-dimensional subcubes T and v ∈ [n]k is at least 3/4.
• All Subcube Heavy Hitters: AllQuery(T ) outputs all joint values v that return YES to Query(T, v). This is conditioned on the algorithm used for Query(T, v).
It is important to emphasize that the stream is presented (in a single pass or constant passes) to the algorithm before
the algorithm receives any query.
1 The gap constant 4 can be narrowed arbitrarily and the success probability can be amplified to 1 − δ as needed, and we omit these factors in
the discussions.
1
Subcube heavy hitters are relevant wherever one dimensional heavy hitters have found applications: combination
of source and destination IP addresses forms the subcube heavy hitters that detect network attacks; combination of
stores, sales quarters and nature of products forms the subcube heavy hitters that might be the pattern of interest
in the data, etc. Given the omnipresence of multiple dimensions in digital analytics, arguably, subcube heavy hitters
limn the significant data properties far more than the single dimensional view.
Related works The problem we address is directly related to frequent itemset mining studied in the data mining
community. In frequent itemset mining, each dimension is binary (n = 2), and we consider Query(T, v) where
v = (1, . . . , 1) := Uk . It is known that counting all maximal subcubes T that have a frequent itemset, i.e., fT (Uk ) ≥
γ, is #P -complete [21]. Furthermore, finding even a single T of maximal size such that fT (Uk ) ≥ γ is NP-hard
[9, 13]. Recently, Liberty et al. showed that any constant-pass streaming algorithm answering Query(T, Uk ) requires
Ω(kd/γ · log(d/k)) bits of memory [13]. In the worst case, this is Ω(d2 /γ) for large k, ignoring the polylogarithmic
factors. For this specific problem, sampling algorithms will nearly meet their lower bound for space. Our problem is
more general, with arbitrary n and v.
Our contributions Clearly, the case k = 1 can be solved by building one of the many known single dimensional
data structures for the heavy hitters problem on each of the d dimension; the k = d case can be thought
of as a
giant single dimensional problem by linearizing the space of all values in [n]k ; for any other k, there are kd distinct
choices for subcube T , and these could
be treated as separate one-dimensional problems by linearizing each of the
subcubes. In general, this entails kd and log(nd ) cost in space or time bounds over the one-dimensional case, which
we seek to avoid. Also, our problem can be reduced to the binary case by unary encoding each dimension by n bits,
and solving frequent itemset mining: the query then has kn dimensions. The resulting bound will have an additional
n factor which is large.
First, we observe that the reservoir sampling approach [18] solves subcube heavy hitters problems more efficiently compared to the approaches mentioned above. Our analysis shows that the space we use is within polylogarithmic factors of the lower bound shown in [13] for binary dimensions and query vector Uk , which is a special
case of our problem. Therefore, the subcube heavy hitters problem can be solved using Õ(kd/γ) space. However,
this is Ω(d2 ) in worst case.
Our main contribution is to avoid this quadratic bottleneck for finding subcube heavy hitters. We adopt the notion
that there is an underlying probabilistic model behind the data, and in the spirit of the Naive Bayes model, we assume
that the dimensions are nearly (not exactly) mutually independent given an observable latent dimension. This could
be considered as a low rank factorization of the dimensions. In particular, one could formalize this assumption by
bounding the total variational distance between the data’s joint distribution and that derived from the Naive Bayes
formula. This assumption is common in statistical data analysis and highly prevalent in machine learning. Following
this modeling, we make two main contributions:
• We present a two-pass, Õ(d/γ)-space streaming algorithm for answering Query(T, v). This improves upon the
kd factor in the space complexity from sampling, without assumptions, to just d with the Naive Bayes assumption,
which would make this algorithm practical for large k. Our algorithm uses sketching in each dimension in one
pass to detect heavy hitters, and then needs a second pass to precisely estimate their frequencies.
• We present a fast algorithm for answering AllQuery(T ) in Õ((k/γ)2 ) time. The naive procedure would take
exponential time Ω((1/γ)k ) by considering the Cartesian product of the heavy hitters in each dimension. Our approach, on the other hand, uses the structure of the Naive Bayes assumption to iteratively construct the subcube
heavy hitters one dimension at a time.
Our work develops the direction of model-based data stream analysis. Model-based data analysis has been effective in other areas. For example, in compressed sensing, realistic signal models that include dependencies between
values and locations of the signal coefficients improve upon unconstrained cases [7]. In statistics, using tree constrained models of multidimensional data sometimes improves point and density estimation. In high dimensional
distribution testing, model based approach has also been studied to overcome the curse of dimensionality [6].
In the data stream model, [10, 2, 3] studied the problem of testing independence. McGregor and Vu [15] studied
the problem of evaluating Bayesian Networks. In another work, Kveton et al. [11] assumed a tree graphical model
and designed a one-pass algorithm that estimates the joint frequency; their work however only solved the k = d
case for the joint frequency estimation problem. Our model is a bit different and more importantly, we solve the
2
subcube heavy hitters problem (addressing all the kd subcubes) which prior work does not solve. In following
such a direction, we have extended the fundamental heavy hitters problem to higher dimensional data. Given that
many implementations already exist for the sketches we use for one-dimensional heavy hitters as a blackbox, our
algorithms are therefore easily implementable.
Background on the Naive Bayes model and its use in our context. The Naive Bayes Model [17] is a Bayesian
network over d features X1 , . . . , Xd and a class variable Y . This model represents a joint probability distribution of
the form
Pr [X1 = x1 , . . . , Xd = xd , Y = y]
= Pr [Y = y]
d
Y
Pr [Xj = xj | Y = y] ,
j=1
which means that the values of the features are conditionally independent given the value of the class variable. The
simplicity of the Naive Bayes model makes it a popular choice in text processing and information retrieval [12, 14],
with state-of-the-art performance in spam filtering [1], text classification [12], and others.
Empirical study. We perform detailed experimental study of subcube heavy hitters. We use a synthetic dataset
where we generate data that confirms to the Naive Bayes model. We then experiment with real data sets from
Yandex (Search) and Adobe (Marketing Cloud) which give multidimensional analytics streams. We experiment with
the reservoir sampling based algorithm as a benchmark that works without modeling assumptions, and our two-pass
subcube heavy hitters algorithm that improves upon it for data that satisfies the model. We also adopt our approach
to give a simpler one-pass algorithm for which theoretical guarantees is weaker. Our experiments show substantial
improvement of the model-based algorithms over the benchmark for synthetic as well as real data sets, and further
show the benefits of the second pass.
2
The Sampling Algorithm
In this section, we show that sampling
solves the problem efficiently compared to running one-dimensional heavy
hitters algorithms for each of kd k-dimensional subcubes independently. It also matches the lower bound in [13]
up to polylogarithmic factors.
Algorithm details. The algorithm samples m0 = Õ(γ −1 kd) random items z1 , . . . , zm0 from the stream using
Reservoir sampling [18]. Let S = {z1 , . . . , zm0 } be the sample set. Given Query(T, v), we output YES if and only if
the sample frequency of v, denoted by fˆT (v), is at least γ/2. Specifically,
|{xi : xi ∈ S and xi,T = v}|
.
fˆT (v) :=
m0
For all subcubes T and joint values v of T , the expected sample frequency fˆT (v) is fT (v). Intuitively, if v is a
frequent joint values, then its sample frequency fˆT (v) ≈ fT (v); otherwise, fˆT (v) stays small.
Let us fix a k-dimensional subcube T and suppose that for all v ∈ [n]k , we have
max{γ, fT (v)}
fˆT (v) = fT (v) ±
.
4
(1)
It is then straightforward to see that if fT (v) < γ/4, then fˆT (v) < γ/4 + γ/4 = γ/2. Otherwise, if fT (v) ≥ γ,
then fˆT (v) ≥ 3fT (v)/4 ≥ 3γ/4 > γ/2. Hence, we output YES for all v where fˆT (v) ≥ γ/2, and output NO
otherwise.
Lemma
Pn2.1. (Chernoff bound) Let X1 , · · · , Xn be independent or negatively correlated binary random variables. Let
X = i=1 Xi and µ = E [X]. Then,
Pr [|X − µ| ≥ µ] ≤ exp(− min{2 , }µ/3) .
3
Recall that S = {z1 , z2 , . . . , zm0 } is the sample set returned by the algorithm. For a fixed v ∈ [n]k , we use Zi
as the indicator variable for the event zi,T = v. Since we sample without replacement, the random variables Zi
are negatively correlated. The following lemma shows that Eq. 1 holds for all v and k-dimensional subcubes T via
Chernoff bound.
Lemma 2.2. For all k-dimensional subcubes T and joint values v ∈ [n]k , with probability at least 0.9,
max{γ, fT (v)}
.
fˆT (v) = fT (v) ±
4
Proof. Let m0 = cγ −1 log(dk · nk ) for some sufficiently large constant c. We first consider a fixed v ∈ [n]k and define
the random variables Zi as above, i.e., Zi = 1 if zi,T = v. Suppose fT (v) ≥ γ. Appealing to Lemma 2.1, we have
0
m
X
fT (v)
Zi
− fT (v) ≥
Pr
0
m
4
i=1
fT (v)
= Pr fˆT (v) − fT (v) ≥
4
fT (v)m0
1
≤ exp −
≤
.
3 × 16
10dk nk
On the other hand, if fT (v) < γ/4, then
h
γi
γ
m0
ˆ
Pr fT (v) − fT (v) ≥
≤ exp −
fT (v)
4
4fT (v)
3
1
≤
.
10dk nk
Therefore, by taking the union bound over all kd · nk ≤ dk · nk possible combinations of k-dimensional subcubes
and the corresponding joint values v ∈ [n]k , we deduce the claim.
We therefore could answer all Query(T, v) correctly with probability at least 0.9 for all joint values v ∈ [n]k
and k-dimensional subcubes T . Because storing each sample zi requires Õ(d) bits of space, the algorithm uses
Õ(dkγ −1 ) space. We note that answering Query(T, v) requires computing fˆT (v) which takes O(|S|) time. We can
answer AllQuery(T ) by computing fˆT (v) for all joint values v of coordinates T that appear in the sample set which
2
will take O(|S| ) time. We summarize the result as follows.
Theorem 2.3. There exists a 1-pass algorithm that uses Õ(dkγ −1 ) space
and solves k-dimensional
subcube heavy
−1 2
−1
hitters. Furthermore, Query(T, v) and AllQuery(T ) take Õ(dkγ ) and Õ dkγ
time respectively.
3
The Near-Independence Assumption
The near-independence assumption. Suppose the random variables representing the dimensions X1 , X2 , . . . , Xd
are near independent. We show that there is a 2-pass algorithm that uses less space and has faster query time. At a
high level, we make the assumption that the joint probability is approximately factorized
f{1,...,d} (v) ≈ f1 (v1 )f2 (v2 ) · · · fd (vd ) .
More formally, we assume that the total variation distance is bounded by a small quantity α. Furthermore, we
assume that α is reasonable with respect to γ that controls the heavy hitters. For example, α ≤ γ/10 will suffice.
The formal near-independence assumption is as follows: There exists α ≤ γ/10 such that for all subcubes T ,
max
v∈[n]|T |
fT (v) −
|T |
Y
i=1
We observe that:
4
fTi (vi ) < α .
• If fT (v) ≥ γ, then
|T |
Y
fTi (vi ) ≥ fT (v) − γ/10 > γ/2 .
i=1
• If fT (v) < γ/4, then
|T |
Y
fTi (vi ) ≤ fT (v) + γ/10 < γ/2 .
i=1
Thus, it suffices to output YES to Query(T, v) if and only if the marginals product
nience, let
λ := γ/2 .
Q|T |
i=1
fTi (vi ) ≥ γ/2. For conve-
Algorithm details. We note that simply computing fi (x) for all coordinates i ∈ [d] and x ∈ [n] will need Ω(dn)
space. To over come this, we make following simple but useful observation. We observe that if v is a heavy hitter in
the subcube T and if T 0 is a subcube of T , then vT 0 is a heavy hitter in the subcube T 0 .
Lemma 3.1. For all subcubes T ,
|T |
Y
fTi (vi ) ≥ λ implies
i=1
Y
fTi (vi ) ≥ λ
i∈V
for all V ⊆ [|T |] (i.e., {Ti : i ∈ V} is a subcube of T ).
The proof is trivial since all fTi (vi ) ≤ 1. Therefore, we have the following corollary.
Corollary 3.2. For all subcubes T ,
|T |
Y
fTi (vi ) ≥ λ implies fTi (vi ) ≥ λ for all i ∈ [|T |] .
i=1
We therefore only need to compute fi (x) if x is a heavy hitter in coordinate i. To this end, for each coordinate
i ∈ [d], by using (for example) Misra-Gries algorithm [16] or Count-Min sketch [5], we can find a set Hi such that if
fi (x) ≥ λ/2, then x ∈ Hi and if fi (x) < λ/4, then x ∈
/ Hi . In the second pass, for each x ∈ Hi , we compute fi (x)
exactly to obtain
Si := {x ∈ [n] : fi (x) ≥ λ} .
Q|T |
We output YES to Query(T, v) if and only if all vi ∈ STi and i=1 fTi (vi ) ≥ λ. Note that if v ∈ Si , then fi (v)
is available to the algorithm since it is computed exactly in the second pass. The detailed algorithm is as follows.
1. First pass: For each coordinate i ∈ [d], use Misra-Gries algorithm to find Hi .
2. Second pass: For each coordinate i ∈ [d], compute fi (x) exactly for each x ∈ Hi to obtain Si .
3. Output YES to Query(T, v) if and only if vi ∈ STi for all i ∈ [|T |] and
|T |
Y
fTi (vi ) ≥ λ .
i=1
Theorem 3.3. There exists a 2-pass algorithm that uses Õ(dγ −1 ) space and solves subcube heavy hitters under the
near-independence assumption. The time to answer Query(T, v) and AllQuery(T ) are Õ(k) and Õ(kγ −1 ) respectively
where k is the dimensionality of T .
5
Proof. The first pass uses Õ(dλ−1 ) space since Misra-Gries algorithm uses Õ(λ−1 ) space for each coordinate i ∈ [d].
Since the size of each Hi is O(λ−1 ), the second pass also uses Õ(dλ−1 ) space. Recall that λ = γ/2. We then conclude
that the algorithm uses Õ(dγ −1 ) space.
For an arbitrary Query(T, v), the algorithm’s correctness follows immediately from Corollary 3.2 and the observation that if vi ∈ STi , then fTi (vi ) is available since it was computed exactly in the second pass. Specifically,
if
|T |
Y
(2)
fTi (vi ) ≥ λ ,
i=1
then vi ∈ STi for all i ∈ [|T |] and we could verify the inequality and output YES. On the other hand, suppose Eq. 2
does not hold. Then, if vi ∈
/ STi for some i, we correctly output NO. But if all vi ∈ STi , then we are able to verify
that the inequality does not hold (and correctly output NO).
The parameter k only affects the query time. We now analyze the time to answer Query(T, v) and AllQuery(T )
for a k-dimensional subcube T .
We can easily see that Query(T, v) takes Õ(k) time as we need to check if all vi ∈ STi (e.g., using binary
Qk
searches) and compute i=1 fTi (vi ).
Next, we exhibit a fast algorithm to answer AllQuery(T ). We note that naively checking all combinations
(v1 , · · · , vk ) in ST1 × ST2 × · · · × STk takes exponential Ω(γ −k ) time in the worst case.
Our approach figures out the heavy hitters gradually and takes advantage of the near-independence assumption.
In particular, define
Wj := {v ∈ [n]j : fT1 (v1 ) · · · fTj (vj ) ≥ λ} .
Recall that the goal is to find Wk . Note that W1 = S1 is obtained directly by the algorithm. We now show that it
is possible to construct Wj+1 from Wj in Õ(λ−1 ) time which in turn means that we can find Wk in Õ(kλ−1 ) time.
We use the notation T[j] := {T1 , . . . , Tj } and v[j] := (v1 , v2 , . . . , vj ).
Qj
We note that |Wj | ≤ 5/(4λ). This holds since if y ∈ Wj , then i=1 fTi (yi ) ≥ λ. Appealing to the nearindependence assumption, we have
fT[j] (y) ≥
j
Y
fTi (yi ) − α ≥ λ − α ≥ 4/5 · λ .
i=1
For each y ∈ Wj , we collect all x ∈ Sj+1 such that
j
Y
!
fTi (yi ) fTj+1 (x) ≥ λ
i=1
and put (y1 , · · · , yj , x) into Wj+1 . Since |Wj | ≤ 5/4 · λ−1 and |Sj+1 | ≤ λ−1 , this step obviously takes O(λ−2 )
Qj
time. However, by observing that there could be at most λ−1 i=1 fTi (yi ) such x for each y ∈ Wj , the upper bound
for the number of combinations of x and y is
j
X 1
X 1Y
fTi (yi ) ≤
(fT (y) + α)
λ i=1
λ [j]
y∈Wj
y∈Wj
=
X α
X fT[j] (y)
+
λ
λ
y∈Wj
y∈Wj
1
3
≤ |Wj | + ≤ .
λ
λ
P
The last inequality follows from the assumption that α ≤ λ/5 and y∈Wj fT[j] (y) ≤ 1. Thus, the algorithm can
find Wj+1 given Wj in Õ(λ−1 ) time. Hence, we obtain Wk in Õ(kλ−1 ) = Õ(kγ −1 ) time. The correctness of this
procedure follows directly from Lemma 3.1 and induction since v = (v1 , . . . , vj+1 ) ∈ Wj+1 implies that v[j] ∈ Wj
and vj+1 ∈ Sj+1 . Thus, by checking all combinations of y ∈ Wj and x ∈ Sj+1 , we can construct Wj+1 correctly.
6
4
The Naive Bayes Assumption
The Naive Bayes assumption. In this section, we focus on the data streams inspired by the Naive Bayes model
which is strictly more general than the near-independence assumption. In particular, we assume that the coordinates
are near-independent given an extra (d+1)th observable class coordinate that has a value in {1, . . . , `}. The (d+1)th
coordinate is also often referred to as the latent coordinate.
As in typical in Naive Bayes analysis, we assume ` is a constant but perform the calculations in terms of ` so its
role in the complexity of the problem is apparent.
Informally, this model asserts that the random variables representing coordinates X1 , . . . , Xd are near independent conditioning on a the random variable Xd+1 that represents the class coordinate.
We introduce the following notation
{xi : xi,T = v ∧ xi,{d+1} = z}
{xi : xi,{d+1} = z}
fT | d+1 (v | z) :=
= Pr [XT = v | Xd+1 = z] .
In other words, fT | d+1 (v | z) is the frequency of the joint values v in the T coordinates among the stream items
where the class coordinate d + 1 has value z.
The formal Naive Bayes assumption is as follows: There exists α ≤ γ/10 such that for all subcubes T ,
max
v∈[n]|T |
X
fT (v) −
fd+1 (z)
|T |
Y
fTi | d+1 (vi | z) < α .
i=1
z∈[`]
Algorithm details. As argued in the previous section, it suffices to output YES to Query(T, v) if and only if
X
fd+1 (z)
|T |
Y
fTi | d+1 (vi | z) ≥ γ/2 = λ .
i=1
z∈[`]
However, naively computing all fi | d+1 (v|z) uses Ω(`dn) space. We circumvent this problem by generalizing Lemma
3.1 as follows. If a joint values v is a heavy hitter in a subcube T in the Naive Bayes formula and T 0 is a subcube of
T , then vT 0 is a heavy hitter in the subcube T 0 .
Lemma 4.1. For all subcubes T ,
q(v) :=
X
fd+1 (z)
X
fTi | d+1 (vi | z) ≥ λ
i=1
z∈[`]
implies
|T |
Y
fd+1 (z)
Y
fTi | d+1 (vi | z) ≥ λ
i∈V
z∈[`]
for all V ⊆ [|T |] (i.e., {Ti : i ∈ V} is a subcube of T ).
Proof. For a fixed z, observe that
X
fTj | d+1 (yj | z) = 1 .
yj ∈[n]
7
Suppose q(v) ≥ λ and consider an arbitrary V ⊆ [|T |]. We have
X
Y
fd+1 (z)
fTi | d+1 (vi | z)
z∈[`]
i∈V
X
Y
=
≥
fd+1 (z)
i∈V
j ∈V
/
X
Y
Y
fd+1 (z)
X
fTi | d+1 (vi | z)
i∈V
fd+1 (z)
|T |
Y
X
z∈[`]
z∈[`]
=
fTi | d+1 (vi | z)
Y
fTj | d+1 (yj | z)
yj ∈[n]
fTj | d+1 (vj | z)
j ∈V
/
fTi | d+1 (vi | z) = q(v) ≥ λ .
i=1
z∈[`]
An alternative proof is by noticing that q(v) is a valid probability density function of |T | variables. The claim follows
by marginalizing over the the variables that are not in V.
Setting V = {i} for each i ∈ [|T |] and appealing to the fact that
X
X
fd+1 (z)fTi | d+1 (vi | z) =
f{Ti ,d+1} ((vi , z)) = fTi (vi ) ,
z∈[`]
z∈[`]
we deduce the following corollary.
Corollary 4.2. For all subcubes T ,
X
z∈[`]
fd+1 (z)
|T |
Y
fTi | d+1 (vi | z) ≥ λ implies fTi (vi ) ≥ λ
i=1
for all i ∈ [|T |].
Therefore, we only need to compute fi | d+1 (x | z) for all coordinates i ∈ [d], values z ∈ [`] if x is a heavy hitter
of coordinate i. Similar to the previous section, for each dimension i ∈ [d], we find Hi in the first pass and use Hi
to find Si in the second pass. Appealing to Corrollary 4.2, we deduce that if
q(v) :=
X
fd+1 (z)
z∈[`]
|T |
Y
fTi | d+1 (vi | z) ≥ λ
i=1
then for all i = 1, 2, . . . , |T |, we have fTi (vi ) ≥ λ which in turn implies that vi ∈ STi . Therefore, we output YES to
Query(T, v) if and only if all vi ∈ STi and q(v) ≥ λ.
To this end, we only need to compute fi | d+1 (x | z) and fd+1 (z) for all x ∈ Hi , z ∈ [`], and i ∈ [d]. The detailed
algorithm is as follows.
1. First pass:
(a) For each value z ∈ [`], compute fd+1 (z) exactly.
(b) For each coordinate i ∈ [d], use Misra-Gries algorithm to find Hi .
2. Second pass:
(a) For each coordinate i ∈ [d] and each value x ∈ Hi , compute fi (x) exactly to obtain Si .
(b) For each value z ∈ [`], coordinate i ∈ [d], and x ∈ Hi , compute fi | d+1 (x | z) exactly.
3. Output YES to Query(T, v) if and only if vi ∈ STi for all i ∈ [|T |] and
X
z∈[`]
fd+1 (z)
|T |
Y
fTi | d+1 (vi | z) ≥ λ .
i=1
8
Theorem 4.3. There exists a 2-pass algorithm that uses Õ(`dγ −1 ) space and solves subcube heavy hitters under the
Naive Bayes assumption. The time to answer Query(T, v) and AllQuery(T ) are Õ(`k) and O(`(k/γ)2 ) respectively
where k is the dimensionality of T .
Proof. The space to obtain Hi and Si over the two passes is Õ(dλ−1 ). Additionally, computing fi | d+1 (x | z) for all
i ∈ [d], z ∈ [`], and x ∈ Hi requires Õ(`dλ−1 ) bits of space. The overall space we need is therefore Õ(`dλ−1 ) =
Õ(`dγ −1 ).
The correctness of answering an arbitrary Query(T, v) follows directly from Corollary 4.2. Specifically, if
X
fd+1 (z)
|T |
Y
fTi | d+1 (vi | z) ≥ λ ,
(3)
i=1
z∈[`]
then, vi ∈ STi ⊆ HTi for all i ∈ [|T |] as argued. Hence, fTi | d+1 (vi | z) is computed exactly in the second pass for
all z ∈ [`]. As a result, we could verify the inequality and output YES. On the other hand, if Eq. 3 does not hold.
Then, if some vi ∈
/ STi , we will correctly output NO. Otherwise if all vi ∈ STi , then we can compute the left hand
side and verify that Eq. 3 does not hold (and correctly output NO).
Obviously, Query(T, v) takes Õ(`k) time for a k-dimensional subcube T . We now exhibit a fast algorithm to
answer AllQuery(T ) for a k-dimensional subcube T . Define
Wj := {v ∈ [n]j :
X
fd+1 (z)
j
Y
fTi | d+1 (vi | z) ≥ λ} .
i=1
z∈[`]
Recall that the goal is to find Wk . We note that W1 = S1 is obtained directly by the algorithm. Next, we show how
to obtain Wj+1 in Õ(λ−2 ) time from Wj . Note that |Wj | ≤ 5/(4λ) because if y ∈ Wj , then
X
fT1 | d+1 (y1 | z) · · · fTj | d+1 (yj | z)fd+1 (z) ≥ λ
z∈[`]
and hence fT[j] (y) ≥ λ − α = 4/5 · λ−1 according to the Naive Bayes assumption. This implies that |Wj | ≤ 5/(4λ).
For each (v1 , · · · , vj ) in Wj , we collect all vj+1 ∈ Sj+1 such that
X
fT1 | d+1 (v1 | z) · · · fTj+1 | d+1 (vj+1 | z)fd+1 (z) ≥ λ
z∈[`]
and put (v1 , · · · , vj+1 ) to Wj+1 . Since |Wj | ≤ 5/(4λ) and |Sj+1 | ≤ 1/λ, this step obviously takes Õ(`kλ−2 ) time.
Since we need to do this for j = 2, 3, . . . , k, we attain Wk in Õ(`(k/γ)2 ) time. The correctness of this procedure
follows directly from Lemma 4.1 and induction since (v1 , . . . , vj+1 ) ∈ Wj+1 implies that (v1 , . . . , vj ) is in Wj and
vj+1 is in Sj+1 . Since we check all possible combinations of (v1 , . . . , vj ) ∈ Wj and vj+1 ∈ Sj+1 , we guarantee to
construct Wj+1 correctly.
5
Experimental study
Overview. We experiment with our algorithms on a synthetic dataset generated from a Naive Bayes model, and
two real-world datasets from Adobe Marketing Cloud2 and Yandex. We thoroughly compare the following approaches:
• The sampling method (Sampling) in Section 2.
• The 2-pass algorithms (TwoPassAlg) described in Section 3 and 4 depending on the context of the experiment.
• The Count-Min sketch heuristic (Heuristic): this heuristic uses Count-Min sketch’s point query estimation (see
[5]) to estimate the frequencies given by the near-independence formula (instead of making a second pass
through the stream to compute their exact values). We note that this approach has no theoretical guarantee.
2 http://www.adobe.com/marketing-cloud.html
9
We highlight the main differences between the theoretical algorithms in Sections 3 and 4 and the actual implementation:
• Instead of running our algorithms with the theoretical memory bounds, we run and compare them for different
memory limits. This approach is more practical and natural from the implementation perspective.
• In theory, Sampling and TwoPassAlg use a fixed threshold γ ∗ = γ/2 to decide between outputting YES or NO.
We however experiment with different values of γ ∗ which is helpful when the memory is more limited or when
the assumptions are not perfect in real data.
The heavy hitters threshold γ is carefully chosen so that the proportion of the number of heavy hitters to the total
number joint values to be reasonably small, i.e., approximately at most 1% in this paper. Therefore, we use different
values of γ for each dataset (see Table 1 for the actual parameter values).
5.1
Synthetic dataset
The synthetic dataset is sampled from a pre-trained Naive Bayes model that is used to estimate the probability of
a page view. The model was provided by [11] and built on the same Clickstream dataset that we used in Section
5.2. The coordinates consist of one class variable Z and five feature variables (X1 , . . . , X5 ) with high cardinalities.
The dataset strongly follows the property that X1 , . . . , X5 are conditionally independent given Z. Specifically, the
variables and their corresponding approximate cardinalities are: country (7), city (10,500), page name (8,500), starting
page name (6,400), campaign (3,500), browser (300) where country is the class variable.
Warm up experiment We first evaluate Sampling and Heuristic on this synthetic dataset. As mentioned earlier,
we compare the performance of the two approaches for each fixed memory size.3 We take a subset of approximately
135, 000 records conditioned on a fixed and most frequent value of Z so that X1 , . . . , X5 are independent in this
subset. We then run experiments on three different subcubes: {X1 , X2 , X3 }, {X2 , X3 , X4 }, and {X3 , X4 , X5 }.
In this warm up experiment, the main goal is not to find the heavy hitters but to compare the accuracy of the
heavy hitters frequency estimations given by Heuristic and Sampling. We measure the performance via the mean
square error (MSE), the mean absolute error (MAE), and the mean absolute percentage error (MAPE). To do this,
the true frequencies were pre-computed. We use the frequencies of the top 10 heavy hitters in each of the above
subcubes. The results (see Figure 1) indicate that Heuristic outperforms Sampling when restricted to small memory.
This warm up experiment gives evidence that knowing the underlying distribution structure helps improving small
space heuristic’s performance in estimating the heavy hitters frequencies.
6
Sampling
Heuristic
2
Sampling
Heuristic
4
14
16
0.00
0.02
0.04 0.06
Memory size
0.08
0.10
log(MAPE)
12
Sampling
Heuristic
1
5
10
log(MAE)
log(MSE)
8
6
0
1
7
2
8
3
0.00
0.02
0.04 0.06
Memory size
0.08
0.10
0.00
0.02
0.04 0.06
Memory size
0.08
0.10
Figure 1: Warm up experiment on synthetic data. Memory size ranges from 0.1% to 10% of data size. We report the error as a
function of memory size.
Experiment with the near-independence assumption. We compare performance of the three aforementioned
methods on finding heavy hitters under the near-independence assumption. In this experiment, we use the same
subset of data and subcubes as in the previous experiment. We fix the memory to be 2% of data size.
3 We compute the memory use by the Sampling as the product of dimension and the sample size. The memory used by Heuristic is computed
as the product of dimension and the Count-Min sketch’s size.
10
Dataset
Synthetic (fixed Z)
Synthetic (whole)
Clickstream
Yandex
Mem.
2%
2%
10%
0.2%
#Subcubes
3
3
4
8
γ
0.002
0.002
0.002
0.1
#HH
29.7
28.7
42.0
2.2
HH ratio
0.079%
0.054%
0.165%
1.65%
Table 1: Parameter values for each experiment.
(The columns correspond to memory size relative to the dataset, number of the experimented subcubes, average number of heavy
hitters, average percentage of heavy hitters.)
We measure the performance, for different values of γ ∗ , based on the number of true positives and false positives.
As shown in Figure 2, for small memory, both Heuristic and TwoPassAlg manage to find more heavy hitters than
Sampling. In terms of false positives, TwoPassAlg beats both Heuristic and Sampling for smaller space. One possible
explanation is that when γ ∗ is small (close to γ), TwoPassAlg, with the advantage of the second pass, accurately
estimates frequencies of potential heavy hitters whereas the other two methods, especially Heuristic, overestimate
the frequencies and therefore report more false positives. For larger γ ∗ , false positives become less likely and all
three approaches achieve similar performances. In general, TwoPassAlg obtains the best performance as seen in
the ROC curve.
Sampling
Heuristic
TwoPassAlg
70
60
50
40
90
Sampling
Heuristic
TwoPassAlg
1000
800
False Positives(s)
True Positive(s)
80
80
True Positive(s)
90
600
400
0.006
*
0.008
0.010
50
Sampling
Heuristic
TwoPassAlg
30
0
0.004
60
40
200
30
70
0.004
0.006
*
0.008
0.010
0
200
400
600
800
False Positive(s)
1000
Figure 2: Near-independence experiment on synthetic dataset. We measure the performance based on the number of true and
false positives (as a function of γ ∗ ), and the ROC curve.
Experiment with the Naive Bayes assumption We use the whole dataset of approximately 168, 000 records
without fixing Z and keep other settings unchanged. We only compared the performance of TwoPassAlg and
Sampling because the conditional probabilities cannot be directly derived from Heuristic. In Figure 3, we observe
that when restricted to small memory, TwoPassAlg attains a better performance by reporting more true heavy hitters
and fewer false heavy hitters. As we allow more space, the performance of Sampling improves as predicted by our
theoretical analysis.
Sampling
TwoPassAlg
80
50
40
800
600
400
30
200
20
0
0.002
0.004
0.006
*
0.008
0.010
70
True Positive(s)
False Positive(s)
60
Sampling
TwoPassAlg
80
1000
70
True Positive(s)
Sampling
TwoPassAlg
1200
60
50
40
30
0.002
0.004
0.006
*
0.008
0.010
20
0
200
Figure 3: Naive Bayes experiment on synthetic dataset.
11
400 600 800 1000 1200
False Positive(s)
5.2
Clickstream dataset
To evaluate TwoPassAlg on real data, we use an advertising dataset called Clickstream Data Feeds from Adobe Marketing Cloud. The approximate dataset size is 168, 000 and all values have been anonymized in advance.
There are 19 high cardinality variables grouped by categories as follows: geography info (city, region, country,
domain, carrier), page info (page name, start page name, first-hit page name), search info (visit number, referrer,
campaign, keywords, search engine), external info (browser, browser width/height, plugins, language, OS).
We avoid obvious correlated features in the query subcubes, e.g., “search engine” and “keywords” are highly
correlated. For example, some highly correlated variables and their correlations are: start page name & first-hit page
name (0.67), browser & OS (0.40), region & country (0.32), search engine & country (0.27).
We carefully select a subset of coordinates that may follow the near independence assumption to query on.
For instance, we show our experiment results for the following subcubes, along with the number of heavy hitters
recorded: {region, page name , language}, {region, campaign, plugins}, {carrier, first-hit page name, plugins},
{carrier, keywords, OS }.
Since strong independence property is not guaranteed in this real dataset, we increase memory size to 10% of
the data size in order to obtain better estimation for all methods. Recall that the memory used by Sampling and
TwoPassAlg is partially determined by the number of dimensions and therefore it is reasonable to use a relatively
larger memory size.
In this experiment, all three algorithms are able to find most true heavy hitters (see Figure 4), but TwoPassAlg
returns far fewer false positives than the other two methods when γ ∗ is small. In addition, TwoPassAlg reaches zero
false positive for reasonably large γ ∗ . We can see in the ROC curve that TwoPassAlg performs slightly better than
Heuristic and much better than Sampling for small space.
Sampling
Heuristic
TwoPassAlg
1000
False Positives(s)
True Positive(s)
160
140
120
180
Sampling
Heuristic
TwoPassAlg
1200
800
160
True Positive(s)
180
600
400
200
100
0.004
0.006
*
0.008
0.010
120
Sampling
Heuristic
TwoPassAlg
100
0
0.002
140
0.002
0.004
0.006
*
0.008
0.010
0
200
400 600 800 1000 1200
False Positive(s)
Figure 4: Near-independence experiment with Clickstream dataset.
Sampling
TwoPassAlg
16
14
12
10
True Positive(s)
20
False Positive(s)
True Positive(s)
25
Sampling
TwoPassAlg
16
15
10
0.2
*
0.3
0.4
0.5
Sampling
TwoPassAlg
8
0
0.1
12
10
5
8
14
0.1
0.2
*
0.3
0.4
0.5
0
5
10
15
False Positive(s)
20
25
Figure 5: Naive Bayes experiment with Yandex dataset.
5.3
Yandex dataset
Finally, we consider the Yandex dataset [20] which is a web search dataset (with more than 167 millions data points).
Each record in the dataset contains a query ID, the corresponding date, the list of 10 displayed items, and the corresponding click indicators of each displayed item. In the pre-processing step, we extracted 10 subsets from the whole
dataset according to top 10 frequent user queries. The sizes of subsets range from 61, 000 to 454, 000. In each subset,
12
we treat the first 10 search results as variables X1 , .., X10 and “day of query” as the latent variable Z. We conjecture
that the search results X1 , .., X10 are approximately independent conditioned on a given day Z. We observe that
web patterns typically experience heavy weekly seasonality and these search results largely depend on user query
time for some fixed query. We proceed to evaluate TwoPassAlg under the Naive Bayes assumption on this dataset.
We consider 8 subcubes in the form {Xi , Xi+1 , Xi+2 } and deliberately set a smaller memory size for this experiment because the cardinality of this dataset is relatively low. We note that different subsets of data may have
different number of heavy hitters, so we take the average over 10 subsets as the final result.
We report the results in Figure 5. We observe that both Sampling and TwoPassAlg are able to find most true
heavy hitters. However, TwoPassAlg performs significantly better in terms of false positives.
6
Concluding Remarks
Our work demonstrates the power of model-based approach for analyzing high dimensional data that abounds in
digital analytics applications. We exhibit algorithms, with fast query time, that overcome worst case space lower
bound under the classical Naive Bayes assumption. Our approach to subspace heavy hitters opens several directions
for further study. For example,
• Can heavy hitters be detected efficiently under more general models?
• Can these models be learned or fitted over data streams with polylogarithmic space? We believe this is an
algorithmic problem of great interest and will have applications in machine learning beyond the context here.
• We assumed that we observe the latent dimension. Can this be learned from the data stream?
• Can the model-based approach be extended to other problems besides heavy hitters, including clustering, anomaly
detection, geometric problems and others which have been studied in the streaming literature.
References
[1] Ion Androutsopoulos, Georgios Paliouras, Vangelis Karkaletsis, Georgios Sakkis, Constantine D. Spyropoulos,
and Panagiotis Stamatopoulos. Learning to filter spam e-mail: A comparison of a naive bayesian and a memorybased approach. CoRR, cs.CL/0009009, 2000.
[2] Vladimir Braverman, Kai-Min Chung, Zhenming Liu, Michael Mitzenmacher, and Rafail Ostrovsky. AMS without 4-wise independence on product domains. In STACS, volume 5 of LIPIcs, pages 119–130. Schloss Dagstuhl
- Leibniz-Zentrum fuer Informatik, 2010.
[3] Vladimir Braverman and Rafail Ostrovsky. Measuring independence of datasets. In STOC, pages 271–280. ACM,
2010.
[4] Graham Cormode.
Finding frequent items in data streams.
http://dmac.rutgers.edu/
Workshops/WGUnifyingTheory/Slides/cormode.pdf, 2008. DIMACS Workshop.
[5] Graham Cormode and S. Muthukrishnan. An improved data stream summary: The count-min sketch and its
applications. In LATIN, volume 2976 of Lecture Notes in Computer Science, pages 29–38. Springer, 2004.
[6] Constantinos Daskalakis, Nishanth Dikkala, and Gautam Kamath. Testing ising models. CoRR, abs/1612.03147,
2016.
[7] Marco F. Duarte, Volkan Cevher, and Richard G. Baraniuk. Model-based compressive sensing for signal ensembles. In Communication, Control, and Computing, 2009. Allerton 2009. 47th Annual Allerton Conference on, pages
244–250. IEEE, 2009.
[8] Amit Goyal, Hal Daumé III, and Suresh Venkatasubramanian. Streaming for large scale NLP: language modeling.
In Human Language Technologies: Conference of the North American Chapter of the Association of Computational
Linguistics, Proceedings, May 31 - June 5, 2009, Boulder, Colorado, USA, pages 512–520. The Association for
Computational Linguistics, 2009.
13
[9] Matthew Hamilton, Rhonda Chaytor, and Todd Wareham. The parameterized complexity of enumerating frequent itemsets. In IWPEC, volume 4169 of Lecture Notes in Computer Science, pages 227–238. Springer, 2006.
[10] Piotr Indyk and Andrew McGregor. Declaring independence via the sketching of sketches. In SODA, pages
737–745. SIAM, 2008.
[11] Branislav Kveton, Hung Hai Bui, Mohammad Ghavamzadeh, Georgios Theocharous, S. Muthukrishnan, and
Siqi Sun. Graphical model sketch. In ECML/PKDD (1), volume 9851 of Lecture Notes in Computer Science, pages
81–97. Springer, 2016.
[12] David D. Lewis, Yiming Yang, Tony G. Rose, and Fan Li. RCV1: A new benchmark collection for text categorization research. Journal of Machine Learning Research, 5:361–397, 2004.
[13] Edo Liberty, Michael Mitzenmacher, Justin Thaler, and Jonathan Ullman. Space lower bounds for itemset frequency sketches. In PODS, pages 441–454. ACM, 2016.
[14] Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schütze. Introduction to Information Retrieval.
Cambridge University Press, 2008.
[15] Andrew McGregor and Hoa T. Vu. Evaluating bayesian networks via data streams. In COCOON, volume 9198
of Lecture Notes in Computer Science, pages 731–743. Springer, 2015.
[16] Jayadev Misra and David Gries. Finding repeated elements. Sci. Comput. Program., 2(2):143–152, 1982.
[17] Stuart J. Russell and Peter Norvig. Artificial Intelligence - A Modern Approach. Pearson Education, 2010.
[18] Jeffrey Scott Vitter. Random sampling with a reservoir. ACM Trans. Math. Softw., 11(1):37–57, 1985.
[19] David P. Woodruff. New algorithms for heavy hitters in data streams (invited talk). In ICDT, volume 48 of
LIPIcs, pages 4:1–4:12. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2016.
[20] Yandex personalized web search challenge.
challenge, 2013.
https://www.kaggle.com/c/yandex-personalized-web-search-
[21] Guizhen Yang. The complexity of mining maximal frequent itemsets and maximal frequent patterns. In KDD,
pages 344–353. ACM, 2004.
14
| 8cs.DS
|
arXiv:1610.03729v1 [cs.SY] 12 Oct 2016
Scheduling of Event-Triggered Networked Control
Systems using Timed Game Automata
Dieky Adzkiya1 and Manuel Mazo, Jr.2
1
Department of Mathematics, Institut Teknologi Sepuluh
Nopember, Indonesia. Part of this work was done in the Delft
Center for Systems and Control, TU Delft, The Netherlands.
(e-mail: [email protected])
2
Delft Center for Systems and Control, TU Delft - Delft University
of Technology, The Netherlands. (e-mail: [email protected]).
Abstract
We discuss the scheduling of a set of networked control systems implemented over a shared communication network. Each control loop is
described by a linear-time-invariant (LTI) system with an event-triggered
implementation. We assume the network can be used by at most one
control loop at any time instant and after each controller update, a predefined channel occupancy time elapses before the network is available. In
our framework we offer the scheduler two options to avoid conflicts: using
the event-triggering mechanism, where the scheduler can choose the triggering coefficient; or forcing controller updates at an earlier pre-defined
time. Our objective is avoiding communication conflict while guaranteeing stability of all control loops. We formulate the original scheduling
problem as a control synthesis problem over a network of timed game automata (NTGA) with a safety objective. The NTGA is obtained by taking
the parallel composition of the timed game automata (TGA) associated
with the network and with all control loops. The construction of TGA
associated with control loops leverages recent results on the abstraction of
timing models of event-triggered LTI systems. In our problem, the safety
objective is to avoid that update requests from a control loop happen
while the network is in use by another task. We showcase the results in
some examples.
1
Introduction
Networked control systems (NCSs) are spatially distributed systems in which
the communication between sensors, actuators and controllers occurs through a
shared band-limited digital communication network, as shown in Fig. 1. Such
1
Sensor 1
Plant 1
Actuator 1
...
Sensor N
Plant N
Actuator N
Network
Scheduler
Controller N
Controller 1
Figure 1: Topology of a set of N networked control systems.
structures bring many advantages, for instance reduced wiring and maintenance
costs as well as an increased flexibility and reconfigurability. NCSs occur in
numerous applications, including power systems [1], aircrafts and automobiles
[2], and process control [3].
Notice that NCSs are implemented over shared communication resources,
most often over digital channels. The impact of such communication infrastructures on control systems has been studied in the last decade [4, 5, 6, 7].
In particular the applicability of wireless communications in NCSs has been
discussed in [8, 9, 10] among others. The delays introduced by these shared
resources on the feedback loops are critical to guarantee stability and performance. Furthermore, when several control loops are implemented over a shared
communication channel, bandwidth becomes a scarce resource, the usage of
which needs to be minimized by each controller.
For these reasons, the traditional time-triggered controller implementations,
i.e. based on periodic sampling, are not suitable anymore. With the objective
of minimizing the bandwidth usage, event-based approaches resulting in aperiodic controller updates have been proposed in [11, 12, 13, 14, 15, 16]. These
aperiodic paradigms introduce a new challenge in the design: the scheduling of
transmissions.
A communication network has finitely many channels. The number of channels represents the maximum number of messages that can be sent simultaneously over the network. If the number of channels equals the number of control
loops, we do not need any scheduler because every control loop has its own
communication channel. However in practice, the number of channels is smaller
than the number of control loops. Thus we need a scheduler to decide which
control loop has access to the network at any time instant while guaranteeing
stability of all control loops. Additionally the scheduler can optimize a certain
combination of control performance and bandwidth usage. In the context of
periodic traffic sources, or aperiodic with known deadlines a-priori, there are
well-studied scheduling techniques that are capable of guaranteeing certain delay bounds for the traffic [17, 18, 19]. In the event-based context, i.e. sporadic
traffic with unknown deadlines, the problem has been less studied and becomes
more challenging [20, 21, 22, 23]. In [21, 22, 23], the authors propose a joint design (codesign) of a control law and a scheduling law for several types of NCSs.
Although the co-design strategy can improve the control performance signifi-
2
cantly, if a new control loop is introduced to the NCSs, the whole co-design
procedure has to be executed again, which can be extremely time consuming.
In order to mitigate this issue, we propose an approach that separates the design
of controllers and schedulers.
In this paper, we design a scheduler for a set of NCSs over a shared communication network (cf. Fig. 1). Our objective is avoiding communication conflict
while guaranteeing stability of all NCSs. Each NCS is a linear-time-invariant
(LTI) system. The controllers are implemented in an event-triggered fashion
where the delays between reading the state and updating the actuators are ignored [12]. With respect to the shared communication network, we assume it
has a single communication channel. Furthermore after each controller update,
a pre-defined channel occupancy time elapses before the network is available.
We consider schedulers that after each transmission of measurements, decide
the policy for the next update. These policies can be to either let the next
update be decided based on a triggering mechanism (to be chosen among a
set of them guaranteeing different performances) or forced to be at an earlier
pre-defined time. If we do not allow the scheduler to force earlier controller
updates, the bandwidth usage is decreased at the cost of worse control performance (or slower convergence). On the other hand if we allow the scheduler
to force earlier controller updates, we obtain a better control performance (or
faster convergence) at the cost of increased bandwidth usage. In this case, nothing prevents the scheduler from always using earlier updates and never use the
event-triggering mechanism. This may result in an undesired over-use of the
communication channel, and could be prevented by introducing costs to the
model, which would result in priced timed game automata (PTGA). Unfortunately to the best of our knowledge, no results are available in the literature
allowing the synthesis of strategies over PTGA with safety objectives. Only the
synthesis of strategies over PTGA with a reachability objective are available
[24, 25]. Thus we propose an alternative approach to prevent the undesired
schedule by limiting the consecutive earlier updates.
The scheduling problem can be formulated as a timed safety game: given
a model and a set of bad states, we seek to construct a strategy such that
the model supervised by the strategy constantly avoids the bad states. In our
problem, the safety problem at hand is to avoid that update requests from a
control loop happen while the network is in use by another task. We focus
our attention to the design of a scheduler by leveraging techniques originally
developed for network of timed game automata (NTGA) [26]. An NTGA is the
parallel composition of timed game automata (TGA), which are timed safety
automata (TSA) in with the set of actions is partitioned into controllable and
uncontrollable actions. We choose NTGA modeling framework because of the
following two reasons. First of all, it allows us to extend the methods in [27].
The authors of [27] discuss formal abstraction of the timing behavior of LTI systems with event-triggered implementation as TSA. The second reason is that
the solution of timed safety game over NTGA can be computed by using backward algorithms [28, 26] or on-the-fly algorithms [29]. Moreover the algorithms
have been implemented in some freely available software tools [30, 29]. The
3
procedure to generate the scheduler (or the strategy) is as follows. First, we
construct an NTGA from a set of NCSs. The NTGA is obtained by taking
the parallel composition of the TGA associated with the network and with all
control loops. Then we characterize the bad states, i.e. states corresponding
to a communication conflict. Finally the scheduler is defined as the solution of
timed safety game over the NTGA.
Timed automata (TA) [31] are a general modeling framework for a wide
range of real-time systems, such as in web services [32], audio/video protocols
[33], bounded retransmission protocols [34], collision avoidance protocols [35,
36] and commercial field bus protocols [37]. Timed safety automata (TSA)
[38] are a simplified version of TA. In order to enforce progress properties,
TSA use local invariant conditions whereas TA use Büchi or Muller accepting
conditions. The scheduling problem over TA and its variants has been already
studied in the literature e.g. applied to the scheduling of a steel plant, a job shop
and a task graph in [39], [40], and [41], respectively. Furthermore the optimal
scheduling of a production w.r.t. a predefined cost for a finite time horizon has
been investigated in [42, 43]. In this case, the models are TA with weights (or
costs) on both locations and edges, so called priced timed automata in [44] and
weighted timed automata in [45]. Finally the optimal scheduling for infinite
time horizon is discussed in [46].
The rest of this manuscript is structured as follows. Section 2 recalls some
modeling frameworks and preliminaries. Section 3 proposes a procedure to
synthesize a conflict-free scheduler. Two experimental results are discussed in
Section 4. The first experiment allows the scheduler to force earlier controller
updates where the number of consecutive earlier updates is limited to 4. The
second experiment does not allow the scheduler to force earlier controller updates. In this experiment, the scheduler can choose the triggering coefficient
among three choices. Finally the conclusions and possible future research directions are summarized in Section 5.
2
2.1
Models and Preliminaries
Timed Automata
A timed automaton (TA) [31] is a finite automaton (namely, a directed graph
containing finitely many nodes and finitely many labeled edges) extended with
real-valued variables, which is usually employed to model real-time systems.
The real-valued variables model logical clocks, that are initialized to zero when
the system is started and thereafter increase synchronously at the same rate.
We shall refer to these variables as simply “clocks”. Clock constraints are used
to restrict the behavior of the automaton. An edge transition can be taken when
the edge is enabled. Edges are enabled if the values of the clocks satisfy the
guard conditions associated with the edge. Additionally, some clocks may be
reset to zero when an edge is taken. Originally, Büchi and Muller accepting conditions are used to enforce progress properties [31]. A simplified version called
4
timed safety automata [38] uses local invariant conditions to specify progress
properties. In this work, we focus on timed safety automata and refer them as
timed automata for simplicity.
We define C as the set of finitely many clocks, Act as the set of finitely many
actions and N0 as the set of natural numbers including zero {0, 1, . . . }. A clock
constraint is a conjunctive formula of atomic constraints of the form x ⊲⊳ n or
x − y ⊲⊳ n for x, y ∈ C, ⊲⊳∈ {≤, <, =, >, ≥} and n ∈ N0 . Clock constraints will
be used as guards on edges and location invariants. We use B(C) to denote the
set of clock constraints.
Definition 1 (Timed Automaton). A timed automaton TA is a sextuple
(L, ℓ0 , Act, C, E, Inv) where
• L is a set of finitely many locations (or nodes);
• ℓ0 ∈ L is an initial location;
• Act is a set of finitely many actions;
• C is a set of finitely many real-valued clocks;
• E ⊆ L × B(C) × Act × 2C × L is a set of edges;
• Inv : L → B(C) assigns invariants to locations.1
Location invariants are restricted to constraints that are downwards closed, in
the form: c ≤ n or c < n where c is a clock and n ∈ N0 .
g,a,r
Sometimes we write ℓ ✲ ℓ′ when (ℓ, g, a, r, ℓ′ ) ∈ E. Furthermore we write
ℓ ✲ ℓ′ to denote the existence of an edge from ℓ to ℓ′ with arbitrary labels.
The semantics of a TA are defined as a transition system where a state
consists of the current location and the current value of clocks. There are
two types of transitions between states depending on whether the automaton:
delays for some time (a delay transition), or takes an enabled edge (a discrete
transition).
To keep track of clock values, we use functions known as clock assignments
u : C → R≥0 and we employ u g (u satisfies g) to denote that the clock values
of u satisfy the guard g. For d ∈ R≥0 , let u + d denote the clock assignment
that maps all c ∈ C to u(c) + d. For a set of clocks c ⊆ C, let u[c] denote the
clock assignment that maps all clocks in c to 0 and agrees with u for the rest
of clocks in C \ c.
Definition 2 (Operational Semantics). The semantics of a timed automaton is
a transition system (also known as a timed transition system) in which states
are pairs of location ℓ and clock assignment u, and transitions are defined by
the rules:
d
✲ (ℓ, u + d) if u Inv(ℓ) and (u + d) Inv(ℓ)
• Delay transition: (ℓ, u)
TS
for a non-negative real number d ∈ R≥0 ;
1 Recall
that 2C denotes the power set of C.
5
• Discrete transition: (ℓ, u)
a
✲ (ℓ′ , u′ ) if ℓ
TS
g,a,r
✲ ℓ′ , u g, u′ = u[r] and
u′ Inv(ℓ′ ).
A run of a timed automaton is a sequence of alternating delay and discrete
transitions in the transition system.
We denote by Runs(TA) the set of runs of timed automaton TA starting from
the initial state (ℓ0 , u0 ) where u0 is a clock assignment that maps all c ∈ C to 0.
Additionally, if ρ is a finite run, the last state of the run is denoted by last (ρ).
The set of actions Act (cf. Definition 1) is assumed to consists of symbols
for input actions a?, output actions a! and internal actions ∗. Synchronous
communication between different TA is done by hand-shake synchronization
using input and output actions.
To model concurrent systems, several TAs can be extended with parallel
composition that takes into account the synchronous communication. Parallel
composition of TAs is also called network of timed automata (NTA). Essentially
the parallel composition of a set of TAs is the product of the TAs. Building
the product timed automaton is an entirely syntactical but computationally
expensive operation. The reader is referred to [47, Sec. 5] for an example on the
composition of two TAs.
The semantics of an NTA are defined as a transition system where a state
consists of a vector of current locations and the current value of clocks in all
TAs [48].
2.2
Timed Game Automata
A timed game automaton is a timed automaton in which the set of actions
is partitioned into controllable and uncontrollable actions. The former are actions that can be triggered by the controller, whereas the latter only by the
environment/opponent.
Definition 3 (Timed Game Automaton). A timed game automaton TGA is a
septuple (L, ℓ0 , Actc , Actu , C, E, Inv) where
• (L, ℓ0 , Actc ∪ Actu , C, E, Inv) is a timed automaton;
• Actc is a set of controllable actions;
• Actu is a set of uncontrollable actions;
• Actc ∩ Actu = ∅.
Similar to TA, TGA can also be extended with parallel composition (essentially the synchronized cartesian product of TGA). The parallel composition of
TGAs is called a “network of timed game automata” (NTGA) which is formally
defined as:
6
Definition 4 (Parallel Composition). Let TGAi = (Li , ℓi0 , Actic , Actiu , C i , E i , Invi )
be a timed game automaton for i ∈ {1, . . . , n}. The parallel composition of
TGA1 , . . . , TGAn denoted by TGA1 | · · · | TGAn is a timed game automaton
TGA = (L, ℓ0 , Actc , Actu , C, E, Inv) where
• L = L1 × · · · × Ln ;
• ℓ0 = (ℓ10 , . . . , ℓn0 );
Sn
• Actc = {∗} ∪ i=1 {a ∈ Actic | a is an internal action};
Sn
• Actu = {⊛} ∪ i=1 {a ∈ Actiu | a is an internal action};
• C = C1 ∪ · · · ∪ Cn;
• E is defined according to the following two rules:
– a TA makes a move on its own via its internal action: the edge is
controllable iff the internal action is controllable;
– two TAs move simultaneously via a synchronizing action: the edge is
controllable iff both input and output actions are controllable (i.e. the
environment has priority over the controller);
• Inv((ℓ1 , . . . , ℓn )) = Inv1 (ℓ1 ) ∧ · · · ∧ Invn (ℓn ).
In the parallel composition of TGAs, a pair of input and output actions is
denoted as a single action. Thus the sets Actc and Actu do not contain any input
and output actions. A synchronizing action should be defined as an element
of Actc if it is controllable and an element of Actu if it is not controllable. In
Definition 4, let us remark that both Actc and Actu do not contain synchronizing
actions for simplicity. Any controllable synchronizing action is denoted by ∗,
whereas any uncontrollable synchronizing action is denoted by ⊛.
Given an NTGA, we are interested in solving the following safety objective:
is it possible to find a strategy for the triggering of controllable actions guaranteeing that a set of pre-specified bad states are never reached regardless of what
and when uncontrollable actions take place? More formally given an NTGA and
a set of bad states A, we seek to construct a strategy f such that the NTGA
supervised by f constantly avoids A.
A strategy [26] is a function that during the course of a game constantly
gives information about what the controller should do in order to win the game.
At any given situation, the strategy could suggest the controller to either “take
a particular controllable action” or “do nothing at this point in time”, i.e. delay,
which will be denoted by the symbol (controllable action) λ.
Definition 5 (Strategy [29, Definition 3]). Let TGA = (L, ℓ0 , Actc , Actu , C, E, Inv)
be a timed game automaton. We define TA = (L, ℓ0 , Actc ∪ Actu , C, E, Inv) as
the timed automaton derived from the timed game automaton. A strategy f over
TGA is a partial function from Runs(TA) to Actc ∪ {λ} s.t. for every finite run
f (ρ)
ρ, if f (ρ) ∈ Actc then last (ρ) ✲ (ℓ′ , u′ ) for some (ℓ′ , u′ ).
TS
7
A strategy f over TGA is called state-based or memoryless whenever last (ρ) =
last (ρ′ ) implies f (ρ) = f (ρ′ ), for each ρ, ρ′ ∈ Runs(TA). The restricted behavior
of an NTGA controlled with some strategy f is defined by the notion of outcome
[28].
A strategy f is winning from a state if all maximal runs [29, p. 70] in the
outcome originated from that state are winning. A state is winning if there exists
a winning strategy f from that state. The winning states can be computed by
using backward algorithms [28, 26] or on-the-fly algorithms [29]. Software tools
are also available that solve safety control problems, e.g. the implementation
from Verimag [30] or UPPAAL-Tiga [29], which implement the backward and
on-the-fly algorithms respectively.
2.3
Event Triggered Control Systems
We consider linear-time-invariant (LTI) systems of the form
˙ = Aξ(t) + Bυ(t),
ξ(t)
ξ(t) ∈ Rn ,
υ(t) ∈ Rm
(1)
where A and B are matrices of appropriate dimensions. We assume the existence
of linear state-feedback laws υ(t) = Kξ(t) rendering the closed-loop system
globally asymptotically stable, where K is a matrix of appropriate dimensions.
Assume a sample-and-hold implementation of the control law is in place
keeping the input signal constant between update times, i.e.
t ∈ [tk , tk+1 [,
υ(t) = Kξ(tk ),
(2)
where t0 , t1 , . . . is a divergent sequence of update times. For simplicity of presentation, we ignore the presence of delays between reading the state and updating
the actuators. The interested reader is referred to [12] for more details, including
accounting for delays.
In event-triggered implementations, the sequence of update times is decided
on run-time based on the state of the plant [12]. Let ξ(t) represent the solution of
(1)-(2). We define an auxiliary variable e(t) representing the difference between
the sampled state ξ(tk ) and the current state ξ(t) of the system:
e(t) = ξ(tk ) − ξ(t),
t ∈ [tk , tk+1 [,
k ∈ N0 .
The event-triggering approach in [12], proposes the following sampling triggering
law:
tk+1 = min{t | t > tk and |e(t)|2 ≥ σ|ξ(t)|2 },
(3)
where σ ∈]0, σ̄[⊂ R+ is the triggering coefficient, which establishes a trade off
between quality of control (convergence rate to the equilibrium) and the amount
of transmissions triggered. The inter-sample time of the state x, denoted by
τσ (x), is defined as the time between consecutive updates when the sampled
state is x:
τσ (x) = min{t | |e(t)|2 ≥ σ|ξ(t)|2 and ξ(0) = x}.
8
(4)
2.4
Abstraction of Event Triggered Control Systems as
Timed Automata
In [27], the authors propose an approach to characterize the sampling behavior
of LTI systems with event-triggered implementation as TAs. The approach
abstracts the spatial and temporal dependencies of the original system. The
following definitions summarize the approach.
Definition 6 (Flow Pipe). The set of reachable states or the flow pipe at the
time interval [t1 , t2 ] from a set of initial states X0 is denoted by
[
{ξ(t) | ξ(0) ∈ X0 }.
X[t1 ,t2 ] (X0 ) =
t∈[t1 ,t2 ]
Definition 7 ([27]). A timed automaton abstracting the triggering timing behavior of system (1)-(2) with triggering coefficient σ is given by
TAσ = (Lσ , ℓσ0 , Actσ , C σ , E σ , Invσ ) where
• Lσ = {Rσ1 , . . . , Rσq };
• ℓσ0 = Rσs such that ξ(0) ∈ Rσs ;
• Actσ = {∗};
• C σ = {c};
• (Rσs , τ σs ≤ c ≤ τ̄sσ , ∗, {c}, Rσt ) ∈ E σ if X[τ σs ,τ̄sσ ] (Rσs ) ∩ Rσt 6= ∅;
¯
¯
• Invσ (Rσs ) = {c | 0 ≤ c ≤ τ̄sσ } for all s ∈ {1, . . . , q}.
Each location Rσs of this TA, is associated with a set of possible states x
of the system (1). We abuse slightly notation denoting both the location and
the associated region with the same symbol Rσs . The suggestion in [27] is to
partition the state space of the control system in conic regions pointed at the
origin, each of which would be associated to a location of the TA. The TA
has one clock variable c that represents the time elapsed since the last update.
According to [12], given a fixed sampled state, the inter-sample time is uniquely
defined, i.e. it is deterministic. In general, when the sampled state is different,
the inter-sample time is also different. The notation τ σs and τ̄sσ represents the
¯
lower and upper bounds of the inter-sample time for sampled states in Rσs .
In [27] it is shown formally that the TA abstracts the timing behavior of the
event-triggered system, implying that:
∀s ∈ {1, . . . , q}, ∀x ∈ Rσs : τσ (x) ∈ [τ σs , τ̄sσ ].
¯
Remark 1. In principle it could happen that τσ (x) = ∞ for some states x
of the system. In practice, one would always impose a maximum time between
transmissions to maintain a minimum level of feedback. This practical solution
is suggested in [27] to guarantee having always τ̄sσ < ∞, as otherwise the TA
model would become useless for scheduling purposes.
9
τ σ1 ≤ c ≤ τ̄1σ
¯ c := 0
τ σ1 ≤ c ≤ τ̄1σ
¯ c := 0
Rσ1
0 ≤ c ≤ τ̄1σ
τ σ2 ≤ c ≤ τ̄2σ
¯ c := 0
Rσ2
τ σ2 ≤ c ≤ τ̄2σ
¯ c := 0
0 ≤ c ≤ τ̄2σ
Figure 2: A timed automaton modeling a control loop with event-triggered implementation. The unsourced arrow indicates the initial location. The internal
action ∗ is omitted in the figure.
From Definition 7 it is also trivial to see that if ξ(t) = x ∈ Rσs then ξ(t +
τσ (x)) ∈ Rσj , with Rσj being one of the end locations in the set of edges with
starting location Rσs . The outgoing edges of Rσs are enabled if the time elapsed
since the last update is between τ σs and τ̄sσ . Only one action denoted by ∗ is
¯
present in this model, and since taking any edge is interpreted as updating the
input value, all edges are labeled with action ∗ and reset the clock variable.
Note that the system may remain in location Rσs for at most τ̄sσ time units,
as a triggering event is guaranteed to happen before that instant. A graphical
representation of a simple TA of the form of those from Definition 7 is shown
in Fig. 2.
3
Scheduling of Event-Triggered Control Systems
Consider a set of event-triggered networked control systems (NCSs) sharing
a common communication channel (cf. Fig. 1). Each control loop consists of
a sensor, a plant, an actuator, and a controller, interconnected through the
shared communication network. Assume that the network can be used by at
most one control loop at any time instant. If several control loops request access
to the channel while the network is in use a conflict arises, and at most one
control loop will be chosen nondeterministically to access the network. While
in time-triggered control systems these type of problems can be prevented by
appropriate scheduling, when one or several control-loops are event-triggered
a-priori scheduling is a much more challenging task because of the unknown
update times.
In this section, we propose an approach based on NTGA to avoid such conflicts. We consider schedulers that after each update of a control loop (transmission of measurements, computation of control and transmission of actuation
signal to actuators) decide whether the next update time of each control loop
should:
• be based on a triggering mechanism selected from a set of finitely many
triggering coefficients {σ1 , . . . , σp }; or
10
• forced to be at a pre-defined time (earlier than the minimum expected
inter-sample time).
We synthesize scheduling strategies by: constructing an NTGA from a set of
NCSs (cf. Section 3.1), characterizing the bad states, i.e. states corresponding to
a communication conflict (cf. Section 3.2), and finally synthesizing a supervising
strategy ensuring that the NTGA avoids the bad states.
3.1
Model: Network of Timed Game Automata
In what follows, we describe the procedure employed to construct an NTGA
from a given set of event-based NCSs. We start constructing a TGA associated
with the shared communication network. Next, for each event-triggered control
loop, we generate a TGA as a modification of the TA described in Section 2.4.
Finally, the NTGA is obtained by taking the parallel composition of the TGA
associated with the network and with all control loops.
3.1.1
Communication Network
Denote the TGA corresponding to the shared communication network by TGAnet
(cf. Definition 8). TGAnet has three locations Idle, InUse and Bad , where the
initial location is Idle (cf. Fig. 3). The location Idle represents the network
being available, InUse represents the network being used by a control loop and
Bad represents a conflict occured. The active location changes from Idle to
InUse when a control loop requests access to the channel to perform an update,
which also forces the reset of the clock variable c. The channel is occupied
for ∆ time units before the network is freed again to service the control tasks.
During this time, the active location is InUse, and after that time the active location changes to Idle. When the active location is InUse and another control
loop requests access then the active location changes to Bad . Once the network enters the location Bad , the network cannot leave the location, i.e. Bad
is an absorbing location. Notice that this is a somewhat conservative model,
as we consider every control loop occupies the channel the whole time ∆. One
could trivially adjust this simple model, and the subsequent work, to associate
different occupancy times to different control loops.
Definition 8. Let ∆ represent the maximum channel occupancy time, a timed
game automaton associated with the communication network is given by TGAnet =
net
net
net
(Lnet , ℓnet
, E net , Invnet ) where
0 , Actc , Actu , C
• Lnet = {Idle, InUse, Bad };
• ℓnet
= Idle;
0
• Actnet
= {∗};
c
• Actnet
u = {up?};
• C net = {c};
11
up?
c := 0
up?
up?
Idle
InUse
c == ∆
Bad
0≤c≤∆
Figure 3: A timed game automaton modeling the shared communication network. The solid and dashed arrows represent controllable and uncontrollable
edges, respectively.
• E net = {(Idle, true, up?, {c}, InUse), (InUse, c = ∆, ∗, ∅, Idle),
(InUse, true, up?, ∅, Bad ), (Bad , true, up?, ∅, Bad )};
• Invnet (InUse) = {c | 0 ≤ c ≤ ∆},
Invnet (Idle) = {c | c ≥ 0}, Invnet (Bad ) = {c | c ≥ 0}.
The guard true represents a condition that is always satisfied, for example c ≥ 0.
3.1.2
Control Loops
Given a control loop, we construct the timed game automata TGAcl allowing
a supervisor (scheduler) to either: force earlier controller updates than those
dictated by the event-triggering mechanism, or choose a triggering coefficient
for the event-triggering mechanism.
σ
Definition 9. Consider a set of timed automata TAσj = (Lσj , ℓ0 j , Actσj , C σj , E σj , Invσj )
generated from an event-triggered control loop with triggering coefficient σj ∈
σ
]0, σ̄[ for j ∈ {1, . . . , p} and assume that Rsσ1 = · · · = Rs p for all s ∈ {1, . . . , q}.
Consider also a set of earlier update time parameters {d1 , d¯1 , . . . , dq , d¯q }, such
¯
¯
that
σj
¯
∀s ∈ {1, . . . , q} ∃ j ∈ {1, . . . , p} : ds ≤ τ s .
¯
Then, the timed game automata TGAcl is given by
cl
cl
cl
cl
cl
TGAcl = (Lcl , ℓcl
0 , Actc , Actu , C , E , Inv ) where
Sq
Sp
• Lcl = j=1 Lσj ∪ s=1 {Rs , Ear s };
σ1
• ℓcl
0 = Rs such that ξ(0) ∈ Rs ;
Sp
σ1
• Actcl
∪ j=1 {acl
c = Act
j };
• Actcl
u = {up!};
• C cl = C σ1 ;
Sq Sp
Sq S
σj
• E cl = s=1 t∈Es {(Ear s , c = 0, up!, ∅, Rt )}∪ s=1 j=1 {(Rs , c = 0, acl
j , ∅, Rs ),
S
S
S
σ
(Rs j , ds ≤ c ≤ d¯s , ∗, {c}, Ear s )} ∪ qs=1 pj=1 {t|(Rs →Rt )∈E σj }
σ¯ σ
σ
{(Rs j , τ s j ≤ c ≤ τ̄s j , up!, {c}, Rt )};
¯
12
Rσ1
R1
Ear2
Ear1
R2
Rσ2
Figure 4: A timed game automaton modeling a control loop with an eventtriggered implementation. The labels over edges and location invariants are not
shown for simplicity.
σ
σ
• Invcl (Rs j ) = {c | c ≤ τ̄s j }, Invcl (Rs ) = {c | c = 0},
Invcl (Ear s ) = {c | c = 0}.
In model just introduced, we use separate locations associated to each triggering coefficient and introduce the additional locations Rs and Ear s for s ∈
σ
{1, . . . , q}. Both the locations Rs and Rs j represent that the sampled state is
in Rσs 1 . In location Rs the scheduler has not chosen the triggering coefficient,
σ
whereas in location Rs j the scheduler has chosen triggering coefficient σj . Since
σ
the scheduler can choose the triggering coefficient, the edges from Rs to Rs j are
labeled with the controllable action aj . After choosing the triggering coefficient,
the scheduler is allowed to either: force earlier controller updates, or use the
event-triggering mechanism (based on the chosen triggering coefficient).
If the scheduler decides to use the event-triggering mechanism, while staying
σ
in location Rs j , the strategy is “do nothing”. This ensures that the outgoing
σ
σ
edge to Ear s is not taken. When the value of c is between τ s j and τ̄s j , the eventσj
¯
triggering mechanism is activated. In this case, the edges from Rs to Rt labeled
with the uncontrollable action up! are enabled. Recall that the scheduler cannot
choose the exact update time when using the event-triggering mechanism. This
also implies that the scheduler cannot choose the region containing the next
sampled state.
If the scheduler decides to force earlier controller updates, the scheduler will
take the edge to Ear s when that edge is enabled. In this case, the scheduler is
σ
able to choose the exact update time. Thus, the edges from Rs j to Ear s are
labeled with the controllable action ∗. In location Ear s , the time cannot elapse
and one of the outgoing edges has to be taken immediately. Since the scheduler
cannot choose the region containing the next sampled state, the outgoing edges
of Ear s are labeled with the uncontrollable action up!. The outgoing edges
are defined as follows: there exists an edge from Ear s to Rt if t ∈ Es := {t |
X[ds ,d̄s ] (Rσs 1 ) ∩ Rσt 1 6= ∅}. A graphical representation of a TGA generated by
¯
Definition
9 is shown in Fig. 4.
In this subsection, we assume the initial conditions of the LTI system are a
subset of a region. If the initial conditions are intersected with a set of regions
Rinit ⊆ {Rσ1 1 , . . . , Rσq 1 }, we can modify the TGA generated by Definition 9 as
follows. Introduce a new location called R0 with invariant {c = 0} and define
13
Rσ1
R1
Ear2
R0
Ear1
R2
Rσ2
Figure 5: A timed game automaton modeling a control loop where the initial
states are intersected with both conic regions. The labels over edges and location
invariants are not shown for simplicity.
R0 as the initial location. Then, define edges from R0 to every location in Rinit
with guard c = 0, action ⊛ and without resetting the clock. Finally, action ⊛ is
defined as an uncontrollable action. In the above modification, the environment
has to choose one of the locations corresponding to initial conditions when the
system is started. A graphical representation of the TGA representing this
situation is depicted in Fig. 5.
Proposition 1. Switching between different triggering coefficients or triggering
earlier does not hinder stability.
Proof. Consider Lyapunov function V : Rn → R≥0 satisfying V̇ (ξ(t)) ≤ −λc V (ξ(t)),
with λc > 0 for the system (1) with continuous feedback u(t) = Kξ(t). It has
been shown in [12] that selecting a triggering coefficient 0 < σ < σ̄, with σ̄ an appropriate constant depending on the LTI dynamics and the state-feedback gain,
the event-triggered controller implementation (2)-(3) satisfies ∀t : V̇ (ξ(t)) ≤
−λe (σ)V (ξ(t)), with λc > λe (σ1 ) > λe (σ2 ) > 0 for 0 < σ1 < σ2 < σ̄.
In fact, the triggering mechanism guarantees that V̇ (ξ(t)) ≤ −λe (σ)V (ξ(t))
in the interval t ∈ [tk , tk + τσ (x)], for all x = ξ(tk ). Since σj < σ̄ for all
j ∈ {1, . . . , p}, switching between different triggering coefficients guarantees
that ∀t : V̇ (ξ(t)) ≤ −λe σ̃V (ξ(t)), with σ̃ := maxj∈{1,...,p} σj .
Finally, if the system is forced to employ earlier triggering the assumption:
σ
∀s ∈ {1, . . . , q} ∃ j ∈ {1, . . . , p} : d¯s ≤ τ s j guarantees that the update occurs
¯
in the interval [tk , tk + τσ̃ (x)], and thus ∀t : V̇ (ξ(t)) ≤ −λe σ̃V (ξ(t)), which
concludes the proof.
As it was mentioned in the beginning of Section 3.1, the NTGA associated
with a set of NCSs, denoted by TGANCSs , is obtained by taking the parallel
composition of the TGA associated with the network and with all control loops.
In other words, TGANCSs := TGAnet | TGAcl1 | · · · | TGAclN where TGAcli , i ∈
{1, . . . , N }, represents the TGA associated with the i-th control loop. The state
of TGANCSs is described by a (2N + 2)-tuple (ℓnet , ℓ1 , . . . , ℓN , unet , u1 , . . . , uN )
where ℓnet is the location of TGAnet , ℓi is the location of TGAcli , unet is the
14
clock assignment of TGAnet and ui is the clock assignment of TGAcli , for i ∈
{1, . . . , N }.
3.2
Specification: Safety
We are interested in finding a strategy such that the trajectories of the NTGA
never enter the states corresponding to a conflict. Recall that a conflict corresponds to the following situation: a control loop is requesting updates when the
communication network is busy. In our NTGA model, conflicts are captured by
the active location of TGAnet becoming Bad . Thus the set of states we aim at
avoiding A contains all states such that the location of TGAnet is Bad , i.e.
A = {(ℓnet , ℓ1 , . . . , ℓN , unet , u1 , . . . , uN ) | ℓnet = Bad }.
3.3
Limiting the Consecutive Earlier Updates
If we allow the scheduler to force earlier controller updates, nothing prevents the
scheduler from always using such type of updates and never employ the eventtriggering mechanism. In this section, we discuss an approach to prevent the
undesired schedule by limiting the consecutive earlier updates. In this approach,
once a pre-specified limit has been reached, the scheduler is forced to use an
event-triggering mechanism.
For this purpose we employ global integer variables, which are an extended
feature of UPPAAL-Tiga modeling language to ease the modeling task but not
part of the standard definition of TGA (cf. Definition 3). We define a global
integer constant earMax representing the maximum consecutive earlier updates,
and a global integer variable earNum to be used as a counter of consecutive
earlier updates. A variable is global if it can be accessed by all TGAs. Finally,
the resulting TGA is defined as follows.
σ
Definition 10. Consider a set of timed automata TAσj = (Lσj , ℓ0 j , Actσj , C σj , E σj , Invσj )
generated from an event-triggered control loop with triggering coefficient σj ∈
σ
]0, σ̄[ for j ∈ {1, . . . , p} and assume that Rσs 1 = · · · = Rs p for all s ∈ {1, . . . , q}.
Consider also some constant earMax and a set of earlier update time parameters
{d1 , d¯1 , . . . , dq , d¯q }, such that
¯
¯
∀s ∈ {1, . . . , q} ∃ j ∈ {1, . . . , p} : d¯s ≤ τ σs j .
¯
Then, the timed game automata with options for earlier update, choice of triggering coefficients and limiting the consecutive earlier updates is given by TGAclim =
(Lclim , ℓclim
, Actclim
, Actclim
, C clim , E clim , Invclim ) where
0
c
u
Sq
Sp
• Lclim = j=1 Lσj ∪ s=1 {Rs , Ear s };
• ℓclim
= Rs such that ξ(0) ∈ Rσs 1 ;
0
Sp
• Actclim
= {up!} ∪ Actσ1 ∪ j=1 {aclim
};
c
j
15
• Actclim
= ∅;
u
• C clim = C σ1 ;
S
S
• E clim = qs=1 t∈Es {(Ear s , c = 0, up!, ∅, Rt )}∪
Sq Sp
σ
clim
, ∅, Rs j ),
s=1
j=1 {(Rs , c = 0, aj
σ
(Rs j , (ds ≤ c ≤ d¯s ) ∧ (earNum < earMax),
S∗,
Sp S
¯
q
{c} ∧ (earNum := earNum + 1), Ear s )} ∪ s=1 j=1 {t|(Rs →Rt )∈E σj }
σ
σ
σ
{(Rs j , τ s j ≤ c ≤ τ̄s j , up!, {c} ∧ (earNum := 0), Rt )};
¯
σ
σ
• Invclim (Rs j ) = {c | c ≤ τ̄s j }, Invclim (Rs ) = {c | c = 0}.
There are two differences between Definition 10 and Definition 9. First, in
σ
the edges from Rs j to Ear s , we add condition earNum < earMax to the guard
and add statement earNum := earNum + 1 to the reset. The additional condition
is used to guarantee that the counter of consecutive earlier updates is always
smaller than or equal to its maximum. The statement on the reset is used to
increase the counter variable by one, once an earlier update happens. Recall that
an earlier update happens when one of these edges is taken (cf. Section 3.1.2).
σ
Second, in the edges from Rs j to Rt , we add the statement earNum := 0 to the
reset. Recall that taking these edges represents the event-triggering mechanism
is used (cf. Section 3.1.2). Thus the counter of consecutive earlier updates is
reset to zero. Notice that the variable earNum takes values in {0, 1, . . . , earMax}.
Remark 2. Note that with the presented implementation, either of the control
loops may exhibit an arbitrary number of consecutive earlier triggerings. This
is because the maximum number of consecutive earlier triggerings being a global
counter. The counter is reset to zero whenever any of the control loops runs in
event-triggered fashion. By employing more counters, one could easily generalize
this idea to limit the number of consecutive earlier triggerings for each loop.
After these modifications, the NTGA associated to the set of NCSs becomes TGANCSs := TGAnet | TGAclim1 | · · · | TGAclimN where TGAclimi
represents the TGA associated with the i-th control loop for i ∈ {1, . . . , N }.
In this new NTGA, the state of TGANCSs is described by a (2N + 3)-tuple
(ℓnet , ℓ1 , . . . , ℓN , unet , u1 , . . . , uN , earNum) which includes the additional counter
earNum. It follows that the bad states A are now given by
{(ℓnet , ℓ1 , . . . , ℓN , unet , u1 , . . . , uN , earNum) | ℓnet = Bad }.
3.4
Scheduler Operation
Formally, a scheduler implements a strategy f , see Definition 5, for the NTGA
TGANCSs . The strategy f is applied to TGANCSs providing, based on the run ρ
of TGANCSs up to that time instant the controllable action f (ρ) that guarantees
the satisfaction of the desired specification.
This means in practice that after each discrete transition of the NTGA, i.e.
every time a transmission is placed on the network, first the strategy chooses a
16
triggering coefficient. Then the strategy decides which control loop is updated
and also its update mechanism: early or event triggered. After such a transition,
and possibly after some time elapses, the environment chooses the conic region
containing the next sampled state, which results in a discrete transition of the
NTGA, and the procedure is repeated.
Example 1. Let us illustrate the use of strategies on an example consisting of
two control loops, two triggering coefficients {σ1 , σ2 } and an option for earlier
updates. The initial location of the first and second control loop is R1 and R2 ,
respectively. Initially the run of TGANCSs is
ρ0 = (Idle, R1 , R2 , 0, 0, 0).
After each update, the scheduler selects a triggering coefficient, according to the
strategy f . Suppose that the scheduler chooses σ2 for the first control loop, i.e.
f (ρ0 ) = acl1
2 . The resulting run is
acl1
2
✲ (Idle, Rσ2 , R2 , 0, 0, 0).
1
ρ1 = ρ0
TS
If the scheduler chooses σ1 for the second control loop, i.e. f (ρ1 ) = acl2
1 , the run
becomes
0
acl2
1
TS
TS
✲ (Idle, Rσ2 , R2 , 0, 0, 0)
1
ρ2 = ρ1
✲ (Idle, Rσ2 , Rσ1 , 0, 0, 0).
1
2
Then the scheduler follows the strategy to decide which control loop is updated
and also its update mechanism: early or event triggered. Suppose that the strategy decides to update the first control loop earlier at time d1 . First, the scheduler
¯
delays the system
ρ3 = ρ2
d1
¯✲
TS
(Idle, Rσ1 2 , Rσ2 1 , d1 , d1 , d1 ).
¯ ¯ ¯
Then an earlier update is performed, i.e. f (ρ3 ) = ∗. Notice that action ∗ is the
internal action associated with the first control loop. We have run
ρ4 = ρ3
∗
✲ (Idle, Ear 1 , Rσ1 , d1 , 0, d1 ).
2
¯
¯
TS
Since time cannot elapse in Ear 1 , the environment has to choose the conic region
containing the next sampled state immediately. If the environment chooses R3 ,
the result is run
ρ5 = ρ4
0
up
TS
TS
✲ (Idle, Ear 1 , Rσ1 , d1 , 0, d1 )
2
¯
¯
✲ (InUse, R3 , Rσ1 , 0, 0, d1 ).
2
¯
Notice that TGAnet and the TGA associated with the first control loop move
simultaneously via synchronizing action up. Input action up? belongs to TGAnet ,
whereas output action up! belongs to the TGA corresponding to the first control
loop. Then the scheduler follows the strategy to select a triggering coefficient for
17
the first control loop, for example σ1 , i.e. f (ρ5 ) = acl1
1 . The network is available
again after ∆ time units, resulting in the runs:
cl1
0
✲ (InUse, R3 , Rσ1 , 0, 0, d1 ) a1✲
2
TS
TS
¯
∆
σ1
σ1
(InUse, R3 , R2 , 0, 0, d1 ) ✲ (InUse, Rσ3 1 , Rσ2 1 , ∆, ∆, d1 + ∆),
TS
¯
¯
ρ6 = ρ5
while the network is being used, and
ρ7 = ρ6
∗
✲ (Idle, Rσ1 , Rσ1 , ∆, ∆, d + ∆)
3
2
¯1
TS
once the network is released. Notice that action ∗ is the internal action associated with TGAnet .
Note that this kind of scheduler is a centralized scheduler that needs to have a
perfect overview of the transmissions placed on the network, and the control loop
responsible for it. Furthermore, given that the locations of TGANCSs are related
to the actual sampled states transmitted through the network, the scheduler also
needs to be able to read the content of the transmitted data.
4
Case Study
We showcase the results in an example comprising two event-triggered NCSs
sharing the same communication network. The first control loop is given by
[12, p. 1683]
˙ξ = 0 1 ξ + 0 υ,
−2 3
1
(5)
υ = 1 −4 ξ.
The second control loop is given by [49, p. 1699]
−0.5 0
1
ξ̇ =
ξ+
υ,
0
3.5
1
υ = 1.02 −5.62 ξ.
(6)
In the sequel, we discuss two experimental results for the above example.
Each experiment is characterized by four parameters: the number of conic
regions q, the set of triggering coefficients {σ1 , . . . , σq }, the set of earlier update parameters {d1 , d¯1 , . . . , dq , d¯q } and maximum consecutive earlier triggering
¯
¯
earMax.
4.1
Limiting the Consecutive Earlier Updates
In this experiment, the minimum channel occupancy time is ∆ = 0.005, the
number of conic regions is q = 200 and there is one triggering coefficient σ1 =
0.05. The input value can be updated 0.005 time units before the lower bound
18
0
2
4
6
8
10
0
2
4
6
8
10
Figure 6: Status of the shared communication network up to 10 time units. The
bars on the top and on the bottom of the x axis represent the network is being
used by (6) and (5), respectively. The top and bottom plots represent the result
of experiments discussed in Sections 4.1 and 4.2, respectively.
in all regions, i.e. ds = τ s − 0.005 and d¯s = τ s for all s ∈ {1, . . . , q}. The
¯
¯
¯
maximum consecutive earlier triggering is earMax = 4.
We create a model in UPPAAL-Tiga according to Definition 10: the TGA
for (5), (6) and the shared communication network are denoted tgaT, tgaH and
net, respectively. The specification is given by
control: A[] not( net.Bad )
A strategy is generated with UPPAAL-Tiga to satisfy this specification. To
illustrate the type of strategies synthesized, we show in the following a fragment of the strategy generated by UPPAAL-Tiga for the situation in which the
locations of tgaT, tgaH and net are R1σ1 , Rσ1 1 and Idle, respectively.
State: ( tgaT.R1a1 tgaH.R1a1 net.Idle )
earNum=3
When you are in (25<=tgaH.c && tgaT.c<65 && tgaH.c-tgaT.c<=-35)
|| (85<tgaT.c && 25<=tgaH.c && tgaT.c<105 && tgaH.c<=30)
|| (38<tgaT.c && 25<=tgaH.c && tgaT.c-tgaH.c<=30 && tgaH.c<=30)
|| (25<=tgaH.c && tgaT.c<31 && tgaH.c-tgaT.c<=-5)
|| (25<=tgaH.c && tgaT.c-tgaH.c<=-5 && tgaH.c<=30),
take transition tgaH.R1a1->tgaH.Ear1
{ c >= 25 && c <= 30 && earNum < earMax, up!, 1 }
net.Idle->net.InUse { 1, up?, c := 0 }
When you are in (105<=tgaT.c && tgaT.c <=111 && tgaH.c<25),
take transition tgaT.R1a1->tgaT.Ear1
{ c >= 105 && c <= 111 && earNum < earMax, up!, 1 }
net.Idle->net.InUse { 1, up?, c := 0 }
As shown above, two different conditions, based on the clock values of tgaT,
clock values of tgaH and the difference of clock values in tgaT and tgaH, can be
appreciated: if the first one is satisfied, an early update is forced for tgaH where
the inter-sample time is between 25 and 30; if the second condition is satisfied,
an early update is forced for tgaT where the inter-sample time is between 105
and 111. If none of the conditions are satisfied, no early update is forced, i.e.
the strategy is to let time elapses for both loops.
19
ξ1
t
2
4
6
8
10
−100
100
ξ2
50
t
2
−50
400
200
−200
4
6
8
10
υ
t
2
4
6
8
10
Figure 7: The state and input trajectories for the experiment in Section 4.1.
The solid and dashed lines are the trajectories of (6) and (5), respectively.
The strategy generated by UPPAAL-Tiga was applied to the two NCSs (5)(6), with both systems initialized at the state [1, 100]T , corresponding to R1 in
both of the timing abstractions for the systems. The network status is shown
in Fig. 6 (top), where long and short bars represent event-triggered and earlier
update mechanisms, respectively. Note that while either of the control loops may
exhibit an arbitrary number of consecutive triggerings, the maximum number
of consecutive earlier triggerings is respected to be below 4 as this counter is
a shared (global) one that is reset to zero whenever any of the two loops run
in event-triggered fashion. During the time horizon of 10, the input of (5)
is updated 63 times consisting of 50 earlier updates and 13 event-triggering
mechanisms. For (6), the input is updated 152 times consisting of 121 earlier
updates and 31 event-triggering mechanisms. The state and input trajectories
are shown in Fig. 7.
4.2
Choice of Triggering Coefficients
In this experiment, the minimum channel occupancy time is ∆ = 0.005, the
number of conic regions is q = 200 and there are three triggering coefficients
σ1 = 0.01, σ2 = 0.03 and σ3 = 0.09.
We create again a model in UPPAAL-Tiga according to Definition 9: the
TGA for (5), (6) and the shared communication network are denoted tgaT, tgaH
and net, respectively. The specification is the same as in Section 4.1 and again
we generate a strategy using UPPAAL-Tiga. The following is a fragment of the
strategy generated by UPPAAL-Tiga when the location of tgaT, tgaH and net
is R37 , Rσ381 and Idle, respectively.
20
ξ1
t
2
4
6
8
10
−100
100
ξ2
50
t
2
−50
400
200
−200
4
6
8
10
υ
t
2
4
6
8
10
Figure 8: The state and input trajectories for the experiment in Section 4.2.
The solid and dashed lines are the trajectories of (6) and (5), respectively.
State: ( tgaT.R37 tgaH.R38a1 net.Idle )
When you are in (tgaT.c==0 && 145<tgaH.c &&
take transition tgaT.R37->tgaT.R37a1 { c ==
When you are in (tgaT.c==0 && 65<=tgaH.c &&
take transition tgaT.R37->tgaT.R37a2 { c ==
When you are in (tgaT.c==0 && 102<tgaH.c &&
|| (tgaT.c==0 && 5<=tgaH.c && tgaH.c<65),
take transition tgaT.R37->tgaT.R37a3 { c ==
tgaH.c<=154),
0, tau, 1 }
tgaH.c<=102),
0, tau, 1 }
tgaH.c<=145)
0, tau, 1 }
Notice that now there are three conditions: if the i-th condition is satisfied, the
location of tgaT is forced to transit to Rσ37i for i ∈ {1, 2, 3}.
The strategy generated by UPPAAL-Tiga is applied to the NCSs (5)-(6),
both with the state initialized at [1, 100]T , corresponding in both cases with
the initial location is R1 . The network status is shown in Fig. 6 (bottom).
Short, medium and long bars represent event-triggered with triggering coefficient equals σ1 , σ2 and σ3 , respectively. During the first 10 time units, the input
of (5) is updated 84 times consisting of 56 updates using σ1 , 6 updates using σ2
and 22 updates using σ3 . For (6), the input is updated 182 times consisting of
106 updates using σ1 , 26 updates using σ2 and 50 updates using σ3 . The state
and input trajectories are shown in Fig. 8.
5
Discussion and Future Work
We have provided an approach to synthesize conflict-free scheduling policies for
sets of networked control systems (NCSs) with the possibilities of updating the
21
input value according to an event-triggering mechanism, selectable from a set
of them, or earlier than the time dictated by such an event-triggering rule. As
indicated in Section 3.3 the main limitations of this proposed scheduling scheme
is its centralized nature, and the fact that the scheduler needs to be able to read
the content of the messages sent through the network. The approach is nonetheless applicable to many setups encountered in practice in which different control
systems are interconnected through a bus, e.g. CAN, EtherCAT or FlexRay. In
such systems, every element connected to the bus can see the traffic flowing
through the network. While it may be the case that the precise content of
messages is not available (e.g. for potential security reasons), it is worth noting
that for the scheduler only the abstracted state, i.e. the region Rs , is relevant.
Therefore we can envisage implementations or practical applications in which
this sort of scheduling could be easily adopted.
In wireless settings, the event-triggered paradigm offers great benefits for
energy consumption reduction, but network topologies can be in general more
complex than a simple bus type of configuration. Therefore, interesting extensions of this work to allow decentralized scheduling, possibly including network
topological constraints, would enable broader applicability of these techniques.
Current and future work is focusing on these issues, extensions of the abstraction of the timing of event-triggered systems beyond LTI systems with statefeedback, and on the implementation of a tool-box automating the whole timing
abstraction and scheduler synthesis proposed in [27] and the current paper respectively.
References
[1] J. Liu, A. Gusrialdi, D. Obradovic, and S. Hirche, “Study on the effect of
time delay on the performance of distributed power grids with networked
cooperative control,” in Proceedings of the 1st IFAC Workshop on Estimation and Control of Networked Systems, 2009, pp. 168–173.
[2] H. A. Thompson, “Wireless and internet communications technologies for
monitoring and control,” Control Engineering Practice, vol. 12, no. 6, pp.
781–791, 2004.
[3] D. Lehmann and J. Lunze, “Extension and experimental evaluation of
an event-based state-feedback approach,” Control Engineering Practice,
vol. 19, no. 2, pp. 101–112, 2011.
[4] P. Antsaklis and J. Baillieul, “Special issue on technology of networked
control systems,” Proceedings of the IEEE, vol. 95, no. 1, pp. 5–8, Jan.
2007.
[5] G. Nair, F. Fagnani, S. Zampieri, and R. Evans, “Feedback control under
data rate constraints: An overview,” Proceedings of the IEEE, vol. 95, no. 1,
pp. 108–137, Jan. 2007.
22
[6] J. P. Hespanha, P. Naghshtabrizi, and Y. Xu, “A survey of recent results in
networked control systems,” Proceedings of the IEEE, vol. 95, pp. 138–162,
Jan. 2007.
[7] D. Hristu-Varsakelis and W. S. Levine, Eds., Handbook of networked
and embedded control systems, ser. Control Engineering. Boston, MA:
Birkhäuser Boston Inc., 2005.
[8] K.-E. Årzén, A. Bicchi, S. Hailes, K. Johansson, and J. Lygeros, “On the design and control of wireless networked embedded systems,” in IEEE Computer Aided Control System Design,, Oct. 2006, pp. 440–445.
[9] M. Rabi and K. H. Johansson, “Event-triggered strategies for industrial
control over wireless networks,” in Proc. 4th Annual Int. Conf. Wireless
Internet. ICST (Institute for Computer Sciences, Social-Informatics and
Telecommunications Engineering), 2008, p. 34.
[10] M. Mazo Jr. and P. Tabuada, “Decentralized event-triggered control over
wireless sensor/actuator networks.” IEEE Transactions on Automatic Control, Special issue on Wireless Sensor Actuator Networks, vol. 56, no. 10,
pp. 2456–2461, Oct. 2011.
[11] K. Åström and B. Bernhardsson, “Comparison of Riemann and Lebesgue
sampling for first order stochastic systems,” in Proceedings of the 41th IEEE
Conference on Decision and Control (CDC’02), vol. 2, Dec. 2002, pp. 2011–
2016.
[12] P. Tabuada, “Event-triggered real-time scheduling of stabilizing control
tasks,” IEEE Transactions on Automatic Control, vol. 52, no. 9, pp. 1680–
1685, Sep. 2007.
[13] M. Velasco, J. Fuertes, and P. Marti, “The self triggered task model for realtime control systems,” in Proceedings of the 24th IEEE Real-Time Systems
Symposium (Work in Progress), 2003, pp. 67–70.
[14] W. Heemels, J. Sandee, and P. van den Bosch, “Analysis of event-driven
controllers for linear systems,” International Journal of Control, vol. 81,
no. 4, pp. 571–590, 2008.
[15] A. Anta and P. Tabuada, “To sample or not to sample: Self-triggered
control for nonlinear systems,” IEEE Transactions on Automatic Control,
vol. 55, pp. 2030–2042, Sep. 2010.
[16] M. Mazo Jr., A. Anta, and P. Tabuada, “An ISS self-triggered implementation of linear controller,” Automatica, vol. 46, pp. 1310–1314, Aug. 2010.
[17] B. Sprunt, L. Sha, and J. Lehoczky, “Aperiodic task scheduling for hardreal-time systems,” Real-Time Systems, vol. 1, no. 1, pp. 27–60, 1989.
23
[18] G. C. Buttazzo, Hard real-time computing systems: predictable scheduling
algorithms and applications. Springer, 2011, vol. 24.
[19] G. C. Walsh and H. Ye, “Scheduling of Networked Control Systems,” IEEE
Control Systems Magazine, 2001.
[20] A. Cervin and T. Henningsson, “Scheduling of event-triggered controllers
on a shared network,” in Proceedings of the 47th IEEE Conference on Decision and Control (CDC’08), Dec. 2008, pp. 3601–3606.
[21] S. Al-Areqi, D. Gorges, S. Reimann, and S. Liu, “Event-based control and
scheduling codesign of networked embedded control systems,” in Proceedings of the 32nd American Control Conference (ACC’13), Jun. 2013, pp.
5299–5304.
[22] S. Al-Areqi, D. Gorges, and S. Liu, “Stochastic event-based control and
scheduling of large-scale networked control systems,” in Proceedings of the
European Control Conference, Jun. 2014, pp. 2316–2321.
[23] S. Reimann, S. Al-Areqi, and S. Liu, “An event-based online scheduling
approach for networked embedded control systems,” in Proceedings of the
32nd American Control Conference (ACC’13), Jun. 2013, pp. 5326–5331.
[24] P. Bouyer, F. Cassez, E. Fleury, and K. Larsen, “Optimal strategies in
priced timed game automata,” in Foundations of Software Technology and
Theoretical Computer Science (FSTTCS’04), ser. Lecture Notes in Computer Science, K. Lodaya and M. Mahajan, Eds. Springer, Heidelberg,
2005, vol. 3328, pp. 148–160.
[25] ——, “Synthesis of optimal strategies using HyTech,” Electronic Notes in
Theoretical Computer Science, vol. 119, no. 1, pp. 11–31, 2005.
[26] O. Maler, A. Pnueli, and J. Sifakis, “On the synthesis of discrete controllers
for timed systems,” in Proc. 12th Symp. on Theoretical Aspects of Computer
Science (STACS’95), ser. Lecture Notes in Computer Science, E. Mayr and
C. Puech, Eds. Springer, Heidelberg, 1995, vol. 900, pp. 229–242.
[27] A. Sharifi Kolarijani and M. Mazo Jr., “A formal traffic characterization of
LTI event-triggered control systems,” CoRR, vol. abs/1503.05816, 2015.
[28] L. De Alfaro, T. Henzinger, and R. Majumdar, “Symbolic algorithms for
infinite-state games,” in Concurrency Theory (CONCUR’01), ser. Lecture
Notes in Computer Science, K. Larsen and M. Nielsen, Eds. Springer,
Heidelberg, 2001, vol. 2154, pp. 536–550.
[29] F. Cassez, A. David, E. Fleury, K. Larsen, and D. Lime, “Efficient on-thefly algorithms for the analysis of timed games,” in Concurrency Theory
(CONCUR’05), ser. Lecture Notes in Computer Science, M. Abadi and
L. de Alfaro, Eds. Springer, Heidelberg, 2005, vol. 3653, pp. 66–80.
24
[30] E. Asarin, O. Maler, A. Pnueli, and J. Sifakis, “Controller synthesis for
timed automata,” in Proc. IFAC Symp. on System Structure & Control,
1998, pp. 469–474.
[31] R. Alur and D. Dill, “A theory of timed automata,” Theoretical Computer
Science, vol. 126, no. 2, pp. 183–235, 1994.
[32] A. Ravn, J. Srba, and S. Vighio, “Modelling and verification of web services
business activity protocol,” in Tools and Algorithms for the Construction
and Analysis of Systems (TACAS’11), ser. Lecture Notes in Computer Science, P. Abdulla and K. Leino, Eds. Springer, Heidelberg, 2011, vol. 6605,
pp. 357–371.
[33] K. Havelund, A. Skou, K. Larsen, and K. Lund, “Formal modeling and
analysis of an audio/video protocol: an industrial case study using UPPAAL,” in Proceedings of the 18th IEEE Real-Time Systems Symposium
(RTSS’97), Dec. 1997, pp. 2–13.
[34] P. D’Argenio, J.-P. Katoen, T. Ruys, and J. Tretmans, “The bounded
retransmission protocol must be on time!” in Tools and Algorithms for
the Construction and Analysis of Systems (TACAS’97), ser. Lecture Notes
in Computer Science, E. Brinksma, Ed. Springer, Heidelberg, 1997, vol.
1217, pp. 416–431.
[35] L. Aceto, A. Burgueño, and K. Larsen, “Model checking via reachability
testing for timed automata,” in Tools and Algorithms for the Construction
and Analysis of Systems (TACAS’98), ser. Lecture Notes in Computer Science, B. Steffen, Ed. Springer, Heidelberg, 1998, vol. 1384, pp. 263–280.
[36] H. Jensen, K. Larsen, and A. Skou, “Modelling and analysis of a collision avoidance protocol using SPIN and UPPAAL,” BRICS Report Series,
vol. 3, no. 24, 1996.
[37] A. David and W. Yi, “Modelling and analysis of a commercial field bus
protocol,” in 12th Euromicro Conference on Real-Time Systems, 2000, pp.
165–172.
[38] T. Henzinger, X. Nicollin, J. Sifakis, and S. Yovine, “Symbolic model checking for real-time systems,” Information and Computation, vol. 111, no. 2,
pp. 193–244, 1994.
[39] A. Fehnker, “Scheduling a steel plant with timed automata,” in Proc. 6th
Int. Conf. Real-Time Computing Systems and Applications (RTCSA’99),
1999, pp. 280–286.
[40] Y. Abdeddaı̈m and O. Maler, “Job-shop scheduling using timed automata?” in Computer Aided Verification (CAV’01), ser. Lecture Notes
in Computer Science, G. Berry, H. Comon, and A. Finkel, Eds. Springer,
Heidelberg, 2001, vol. 2102, pp. 478–492.
25
[41] Y. Abdeddaı̈m, A. Kerbaa, and O. Maler, “Task graph scheduling using
timed automata,” in Proc. Int. Parallel and Distributed Processing Symposium (IPDPS’03), Apr. 2003, pp. 8 pp.–.
[42] G. Behrmann, E. Brinksma, M. Hendriks, and A. Mader, “Scheduling lacquer production by reachability analysis – a case study,” in Proceedings of
the 16th IFAC World Congress. Elsevier, 2005.
[43] ——, “Production scheduling by reachability analysis - a case study,”
in Proc. 19th IEEE Int. Parallel and Distributed Processing Symposium
(IPDPS’05), Apr. 2005, pp. 140a–140a.
[44] G. Behrmann, A. Fehnker, T. Hune, K. Larsen, P. Pettersson, J. Romijn,
and F. Vaandrager, “Minimum-cost reachability for priced time automata,”
in Hybrid Systems: Computation and Control (HSCC’01), ser. Lecture Notes in Computer Science, M. Di Benedetto and A. SangiovanniVincentelli, Eds. Springer, Heidelberg, 2001, vol. 2034, pp. 147–161.
[45] R. Alur, S. La Torre, and G. Pappas, “Optimal paths in weighted timed
automata,” in Hybrid Systems: Computation and Control (HSCC’01), ser.
Lecture Notes in Computer Science, M. Di Benedetto and A. SangiovanniVincentelli, Eds. Springer, Heidelberg, 2001, vol. 2034, pp. 49–62.
[46] P. Bouyer, E. Brinksma, and K. Larsen, “Optimal infinite scheduling for
multi-priced timed automata,” Formal Methods in System Design, vol. 32,
no. 1, pp. 3–23, 2008.
[47] J. Bengtsson and W. Yi, “Timed automata: Semantics, algorithms and
tools,” in Lectures on Concurrency and Petri Nets, ser. Lecture Notes in
Computer Science, J. Desel, W. Reisig, and G. Rozenberg, Eds. Springer,
Heidelberg, 2004, vol. 3098, pp. 87–124.
[48] G. Behrmann, A. David, and K. Larsen, “A tutorial on Uppaal,” in Formal
Methods for the Design of Real-Time Systems (SFM-RT’04), ser. Lecture
Notes in Computer Science, M. Bernardo and F. Corradini, Eds., vol. 3185.
Springer, Heidelberg, Sep. 2004, pp. 200–236.
[49] L. Hetel, A. Kruszewski, W. Perruquetti, and J.-P. Richard, “Discrete and
intersample analysis of systems with aperiodic sampling,” IEEE Transactions on Automatic Control, vol. 56, no. 7, pp. 1696–1701, Jul. 2011.
26
| 3cs.SY
|
IEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x
1
LETTER
Local Modules in Imperative Languages
arXiv:1701.05034v4 [cs.PL] 19 Oct 2017
Keehang KWON† , Member and Daeseong KANG†† , Nonmember
SUMMARY
We propose a notion of local modules for imperative langauges. To be specific, we introduce a new implication
statement of the form D ⊃ G where D is a module (i.e., a set
of procedure declarations) and G is a statement. This statement
tells the machine to add D temporarily to the program in the
course of executing G. Thus, D acts as a local module and will be
discarded after executing G. It therefore provides efficient program management. In addition, we describe a new constructive
module language to improve code reuse. Finally, we describe a
scheme which improves the heap management in traditional languages. We illustrate our idea via Cmod , an extension of the core
C with the new statement.
key words: local modules, program management, memory management.
1.
by its most recent declaration. On the other hand, most
modern languages allow local procedures within nested
procedures. These approaches are based on static scoping in granting access and requires no run-time loading
and unloading. The main advantages of our approach
is the following:
(1)
It allows local procedures at the statement level,
whereas other languages allow local procedures
only at the procedural level. Thus our langauge
provides the programmer more flexibility.
(2)
It leads to efficient program/memory management
due to loading and unloading. This is not negligible when local procedures are big.
(3)
It has a simple, natural syntax and semantics due
to dynamic scoping of local procedures. In contrast, it is well-known that other systems have
awkward and complicated semantics mainly due
to static scoping. Consequently, these systems are
very difficult to read, write, implement and reason
about.
Introduction
Efficient program management in imperative programming – C, its extension [4] and Java – is an important
issue. Yet this has proven a difficult task, relying on adhoc techniques such as various cache/page replacement
algorithms.
An analysis shows that this difficulty comes from
the fact that the machine has no idea as to which modules are to be used in the near future. Therefore module
management can be made easier by making the programmer specify which modules are to be used. Toward this end, inspired by [5]–[7], we propose a new
implication statement D ⊃ G, where D is a set of procedure declarations and G is a statement. This has the
following execution semantics: add D temporarily to
the current program in the course of executing G. In
other words, the machine loads D to the current program, executes G, and then unloads D from the current
program. This implication statement is closely related
to the let-expression in functional languages and the
implication goals in logic languages [6], [7].
Thus D acts as local procedures to G in that D is
hidden from the rest. Our approach calls for a new runtime stack called program stack, as it requires run-time
loading and unloading. It follows that local procedures
in our language is (stack) dynamic scoped in the sense
that the meaning of a procedure is always determined
Manuscript received January 1, 2003.
Manuscript revised January 1, 2003.
Final manuscript received January 1, 2003.
†
The authors are with Computer Eng., DongA University. email:[email protected]
††
The author is with Electronics Eng., DongA University.
On the negative side, it requires a little run-time overhead for loading and unloading.
In the sequel, a module is nothing but a set of procedures with a name. Our notion of local procedures
extends to a notion of local (occurrences of ) modules in
a straightforward way. That is, we propose a new module implication statement /m ⊃ G, where m is a module
name and G is a statement. This has the following execution semantics: add a (local occurrence of) module
m temporarily to the current program in the course of
executing G. Note that our modules are stack dynamic
in the sense that they are loaded/unloaded in the program in a stack fashion. This leads naturally to the dynamic scoping for procedure names. In contrast, most
imperative languages have a module language which is
typically based on the notion of static modules with
no run-time loading and unloading, leading naturally
to static scoping. It is well-known that static scoping
causes the naming problem among procedures across
independent modules.
Our module system has some advantages over
other popular module systems in imperative languages.
(1)
It allows the programmer to load and unload other
IEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x
2
modules due to the module implication statement.
This leads to the dynamic scoping for procedure
names. In contrast, this is traditionally impossible
in other languages, leading to the static scoping
and the naming problem.
(2)
Dynamic scoping leads to a simple, natural syntax
and semantics, as there is no naming problem.
(3)
As we shall see later, it allows mutually recursive
modules thanks to dynamic scoping.
In addition, we add a novel module language to
improve code reuse. Finally, we propose a variant of
the implication statement which considerably simplifies
the heap management.
This paper extends a C-like language with the new
implication statement. We focus on the minimum core
of C. The remainder of this paper is structured as follows. We describe Cmod , an extension of C with a new
statement in Section 2. In Section 3, we present an example of Cmod . In Section 4, we describe a constructive
module language for enhancing code reuse. In Section
5, we propose a scheme that improves the heap management. Section 6 concludes the paper.
2.
the interpreter tries to find a matching procedure for a
procedure call A inside the module D by decomposing
D into a smaller unit (via rule (4)-(5)) and reducing D
to its instance (via rule (2)) and then backchaining on
the resulting definition (via rule (1)). To be specific,
the rule (2) basically deals with argument passing: it
eliminates the universal quantifier x in ∀xD by picking a value t for x so that the resulting instantiation,
[t/x]D, matches the procedure call A. The notation S
seqand R denotes the sequential execution of two tasks.
To be precise, it denotes the following: execute S and
execute R sequentially. It is considered a success if both
executions succeed. Similarly, the notation S parand R
denotes the parallel execution of two tasks. To be precise, it denotes the following: execute S and execute
R in any order. It is considered a success if both executions succeed. The notation S ← R denotes reverse
implication, i.e., R → S.
Definition 1. Let G be a statement and let P be the
program. Then the notion of executing hP, Gi and producing a new program P ′ – ex(P, G, P ′ ) – is defined as
follows:
(1)
bc((A = G1 ), P, A, P1 ) ←
ex(P, G1 , P1 ). % A matching procedure for A is
found.
(2)
bc(∀xD, P, A, P1 , ) ←
bc([t/x]D, P, A, P1 ). % argument passing
(3)
bc(D1 ∧ D2 , P, A, P1 ) ←
bc(D1 , P, A, P1 ). % look for a matching procedure
in D1 .
(4)
bc(D1 ∧ D2 , P, A, P1 ) ←
bc(D2 , P, A, P1 ). % look for a matching procedure
in D2
(5)
ex(P, A, P1 ) ← (Di ∈ P) parand bc(Di , P, A, P1 ),
provided that Di is the first module in the stack,
which contains a declaration of A. % A is a procedure call
(6)
ex(P, true, P). % True is always a success.
(7)
ex(hS, θi, x = E, hS, θ⊎{(x, E ′ )}i) ← eval(P, E, E ′ ).
% evaluate E to get E ′ . Here, ⊎ denotes a set
union but (x, V ) in θ will be replaced by (x, E ′ ).
(8)
ex(P, G1 ; G2 , P2 ) ←
ex(P, G1 , P1 ) seqand ex(P1 , G2 , P2 ). % a sequential composition
(9)
ex(hS, θi, D ⊃ G1 , P1 ) ←
ex((hD :: S, θi, G1 , P1 ). % add D to the top of the
program stack S.
The Core Language
The language is core C with procedure definitions. It
is described by G- and D-formulas given by the BNF
syntax rules below:
G ::= true | A | x = E | G; G | D ⊃ G
D ::= A = G | ∀x D | D ∧ D
In the above, A represents a head of a procedure declaration p(x1 , . . . , xn ) where x1 , . . . , xn are parameters.
A D-formula is a set of procedure declarations. In
the transition system to be considered, a G-formula
will function as a statement and a list of D-formulas
enhanced with the machine state (a set of variablevalue bindings) will constitute a program. Thus, a
program is a pair of two disjoint components, i.e.,
hD1 :: . . . :: Dn :: nil, θi where D1 :: . . . :: Dn :: nil
is a stack of D-formulas and θ represents the machine
state. θ is initially empty and will be updated dynamically during execution via the assignment statements.
We will present an interpreter for our language via
natural semantics [3]. Note that our interpreter alternates between the execution phase and the backchaining phase.
In the execution phase (denoted by
ex(P, G, P ′ )), it executes a statement G with respect
to P and produce a new program P ′ by reducing G to
simpler forms. The rules (6)-(9) deal with this phase.
If G becomes a procedure call, the machine switches to
the backchaining mode. This is encoded in the rule (5).
In the backchaining mode (denoted by bc(D, P, A, P ′ )),
(10) ex(hS, θi, /m ⊃ G1 , P1 ) ←
ex((hD :: S, θi, G1 , P1 ). % add D to the top of the
program stack S.
LETTER
3
If ex(P, G, P1 ) has no derivation, then the interpreter
returns the failure. The rule (9) deals with the new
feature.
3.
Examples
In our language, a module is simply a set of procedures
associated with a name. Below the keyword module
associates a name to a D-formula. The following module Emp has a procedure Age which sets the variable
named age, whose value represents the employee’s age.
Similarly, the module Bank is defined with the procedures Deposit, Withdraw, Balance.
module Emp.
Age(emp) =
switch (emp) {
case tom: age = 31; break;
case kim: age = 40; break;
case sue: age = 22; break;
default: age = 0; break;
}
module Bank.
Deposit(name,amount) = . . .
Withdraw(name,amount) = . . .
Balance(name) = . . .
Now consider executing the following main statement G from an empty program.
% first task using module EmpAge
( Emp ⇒
(Age(tom); print(age);
Age(kim); print(age);
Age(sue); print(age)))
;
% second task using module Bank
( Bank ⇒
deposit(tom,$100))
Execution proceeds as follows: Initially the program
is empty. Then, the machine loads the module Emp
to the program, printing the ages of three employees
– Tom, Kim and Sue –, and then unloads the module
Emp. Then, the machine loads the module Bank to
the program, deposits $100 to Tom’s account, and then
unloads the module Bank. Note that the module Emp
is available to the first task only, while Bank to the
second task only.
As the second example, let us consider two mutually recursive modules Ev and Od. The module Ev
has a procedure Even(x) which returns true if x is even.
Similarly, the module Od is defined with the procedure
Odd(x).
module Ev.
Even(x) = if x == 0, true else Od ⇒ Odd(x-1);
module Od.
Odd(x) = if x == 1, true else Ev ⇒ Even(x-1);
Now consider executing even(9) from the module
Ev. Execution proceeds as follows: Initially the program is empty. Then, the machine loads the module
Emp to the program, printing the ages of three employees – Tom, Kim and Sue –, and then unloads the
module Emp. Then, the machine loads the module
Bank to the program, deposits $100 to Tom’s account,
and then unloads the module Bank. Note that the
module Emp is available to the first task only, while
Bank to the second task only.
4.
A Constructive Module Langauge
Modern languages typically support code reuse via inheritance. We propose a constructive approach to code
reuse as an alternative to inheritance. To begin with,
our language provides a special macro function / which
binds a name to a set of method (and constant) declarations. This macro function serves to represent programs in a concise way. For example, given two macro
definitions /p = f (x) = x and /q = g(x) = 0, the
notation /p ∧ /q represents f (x) = x ∧ g(x) = 0. Here
∧ means the accumulation of two modules.
In addition to ∧, our module language provides a
IEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x
4
renaming operation of the form ren(b, a)D which replaces b by a in a module D and ∀xD for universal
generalization. There are other useful operations such
as private f D (reuse D with making f private) and
shareD (reuse D via sharing, not copying) but we will
not discuss them further here.
Now let us consider macro processing. Macro definitions are typically processed before the execution but
in our setting, it is possible to process macros and execute regular programs in an interleaved fashion. We
adopt this approach below.
We reconsider the language in Section 2.
(7)
ex(P, A, P1 ) ← (Di ∈ P) parand bc(Di , P, A, P1 ),
provided that Di is the first module in the stack,
which contains a declaration of A. % A is a procedure call
(8)
ex(P, true, P). % True is always a success.
(9)
ex(hS, M, θi, x = E, hS, M, θ ⊎ {(x, E ′ )}i) ←
eval(P, E, E ′ ).
% evaluate E to get E ′ . Here, ⊎ denotes a set
union but (x, V ) in θ will be replaced by (x, E ′ ).
(10) ex(P, G1 ; G2 , P2 ) ←
ex(P, G1 , P1 ) seqand ex(P1 , G2 , P2 ). % a sequential composition
G ::= true | A | x = E | G; G | D ⊃ G | /n : M ⊃ G
D ::= A = G | /n | ren(a, b)D | ∀x D | D ∧ D
(11) ex(hS, M, θi, D ⊃ G1 , P1 ) ←
M ::= /n = D | M ∧ M
ex((hD :: S, M, θi, G1 , P1 ). % add D to the top of
the program stack S.
In the above, n is a name and A represents a head of
a procedure declaration p(x1 , . . . , xn ) where x1 , . . . , xn
are parameters. A D-formula is a set of procedure declarations. An M -formula is called macro definitions and
M is a list of M -formulas. In the transition system to
be considered, a G-formula will function as a statement
and a list of D-formulas, a list of M -formulas and the
machine state (a set of variable-value bindings) will constitute a program. Thus, a program is a pair of three
disjoint components, i.e., hD1 :: . . . :: Dn :: nil, M, θi
where θ represents the machine state. θ is initially
empty and will be updated dynamically during execution via the assignment statements.
Definition 2. Let G be a statement and let P be the
program. Then the notion of executing hP, Gi and producing a new program P ′ – ex(P, G, P ′ ) – is defined as
follows:
(1)
bc((A = G1 ), P, A, P1 ) ←
ex(P, G1 , P1 ). % A matching procedure for A is
found.
(2)
bc(∀xD, P, A, P1 , ) ←
bc([t/x]D, P, A, P1 ). % argument passing
(3)
bc(D1 ∧ D2 , P, A, P1 ) ←
bc(D1 , P, A, P1 ). % look for a matching procedure
in D1 .
(4)
bc(D1 ∧ D2 , P, A, P1 ) ←
bc(D2 , P, A, P1 ). % look for a matching procedure
in D2
(5)
bc(ren(a, b)D, P, A, P1 ) ←
bc([b/a]D, P, A, P1 ). % renaming operation
(6)
bc(/n, P, A, P1 ) if bc(D, P, A, P1 ) and (/n = D) ∈
M. % we assume it chooses the most recent macro
definition.
(12) ex(hS, M, θi, /n : M ⊃ G1 , P1 ) if ex(h/n :: S, M ::
M, θi, G1 , P1 ) % Add new macros to the front of
M. Here :: is a list constructor.
If ex(P, G, P1 ) has no derivation, then the interpreter
returns the failure.
5.
Improving Heap Management
Our earlier discussions in Section 2 are based on dynamic procedure binding. More interestingly, our notion of implication statements can be applied equally
well to static procedure/data binding.
To be specific, allocation and deallocation of
(heap) objects – and accessing them through pointers
– occur frequently in traditional imperative languages
with static procedure/data binding. This includes
malloc-free for memory management, new-dispose for
objects, new-delete for heap objects (arrays, records,
etc). Unfortunately, allocation and deallocation constructs are unstructured. Using allocation and deallocation carelessly leads to serious problems.
Towards an efficient yet robust memory management, we need to impose some restrictions on the use
of allocation and deallocation by providingng a highlevel statement. To be specific, we propose to use the
following pointer-implication statement of the form
(p = new obj) ⊃ G
where p is a pointer and obj is a program object
or a data object (arrary, record, etc). This is a variant
of the implication statement in Section 1 with the following semantics: It creates an object obj with p being
a pointer to obj, executes the statement G and then
deallocate obj and p. To avoid additional complications, we assume that pointers can only be initialized
but not manipulated.
The following code is an example where arrayptr1
LETTER
5
and int[100] are available in both the statements S1 and
S2 , while arrayptr2 and int[1000] are available only in
S2 .
% heap allocation
(arrayptr1 = new int[100] ⊃
(arrayptr2 = new int[1000] ⊃ S1 )
S2 )
Our system requires considerable change to memory management: it needs to maintain three different
categories of memory: program/data stack, run-time
stack and the heap. Program/data stack is a new component and it – instead of the heap – is used to maintain
program/data objects created via the new construct.
Thus it replaces most of the works done by the heap.
Program/data stack has considerable advantages over
the heap: it is more efficient and can simplify several
complications caused by the heap including garbage collection, heap fragmentation, and dangling pointers.
6.
Conclusion
In this paper, we have proposed a simple extension to
imperative languages. This extension introduced an
implication statement D ⊃ G where D is a module
and G is a statement. This statement makes D local
to G. It therefore maintains only the active modules in
the current program context.
7.
Acknowledgements
This work was supported by Dong-A University Research Fund.
References
[1] J. Alglave and L. Maranget and M. Tautschnig, “Herding
cats: Modelling, simulation, teating and data mining for
weak memory”, ACM Transactions on Programming Languages and Systems, vol.36, no.2, pp.1–74, 2014.
[2] G. Boudol and G. Petri, “Relaxed memory models: an operational approach”, In POPL, pp.392–403, ACM, 2009.
[3] G. Kahn, “Natural Semantics”, In the 4th Annual Symposium on Theoretical Aspects of Computer Science, LNCS vol.
247, 1987.
[4] K. Kwon, S. Hur and M. Park, “Improving Robustness via
Disjunctive Statements in Imperative Programming”, IEICE
Transations on Information and Systems, vol.E96-D,No.9,
pp.2036-2038, September, 2013.
[5] J. Hodas and D. Miller, “Logic Programming in a Fragment
of Intuitionistic Linear Logic”, Information and Computation, vol.110, No.2, pp.327-365, 1994.
[6] D. Miller, G. Nadathur, F. Pfenning, and A. Scedrov, “Uniform proofs as a foundation for logic programming”, Annals
of Pure and Applied Logic, vol.51, pp.125–157, 1991.
[7] D. Miller, G. Nadathur, Programming with higher-order
logic, Cambridge University Press, 2012.
| 6cs.PL
|
Good Code Sets from Complementary Pairs via
Discrete Frequency Chips
Ravi Kadlimatti 1,*and Adly T. Fam 2
Advanced Wireless Systems Research Center, State University of New York at Oswego, Oswego, NY
13126, USA; [email protected]
2
Department of Electrical Engineering, University at Buffalo, The State University of New York, Buffalo,
NY 14260, USA; [email protected]
* Correspondence: [email protected]; Tel.: +1-716-491-8985
1
Abstract: It is shown that replacing the sinusoidal chip in Golay complementary code pairs by
special classes of waveforms that satisfy two conditions, symmetry/anti-symmetry and
quazi-orthogonality in the convolution sense, renders the complementary codes immune to
frequency selective fading and also allows for concatenating them in time using one frequency
band/channel. This results in a zero-sidelobe region around the mainlobe and an adjacent region of
small cross-correlation sidelobes. The symmetry/anti-symmetry property results in the
zero-sidelobe region on either side of the mainlobe, while quasi-orthogonality of the two chips
keeps the adjacent region of cross-correlations small. Such codes are constructed using discrete
frequency-coding waveforms (DFCW) based on linear frequency modulation (LFM) and piecewise
LFM (PLFM) waveforms as chips for the complementary code pair, as they satisfy both the
symmetry/anti-symmetry and quasi-orthogonality conditions. It is also shown that changing the
slopes/chirp rates of the DFCW waveforms (based on LFM and PLFM waveforms) used as chips
with the same complementary code pair results in good code sets with a zero-sidelobe region. It is
also shown that a second good code set with a zero-sidelobe region could be constructed from the
mates of the complementary code pair, while using the same DFCW waveforms as their chips. The
cross-correlation between the two sets is shown to contain a zero-sidelobe region and an adjacent
region of small cross-correlation sidelobes. Thus, the two sets are quasi-orthogonal and could be
combined to form a good code set with twice the number of codes without affecting their
cross-correlation properties. Or a better good code set with the same number codes could be
constructed by choosing the best candidates form the two sets. Such code sets find utility in
multiple input-multiple output (MIMO) radar applications.
Keywords: discrete frequency-coding waveform; linear FM; chirp; piecewise LFM; complementary
code pair; symmetry; anti-symmetry; good code sets; MIMO radar
1. Introduction
1.1. Background
Good aperiodic codes are characterized by a narrow mainlobe and small sidelobes. Smaller
autocorrelation peak sidelobes reduce the probability of false alarm, while a narrower mainlobe
enhances the range resolution. Such properties are desirable in certain communications applications
like preamble synchronization and in most radar applications.
Let 𝑥[𝑛] be a code of length 𝑁. Its aperiodic autocorrelation (ACF) can be represented by,
𝑁−1
𝑅[𝑛] = ∑ 𝑥[𝑘]𝑥 ∗ [𝑛 + 𝑘]
𝑘=0
𝑍
In the 𝑍-domain, if 𝑥[𝑛] ⇔ 𝑋(𝑧) then, its ACF can be represented by,
(1)
𝑍
𝑅[𝑛] ⇔ 𝑅(𝑧) = 𝑋(𝑧)𝑋(𝑧 −1 ) = 𝑁 + 𝑆(𝑧)
(2)
where 𝑁 and 𝑆(𝑧) represent the mainlobe/peak and the sidelobes respectively.
Let 𝑝(𝑡) = 𝑒 𝑗2𝜋𝑓𝑝 𝑡 , 0 ≤ 𝑡 ≤ 𝑇, represent a sinusoidal chip. 𝑥[𝑛] modulates 𝑝(𝑡) resulting in
𝑥𝑝 (𝑡).
𝑁−1
𝑥𝑝 (𝑡) = ∑ 𝑥[𝑛] 𝑝(𝑡 − 𝑛𝑇)
(3)
𝑛=0
Its ACF can be represented by,
𝑅𝑥 (𝑡) = ∫
𝑡+𝑇
𝑡
If 𝑥𝑝 (𝑡) is sampled at intervals of 𝑇𝑠 =
0, 1, 2, … , (𝑁 − 1)𝑁𝑠 , can be represented by,
1
𝑓𝑠
𝑥𝑝 (𝜏)𝑥𝑝∗ (𝑡 + 𝜏)𝑑𝜏
(4)
, then in the 𝑍 -domain, 𝑥𝑝 (𝑚𝑇𝑠 ) for 𝑚 =
𝑍
𝑥𝑝 (𝑚𝑇𝑠 ) ⇔ 𝑋𝑝 (𝑧) = 𝑋(𝑧 𝑁𝑠 )𝑃(𝑧)
(5)
𝑍
where 𝑝(𝑚𝑇𝑠 ) ⇔ 𝑃(𝑧) and 𝑁𝑠 represents the number of samples in 𝑃(𝑧). ACF of 𝑋𝑝 (𝑧) can be
represented by,
𝑅𝑥 (𝑧) = 𝑋𝑝 (𝑧)𝑋𝑝 (𝑧 −1 )
(6)
𝑅𝑥 (𝑧) = 𝑋(𝑧 𝑁𝑠 )𝑃(𝑧)𝑋(𝑧 −𝑁𝑠 )𝑃(𝑧 −1 )
(7)
𝑅𝑥 (𝑧) = 𝑅(𝑧 𝑁𝑠 )𝑅𝑝 (𝑧)
(8)
𝑅𝑥 (𝑧) = 𝑁𝑅𝑝 (𝑧) + 𝑆(𝑧 𝑁𝑠 )𝑅𝑝 (𝑧)
(9)
where 𝑅𝑝 (𝑧) = 𝑃(𝑧)𝑃(𝑧 −1 ) represents the ACF of the highly sampled chip and 𝑅(𝑧) is given in
Equation (2). It can be observed that the mainlobe and the sidelobes are spread by the ACF of the
chip (𝑅𝑝 (𝑧)).
Some of the well-known good aperiodic codes are the Barker codes, Frank codes [1] and
waveforms based on Costas arrays [2–4]. Barker codes are biphase codes that have unity magnitude
peak sidelobes. However, the longest known Barker code of odd length is of length 13 which gives a
1
peak sidelobe ratio of . Frank codes are polyphase codes obtained by concatenating the rows of a
13
discrete Fourier transform (DFT) matrix. Thus, Frank codes are available for a variety of code lengths
and achieve high peak sidelobe ratios for long code lengths. However, as the code length increases
the number of distinct phases also increases. Costas arrays are completely classified to date up to
order 27, and are known to exist of orders up to 200. They are widely used in radar and sonar
applications because of their thumbtack ambiguity function. The codes presented in this paper are
based on biphase complementary code pair that are available at a variety of code lengths.
One of the widely used approaches to suppress the ACF sidelobes of a waveform/code is via
windowing whereby the peak sidelobe level is reduced, but at the cost of increasing the mainlobe
width. Mismatched filters like the ones introduced in [5–8] could also be used to reduce ACF
sidelobes at the cost of a small loss in signal-to-noise ratio. In [9], it is shown that complete sidelobe
cancellation for a class of aperiodic codes is possible using additive-multiplicative processing of the
matched filter (MF) output. However, the non-linear processing involved in these mismatched filters
incurs additional computational cost and suffer from degraded performance in high noise
environments. Although it is impossible for individual aperiodic codes to have zero sidelobes, Golay
complementary code pair [10], polyphase complementary codes [11] and complementary code
sequences [12] achieve complete sidelobe cancellation via addition of their autocorrelations. Figure 1
shows the transmission and reception of a Golay complementary code pair (say 𝑎[𝑛] and 𝑏[𝑛], of
length 𝑁𝑔 ). 𝑎[𝑛] is transmitted at frequency 𝑓1 (using 𝑐1 (𝑡) = 𝑒 𝑗2𝜋𝑓1 𝑡 ) and 𝑏[𝑛] at 𝑓2 (using
𝑐2 (𝑡) = 𝑒 𝑗2𝜋𝑓2 𝑡 ). At the receiver, the code pair are received in their own matched filters. Addition of
their MF outputs (𝑅𝑎 (𝑡) and 𝑅𝑏 (𝑡)) results in complete cancellation of the sidelobes. However,
frequency selective fading results in unequal attenuation of the two codes. This results in
reemerging sidelobes due to inexact cancellation of the non-zero sidelobes of the code pair, as shown
in Figure 2 for 𝑁𝑔 = 8. Since the sidelobes of the code pair are not small, any inexact cancellation
results in reemerging sidelobes which is highly undesirable. In Figure 2, 𝑅𝑎 (𝑡), 𝑅𝑏 (𝑡) and 𝑅(𝑡)
represent the fading absent case and 𝑅𝑎 ′(𝑡), 𝑅𝑏 ′(𝑡) and 𝑅′(𝑡) represent the fading present case. For
the fading present case, 𝑎(𝑡) and 𝑏(𝑡) are attenuated by 0.95 and 0.85 respectively. Galati et al.
[13] also discusses in detail this effect of unequal attenuation on the complementary code pairs. Liu
et al. [14] describes z-complementary code pairs that achieve a zero-correlation zone around the
mainlobe and minimum possible sidelobe peaks outsize the zero domain. However, these codes also
require two frequencies or channels which makes them vulnerable to the aforementioned effects of
fading. Tang et al. [15] and Li [16] describe (loosely synchronized) LS and (large area) LA codes for
(quasi-synchronous code division multiple access) QS-CDMA systems. These codes also achieve a
zero interference zone in both the autocorrelation and the cross-correlations, but the peak sidelobes
outside the zero-sidelobe region are not small. In this paper, we introduce codes constructed from
biphase Golay complementary code pairs that are available at a variety of lengths. These codes have
a zero-sidelobe domain on either side of the mainlobe, while the sidelobe peaks outside the zero
domain are very small. Since these codes use the same frequency band, they are not affected by the
effects of fading.
Ng
0
a (t )
a [n ]
e
j 2πf1 t
e
am (t )
2N g
a* (T - t )
am (t )+n (t )
e
j 2πfc t
- j 2πfc t
BPF1
R a (t )
Matched
filter
0
R (t )
Ng
0
b [n ]
bm (t )
b (t )
b* (T - t )
bm (t )+n (t )
e j 2πf2 t e j 2πfc t
e - j 2πfc t
BPF2
Channel
Tx
R b (t )
Matched
filter
Rx
Figure 1. Transmitter (Tx) and receiver (Rx) of a Golay complementary code pair.
N
N
2N
0
0
0
g
g
(a) R (t)
g
(b) R (t)
a
(c) R(t)
b
Ng
Ng
2Ng
0
0
0
(d) Ra'(t)
(e) Rb'(t)
(f) R'(t)
Figure 2. (a–c) Plots of matched filter outputs in the absence of frequency selective fading; and (d–f)
plots of matched filter outputs in the presence of fading.
1.2. The Proposed Codes Based on Complementary Pairs via Discrete Frequency Chips
Given any complementary code pair 𝑎[𝑛] and 𝑏[𝑛] , 𝑎[𝑛] + 𝑏[𝑛 − 𝑁𝐷 ] is constructed by
concatenating 𝑎[𝑛] and 𝑏[𝑛] in the time domain using one frequency band/channel. The chips used
for 𝑎[𝑛] and 𝑏[𝑛] satisfy the following two conditions:
1.
2.
They are symmetrical/anti-symmetrical mirror images of each other, i.e., 𝑐(𝑡) with 𝑎[𝑛] and
𝑐(−𝑡)/−𝑐(−𝑡) (or 𝛾𝑐(−𝑡), where 𝛾 = 1 or −1) with 𝑏[𝑛].
They are quasi-orthogonal in the convolution sense.
This concatenated code has a region of zero-sidelobes on either side of the mainlobe and an
adjacent region of small cross-correlations. Due to the symmetry/anti-symmetry property, the two
chips have identical ACF. This results in exact sidelobe cancellation of the complementary code pair,
creating a zero-sidelobe region on either side of the mainlobe. The cross-correlation peak (𝐶𝐶𝑃)
between 𝑐(𝑡) and 𝛾𝑐(−𝑡) is represented by,
𝑡+𝑇
𝐶𝐶𝑃 = max |
∫𝑡
𝛾𝑐(𝜏)𝑐 ∗ (−𝑡 + 𝜏)𝑑𝜏
2𝑁𝑔
𝑡
|
(10)
Since 𝑐(𝑡) and 𝛾𝑐(−𝑡) are quazi-orthogonal in the convolution, 𝐶𝐶𝑃 is small compared to the
mainlobe peak 2𝑁𝑔 . This property makes the adjacent region of cross-correlation sidelobes small.
Discrete frequency-coding waveforms (DFCW) based on linear frequency modulation (LFM,
also known as a chirp signal) and piecewise LFM (PLFM) [17] waveforms satisfy both the
symmetry/anti-symmetry and quasi-orthogonality conditions. Hence, they are used as chips for the
complementary code pair. A discrete frequency-coding waveform, say 𝑐(𝑡), is the sum of 𝑁
contiguous sub-pulses of the same duration Δ𝑇, but not necessarily the same frequency.
𝑐(𝑡) =
1
𝑁−1
∑ 𝑒 𝑗2𝜋𝑓[𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
(11)
where Δ𝑊 is the smallest possible frequency offset between two frequencies, 𝑓[𝑛]Δ𝑊 is the
frequency of the nth sub-pulse and 𝑓[𝑛] ∈ {0, 1, 2, … , 𝑁 − 1}.
DFCW sets find utility in MIMO radars [18] and Orthogonal Netted Radar Systems (ONRS)
[19,20] that improve radar performance through spatial diversity. Several good DFCW sets have
been proposed, such as the ones in [20–22].
In this paper, it is also shown that a quasi-orthogonal set of symmetrical/anti-symmetrical
DFCW waveforms could be used as chips with the same complementary code pair to result in a
good code set with a zero-sidelobe region. Several good code sets are constructed by changing the
slopes/chirp rates of the DFCW chips based on LFM and PLFM waveforms. The code sets
constructed with DFCW chips based on LFM waveforms occupy different bandwidths but have
smaller peak cross-correlations, while the code sets constructed with DFCW chips based on PLFM
waveforms occupy the same bandwidth but have slightly larger peak cross-correlations.
1.3. Construction of a Good Code Set from the Mates of the Compelemntary Code Pair Resulting in Doubling
the Number of Codes in the Set or a Better Good Code Set with Significaltly Smaller Cross-Correlations
Given a complementary code pair, 𝑎[𝑛] and 𝑏[𝑛], there exists the complementary code pair:
𝑏[−𝑛 + 𝑁𝑔 ] and −𝑎[−𝑛 + 𝑁𝑔 ], such that sum of the cross-correlation of 𝑎[𝑛] with 𝑏[−𝑛 + 𝑁𝑔 ] and
that of 𝑏[𝑛] with −𝑎[−𝑛 + 𝑁𝑔 ] results in complete sidelobe cancellation. The code pair: 𝑏[−𝑛 + 𝑁𝑔 ]
and −𝑎[−𝑛 + 𝑁𝑔 ], are called as the mates [23,24] of the complementary code pair: 𝑎[𝑛] and 𝑏[𝑛].
Thus, two good code sets could be constructed from the complementary code pair and their mates
using the same DFCW waveforms as chips. Since these two good code sets will be quasi-orthogonal
to each other, they could be combined to form a larger good code set with double the number of
codes in the set without affecting the cross-correlation properties. Instead of using all the codes in
the two sets, it is shown that a better good code set with significantly reduced cross-correlation
peaks could be constructed by choosing the best candidates form the two sets.
1.4. Previous Research
In [25,26], continuous frequency LFM and PLFM waveforms were used as chips for 𝑎[𝑛] and
𝑏[𝑛]. The current manuscript uses the same code structure as in [25,26], but DFCW waveforms based
on LFM and PLFM waveforms are used as chips for the complementary code pair. These codes find
utility in more modern digital radar systems like the ones discussed in [17–19]. In addition, this
paper also introduces a method of doubling the number of codes in the set without affecting the
cross-correlation properties or constructing a better good code set with significantly smaller
cross-correlations than the ones introduced in [25,26]. This is achieved by constructing good code
sets from a complementary code pair and their mates, while using the same DFCW waveforms as
chips. This is explained in detail in Section 7.
1.5. Paper Structure
Section 2 explains the two properties of symmetry/anti-symmetry and quasi-orthogonality for
the discrete frequency-coding waveform chips based on LFM and PFLM waveforms. Construction
of the proposed codes using DFCW chips is explained in Section 3. Doppler properties of the
proposed codes are discussed in Section 4. Invariance of the zero-sidelobe region under frequency
selective fading is demonstrated in Section 5. Section 6 shows the construction of good code sets.
Section 7 shows the method of doubling the number of codes in the set or constructing a better good
code set followed by conclusion in Section 8.
2. Symmetrical/Anti-Symmetrical DFCW Chips That Are Quasi-Orthogonal
Let 𝑎(𝑡) and 𝑏(𝑡) be constructed by using 𝑐(𝑡), for 0 ≤ 𝑡 ≤ 𝑇, and 𝛾𝑐(−𝑡) (where 𝛾 = 1 or
−1) as chips for the Golay complementary code pair {𝑎[𝑛], 𝑏[𝑛]} of length 𝑁𝑔 , respectively. 𝑎(𝑡)
and 𝑏(𝑡) can be represented as shown in Equation (3).
𝑍
𝑍
If 𝑎[𝑛] ⇔ 𝐴(𝑧) and 𝑏[𝑛] ⇔ 𝐵(𝑧), then
𝑍
𝑎(𝑚𝑇𝑠 ) ⇔ 𝐴𝑐 (𝑧) = 𝐴(𝑧 𝑁𝑐 )𝐶(𝑧)
𝑍
𝑏(𝑚𝑇𝑠 ) ⇔ 𝐵𝑐 (𝑧) = 𝛾𝐵(𝑧 𝑁𝑐 )𝐶(𝑧 −1 )
(12)
(13)
where 𝑎(𝑚𝑇𝑠 ) and 𝑏(𝑚𝑇𝑠 ) represent the sampled versions of 𝑎(𝑡) and 𝑏(𝑡), respectively, 𝑁𝑐 is
𝑍
the number of samples in 𝐶(𝑚𝑇𝑠 ) ⇔ 𝐶(𝑧) and 𝑚 ∈ {0, ±1, ±2, … , ±∞}.
𝐴𝑐 (𝑧) + 𝑧 −𝐷 𝐵𝑐 (𝑧) represents a code constructed by concatenating 𝐴𝑐 (𝑧) and 𝐵𝑐 (𝑧) with a gap
𝐷 in between. At the receiver, this code is passed into two matched filters: 𝐴𝑐 (𝑧 −1 ) which is the
matched filter of 𝐴𝑐 (𝑧), and 𝐵𝑐 (𝑧 −1 ) which is the MF of 𝐵𝑐 (𝑧). The cross-correlation of 𝐴𝑐 (𝑧) +
𝑧 −𝐷 𝐵𝑐 (𝑧) with 𝐴𝑐 (𝑧 −1 ) results in,
𝑅𝐴𝑐 (𝑧) = (𝐴𝑐 (𝑧) + 𝑧 −𝐷 𝐵𝑐 (𝑧))𝐴𝑐 (𝑧 −1 )
(14)
𝑅𝐴𝑐 (𝑧) = 𝐴(𝑧 𝑁𝑐 )𝐶(𝑧)𝐴(𝑧 −𝑁𝑐 )𝐶(𝑧 −1 ) + 𝛾𝑧 −𝐷 𝐵(𝑧 𝑁𝑐 )𝐶(𝑧 −1 )𝐴(𝑧 −𝑁𝑐 )𝐶(𝑧 −1 )
(15)
𝑅𝐴𝑐 (𝑧) = 𝑅𝐴 (𝑧 𝑁𝑐 )𝑅𝑐 (𝑧) + 𝛾𝑧 −𝐷 𝑅𝐵,𝐴 (𝑧 𝑁𝑐 )𝐶 2 (𝑧 −1 )
(16)
where 𝑅𝐴 (𝑧) = 𝐴(𝑧)𝐴(𝑧 −1 ) represents the autocorrelation of 𝑎[𝑛], 𝑅𝑐 (𝑧) = 𝐶(𝑧)𝐶(𝑧 −1 ) represents
the chip ACF (𝐶(𝑧 −1 ) is the MF of 𝐶(𝑧)), 𝑅𝐵,𝐴 (𝑧) = 𝐵(𝑧)𝐴(𝑧 −1 ) represents the cross-correlation of
𝑏[𝑛] with 𝑎[𝑛] and 𝛾𝐶 2 (𝑧 −1 ) represents the cross-correlation of 𝛾𝐶(𝑧 −1 ) with 𝐶(𝑧 −1 ).
Similarly, the cross-correlation of 𝐴𝑐 (𝑧) + 𝑧 −𝐷 𝐵𝑐 (𝑧) with 𝐵𝑐−1 (𝑧) can be represented by,
𝑅𝐵𝑐 (𝑧) = 𝛾𝑅𝐴,𝐵 (𝑧 𝑁𝑐 )𝐶 2 (𝑧) + 𝛾 2 𝑧 −𝐷 𝑅𝐵 (𝑧 𝑁𝑐 )𝑅𝑐 (𝑧)
(17)
where 𝑅𝐴,𝐵 (𝑧) = 𝐴(𝑧)𝐵(𝑧 −1 ) represents the cross-correlation of 𝑎[𝑛] with 𝑏[𝑛] , 𝑅𝐵 (𝑧) =
𝐵(𝑧)𝐵(𝑧 −1 ) represents the ACF of 𝑏[𝑛] and 𝛾𝐶 2 (𝑧) is the cross-correlation of 𝐶(𝑧) with 𝛾𝐶(𝑧).
Delaying 𝑅𝐴𝑐 (𝑧) by 𝐷 and adding it to 𝑅𝐵𝑐 (𝑧) results in,
𝑅𝑠 (𝑧) = 𝑧 −𝐷 𝑅𝐴𝑐 (𝑧) + 𝑅𝐵𝑐 (𝑧)
(18)
Since 𝛾 2 = 1 and 𝑅𝐴 (𝑧) + 𝑅𝐵 (𝑧) = 2𝑁𝑔 , 𝑅𝐴 (𝑧 𝑁𝑐 )𝑅𝑐 (𝑧) + 𝛾 2 𝑅𝐵 (𝑧 𝑁𝑐 )𝑅𝑐 (𝑧) = 2𝑁𝑔 𝑅𝑐 (𝑧), Equation
(17) becomes
𝑅𝑠 (𝑧) = 𝛾𝑅𝐴,𝐵 (𝑧 𝑁𝑐 )𝐶 2 (𝑧) + 2𝑁𝑔 𝑧 −𝐷 𝑅𝑐 (𝑧) + 𝛾𝑧 −2𝐷 𝑅𝐵,𝐴 (𝑧 𝑁𝑐 )𝐶 2 (𝑧 −1 )
(19)
𝑅𝑠 (𝑧) contains the mainlobe ( 2𝑁𝑔 𝑅𝑐 (𝑧) ) that is spread by the chip ACF (i.e., 𝑅𝑐 (𝑧) ), a
zero-sidelobe region on either side of the mainlobe and an adjacent region of cross-correlation
sidelobes (i.e., the first and the last terms in Equation (19)). If 𝛾 = 1, then 𝐶(𝑧) and 𝛾𝐶(𝑧 −1 ) =
𝐶(𝑧 −1 ) represent symmetrical waveforms. If 𝛾 = −1, then 𝐶(𝑧) and 𝛾𝐶(𝑧 −1 ) = −𝐶(𝑧 −1 ) represent
anti-symmetrical waveforms. For both the symmetry ( 𝛾 = 1 ) and anti-symmetry ( 𝛾 = −1 )
conditions, ACF of 𝐶(𝑧), i.e., 𝑅𝑐 (𝑧) = 𝐶(𝑧)𝐶(𝑧 −1 ) is identical to the ACF of 𝛾𝐶(𝑧), i.e., 𝑅𝛾𝑐 (𝑧) =
𝛾 2 𝐶(𝑧 −1 )𝐶(𝑧) = 𝑅𝑐 (𝑧). As a result, exact sidelobe cancellation of the complementary code pair is
achieved resulting in the zero-sidelobe region on either side of the mainlobe as shown in Equation
(19). If 𝐶(𝑧) and 𝛾𝐶(𝑧 −1 ) are also quasi-orthogonal in the convolution sense, i.e., their
cross-correlation peak (as defined in Equation (10)) is small compared to the mainlobe peak (2𝑁𝑔 ),
then the first and the last terms in Equation (19) are small. It should be noted that using a sinusoidal
chip for the complementary code pair will also result in a zero-sidelobe region around the mainlobe
as it satisfies the symmetry/anti-symmetry condition. However, a sinusoidal chip is not
quasi-orthogonal to its symmetrical/anti-symmetrical mirror image. Thus, the proposed codes with
sinusoidal chip achieves zero-sidelobe region but the cross-correlation sidelobes outside the zero
domain are large which is undesirable. Hence, any waveform that satisfies the two conditions,
symmetry/anti-symmetry and quasi-orthogonality, could be used as chips for the complementary
code pair in the proposed code design.
DFCWs based on LFM and PLFM waveforms are constructed by discretizing their
instantaneous frequencies. These will be referred to as Discrete Frequency LFM (DF-LFM) and
Discrete Frequency PLFM (DF-PLFM) waveforms, respectively. The DF-LFM and DF-PLFM
waveforms satisfy the symmetry/anti-symmetry and the quasi-orthogonality conditions required for
the chips in the proposed code design. The following three types of DFCWs are used as chips for the
complementary code pair in the proposed code design:
1.
2.
3.
DF-LFM waveforms.
DF-PLFM waveforms comprised of two up-chirp segments.
DF-PLFM waveforms comprised of an up-chirp followed by a down-chirp as in [21].
Let 𝑢(𝑡) and 𝑑(𝑡) = 𝑢∗ (−𝑡) represent a DF-LFM/DF-PLFM waveform and its symmetrical
mirror image respectively. The DF-LFM waveforms can be represented by,
𝑢(𝑡) =
𝑑(𝑡) =
where 𝑓[𝑛] = 𝑛 , Δ𝑊 =
1
Δ𝑇
𝑇
𝑁−1
1
∑ 𝑒 𝑗2𝜋𝑓[𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
(20)
𝑁−1
1
∑ 𝑒 𝑗2𝜋𝑓[𝑁−1−𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
(21)
, Δ𝑇 = , 𝑛Δ𝑇 ≤ 𝑡𝑛 ≤ (𝑛 + 1)Δ𝑇 and 𝑇 is the time-duration of the
𝑁
DFCW chip.
The DF-PLFM waveforms comprised of two up-chirp segments can be represented by,
1
𝑁
−1
2
∑ 𝑒 𝑗2𝜋𝑓1[𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑢(𝑡) =
1
{
√Δ𝑇
𝑁−1
∑𝑒
𝑛=
𝑁
2
𝑗2𝜋𝑓2 [𝑁−1−𝑛]Δ𝑊𝑡𝑛
(22)
1
∑ 𝑒 𝑗2𝜋𝑓2 [𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑑(𝑡) =
(1−𝑘)2𝑛
𝑁−1
2𝑘(𝑁−1)𝑛
𝑁
(23)
𝑁−1
1
{
where 𝑓1 [𝑛] = 𝑟𝑜𝑢𝑛𝑑 (
𝑁
−1
2
√Δ𝑇
∑𝑒
𝑗2𝜋𝑓1 [𝑁−1−𝑛]Δ𝑊𝑡𝑛
𝑁
𝑛=
2
𝑁
) for 𝑛 = 0, 1, … , − 1, 0 < 𝑘 ≤ 0.5 and 𝑓2 [𝑛] = 𝑟𝑜𝑢𝑛𝑑 ((𝑁 − 1)(1 −
2
𝑁 𝑁
)) for 𝑛 = , + 1, … , 𝑁 − 1.
2 2
The DF-PLFM waveforms comprised of an up-chirp and a down-chirp segment can be
represented by,
1
∑ 𝑒 𝑗2𝜋𝑓1[𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑢(𝑡) =
1
{
√Δ𝑇
1
𝑑(𝑡) =
𝑁
−1
2
(24)
𝑁−1
∑𝑒
𝑛=
𝑗2𝜋𝑓2 [𝑛]Δ𝑊𝑡𝑛
𝑁
2
𝑁
−1
2
𝑁
∑ 𝑒 𝑗2𝜋𝑓2 [ 2 −1−𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
1
√Δ𝑇
(25)
𝑁−1
∑𝑒
𝑗2𝜋𝑓1 [𝑁−1−𝑛]Δ𝑊𝑡𝑛
𝑁
𝑛=
{
2
Figure 3. shows the plots of instantaneous frequencies of the aforementioned DFCW chips. Let
𝑓𝑢 (𝑡) and 𝑓𝑑 (𝑡) represent the instantaneous frequencies of 𝑢(𝑡) and 𝑑(𝑡) , respectively. 𝑅𝑢 (𝑡) ,
𝑅𝑑 (𝑡) and 𝑅𝑢,𝑑 (𝑡) represent the ACF of 𝑢(𝑡), ACF of 𝑑(𝑡) and cross-correlation between 𝑢(𝑡) and
𝑑(𝑡) respectively. They can be represented as shown in Equation (4). Since 𝑢(𝑡) and 𝑑(𝑡) are
symmetrical mirror images of each other, their ACFs are identical, i.e., 𝑅𝑢 (𝑡) = 𝑅𝑑 (𝑡), as shown in
the plots in Figure 4. Since 𝑢(𝑡) and 𝑑(𝑡) are also quasi-orthogonal, max |𝑅𝑢,𝑑 (𝑡)| is very small as
𝑡
shown in their plots in Figure 4. For these plots, 𝑘 = 0.24 and 𝑁 = 32. 𝑘 closes to 0 results in a
DF-LFM/DF-PLFM waveform that is close to a pure sinusoid, while 𝑘 close to 0.5 results in a very
sharp slope.
(N-1)W
(N-1)W
Instantaneous frequency [Hz]
(N-1)W
0
0
0
NT
(a) Time [s]
0
(N/2)T
(b) Time [s]
NT
0
0
(N/2)T
(c) Time [s]
Figure 3. Plots of instantaneous frequencies of (a) discrete frequency linear frequency modulation
(DF-LFM) chips, (b) discrete frequency piecewise linear frequency modulation (DF-PLFM) chips
comprised of two up-chirp segments and (c) DF-PLFM chips comprised of an up-chirp and a
down-chirp segment.
NT
DF-LFM waveform
0
-20
|Ru(t)|
-60
-0.5T
0
0
(a) Time [s]
0.5T
-20
|Rd(t)|
-60
-0.5T
0
0
(b) Time [s]
0.5T
|Ru,d(t)|
-20
-60
-0.5T
0
(c) Time [s]
0.5T
DF-PLFM waveform comprised of two up-chirp segments
0
|Ru(t)|
-20
-60
-0.5T
0
0
(a) Time [s]
0.5T
|Rd(t)|
-20
-60
-0.5T
0
0
(b) Time [s]
0.5T
|Ru,d(t)|
-20
-60
-0.5T
0
(c) Time [s]
0.5T
DF-PLFM waveform comprised of an up-chirp and a down-chirp
0
|Ru(t)|
-20
-60
-0.5T
0
0
(a) Time [s]
0.5T
|Rd(t)|
-20
-60
-0.5T
0
0
(b) Time [s]
0.5T
|Ru,d(t)|
-20
-60
-0.5T
0
(c) Time [s]
0.5T
Figure 4. (a) and (b) Autocorrelation and (c) cross-correlation of the discrete frequency linear
frequency modulation (DF-LFM) and discrete frequency piecewise linear frequency modulation
(DF-PLFM) waveforms.
3. The Proposed Code Based on Complementary Pair Using DF-LFM and DF-PLFM Waveforms
as Chips
Let 𝑎[𝑛] and 𝑏[𝑛] represent a Golay complementary code pair of length 𝑁𝑔 . Let 𝑎(𝑡) and
𝑏(𝑡) represent the waveforms obtained by modulating 𝑢(𝑡) (Equation (20), (22) or (24)) with 𝑎[𝑛]
and 𝑑(𝑡) (Equation (21), (23) or (25)) with 𝑏[𝑛], respectively.
𝑁𝑔
𝑎(𝑡) = ∑ 𝑎[𝑛]𝑢(𝑡 − (𝑛 − 1)𝑇)
𝑛=1
(26)
𝑁𝑔
𝑏(𝑡) = ∑ 𝑏[𝑛]𝑑(𝑡 − (𝑛 − 1)𝑇)
(20)
𝑛=1
where 𝑇 = 𝑁Δ𝑇 is the duration of 𝑢(𝑡) and 𝑑(𝑡).
A new code, 𝑠(𝑡) = 𝑎(𝑡) + 𝑏(𝑡 − 2𝑁𝑔 𝑇), is constructed by concatenating 𝑎(𝑡) and 𝑏(𝑡) with a
delay of 2𝑁𝑔 𝑇 between them, as shown in Figure 5. 𝑠(𝑡) is then carrier (𝑓𝑐 ) modulated before
transmission, resulting in 𝑠𝑚 (𝑡), as shown in Figure 5. The transmitted signal can be represented by,
𝑠𝑚 (𝑡) = (𝑎(𝑡) + 𝑏(𝑡 − 2𝑁𝑔 𝑇)) 𝑒 𝑗2𝜋𝑓𝑐 𝑡
(21)
Figure 6 shows the block diagram of the receiver. The received signal, 𝑥(𝑡) = 𝑠𝑚 (𝑡) + 𝑛(𝑡), is
first carrier demodulated and then bandpass filtered (BPF), giving 𝑥 ′ (𝑡) as shown in Figure 6. Since
both 𝑎(𝑡) and 𝑏(𝑡) occupy the same bandwidth, only one BPF is used. 𝑛~𝑁(0, 𝜎 2 ) represents
additive white Gaussian noise.
𝑥 ′ (𝑡) = 𝑎(𝑡) + 𝑏(𝑡 − 2𝑁𝑔 𝑇) + +𝑛(𝑡)𝑒 −𝑗2𝜋𝑓𝑐𝑡
(22)
In one path, 𝑥 ′ (𝑡) is passed into the matched filter (MF) of 𝑎(𝑡) implemented as a cascade of
two MFs, one for the digital code (𝑎[𝑛]) and the other for the chip (𝑢(𝑡)), as shown in Figure 6.
𝑅𝑥 ′ ,𝑎 (𝑡) = 𝑅𝑎 (𝑡) + 𝑅𝑏,𝑎 (𝑡 − 2𝑁𝑔 𝑇) + 𝑛𝑎 (𝑡)
(30)
where 𝑅𝑎 (𝑡) represents the ACF of 𝑎(𝑡), 𝑅𝑏,𝑎 (𝑡) represents the cross-correlation of 𝑏(𝑡) with 𝑎(𝑡)
and 𝑛𝑎 (𝑡) represents the filtered noise process, 𝑛𝑎 ~𝑁(0, 𝑁𝑔 𝜎 2 ).
𝑁𝑔
𝑁𝑔
𝑅𝑎 (𝑡) = ∑ ∑ 𝑎[𝑛]𝑎 ∗ [𝑁𝑔 − 𝑚] ∫
𝑢(𝜏 − (𝑛 − 1)𝑇)𝑢∗ (𝑡 + 𝜏 + (𝑚 − 1)𝑇)𝑑𝜏
(31)
𝑡
𝑛=1 𝑚=1
𝑁𝑔
𝑡+𝑇
𝑁𝑔
𝑅𝑏,𝑎 (𝑡) = ∑ ∑ 𝑏[𝑛]𝑎∗ [𝑁𝑔 − 𝑚] ∫
𝑡+𝑇
𝑑(𝜏 − (𝑛 − 1)𝑇)𝑢∗ (𝑡 + 𝜏 + (𝑚 − 1)𝑇)𝑑𝜏
(32)
𝑡
𝑛=1 𝑚=1
′
In another path, 𝑥 (𝑡) is passed into the MF of 𝑏(𝑡) which is also implemented as a cascade of
MF of 𝑏[𝑛] and that of 𝑑(𝑡).
𝑅𝑥 ′ ,𝑏 (𝑡) = 𝑅𝑎,𝑏 (𝑡) + 𝑅𝑏 (𝑡 − 2𝑁𝑔 𝑇) + 𝑛𝑏 (𝑡)
(33)
where 𝑅𝑎,𝑏 (𝑡) represents the cross-correlation of 𝑎(𝑡) with 𝑏(𝑡), which can be represented as
shown in Equation (17), 𝑅𝑏 (𝑡) represents the ACF of 𝑏(𝑡) and 𝑛𝑎 (𝑡) represents the filtered noise
process, 𝑛𝑏 ~𝑁(0, 𝑁𝑔 𝜎 2 ).
u (t )
a (t )
a [n ]
b (t )
NgT
a (t )
NgT
NgT
s (t )
b (t )
b [n ]
d (t )
z - 2NgT
sm (t )
e j 2πfc t
Delay
block
Figure 5. Block diagram showing the construction of the proposed code using Golay complementary
code pair with DFCW chips.
(a )
Rx',a(t-2NgT)
1
x (t )
e - j 2πfc t
BPF
1
(c )
x' (t )
a* [Ng-n ]
u* (T - t )
z - 2 N gT
Digital
code MF
DFCW
chip MF
Delay
block
0.1
1
1
R x ,a (t )
1
0.1
0.1
0
(b )
R (t )
1
0.1
b* [Ng-n ]
d* (T - t )
Digital
code MF
DFCW
chip MF
R x ,b (t )
Figure 6. Receiver block diagram.
Delaying 𝑅𝑥 ′ ,𝑎 (𝑡) by 2𝑁𝑔 𝑇 (as shown in Figure 6) and adding it to 𝑅𝑥 ′ ,𝑏 (𝑡) results in,
𝑅(𝑡) = 𝑅𝑎,𝑏 (𝑡) + 𝑅𝑎 (𝑡 − 2𝑁𝑔 𝑇) + 𝑅𝑏 (𝑡 − 2𝑁𝑔 𝑇) + 𝑅𝑏,𝑎 (𝑡 − 4𝑁𝑔 𝑇) + 𝑛𝑎 (𝑡 − 2𝑁𝑔 𝑇) + 𝑛𝑏 (𝑡)
(34)
Since,
2𝑁 𝑅 (𝑡), −𝑇 ≤ 𝑡 ≤ 𝑇
𝑅𝑎 (𝑡) + 𝑅𝑏 (𝑡) = { 𝑔 𝑢
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
(35)
where 𝑅𝑢 (𝑡) = 𝑅𝑑 (𝑡) represents the ACF of the DFCW chip.
𝑅(𝑡) = 𝑅𝑎,𝑏 (𝑡) + 2𝑁𝑔 𝑅𝑢 (𝑡 − 2𝑁𝑔 𝑇) + 𝑅𝑏,𝑎 (𝑡 − 4𝑁𝑔 𝑇) + 𝑛𝑎 (𝑡 − 2𝑁𝑔 𝑇) + 𝑛𝑏 (𝑡)
(36)
Thus, the final output, 𝑅(𝑡), (excluding the noise terms) consists of,
𝑅𝑎,𝑏 (𝑡), 0 ≤ 𝑡 < 2𝑁𝑔 𝑇
0, 2𝑁𝑔 𝑇 ≤ 𝑡 < 3𝑁𝑔 𝑇 − 𝑇
𝑅(𝑡) = 2𝑁𝑔 𝑅𝑢 (𝑡), 3𝑁𝑔 𝑇 − 𝑇 ≤ 𝑡 < 3𝑁𝑔 𝑇 + 𝑇
0, 3𝑁𝑔 𝑇 + 𝑇 ≤ 𝑡 < 4𝑁𝑔 𝑇
𝑅
{
𝑏,𝑎 (𝑡), 4𝑁𝑔 𝑇 ≤ 𝑡 < 6𝑁𝑔 𝑇
(37)
In Equation (37), 2𝑁𝑔 𝑅𝑢 (𝑡) is the mainlobe of 𝑅(𝑡) (𝑅𝑢 (𝑡) is the ACF of the DFCW chip, as
shown in the plots in Figure 4). 2𝑁𝑔 𝑇 ≤ 𝑡 ≤ 3𝑁𝑔 𝑇 − 𝑇 and 3𝑁𝑔 𝑇 + 𝑇 ≤ 𝑡 ≤ 4𝑁𝑔 𝑇 are the
zero-sidelobe regions on either side of the mainlobe. 𝑅𝑎,𝑏 (𝑡) and 𝑅𝑏,𝑎 (𝑡) (defined in Equation (32)),
respectively, represent the small cross-correlation sidelobes. Let 𝐶𝐶𝑃 represent the peak
cross-correlation sidelobe. Since, 𝑅𝑎,𝑏 (𝑡) and 𝑅𝑏,𝑎 (𝑡) are mirror images of each other,
𝐶𝐶𝑃 = max |
𝑡
𝑅𝑎,𝑏 (𝑡)
|
2𝑁𝑔
(38)
The table in Figure 7 shows the plots of |𝑅𝑥 ′ ,𝑎 (𝑡 − 2𝑁𝑔 𝑇)|, |𝑅𝑥 ′ ,𝑏 (𝑡)| and |𝑅(𝑡)|, all normalized
by 2𝑁𝑔 and in the absence of noise, for 𝑁𝑔 = 8 and 𝑁 = 16 for the three DFCW chips in Column 1.
3D-plots of the corresponding cross-correlation peaks (𝐶𝐶𝑃) for a range of 𝑁𝑔 and 𝑁 are shown in
Column 2. Clearly, the ACF plots show the zero-sidelobe region around the mainlobe and the
adjacent region of small cross-correlation sidelobes. From the 3D-plots of 𝐶𝐶𝑃 it can be observed
that 𝐶𝐶𝑃 values are very small, and decrease in magnitude as the code length 𝑁𝑔 and the number
of discrete frequencies 𝑁 in the DF-LFM/DF-PLFM chips increase. In addition, the 𝐶𝐶𝑃 values are
slightly larger for the codes constructed using the DFCW chips based on PLFM waveforms
compared to the 𝐶𝐶𝑃 values for the codes constructed using DFCW chips based on LFM waveform,
but still very small (≈ −30 dB).
If −𝑎(𝑡) and 𝑏(𝑡) are concatenated and transmitted at a second frequency band and received
using a receiver similar to the one described in Figure 6, then the final output of this code can be
represented by,
−𝑅𝑎,𝑏 (𝑡), 0 ≤ 𝑡 < 2𝑁𝑔 𝑇
0, 2𝑁𝑔 𝑇 ≤ 𝑡 < 3𝑁𝑔 𝑇 − 𝑇
𝑅′(𝑡) = 2𝑁𝑔 𝑅𝑢 (𝑡), 3𝑁𝑔 𝑇 − 𝑇 ≤ 𝑡 < 3𝑁𝑔 𝑇 + 𝑇
0, 3𝑁𝑔 𝑇 + 𝑇 ≤ 𝑡 < 4𝑁𝑔 𝑇
−𝑅𝑏,𝑎 (𝑡), 4𝑁𝑔 𝑇 ≤ 𝑡 < 6𝑁𝑔 𝑇
{
Column 1
(39)
Column 2
DF-LFM Chips
0
|Rx',a(t-2NgT)|
-20
-35
0
2NgT
0
3NgT
(a) Time [s]
4NgT
|Rx',b(t)|
-20
-70
2NgT
0
3NgT
(b) Time [s]
-40
-10
CCPa,b [dB]
-70
4NgT
-20
-45
-30
-40
-50
-50
|R(t)|
-60
16
-20
64
32
-70
2NgT
3NgT
(c) Time [s]
8
128
N
4NgT
-55
32
64
16
N
g
DF-PLFM chips comprised of two up-chirp segments
0
|Rx',a(t-2NgT)|
-20
-28
-30
-32
2NgT
0
3NgT
(a) Time [s]
4NgT
|Rx',b(t)|
-20
-70
2NgT
0
3NgT
(b) Time [s]
4NgT
CCPa,b [dB]
-70
-34
-10
-36
-20
-38
-30
-40
-40
-42
-50
|R(t)|
-20
0
-60
16
64
-46
32
32
64
-70
2NgT
3NgT
(c) Time [s]
-44
N
4NgT
128
8
-48
Ng
16
DF-PLFM chips comprised of an up-chirp and a down-chirp
0
-28
|Rx',a(t-2NgT)|
-20
-30
-32
0
2NgT
0
3NgT
(a) Time [s]
-10
|Rx',b(t)|
-20
-70
2NgT
0
3NgT
(b) Time [s]
-34
4NgT
4NgT
CCPa,b [dB]
-70
-36
-20
-38
-30
-40
-40
-42
-50
|R(t)|
-20
-60
16
64
-46
32
32
64
-70
2NgT
3NgT
(c) Time [s]
4NgT
-44
N
128
8
16
Ng
-48
Figure 7. Column 1. (a), (b) and (c) Plots in dB of the normalized matched filter (MF) outputs.
Column 2. Plots of the corresponding cross-correlation sidelobe peaks (𝐶𝐶𝑃) vs. code length (𝑁𝑔 ) vs.
number of discrete frequencies in the DF-LFM/DF-PLFM chips (𝑁).
Clearly, adding Equations (37) and (39) results in complete cancellation of the cross-correlation
sidelobes. Thus, achieving complete zero-sidelobes on either side of the mainlobe. This is shown in
the plots of 𝑅(𝑡), 𝑅′(𝑡) and 𝑅(𝑡) + 𝑅′(𝑡) for 𝑁𝑔 = 8 and 𝑁 = 16 in Figure 8. Frequency selective
fading between the two bands could result in inexact cancellation of the cross-correlation sidelobes
that are adjacent to the zero-sidelobe region. However, since the cross-correlation sidelobes are very
small, any inexact cancellation due to fading will not create undesirable high magnitude sidelobes.
0
|R(t)|
-20
-70
2NgT
0
3NgT
(a) Time [s]
4NgT
|R'(t)|
-20
-70
2NgT
0
3NgT
(b) Time [s]
4NgT
-20
-70
|R(t)+R'(t)|
2NgT
Figure 8. Plots in dB of: (a)
3NgT
(c) Time [s]
|𝑅(𝑡)|
𝑁𝑔
; (b)
4NgT
|𝑅′(𝑡) |
|𝑅(𝑡)+𝑅′(𝑡)|
𝑁𝑔
2𝑁𝑔
; and (c)
.
𝑡
The complementary code pair could also be interlaced in the time domain (i.e., 𝑎 ( ) +
𝑇
𝑡
𝑏 ( − 𝑇)). This scheme results in a transmission without gaps, while still using the same frequency
𝑇
band as in the previous scheme. Fishler et al. [18] and Deng [19] describe this scheme using
continuous LFM and PLFM chips. However, for this interlaced code, the zero-sidelobes and the
small cross-correlation sidelobes in the final output are interlaced as well.
4. Doppler Properties
Figure 9a,b shows the 3D ambiguity function (AF) of the proposed code constructed using
DF-LFM chips for 𝑁𝑔 = 8 and 𝑁 = 16 as a function of the normalized Doppler frequency 𝑓𝑑 𝑇. The
AF can be represented by,
∞
𝑅(𝑡, 𝑓) = ∫ 𝑠𝑓 (𝜏)𝑠 ∗ (𝑡 − 𝜏)𝑑𝜏
−∞
(40)
where 𝑠𝑓 (𝑡) = 𝑠(𝑡)𝑒 𝑗2𝜋𝑓𝑑 𝑡 , for 0 ≤ 𝑡 ≤ 3𝑁𝑔 𝑇, is the transmitted code Doppler shifted by 𝑓𝑑 , 𝑇 is the
duration of the DF-LFM/DF-PLFM chips and 𝑠(𝑡) is the proposed code.
Figure 9. Plots of 3D-ambiguity function in (a) linear scale and (b) in dB scale of the proposed codes
constructed using DF-LFM chips.
It can be observed that Doppler shift adversely affects the cancellation of the sidelobes of the
complementary code pair, resulting in the disappearance of the zero-sidelobe domain on either side
of the mainlobe. The ACF mainlobes of the symmetrical chips shift in opposite directions under
Doppler shift. As a result, ACF of the complementary code pair do not align with each other. Thus,
the proposed codes do not have good Doppler properties. One way to improve the Doppler
properties could be to use triangular FM based waveforms as chips for the complementary code pair
in the proposed code structure. Another method to improve the Doppler properties could be to use
the approach discussed in [27] to construct Doppler resilient Golay complementary code pairs.
5. Effect of Frequency Selective Fading
In the presence of frequency selective fading, the higher frequencies within the frequency
sweep (𝑊 = 𝑁(𝑁 − 1)Δ𝑊Δ𝑇) of the DF-LFM/DF-PLFM waveforms could be slightly attenuated
compared to the lower frequencies. However, autocorrelations of the DF-LFM/DF-PLFM waveforms
remain identical, since they are symmetrical. Hence, the zero-sidelobe region is not affected.
However, the adjacent region of small cross-correlation sidelobes could vary slightly depending on
the attenuation due to fading. However, its effect on the error in the location of the mainlobe is
negligible as will be shown in the simulation results later in this section.
In the presence of frequency selective fading, let the lower and the higher frequency
components of the DF-LFM/DF-PLFM chips be attenuated by 𝛼1 and 𝛼2 respectively, where 𝛼2 <
𝛼1 . The DF-LFM chips (𝑢𝑓 (𝑡) and 𝑑𝑓 (𝑡)) in the presence of fading can be represented by,
𝛼1
√Δ𝑇
𝑢𝑓 (𝑡) =
𝛼2
𝑁1 −1
∑ 𝑒 𝑗2𝜋𝑓[𝑛]Δ𝑊𝑡𝑛
𝑛=0
𝑁−1
(41)
∑ 𝑒
{√Δ𝑇 𝑛=𝑁1
𝑁−1−𝑁1
𝛼2
𝑑𝑓 (𝑡) =
𝑗2𝜋𝑓[𝑛]Δ𝑊𝑡𝑛
∑ 𝑒 𝑗2𝜋𝑓[𝑁−1−𝑁1−𝑛]Δ𝑊𝑡𝑛
√Δ𝑇
𝑛=0
𝑁−1
𝛼1
(42)
𝑗2𝜋𝑓[𝑁−1−𝑛]Δ𝑊𝑡𝑛
∑ 𝑒
{ √Δ𝑇 𝑛=𝑁−𝑁1
The DF-PLFM chips comprised of two up-chirp segments in the presence of fading can be
represented by,
𝛼1
𝑁
−1
2
∑ 𝑒 𝑗2𝜋𝑓1 [𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑢𝑓 (𝑡) =
𝑁−1
𝛼2
{
√Δ𝑇
(43)
∑ 𝑒 𝑗2𝜋𝑓2[𝑁−1−𝑛]Δ𝑊𝑡𝑛
𝑁
𝑛=
2
𝑁
−1
2
𝛼2
∑ 𝑒 𝑗2𝜋𝑓2 [𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑑𝑓 (𝑡) =
𝑁−1
𝛼1
√Δ𝑇
∑𝑒
(44)
𝑗2𝜋𝑓1 [𝑁−1−𝑛]Δ𝑊𝑡𝑛
𝑁
𝑛=
{
2
The DF-PLFM chips comprised of an up-chirp and a down-chirp segment in the presence of
fading can be represented by,
𝛼1
𝑁
−1
2
∑ 𝑒 𝑗2𝜋𝑓1 [𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑢𝑓 (𝑡) =
𝛼2
{
√Δ𝑇
(45)
𝑁−1
∑𝑒
𝑁
𝑛=
2
𝑗2𝜋𝑓2 [𝑛]Δ𝑊𝑡𝑛
𝑁
−1
2
𝛼2
𝑑𝑓 (𝑡) =
𝑁
∑ 𝑒 𝑗2𝜋𝑓2[ 2 −1−𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
(46)
𝑁−1
𝛼1
√Δ𝑇
∑𝑒
𝑗2𝜋𝑓1 [𝑁−1−𝑛]Δ𝑊𝑡𝑛
𝑁
𝑛=
{
2
For the DF-LFM waveforms, all the discrete frequencies 0 ≤ 𝑛 ≤ 𝑁1 − 1, are considered as the
lower frequency component which are attenuated by 𝛼1 , while the discrete frequencies 𝑁1 ≤ 𝑛 ≤
𝑁 − 1 will be considered as the higher frequency component which are attenuated by 𝛼2 . For the
𝑁
DF-PLFM waveforms, all the discrete frequencies 0 ≤ 𝑛 ≤ − 1 are considered as the lower
frequency component and
𝑁
2
2
≤ 𝑛 ≤ 𝑁 − 1, the higher frequency component. Figure 10 shows the
autocorrelations of the DF-LFM chips with and without fading, i.e., 𝑢𝑓 (𝑡) and 𝑑𝑓 (𝑡) and 𝑢(𝑡) and
𝑑(𝑡) for 𝑁 = 16 , 𝑁1 = 12 , 𝑁𝑔 = 16 , 𝛼1 = 0.9792 and 𝛼2 = 0.8470 ( 𝛼1 and 𝛼2 are randomly
generated). Although, fading results in reducing the mainlobe peaks of the autocorrelations,
|𝑅𝑢𝑓 (𝑡)| and |𝑅𝑑𝑓 (𝑡)| (as seen in the plots), they remain identical.
0
0
-20
|Ru(t)|
-60
-20
|Ru (t)|
f
-60
-0.5T
0
0
0.5T
(a) Time [s]
-20
-0.5T
0
|Rd(t)|
-60
0
0.5T
(c) Time [s]
-20
|Rd (t)|
f
-60
-0.5T
0
0.5T
(b) Time [s]
-0.5T
0
0.5T
(d) Time [s]
Figure 10. (a,b) Plots in dB of |𝑅𝑢 (𝑡)| and |𝑅𝑑 (𝑡)|; and (c,d) plots of |𝑅𝑢𝑓 (𝑡)| and |𝑅𝑑𝑓 (𝑡)|.
Figure 11 shows the matched filter outputs, 𝑅𝑥 ′ ,𝑎 (𝑡) and 𝑅𝑥 ′ ,𝑏 (𝑡), and their sum 𝑅(𝑡) (given
by Equations (30), (33) and (37), respectively), in the presence of frequency selective fading. Clearly,
the zero-sidelobe region on either side of the mainlobe is not affected by fading.
0
|Rx',a(t-2NgT)|
-20
-70
2NgT
3NgT
(a) Time [s]
0
4NgT
|Rx',b(t)|
-20
-70
2NgT
3NgT
(b) Time [s]
0
4NgT
|R(t)|
-20
-70
Figure 11. Plots in dB of: (a)
2NgT
3NgT
(c) Time [s]
4NgT
𝑅𝑥′ ,𝑎 (𝑡−2𝑁𝑔 𝑇)
𝑅𝑥′ ,𝑏 (𝑡)
|𝑅(𝑡)|
𝑁𝑔
𝑁𝑔
2𝑁𝑔
; (b)
; and (c)
, in the presence of frequency selective fading.
Root mean square error (𝑅𝑀𝑆𝐸) in the location of the mainlobe is used as a metric to measure of
effectiveness in sidelobe cancellation in the presence of frequency selective fading and noise.
Location of the mainlobe of 𝑅(𝑡) is given by,
𝑅(𝑡)
𝑡̂ = argmax |
|
(47)
2𝑁𝑔
𝑡
where 𝑅(𝑡) is the final output, given by Equation (37).
𝑅𝑀𝑆𝐸 in the location of the mainlobe can be represented by,
𝐾
1 1
𝑅𝑀𝑆𝐸 = ( ∑(𝑡̂𝑖 − 𝑡)2 )
𝑇 𝐾
0.5
(48)
𝑖=1
where 𝐾 is the number of trials, 𝑡 is the actual timing of the mainlobe, 𝑡𝑖̂ is the estimated timing of
the mainlobe for the ith trial in the presence of noise and 𝑇 is the chip duration.
𝑅𝑀𝑆𝐸 vs. signal-to-noise ratio (SNR) plots for the proposed codes are shown in Figure 12.
DF-LFM chips under frequency selective fading
10
10
RMSE
10
10
10
10
10
1
Ng=16, N=32,
Without fading
Ng=16, N=32,
With fading
0
-1
-2
-3
-4
-5
-4
-2
0
2
4
SNR [dB]
6
8
10
12
DF-PLFM chips comprised of two up-chirp segments under frequency selective fading
10
10
RMSE
10
10
10
10
10
10
1
0
Ng=16, N=32,
Without fading
-1
Ng=16, N=32,
With fading
-2
-3
-4
-5
-6
-4
-2
0
2
4
SNR [dB]
6
8
10
12
DF-PLFM chips comprised of an up-chirp and a down-chirp segment under frequency selective fading
10
10
RMSE
10
10
10
10
10
10
1
Ng=16, N=32,
Without fading
0
Ng=16, N=32,
With fading
-1
-2
-3
-4
-5
-6
-4
-2
0
2
4
SNR [dB]
6
8
10
12
Figure 12. Plots of root mean square error (𝑅𝑀𝑆𝐸) vs. SNR for the proposed codes.
SNR is given by 20 log10
2𝑁𝑔
𝜎2
, where 2𝑁𝑔 is the code energy and 𝜎 2 represents the noise power.
SNR is varied by changing 𝜎 2 and 𝐾 = 1000 iterations are performed for each SNR value. 𝛼1 and
𝛼2 are randomly chosen such that 0.6 ≤ 𝛼1 , 𝛼2 ≤ 1. From these plots it can be observed that the
RMSE curve with fading stays very close to the 𝑅𝑀𝑆𝐸 in the absence of fading. This is true for the
DF-LFM and the DF-PLFM chips.
As discussed in the Introduction, complete sidelobe cancellation can be achieved using a
complementary code pair. However, it requires transmission and reception of the complementary
code pair at two frequencies/channels as shown in Figure 1. In the presence of frequency selective
fading, the cancellation of sidelobes is inexact which results in increasing the probability of false
alarm. This can be seen in the plots of 𝑅𝑀𝑆𝐸 vs. SNR in Figure 13 for a Golay complementary code
pair (𝐴, 𝐵) of length 𝑁𝑔 = 16 and 64. 𝐴 is attenuated by 𝛼1 and 𝐵 is attenuated by 𝛼2 such that
0.6 ≤ 𝛼1 , 𝛼2 ≤ 1. It can be observed that frequency selective fading results in increasing the RMSE in
the location of the mainlobe. The proposed codes are resistant to the effect of frequency selective
fading, as shown in the 𝑅𝑀𝑆𝐸 plots in Figure 12.
10
1
Without fading, Ng=16
10
0
With fading, Ng=16
Without fading, Ng=64
RMSE
10
10
10
10
10
-1
With fading, Ng=64
-2
-3
-4
-5
-4
-2
0
2
4
6
8
SNR [dB]
10
12
14
16
Figure 13. Plot of root mean square error (𝑅𝑀𝑆𝐸) vs. signal-to-noise ratio (SNR) for complementary
code pair using sinusoidal chip.
6. Good Code Sets from Complementary Pair Using DF-LFM and DF-PLFM Waveforms as Chips
A good code set is constructed by changing the slopes of the instantaneous frequency (shown in
the plots in Figure 3) of the DF-LFM/DF-PLFM waveforms used as chips for the same
complementary code pair. The parameter 𝛾𝑙 , as shown in the Equations (20)-(25), is varied carefully
to result in DF-LFM and DF-PLFM chips with different slopes. The resulting good code set has a
zero domain on either side of the mainlobe and an adjacent region of small sidelobes. The
cross-correlation between any code pair in the set is very small. The lth code in the set can be
represented by,
𝑠𝑙 (𝑡) = 𝑎𝑙 (𝑡) + 𝑏𝑙 (𝑡 − 2𝑁𝑔 𝑇)
𝑁𝑔
∑𝑛=1
𝑎[𝑛]𝑢𝑙 (𝑡
(49)
𝑁𝑔
∑𝑛=1
𝑏[𝑛]𝑑𝑙 (𝑡
where 𝑎𝑙 (𝑡) =
− (𝑛 − 1)𝑇) , 𝑏𝑙 (𝑡) =
− (𝑛 − 1)𝑇) , and 𝑢𝑙 (𝑡) and 𝑑𝑙 (𝑡)
represent the two symmetrical DF-LFM/DF-PLFM waveforms.
𝑢𝑙 (𝑡) and 𝑑𝑙 (𝑡) for the good code sets from complementary pairs using DF-LFM chips can be
represented by,
𝑢𝑙 (𝑡) =
𝑑𝑙 (𝑡) =
1
𝑁−1
∑ 𝑒 𝑗2𝜋𝑓𝑙[𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
1
(50)
𝑁−1
∑ 𝑒 𝑗2𝜋𝑓𝑙 [𝑁𝑙 −1−𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
(51)
where 𝑓𝑙 [𝑛] = 𝑟𝑜𝑢𝑛𝑑 (
𝑘 ∈ (0, 0.5], Δ𝑊 =
1
Δ𝑇
2𝛾𝑙 (𝑁−1)𝑛
𝑁
) for 𝑛 = 0, 1, 2, … , 𝑁 − 1 , 𝛾𝑙 = 𝑘 +
(𝑙−1)(1−3𝑘)
𝐿−1
, 𝑙 = 0, 1, 2, … , 𝐿 − 1 ,
and 𝑛Δ𝑇 ≤ 𝑡𝑛 ≤ (𝑛 + 1)Δ𝑇.
𝑢𝑙 (𝑡) and 𝑑𝑙 (𝑡) for the good code sets from complementary pairs using DF-PLFM chips
comprised of two up-chirps can be represented by,
𝑁
−1
2
1
∑ 𝑒 𝑗2𝜋𝑓𝑙,1[𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑢𝑙 (𝑡) =
{
√Δ𝑇
∑𝑒
𝑛=
1
1)(1 −
(1−𝛾𝑙 )2𝑛
𝑁−1
𝑁
𝑁 𝑁
𝑁
2
𝑁
−1
2
(53)
𝑁−1
1
{
2𝛾𝑙 (𝑁−1)𝑛
𝑗2𝜋𝑓𝑙,2 [𝑁−1−𝑛]Δ𝑊𝑡𝑛
∑ 𝑒 𝑗2𝜋𝑓𝑙,2[𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑑𝑙 (𝑡) =
where 𝑓𝑙,1 [𝑛] = 𝑟𝑜𝑢𝑛𝑑 (
(52)
𝑁−1
1
√Δ𝑇
∑𝑒
𝑗2𝜋𝑓𝑙,1 [𝑁−1−𝑛]Δ𝑊𝑡𝑛
𝑁
𝑛=
2
𝑁
(𝑙−1)(1−2𝑘)
2
𝐿−1
) for 𝑛 = 0, 1, 2, … , − 1, 𝛾𝑙 = 𝑘 +
, 𝑓𝑙,2 [𝑛] = 𝑟𝑜𝑢𝑛𝑑 ((𝑁 −
)) for 𝑛 = , + 1, … , 𝑁 − 1 and 𝑘 ∈ (0, 0.5].
2 2
𝑢𝑙 (𝑡) and 𝑑𝑙 (𝑡) for the good code sets from complementary pairs using DF-PLFM chips
comprised of an up-chirp and a down-chirp are given by,
𝑁
−1
2
1
∑ 𝑒 𝑗2𝜋𝑓𝑙,1[𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
𝑢𝑙 (𝑡) =
{
√Δ𝑇
1
𝑑𝑙 (𝑡) =
𝑁−1
1
∑𝑒
𝑛=
𝑁
2
𝑁
−1
2
∑ 𝑒 𝑗2𝜋𝑓𝑙,2[𝑁𝑙−1−𝑛]Δ𝑊𝑡𝑛
√Δ𝑇 𝑛=0
1
√Δ𝑇
(54)
𝑗2𝜋𝑓𝑙,2 [𝑛]Δ𝑊𝑡𝑛
𝑁−1
∑𝑒
(55)
𝑗2𝜋𝑓𝑙,1 [𝑁−1−𝑛]Δ𝑊𝑡𝑛
𝑁
𝑛=
{
2
The receiver for each code is as shown in Figure 6. The ACFs of all the codes in the set are as
shown in the plots in Figure 7. Thus, the autocorrelation sidelobe peak (𝐴𝑆𝑃𝑙 ) of the 𝑙𝑡ℎ code in the
set is given by,
𝐴𝑆𝑃𝑙 = 𝐶𝐶𝑃𝑎𝑙,𝑏𝑙
(56)
where 𝐶𝐶𝑃𝑎𝑙 ,𝑏𝑙 is the same as 𝐶𝐶𝑃𝑎,𝑏 which is defined in Equation (38) and shown in the plots in
Figure 7.
Maximum cross-correlation peak (𝑀𝐶𝐶𝑃) and average cross-correlation peak (𝐴𝐶𝐶𝑃) are used
as measures to compare the different good code sets introduced in this paper.
𝑀𝐶𝐶𝑃 =
𝐴𝐶𝐶𝑃 =
max {max|𝑅𝑠𝑙,𝑠𝑘 (𝑡)|}
𝑙,𝑘=1,2,..,𝐿
𝑡
∑𝐿𝑙=1 ∑𝐿𝑘=𝑙 max|𝑅𝑠𝑙 ,𝑠𝑘 (𝑡)|
𝑡
(𝐿2)
(57)
(58)
where 𝑅𝑠𝑙,𝑠𝑘 (𝑡) is the normalized cross-correlation between the code pair {𝑙, 𝑘} in the set.
The table (Column 2) in Figure 14 shows the plots of 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 for each set for different
values of 𝑁, 𝑁𝑔 and 𝐿.
(Column 1) Plots of 𝑓𝑢𝑙 (𝑡), 𝑓𝑑𝑙 (𝑡) vs. time
(Column 2) Plots of 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 vs. 𝐿 for
(𝑓𝑢𝑙 (𝑡) are shown in solid lines and𝑓𝑑𝑙 (𝑡) in
different values of 𝑁𝑔 and 𝑁
lines with dashes)
Good code sets from complementary pairs using DF-LFM waveforms as chips
0
0
N=64, Ng=8
-10
-20
-30
-40
4
5
0
0
NT
MCCP [dB]
0
-10
7
5
6
(c) L
0
-20
-30
6
(b) L
-30
-40
4
N=64, Ng=16
5
N=128, Ng=8
-20
8
N=128, Ng=16
-40
4
𝑘 = 0.24 and 𝑁 = 32
6
(a) L
N=64, Ng=8
-10
ACCP [dB]
N=128, Ng=8
ACCP [dB]
MCCP [dB]
Instantaneous frequency [Hz]
(N-1)W
7
8
N=64, Ng=16
-10
N=128, Ng=16
-20
-30
-40
4
8
7
5
6
(d) L
7
8
Good code sets from complementary pairs using DF-PLFM waveforms
comprised of two up-chirps as chips
-20
-30
-40
4
0
(N/2)T
NT
MCCP [dB]
0
0
-10
5
6
(a) L
7
-30
5
6
(c) L
7
8
N=64, Ng=16
N=128, Ng=16
-30
6
(b) L
N=128, Ng=8
-20
0
N=64, Ng=16
5
N=64, Ng=8
-10
-40
4
8
-20
-40
4
𝑘 = 0.24 and 𝑁 = 32.
N=128, Ng=8
ACCP [dB]
MCCP [dB]
Instantaneous frequency [Hz]
-10
0
N=64, Ng=8
ACCP [dB]
0
(N-1)W
7
-10
-30
-40
4
8
N=128, Ng=16
-20
5
6
(d) L
7
8
Good code sets from complementary pairs using DF-PLFM waveforms
comprised of an up-chirp and a down-chirp as chips
0
-10
-10
-20
N=64, Ng=8
-30
-40
4
N=128, Ng=8
5
0
(N/2)T
𝑘 = 0.24 and 𝑁 = 32
NT
MCCP [dB]
0
0
ACCP [dB]
0
6
(a) L
7
-20
-40
4
N=64, Ng=16
N=128, Ng=16
5
6
(b) L
N=128, Ng=8
-30
5
0
-10
-30
N=64, Ng=8
-20
-40
4
8
ACCP [dB]
MCCP [dB]
Instantaneous frequency [Hz]
(N-1)W
7
8
6
(c) L
7
8
N=64, Ng=16
-10
N=128, Ng=16
-20
-30
-40
4
5
6
(d) L
7
8
Figure 14. Plots of instantaneous frequencies, i.e., 𝑓𝑢𝑙 (𝑡) and 𝑓𝑑𝑙 (𝑡) vs. time in column 1; and (a), (b)
plots of maximum cross-correlation peaks (𝑀𝐶𝐶𝑃) and (c), (d) average cross-correlation peaks
(𝐴𝐶𝐶𝑃) vs. number of codes (𝐿), for various values of code length (𝑁𝑔 ) and number of discrete
frequencies (𝑁) in column 2.
From the plots in Figure 14 it can be observed that the good code sets constructed using
DF-LFM chips have better 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 properties compared to those constructed using
DF-PLFM chips. However, the time bandwidth products of all the codes in the sets constructed
using DF-PLFM chips is the same (i.e., 𝑁(𝑁 − 1)Δ𝑊Δ𝑇 ), while those of the codes in the set
constructed using DF-LFM chips are different. In addition, for a given complementary code pair
code length (𝑁𝑔 ), the cross-correlation sidelobe peaks could be reduced by increasing the number of
discrete frequencies (𝑁) in the chips.
7. Construction of a Good Code Set from the Mates of a Complementary Code Pair
Consider the complementary code pair: 𝐵(𝑧 −1 ) and −𝐴(𝑧 −1 ). The sum of the cross-correlation
of 𝐴(𝑧) with 𝐵(𝑧 −1 ) and that of 𝐵(𝑧) with −𝐴(𝑧 −1 ) will be,
𝐴(𝑧)𝐵(𝑧) − 𝐵(𝑧)𝐴(𝑧) = 0
(59)
Thus, the code pair: 𝐵(𝑧 −1 ) and −𝐴(𝑧 −1 ), are orthogonal to the complementary code pair:
𝐴(𝑧) and 𝐵(𝑧). Such code pairs are called complementary code pair mates, as described in [23,24].
Let 𝑆1 (𝑧) = 𝐴(𝑧 𝑁𝑐 )𝐶(𝑧) + 𝑧 −𝐷 𝐵(𝑧 𝑁𝑐 )𝛾𝐶(𝑧 −1 ) and 𝑆2 (𝑧) = 𝐵(𝑧 −𝑁𝑐 )𝐶(𝑧) − 𝑧 −𝐷 𝐴(𝑧 −𝑁𝑐 )𝛾𝐶(𝑧 −1 )
be the codes constructed by concatenating the complementary code pair mates (𝐴(𝑧), 𝐵(𝑧) and
𝐵(𝑧 −1 ), −𝐴(𝑧 −1 )) using 𝐶(𝑧) and its symmetrical/anti-symmetrical mirror image (𝛾𝐶(𝑧 −1 ), where
𝛾 = ±1) as chips, as described in Section 2. Their autocorrelations can be represented by,
𝑅𝑆1 (𝑧) = 𝛾𝑅𝐴,𝐵 (𝑧 𝑁𝑐 )𝐶 2 (𝑧) + 2𝑁𝑔 𝑧 −𝐷 𝑅𝑐 (𝑧) + 𝛾𝑧 −2𝐷 𝑅𝐵,𝐴 (𝑧 𝑁𝑐 )𝐶 2 (𝑧 −1 )
𝑅𝑆2 (𝑧) = 𝛾𝑅𝐴,𝐵
(𝑧 −𝑁𝑐 )𝐶 2 (𝑧)
+ 2𝑁𝑔 𝑧
−𝐷
𝑅𝑐 (𝑧) − 𝛾𝑧
−2𝐷
𝑅𝐵,𝐴
(𝑧 −𝑁𝑐 )𝐶 2
(𝑧
−1
(60)
)
(61)
Both the autocorrelations consist of a zero-sidelobe region on either side of the mainlobe and an
adjacent region of small cross-correlation sidelobes, as shown in Equation (37) in Section 2. Their
autocorrelation sidelobe peaks (i.e., 𝐶𝐶𝑃, given by (38)) vary with 𝑁 and 𝑁𝑔 as shown in the plots
in Figure 7. The cross-correlation between 𝑆1 (𝑧) and 𝑆2 (𝑧) can be represented by,
𝑅𝑆1 ,𝑆2 (𝑧) = (𝐴(𝑧 𝑁𝑐 )𝐶(𝑧) + 𝑧 −𝐷 𝐵(𝑧 𝑁𝑐 )𝛾𝐶(𝑧 −1 ))(𝐵(𝑧 𝑁𝑐 )𝐶(𝑧 −1 ) − 𝑧 𝐷 𝐴(𝑧 𝑁𝑐 )𝛾𝐶(𝑧1 ))
(62)
Simplifying Equation (62) results in,
𝑅𝑆1,𝑆2 (𝑧) = −𝐴2 (𝑧 𝑁𝑐 )𝐶 2 (𝑧) + 0 + 𝛾𝑧 −2𝐷 𝐵2 (𝑧 𝑁𝑐 )𝐶 2 (𝑧 −1 )
(63)
Clearly, the cross-correlation consists of a zero-sidelobe region and an adjacent region of small
cross-correlation sidelobes. Thus, 𝑆1 (𝑧) and 𝑆2 (𝑧) are quazi-orthogonal in the convolution sense.
The zero-sidelobe domain in the cross-correlation is due to the symmetry/anti-symmetry property of
the chips. Since, the symmetry/anti-symmetry property is immune to frequency selective fading, the
zero-sidelobe domain in the cross-correlation is also immune to frequency selective fading. Figure 15
show the autocorrelations 𝑆1 (𝑧) and 𝑆2 (𝑧) and their cross-correlation. For these plots, 𝑆1 (𝑧) and
𝑆2 (𝑧) are constructed using DF-LFM chips with 𝑁 = 8 and 𝑁𝑔 = 16. These plots clearly show the
zero-sidelobe domain in the cross-correlation between the proposed code constructed using a
complementary code pair and the one constructed using its mate.
Varying the slopes of the DFCW chips in both 𝑆1 (𝑧) and 𝑆2 (𝑧), as shown in Section 6, results in
two good code sets. Since these two sets are quasi-orthogonal to each other, they could be combined
to form a larger set with twice the number of codes, while maintaining the same cross-correlation
properties as those of the individual sets. This can be seen in the plots of maximum cross-correlation
peaks (𝑀𝐶𝐶𝑃) and average cross-correlation peaks (𝐴𝐶𝐶𝑃) in Column 1 of the table in Figures 16 and
17. When compared to the plots in Figure 14, clearly the cross-correlation properties remain the
same, while doubling the number of the codes in the set.
0
|RS1(z)|
-20
-70
2NgT
0
3NgT
(a) Time [s]
4NgT
|RS2(z)|
-20
-70
2NgT
0
3NgT
(b) Time [s]
4NgT
|RS1,S2(z)|
-20
-70
2NgT
3NgT
(c) Time [s]
4NgT
Figure 15. Plots in dB of: (a) 𝑅𝑆1 (𝑧); (b) 𝑅𝑆2 (𝑧); and (c) 𝑅𝑆1,𝑆2 (𝑧), for 𝑁 = 8 and 𝑁𝑔 = 16 using DF-LFM
chips.
(Column 1) Plots of 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 vs. 𝐿
(Column 2) Plots of 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 vs. 𝐿
for the good code set with twice the number of
for the good codes set with better
codes
cross-correlation properties
Good code sets from complementary pairs and their mates using DF-LFM waveforms as chips
0
-30
-40
8
-40
8
12
(a) L
14
16
0
10
14
16
-20
-30
-30
-40
8
-40
8
12
(b) L
14
16
5
6
(a) L
7
-50
4
8
N=64, Ng=16
12
(d) L
14
16
5
6
(c) L
7
8
N=64, Ng=16
-10
N=128, Ng=16
-30
-40
10
-30
0
-20
-50
4
N=128, Ng=8
-20
-40
-10
N=128, Ng=16
MCCP
-20
10
N=64, Ng=16
-10
ACCP
MCCP
12
(c) L
-50
4
0
N=64, Ng=16
N=128, Ng=16
-30
-40
0
-10
N=128, Ng=8
-20
N=64, Ng=8
-10
ACCP
N=128, Ng=8
-20
-30
N=64, Ng=8
-10
ACCP
ACCP
MCCP
-20
10
N=64, Ng=8
-10
N=128, Ng=8
0
0
N=64, Ng=8
MCCP
0
-10
N=128, Ng=16
-20
-30
-40
5
6
(b) L
7
-50
4
8
5
6
(d) L
7
8
Good code sets from complementary pairs and their mates using DF-PLFM waveforms
comprised of two up-chirps as chips
-40
8
-40
8
14
16
10
12
(c) L
14
16
N=128, Ng=16
N=64, Ng=16
-10
-20
-20
-40
-40
8
-40
8
-50
4
12
(b) L
14
16
10
12
(d) L
14
16
ACCP
6
(a) L
7
-50
4
8
N=64, Ng=16
5
6
(c) L
7
8
N=64, Ng=16
-10
N=128, Ng=16
-30
-30
-30
0
-20
-30
10
5
-10
N=128, Ng=16
N=128, Ng=8
-20
-40
0
0
N=64, Ng=16
-30
-40
MCCP
MCCP
-10
12
(a) L
N=128, Ng=8
-20
-50
4
N=64, Ng=8
-10
ACCP
-30
10
N=128, Ng=8
-20
-30
0
N=64, Ng=8
-10
MCCP
-20
0
0
N=64, Ng=8
-10
N=128, Ng=8
ACCP
MCCP
-10
0
N=64, Ng=8
ACCP
0
N=128, Ng=16
-20
-30
-40
5
6
(b) L
7
8
-50
4
Figure 16. Plots of (a), (b) 𝑀𝐶𝐶𝑃 vs. 𝐿 and (c), (d) 𝐴𝐶𝐶𝑃 vs. 𝐿.
5
6
(d) L
7
8
(Column 1) Plots of 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 vs. 𝐿
(Column 2) Plots of 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 vs. 𝐿
for the good code set with twice the number of
for the good codes set with better
codes
cross-correlation properties
Good code sets from complementary pairs and their mates using DF-PLFM waveforms
comprised of an up-chirp and a down-chirp as chips
-40
8
-10
10
12
(a) L
14
16
10
0
N=64, Ng=16
N=128, Ng=16
-30
12
(c) L
14
16
-20
-40
8
-40
8
-50
4
16
10
12
(d) L
14
16
7
-30
-50
4
8
5
6
(c) L
7
8
0
N=64, Ng=16
-10
N=128, Ng=16
-30
-40
14
6
(a) L
N=64, Ng=16
-20
-30
12
(b) L
5
-10
N=128, Ng=16
N=128, Ng=8
-20
-40
0
-30
10
ACCP
MCCP
-20
-50
4
N=64, Ng=16
-10
-20
N=128, Ng=8
ACCP
-40
8
N=64, Ng=8
-10
-40
MCCP
-30
N=64, Ng=8
-10
N=128, Ng=8
-20
-30
0
0
N=64, Ng=8
-10
-20
0
MCCP
N=128, Ng=8
ACCP
MCCP
-10
0
N=64, Ng=8
ACCP
0
N=128, Ng=16
-20
-30
-40
5
6
(b) L
7
8
-50
4
5
6
(d) L
7
8
Figure 17. Plots of (a), (b) 𝑀𝐶𝐶𝑃 vs. 𝐿 and (c), (d) 𝐴𝐶𝐶𝑃 vs. 𝐿.
If a better good code set with the smaller 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 is required, then the best
candidates from the two sets could be selected. This could be achieved by using only half the slope
options for the DFCW waveforms used as chips for the complementary code pair and its mates i.e.,
the DF-LFM/DF-PLFM waveforms with 𝑙 = 0, 2, … , 𝐿 − 1 in Equation (52)–(57). This results in a
good code set with significantly better 𝑀𝐶𝐶𝑃 and 𝐴𝐶𝐶𝑃 values, as shown in the plots in column 2
of the table in Figures 16 and 17. These are significantly smaller than the 𝑀𝐶𝐶𝑃 and the 𝐴𝐶𝐶𝑃
values in Figure 14 for the same number of codes in the set. Thus, constructing the proposed codes
from complementary code pair and their mates, while using the same chips for both pairs, offers the
aforementioned flexibility of choosing between a better good code set and a larger good code set.
Table 1 shows the average autocorrelation and average cross-correlation values for the Deng’s
polyphase [28] good code set and the Deng’s discrete frequency [20] set designs. The autocorrelation
and cross-correlation sidelobe peaks of the codes introduced in this paper are smaller, as shown in
the plots in Figures 7 and 17 respectively. The proposed codes constructed via concatenating biphase
Golay complementary code pairs and using discrete frequency chips results in a zero-sidelobe
domain and an adjacent region of very small cross-correlation sidelobes. In [28], it is shown that the
average autocorrelation sidelobe and cross-correlation peaks decrease with increase in code length
1
and approach 𝑂( ) for large 𝑁. Their discrete frequency good code set [20] being wideband,
√𝑁
achieve much smaller sidelobe peaks. In the case of the good code set design proposed in this paper,
for a given code length (𝑁𝑔 ), increasing the number of discrete frequencies (𝑁) in the chips results in
a better code set.
Good code set
Deng’s polyphase set (code length
= 128 and 𝐿 = 3 codes)
Deng’s discrete frequency set (# of
discrete frequencies = 128 and 𝐿 =
3 codes)
Proposed code set with 𝑁𝑔 = 16,
𝑁 = 32 and 𝐿 = 4
Average autocorrelation peak
(dB)
Average cross-correlation
peak (dB)
−20.9606
−19.1525
−32.2641
−32.2522
−39.8280
−37.9926
Table 1. Table containing the average autocorrelation and cross-correlation peaks for the proposed
set in comparison with Deng’s polyphase and discrete frequency code sets.
8. Conclusions
In this paper, it is shown that, by replacing the sinusoidal chips in the complementary code pair
with waveforms that satisfy two conditions, symmetry/anti-symmetry and quazi-orthogonality in
the convolution sense, allows for concatenating them in the time domain using one frequency
band/channel. This results in a zero sidelobe domain around the mainlobe and an adjacent region of
small cross-correlation sidelobes. Symmetry/anti-symmetry property results in the zero-sidelobe
region, while the quasi-orthogonality property makes the adjacent region of cross-correlation
sidelobes small. Discrete frequency coding waveform (DFCW) based on LFM and PLFM waveforms
are used as chips, since they satisfy both the symmetry/anti-symmetry and quasi-orthogonality
conditions. Since, frequency selective fading does not affect the symmetry/anti-symmetry property,
the zero-sidelobe region is resistant to the effects of fading. A good code set with a zero-sidelobe
region is then constructed by varying the slopes of the DFCW chips, while using the same
complementary code pair. Such good code sets are constructed using a set of quasi-orthogonal
DFCW chips based on LFM and PLFM waveforms. It is also shown that mates of the complementary
code pair could be used to generate a second good code set using the same DFCW chips. These two
sets are shown to be quasi-orthogonal with a zero-sidelobe domain. Thus, resulting in a larger good
code set with twice the number of codes, while maintaining the same cross-correlation properties. Or
a better good code set could be constructed by choosing the best candidates from the two sets.
As a topic for further research, the proposed codes could be constructed using polyphase
versions (as shown in [29,30]) of the LFM/PLFM waveforms as chips with the complementary code
pairs and their mates. Exploring other waveforms or codes that satisfy the two conditions of
symmetry/anti-symmetry and quazi-orthogonality and improving the Doppler properties of these
codes are the other topics of future research.
Author Contributions: Ravi Kadlimatti and Adly T. Fam conceived and designed the proposed good code set.
Ravi Kadlimatti performed the simulations.
Conflicts of Interest: The authors declare no conflict of interest.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
Frank, R.C. Polyphase codes with good nonperiodic correlation properties. IEEE Trans. Inf. Theory 1963, 9,
43–45.
Costas; J.P. A study of a class of detection waveforms having nearly ideal range-Doppler ambiguity
properties. Proc. IEEE 1984, 72, 996—1009.
Golomb, S.W.; Taylor, H. Construction and properties of Costas arrays. Proc. IEEE 1984, 72, 1143—1163.
Barker, L.; Drakakis, K.; Rickard, S. On the Complexity of the Verification of the Costas Property. Proc.
IEEE 2009, 97, 586–593.
Fam, A.T.; Sarkar, I.; Poonnen, T. Area and power efficient mismatched filters based on sidelobe inversion.
In Proceedings of the IEEE 2008 Radar Conference, Rome, Italy, 26–30 May 2008; pp. 1–6.
Akbaripour, A; Bastani, M.H. Range Sidelobe Reduction Filter Design for Binary Coded Pulse
Compression System. IEEE Trans. Aerosp. Electron. Syst. 2012, 48, 348–359.
Jung, K.T.; Kim, C.J.; Lim, C.H.; Lee, H.S.; Kwag, Y.K. Design of optimum mean square sidelobe
suppression filters for Barker codes. In Proceedings of the 1992 International Conference on Radar,
Brighton, UK, 12–13 October 1992; pp. 530–533.
Rihaczek, A.W.; Golden, R.M. Range Sidelobe Suppression for Barker Codes. IEEE Trans. Aerosp. Electron.
Syst. 1971, AES-7, 1087–1092.
Fam, A.T.; Qazi, F.A; Kadlimatti, R. Zero Sidelobe Aperiodic Codes via Additive-Multiplicative
Mismatched Filtering. In Proceedings of the 2013 IEEE Military Communications Conference (MILCOM
2013), San Diego, CA, USA, 18–20 November 2013; pp. 829–836.
Golay, M.J.E. Complementary series. IRE Trans. Inf. Theory 1961, 7, 82–87.
Frank, R. Polyphase complementary codes. IEEE Trans. Inf. Theory 1980, 26, 641–647.
Jafari, A.H.; O′Farrell, T. Performance Evaluation of Spatial Complementary Code Keying Modulation in
MIMO Systems. In Proceedings of the 2015 IEEE 81st Vehicular Technology Conference (VTC Spring),
Glasgow, UK, 11–14 May 2015; pp. 1–5.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
28.
29.
30.
Galati, G.; Pavan, G. Range Sidelobes Suppression in Pulse-Compression Radar using Golay Pairs: Some
Basic Limitations for Complex Targets. IEEE Trans. Aerosp. Electron. Syst. 2012, 48, 2756–2760.
Liu, Z.; Parampalli, U.; Guan, Y.L. Optimal Odd-Length Binary Z-Complementary Pairs. IEEE Trans. Inf.
Theory 2014, 60, 5768–5781.
Tang, X.; Mow, W.H. Design of spreading codes for quasi-synchronous CDMA with intercell interference.
IEEE J. Sel. Areas Commun. 2006, 24, 84–93.
Li, D. The perspectives of large area synchronous CDMA technology for the fourth-generation mobile
radio. IEEE Commun. Mag. 2003, 41, 114–118.
Qazi, F.A.; Fam, A.T. Good code sets based on Piecewise Linear FM. In Proceedings of the 2012 IEEE Radar
Conference, Atlanta, GA, USA, 7–11 May 2012; pp. 522–527.
Fishler, E.; Haimovich, A.; Blum, R.; Chizhik, D.; Cimini, L.; Valenzuela, R. MIMO radar: An idea whose
time has come. In Proceedings of the 2004 IEEE Radar Conference, Philadelphia, PA, USA, 26–29 April
2004; pp. 71–78.
Deng, H. Orthogonal netted radar systems. IEEE Aerosp. Electron. Syst. Mag. 2012, 27, 28–35.
Deng, H. Discrete frequency-coding waveform design for netted radar systems. IEEE Signal Proc. Lett.
2004, 11, 179–182.
Qazi, F.A.; Fam, A.T. Discrete Frequency-Coding Waveform sets based on Piecewise Linear FM. In
Proceedings of the 2014 IEEE Radar Conference, Cincinnati, OH, USA, 19–23 May 2014; pp. 469–473.
Liu, B.; He, Z.; He, Q. Optimization of Orthogonal Discrete Frequency-Coding Waveform Based on
Modified Genetic Algorithm for MIMO Radar. In Proceedings of the 2007 International Conference of
Communications, Circuits and Systems (ICCCAS), Kokura, Japan, 11–13 July 2007; pp. 966–970.
Harris. F.; Dick. C. A versatile Filter Structure to Generate and Compress Binary and Polyphase
Complementary Spreading Codes. In Proceedings of the 2004 SDR Technical Conference and Product
Exposition, Phoenix, AZ, USA, 15–18 November 2004.
Suehiro, N. Complete complementary code composed of N-multiple-shift orthogonal sequences. IECE
Trans. 1982, J65 A, 1247–1253.
Fam, A.T.; Kadlimatti, R. Complementary code pairs that share the same bandwidth via symmetrical
linear FM chips. In Proceedings of the 2015 IEEE Military Communications Conference (MILCOM),
Tampa, FL, USA, 26–28 October 2015; pp.820–825.
Kadlimatti, R.; Fam, A.T. Good code sets from complementary pairs via symmetrical/anti-symmetrical
chips. IEEE Trans. Aerosp. Electron. Syst. 2016, 52, 1327–1339.
Pezeshki, A.; Calderbank, A.R.; Moran, W.; Howard, S.D. Doppler Resilient Golay Complementary
Waveforms. IEEE Trans. Inf. Theory 2008, 54, 4254–4266.
Deng, H. Polyphase code design for Orthogonal Netted Radar systems. IEEE Trans. Signal Proc. 2004, 52,
3126–3135.
Lewis, B.L.; Kretschmer, F.F. Linear Frequency Modulation Derived Polyphase Pulse Compression Codes.
IEEE Trans. Aerosp. Electron. Syst. 1982, AES-18, 637–641.
Qazi, F.A.; Fam, A.T. Doppler tolerant and detection capable polyphase code sets. IEEE Trans. Aerosp.
Electron. Syst. 2015, 51, 1123–1135.
| 7cs.IT
|
Data-driven and Model-based Verification:
a Bayesian Identification Approach
arXiv:1509.03347v1 [cs.SY] 10 Sep 2015
Sofie Haesaert a, Paul M.J. Van den Hof a, Alessandro Abate b
a
Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands
b
Department of Computer Science, University of Oxford, Oxford, United Kingdom
Abstract
This work develops a measurement-driven and model-based formal verification approach, applicable to systems with partly
unknown dynamics. We provide a principled method, grounded on reachability analysis and on Bayesian inference, to compute
the confidence that a physical system driven by external inputs and accessed under noisy measurements, verifies a temporal
logic property. A case study is discussed, where we investigate the bounded- and unbounded-time safety of a partly unknown
linear time invariant system.
Key words: Temporal logic properties, Bayesian inference, Linear time-invariant models, Model-based verification,
Data-driven validation, Statistical model checking,
1
Introduction
The design of complex, high-tech, safety-critical systems such as autonomous vehicles, intelligent robots,
and cyber-physical infrastructures, demands guarantees
on their correct and reliable behaviour. Correct functioning and reliability over models of systems can be attained by the use of formal methods. Within the computer sciences, the formal verification of software and
hardware has successfully led to industrially relevant and
impactful applications [13]. Carrying the promise of a
decrease in design faults and implementation errors and
of correct-by-design synthesis, the use of formal methods, such as model checking [13], has become a standard
in the avionics, automotive, and railway industries [34].
Life sciences [6,14] and general engineering applications
[5,11] have also recently pursued the extension of these
successful techniques from the computer science: this has
required a shift from finite-state to physical and cyberphysical models that are of practical use in nowadays
science and technology [23,32].
The strength of formal techniques, such as model checking, is bound to the fundamental requirement of having
access to a given model, obtained from the knowledge
Email addresses: [email protected] (Sofie Haesaert),
[email protected] (Paul M.J. Van den Hof),
[email protected] (Alessandro Abate).
of the behaviour of the underlying system of interest. In
practice, for most physical systems the dynamical behaviour is known only in part: this holds in particular
with biological systems [1] or with classes of engineered
systems where, as a consequence, the use of uncertain
control models built from data is a common practice [22].
Only limited work within the formal methods community deals with the verification of models with partly
unknown dynamics. Classical results [4,19] consider the
verification problem for non-stochastic models described
by differential equations and with bounded parametric
uncertainty. Similarly, but for continuous time probabilistic models, [9,10] explore the parameter space with
the objective of model verification (respectively statistical or probabilistic). Whenever full state measurements
of the system are available, Statistical Model Checking (SMC) [31,24] replaces model(-based) checking procedures with empirical testing of formalised properties.
SMC is limited to fully observable stochastic systems
with little or no non-determinism, and may require the
gathering a large set of measurements. Extensions towards the inclusion of non-determinism have been studied in [18,25], with preliminary steps towards Markov decision processes. Related to SMC techniques, but bound
to finite state models, [12,27,30] assume that the system
is encompassed by a finite-state Markov chain and efficiently use data to learn the corresponding model and to
verify it. Similarly, [3,8] employ machine learning tech-
14 September 2015
representing S, is assumed to be an element of G, namely
θ0 ∈ Θ: as an example, model sets G obtained through
first principles adhere to this classical assumption.
niques to infer finite-state Markov models from data over
specific logical formulae.
An alternative approach, allowing both partly unknown
dynamics over uncountable (continuous) variables and
noisy output measurements, is the usage of a Bayesian
framework relating the confidence in a formal property
to the uncertainty of a model built from data. When
applied on nonlinearly parameterised linear time invariant (LTI) models this approach introduces huge computational problems, which as proposed in [16], can only
be mitigated by statistical methods. Instead, to obtain
reliable and numerical solutions, we propose the use of
linearly parameterised model sets defined through orthonormal basis functions to represent these partially
unknown systems. This is a broadly used framework in
system identification [21,22]: it allows for the incorporation of prior knowledge, while maintaining the benefits (computational aspects) of linear parameterisations.
Practically, it has been widely used for the modelling of
physical systems, such as the thermal dynamics of buildings [35]. In contrast, in this paper we pursue a promising new numerical approach: instead of employing directly a nonlinearly parameterised model, we embed it
in a linearly parameterised one via a series expansion of
orthonormal basis functions.
Samples can be drawn from the underlying physical system via a measurement set-up, as depicted in Figure 1.
An experiment consists of a finite number (Ns ) of inputoutput samples drawn from the system, and is denoted
s
by Z Ns = {u(t)ex , ỹ(t)ex }N
t=1 , where u(t)ex ∈ U is the
input for the experiment and ỹ(t)ex is a (possibly noisy)
measurement of y0 (t)ex . In general, the measurement
noise can enter non-additively and be a realisation of a
stationary stochastic process. 1 We assume that at the
beginning of the measurement procedure (say at t = 0),
the initial condition of the system, encompassed by the
initial state x(0)ex of models in M, is either known, or,
when not known, has a structured uncertainty distribution based on the knowledge of past inputs and/or
outputs. As reasonable, we implicitly consider only welldefined problems, such that for any model representing
the system, given a signal input u(t)ex and an (uncertainty distribution for) x(0)ex , the probability density
distribution of the measured signal can be fully characterised.
u(t)ex
In this contribution we further analyse and extend the
related results in [17], obtained for a time-bounded subset of temporal logic properties, to unbounded-time temporal logic properties, and analyse their robustness.
2
S
y0 (t)ex
e(t)
ỹ(t)ex
Fig. 1. System and measurement setup. In the measurement
setup (grey box) the measured output ỹ(t)ex includes the
system output y0 (t)ex and the measurement noise e(t). Data
collected from experiments comprises the input u(t)ex and
the measured output ỹ(t)ex signals.
General Framework and Problem Statement
In this section, we provide a novel methodology to verify
whether a system S satisfies a specification ψ, formulated
in a suitable temporal logic, by integrating the partial
knowledge of the system dynamics with data obtained
from a measurement set-up around the system.
The end objective is to analyse the behaviour of system
S. We consider properties encoded as specifications ψ
and expressed in a temporal logic of choice (to be detailed shortly). Let us remark that the behaviour of S to
be analysed is bound to a set of operating conditions that
are pertinent to the verification problem and that will
be indexed with ver: this comprises the set of possible
input signals u(t)ver (e.g., a white or coloured noise signal, or a non-deterministic signal u(t)ver ∈ Uver ⊆ U),
and of the set of initial states x(0)ver ∈ Xver for the
mathematical models M reflecting past inputs and/or
outputs of the system. The system satisfies a property
if the “true” model representing it satisfies it, namely
S ψ if and only if M(θ0 ) ψ.
Let us further clarify this framework. Let us denote with
S a physical system, or equivalently the associated dynamical behaviour. A signal input u(t) ∈ U, t ∈ N, captures how the environment acts on the system. Similarly,
an output signal y0 (t) ∈ Y indicates how the system interacts with the environment, or alternatively how the
system can be measured. Note that the input and output
signals are assumed to take values over continuous domains. The system dynamics can be described via mathematical models, which express the behavioural relation
between its inputs and outputs. The knowledge of the
behaviour of the system is often limited or uncertain,
making it impossible to analyse its behaviour via that
of a “true” model. In this case, a-priori available knowledge allows to construct a model set G with elements
M ∈ G: this model class supports the structured uncertainty as a distribution over a parameterisation θ ∈ Θ,
G = {M(θ)|θ ∈ Θ}. The unknown “true” model M(θ0 )
1
Both the operating conditions of the experiment, that
is the input signal u(t)ex and the initial state x(0)ex , and
the measurements have been indexed with ex to distinguish
them from the operating conditions of interest for verification, to be discussed shortly.
2
which presumes an uncertainty distribution p (θ) over
the parameter set Θ, representing the prior knowledge.
In this work we consider the satisfaction of a property
M(θ) ψ as a binary-valued mapping from the parameter space Θ. More generally, when in addition to the
measurements of the system also its transitions are disturbed by stochastic noise, then property satisfaction is
a mapping from the parameter space Θ to the interval
[0, 1], and quantifies the probability that the model M(θ)
satisfies the property. This mapping generalises the definition of the satisfaction function introduced in [9], and
is now stated as follows.
The statement can be formally derived based on standard Bayesian calculus, as in [26]. We have chosen to
employ a Bayesian framework, as per (3), since it allows
to reason explicitly over the uncertain knowledge on the
system and to work with the data acquired from the measurement setup. This leads to the efficient incorporation
of the available knowledge and to its combination with
the data acquisition procedure, in order to compute the
confidence on the validity of a given specification over
the underlying system. As a special instance, this result
can be employed for Bayesian hypothesis testing [36].
As long as the mapping fψ is measurable, the models
in the model set (and hence the system represented by
it) can be characterised by either probabilistic or nonprobabilistic dynamics.
Definition 1 (Satisfaction Function) Let G be a set
of models M that is indexed by a parameter θ ∈ Θ,
and let ψ be a formula in a suitable temporal logic. The
satisfaction function fψ : Θ → [0, 1] associated with ψ
is
fψ (θ) = P (M(θ) ψ) .
(1)
Let us assume that the satisfaction function fψ is measurable and entails a decidable verification problem (e.g.,
a model checking procedure) for all θ ∈ Θ.
Remark 2 In statistical model checking [24,31], the
objective is to replace the computationally tolling verification of a system over bounded-time properties by
the empirical (statistical) testing of the relevant specifications over finite executions drawn from the system.
In contrast, our problem statement tackles the problem
of efficiently incorporating data with prior knowledge,
for the formal (deductive) verification of the behaviour
of a system with partly unknown dynamics – as such
our overall verification approach is, as claimed, both
data-driven and model-based. Moreover, by separating
the operational conditions in an experiment from those
of importance for the verification procedure, the system
can be verified over non-deterministic inputs, encompassing as such both controller and disturbance inputs,
or modelling errors.
Problem 1 For a partly unknown physical system S,
under prior knowledge on the system given as a parameterised model class G supporting an uncertainty
distribution over the parameterisation, gather possibly noisy data drawn from the measurement setup and
verify properties on S expressed in a temporal logic of
choice, with a formal quantification of the confidence of
the assertion.
2.1
A Bayesian Framework for Data-driven Modelling
and Verification
Consider Problem 1. Denote loosely with P (·) and p (·)
respectively a probability measure and a probability
density function, both defined over a continuous domain. We employ Bayesian probability calculus [26] to
express the confidence in a property as a measure of the
uncertainty distribution defined the set G. By adopting
the Bayesian framework, uncertainty distributions are
handled as probability distributions of random variables. Therefore the confidence in a property is computed as a probability measure P (·) via the densities
p (·) over the uncertain variables.
2.2
The Bayesian approach is widely applicable to different
types of properties and models, however its computational complexity might in practice limit its implementation. In the literature the satisfaction function is related to the exploration of a parameter set over the validity of a formal property fψ (θ), and has been studied for
autonomous models in continuous time in [4,15,19]. Analytical solutions to the parametric inference equation
(3) can be found if the prior is a conjugate distribution.
For linear dynamical systems, closed-form solutions are
given inter alia in [28]. In general (2)-(3) in Proposition 1
lack analytical solutions, and the assessment of the satisfaction function (1) may be computationally intensive.
Statistical methods such as the one proposed in [16] on
a similar Bayesian approach lead to involved computations and introduce additional uncertainty from Monte
Carlo techniques.
Proposition 1 (Bayesian Confidence) Given
a
specification ψ and a data set Z Ns , the confidence that
S ψ can be quantified via inference as
R
(2)
P S ψ | Z Ns = Θ fψ (θ)p θ|Z Ns dθ .
where fψ is the satisfaction function given in (1).
The
a-posteriori uncertainty distribution p θ|Z Ns , given
the data set Z Ns , is based on parametric inference over
θ as
p(Z Ns |θ )p(θ)
p θ|Z Ns = R
,
(3)
Ns
Θ
p(Z
Computational Approaches
On the contrary, in the next section, we propose a novel
computational approach over discrete-time linear time-
|θ)p(θ)dθ
3
Denote the k-bounded and unbounded invariance operVk
ator as ✷k ψ = i=0 i ψ and ✷ψ = ¬(true U ¬ψ), respectively.
invariant systems. By exploiting linear parameterisations analytical solutions of both the parametric inference and the satisfaction function are characterised for
properties expressed within a fragment of a temporal
logic.
3
Of interest are formal properties encoded on the inputoutput behaviour of the system, and over a time horizon
t ≥ 0. The output y0 (t)ver ∈ Y is labeled by a map
L : Y → Σ, which assigns letters α in the alphabet Σ via
half spaces on the output, as
LTL Verification of LTI systems
Consider a system S that can be represented by a class
of finite-dimensional dynamical models that evolve in
discrete-time, and are linear, time-invariant (LTI), and
not probabilistic. These models depend on input and
output signals ranging over Rm and Rp , respectively,
and on variables xS (t) taking values in an Euclidean
space, xS (t) ∈ X ⊆ Rn , where n, the state dimension,
is the model order. The behaviour of such a system is
encompassed by state-space models (AS , BS , CS , DS ) as
S:
(
xS (t + 1) = AS xS (t) + BS u(t),
y0 (t)
= CS xS (t) + DS u(t),
L(y0 (t)ver ) = α ∈ Σ ⇔
(4)
3.1
πt
πt
(negation) πt
(conjunction) πt
(disjunction) πt
(next) πt
(until) πt
true
p
¬ψ
ψ1 ∧ ψ2
ψ1 ∨ ψ2
ψ
ψ1 U ψ2
Api y0 (t)ver ≤ bpi , (5)
Model Set Selection
As a first step we need to embed the a-priori available
knowledge on the underlying system within a parameterised model set, under a prior distribution. The use
of linearly parameterised model sets defined through
orthonormal basis functions to represent partially unknown systems is a broadly used framework in system
identification: it allows for the incorporation of prior
knowledge, while maintaining the benefits (computational aspects) of linear parameterisations. Practically,
it has been widely used for the modelling of physical systems, such as the thermal dynamics of buildings [35,29].
Note that although the goal of parameter exploration in
formal verification has recently attracted quite some attention [4,15,19], there are as of yet no general scalable
results for the computation of the satisfaction function for nonlinearly-parameterised discrete-time LTI
models. Whilst in general linear time-invariant models
with uncertain parameters do not map onto a linearlyparameterised model set, we argue that a linearlyparameterised model set can encompass a relevant class
of models. For instance, any asymptotically stable LTI
model can be represented uniquely by its (infinite)
impulse response [20], and the coefficients of the impulse response define a linear parameterisation for this
model. Further, the coefficients of the impulse response
converge to zero, so that a truncated set of impulse coefficients can provide a good approximate model set with
a finite-dimensional, linear parameterisation. This is
only one possible instance of modelling by a finite set of
orthonormal basis functions [21, Chapters 4 and 7],[33],
which can be selected to optimally incorporate prior
knowledge: we conclude that, as an alternative to the use
of a nonlinearly parameterised set of models, structural
information (even when inexact) can be used to select a
System properties are expressed, over a finite set of
atomic propositions pi ∈ AP , i = 1, . . . , |AP |, in Lineartime Temporal Logic [2]. LTL formulae are built recursively via the syntax ψ ::= true | p | ¬ψ | ψ ∧ ψ | ψ ∨ ψ |
+
ψ | ψ U ψ. Let π = π(0), π(1), π(2), . . . ∈ ΣN be a
string composed of letters from the alphabet Σ = 2AP ,
and let πt = π(t), π(t + 1), π(t + 2), . . . be a subsequence
of π, then the satisfaction relation between π and ψ is
denoted as π ψ (or equivalently π0 ψ). The semantics for the satisfaction are defined recursively over πt
and the LTL syntax as
(true)
pi ∈α
for given Api ∈ R1×p , bpi ∈ R that is, sets of atomic
propositions are associated to polyhedra over Y ⊂ Rp .
Let us underline that properties are defined over the behaviour y0 (t)ver of the system, and not over the noisy
measurements ỹ(t)ex of the system in the measurement
setup. Additionally, for the verification problem the input signal is modelled as a bounded signal u(t) ∈ Uver ,
and represents possible external non-determinism of the
environment acting on the system.
where matrices AS , BS , CS , DS are of appropriate dimensions. Let us remark that LTI systems represent
the most common modelling framework in control theory, a key framework leading towards generalisations
to more complicated (e.g., nonlinear) dynamical models. The experimental measurement setup, as depicted
in Figure 1, consists of the signals u(t)ex and ỹ(t)ex =
y0 (t)ex + e(t), representing the inputs and the measured
outputs, respectively, and where e(t) is an additive zeromean, white, Gaussian-distributed measurement noise
with covariance Σe that is uncorrelated from the inputs. Ns samples are collected within a data set Z Ns =
s
{u(t)ex , ỹ(t)ex }N
t=1 .
(atomic prop.)
V
⇔ true
⇔ p ∈ π(t)
⇔ πt 6 ψ
⇔ πt ψ1 and πt ψ2
⇔ πt ψ1 or πt ψ2
⇔ πt+1 ψ
⇔ ∃ i ∈ N : πt+i ψ2 ,
and ∀j ∈ N :
0 ≤ j < i, πt+j ψ1
4
1
= p
exp
|Σe |Ns (2π)pNs
Ns
1X
(ŷ(t, θ) − ỹ(t)ex )T Σ−1
(ŷ(t,
θ)
−
ỹ(t)
)
−
ex
e
2 t=1
set of orthonormal basis functions, whose finite truncation defines a finite-dimensional linearly-parameterised
model set indexed over the coefficients of the basis
functions. Thus, in the following we consider a linearly
parameterised model set G that encapsulates system S,
and specifically G = {(A, B, C(θ), D(θ)), θ ∈ Θ}.
and can be directly usedin Proposition 1. This conditional density p Z Ns |θ depends implicitly on the
given initial state x(0)ex and, for the case of a given
uncertainty distribution for x(0)ex , p Z Ns |θ should be
marginalised as a latent variable [28]. The a-posteriori
uncertainty distribution is obtained as the analytical
solution of the parametric inference in (3) [28].
A system, or equivalently the mathematical model that
represents it, satisfies a property if all the words generated by the model satisfy that property. Since properties
are encoded over the external (input-output) behaviour
of the system S, which is the behaviour of M(θ0 ), θ0 ∈ Θ,
we can equivalently assert that any property ψ is verified by the system, S ψ, if and only if it is verified
by the unknown model representing the system, namely
M(θ0 ) ψ. Introduce Θψ to be the feasible set of parameters, such that for every parameter in that set the
property ψ holds, i.e., ∀θ ∈ Θψ : M(θ) ψ. As such Θψ
is characterised as the level set of the satisfaction function fψ , Θψ = {θ ∈ Θ : fψ (θ) = 1}.
3.2
Recall now that for a given specification ψ, we seek to
determine a feasible set of parameters Θψ , such that
the corresponding models admit property ψ, namely
M(θ) ψ, ∀θ ∈ Θψ . Since models M(θ) have a linearlyparameterised state space realisation as per (6), it follows that when the set of initial states and inputs Xver
and Uver are bounded polyhedra, the verification of a
class of safety properties expressed by formulae with labels as in (5) leads to a set of feasible parameters Θψ that
is a polyhedron, which can be easily computed. More
precisely, the following theorem can be derived.
Safety Verification of Bounded-time Properties
Models M in the class G have the following representation (A, B, C(θ), 0):
M(θ) :
(
x(t + 1) = Ax(t) + Bu(t),
ŷ(t, θ)
= C(θ)x(t),
Theorem 3 ([17]) Given a bounded polyhedral set (or
equivalently a polytope) of initial states x(0) ∈ Xver
and of inputs u(t) ∈ Uver for t ≥ 0, and considering a labelling map as in (5), then the feasible set Θψ
of the parameterised model set (6) results in a polyhedron for properties ψ composed of the LTL fragment
ψ ::= α| ψ|ψ1 ∧ ψ2 , with α ∈ Σ.
(6)
and are parameterised by θ ∈ Θ ⊂ Rpn :θ = vec(C) with
a prior probability distribution p (θ). In addition to this
strictly proper model class we will also allow for proper
model (A, B, C(θ), D(θ)) where both the C and the Dmatrices are parameterised and the parameterisation is
θ = vec([C D])). For a given initial condition x(0) and
input sequence, the output of the “true” model ŷ(t, θ0 )
is equal to the system output y0 (t).
Proof [of Theorem 3] Let ⊗ denote the Kronecker product. Consider the input set Uver to be the convex hull
of U , i.e. conv(U ) = Uver . Similarly let the set of initial
states be conv(Xver ) = Xver . Let the model set be given
as M(θ) = (A, B, C(θ), D). We will temporarily assume
that D is set equal to zero. Afterwards we will show how
to work with a parameterised D. Note that the syntax
fragment ψ ::= α| ψ|ψ1 ∧ ψ2 with α ∈ Σ = 2AP is
equivalent to ψ ::= p| ψ|ψ1 ∧ ψ2 with p ∈ AP .
Given a measurement set-up as in Figure 1 with unknown parameter θ0 . Then u(t)ex and ỹ(t)ex represent the input and the measured output, respectively,
and e(t) is an additive zero-mean, white, Gaussiandistributed measurement noise with covariance Σe that
is uncorrelated from the input. Furthermore u(t) is
assumed to be uncorrelated with the noise e(t). From
this set-up Ns samples are collected in a data set
s
Z Ns = {u(t)ex , ỹ(t)ex }N
t=1 .
1. We claim that for every specification ψ composed
from the syntax fragment ψ ::= p| ψ|ψ1 ∧ψ2 and θ ∈ Θ,
the words generated by a model M(θ) = (A, B, C(θ), 0)
with state x(t) satisfy the specification ψ, denoted <
M(θ), x(t) > ψ, if and only if
Therefore given the operating conditions of the experiment set-up the measured signal ỹ(t)ex can be fully characterised: its probability density, conditional on the parameters θ, is
T
Inψ ⊗ x(t) Nψ + Kψ θ ≤ Bψ .
(7)
The matrices Nψ ∈ Rnnψ ×np , Kψ ∈ Rnψ ×np , Bψ ∈ Rnψ
in the above satisfaction relation have dimensions that
are functions of the parametrisation and of the property
dependent “dimension” nψ , and are obtained inductively
Ns
Y
p Z Ns |θ =
p (ỹ(t)ex |θ)
t=1
5
over the syntax of the specification.
For any atomic propositions the model starting from
state x(t) satisfies a property pi , i.e., < M(θ), x(t) >
pi ⇔ Api y ≤ bpi , with Api ∈ R1×p and bpi ∈ R we construct the matrices Npi , Kpi and Bpi as follows. Consider y(t) for a given x(t) then
The and operation ψ1 ∧ ψ2 for (Nψ1 , Kψ1 ,Dψ1 ,bψ1 ) and
(Nψ2 , Kψ2 ,Dψ2 ,bψ2 ) with nψ1 ∧ψ2 = (nψ1 + nψ2 ) gives
Nψ1 ∧ψ2 =
Api y(t) = Api C(θ)x(t) = x(t)T (In ⊗ Api )θ.
"
N ψ1
N ψ2
#
, Kψ1 ∧ψ2
#
"
#
"
Bψ 1
K ψ1
.
, Bψ1 ∧ψ2 =
=
Bψ 2
K ψ2
This can be derived from
n×np
, Kpi = O1×np ∈
This yields Npi = (In ⊗ Api ) ∈ R
R1×np , and Bpi = bpi ∈ R1×1 .
The next operation
ψ1 with matrices (Nψ1 ,Kψ1 ,
Dψ1 ,bψ1 ) yields matrices
< M(θ), x(t) > ψ1 ∧ ψ2
T
^
Inψi ⊗ x(t) Nψi + Kψi θ ≤ Bψi
⇔
i∈{1,2}
N
K
B
ψ1
ψ1
ψ1
= 1|U| ⊗ Inψ1 ⊗ AT Nψ1 ,
T
= U Inψ1 ⊗ B Nψ1 + 1|U| ⊗ Kψ1 ,
= 1|U| ⊗ Bψ1 ,
⇔ Inψ1 ∧ψ2
where the i-th set of nψ1 rows of U ∈ R|U|nψ1 ×m is defined as
Inψ1 ⊗ uTi with ui ∈ U
and where n
ψ1
2. The matrix-valued function
T
Inψ ⊗ x(0) Nψ + Kψ θ
= |U |nψ1 . This can be derived as
is affine in xT (0) (for a fixed θ), therefore its value at the
initial condition x(0) ∈ Xver is a convex combination of
the function values at the vertices Xver of Xver . Thus the
satisfaction relation < M(θ), x(0) > ψ represented by
the multi-affine inequality holds uniformly over x(0) ∈
Xver if and only if it holds for the vertices of Xver .
This gives a set of affine inequalities in θ, thus the feasible
set Θψ is a polyhedron and is given as
< M(θ), x(t) > ψ ⇔ ∀u(t) ∈ Uver :
T
Inψ1 ⊗ x(t + 1) Nψ1 + Kψ1 θ ≤ Bψ1 ,
⇔ ∀u(t) ∈ Uver :
T
Inψ1 ⊗ Ax(t) Nψ1
T
+ Inψ1 ⊗ Bu(t) Nψ1 + Kψ1 θ ≤ Bψ1 .
(
Since the above is an affine function in u(t), the image
of every u(t) ∈ conv(U ) = Uver can be expressed as a
convex combination of the values at the vertices ui ∈ U ,
c.f. [6]. Then an equivalent expression is
+ Inψ1
ψ,
and b
ψ,
xi ∈Xver
Inψ ⊗ xi
T
N ψ + K ψ θ ≤ Bψ
)
.
3. To prove Theorem 3 we need to extend the results to
models with parameterised D. The dynamics of model
(A, B, C, D) with both C and D fully parameterised can
be reformulated as
which
can be rewritten as
T
T
⇔ 1|U| ⊗ Inψ1 ⊗ Ax(t) Nψ1 + U Inψ1 ⊗ B Nψ1
+ 1|U| ⊗ Kψ1 θ ≤ 1|U| ⊗ Bψ1 .
Having obtained K ψ , D
first term to obtain N ψ :
θ∈Θ:
^
The set Θψ is a polyhedron, since it is formed by a finite
set of half spaces.
T
Inψ1 ⊗ Ax(t) Nψ1
T
T
Inψ1 ⊗ B Nψ1 + Kψ1 θ ≤ Bψ1
⊗ ui
⇔ ∀ui ∈ U :
#
#!
"
# "
"
T N ψ1
Bψ 1
K ψ1
.
θ≤
+
⊗ x(t)
Bψ 2
K ψ2
N ψ2
"
#
x(t + 1)
u(t + 1)
=
"
#"
#
A B x(t)
0 0
u(t)
+
" #
0
I
u(t + 1)
h
i
y(t) = C D x(t).
now rewrite the
Using the new matrices (Ã, B̃, C̃(θ), 0) the obtained results still hold. For part 2. set of vertices Xver needs to
be extended with the vertices of U as Xver × U .
✷
1|U| ⊗ Inψ1 ⊗ x (t) Inψ1 ⊗ AT Nψ1
= I|U| 1|U| ⊗ Inψ1 ⊗ xT (t) Inψ1 ⊗ AT Nψ1
= I|U|nψ1 ⊗ xT (t) 1|U| ⊗ Inψ1 ⊗ AT Nψ1 .
T
In the computation of the feasible set, the faces of the
polyhedron Θψ are shown to be a function of the ver-
6
Table 1
Mean (µ) and variance (σ 2 ) of the confidence obtained from
100 experiments with 200 measurements each.
tices 2 of the bounded set of initial states Xver and of
the set of inputs Uver , and are also expected to grow in
number as a function of the time horizon of the property.
The result in Theorem 3 is valid for any finite composition of the LTL fragment ψ ::= α| ψ|ψ1 ∧ ψ2 , as such
it only holds for finite horizon properties. Properties defined over the infinite horizon will be the objective of
Section 3.4.
3.3
θ0
h
iT
-1 -1
h
iT
-1 0
h
iT
-1 1
Case Study: Bounded-Time Safety Verification
Consider a system S and verify whether hthe output
i
y0 (t)ver remains within the interval I = −0.5, 0.5 ,
labeled as ι, for the next 5 time steps, under u(t)ver ∈
Uver = [−0.2, 0.2] and x(0)ver ∈ {02 } = Xver . Introduce accordingly the alphabet Σ = {ι, τ } and the
labelling map L : L(y) = ι, ∀y ∈ I, L(y) = τ, ∀y ∈ Y\ I.
Now Vcheck whether the following LTL property holds:
S 5i=1 ( )i ι. We assume that system S can be represented as an element of a model set G with transfer
functions characterised by second-order Laguerre-basis
ones [20] (a special case of orthonormal basis functions), which translates to the following parameterised
state-space representation:
#
√
1 − a2
x(t + 1)=
x(t) +
u(t),
√
1 − a2 a
(−a) 1 − a2
"
a
0
#
"
µ
σ2
0.348
0.073
h
0.060
h
0.086
h
0.705
0.492
θ0
µ
σ2
1 -1
iT
0.491
0.085
1 0
iT
0.730
0.056
1 1
iT
0.339
0.065
represents the confidence in the safety of the system, as
displayed in Table 1 via mean and variance terms.
θ2
1
0
−1
−2
3.4
(8)
T
0
θ1
2
Fig. 2: Feasible set
of parameters in Θ,
and
contour
lines
of
the
quantity
p θ|Z Ns ,
obtained
for θ0 = [1 0]T .
Verifying Unbounded-Time Properties Using Invariant Sets
In this section we extend the approach unfolded in Section 3.2, to hold on the LTL fragment ψ ::= α| ψ|ψ1 ∧
ψ2 with additionally the unbounded invariance (safety)
operator. Recall the form of the k-bounded and of the unVk
bounded invariance operators, namely ✷k ψ = i=0 i ψ
and ✷ψ = ¬(true U ¬ψ) respectively. The extension
from a k-bounded operator, covered by the result in Theorem 3, to the unbounded invariance one, is based on
the concept of robust positive invariance [7, Def. 4.3],
recalled next.
ŷ(t, θ) = θ x(t) .
The parameter set is chosen as θ ∈ Θ = [−10, 10]2,
whereas the coefficient a is chosen to be equal to 0.4.
We select, as prior available knowledge on the system, a
uniform distribution p (θ) on the model class, and pick a
known variance σe2 = 0.5 for the white additive noise on
the measurement. The set of feasible parameters Θψ ⊂ Θ
is represented in Figure 2 and is computed according to
Theorem 3. Based on the prior available knowledge, the
confidence associated to θ0 ∈ Θψ amounts to 0.0165 3 .
In comparison to this value, after doing an experiment
on the system with “true parameter” θ0 = [1 0]T (Figure 2) and with input signal u(t)ex , a realisation of a
white noise with a uniform distribution over [−0.2, 0.2],
and measuring ỹ(t)ex for 200 consecutive time instances
the uncertainty distribution is refined as p θ|Z Ns . The
resulting confidence (2) in the property is increased to
0.779.
Along this line of experiments, we have repeated the test
100 times, for several instances of the parameter θ0 characterising the underlying system S. In all instances, after
obtaining 200 measurements the a-posteriori confidence
Definition 2 For the system x(t + 1) = Ax(t) + Bu(t),
the set S ⊆ X is said to be robustly positively invariant
if, for all x(0) ∈ S and u(t) ∈ U, the condition x(t) ∈ S
holds for all t ≥ 0.
Recall that the feasible set Θψ is defined as the set of parameters for which property ψ holds, namely ∀θ ∈ Θψ :
M(θ) ψ. The satisfaction relation M(θ) ψ depends
implicitly on the set of initial states x(0) ∈ Xver and
on the set of inputs Uver . Let us extend the definition
of the feasible set to explicitly account for its dependence on the set of initial conditions: given a bounded
and convex set S ⊂ X, let Θψ (S) be defined as the set
of parameters in Θ for which the parameterised models
M(θ) initialised with x(0) ∈ S satisfy ψ over input signals u(t) ∈ Uver t ≥ 0. Hence the feasible set Θψ can
be written as a function of the set of initial states Xver ,
that is Θψ (Xver ). Thus the extended map Θψ (·) takes
2
A polytope can be written as the convex hull of a finite
set of vertices.
3
This is obtained by numerical computation of (2) with
probability distribution p (θ). ntegrals are solved via the numerical integration tool in Matlab.
7
Assume that Uver includes the origin, and denote the
forward reachability mappings initialised with R(0) :=
{0n } ⊂ X as
subsets of the state space into subsets of the parameter
space. Note that if S is a robustly positively invariant
set that includes the set of initial states Xver ⊆ S, then
for all θ ∈ Θψ (S) the models M(θ) satisfy ψ over all
infinite-time model traces x(t): this allows to state that
M(θ) ✷ψ. We can show that the following holds.
X
R(i) := Post(R(i−1) ),
with set operation Post(X) := {x′ = Ax + Bu, x ∈
X, u ∈ U}. Denote the limit reachable set as R∞ =
limi→∞ R(i) . From literature we recall that properties of
these i-step reachable sets, as given in [7] include the following: for a reachable pair (A, B) and an asymptotically
stable matrix A, the ∞-reachable set R∞ is bounded
and convex [7, Proposition 6.9]. The k-step reachable
set converges to the ∞-reachable set via (9), since it is
monotonically increasing R(i) ⊆ R(i+1) . Moreover, R∞
is the minimal robustly positively invariant set for the
system, so that any positively invariant set includes R∞
[7, Proposition 6.13]. Thus, starting from x(0) = 0n , all
x(t) ∈ R∞ , and furthermore ofinterest to this work we
conclude that Θ✷k ψ= Θψ R(k) and Θ✷ψ = Θψ R∞ .
Θ
Lemma 4 The function Θψ (·) : 2 → 2 , for specifications obtained as ψ ::= α | ψ | ψ1 ∧ ψ2 , is monotonically decreasing: that is if S1 ⊆ S2 , then Θψ (S2 ) ⊆
Θψ (S1 ).
Proof We leverage the notation used in the proof of
Theorem 1. Provided that the parameterised model is
given as (A, B, C(θ), 0), we show that any θ ∈ Θψ (S2 ) is
also an element of θ ∈ Θψ (S1 ). Suppose S2 has a finite
number of vertices xi ∈ V (S2 ), then for any θ ∈ Θψ (S2 ):
V
xi ∈V(S2 )
(Inψ ⊗ xi )T Nψ + Kψ θ ≤ Bψ
and for every x ∈ S2
Feasible set for invariance properties under polytopic
sets of initial states
(Inψ ⊗ x)T Nψ + Kψ θ ≤ Bψ .
More generally, if Xver ⊆ R∞ and ceteris paribus, then
R∞ is the minimal robustly positively invariant set that
includes Xver , and Θψ (R∞ ) = Θ✷ψ . For finite iterations the reachable sets R(i) are polytopes, and if R(i) =
R(i+1) , then R(i) = R∞. Though the iterations can stop
in finite time, in general the number of iterations to obtain R∞ can be infinite. Whilst the minimal robustly
positively invariant set is not necessarily closed or a polytope, there exist methods to approximate R∞ as detailed
in [7]. For instance, for stable systems, R(k) is shown to
converge to R∞ , in the sense that for all ǫ > 0 there exists k̄ such that for k ≥ k̄, R(k) ⊆ R∞ ⊆ (1 + ǫ)R(k) [7,
Proposition 6.9].
Since the vertices xj ∈ V (S1 ) are also elements of S2 ,
then
V
xj ∈V(S1 )
(9)
(Inψ ⊗ xj )T Nψ + Kψ θ ≤ Bψ
and θ ∈ Θψ (S1 ). This reasoning can be trivially extended to include parameterised D matrices. Increasing
the number of vertices of S1 and S2 , does not change
the result, hence the same holds if S1 and S2 are convex
sets.
✷
Based on the result in Lemma 4, we conclude that the
maximal feasible set Θ✷ψ is obtained as a mapping from
the minimal robustly positively invariant set S that includes Xver : Θ✷ψ = Θψ (S). This leads next to consider
under which conditions such minimal robustly positively
invariant set S can be exactly computed or approximated.
Recall that the maximal feasible set Θ✷ψ is obtained
as a mapping from the minimal robustly positively invariant set S including Xver , that is Θ✷ψ = Θψ (S).
Let us extend the study to the case where the conditions Xver = {0n } or its extension Xver ⊆ R∞ do not
apply, while the condition on the bounded set Uver is
maintained, that is 0 ∈ Uver . Consider the more general case where the set of initial states is a polytope
but not necessarily a subset of R∞ . Denote the union
of the forward reachability mappings initialised with
(0)
RXver := Xver ⊆ X as
Feasible set for invariance properties with Xver = {0n }
For Xver = {0n }, assuming a bounded interval Uver with
the origin in its interior, and under some basic assumptions on the dynamics (to be shortly discussed), the minimal robustly positively invariant set can be shown to be
a bounded and convex set that includes the origin [7].
Maintaining the condition of Uver being bounded and
having the origin in its interior, we first consider the case
that Xver = {0n } and characterise S via tools available
from set theory in systems and control; thereafter we
look at extensions to more general sets of initial states
Xver .
(i)
(i−1)
(i−1)
RXver := RXver ∪ Post(RXver ) .
(10)
This set is also known in the literature as the reach tube.
The corresponding set for infinite time is denoted as
(i)
∞
R∞
Xver = limi→∞ RXver . Notice that if Xver ⊆ R , then
∞
∞
R = RXver . The iteration is monotonically increasing
(i)
(i+1)
(i)
(i+1)
RXver ⊆ RXver , and whenever RXver = RXver it stops
8
(i)
after a finite number of iterations with R∞
Xver = RXver .
Of course, also in this more general case, the number of
iterations can be unbounded, however the convergence
properties of R(i) extend seamlessly to the case of sets
(i)
(i)
RXver . Since RXver is a union of polytopes, it is not guaranteed to be a convex set. Still, it can be shown via the
proof of Theorem 3 that the computation of the
feasible
set Θψ (S) boils down to that of Θψ conv(S) .
if ǫθ ≥
Proof 1. Θψ (R + ǫx B) ⊆ Θψ (R)
Based on the definition of this set (c.f. the proof of Theorem 3), the set operation Θψ (·) is monotonically decreasing. Therefore Θψ (R + ǫx B) ⊆ Θψ (R) holds.
Remark 5 Let us illustrate the convergence property
(i)
for sets RXver as follows. For every vertex xi (0) ∈ Xver ,
select a decomposition xir + xis with xir ∈ R∞ , which
minimises kxis k for a chosen vector norm k · k. Since
every element x(0) ∈ Xver is a convex combination of
the vertices xi (0), it follows that for all x(0) ∈ Xver :
x(0) =
X
i
i
ai x (0) =
X
ai xir (0)
+
i
X
ǫx ǫp maxi (kvi k)2
kAp k2
, for ǫp := max
.
p∈AP |bp |
1 + ǫx ǫp maxi (kvi k)
2. Θψ (R) ⊆ Θψ (R + ǫx B) + ǫθ B
Consider the case where the model is (A, B, C(θ), 0). To
prove (11), we first find a ǫθ as a function of ǫx such that
Θψ (R) ⊆ Θψ (R + ǫx B) + ǫθ B.
Let vi be the vertices of the polytope vi ∈ V (Θψ (R)),
then (12) holds if and only if vi ∈ Θψ (R + ǫx B) + ǫθ B.
Equivalently, this means that there exists a rθ ∈ ǫθ B
such that vi − rθ ∈ Θψ (R + ǫx B). This is equivalent to
demanding that for every xTj ∈ V (R), vi ∈ V (Θψ (R))
and rx ∈ ǫx B, there exists a vector rθ ∈ ǫθ B:
ai xis (0)
i
∈ conv(xir (0)) + conv(xis (0)) ⊆ R∞ + X̄ver ,
P
with i ai = 1 for ai ≥ 0 and where X̄ver = conv(xis (0)).
We obtain that Xver ⊆ R∞ + X̄ver , and that the minimal positively invariant set R∞
Xver can be bounded by
Sk
i
∞
R + limk→∞ i=0 A X̄ver . Under condition of asymptotic stability on A, necessary for R∞ to be a bounded
and convex polytope, Ai X̄ver will converge to {0n }.
(k)
Thus, the iteration RXver is monotonically increasing
and bounded, hence it converges. If X̄ver includes the
origin in its interior then there exists a finite iteration
Sk+1
Sk
such that i=0 Ai X̄ver = i=0 Ai X̄ver . Moreover, for
any reachable pair (A, B) and asymptotically stable A,
the closure of the minimal robustly positively invariant
set R∞
Xver includes the origin.
(Inψ ⊗ (xTj + rxT ))Nψ + Kψ (vi − rθ ) ≤ Bψ
⇔
(Inψ ⊗ xTj )Nψ + Kψ (vi − rθ )
+ (Inψ ⊗ rxT )Nψ (vi − rθ ) ≤ Bψ .
Take (vi − rθ ) = (1 − αi )vi with αi ∈ [0, 1), then
(Inψ ⊗ xTj )Nψ + Kψ (1 − αi )vi
+ (Inψ ⊗ rxT )Nψ (1 − αi )vi ≤ Bψ
(1 − αi )(Inψ ⊗ rxT )Nψ vi ≤ αi Bψ .
⇔
(13)
Separate the matrix Nψ and Bψ into its block matrices
Nψj = [Nψ ]{1+(j−1)n:nj}×{1:n} and B j = [Bψj ]j such that
inequality (13) is equivalent to the set of inequalities
Robust approximations of the feasible set via Θψ (·)
In order to exploit convergence in the computation of the
feasible set for invariance properties, we need to bound
the error incurred with the use of approximations of the
∞
sets R∞
Xver or R . Let B denote a unit ball centred at
the origin and let the Hausdorff distance between sets
R1 and R2 be defined as
(1 − αi )rxT Nψj vi′ ≤ αi bj , for j = 1, . . . , nψ
αi
⇔ rxT Nψj vi′ ≤
bj .
(1 − αi )
(14)
(15)
Given that 0 ∈ Θψ (R), it follows that bj ≥ 0 for j =
1, . . . , nψ
δH (R1 , R2 ) = inf{ǫ ≥ 0|R1 ⊆ R2 + ǫB, R2 ⊆ R1 + ǫB}.
max rxT Nψj vi′ (bj )−1 ≤
j
We can show that the following holds.
αi
.
(1 − αi )
The term on the left can be upper bounded based on the
Cauchy-Schwarz inequality
Lemma 6 Consider a polytope R, and a property ψ
comprised of ψ ::= α| ψ|ψ1 ∧ψ2 , with α ∈ Σ, for which
Θψ (R) is a non-empty polytope with vertices vi and the
origin in its interior. Let A be bounded as kAk2 ≤ 1.
Then for any ǫx ≥ 0,
Θψ (R + ǫx B) ⊆ Θψ (R) ⊆ Θψ (R + ǫx B) + ǫθ B
(12)
max rxT Nψj vi′ (bj )−1 ≤ max k(Nψj )T rx k2 kvi′ k2 (bj )−1
j
≤
j
j T
max k(Nψ ) k2 krx k2 kvi′ k2 (bj )−1
j
≤ ǫx ǫp kvi′ k2 .
(11)
9
and krx k2 ≤ ǫx
0n in its interior. The generalisation to the case dealing
with an Hausdorff distance of the feasible set for invariance properties with a set of inputs 0 6∈ Uver is outside
of the scope of this work.
The last inequality follows from the introduction of the
precision of the labelling, denoted as ǫp , and defined as
ǫp = max
p∈AP
kAp k2
.
|bp |
(16)
Convergence properties
Remember that kL ⊗ Kk2 = kLk2 kKk2 . Then based on
Theorem 3 and on the condition kAk2 ≤ 1, it can be
shown that
max k(Nψj )T k2 |bj |−1 ≤ max
j
p∈AP
We can employ Lemma 6 to bound the Hausdorff dis(k)
tance between Θψ (RXver ) and Θ✷ψ . If Xver = {0n } and
the spectral radius of A is strictly less than 1 (that is
ρ(A) < 1), then the Hausdorff distance can be bounded
as
kAp k2
.
|bp |
αi
Note that (1−α
monotonically increases with αi for
i)
αi ∈ [0, 1). Therefore a bound on αi can be found as
δH (R(k) , R∞ ) ≤ ǫ(k) := kAk k2 max (|u|)c1 ,
u∈U
P∞
i
with c1 a bound on
i=0 kA Bk, which is the peakto-peak performance of the dynamical system formed
by (A, B). In case that Xver 6⊆ R∞ then the forward
reachable iteration can be rewritten as
αi = (ǫx ǫp kvi k)/(1 + ǫx ǫp kvi k) for j = 1, . . . , nψ . (17)
It follows that (12) holds if
ǫθ = max(kvi k2 )
ǫx ǫp max(kvi k2 )
.
1 + ǫx ǫp max(kvi k2 )
(19)
(18)
(k)
RXver =
For the case that the model is parameterised in both
S and D, i.e., (A, B, C(θ), D(θ)) the derivation is a bit
more cumbersome but can be repeated with no change
to the end result.
✷
k
[
i=0
Ai Xver + R(k) .
The Hausdorff norm can be bounded as
(k)
k+1 k δ (X
δH (RXver , R∞
2 H
ver , {0n }).
Xver ) ≤ ǫ(k) + kA
Note that for ρ(A) < 1 the norm kAk k2 → 0 for k → ∞.
(k)
In case the conditions of Lemma 6 on RXver ⊆ X and
(k)
Θψ RXver hold, the Hausdorff distance δH (Θ✷k ψ , Θ✷ψ )
can be bounded by
kAk k2 max(kvi k)2 ǫp max (|u|)c1 + kAkδH (Xver , {0n }) .
Let us briefly discuss the conditions under which Lemma
6 is applicable. The condition that Θψ (R) is not empty
is raised to avoid the trivial case where Θψ (R) = ∅ (11)
holds for all ǫθ . The condition that Θψ (R) is a polytope
and hence bounded is necessary to obtain a bounded
Hausdorff distance. This distance quantifies the difference between two sets, and is a necessary step to bound
the approximation error. The requirement that Θψ (R)
includes the origin is a sufficient condition and relates
to well-posedness for bounded input sets including the
origin. When considering invariance properties defined
for 0 ∈ Uver and for any polytope Xver , the requirement
that 0n ∈ Θψ (·) is necessary for Θ✷ψ to be non-empty:
this can be intuitively illustrated by noting that under
an assumption of asymptotic stability for A, for any θ
and for u(·) = 0 the output ŷ(t, θ) of the model in (6)
converges to 0. Hence for a property to be satisfied under these conditions it should at least hold for the zero
output, which is equivalent to demanding that it holds
for θ = 0n . For any atomic proposition pi ∈ AP (see
Equation (5)) it can be shown that there is an invertible mapping between the row vectors, proportional to
the normals of the faces of the polyhedral set Θpi (x(0)),
and the initial state x(0). Therefore, if R(k) has the origin in its interior, then Θpi (R(k) ) has to be bounded,
and as a consequence so has any feasible set comprising
this atomic proposition. This holds for k ≥ n if (A, B) is
a reachable pair and if Uver has 0 in its interior. Under
(k)
the same conditions there exists a k such that RXver has
i
u∈U
(20)
Use in the verification of unbounded-time properties
Based on the convergence properties of the feasible set,
the asymptotic behaviour of the confidence computed in
Proposition 1 can be stated.
Corollary 7 (Convergence) Under the conditions
of Lemma 6, for a Gaussian distribution p (θ) ∼
N (µθ , Rθ ) with a covariance Rθ ≻ 0, P θ ∈ Θ✷k ψ →
P (θ ∈ Θ✷ψ ) for k → ∞.
Proof [of Corollary 7] For a strictly positive Rθ , the
Gaussian density distribution takes finite values over
the parameter space, therefore the convergence of a
monotonically-decreasing polytope over the parameter
space induces the convergence of the associated probability measure.
✷
Theorem 3 can now be generalised to include unboundedtime invariance properties as follows.
10
Based on R(k) , the feasible set for the k-bounded
invariance ✷k ι can be computed as Θ✷k ι = Θι R(k) . The feasible sets Θ✷k ι with k = 1, . . . , 20 are plotted in Figure
3b. Observe that the feasible set Θ✷1 ι is not bounded,
but for k ≥ 2 the feasible sets are bounded and, as expected, decrease in size with time. In Figure 4 (middle
plot) bounds on the Hausdorff distances δH (Θ✷ι , Θ✷k ι )
are given for k = 2, . . . , 20 (no finite bound is computed
for the index k = 1, since for that instance the feasible set is not bounded). Let us conclude this case study
looking at confidence quantification, as a function of the
time horizon. Figure 4 (lower plot) represents
the confidence over the property P θ ∈ Θ✷k ι | Z Ns , for indices
k = 1, . . . , 20. Unlike the case discussed in Section 3.3,
which focused on looking at statistics of the confidence
via mean and variance drawn over multiple experiments,
we zoom in on asymptotic properties by considering a
data set Z Ns comprising a single trace made up of 200
measurements, simulated under the same conditions as
in Section 3.3, and with θ0 = [1 0]T . From the
resulting probability density distribution p θ | Z Ns , it may
be observed that the confidence converges rapidly to a
nonzero value.
Theorem 8 Consider a polytopic set of initial states
x(0) ∈ Xver , inputs u(t) ∈ Uver for t ≥ 0, and a
labelling map as in (5). Let R̂∞
Xver be a polytopic superset of the minimal robustly positively invariant set
that includes Xver , denoted as R∞
Xver ; then the feasible set admits a polyhedral subset Θ̂ψ ⊂ Θψ for every specification ψ expressed within the LTL fragment
∞
ψ := α| ψ|ψ1 ∧ ψ2 |✷ψ, and if R̂∞
Xver = RXver then
Θ̂ψ = Θψ .
Proof Every property φ ::= p| ψ|ψ1 ∧ ψ2 |✷ψ with
p ∈ AP can be rewritten as ✷ψ1 ∧ ψ2 where ψ1 and ψ2
have syntax ψ ::= p| ψ|ψ1 ∧ ψ2 .
For the set of initial states Xver , a property ψ is invariant
hM(θ), x(0)i ✷ψ, ∀x(0) ∈ Xver
∞
if and only if ∀x ∈ R∞
Xver : hM(θ), xi ψ. Let R̂Xver be
∞
a polytopic superset of RXver with a finite set of vertices
vR ∈ VR , then the subset approximation of the feasible
set Θ✷ψ follows as Θ✷ψ ⊇ Θ̂✷ψ =
θ∈Θ:
^
vR ∈VR
T
(Inψ ⊗ vR
)Nψ + Kψ θ ≤ bψ
3.6
)
The discussed approach based on polytopes allows for
analytical expressions of the feasible set, however the
implementation may not scale to models with very large
dimension: in particular, the number of half-planes characterising the feasible set may increase with the time
bound of the LTL formula ψ (that is, with the repeated
application of the
operator), and with the cardinality of the atomic propositions in the alphabet Σ. Still,
note that these computations are essentially quite similar to known reachability computations, therefore the
method is extendable well beyond the 2-dimensional case
study, especially when applying sophisticated reachabil-
∞
where Θ̂✷ψ ⊆ Θ✷ψ . Note that if R̂∞
Xver = RXver then
Θ̂✷ψ = Θ✷ψ . The feasible set of ✷ψ1 ∧ ψ2 is equal to
Θ✷ψ1 ∧ψ2 = Θ✷ψ1 ∩Θψ2 . And Θ✷ψ1 ∧ψ2 can be upper and
lower bounded as Θ̂✷ψ1 ∩Θψ2 ⊆ Θ✷ψ1 ∧ψ2 ⊆ Θ✷k ψ1 ∩Θψ2
with k ∈ N. This proves Theorem 8 for the case where
the model is (A, B, C(θ), 0). The additional parameterisation of D does not change the reasoning.
✷
The extension beyond the LTL fragment discussed above
may lead to feasible sets that are in general not convex
and are therefore beyond the scope of this work.
0.5
5
x2
3.5
Discussion on the Generalisation of the Results
Case Study (cont.): Unbounded-Time Safety Verification
0
−0.5
We study convergence properties for the safety specification ι considered in the case study in Section 3.3
maintaining the same operating conditions as before for
the safety verification and the experiment. In Figure 3a
the forward reachability sets R(k) with k = 1, . . . , 20
are obtained for the model dynamics in (8). Figure 4
(upper plot) displays bounds ǫ(k) on the Hausdorff distances δH (R(k) , R∞ ) computed with (19): starting from
a slanted line segment for R(1) as in Figure 3a, it can be
observed that the forward reachable sets R(k) converge
rapidly, as confirmed with the error bound displayed in
Figure 4 (upper plot).
θ2
(
−0.2 0
x1
0.2
(a) The first 20 iterations
of the forward reachable set
R(k) , k = 1, . . . , 20 for the
case study. The reachable sets
grow in size from dark grey
(k = 1) to light grey (k = 20),
so that R(k−1) ⊆ R(k) .
0
−5
−5
0
θ1
5
(b) The feasible sets for
the k-bounded invariance
property ✷k ι, with k =
1, . . . , 20, obtained for the
case study.
Fig. 3. Reachable and feasible sets for unbounded-time verification problem.
11
signal interaction for efficient data usage over general
classes of models, as initially attempted in [17]. Additionally, the design of control policies that optimise properties of interest over partly unknown systems is topic
of current work.
ǫ(k) ≥ δH (R(k) , R(∞) )
0.5
0
5
10
15
20
k
ǫθ (k) ≥ δH (Θ✷ι , Θ✷k ι )
3
1.5
0
5
10
15
References
[1] A. Abate, R. C. Hillen, and S. A. Wahl. Piecewise affine
approximation of fluxes and enzyme kinetics from in-vivo 13 C
labeling experiments. International Journal of Robust and
Nonlinear Control, pages 1120–1139, 2012. Special Issue on
System Identification for Biological Systems.
20
k
P θ ∈ Θ✷k ι | Z Ns
1
0.9
0.8
0
5
10
[2] C. Baier and J.-P. Katoen. Principles of model checking.
MIT Press, 2008.
15
[3] E. Bartocci, L. Bortolussi, and G. Sanguinetti. Learning
temporal logical properties discriminating ECG models of
cardiac arrhythmias. CoRR, abs/1312.7523, 2013.
20
[4] G. Batt, C. Belta, and R. Weiss. Model checking genetic
regulatory networks with parameter uncertainty. In HSCC,
pages 61–75. Springer, 2007.
Fig. 4. (Upper plot) Error bound on the approximation
level of the k-th forward reachable sets, which is such
that R(∞) ⊆ R(k) + ǫ(k) for k = 1, . . . , 20. (Middle plot)
The Hausdorff distance ǫθ (k) between Θ✷k ψ and Θ✷ψ with
k = 2, . . . , 20, obtained for the case study.(Lower plot) Confidence that S ✷k ι for k = 1, . . . , 20 for the case in Section 3.3, with a new experiment consisting of 200 samples
collected as Z Ns .
[5] C. Belta, A. Bicchi, M. Egerstedt, E. Frazzoli, E. Klavins,
and G. J. Pappas. Symbolic planning and control of robot
motion [grand challenges of robotics]. Robotics Automation
Magazine, IEEE, pages 61–70, Mar. 2007.
[6] C. Belta, L. C. G. J. M. Habets, and V. Kumar. Control of
multi-affine systems on rectangles with applications to hybrid
biomolecular networks. In Conf.on CDC, pages 534–539,
2002.
ity analysis tools in the literature. Therefore the discussed limitations related to the current implementation
of the approach, ought to be dealt with in the future
by the use of tailored and less naı̈ve computational approaches.
[7] F. Blanchini and S. Miani. Set-Theoretic Methods in Control.
Birkhäuser Basel, 1st edition, 2007.
[8] L. Bortolussi and G. Sanguinetti. Learning and designing
stochastic processes from logical constraints. In QEST, pages
89–105. Springer, 2013.
[9] L. Bortolussi and G. Sanguinetti. Smoothed model checking
for uncertain continuous time Markov chains.
CoRR,
abs/1402.1450, 2014.
In the discussion of model selection, we hinted at possible generalisation beyond linearly-parameterised model
sets. Future extension will deal with hybrid models, since
when systems are not linear, their (local) behaviour is
often well approximated with piecewise-linear dynamical models.
[10] L. Brim, M. Češka, S. Dražan, and D. Šafránek. Exploring
parameter space of stochastic biochemical systems using
quantitative model checking. In N. Sharygina and H. Veith,
editors, CAV, volume 8044 of LNCS, pages 1–17. Springer,
2013.
This paper has discussed the formal verification of physical systems with partly unknown dynamics, by introducing a Bayesian framework allowing for the efficient incorporation of measurement data and prior information
within a verification procedure based on safety analysis.
This formal approach has allowed for the computation
of the confidence level over the validity of a property of
interest on the unknown system. The method has been
applied to the verification of LTI models of systems over
bounded and unbounded safety properties, and its computational overhead has been discussed at length.
[11] J. W. Burdick, N. du Toit, A. Howard, C. Looman, J. Ma,
R. M. Murray, and T. Wongpiromsarn. Sensing, navigation
and reasoning technologies for the DARPA urban challenge.
Technical report, DTIC Document, 2007.
[12] Y. Chen and T. D. Nielsen. Active learning of Markov
decision processes for system verification. In Conf. on
Machine Learning and Applications, pages 289–294, 2012.
[13] E. M. Clarke. The birth of model checking. In 25 Years of
Model Checking, pages 1–26, 2008.
[14] D. Del Vecchio and E. D. Sontag. Engineering principles
in bio-molecular systems: From retroactivity to modularity.
European Journal of Control, pages 389 – 397, 2009.
[15] G. Frehse, S. K. Jha, and B. H. Krogh. A counterexampleguided approach to parameter synthesis for linear hybrid
automata.
In HSCC, pages 187–200. Springer Berlin
Heidelberg, 2008.
Looking forward, current work targets the extension of
the applicability of tractable solutions to model-based
and data-driven verification over complex physical systems. We are presently working to extensions of the
considered set of logic formulae of interest, and plan to
employ experiment design to optimise the input-output
[16] B. M. Gyori, D. Paulin, and S. K. Palaniappan. Probabilistic
verification of partially observable dynamical systems.
CoRR, abs/1411.0976, 2014.
12
[17] S. Haesaert, P. M. J. Van den Hof, and A. Abate. Datadriven property verification of grey-box systems by Bayesian
experiment design. In American Control Conference, pages
1800–1805, 2015.
[36] P. Zuliani, A. Platzer, and E. M. Clarke. Bayesian statistical
model checking with application to Stateflow/Simulink
verification. Formal Methods in System Design, pages 338–
367, 2013.
[18] D. Henriques, J. G. Martins, P. Zuliani, A. Platzer, and
E. M. Clarke. Statistical model checking for Markov decision
processes. In QEST, pages 84–93, 2012.
Derivation of the Bounds in Section 3.4
[19] T. Henzinger and H. Wong-Toi. Using hytech to synthesize
control parameters for a steam boiler. In Formal Methods
for Industrial Applications, pages 265–282. Springer Berlin
Heidelberg, 1996.
1. Hausdorff distance of forward reachable mappings. We only sketch the method to bound the Hausdorff distance, whereas a more formal derivation can be
found in the literature on robustly positively invariant
sets [7].
[20] P. S. C. Heuberger, P. M. J. Van den Hof, and O. H. Bosgra. A
generalized orthonormal basis for linear dynamical systems.
Automatic Control, IEEE Transactions on, 40(3):451–465,
1995.
The k-step forward reachable set equals
[21] P. S. C. Heuberger, P. M. J. Van den Hof, and B. Wahlberg.
Modelling and identification with rational orthogonal basis
functions. Springer London, 2005.
[22] H. Hjalmarsson. From experiment design to closed-loop
control. Automatica, pages 393–438, 2005.
R(k) :=
[23] E. A. Lee. Cyber physical systems: Design challenges. In
Proc. of Object Oriented Real-Time Distributed Computing,
pages 363–369. IEEE Computer Society, 2008.
i
k X
[
i=1
j=1
Aj−1 Bu(i − j), for u(j) ∈ Uver
.
For 0 ∈ Uver , the minimal invariant set R∞ can be
written as
∞
i−1
X
X
Aj Bu(j) + Ai
Ak Bu(k), for u(·) ∈ Uver .
R(∞) :=
[24] A. Legay, B. Delahaye, and S. Bensalem. Statistical model
checking: An overview.
In H. Barringer, Y. Falcone,
B. Finkbeiner, K. Havelund, I. Lee, G. Pace, G. Roşu,
O. Sokolsky, and N. Tillmann, editors, Runtime Verification,
volume 6418 of LNCS, pages 122–135. Springer Berlin
Heidelberg, 2010.
j=0
k=0
(.1)
[25] A.
Legay
and
S. Sedwards. Lightweight Monte Carlo algorithm for Markov
decision processes. CoRR, abs/1310.3609, 2013.
If the spectral radius of a A is strictly smaller than 1,
ρ(A) < 1, then
[26] D. V. Lindley. The philosophy of statistics. Journal of the
Royal Statistical Society: Series D (The Statistician), pages
293–337, 2000.
R(∞) ⊆ R(k) + ǫ(k)B,
[27] H. Mao and M. Jaeger.
Learning and model-checking
networks of I/O automata. In Proc. of Asian Conference on
Machine Learning, 2012.
(.2)
with
[28] V. Peterka. Bayesian Approach to System Identification.
Trends Prog. Syst. Identif., 1981.
Ak
[29] B. C. Reginato, R. Z. Freire, G. H. D. C. Oliveira, N. Mendes,
and O. Abadie, Marc. Predicting the temperature profile
of indoor buildings by using orthonormal basis functions.
In Conf. on Building Performance Simulation Association,
United Kingdom, 2009.
∞
X
i=0
Ai Bu(k) ⊆ ǫ(k)B, for u(·) ∈ Uver .
Note that ǫ(k) is bounded for ρ(A) < 1. For a matrix
A without defective eigenvalues, i.e. where the eigenvectors form a complete basis, this L1 norm can be easily
bounded using the spectral radius of A, by selecting
[30] K. Sen, M. Viswanathan, and G. Agha. Learning continuous
time Markov chains from sample executions. In QEST, pages
146–155, 2004.
[31] K. Sen, M. Viswanathan, and G. Agha. Statistical model
checking of black-box probabilistic systems. In R. Alur and
D. Peled, editors, CAV, volume 3114 of LNCS, pages 202–
215. Springer, 2004.
ǫ(k) =
[32] P. Tabuada. Verification and Control of Hybrid Systems: a
Symbolic Approach. Springer, 2009.
∞
X
|ρ(A)|k
kAi Bk2 |u(k)|.
kBk2 max (|u|) ≥ kAk k2
u∈Uver
1 − |ρ(A)|
i=0
In case that the matrix A is defective, we opt to bound
the L1 -norm by exploitingPabsolute sum of the L2 in∞
duced norm for Ai i → ∞: i=0 kAi k2 . Note that kAi k2
converges to 0 for i → ∞ since ρ(A) < 1, therefore there
exists a finite l such that kAl k2 < 1 and we can upper
bound the absolute sum as
!
! ∞
l−1
∞
X
X
X
l i2
i1
i
kA k2
kA k2
kA k2 ≤
[33] P. M. J. Van den Hof, P. S. C. Heuberger, and J. Bokor.
System identification with generalized orthonormal basis
functions. Automatica, pages 1821–1834, 1995.
[34] M. Y. Vardi. From philosophical to industrial logics. In
Proc. of the Indian Conference on Logic and Its Applications,
pages 89–115, Berlin, Heidelberg, 2009. Springer-Verlag.
[35] G. S. Virk and D. L. Loveday. Model-based control for HVAC
applications. In Conf. on Control Applications, pages 1861–
1866. IEEE, 1994.
i=0
13
i1 =0
i2 =0
=
l−1
X
i1 =0
i1
kA k2
!
1
.
1 − kAl k2
Thus in general, the Hausdorff distance can be bounded
as
δH (R(k) , R(∞) ) ≤ ǫ(k) = kAk k2 max (|u|)c1 ,
u∈Uver
P
l
i1 =0
kAi1 k2
kBk2 for l such that kAl k2 <
with c1 =
1−kAl k2
1. Note that c1 can be replaced by any bound on the L1
norm of the dynamical system formed by (A, B).
In case that Xver 6⊆ R∞ then the forward reachable
iteration can be rewritten as
!
k
[
(k)
i
RXver =
A Xver + R(k) ,
i=0
for which we know that
(∞)
(k)
RXver ⊆ RXver + ǫ(k) + kAkk+1 δH (Xver , {0}).
Thus the Hausdorff norm is upper bounded as
(∞)
(k)
δH (RXver , RXver ) ≤ ǫ(k) + kAk+1 kδH (Xver , {0}).
2. Hausdorff distance on feasible sets. Suppose
(k)
that the conditions in Lemma 6 hold for RXver , then
(k)
we can compute a value for ǫθ such that Θψ (RXver ) ⊆
(k)
Θψ (RXver +ǫx B)+ǫθ B, where ǫx is a bound on the Haus(k)
(∞)
dorff distance δH (RXver , RXver ).
The set operation Θψ (·) is monotonically decreasing,
(k)
therefore Θψ (RXver + ǫ(k)B) ⊆ Θ✷ψ = Θψ R∞
Xver ⊆
(k)
(k)
Θψ RXver = Θ✷k ψ , and Θ✷k ψ ⊆ Θψ (RXver + ǫ(k)B)+
ǫθ B ⊆ Θ✷ψ + ǫθ B, and
Θ✷ψ ⊆ Θ✷k ψ ⊆ Θ✷ψ + ǫθ B.
Based on Lemma 6, with ǫp = maxpi
ǫθ =
|Api |
|bpi | ,
we obtain
ǫx ǫp maxi (kvi k)2
≤ ǫx ǫp max(kvi k)2 .
i
1 + ǫx ǫp maxi (kvi k)
Note that since kAk k2 converges to 0 for k → ∞ for
ρ(A) < 1, and since maxi (kvi k)2 is not increasing, the
error ǫθ also converges to 0.
14
| 3cs.SY
|
arXiv:1312.6461v3 [cs.LG] 19 Feb 2014
Nonparametric Weight Initialization of Neural
Networks via Integral Representation
Sho Sonoda, Noboru Murata
Schools of Advanced Science and Engineering
Waseda University
Shinjuku, Tokyo 169-8555
[email protected],
[email protected]
Abstract
A new initialization method for hidden parameters in a neural network is proposed. Derived from the integral representation of neural networks, a nonparametric probability distribution of hidden parameters is introduced. In this proposal,
hidden parameters are initialized by samples drawn from this distribution, and
output parameters are fitted by ordinary linear regression. Numerical experiments
show that backpropagation with proposed initialization converges faster than uniformly random initialization. Also it is shown that the proposed method achieves
enough accuracy by itself without backpropagation in some cases.
1
Introduction
In the backpropagation learning of a neural network, the initial weight parameters are crucial to its
final estimates. Since hidden parameters are put inside nonlinear activation functions, simultaneous
learning of all parameters by backpropagation is accompanied by a non-convex optimization problem. When the machine starts from an initial point far from the goal, the learning curve easily gets
stuck in local minima or lost in plateaus, and the machine fails to provide good performance.
Recently deep learning schemes draw tremendous attention for their overwhelming high performances for real world problems[1, 2]. Deep learning schemes consist of two stages: pre-training
and fine-tuning. The pre-training stage plays an important role for the convergence of the following fine-tuning stage. In pre-training, the weight parameters are constructed layer by layer, by
stacking unsupervised learning machines such as restricted Boltzmann machines[3] or denoising
autoencoders[4]. Despite the brilliant progress in application fields, theoretical interpretation of the
schemes is still an open question[5].
In this paper we introduce a new initialization/pre-training scheme which could avoid the nonconvex optimization problem. The key concept is the probability distribution of weight parameters
derived from Murata’s integral representation of neural networks[6]. The distribution gives an intuitive idea what the parameters represent and contains information about where efficient parameters
exist. Sampling from this distribution, we can initialize weight parameters more efficiently than just
sampling from a uniform distribution. In fact, for relatively simple or low dimensional problems,
our method by itself attains a high accuracy solution without backpropagation.
De Freitas et al.[7] also introduced a series of stochastic learning methods for neural networks based
on the Sequential Monte Carlo (SMC). In their methods the learning process is iterative and initial
parameters are given by less informative distributions such as normal distributions. On the other
hand we could draw the parameters from a data dependent distribution. Furthermore, in SMC, the
number of hidden units must be determined before the learning, while it is determined naturally in
our method.
1
2
Back ground and related works
One of the most naive initialization heuristics is to draw samples uniformly from an interval [−α, α].
Nguyen and Widrow[8] gave two fundamental points of view. First, since a typical activation function such as sigmoid and hyperbolic tangent is approximated as a linear function at its inflection
point, one should initialize the hidden parameters in such a way that the inputs for each hidden unit
are in the linear region. Second, since each hidden unit determines the slice of the Fourier transformed input space, that is, each individual hidden unit responds selectively to only the inputs whose
spatial frequency is in a particular band, one should initialize hidden parameters in such a way that
the corresponding frequency bands cover the possible input frequencies.
LeCun et al.[9] also emphasized the need to preset parameters in the linear region because parameters outside the linear region have small gradients and stray into more difficult nonlinear regions.
They focused on the curvature of input vectors and proposed to use α ∝ m−1/2 , where m is the fanin, or the dimensionality of input vectors. Shimodaira[10] proposed to initialize parameters such that
corresponding activation regions to cover whole the possible inputs. Linear algebraic techniques are
also employed. For example, Shepanski[11] used the pseudo inverse to determine the parameters of
linear approximated neural networks, and Yam and Chow[12] used the QR decomposition.
Integral transform viewpoints originated from more theoretical backgrounds than linear region viewpoints: the theoretical evaluation of the approximation power of neural networks. In the earliest
stage, purely functional analysis methods were employed. In 1957 Kolmogorov[13] showed that
any multivariate continuous functions can be exactly represented by sums of compositions of different continuous functions of only one variable. Inspired by the Kolmogorov’s theorem, HechtNielsen[14] and Kůrková[15] applied the idea to neural networks, which are sums of compositions
of the same sigmoid function. Sprecher[16] gave more constructive version of the proof and later
implemented the improved proof as a learning algorithm of neural networks[17].
In 1989 the universal approximation property of single layer neural networks has been investigated and the integral transform aspects emerged. Carroll and Dickinson[18] introduced the
Radon transform and Funahashi[19] used the Fourier analysis and the Paley-Weiner theory, whereas
Cybenko[20] employed the Hahn-Banach and Riesz Representation theorems. In the following
years, upper bounds of the approximation error were investigated[21, 22, 6]. Barron[22] refined
the Jones’ result[21] using the weighted Fourier transform. Kůrková[15] later developed the general theory of integral transforms. Inspired by the Barron’s result, Murata[6] introduced a family
of integral transforms defined by ridge functions, which are regarded as a hybrid of the Radon and
wavelet transforms. Candés[23] inherited Murata’s transforms and developed ridgelets, which was
the beginning of the series of multiscale “-lets” analysis[24].
Those multiscale viewpoints also inherits the selective activation properties of neural networks. Denoeux and Lengellé[25] proposed to collect K prototype vectors as initial hidden parameters. Each
prototype pk is drawn from its corresponding cluster Ck , where the clusters {Ck }K
k=1 are formed
in a stereographically projected input space. In this manner each prototype pk comes to selectively
respond to the input vectors x which belongs to the cluster Ck .
This study is based on the integral transform viewpoint, and proposes a new way for practical implementation. Although integral transforms have been well studied as theoretical integral representations of neural networks, practical implementations for training have been merely done. However
integral representations have big advantage over linear region viewpoints in that they can give global
directions how each neural units should behave, while the latter only give local directions.
3
3.1
Nonparametric weight initialization via integral transform
Sampling based two-stage learning
Let g : Rm → R be a neural network with a single hidden layer expressed as
g(x) =
J
X
wj φ (aj · x − bj ) + w0 ,
j=1
2
(1)
where the map φ is called the activation function; aj and bj are called hidden parameters, and wj are
1
output parameters. With an ordinary sigmoid function σ(z) := 1+exp(−z)
, the activation function φ
is supposed to be the sigmoid pair in the form
1
{σ(z + h) − σ(z − h)} , (h > 0),
(2)
φ(z) :=
H
where H := σ(h) − σ(−h) normalizes the maximum value of φ to be one.
We consider an oracle distribution p(a, b) of hidden parameters. If such a distribution exists, we
can sample and fix these hidden parameters according to p(a, b) first, and then we could fit the rest
output parameters by ordinary linear regression. We call this two-stage framework as Sampling
Regression (SR) learning.
The candidates of p(a, b) could be some parametric distributions such as normal distributions or
uniform distributions. In the following sections we derive a data dependent distribution from an
integral representation of neural networks.
3.2
Integral representations of neural networks
Consider approximating a map f : Rm → R with a neural network. Murata[6] defined an integral
transform T of f with respect to a decomposing kernel φd as
Z
1
T (a, b) :=
φd (a · x − b)f (x)dx,
(3)
C Rm
where C is a normalizing constant. Murata also showed that given the decomposing kernel φd , there
exists the associating composing kernel φc such that for any f ∈ L1 (Rm ) ∩ Lp (Rm )(1 ≤ p ≤ ∞),
the inversion formula
Z
2
f (x) = lim
φ∗c (a · x − b)T (a, b)e−|a| dadb in Lp ,
(4)
→0
Rm+1
2
∗
holds (Th.1 in [6]) where · denotes the complex conjugate. The convergence factor e−|a| is
omitted when T ∈ L1 (Rm+1 ), which is attained when f is compactly supported and C m,α -Hölder
continuous with 0 < α ≤ 1 (Th.3 in [6]), or compactly supported and bounded C m+1 -smooth
(Cor.2 in [6]).
In particular one can set a composing kernel φc as a sigmoid pair φ given in Eq.2 and the associating
decomposing kernel as:
(m)
ρ (z)
if m is even
φd (z) =
,
(5)
ρ(m+1) (z) otherwise
where ρ is a nonnegative C ∞ -smooth function whose support is in the interval
[−1,
1]. Such a ρ does
exist and is known as a mollifier[26].The standard mollifier ρ(z) = exp
example.
1
z 2 −1
is a well-known
Hereafter we assume φc is a sigmoid pair and φd is the corresponding derivative of the standard
mollifier. We also assume that our target f is a bounded and compactly supported C (m+1) -smooth
function. Then the integral transform
R T of f is absolutely integrable and the inversion formula is
reduced to the direct form f (x) = Rm+1 φ∗c (a · x − b)T (a, b)dadb.
Let τ (a, b) be a probability distribution function over Rm+1 which is proportional to |T (a, b)|,
and c(a, b) be satisfying c(a, b)τ (a, b) = T (a, b) for all (a, b) ∈ Rm+1 . With this notations, the
inversion formula is rewritten as the expectation form with respect to τ (a, b), that is,
Z
f (x) =
c(a, b)φc (a · x − b)τ (a, b)dadb.
(6)
Rm+1
The expression implies the finite sum
J
1X
c(aj , bj )φc (aj · x − bj ),
gJ (x) :=
J j=1
i.i.d.
(aj , bj ) ∼ τ (a, b)
(7)
converges to f in mean square as J → ∞, i.e. E[gJ ] = f and Var[gJ ] < ∞ holds for any J (Th.2
in [6]). Here gJ is a neural network with 2J hidden units, therefore we can regard the inversion
formula as an integral representation of neural networks.
3
3.3
Practical calculation of the integral transform
Now we attempt to make use of the integral transform |T (a, b)| as an oracle distribution p(a, b) of
hidden parameters. Although the distribution is given in the explicit form as we saw in the preceding
section, further refinements are required for practical calculation.
m
Given a set {(xn , yn )}N
n=1 ⊂ R ×R of input and output pairs, T (a, b) is empirically approximated
as
N
1 X
T (a, b) ≈
φd (a · xn − b)yn ,
(8)
Z n=1
with some constant Z > 0 which is hard to calculate exactly. In fact sampling algorithms such
as the acceptance-rejection method[27] and Markov chain Monte Carlo method[27] work with any
unnormarized distribution because they only evaluate the ratio between probability values. Note
that the approximation converges to the exact T (a, b) in probability by the law of large numbers
only when the input vectors are i.i.d. samples from a uniform distribution.
As a decomposing kernel φd we make use of the k-th order derivative of the standard mollifier
ρ(z) = exp z21−1 where k = m if m is even and k = m + 1 otherwise. The k-th derivative ρ(k) (z)
of the mollifier takes the form
Pk (z)
ρ(k) (z) = 2
ρ(z) (k = 0, 1, 2, · · · ),
(9)
(z − 1)2k
where Pk (z) denotes a polynomial of z which is calculated by the following recurrence formula:
P0 (z) ≡
Pk+1 (z)
=
1 (const.),
Pk0 (z)(z 4 − 2z 2 + 1) + Pk (z) −4kz 3 + 2(2k − 1)z .
(10)
(11)
The higher order derivatives of a mollifier has more rapid oscillations in the neighbourhoods of both
edges of its support.
m
Given a data set D := {(xn , yn )}N
n=1 ⊂ R × R, our Sampling Regression method is summarized
as below:
0. Preliminary stage: Calculate ρ(k) (z) according to Eq.9, Eq.10 and Eq.11, where k = m
if m is even and k = m + 1 otherwise. Then T (a, b) is calculated by Eq.8 with setting
φd = ρ(k) . As we noted above, one can choose arbitrary Z > 0.
1. Sampling stage: Draw J samples {(aj , bj )}Jj=1 from the probability distribution τ (a, b) ∝
|T (a, b)| by acceptance-rejection method, where J denotes the number of hidden (sigmoid
pair) units. Then we obtain the hidden parameters {(aj , bj )}Jj=1 .
2. Regression stage: Let φjn := φd (aj · xn − bj ) for all j = 1, · · · , J and n = 1, · · · , N .
PJ
Solve the system of linear equations yn = j=1 wj φjn + w0 (n = 1 · · · N ) with respect
to {wj }Jj=0 . Then we obtain the output parameters {wj }Jj=0 .
3.4
For more efficient sampling
Generally |T (a, b)| is ill-shaped and sampling from the distribution is difficult. For example in Fig.1
Left, samples drawn from |T (a, b)| of f (x) = sin 2πx with x ∈ [−1, 1] is plotted. Whereas in Fig.1
Right, the same distribution is transformed to another (α, β)-coordinate system (which is explained
below). The support of the distribution is reshaped into a rectangular, which implies sampling from
|T (α, β)| is easier than doing from |T (a, b)|.
This ill-shapeness is formulated as following proposition.
Proposition 3.1. Suppose the objective function f (x) has a compact support, then the support of
its transform T (a, b) is in the region Ω := {(a, b) | |b| ≤ M kak + 1} with M := maxx∈supp f kxk.
Proof. Recall the support of φd is included in the interval [−1, 1], therefore for any a, b and x,
φd (a · x − b) 6= 0 implies |a · x − b| < 1. The latter condition is equivalently deformed to
a · x − 1 < b < a · x + 1, which implies |b| < |a · x| + 1. By the compact support assumption of f ,
4
Figure 1: Sample parameters drawn from |T (a, b)| of sin 2πx case. Red lines indicate the theoretical boundary b = ±(M kak + 1) of the support of |T (a, b)|. Left: |T (a, b)| has a non-convex
support, in which case sampling is inefficient. Right: The same sample points are plotted in the
coordinate transformed (α, β)-space. Coordinate lines are deformed to lattice lines, and |T (α, β)|
has a rectangular support.
taking the maximum with respect to x leads to |b| < M kak + 1. By tracking back the inferences,
for any a, b and x ∈ supp f ,
(a, b) ∈
/ Ω ⇒ φd (a · x − b) = 0.
(12)
Since for any x ∈
/ supp f , the integrand of T (a, b) is always zero, the integration domain of T (a, b)
can be restricted into supp f . Therefore by Eq.12,
T (a, b) 6= 0 ⇒ (a, b) ∈ Ω
(13)
holds, which comes to the conclusion: supp T ⊂ Ω.
In a relatively high dimensional input case, sampling in the coordinate transformed (α, β)-space
a
α
=
,
(14)
b
(M kαk + 1)β
is more efficient than sampling in the (a, b)-space because the shape of the support of |T (a, b)| in the
(α, β)-space is rectangular (see, Fig.1) and therefore the proposal distribution is expected to reduce
miss proposals, out of the support.
In case that the coordinate transform technique is not enough, it is worth sampling from each component distribution. Namely, the empirically approximated |T (a, b)| is bounded above by a mixture
distribution:
|T (a, b)|
≈
N
1 X
yn φd (a · xn − b)
Z n=1
≤
∝
N
1 X
|yn ||φd (a · xn − b)|,
Z n=1
N
X
ηn pn (a, b),
(15)
n=1
where pn (a, b) ∝ |φd (a·xn −b)| is a component distribution and ηn ∝ |yn | is a mixing probabilities.
In addition, an upper bound of φd is given by the form
log |φd (z)| ≤ Az 2 + B,
for some A > 0 and B.
5
(16)
4
Experimental results
We conducted three sets of experiments comparing three types of learning methods:
BP Whole parameters are initialized by samples from a uniform distribution, and trained by
BackPropagation.
SBP Hidden parameters are initialized by Sampling from |T (a, b)|; and the rest output parameters are initialized by samples from a uniform distribution. Then whole parameters are
trained by BackPropagation.
SR Hidden parameters are determined by Sampling from |T (a, b)|; the rest output parameters
are fitted by linear Regression.
In order to compare the ability of the three methods, we conducted three experiments on three different problems: One-dimensional complicated curve regression, Multidimensional Boolean functions
approximation and Real world data classification.
4.1
One-dimensional complicated curve regression - Topologist’s sine curve sin 2π/x
Figure 2: Training results of three methods for fitting the topologist’s sine curve sin 2π/x. Left:
SR (solid black line) by itself achieved the highest accuracy without the iterative learning, whereas
SBP (dashed red line) converged to lower RMSE than BP (dotted green line). Right: The original
curve (upper left) has high frequencies around the origin. SR (upper right) followed such a dynamic
variation of frequency better than other two methods. SBP (lower left) roughly approximated the
curve with noise. BP (lower right) only fitted moderate part of the curve.
First we performed one-dimensional curve regression. The objective function is a two-sided topologists’s sine curve (TSC) f (x) := sin 2π/x defined on the interval [−1, 1] whose indefiniteness at
zero is removed by defining f (0) = 0. The TSC is such a complicated curve whose spatial frequency
gets arbitrary high as x tends to zero. For training, 201 points were sampled from the domain [−1, 1]
in equidistant manner. The number of hidden parameters were fixed to 100 in each model. Note that
relatively redundant quantity of parameters are needed for our sampling initialization scheme to obtain good parameters. The output function was set to linear and the batch learning was performed
by BFGS quasi-Newton method. Uniformly random initialization parameters for BP and SBP were
drawn from the interval [−1, 1]. Sampling from |T (a, b)| was performed by acceptance-rejection
method.
In Fig.2 Left, the Root Mean Squared Error (RMSE) in training phase of three methods are shown.
The solid black line corresponds to the result by SR, which by itself achieved the highest accuracy
without iterative learnings. The dashed red line corresponds to the result by SBP, and it converged
to lower RMSE than that of BP depicted in the dotted green line. In Fig.2 Right, fitting results of
the three methods are shown. As we noted the original curve (upper left) has numerical instability
6
around the origin, therefore it is difficult to fit the curve. SR (upper right) approximated the original
curve well except around the origin, while other two methods, SBP (lower left) and BP (lower right)
could just partly fit the original curve. In this experiment, we examined the flexibility of our method
by fitting a complicated curve. The experimental result supports that the oracle distribution gave
advantageous directions.
4.2
Multidimensional Boolean functions approximation - Combined AND, OR and XOR
Second we performed a binary problem with twodimensional input and three-dimensional output. Output vectors are composed of three logical functions:
F (x, y) := (xANDy, xORy, xXORy). Therefore
the total number of data is just four: (x, y) ∈
{(0, 0), (0, 1), (1, 0), (1, 1)}. The number of hidden
units were fixed to 10. The output function was set to
sigmoid and the loss function was set to cross-entropy.
Uniformly random initialization parameters for BP and
SBP were drawn from the interval [−1, 1]. Sampling
from |T (a, b)| was performed by acceptance-rejection
method.
In Fig.3 both the cross-entropy curves and classification error rates are depicted in thin and thick lines respectively. The solid black line corresponds to the results by SR, which achieved the perfectly correct answer from the beginning. The dashed red line cor- Figure 3: Cross-entropy curves (thin lines)
responds to the results by SBP, which also attained and classification error rates (thick lines).
the perfect solution faster than BP. The dotted green SR (solid black line) achieved the perfectly
line corresponds to the results by BP, which cost 100 correct answer from the beginning. SBP
iterations of learning to give the correct answer. In (dashed red line) also attained the perfect
this experiment we have validated that the proposed solution faster than BP. BP (dotted green
method works well with multiclass classification prob- line) costed 100 iterations of learning to
lems. The quick convergence of SBP indicates that give the correct answer.
|T (a, b)| contains advantageous information on the training examples to the uniform distribution.
4.3
MNIST
Finally we examined a real classification problem using the MNIST data set[28]. The data set consists
of 60, 000 training examples and 10, 000 test examples. Each input vector is a 256-level gray-scaled
(28 × 28 =)784-pixel image of a handwritten digit.
The corresponding label is one of 10 digits. We implemented these labels as 10-dimensional binary vectors
whose components are chosen randomly with equivalent probability for one and zero. We used randomly
sampled 15, 000 training examples for training and
whole 10, 000 testing examples for testing. The number of hidden units were fixed to 300, which is the
same size as used in the previous study of LeCun et
al.[29]. Note that J sigmoid pairs corresponds to 2J
sigmoid units, therefore we used 150 sigmoid pairs for
SR and SBP, and 300 sigmoid units for BP. The output function was set to sigmoid and the loss function
was set to cross-entropy. In obedience to LeCun et
al.[9], input vectors were normalized and randomly initialized parameters for BP and SBP were drawn from
uniform distribution with mean zero and standard deviation 784−1/2 ≈ 0.0357.
7
Figure 4: Test classification error rates
for MNIST dataset. SR (black real line)
marked 23.0% at the beginning and finished 9.94%, the error reascent suggests
that SR may have overfitted. SBP (red
dashed line) reduced the fastest and finished 8.30%. BP (green dotted line) declined the slowest and finished 8.77%.
Direct sampling from |T (a, b)| is numerically difficult because the differential order of its decomposing kernel φd piles up as high as 784-th order. We abandoned rigorous sampling and tried sampling
from a mixture annealed distribution. As described in Eq.15, we regarded |T (a, b)| as a mixture of
|φd (a · xn − b)|. By making use of the log boundary given by Eq.16, we numerically approximated
φd (z) from above
log |φd (z)| ≤ 2800z 2 − 800, (|z| < 1),
(17)
and drew samples from an easier component distribution pn (a, b) ∝ exp{2800(a · xn − b)2 −
800}. Details of the sampling technique is explained in A.2. The sampling procedure scales linearly
with the dimensionality of the input space (784) and the number of required hidden units (150)
respectively. In particular it scales constantly with the number of the training examples.
The following linear regression was conducted by singular value demcomposition (SVD), which
generally costs O(mn2 ) operations, assuming m ≥ n, for decomposing a m × n-matrix. In our case
m corresponds to the number of the training examples (15, 000) and n corresponds to the number of
hidden units (300). At last backpropagation learning was performed by stochastic gradient descent
(SGD) with adaptive learning rates and diagonal approximated Hessian[30]. The experiment was
performed in R[31] on a Xeon X5660 2.8GHz with 50GB memory.
In Fig.4 the classification error rates for test examples are depicted. The black real line corresponds
to the results by SR, which marked the lowest error rate (23.0%) of the three at the beginning, and
finished 9.94% after 45, 000 iterations of SGD training. The training process was not monotonically
decreasing in the early stage of training, it appears that the SR initialization overfitted to some extent.
The red dashed line corresponds to the results by SBP, which marked the steepest error reduction in
the first 5, 000 iterations of SGD training and finished 8.30%. The green dotted line corresponds to
the results by BP, which declined the slowest in the early stage of training and finished 8.77%.
In Tab.1 the training time from initialization through SGD training is listed. The sampling step in
SR ran faster than the following regression and SGD steps. In addition, the sampling time of SR
and SBP was as fast as the sampling time of BP. As we expected, the regression step in SR, which
scales linearly with the amount of the data, cost much more time than the sampling step did. The
SGD step also cost, however each step cost around merely 0.05 seconds, and it would be shorten if
the initial parameters had better accuracy.
Table 1: Training Times for MNIST
Method
SR
SBP
BP
Sampling [s]
1.15 × 10−2
1.14 × 10−2
1.15 × 10−2
Regression [s]
2.60
-
BP by SGD (45, 000 itrs.) [s]
2.00 × 103
2.31 × 103
2.67 × 103
In this experiment, we confirmed that the proposed method still works for real world data with the
aid of an annealed sampling technique. Although SR showed an overfitting aspects, the fastest convergence of SBP supports that the oracle distribution gave meaningful parameters, and the annealed
sampling technique could draw meaningful samples. Hence the overfitting of SR possibly comes
from regression step, which suggests the necessity for further blushing up of regression technique.
In addition, our further experiments also indicated that when the number of hidden units increased
to 6, 000, the initial test error rate scored 3.66%, which is smaller than the previously reported error
rates 4.7% by LeCun et al.[29] with 300 hidden units.
5
Conclusion and future directions
In this paper, we introduced a two-stage weight initialization method for backpropagation: sampling hidden parameters from the oracle distribution and fitting output parameters by ordinary linear regression. Based on the integral representation of neural networks, we constructed our oracle
distributions from given data in a nonparametric way. Since the shapes of those distributions are
not simple in high dimensional input cases, we also discussed some numerical techniques such as
the coordinate transform and the mixture approximation of the oracle distributions. We performed
three numerical experiments: complicated curve regression, Boolean function approximation, and
8
handwritten digit classification. Those experiments show that our initialization method works well
with backpropagation. In particular for the low dimensional problems, well-sampled parameters
by themselves achieve good accuracy without any parameter updates by backpropagation. For the
handwritten digit classification problem, the proposed method works better than random initialization.
Sampling learning methods inevitably come with redundant hidden units since drawing good samples usually requires a large quantity of trial. Therefore the model shrinking algorithms such as
pruning, sparse regression, dimension reduction and feature selection are naturally compatible to
the proposed method.
Although plenty of integral transforms have been used for theoretical analysis of neural networks,
numerical implementations, in particular sampling approaches are merely done. Even theoretical
calculations often lack practical applicability, for example a higher order of derivative in our case,
each integral representation interprets different aspects of neural networks. Further Monte Carlo
discretization of other integral representations is an important future work.
In the deep learning context, it is said that the deep structure remedies the difficulty of a problem by
multilayered superpositions of simple information transformations. We conjecture that the complexity of high dimensional oracle distributions can be decomposed into relatively simple distributions
in each layer of the deep structure. Therefore, extending our method to the multilayered structure is
our important future work.
Acknowledgments
The authors are grateful to Hideitsu Hino for his incisive comments on the paper. They also thank
to Mitsuhiro Seki for having constructive discussions with them.
References
[1] Q. Le, M. Ranzato, R. Monga, M. Devin, K. Chen, G. Corrado, J. Dean, and A. Ng. Building high-level
features using large scale unsupervised learning. In J. Langford and J. Pineau, editors, ICML2012, pages
81–88, 2012.
[2] G. E. Dahl, D. Yu, L. Deng, and A. Acero. Context-dependent pre-trained deep neural networks for largevocabulary speech recognition. IEEE Transactions on Audio, Speech & Language Processing, 20(1),
2012.
[3] G. E. Hinton, S. Osindero, and Y.-W. Teh. A fast learning algorithm for deep belief nets. Neural Computation, 18(7):1527–1554, 2006.
[4] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle. Greedy layer-wise training of deep networks. In
Bernhard Scholkopf, John Platt, and Thomas Hoffman, editors, NIPS ’06, pages 153–160, 2007.
[5] Y. Bengio., A. Courville., and P. Vincent. Representation learning: A review and new perspectives. PAMI,
35(8):1798–1828, 2013.
[6] N. Murata. An integral representation of functions using three-layered networks and their approximation
bounds. Neural Networks, 9(6):947–956, 1996.
[7] J. F. G. de Freitas, M. Niranjan, A. H. Gee, and A. Doucet. Sequential monte carlo methods to train neural
network models. Neural Computation, 12(4):955–993, 2000.
[8] D. Nguyen and B. Widrow. Improving the learning speed of 2-layer neural networks by choosing initial
values of the adaptive weights. In IJCNN 1990, volume 3, pages 21–26, 1990.
[9] Y. LeCun, L. Bottou, G. B. Orr, and K.-R. Müller. Efficient backprop. In G. Montavon, G. B. Orr, and
K.-R. Müller, editors, Neural Networks: Tricks of the Trade (2nd ed.), pages 9–48. Springer, 2012.
[10] H. Shimodaira. A weight value initialization method for improving learning performance of the backpropagation algorithm in neural networks. In ICTAI1994, pages 672–675, 1994.
[11] J. F. Shepanski. Fast learning in artificial neural systems: multilayer perceptron training using optimal
estimation. In ICNN1988, volume 1, pages 465–472, 1988.
[12] J. Y. F. Yam and T. W. S. Chow. A weight initialization method for improving training speed in feedforward neural network. Neurocomputing, 30(1-4):219–232, 2000.
[13] A. N. Kolmogorov. On the representation of continuous functions of several variables by superposition
of continuous functions of one variable and addition. Doklady Akademii Nauk SSSR, 114:369–373, 1957.
9
[14] R. Hecht-Nielsen. Kolmogorov’s mapping neural network existence theorem. In ICNN1987, volume III,
pages 11–13, 1987.
[15] V. Kůrková. Kolmogorov’s theorem is relevant. Neural Computation, 3(4):617–622, 1991.
[16] D. A. Sprecher. On the structure of continuous functions of several variables. Transactions of the AMS,
115:340–355, 1965.
[17] D. A. Sprecher. A numerical implementation of kolmogorov’s superpositions.
9(5):765–772, 1996.
Neural Networks,
[18] S. M. Carroll and B. W. Dickinson. Construction of neural nets using the radon transform. In IJCNN
1989, volume 1, pages 607–611, 1989.
[19] K. Funahashi. On the approximate realization of continuous mappings by neural networks. Neural
Networks, 2(3):183–192, 1989.
[20] G. Cybenko. Approximation by superpositions of a sigmoidal function. MCSS, 2(4):303–314, 1992.
[21] L. K. Jones. A simple lemma on greedy approximation in hilbert space and convergence rates for projection pursuit regression and neural network training. The Annals of Statistics, 20(1):608–613, 1992.
[22] A.R. Barron. Universal approximation bounds for superpositions of a sigmoidal function. IEEE Transactions on Information Theory, 39(3):930–945, 1993.
[23] E. J. Candes. Ridgelets: Theory and Applications. PhD thesis, Standford University, 1998.
[24] J. Fadili and J. L. Starck. Curvelets and ridgelets. In R. A. Meyers, editor, Encyclopedia of Complexity
and Systems Science, volume 3, pages 1718–1738. Springer New York, 2009.
[25] T. Denoeux and R. Lengelle. Initializing back propagation networks with prototypes. Neural Networks,
6(3):351–363, 1993.
[26] K. O. Friedrichs. The identity of weak and strong extensions of differential operators. Transactions of the
AMS, 55(1):132–151, 1944.
[27] C. M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006.
[28] C. J. C. Burges Y. LeCun, C. Cortes. The mnist database of handwritten digits. http://yann.lecun.
com/exdb/mnist/.
[29] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition.
In Proceedings of the IEEE, volume 11, pages 2278–2324, 1998.
[30] L. Bottou. Online algorithms and stochastic approximations. In D. Saad, editor, Online Learning and
Neural Networks. Cambridge University Press, 1998.
[31] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical
Computing, Vienna, Austria, 2013.
Supplementary materials
A
Sampling recipes
Sampling hidden parameter (a, b)’s from the oracle distribution p(a, b) demands a little ingenuity. In our experiments, we have implemented two sampling procedures: a rigorous but naive, computationally inefficient way
and an approximative/ad hoc but quick and well-performing way. Although both work quickly and accurately
in a low dimensional input problem, only the latter works in a high dimensional problem such as MNIST.
A.1
Sampling from rigorous oracle distribution
Given a decomposing kernel φd (z) := ρ(m) (z), we employed acceptance-rejection (AR) method directly on
rigorous sampling from p(a, b) On a proposal distribution q(a, b), we employed uniform distribution. We
assume here that the support Ω of proposal distribution q(a, b) has been adjusted to cover the mass of p(a, b) as
tight as possible, and the infimum k := inf p(a, b)/q(a, b) has been estimated. Then our sampling procedure
is conducted according to the following Alg.1.
Note that in a high dimensional case, the estimation accuracy of k and the tightness of Ω affects the sampling
efficiency and accuracy materially. In fact, the expectation number of trial to obtain one sample by AR is k
times, which gets exponentially large as the dimensionality increases. Since the support of the oracle distribution p(a, b) is not rectangular, sampling from coordinate transformed p(α, β) remedies the difficulty. In
addition, the high order differentiation in the decomposing kernel φd cause numerical unstability.
10
Algorithm 1 Rigorous sampling according to ordinary acceptance-rejection method.
repeat
draw proposal point (a∗ , b∗ ) ∼ q(a, b).
draw uniformly random value u from the interval [0, 1].
p(a∗ ,b∗ )
if u ≤ kq(a
∗ ,b∗ ) then
return (a∗ , b∗ ) {accept}
else
do nothing {reject}
end if
until acceptance occurs.
A.2
Sampling from mixture annealed distribution
In order to overcome the high dimensional sampling difficulty, we approximately regarded p(a, b) as a mixture
P
distribution p(a, b) ≈ N
n=1 ηn pn (a, b) (as described in Eq.15) and conducted two-step sampling: first choose
one component distribution pn (a, b) according to the mixing probability ηn ∝ |yn |, second draw a sample
(a, b) from chosen component distribution pn (a, b).
Sampling from pn (a, b) ∝ |φd (a·xn −b)| holds another difficulty due to its high order differentiation in φd (z).
According to its upper bound evaluation (Eq.16), a high order derivative ρ(m) (z)(= φd (z)) has its almost all
mass around both edge of its domain interval [−1, 1] and almost no mass in the middle of the domain (see
Fig.5 Left). Hence we approximated, or annealed, ρ(m) (z) by a beta distribution, which could model extreme
skewness of ρ(m) (z) (e.g., Beta(z; 100, 3); see Fig.5 Right). Then we conducted further steps of sampling:
first sample z ∈ [−1, 1] according to the annealing beta distribution, then sample a and b under the restriction
z = a · xn − b.
Figure 5: 10-th order derivative ρ(10) (z) of mollifier. Left: ρ(10) (z) has almost all mass, with high
frequency, at both ends, and no mass in the middle of domain. Right: The right half of |ρ(10) (z)| is
approximated by beta distribution Beta(z; 100, 3) (red line).
Obviously the mixture approximation gives rise to poor restriction and virtual indefiniteness of (a, b). Since the
rigorous computation establishes all relations between (a, b) and all xn ’s, whereas the mixture approximation
does just one relation between (a, b) and one particular xn . We introduced two additional assumptions. First,
a is parallel to given xn . Since a always appears in the form a · xn , only the parallel component of a could
have any effect (on one particular xn ), hence we eliminated the extra freedom in the orthogonal component.
Second, the norm a := kak has similar scale to the distances kxn − xm k between input vectors. Since a
controls the spatial frequency of a hidden unit, it determines how broad the hidden unit covers the part of the
input space. Namely, a controls which input vectors are selectively responded by the unit. Therefore, in order
to avoid such an isolation case that an unit responds for only one input vector, we assumed a is no smaller than
the distance between input vectors. In this procedure we set a as a distance kxn − xm k of randomly selected
two input examples xn and xm . We denote this procedure simply by a ∼ p(kx − x0 k). Once a is fixed with
these assumptions, b is determined as b = a · xn − z.
11
Given shape parameters α, β of the beta distribution Beta(z; α, β), one cycle of our second sampling method
is summarized as Alg.2. This method consists of no more expensive steps. It scales linearly with the dimensionality of the input space and the number of required sample parameters respectively. Moreover, it does not
depends on the size of the training data.
Algorithm 2 Quick sampling from mixture annealed distribution (for high dimensional use.)
choose a suffix n of xn according to the mixing probability ηn .
draw ζ ∼ Beta(z; α, β) and k ∼ Bernoulli(k; p = 0.5)
z ← (−1)k ζ
set length a ∼ p(x − x0 ).
a ← axn /kxn k.
b ← a · xn − z.
12
| 9cs.NE
|
1
The Explicit Coding Rate Region of
Symmetric Multilevel Diversity Coding
arXiv:1801.02376v1 [cs.IT] 8 Jan 2018
Tao Guo and Raymond W. Yeung
Abstract
It is well known that superposition coding, namely separately encoding the independent sources, is optimal for
symmetric multilevel diversity coding (SMDC) (Yeung-Zhang 1999). However, the characterization of the coding rate
region therein involves uncountably many linear inequalities and the constant term (i.e., the lower bound) in each
inequality is given in terms of the solution of a linear optimization problem. Thus this implicit characterization of
the coding rate region does not enable the determination of the achievability of a given rate tuple. In this paper,
we first obtain closed-form expressions of these uncountably many inequalities. Then we identify a finite subset of
inequalities that is sufficient for characterizing the coding rate region. This gives an explicit characterization of the
coding rate region. We further show by the symmetry of the problem that only a much smaller subset of this finite
set of inequalities needs to be verified in determining the achievability of a given rate tuple. Yet, the cardinality of
this smaller set grows at least exponentially fast with L. We also present a subset entropy inequality, which together
with our explicit characterization of the coding rate region, is sufficient for proving the optimality of superposition
coding.
Index Terms
Symmetric multilevel diversity coding, superposition coding, network coding, closed-form, distributed data storage, robust network communication.
I. I NTRODUCTION
Symmetric multilevel diversity coding (SMDC) was introduced by Roche et al. [1] for applications in distributed
data storage and robust network communication. It is also a special case of multi-source network coding [2], [3]. In
this problem, there are L (L ≥ 2) independent discrete memoryless sources {Xl (t) : t = 1, 2, · · · }, l = 1, 2, · · · , L,
where for each l, Xl (t) are independent and identically distributed copies of a generic random variable Xl . The
importance of the sources is in the order X1 (t) > X2 (t) > · · · > XL (t). The sources are encoded by L encoders.
There are totally 2L − 1 decoders, each of which has access to a distinct subset of the encoders. A decoder which
can access any α encoders, called a Level α decoder, is required to reconstruct the first α sources. Such a system
is called a symmetric L-level diversity coding system. The system is symmetric in the sense that the problem is
unchanged by permuting the L encoders, which is evident from the reconstruction requirements of the decoders.
The SMDC problem was treated for L = 3 in [1] and in full generality by Yeung and Zhang [4], where a coding
method called superposition coding was proved to be optimal. In this method, the independent sources {Xl (t)},
l = 1, 2, · · · , L are encoded separately. The problem has subsequently been generalized in different directions. The
2
secure communication setting was considered by Balasubramanian et al. [5] and Jiang et al. [6]. In [6], they also
extended the original SMDC setting by introducing an all-access encoder which is accessible by all the decoders.
In both of the above settings, superposition coding is shown to be optimal. Xiao et al. [7] studied the problem
of distributed multilevel diversity coding where each source is decomposed into L components, each of which is
accessed by one distinct encoder. Tian and Liu [8] considered the problem with regeneration, where the storage
versus repair-bandwidth tradeoff was investigated. Mohajer et al. [9] considered the asymmetric multilevel diversity
coding problem and proved that superposition coding is in general suboptimal. In the current paper, we focus on
some fundamental issues pertaining to the original SMDC problem discussed in [1], [4].
It was proved in [1] that superposition coding is optimal for L = 3, and the corresponding coding rate region,
referred to as the superposition coding rate region, can be explicitly characterized by 10 linear inequalities in the
coding rates of the 3 encoders. Thus, the achievability of any given rate triple can be determined by verifying these
10 inequalities.
However, the optimality proof in [1] is not readily generalizable to a general L. Here is an outline of the proof in
[1]. The superposition coding rate region is first characterized by the aforementioned 10 inequalities. This involves
the determination of the extreme points of the region. Then the necessity of these 10 inequalities are established by
means of conventional techniques for proving converse coding theorems. The difficulty for generalizing the proof
to a general L is two-fold:
1) It is observed through computation that the number of extreme points of the superposition coding rate region
grows with L. As such, it is impossible to determine all of them for a general L.
2) For a fixed L, once the superposition coding rate region is characterized by a finite set of linear inequalities,
their necessity needs to be proved. With conventional techniques, this needs to be done for each inequality in
a way that depends on the coefficients of coding rates. It is observed through computation that the number of
these inequalities grows with L. Therefore, for a general L, it is not possible to prove the necessity of all of
these inequalities.
For a fixed L, the extreme points of the superposition coding rate region and the set of rate constraints characterizing
the region can in principle be found by computation. However, the complexity grows very quickly with L and
becomes prohibitive for L larger than 5 or so.
In [4], the optimality of superposition coding was established for a general L by means of a highly sophisticated
method that does not involve any explicit characterization of the coding rate region. Instead of a fixed L, the problem
is tackled for a general L. As L is not fixed, the number of linear inequalities needed for the characterization
of the superposition coding rate region may be unbounded. To get around the problem, the coding rate region is
characterized by an uncountable collection of linear inequalities, where for each inequality, the coefficients associated
with the rates are arbitrary nonnegative real numbers with at least one of them being nonzero. The constant terms
(i.e., the lower bounds) in these inequalities are given implicitly in terms of the solution of a common linear
optimization problem with the coefficients associated with the rates as parameters. In other words, although the
coding rate region is characterized by uncountably many linear inequalities, they have a common form and the
necessity of these inequalities can be established in a unified manner.
3
Although the optimality of superposition coding for a general L has been established in [4], this result does
not yield an explicit characterization of the coding rate region for any fixed L. In particular, it does not enable
the determination of the achievability of a given rate tuple, even for a fixed L, for the following two reasons.
First, the characterization of the coding rate region in [4] involves an uncountable number of inequalities. Second,
each inequality in the characterization is implicit, and can be made explicit only by solving a linear optimization
problem.
In the present paper, we develop fundamental results pertaining to SMDC. Our main contributions are summarized
as follows:
1) We obtain an explicit characterization of the coding rate region for a general L. This is done by first solving
in closed form the linear optimization problem in [4] that gives an implicit characterization of the coding rate
region. Then among all the uncountably many inequalities involved in characterizing the coding rate region,
we identify a finite subset that is sufficient for characterizing the coding rate region. It is further proved that
there is no redundancy in this finite set of inequalities. Thus for a fixed L, the achievability of any given rate
tuple can be determined.
2) By taking advantage of the symmetry of the problem, we show that in determining the achievability of a
given rate tuple, it suffices to verify a much smaller subset of the set of inequalities identified in 1). Yet, the
cardinality of this smaller set of inequalities grows at least exponentially fast with L. This reveals the inherent
complexity of the problem.
3) A subset entropy inequality, which plays a key role in the converse proof in [4], requires a painstaking and
extremely technical proof. We present a weaker version of this subset entropy inequality whose proof is
considerably simpler. With our explicit characterization of the coding rate region, this weaker version of the
subset entropy inequality is sufficient for proving the optimality of superposition coding.
The rest of the paper is organized as follows. We first formulate the problem and state some existing results in
Section II. In Section III, we present a closed-form solution of the linear optimization problem in [4] and establish
some basic properties of the solution. In Section IV, we identify a finite set of inequalities that characterizes the
superposition coding rate region and show that this set contains no redundancy. In Section V, we further identify
a subset of inequalities we need to verify in determining the achievability of a given rate tuple. We also provide a
lower bound and an upper bound on the cardinality of this set. In Section VI, we present a weaker version of the
subset entropy inequality in [4]. We conclude the paper in Section VII. Some essential proofs can be found in the
appendices.
II. P ROBLEM F ORMULATION
AND
E XISTING R ESULTS
A. Problem Formulation
Let L = {1, 2, · · · , L}, where L ≥ 2. Let t be the time index and
X1 (t), X2 (t), · · · , XL (t) : t = 1, 2, · · ·
be
a collection of L independent discrete memoryless information sources with an L-tuple of generic random variables
(X1 , X2 , · · · , XL ) taking values in X1 × X2 × · · · × XL , where Xi , i ∈ L are finite alphabets. In the sequel, we
use boldfaced letters to denote vectors of length n, for example X1 = (X1 (1), X1 (2), · · · , X1 (n)). There are L
4
encoders, indexed by L, each of which can access all the L information sources. There are also 2L − 1 decoders.
For each U ⊆ L such that U 6= ∅, Decoder-U can access the subset of encoders indexed by U. For 1 ≤ α ≤ L
and U such that |U| = α, Decoder-U can reconstruct the first α sources X1 , X2 , · · · , Xα perfectly in the usual
Shannon sense.
An (n, M1 , M2 , · · · , ML ) code is defined by the encoding functions
El :
L
Y
Xin → {1, 2, · · · , Ml }, for l ∈ L
(1)
i=1
and decoding functions
DU :
Y
{1, 2, · · · , Ml } →
|U |
Y
Xin , for U ⊆ L and U 6= ∅.
(2)
i=1
l∈U
Let Wl = El (X1 , X2 , · · · , XL ) be the output of Encoder-l and WU = (Wi : i ∈ U) for U ⊆ L. A nonnegative rate
tuple (R1 , R2 , · · · , RL ) is achievable if for any ǫ > 0, there exists for sufficiently large n an (n, M1 , M2 , · · · , ML )
code such that
1
log Ml ≤ Rl + ǫ, ∀ l ∈ L,
n
(3)
Pr{DU (WU ) 6= (X1 , X2 , · · · , Xα )} ≤ ǫ,
(4)
and
for all α = 1, 2, · · · , L and U ⊆ L such that |U| = α. The achievable rate region R is defined as the collection of
all achievable rate tuples.
B. Existing Results
We adopt the terminologies and notations in [4]. Let Rsup be the rate region induced by superposition coding.
Then Rsup is the set of nonnegative rate tuples (R1 , R2 , · · · , RL ) such that
Rl =
L
X
rlα , for l ∈ L
(5)
α=1
for some rlα ≥ 0, 1 ≤ α ≤ L, satisfying
X
rlα ≥ H(Xα ), for all U ⊆ L and |U| = α.
(6)
l∈U
Let λ = (λ1 , λ2 , · · · , λL ) and
RL
+ = {λ : λ 6= 0 and λi ∈ R, λi ≥ 0 for i ∈ L}.
(7)
L
Let Ωα
L = v ∈ {0, 1} : |v| = α , where |v| is the Hamming weight of a vector v = (v1 , v2 , · · · , vL ). Note that
there is a one-to-one correspondence between a vector v ∈ {0, 1}L and Decoder-U, where U = {i : vi = 1}. For
any λ ∈ RL
+ and α ∈ L, let fα (λ) be the optimal solution to the following optimization problem:
X
fα (λ) , max
cα (v)
(8)
v∈Ωα
L
s.t.
X
cα (v)v ≤ λ
(9)
v∈Ωα
L
cα (v) ≥ 0, ∀v ∈ Ωα
L.
(10)
5
Note that the functions fα (·) and cα (·) above depend on L, but for simplicity we omit this dependency in the
notations. Thus, if the length of λ is given, then fα (λ) can be defined accordingly. A set {cα (v) : v ∈ Ωα
L } is
called an α-resolution for λ if (9) and (10) are satisfied and it will be abbreviated as {cα (v)} if there is no
ambiguity. Furthermore, an α-resolution is called optimal if it achieves the optimal value fα (λ).
Let Rh be the collection of nonnegative rate tuples (R1 , R2 , · · · , RL ) such that
L
X
λl Rl ≥
L
X
fα (λ)H(Xα ), for all λ ∈ RL
+.
(11)
α=1
l=1
It was proved in [4] that the superposition region Rsup can be alternatively characterized by Rh . This means that in
addition to being the optimal value of the optimization problem in (8), for every fixed λ ∈ RL
+ , fα (λ) also gives
a tight linear outer bound on Rsup via (11). It was further proved in [4] that Rh is an outer bound on R. Then
Rsup ⊆ R ⊆ Rh
(12)
R = Rh = Rsup ,
(13)
which implies
i.e., superposition coding is optimal.
The following lemma is a direct consequence of Lemma 4 and 7 in [4]. It will be used in the proof of our main
result in the next section.
Lemma 1. Assume λ1 ≥ λ2 ≥ · · · ≥ λL . For α ≥ 2, if λ1 ≤
λ2 +λ3 +···+λL
,
α−1
then fα (λ) =
1
α
PL
i=1
λi .
III. O PTIMAL α- RESOLUTION
For any λ ∈ RL
+ and any permutation ω on {1, 2, · · · , L}, with an abuse of notation, we denote λω(1) ,
λω(2) , · · · , λω(L) by ω(λ). For any α ∈ L, due to the symmetry of the system, it is intuitive that the values of
fα (ω(λ)) are the same for all ω. This important property of fα (λ) is formally proved in the following lemma.
Lemma 2. fα ω(λ) = fα (λ) for α ∈ L.
Proof. For any α ∈ L, let {cα (v) : v ∈ Ωα
L } be an optimal α-resolution for λ. Then we have
X
cα (v)v
≤
λ,
(14)
cα (v)
≥
0, ∀v ∈ Ωα
L,
(15)
X
cα (v).
(16)
v∈Ωα
L
and
fα (λ) =
v∈Ωα
L
Let
that
P
v∈Ωα
L
cα (v)v = λ̃. Then by (14), we have λ̃ ≤ λ. For any permutation ω on {1, 2, · · · , L}, we can check
X
v∈Ωα
L
cα (v) ω(v) = ω
X
v∈Ωα
L
cα (v)v
= ω(λ̃) ≤ ω(λ).
(17)
6
For any v ∈ Ωα
L , let
It is immediate that for all v ∈ Ωα
L,
c′α ω(v) = cα (v).
(18)
c′α ω(v) ≥ 0.
(19)
α
α
α
Since ω is a one-to-one mapping from Ωα
L to ΩL , we have v ∈ ΩL if and only if ω(v) ∈ ΩL for any ω. Thus,
X
X
c′α ω(v) ω(v) =
c′α ω(v) ω(v)
v∈Ωα
L
ω(v)∈Ωα
L
X
=
cα (v) ω(v)
v∈Ωα
L
≤ ω λ ,
(20)
where the inequality follows from (17). By (19) and (20), we see that {c′α ω(v) : v ∈ Ωα
L } is an α- resolution
for ω(λ). In light of the definition of fα (λ) in (8)-(10), we have
X
X
X
c′α ω(v) =
fα ω(λ) ≥
c′α ω(v) =
cα (v) = fα (λ),
v∈Ωα
L
ω(v)∈Ωα
L
(21)
v∈Ωα
L
and so
fα ω(λ) ≥ fα (λ).
(22)
fα ω −1 ω(λ) ≥ fα ω(λ) .
(23)
Let ω −1 be the inverse permutation of ω. By the same argument, we can obtain
Since ω −1 (ω(λ)) = λ, we see from (22) and (23) that
The lemma is proved.
fα (λ) = fα ω(λ)
for all α ∈ L.
(24)
If a vector λ satisfies
λ1 ≥ λ2 ≥ · · · ≥ λL ,
(25)
we call λ an ordered vector. Throughout this section except for Lemma 8, we assume without loss of generality
that λ is an ordered vector. For any α ∈ L, it is easy to see that
fα (µλ) = µfα (λ)
(26)
for all µ ∈ R such that µ > 0. Thus, we will consider only λ’s whose minimum nonzero element is equal to 1.
Then there exists a ζ ∈ L such that
λ1 ≥ λ2 ≥ · · · ≥ λζ = 1
(27)
and λi = 0 for all i = ζ + 1, ζ + 2, · · · , L.
Fix λ, it is easy to see that
f1 (λ) =
L
X
i=1
λi ,
(28)
7
and
fζ (λ) = 1,
(29)
fα (λ) = 0.
(30)
and for α ≥ ζ + 1,
For other cases, determining the value of fα (λ) is highly nontrivial.
For α ∈ L and β = 0, 1, · · · , α − 1, let
gα,λ (β) =
L
X
1
λi .
α−β
(31)
i=β+1
Let βα∗ be a value of β (not necessarily unique) that achieves the minimum minβ∈{0,1,··· ,α−1} gα,λ (β), i.e.,
gα,λ (βα∗ ) =
min
β∈{0,1,··· ,α−1}
gα,λ (β).
(32)
The following theorem, a main result of the current paper, gives a closed-form solution for fα (λ).
Theorem 1. For any α ∈ L, fα (λ) = gα,λ (βα∗ ).
Proof. Fix an α ∈ L, and denote βα∗ by β ∗ for simplicity. We prove the theorem by proving i) fα (λ) ≤ gα,λ (β ∗ );
ii) there exists a solution for the optimization problem (8) that achieves gα,λ (β ∗ ), so that fα (λ) ≥ gα,λ (β ∗ ).
i) fα (λ) ≤ gα,λ (β ∗ ).
For 0 ≤ β ≤ α − 1, let eβ be an L-vector with the first β components being 0 and the last L − β components
Pβ
being 1. For any v ∈ Ωα
L , since
i=1 vi ≤ β, we have
v · eβ =
L
X
vi ≥ α − β.
(33)
i=β+1
Then for any solutions {cα (v)} to the optimization problem in (8), we have
L
X
= λ · eβ
λi
(34)
i=β+1
≥
=
X
v∈Ωα
L
cα (v)v · eβ
(35)
X
cα (v)(v · eβ )
(36)
X
cα (v)(α − β)
(37)
v∈Ωα
L
≥
v∈Ωα
L
= (α − β)
X
cα (v).
(38)
v∈Ωα
L
This implies that
fα (λ) ≤
L
X
1
λi = gα,λ (β), for all 0 ≤ β ≤ α − 1.
α−β
(39)
i=β+1
Thus, we have
fα (λ) ≤ gα,λ (β ∗ ).
(40)
8
ii) fα (λ) ≥ gα,λ (β ∗ ).
We now show that there exists a solution that achieves gα,λ (β ∗ ). For any α ∈ L and β ∗ ∈ {0, 1, · · · , α − 2},
by (32), we have
1
α − β∗
L
X
λi ≤
L
X
1
λi ,
α − (β ∗ + 1)
∗
(41)
i=β +2
i=β ∗ +1
which is equivalent to
λβ ∗ +1 ≤
L
X
1
λi .
(α − β ∗ ) − 1
∗
(42)
i=β +2
Denote the (L − β ∗ )-vector (λβ ∗ +1 , λβ ∗ +2 , · · · , λL ) by λ′ . In view of (42), by Lemma 1, (31), and (32), we
have
fα−β ∗ (λ′ ) =
1
α − β∗
L
X
λi = gα,λ (β ∗ ).
(43)
i=β ∗ +1
In view of (28) and (31) with β = β ∗ , it is easy to check that (43) is also satisfied for β ∗ = α − 1. Without
o
n
∗
be an optimal (α − β ∗ )-resolution for λ′ . Then it follows
loss of generality, let cα−β ∗ (u) : u ∈ Ωα−β
∗
L−β
from (43) that
X
cα−β ∗ (u) = fα−β ∗ (λ′ ) = gα,λ (β ∗ ).
(44)
α−β ∗
u∈ΩL−β
∗
For any v ∈ Ωα
L , let
cα (v) =
Then we have
cα−β ∗ (u),
0,
X
cα (v) =
v∈Ωα
L
∗
if v = (11 · · · 1u) for some u ∈ Ωα−β
L−β ∗
(45)
otherwise.
X
cα−β ∗ (u) = gα,λ (β ∗ ).
(46)
α−β ∗
u∈ΩL−β
∗
Again, by (32), we have
L
X
1
1
λi ≥
α − (β ∗ − 1)
α
−
β∗
∗
i=β
Then
λβ ∗ ≥
1
α − β∗
L
X
L
X
λi .
(47)
i=β ∗ +1
λi = gα,λ (β ∗ ),
(48)
i=β ∗ +1
where the equality above follows from (31). Thus,
λ1 ≥ λ2 ≥ · · · ≥ λβ ∗ ≥ gα,λ (β ∗ ).
(49)
For i = 1, 2, · · · , β ∗ , since cα (v) = 0 if vi = 0, we have
X
v∈Ωα
L :vi =1
cα (v) =
X
v∈Ωα
L
cα (v) = gα,λ (β ∗ ) ≤ λi ,
(50)
9
where the second equality follows from (46). For i = β ∗ + 1, β ∗ + 2, · · · , L,
X
cα (v)
X
=
v∈Ωα
L :vi =1
cα (v) +
v∈Ωα
L : vi =1,
(v1 ,··· ,vβ ∗ )=1
cα (v)
v∈Ωα
L : vi =1,
(v1 ,··· ,vβ ∗ )6=1
X
=
X
cα−β ∗ (u) + 0
∗
α−β
u∈ΩL−β
∗ : ui−β ∗ =1
≤ λi ,
since
(51)
o
n
∗
is an optimal (α − β ∗ )-resolution for λ′ . From (46), (50), and (51), we can
cα−β ∗ (u) : u ∈ Ωα−β
L−β ∗
∗
see that {cα (v) : v ∈ Ωα
L } defined by (45) is an α-resolution for λ that achieves gα,λ (β ). Thus, we have
fα (λ) ≥ gα,λ (β ∗ ).
(52)
The following lemma provides an important insight into the minimum in (32).
Lemma 3. For any α ∈ {2, 3, · · · , L} and 0 ≤ β ≤ α − 2,
(i) if gα,λ (β) ≥ gα,λ (β + 1), then gα,λ (0) ≥ gα,λ (1) ≥ · · · ≥ gα,λ (β + 1);
(ii) if gα,λ (β) ≤ gα,λ (β + 1), then gα,λ (β) ≤ gα,λ (β + 1) ≤ · · · ≤ gα,λ (α − 1).
Remark 1. In Lemma 3, if all the non-strict inequalities are replaced by strict inequalities, the lemma remains valid.
This alternative version of Lemma 3 can be proved by modifying the proof below accordingly.
Proof of Lemma 3. In the following, we only prove (ii). The proof for (i) can be obtained similarly.
For α = 2, the lemma is immediate. For 3 ≤ α ≤ L and β = α − 2, (ii) is immediate. For 0 ≤ β ≤ α − 3, from
the definition of gα,λ (·) in (31), the condition gα,λ (β) ≤ gα,λ (β + 1) is equivalent to
L
L
X
X
1
1
λi ≤
λi ,
α−β
α − (β + 1)
i=β+1
or
λβ+1 ≤
(53)
i=β+2
L
X
1
λi .
α − (β + 1)
(54)
i=β+2
Then by the assumption in (25), we have
λβ+2 ≤
L
X
1
λi ,
α − (β + 1)
(55)
i=β+2
or
λβ+2 ≤
L
X
1
λi ,
α − (β + 2)
(56)
i=β+3
which is also equivalent to
L
L
X
X
1
1
λi ≤
λi .
α − (β + 1)
α − (β + 2)
i=β+2
(57)
i=β+3
From (31), we have
gα,λ (β + 1) ≤ gα,λ (β + 2).
(58)
10
Then we see inductively that for all β + 1 ≤ β ′ ≤ α − 2,
gα,λ (β ′ ) ≤ gα,λ (β ′ + 1).
(59)
The lemma is proved.
For any α ∈ {2, 3, · · · , L} and any β ∈ {0, 1, · · · , α − 1}, we can readily see from Lemma 3 that βα∗ = β if
and only if
gα,λ (0) ≥ gα,λ (1) ≥ · · · ≥ gα,λ (β)
(60)
gα,λ (β) ≤ gα,λ (β + 1) ≤ · · · ≤ gα,λ (α − 1).
(61)
and
This provides a method to find the optimal value βα∗ quickly. We only need to compare gα,λ (β) and gα,λ (β + 1)
for β = 0, 1, · · · , α − 2 successively and stop at the first β such that gα,λ (β) ≤ gα,λ (β + 1). Then this β gives a
value of βα∗ that achieves the minimum in (32).
Lemma 4. 0 = β1∗ ≤ β2∗ ≤ · · · ≤ βL∗ .
Proof. It is easy to see from (28) that β1∗ = 0. This implies that β1∗ ≤ β2∗ . Now, we prove the lemma by showing
∗
that βα−1
≤ βα∗ for any 3 ≤ α ≤ L. If βα∗ ∈ {α − 2, α − 1}, since for a fixed α ∈ L we have 0 ≤ β ≤ α − 1, it
is obvious that
∗
≤ α − 2 ≤ βα∗ .
βα−1
(62)
Otherwise, βα∗ ∈ {0, 1, · · · , α − 3}. Since βα∗ achieves the minimum in (32), we have
1
α − βα∗
L
X
λi ≤
∗ +1
i=βα
L
X
1
λi ,
α − (βα∗ + 1)
∗
(63)
i=βα +2
which is equivalent to
λ
∗ +1
βα
L
X
1
≤
λi .
α − (βα∗ + 1)
∗
(64)
i=βα +2
This implies that
λβα∗ +1 ≤
L
X
1
λi ,
α − (βα∗ + 2)
∗
(65)
i=βα +2
which is equivalent to
L
L
X
X
1
1
λ
≤
λi .
i
α − (βα∗ + 1)
α − (βα∗ + 2)
∗
∗
i=βα +1
(66)
i=βα +2
Thus, we have
1
(α − 1) − βα∗
L
X
∗ +1
i=βα
λi ≤
L
X
1
λi ,
(α − 1) − (βα∗ + 1)
∗
(67)
i=βα +2
which by (31) implies that
gα−1,λ (βα∗ ) ≤ gα−1,λ (βα∗ + 1).
(68)
11
By the discussion following Lemma 3, we conclude that
∗
βα−1
≤ βα∗ .
(69)
Lemma 5. Let λ1 = (λ1,1 , λ1,2 , · · · , λ1,L ) and λ2 = (λ2,1 , λ2,2 , · · · , λ2,L ) be two vectors such that λ1,1 > λ2,1
and λ1,i = λ2,i for all 2 ≤ i ≤ L. For any α0 ∈ L, if fα0 (λ1 ) = fα0 (λ2 ), then fα (λ1 ) = fα (λ2 ) for all α ≥ α0 .
Proof. For α ∈ L, let βα1 and βα2 be the values that achieve fα (λ1 ) and fα (λ2 ), respectively. The condition
fα0 (λ1 ) = fα0 (λ2 ) implies that
βα1 0 = βα2 0 ≥ 1,
since otherwise
(70)
L
L
1 X
1 X
λ1,i =
λ2,i ,
α0 i=1
α0 i=1
(71)
which is a contradiction to the assumption that λ1,1 > λ2,1 and λ1,i = λ2,i for 2 ≤ i ≤ L. For all α ≥ α0 , by
Lemma 4, we have βα1 ≥ 1 and βα2 ≥ 1. Then by Theorem 1, we have
fα (λ1 ) = fα (λ2 ) =
min
β∈{1,2,··· ,α−1}
gα,λ1 (β).
(72)
This proves the lemma.
Let λ[1] be the L-vector with the first component being 1 and the rest being 0.
Lemma 6. If λ1 >
PL
i=2
λi , let λ′ =
PL
λi , λ2 , λ3 , · · · , λL . Then for all α ∈ L,
L
X
fα (λ) = λ1 −
λi fα λ[1] + fα (λ′ ).
i=2
(73)
i=2
Proof. By Theorem 1, we have
f2 (λ′ ) =
L
X
λi .
(74)
i=2
The condition λ1 >
PL
i=2
λi implies that
g2,λ (0) > g2,λ (1).
(75)
Again by Theorem 1, we have
f2 (λ) = g2,λ (1) =
L
X
λi .
(76)
i=2
Then
f2 (λ) = f2 (λ′ ),
(77)
fα (λ) = fα (λ′ ), for all 2 ≤ α ≤ L.
(78)
and by Lemma 5, we have
12
For 2 ≤ α ≤ L, since fα (λ[1] ) = 0, the equation (73) is satisfied by virtue of (78). For α = 1, we can check that
f1 (λ) =
L
X
λi
i=1
=
λ1 −
L
X
λi
i=2
=
λ1 −
L
X
i=2
so that (73) is also satisfied. This proves the lemma.
λi
!
·1+2
!
f1 (λ[1] ) + f1 (λ′ ),
L
X
λi
i=2
(79)
Lemma 7. For any η ∈ {1, 2, · · · , L − 1},
P
(i) if λ1 ≤ η1 L
i=2 λi , then fα (λ) = gα,λ (0), for α = 1, 2, · · · , η + 1;
P
L
(ii) if λ1 ≥ η1 i=2 λi , then fα (λ) = fα−1 (λ2 , λ3 , · · · , λL ), for α = η + 1, η + 2, · · · , L.
Remark 2. If λ1 =
1
η
PL
i=2
λi , we have from the lemma that
L
fη+1 (λ) = fη (λ2 , λ3 , · · · , λL ) =
In this case,
fα (λ) =
1 PL λi ,
i=1
α
1 X
λi ,
η + 1 i=1
for α ≤ η + 1
(80)
(81)
fα−1 (λ2 , λ3 , · · · , λL ), for α ≥ η + 1.
Proof. We first prove (i). For α ≤ η + 1, it is easy to check that
1
1
1
1+
≤
.
α
η
α−1
(82)
Thus,
L
1X
λi
α i=1
L
=
≤
=
1
1X
λi
λ1 +
α
α i=2
L
L
1X
1 1X
λi +
λi
α η i=2
α i=2
L
1
1 X
λi
1+
α
η i=2
(83)
L
1 X
λi ,
(84)
α − 1 i=2
P
where (83) follows from the assumption that λ1 ≤ η1 L
i=2 λi and (84) follows from (82). Then by the discussion
≤
following Lemma 3, we have
L
fα (λ) =
1X
λi = gα,λ (0).
α i=1
Next, we prove (ii). For α ≥ η + 1, it is easy to check that
1
1
1
1+
≥
.
α
η
α−1
(85)
(86)
13
Similar to the derivation of (84), with the assumption that λ1 ≥
L
1
η
L
PL
i=2
λi , (86) implies that
1 X
1X
λi ≥
λi .
α i=1
α − 1 i=2
Thus, we have
fα (λ) =
=
=
min
β∈{0,1,··· ,α−1} α
1
−β
L
X
(87)
λi
i=β+1
L
X
1
min
λi
β∈{1,2,··· ,α−1} α − β
i=β+1
L
X
1
min
λi
β∈{0,1,··· ,α−2} (α − 1) − β
(88)
i=β+2
=
fα−1 (λ2 , λ3 , · · · , λL ),
(89)
where (88) follows from (87). This proves the lemma.
The following lemma implies that fα (λ) is a concave function of λ ∈ RL
+ for all α ∈ L. Note that the vectors
in this lemma are not necessarily ordered.
Lemma 8. For any α ∈ L,
fα (µ1 λ1 + µ2 λ2 ) ≥ µ1 fα (λ1 ) + µ2 fα (λ2 )
(90)
for any λ1 , λ2 ∈ RL
+ and µ1 , µ2 ≥ 0.
Proof. Let λ1 = (λ1,1 , λ1,2 , · · · , λ1,L ) and λ2 = (λ2,1 , λ2,2 , · · · , λ2,L ). Let π1 (·), π2 (·) be two permutations of
{1, 2, · · · , L} such that
λ1,π1 (1) ≥ λ1,π1 (2) ≥ · · · ≥ λ1,π1 (L)
(91)
λ2,π2 (1) ≥ λ2,π2 (2) ≥ · · · ≥ λ2,π2 (L) .
(92)
and
Denote the ordered vectors by π1 (λ1 ) and π2 (λ2 ), respectively. For any β = 0, 1, · · · , α − 1, it is easy to see that
L
L
X
X
1
1
λ1,i ≥
λ1,π1 (i)
α−β
α−β
i=β+1
and
L
L
X
X
1
1
λ2,i ≥
λ2,π2 (i) .
α−β
α−β
i=β+1
Thus, we have
(94)
i=β+1
L
L
X
X
1
1
(µ1 λ1,i + µ2 λ2,i ) ≥
(µ1 λ1,π1 (i) + µ2 λ2,π2 (i) ).
α−β
α−β
i=β+1
(93)
i=β+1
(95)
i=β+1
This implies that
fα (µ1 λ1 + µ2 λ2 ) ≥ fα (µ1 π1 (λ1 ) + µ2 π2 (λ2 )).
(96)
14
For any α ∈ L, it is easy to check that
fα (π1 (λ1 )) = fα (λ1 )
(97)
fα (π2 (λ2 )) = fα (λ2 ).
(98)
and
Therefore, if the lemma holds for any ordered vectors λ1 and λ2 , then the lemma holds for any vectors λ1 and
λ2 (not necessarily ordered), because
fα (µ1 λ1 + µ2 λ2 ) ≥
fα (µ1 π1 (λ1 ) + µ2 π2 (λ2 ))
≥
µ1 fα (π1 (λ1 )) + µ2 fα (π2 (λ2 ))
=
µ1 fα (λ1 ) + µ2 fα (λ2 ).
(99)
Thus without loss of generality, we assume that λ1 and λ2 are ordered. Then for any β = 0, 1, · · · , α − 1, we
have from Theorem 1 that
L
X
1
λ1,i ≥ fα (λ1 )
α−β
(100)
i=β+1
and
L
X
1
λ2,i ≥ fα (λ2 ),
α−β
(101)
i=β+1
which implies
L
X
1
(µ1 λ1,i + µ2 λ2,i ) ≥ µ1 fα (λ1 ) + µ2 fα (λ2 ).
α−β
(102)
i=β+1
By taking the minimum over all β = 0, 1, · · · , α − 1, we obtain
L
1
X
min
(µ1 λ1,i + µ2 λ2,i ) ≥ µ1 fα (λ1 ) + µ2 fα (λ2 ),
β∈{0,1,··· ,α−1} α − β
(103)
i=β+1
which by Theorem 1 is equivalent to
fα (µ1 λ1 + µ2 λ2 ) ≥ µ1 fα (λ1 ) + µ2 fα (λ2 ).
(104)
This proves the lemma.
IV. T HE M INIMUM S UFFICIENT S ET OF I NEQUALITIES
Even though the superposition region Rsup can be explicitly characterized by Theorem 1, an uncountable number
of inequalities are involved. For a fixed L, among all these inequalities, only a finite number of them are needed
because Rsup is a polytope. In this section, we provide a method to determine this minimum sufficient set of
inequalities.
For any λ ∈ RL
+ , let π(·) be a permutation of {1, 2, · · · , L} such that
λπ(1) ≥ λπ(2) ≥ · · · ≥ λπ(L) .
(105)
15
Recall that we consider only λ’s whose minimum nonzero element is equal to 1. Let ζ ∈ L be such that
λπ(ζ) = 1
(106)
λπ(j) = 0.
(107)
and for j = ζ + 1, ζ + 2, · · · , L,
Toward listing all the inequalities defining Rsup , we consider a certain finite subset of RL
+ defined as follows. Let
GL be the collection of all λ ∈ RL
+ such that for j = ζ − 1, ζ − 2, · · · , 1,
L
L
L
X
X
X
1
1
λπ(i) , · · · ,
λπ(i) ,
λπ(i) ,
λπ(j) ∈
2
θj+1 + 1
i=j+1
i=j+1
(108)
i=j+1
where θζ = 0 and for j = ζ − 2, · · · , 1, θj+1 is the integer such that
λπ(j+1) =
1
θj+1
L
X
λπ(i) .
(109)
i=j+2
Here, (106), (108), and (109) not only defines GL but in fact provides a method to exhaust all λ ∈ GL . For ζ = 1,
the only possible λ are λ[1] = (1, 0, 0, · · · ) and its permutations. For ζ ≥ 2, starting with λπ(ζ) = 1, the values of
λπ(ζ−1) , λπ(ζ−2) , · · · , λπ(1) can be chosen recursively according to (108). It is easy to check that
θj ∈ {1, 2, · · · , θj+1 + 1}
(110)
1 ≤ θj ≤ ζ − j
(111)
and
for 1 ≤ j ≤ ζ − 1. Furthermore, for the last element of the set in (108) which is the smallest in the set, we have
L
L
L
X
X
X
1
1
1
λπ(i)
λπ(i) +
λπ(i) =
θj+1 + 1 i=j+1
θj+1 + 1 θj+1 i=j+2
i=j+2
=
=
L
θj+1 + 1 X
λπ(i)
θj+1 + 1 θj+1 i=j+2
1
1
θj+1
L
X
λπ(i)
i=j+2
= λπ(j+1) ,
(112)
so that λπ(j) ≥ λπ(j+1) as required by (105). Also, we see from (108) that
1) for ζ ≥ 2, λπ(ζ−1) = λπ(ζ) = 1;
2) for ζ ≥ 3, λπ(j+1) is always a possible choice for λπ(j) for 1 ≤ j ≤ ζ − 2.
Denote the cardinality of GL by SL . Let GL0 = {λ ∈ GL : λ is ordered }, and denote its cardinality by |GL0 | = SL0 .
For the ease of notation, we let
with λ(1) = λ[1] and
o
n
0
GL0 = λ(1) , λ(2) , · · · , λ(SL )
(113)
o
n
0
0
GL = GL0 ∪ λ(SL +1) , λ(SL +2) , · · · , λ(SL ) .
(114)
16
In other words, the set GL is the collection of all possible permutations of the vectors in GL0 .
For i = 1, 2, · · · , SL , let πi (·) be a permutation of {1, 2, · · · , L} such that
(i)
(i)
(i)
λπi (1) ≥ λπi (2) ≥ · · · ≥ λπi (L) .
(115)
For any λ ∈ RL
+ , let f (λ) = f1 (λ), f2 (λ), · · · , fL (λ) . The following lemma is important for the proof of our
main theorem.
[1]
Lemma 9. Consider any ordered vector λ ∈ RL
+ such that λ 6= λ . Assume there exists ci ≥ 0, i = 1, 2, · · · , SL
such that
SL
X
ci λ(i) , f (λ(i) ) .
λ, f (λ) =
(116)
i=1
Let I = i ∈ {1, 2, · · · , SL } : ci 6= 0 . For any η ∈ {1, 2, · · · , ζ − 1},
PL
PL
(i)
(i)
(i) if λ1 ≤ η1 j=2 λj , then λπi (1) ≤ η1 j=2 λπi (j) for all i ∈ I;
PL
PL
(i)
(i)
(ii) if λ1 ≥ η1 j=2 λj , then λπi (1) ≥ η1 j=2 λπi (j) for all i ∈ I.
Remark 3. In the above, since λ is ordered, we have
L
λ1 ≥
1 X
λj .
ζ − 1 j=2
(117)
Therefore, when η = ζ − 1, the condition in (i) can only be satisfied with an equality, i.e., λ1 =
Proof. See Appendix A.
1
ζ−1
PL
j=2
λj .
The assumption that λ(i) ∈ GL for 1 ≤ i ≤ SL is not invoked in the proof of Lemma 9. By taking this assumption
into account, Lemma 9 can be further strengthened with the following setup. For any ordered vector λ ∈ RL
+ not
equal to λ[1] , by the constraint in (117), there exists a unique η ∈ {1, 2, · · · , ζ − 1} such that
L
L
1 X
1X
λj ≤ λ1 <
λj .
η j=2
η − 1 j=2
(118)
In the sequel, we adopt the convention that
∞, if c 6= 0
1
·c=
0
1,
(119)
if c = 0.
Then the upper bound in (118) is ∞ when η = 1.
[1]
Lemma 10. For any ordered vector λ ∈ RL
+ such that λ 6= λ , assume there exists ci ≥ 0, i = 1, 2, · · · , SL such
that
SL
X
ci λ(i) , f (λ(i) ) .
λ, f (λ) =
(120)
i=1
Then for all i ∈ I,
(i)
λπi (1) ∈
L
1 X
η
j=2
(i)
λπi (j) ,
1
η−1
L
X
j=2
(i)
λπi (j)
,
(121)
17
(i)
where η depends on λ and is defined in (118). In particular, if the lower bound in (118) is tight, then λπi (1) =
PL
(i)
1
j=2 λπi (j) for all i ∈ I.
η
Proof. The lemma can be easily obtained from Lemma 9. See details in Appendix B.
(i)
Remark 4. For all 1 ≤ i ≤ SL , λπi (1) can in general take one of the θ2 + 1 values prescribed in (108). However,
(i)
under the constraint (120), the above lemma asserts that for all i ∈ I, λπi (1) can only take one of the two values
prescribed in (121).
Let R∗ be the collection of nonnegative rate tuples (R1 , R2 , · · · , RL ) such that
L
X
l=1
λl Rl ≥
L
X
fα (λ)H(Xα ), for all λ ∈ GL .
(122)
α=1
The next theorem shows that R∗ provides an equivalent characterization of Rsup . Note that R∗ is the intersection
of only a finite set of halfspaces. Thus, R∗ is more explicit than Rh . For L = 1, 2, · · · , 5, all the rate constraints
of R∗ with ordered coefficient vectors are listed in Appendix I.
Theorem 2. Rsup = R∗ .
∗
∗
Proof. We prove the theorem by showing that Rh = R∗ . Since GL ⊆ RL
+ , we have Rh ⊆ R . To show R ⊆ Rh ,
we consider the following. Define three sets of (2L)-vectors by
FL1 = (λ, f (λ)) : λ ∈ RL
+ ,
(123)
FL2 = {(λ, f (λ)) : λ ∈ GL } ,
(124)
FL3 = (λ, f (λ)) : λ ∈ GL0 .
(125)
and
Note that none of FL2 and FL3 is a vector space since they are not closed under vector addition. We prove R∗ ⊆ Rh
by proving the claim that any (λ, f (λ)) ∈ FL1 is a conic combination of the vectors in FL2 .
Without loss of generality, we consider only λ such that λ1 ≥ λ2 ≥ · · · ≥ λL , and show that (λ, f (λ)) for any
such λ is a conic combination of the vectors in FL3 . We prove the claim by induction on L for L ≥ 1. Since we
consider only λ’s whose minimum nonzero element is equal to 1, it is easy to see that F11 = F13 = {(1, 1)} and
thus the claim is true for L = 1.
Assume the claim is true for L = N . We will show that the claim is true for L = N + 1. This can readily be
+1
+1
verified for λ ∈ RN
such that ζ = 1. Thus, we consider only λ ∈ RN
such that ζ ≥ 2. For any ordered vector
+
+
λN = (λ2 , λ3 , · · · , λN +1 ) ∈ RN
+ , let λN +1 = (λ1 , λ2 , · · · , λN +1 ) where λ1 ≥ λ2 . By the induction hypothesis,
0
there exist ci ≥ 0, i = 1, 2, · · · , SN
such that
0
(λN , f (λN )) =
SN
X
i=1
(i)
(i)
ci λN , f (λN ) ,
(126)
18
(i)
(i)
(i)
(i)
(i)
0
0
are distinct elements of GN
. Let λN = (λ2 , λ3 , · · · , λN +1 ). Recall that I = {i ∈
where λN , i = 1, 2, · · · , SN
0
0
{1, 2, · · · , SN } : ci 6= 0} in Lemma 9. For simplicity, let ci = 0 for all i ∈ {SN
+ 1, SN
+ 2, · · · , SN }. For any
i ∈ I, by Lemma 10, we have
(i)
λ2 ∈
+1
1 N
X
η′
j=3
N
+1
X
1
(i)
(i)
,
λj
λj , ′
η − 1 j=3
(127)
where η ′ ∈ {1, 2, · · · , N − 1} is unique and determined by
N +1
N +1
1 X
1 X
λj .
λ
≤
λ
<
j
2
η ′ j=3
η ′ − 1 j=3
1
η′
Since the second inequality in (128) is equivalent to λ2 <
for λ1 :
PN +1
j=2
(128)
λj , we consider the following three cases
PN +1
λj < λ1 ≤ j=2 λj ;
P +1
b) λ2 ≤ λ1 ≤ η1′ N
j=2 λj ;
PN +1
c) λ1 > j=2 λj .
PN +1
PN +1
Case a): If η1′ j=2 λj < λ1 ≤ j=2 λj , there exists a unique ϕ ∈ {1, 2, · · · , η ′ − 1} such that
a)
1
η′
PN +1
j=2
N +1
N +1
1 X
1 X
λj < λ1 ≤
λj .
ϕ + 1 j=2
ϕ j=2
(129)
Then by Lemma 7, we have
fα (λN +1 ) =
1 PN +1 λj , for 1 ≤ α ≤ ϕ + 1
j=1
α
For all i ∈ I, let
(1,i)
λ1
=
N +1
1 X (i)
λ
ϕ + 1 j=2 j
and
(2,i)
(131)
N +1
1 X (i)
λ .
ϕ j=2 j
=
λ1
(130)
for ϕ + 2 ≤ α ≤ N + 1.
fα−1 (λN ),
(132)
For j ∈ {2, 3, · · · , N + 1}, for notational simplicity, let
(1,i)
λj
(2,i)
= λj
(i)
= λj .
(133)
Let
(1,i)
(1,i)
(2,i)
(2,i)
λN +1 = (λ1
(1,i)
, λ2
(1,i)
, · · · , λN +1 )
(134)
and
λN +1 = (λ1
(2,i)
, λ2
(2,i)
, · · · , λN +1 ).
(1,i)
(135)
(2,i)
0
From (127), (108), and the range of ϕ, we can check that λN +1 , λN +1 ∈ GN
+1 . By Remark 2 following Lemma 7,
we have
(1,i)
fα (λN +1 ) =
1 PN +1 λ(1,i) , for 1 ≤ α ≤ ϕ + 2
j
j=1
α
fα−1 (λ(i) ),
N
for ϕ + 2 ≤ α ≤ N + 1,
(136)
19
and
(2,i)
fα (λN +1 ) =
1 PN +1 λ(2,i) , for 1 ≤ α ≤ ϕ + 1
j
j=1
α
fα−1 (λ(i) ),
for ϕ + 1 ≤ α ≤ N + 1.
N
(1,i)
Consider the conic combination of λ1
X
(137)
for i ∈ I,
(1,i)
N +1
1 X X (i)
λj
ci
ϕ+1
j=2
=
ci λ1
(138)
i∈I
i∈I
N +1
1 X X
(i)
ci λj
ϕ + 1 j=2
=
i∈I
1
ϕ+1
=
<
λ1 ,
N
+1
X
λj
(139)
j=2
(140)
where (138) follows from (131), (139) follows from (126), and (140) follows from (129). Similarly, from (132),
(126), and (129), we have
X
(2,i)
ci λ1
≥ λ1 .
(141)
i∈I
Let u1 =
P
(1,i)
i∈I ci λ1
and u2 =
P
(2,i)
.
i∈I ci λ1
(i)
λ1 =
Then we have u1 < λ1 ≤ u2 . For all i ∈ I, let
u2 − λ1 (1,i) λ1 − u1 (2,i)
λ
+
λ
.
u2 − u1 1
u2 − u1 1
(142)
It is easy to check that
X
(i)
ci λ1 = λ1 .
(143)
i∈I
(1)
Let ci
that
=
u2 −λ1
u2 −u1
(2)
· ci and ci
=
λ1 −u1
u2 −u1
(1)
· ci . It is readily seen that ci
(2)
and ci
c(1) + c(2) = ci
i
i
c(1) λ(1,i) + c(2) λ(2,i) = ci λ(i) .
1
1
1
i
i
are nonnegative, and we can check
(144)
Then we have from (130), (143), and (144) that
λN +1 =
X
i∈I
(1) (1,i)
(2) (2,i)
ci λN +1 + ci λN +1 .
(145)
20
Following (130), we have for 1 ≤ α ≤ ϕ + 1 that
fα (λN +1 )
N +1
1 X
λj
α j=1
=
(146)
N +1
1 XX
(i)
ci λj
α j=1
i∈I
"
#
N
+1
1 X X (1) (1,i)
(2) (2,i)
=
ci λj + ci λj
α j=1
i∈I
N
+1
+1
X (1) 1 N
X
X
1
(2)
(1,i)
(2,i)
ci
=
+ ci
λ
λ
α j=1 j
α j=1 j
i∈I
i
X h (1)
(2,i)
(1,i)
(2)
=
ci fα (λN +1 ) + ci fα (λN +1 )
=
(147)
(148)
(149)
i∈I
where (147) follows from (126), (148) follows from (133) and (144), and (149) follows from (136) and (137).
Similarly, for ϕ + 2 ≤ α ≤ N + 1, following (130), we have
fα (λN +1 )
= fα−1 (λN )
X
(i)
=
ci fα−1 (λN )
(150)
(151)
i∈I
=
X
(1)
ci
(2)
+ ci
i∈I
=
(i)
fα−1 (λN )
(152)
i
X h (1)
(2,i)
(2)
(1,i)
ci fα (λN +1 ) + ci fα (λN +1 ) ,
(153)
i∈I
where (151) follows from (126), (152) follows from (144), and (153) follows from (136) and (137). In other words,
(149) or (153) holds for all 1 ≤ α ≤ N + 1. Summarizing the above, we have
X (1) (1,i)
(1,i)
(2)
(2,i)
(2,i)
λN +1 , f (λN +1 ) =
ci
λN +1 , fα (λN +1 ) + ci
λN +1 , fα (λN +1 ) ,
(154)
i∈I
3
and thus λN +1 , f (λN +1 ) is a conic combination of vectors in FN
+1 .
P
P
N +1
N +1
1
1
Case b): If λ2 ≤ λ1 ≤ η′ j=2 λj , since the condition η′ j=3 λj ≤ λ2 in (128) is equivalent to
λ2 ≥
we have
By Lemma 7, we obtain
I1 ∪ I2 = I. For i ∈ I1 , let
(155)
N +1
N +1
1 X
1 X
λ
≤
λ
≤
λj .
j
1
η ′ + 1 j=2
η ′ j=2
fα (λN +1 ) =
In light of (127), let I1 =
N +1
1 X
λj ,
η ′ + 1 j=2
n
(i)
i ∈ I : λ2
(156)
1 PN +1 λj , for 1 ≤ α ≤ η ′ + 1
j=1
α
for η ′ + 2 ≤ α ≤ N + 1.
n
PN +1 (i) o
(i)
= η′1−1 j=3 λj
and I2 = i ∈ I : λ2 =
fα−1 (λN ),
(1,i)
λ1
(2,i)
= λ1
=
N +1
1 X (i)
λ .
η ′ j=2 j
(157)
1
η′
PN +1
j=3
(i)
λj
o
, where
(158)
21
For i ∈ I2 , let
(1,i)
λ1
=
N +1
1 X (i)
λ
η ′ + 1 j=2 j
(159)
N +1
1 X (i)
λ .
η ′ j=2 j
(160)
and
(2,i)
=
λ1
(1,i)
(2,i)
0
Again, from (127) and (108), we can check that λN +1 , λN +1 ∈ GN
+1 for all i ∈ I. By Remark 2 following
Lemma 7, we have for i ∈ I1 that
(1,i)
(2,i)
fα (λN +1 ) = fα (λN +1 ) =
1 PN +1 λ(1,i) ,
j=1
α
j
fα−1 (λ(i) ),
N
and for i ∈ I2 ,
(1,i)
fα (λN +1 ) =
P
1 N +1 λ(2,i) ,
α
j
j=1
fα−1 (λ(i) ),
N
and
(2,i)
fα (λN +1 ) =
1 PN +1 λ(2,i) ,
α
j
j=1
fα−1 (λ(i) ),
N
for 1 ≤ α ≤ η ′ + 1
(161)
for η ′ + 1 ≤ α ≤ N + 1,
for 1 ≤ α ≤ η ′ + 2
(162)
for η ′ + 2 ≤ α ≤ N + 1,
for 1 ≤ α ≤ η ′ + 1
(163)
for η ′ + 1 ≤ α ≤ N + 1.
Following from (158) and (159), we have
X
(1,i)
ci λ1
=
i∈I
X
ci
i∈I1
=
N +1
N +1
1 X (i) X
1 X (i)
λ
+
λ
c
i
j
η ′ j=2
η ′ + 1 j=2 j
i∈I2
N
N +1
+1
X
X 1
X
1 X (i)
1
1
(i)
λ +
λj
1+ ′
ci ′
ci ′ 1 + ′
η
η − 1 j=3 j
η +1
η
j=3
=
=
N +1
N +1
1 X (i) X 1 X (i)
+
λ
λ
c
i
η ′ − 1 j=3 j
η ′ j=3 j
i∈I2
i∈I1
X
X
(i)
(i)
ci λ2 +
ci λ2
X
(164)
i∈I2
i∈I1
ci
i∈I1
(165)
(166)
i∈I2
=
λ2
(167)
≤
λ1 ,
(168)
where (164) and (166) follow from the definition of I1 and I2 . Similar to (146)-(149), we have
X
(2,i)
ci λ1
≥ λ1 .
(169)
i∈I
(1)
For i ∈ I, similar to (142)-(144), let ci
(i)
=
λ1 =
u2 −λ1
u2 −u1
(2)
· ci , ci
=
λ1 −u1
u2 −u1
· ci , and
u2 − λ1 (1,i) λ1 − u1 (2,i)
λ
+
λ
.
u2 − u1 1
u2 − u1 1
(170)
We can check that
X
i∈I
(i)
ci λ1 = λ1
(171)
22
and for all i ∈ I,
c(1) + c(2) = ci
i
i
Then similar to (145)-(154), we have
c(1) λ(1,i) + c(2) λ(2,i) = ci λ(i) .
1
1
1
i
i
X (1) (1,i)
(2,i)
(2,i)
(2)
(1,i)
λN +1 , fα (λN +1 ) .
λN +1 , fα (λN +1 ) + ci
λN +1 , f (λN +1 ) =
ci
(172)
(173)
i∈I
Case c): If λ1 >
PN +1
j=2
λj , let λ′N +1 =
P
N +1
j=2
(1)
λj , λ2 , · · · , λN +1 and λN +1 be the (N + 1)-vector with the
first component being 1 and the rest being 0. From Lemma 6, we have
N
+1
X
(1)
(1)
λN +1 , f (λN +1 ) + λ′N +1 , f (λ′N +1 ) .
λj
(λN +1 , f (λN +1 )) = λ1 −
(174)
j=2
It is easy to see that
(1)
(1)
3
λN +1 , f (λN +1 ) ∈ FN
+1 .
(175)
Note that λ′N +1 satisfies the condition for Case a) provided that η ′ 6= 1. Otherwise, it satisfies the condition for
3
Case b). Thus we see that λ′N +1 , f (λ′N +1 ) is always a conic combination of the vectors in FN
+1 . This implies
3
that (λN +1 , f (λN +1 )) is a conic combination of the vectors in FN
+1 , as is to be proved.
For any λ ∈ RL
+ , let λL−1 = (λ2 , λ3 , · · · , λL ) and f (λL−1 ) = f1 (λL−1 ), f2 (λL−1 ), · · · , fL−1 (λL−1 ) . The
following lemma provides a method for finding a set of conic combination coefficients for λL−1 , f (λL−1 ) from
the conic combination for λ, f (λ) .
[1]
Lemma 11. Consider any ordered vector λ ∈ RL
+ such that λ 6= λ . Assume there exists ci ≥ 0, i = 1, 2, · · · , SL
such that
SL
X
ci λ(i) , f (λ(i) ) ,
λ, f (λ) =
(176)
i=1
Then we have
SL
X
(i)
(i)
(i)
(i)
(i)
(i)
ci (λ2 , λ3 , · · · , λL ), f (λ2 , λ3 , · · · , λL ) .
λL−1 , f (λL−1 ) =
(177)
i=1
Proof. See Appendix C.
(i)
(i)
(i)
(i)
(i)
Lemma 12. For any λ(i) ∈ GL , if λ2 , λ3 , · · · , λL ∈ GL−1 , then λ1 = 0 or λπi (1) .
Proof. See Appendix D.
Lemma 13. For any i0 ∈ {1, 2, · · · , SL }, there does not exist (c1 , c2 , · · · , cSL ) ∈ RS+L such that ci0 = 0 and
SL
X
ci λ(i) , f (λ(i) ) .
λ(i0 ) , f (λ(i0 ) ) =
(178)
i=1
Proof. See Appendix E.
Theorem 2 gives a rate region R∗ that simplifies the characterization of the superposition region. The following
theorem shows that there is no redundancy in the specification of R∗ .
23
Theorem 3. For the inequalities specifying R∗ in (122), none of them is implied by the others.
Proof. It suffices to show the following: for any i0 ∈ {1, 2, · · · , SL }, if
λ(i0 ) =
SL
X
ci λ(i)
(179)
i=1
for some (c1 , c2 , · · · , cSL ) ∈ RS+L such that ci0 = 0, then the inequality
L
X
(i0 )
λl
Rl ≥
SL
X
L
X
ci
i=1
fα (λ(i0 ) )H(Xα )
(180)
α=1
l=1
is tighter than the inequality
L
X
(i)
λl Rl
l=1
!
≥
SL
X
ci
i=1
L
X
!
fα (λ(i) )H(Xα ) .
α=1
In view of (179), the LHS of (181) is equal to the LHS of (180), and (181) can be rewritten as
!
SL
L
L
X
X
X
(i0 )
(i)
ci fα (λ ) H(Xα ).
λl Rl ≥
α=1
l=1
(182)
i=1
Thus we need to show that the RHS of (180) is always greater than the RHS of (182), i.e.,
!
SL
L
L
X
X
X
(i)
(i0 )
ci fα (λ ) H(Xα )
fα (λ )H(Xα ) >
α=1
(181)
α=1
(183)
i=1
for all possible values of H(Xα ), α ∈ L. By Lemma 8, (179) implies that
SL
X
fα (λ(i0 ) ) ≥
ci fα (λ(i) ), for all α ∈ L.
(184)
i=1
Upon multiplying by H(Xα ) and summing over all α ∈ L, we obtain
L
X
fα (λ(i0 ) )H(Xα ) ≥
α=1
L
X
α=1
SL
X
!
ci fα (λ(i) ) H(Xα ),
i=1
(185)
which is equivalent to (183) except that the inequality above is nonstrict. Thus to prove (183), we only need to
show that there exists at least one α ∈ L such that
fα (λ(i0 ) ) >
SL
X
ci fα (λ(i) ).
(186)
i=1
Assume the contrary is true, i.e. equality holds in (184) for all α ∈ L. Then this implies
λ
(i0 )
, f (λ
(i0 )
SL
X
ci λ(i) , f (λ(i) ) ,
) =
(187)
i=1
which is a contradiction to Lemma 13. This completes the proof of the theorem.
V. C HECKING
THE
ACHIEVABILITY
OF
A R ATE T UPLE
The implicit characterization of Rsup obtained in [4] does not provide a method to check the achievability of
a given rate tuple because it involves an uncountable number of inequalities. In the last section, we provided an
explicit characterization of Rsup that contains only a finite number of inequalities. This makes it possible to check
the achievability of any given rate tuple.
24
In this section, we determine the number of inequalities that we need to check. Even though there is no redundancy
in the set of inequalities in (122) that specifies R∗ , by taking advantage of the symmetry of the problem, we in fact
do not need to check all these inequalities. The next lemma gives the set of the inequalities we need to check. As
we will see, the number of such inequalities is significantly smaller than the total number of inequalities specifying
R∗ .
For any R = (R1 , R2 , · · · , RL ) ∈ RL
+ and any permutation ω on {1, 2, · · · , L}, recall from the beginning of
Section III that
ω(R) = Rω(1) , Rω(2) , · · · , Rω(L) .
(188)
Due to the symmetry of the problem, for any λ ∈ GL , the inequality
L
X
λl Rl ≥
L
X
fα (λ)H(Xα )
(189)
α=1
l=1
implies the inequality
L
X
λω(l) Rω(l) ≥
L
X
α=1
l=1
fα ω(λ) H(Xα ),
(190)
and vice versa. Thus, R is achievable if and only if ω(R) is achievable for all ω. As such, we only need to consider
rate tuples R ∈ RL
+ such that
R1 ≤ R2 ≤ · · · ≤ RL .
(191)
Lemma 14. For any nonnegative rate tuple R such that (191) is satisfied and any λ ∈ RL
+ , we have
L
X
λi Ri ≥
i=1
L
X
λπ(i) Ri .
(192)
i=1
Proof. See Appendix H.
From Lemma 2, we can see that RHS of the inequality in (122) does not change with λ replaced by π(λ). Thus,
in order to check the achievability of a given rate tuple (assume satisfying (191)), by Lemma 14, we only need to
check those inequalities for which the coefficients are in descending order, i.e.
λ1 ≥ λ2 ≥ · · · ≥ λL .
(193)
All the other inequalities are redundant for this rate tuple. Then, the number of inequalities we need to check is
only SL0 , which is bounded in the following theorem.
Theorem 4. 2L−1 ≤ SL0 ≤ L!.
Remark 5. We ran a program on a notebook computer to list all the SL0 inequalities for all L ≤ 15. For L = 16,
the computation involved appears to be prohibitive.
Proof. We can see from Appendix I that S10 = 1 and S20 = 2. It is easy to check that the theorem is true for L = 1
and L = 2.
25
For L ≥ 2, let λ = (λ1 , λ2 , · · · , λL ) and GL∗ =
λ ∈ GL0 : λL = 0 . For any λ ∈ GL∗ , it is easy to check
0
0
, we have
from (108) that (λ1 , λ2 , · · · , λL−1 ) ∈ GL−1
. On the other hand, for any (λ1 , λ2 , · · · , λL−1 ) ∈ GL−1
0
(λ1 , λ2 , · · · , λL−1 , 0) ∈ GL∗ . Thus, there is a one-to-one correspondence between GL∗ and GL−1
, which implies that
0
|GL∗ | = SL−1
.
(194)
For k ≥ 2, let Dk = |Gk0 \Gk∗ |. By (194) and the fact that Gk∗ ⊆ Gk0 , we have
0
,
Dk = Sk0 − Sk−1
(195)
which implies that
SL0 = S10 +
L
X
Dk .
(196)
k=2
Now, we only need to calculate Dk for k ≥ 2. For any (λL−k+1 , λL−k+2 , · · · , λL ) ∈ Gk0 \Gk∗ , where λL = 1, we can
∗
0
\Gk−1
by construction. Thus, all (λL−k+1 , λL−k+2 , · · · , λL ) ∈
see from (108) that (λL−k+2 , λL−k+3 , · · · , λL ) ∈ Gk−1
0
∗
Gk0 \Gk∗ can be generated from (λL−k+2 , λL−k+3 , · · · , λL ) ∈ Gk−1
\Gk−1
with a proper choice of λL−k+1 . Since
λL = 1, we have ζ = L. Recall from (109) that θL = 0 and for j = L − 1, L − 2, · · · , L − k + 1, θj is the integer
such that
L
1 X
λi .
λj =
θj i=j+1
(197)
According to (197), the k-vector (λL−k+1 , λL−k+2 , · · · , λL ) ∈ Gk0 \Gk∗ is uniquely determined by the tuple (θL−k+1 ,
· · · , θL−1 , θL ). Thus Dk is equal to the cardinality of the set
Θk = (θL−k+1 , · · · , θL−1 , θL ) : 1 ≤ θj ≤ θj+1 + 1 for j = L − 1, L − 2, · · · , L − k + 1 .
(198)
By straightforward counting, we can obtain
Dk = |Θk | =
θX
L +1
0
X
θL−k+2 +1
θL−1 +1
X
···
X
1.
(199)
θL−k+1 =1
θL =0 θL−1 =1 θL−2 =1
Now we bound Dk according to (199). Observe that θL = 0 and θL−1 = 1 always hold. Then for k ≥ 3, (199)
can be rewritten as
Dk =
θL−2 +1
2
X
X
θL−k+2 +1
θL−2 =1 θL−3 =1
Let
(1)
Dk =
2
X
2
X
2
X
3
X
X
···
(2)
Dk =
2
X
···
θL−2 =1 θL−3 =1
and
1.
(200)
1
(201)
1.
(202)
θL−k+1 =1
θL−k+1 =1
···
θL−2 =1 θL−3 =1
k−1
X
θL−k+1 =1
From (111), it is easy to check that
(1)
(2)
Dk ≤ Dk ≤ Dk ,
(203)
and we have
(1)
Dk = 2k−2
(204)
26
and
(2)
Dk = (k − 1)! .
Thus, for L ≥ 3, we have
L
X
(1)
Dk =
k=3
and
L
X
L
X
(205)
2k−2 = 2L−1 − 2
(206)
k=3
(2)
=
Dk
k=3
L
X
(k − 1)!
k=3
≤ (L − 1)! × (L − 2)
≤ L! − 2.
(207)
Then, by (196), (203), and the fact that S10 = 1 and D2 = 1, we have for L ≥ 3 that
2L−1 ≤ SL0 ≤ L! .
(208)
This proves the theorem.
VI. S UBSET E NTROPY I NEQUALITY
In [4], the proof of the optimality of superposition coding was established through a subset entropy inequality,
namely Theorem 3 therein. As we will point out, this subset entropy inequality is in fact a generalization of Han’s
inequality [10]. The proof of Theorem 3 in [4], however, is painstaking. In this section, we present a weaker version
of this theorem, namely Theorem 5 below, whose proof is considerably simpler. With our explicit characterization
of Rsup in Theorem 2, Theorem 5 is sufficient for proving the optimality of superposition coding.
Theorem 5 (Subset entropy inequality). Let L ≥ 2 and for any u ∈ {0, 1}L , let Hu = H Wi : ui = 1 . For any
λ ∈ GL , there exists {cα (u)}, α ∈ L, where {cα (u)} is an optimal α-resolution for λ, such that
X
X
cα−1 (u)Hu ≥
cα (u)Hu
(209)
u∈Ωα
L
u∈Ωα−1
L
for all α = 2, 3, · · · , L.
Remark 6. Theorem 3 in [4] is the same as Theorem 5 above except that the former is for all λ ∈ RL
+ . By the
explicit characterization of Rsup in Theorem 2, namely R∗ , Theorem 5 is sufficient for proving the tightness of R∗ .
Remark 7. For α ∈ L and u ∈ Ωα
L , let
c̃α (u) =
1
.
(210)
L−1
α−1
It is not difficult to see that for all α ∈ L, {c̃α (u) : u ∈ Ωα
L } is the unique optimal α-resolution for λ = 1. Then
(209) in Theorem 5 becomes
1
L−1
α−2
X
u∈Ωα−1
L
Hu ≥
1
L−1
α−1
X
Hu ,
(211)
u∈Ωα
L
which is Han’s inequality [10]. It was proved in [6] that both Han’s inequality and the subset entropy inequality
in [4] can be established from the subset entropy inequality of Madiman and Tetali [11].
27
Proof of Theorem 5. By symmetry, we only have to prove the theorem for λ ∈ GL0 . We will prove the theorem by
induction on L. It is easy to check that the theorem is true for L = 2. Assume the theorem is true for L = N − 1,
0
we will show that the theorem is also true for L = N . This can be readily verified for λ ∈ GN
such that ζ = 1.
0
Thus, we only need to consider λ ∈ GN
such that ζ ≥ 2.
0
0
For any λN = (λ1 , λ2 , · · · , λN ) ∈ GN
, by the construction in (108), we have λN −1 = (λ2 , λ3 , · · · , λN ) ∈ GN
−1 .
For α ∈ {1, 2, · · · , N − 1}, by the induction hypothesis, let c̃α (u) : u ∈ Ωα
N −1 be an optimal α-resolution for
λN −1 such that (209) is satisfied for all α = 2, 3, · · · , N − 1. Now we need to design a proper optimal α-resolution
{cα (w) : w ∈ Ωα
N } for λN that satisfies (209) for all α = 2, 3, · · · , N .
From (108), there exists a θ ∈ {1, 2, · · · , N − 1} such that
N
λ1 =
1X
λi .
θ i=2
(212)
For any u ∈ {0, 1}N −1 and w ∈ {0, 1}N , let u = (u2 , u3 · · · , uN ) and w = (w1 , w2 , · · · , wN ). For α =
1, 2, · · · , N , we now construct an α-resolution for λN in (i) and (ii) in the following.
(i) Design {cα (w)} for α = θ + 1, θ + 2, · · · , N .
For α ≥ θ + 1 and w ∈ Ωα
N , let
cα (w) =
c̃α−1 (u), if w = (1, u), u ∈ Ωα−1
N −1
0,
From (212), we have
(213)
otherwise.
N
1 X
λi
λ1 ≥
α − 1 i=2
for all α = θ + 1, θ + 2, · · · , N.
(214)
Lemma 9 in [4] states that {cα (w)} is an optimal α-resolution for λN if
N
λ1 >
1 X
λi
α − 1 i=2
for all α = θ + 1, θ + 2, · · · , N.
(215)
We observe that the lemma can be strengthened by replacing the strict inequality in (215) by a non-strict
inequality (i.e., the condition in (214)) with essentially no change in the proof. Thus, by invoking this
strengthened version of the lemma, we conclude that {cα (w)} is an optimal α-resolution for λN .
For all α = θ + 2, θ + 3, · · · , N , following the steps leading to (48) in [4], we can check that
X
cα−1 (w)Hw ≥
w∈Ωα−1
N
X
cα (w)Hw .
(216)
w∈Ωα
N
(ii) Design {cα (w)} for α = 1, 2, · · · , θ.
For α = 2, 3, · · · , θ + 1 and any optimal α-resolution {cα (w) : w ∈ Ωα
N } for λN , we claim that there
exists an optimal (α − 1)-resolution {cα−1 (w) : w ∈ Ωα−1
N } for λN such that (209) is satisfied. Since
P
N
1
λ1 ≤ α−1
i=2 λi , this is exactly the first case of the proof of Proposition 1 in [4], which is relatively
straightforward.
0
In (i) and (ii) above, we have constructed an optimal α-resolution {cα (w)} for any λN ∈ GN
that satisfies (209)
for all α = 2, 3, · · · , N . This proves the theorem.
28
VII. C ONCLUSION
In this paper, we studied the problem of SMDC for which superposition coding was proved to be optimal in [1],
[4]. We enhanced their results by obtaining in closed form the minimum set of inequalities that is needed for
characterizing Rsup , the superposition coding rate region. We further show by the symmetry of the problem that
only a much smaller subset of these inequalities needs to be verified in determining the achievability of a given rate
tuple. Yet, the cardinality of this smaller set grows at least exponentially fast with L, thus revealing the inherent
complexity of the problem. A subset entropy inequality, which plays a key role in the converse proof in [4], requires
a painstaking and extremely technical and proof. We present a weaker version of this subset entropy inequality
whose proof is considerably simpler. With our explicit characterization of the coding rate region, this weaker version
of the subset entropy inequality is sufficient for proving the optimality of superposition coding. Some of our results
may be extensible to the more general settings in [5]–[9].
A PPENDIX A
P ROOF
L EMMA 9
PL
We first prove (i). By Lemma 7 (i), the condition λ1 ≤ η1 j=2 λj implies that
OF
L
1 X
λj .
η + 1 j=1
fη+1 (λ) =
(217)
In the following, we prove the claim by contradiction. Assume there exists a nonempty subset I1 ⊆ I such that
i ∈ I1 if and only if
(i)
λπi (1) >
L
1 X (i)
,
λ
η j=2 πi (j)
(218)
which is equivalent to
L
X
(i)
λπi (j)
j=1
or
L
1 X (i)
,
λ
> 1+
η j=2 πi (j)
L
(219)
L
1 X (i)
1 X (i)
.
λπi (j) >
λ
η + 1 j=1
η j=2 πi (j)
(220)
For all i ∈ I, by Lemma 7 (ii), the condition in (218) implies that
By Theorem 1, we have
(i)
(i)
(i)
fη+1 λ(i) = fη λπi (2) , λπi (3) , · · · , λπi (L) .
(221)
L
(i)
(i)
(i)
fη (λπi (2) , λπi (3) , · · · , λπi (L) ) ≤
1 X (i)
.
λ
η j=2 πi (j)
(222)
Then by (221), (222), and (220) we obtain
L
L
1 X (i)
1 X (i)
fη+1 (λ ) <
λπi (j) =
λ .
η + 1 j=1
η + 1 j=1 j
(i)
(223)
For i ∈ I\I1 , we have
(i)
λπi (1) ≤
L
1 X (i)
,
λ
η j=2 πi (j)
(224)
29
which by Lemma 7 (i) implies that
L
fη+1 (λ(i) ) =
1 X (i)
λ .
η + 1 j=1 j
(225)
Thus, we have from (116) that
fη+1 (λ) =
SL
X
ci fη+1 (λ(i) )
i=1
=
X
ci fη+1 (λ(i) ) +
i∈I1
<
=
X
ci fη+1 (λ(i) )
i∈I\I1
L
X
1
(i)
λj
ci
η
+
1
j=1
i=1
!
S
L
L
1 X X
(i)
ci λj
η + 1 j=1 i=1
SL
X
L
=
1 X
λj ,
η + 1 j=1
(226)
where the inequality follows from (223) and (225). This is a contradiction to (217). Thus, the assumption in (218)
is false and we have for all i ∈ I that
(i)
λπi (1) ≤
L
1 X (i)
.
λ
η j=2 πi (j)
(227)
Next, we prove (ii) by contradiction. Assume there exists a nonempty subset I2 ⊆ I such that i ∈ I2 if and only
if
(i)
λπi (1) <
L
1 X (i)
,
λ
η j=2 πi (j)
(228)
which is equivalent to
L
L
1 X (i)
1 X (i)
,
λπi (j) <
λ
η + 1 j=1
η j=2 πi (j)
or
gη+1,λ(i) (0) < gη+1,λ(i) (1).
(229)
(230)
For any i ∈ I2 , by Lemma 7 (i), (228) implies that
fη+1 (λ(i) ) = gη+1,λ(i) (0).
(231)
For any t ∈ {1, 2, · · · , η}, in light of (230), by applying the alternative version of Lemma 3 (ii) (see the remark
below Lemma 3), we obtain
fη+1 (λ(i) ) = gη+1,λ(i) (0) < gη+1,λ(i) (1) < · · · < gη+1,λ(i) (t).
(232)
Then it follows from the definition of πi (·) in (115) that
fη+1 (λ(i) ) < gη+1,λ(i) (t) =
L
L
X
X
1
1
(i)
(i)
λπi (j) ≤
λ ,
η + 1 − t j=t+1
η + 1 − t j=t+1 j
(233)
30
and so
fη+1 (λ(i) ) <
L
X
1
(i)
λ .
η + 1 − t j=t+1 j
(234)
For all i ∈ I\I2 and any t ∈ {1, 2, · · · , η}, by Theorem 1, similar to (233), we have
fη+1 (λ(i) ) ≤ gη+1,λ(i) (t) =
L
L
X
X
1
1
(i)
(i)
λπi (j) ≤
λ ,
η + 1 − t j=t+1
η + 1 − t j=t+1 j
and so
fη+1 (λ(i) ) ≤
(235)
L
X
1
(i)
λ .
η + 1 − t j=t+1 j
(236)
Thus, by (116), we have for any t ∈ {1, 2, · · · , η} that
fη+1 (λ)
=
SL
X
ci fη+1 (λ(i) )
i=1
<
=
L
X
1
(i)
ci
λj
η
+
1
−
t
i=1
j=t+1
SL
X
L
X
1
λj
η + 1 − t j=t+1
= gη+1,λ (t),
(237)
where the inequality follows from (234) and (236).
PL
The condition λ1 ≥ η1 j=2 λj is equivalent to
gη+1,λ (0) ≥ gη+1,λ (1).
(238)
Then by Theorem 1, we have
fη+1 (λ) =
min
β∈{0,1,··· ,η}
gη+1,λ (β) =
min
β∈{1,2,··· ,η}
gη+1,λ (β).
(239)
Thus, there must exist a t ∈ {1, 2, · · · , η} such that
fη+1 (λ) = gη+1,λ (t),
(240)
which is a contradiction to (237). Thus the assumption in (228) is false and we have for all i ∈ I that
(i)
λπi (1) ≥
L
1 X (i)
.
λ
η j=2 πi (j)
(241)
A PPENDIX B
P ROOF
OF
L EMMA 10
(i)
1
η
For all i ∈ I, since λ(i) ∈ GL in light of (108), we only need to prove λπi (1) ≤
PL
(i)
j=2 λπi (j) .
1
η−1
PL
j=2
(i)
(i)
λπi (j) and λπi (1) ≥
(i)
We first prove the upper bound on λπi (1) . For η = 1 and i = 1, we have
L
(i)
λπi (1) = 1 =
1 X (i)
.
λ
η − 1 j=2 πi (j)
(242)
31
For η = 1 and i ∈ I\{1}, it is obvious that
L
(i)
λπi (1) <
1 X (i)
= ∞.
λ
η − 1 j=2 πi (j)
(243)
For η ∈ {2, 3, · · · , ζ − 1}, the upper bound in (118) can be rewritten as
λ1 <
L
1 X
λj ,
η ′ j=2
(244)
where η ′ = η − 1 and η ′ ∈ {1, 2, · · · , ζ − 2}. By Lemma 9 (i), this implies that
(i)
λπi (1) ≤
L
L
1 X (i)
1 X (i)
.
=
λ
λ
η ′ j=2 πi (j)
η − 1 j=2 πi (j)
(245)
(i)
Thus, the upper bound on λπi (1) is proved.
(i)
Now we prove the lower bound on λπi (1) . For η ∈ {1, 2, · · · , ζ − 1}, the lower bound in (118) is
λ1 ≥
L
1X
λj ,
η j=2
(246)
and so by Lemma 9 (ii), we have
(i)
λπi (1) ≥
L
1 X (i)
.
λ
η j=2 πi (j)
(247)
If the lower bound in (118) is tight, it follows immediately from Lemma 9 that for any η ∈ {1, 2, · · · , ζ − 1},
(i)
λπi (1) =
L
1 X (i)
.
λ
η j=2 πi (j)
(248)
This proves the lemma.
A PPENDIX C
P ROOF
OF
L EMMA 11
We only need to show that for α = 1, 2, · · · , L − 1,
SL
X
(i)
(i)
(i)
ci fα (λ2 , λ3 , · · · , λL ).
(249)
By (118) and Lemma 10, we have for i ∈ I that
L
L
1 X
X
1
(i)
(i)
(i)
λπi (1) ∈
λπi (j) ,
λπi (j) .
η
η − 1 j=2
j=2
(250)
fα (λL−1 ) =
i=1
Consider the following two cases:
i) α = 1, 2, · · · , η − 1;
ii) α = η, η + 1, · · · , L − 1.
Case i): For α = 1, 2, · · · , η − 1, if η = 2, then α can only be 1 and it is easy to see that
f1 (λL−1 ) =
L
X
j=2
λj =
SL
L X
X
j=2 i=1
(i)
ci λj
=
SL
X
i=1
ci
L
X
j=2
(i)
λj
=
SL
X
i=1
(i)
(i)
(i)
ci f1 (λ2 , λ3 , · · · , λL ).
(251)
32
If η > 2, consider the following. The second inequality in (118) is equivalent to
L
L
1 X
1X
λj <
λj
η j=1
η − 1 j=2
or
(252)
gη,λ (0) < gη,λ (1).
(253)
By applying the alternative version of Lemma 3 (ii) (see the remark below Lemma 3), we obtain
gη,λ (1) < gη,λ (2),
(254)
which is equivalent to
L
L
1 X
1 X
λj <
λj
η − 1 j=2
η − 2 j=3
or
(255)
L
λ2 <
1 X
λj .
η − 2 j=3
(256)
Then by Lemma 7 (i), we have
fα (λL−1 ) = gα,λL−1 (0) =
L
1X
λj .
α j=2
(257)
Since (250) implies
L
(i)
λπi (1) ≤
1 X (i)
,
λ
η − 1 j=2 πi (j)
(258)
similar to (252)-(256) (with all <’s replaced by ≤’s), we have
L
(i)
λπi (2) ≤
1 X (i)
.
λ
η − 2 j=3 πi (j)
(259)
Thus, following (258) and (259), we have
L
(i)
λπi (1)
≤
≤
1 X (i)
λ
η − 1 j=2 πi (j)
1
η−1
X
L
1
(i)
λπi (j)
+1
η−2
j=3
L
=
≤
1 X (i)
λ
η − 2 j=3 πi (j)
1
η−2
X
(i)
λπi (j)
(260)
j∈{2,3,··· ,L}\{j0 }
for any j0 ∈ {2, 3, · · · , L}. Let πi′ (·) be a permutation of {2, 3, · · · , L} defined as follows:
a) if πi (1) = 1, then πi′ (j) = πi (j) for all j ∈ {2, 3, · · · , L};
b) if πi (j0 ) = 1 for some j0 ∈ {2, 3, · · · , L}, then
πi (j − 1), for j = 2, 3, · · · , j0
πi′ (j) =
πi (j),
for j = j0 + 1, · · · , L.
(261)
33
It is easy to check that
(i)
(i)
(i)
λπ′ (2) ≥ λπ′ (3) ≥ · · · ≥ λπ′ (L) .
(262)
i
i
i
If a) holds, then by (259), we have
L
L
1 X (i)
1 X (i)
λπi (j) =
λ ′ .
η − 2 j=3
η − 2 j=3 πi (j)
(i)
(i)
λπ′ (2) = λπi (2) ≤
i
(263)
If b) holds, then by (260), we have
(i)
(i)
λπ′ (2) = λπi (1) ≤
i
L
1
η−2
(i)
X
λπi (j) =
j∈{2,3,··· ,L}\{j0 }
1 X (i)
λ ′ .
η − 2 j=3 πi (j)
(264)
Summarizing the two cases, we see that
L
(i)
λπ′ (2) ≤
i
1 X (i)
λ ′
η − 2 j=3 πi (j)
(265)
always holds. By Lemma 7 (i), this implies that
L
(i)
(i)
(i)
fα (λ2 , λ3 , · · · , λL ) =
L
1 X (i)
1 X (i)
λπ′ (j) =
λ .
i
α j=2
α j=2 j
(266)
Following (257), we have
L
1X
λj
α j=2
fα (λL−1 ) =
(267)
!
SL
L
1X X
(i)
ci λj
α j=2 i=1
SL
L
X
X
1
(i)
ci
λj
α
i=1
j=2
=
=
SL
X
=
(i)
(i)
(268)
(269)
(i)
ci fα (λ2 , λ3 , · · · , λL ),
(270)
i=1
where (268) follows from (176) and (270) follows from (266).
Case ii): For α = η, η + 1, · · · , L − 1, by Lemma 7 (ii), the first inequality in (118) implies that
fα (λL−1 ) = fα+1 (λ).
(i)
For any i ∈ {1, 2, · · · , SL }, since (250) implies λπi (1) ≥
1
η
PL
(271)
(i)
j=2
λπi (j) , we have by Lemma 7 (ii) that
(i)
(i)
(i)
fα+1 (λ(i) ) = fα λπi (2) , λπi (3) , · · · , λπi (L) .
(272)
From the definition of πi′ (·), it is readily seen that
(i)
(i)
λπi (j) ≤ λπ′ (j) , for all j = 2, 3, · · · , L.
(273)
i
Thus, we have for any β = 1, 2, · · · , α − 1 that
L
L
X
X
1
1
(i)
(i)
λπi (j) ≤
λπ′ (j) .
i
α−β
α−β
j=β+1
j=β+1
(274)
34
By Theorem 1, this implies that
(i)
(i)
(i)
(i)
(i)
(i)
fα λπi (2) , λπi (3) , · · · , λπi (L) ≤ fα λπ′ (2) , λπ′ (3) , · · · , λπ′ (L) ,
(275)
i
i
i
and thus by (272), we have
(i)
(i)
(i)
fα+1 (λ(i) ) ≤ fα λ2 , λ3 , · · · , λL .
Following (271), we have
fα (λL−1 ) =
(276)
fα+1 (λ)
SL
X
=
(277)
ci fα+1 (λ(i) )
(278)
i=1
SL
X
≤
(i)
(i)
(i)
ci fα λ2 , λ3 , · · · , λL
i=1
≤
fα
SL
X
(i)
(i)
ci λ2 , λ3 , · · ·
i=1
=
fα
SL
X
(i)
ci λ2 ,
SL
X
=
fα (λ2 , λ3 , · · · , λL )
=
fα (λL−1 ),
!
(i)
, λL
(i)
ci λ3 , · · ·
i=1
i=1
(279)
,
SL
X
i=1
(280)
(i)
ci λL
!
(281)
(282)
(283)
where both (278) and (282) follow from (176), (279) follows from (276), and (280) follows from Lemma 8. Upon
observing that the LHS of (277) is the same as the RHS of (283), we conclude that the inequalities in both (279)
and (280) are tight, and hence
fα (λL−1 ) =
SL
X
i=1
The lemma is proved.
(i)
(i)
(i)
ci fα λ2 , λ3 , · · · , λL .
(284)
A PPENDIX D
P ROOF
L EMMA 12
(i)
(i)
(i)
Fix i ∈ {1, 2, · · · , SL } and assume that λ2 , λ3 , · · · , λL ∈ GL−1 . Let γj , j = 1, 2, · · · , ζ − 1 be the integer
OF
such that
(i)
λπi (j) =
L
1 X (i)
λπi (k) ,
γj
(285)
k=j+1
and let γζ = 0. Note that the role of γj for λ(i) is the same as the role of θj for λ (cf. (109)). Also note that ζ
and γj depend on i, but since we fix i, this dependence is omitted to simplify notation.
Let j0 ∈ {1, 2, · · · , L} be such that
(i)
(i)
λπi (j0 ) = λ1 .
(286)
(i)
If j0 ≥ ζ + 1, λ1 = 0 and thus the lemma is proved. If j0 = 1, the lemma is immediate from (286). If 2 ≤ j0 ≤ ζ,
(i)
(i)
we claim that γj = γj0 + (j0 − j) and λπi (j) = λπi (j0 ) for all j = j0 , j0 − 1, · · · , 1. Then the lemma follows from
35
the claim for j = 1. In the following, we prove the claim by induction on j for j ≤ j0 . The claim is immediate
for j = j0 . Assume the claim is true for j = j0 , j0 − 1, · · · , N for some N ∈ {2, 3, · · · , j0 }, and we will show
that the claim is also true for j = N − 1. By the induction hypothesis, we have
γN = γj0 + (j0 − N ),
(287)
and for all j = j0 , j0 − 1, · · · , N ,
L
X
1
γj0
(i)
(i)
λπi (j) = λπi (j0 ) =
(i)
λπi (k) .
(288)
k=j0 +1
By (108) and (110), there exists
γN −1 ∈ {1, 2, · · · , γj0 + (j0 − N + 1)}
such that
(i)
λπi (N −1)
L
X
1
=
γN −1
(i)
=
=
=
1
γN −1
L
X
(i)
λπi (k) .
(290)
k=N
Thus, we have
λπi (N −1)
(289)
(i)
λπi (k)
k=N
j0
X
1
1+
γj0
1
γN −1
k=N
!
L
X
(i)
λπi (k)
(291)
k=j0 +1
L
γj0 + (j0 − N + 1) X (i)
λπi (k) ,
γj0 γN −1
(292)
k=j0 +1
where (291) follows from (288). In light of (286) and j0 ≥ 2, recall the definition of πi′ (·) in (261). With the
(i)
(i)
(i)
′
assumption that λ2 , λ3 , · · · , λL ∈ GL−1 , by (108), there exists an integer γN
−1 such that
(i)
λπ′ (N )
i
=
or
(i)
λπi (N −1) =
1
′
γN
−1
L
X
1
′
γN
−1
jX
0 −1
(i)
λπ′ (k)
(i)
λπi (k) +
k=N
L
X
(i)
k=N
(i)
λπi (k)
>
jX
0 −1
k=N
(i)
λπi (k) +
(i)
k=j0 +1
Comparing the RHS of (290) and (294), since λπi (j0 ) ≥ 1, we have
L
X
(293)
i
k=N +1
L
X
λπi (k) .
(i)
λπi (k) .
(294)
(295)
k=j0 +1
′
Since the LHS of (290) and (294) are the same, we see that γN
−1 < γN −1 , which implies that
′
γN
−1 ≤ γj0 + (j0 − N ).
(296)
36
Following (294), we have
(i)
λπi (N −1)
=
=
=
1
′
γN
−1
1
′
γN
−1
jX
0 −1
(i)
λπi (k)
+
k=j0 +1
k=N
1+
L
X
jX
0 −1
k=N
1
γj0
!
L
X
(i)
λπi (k)
(i)
λπi (k)
(297)
k=j0 +1
L
γj0 + (j0 − N ) X (i)
λπi (k) ,
′
γj0 γN
−1
(298)
k=j0 +1
where (297) follows from (288). Comparing (292) and (298), it is easy to see that
′
γN
−1 =
γN −1 [γj0 + (j0 − N )]
.
γj0 + (j0 − N + 1)
(299)
Since γj0 + (j0 − N ) and γj0 + (j0 − N + 1) are coprime and γN −1 ≤ γj0 + (j0 − N + 1) by (289), we have
γN −1 = γj0 + (j0 − N + 1).
(i)
(300)
(i)
Substituting (300) into (292) and invoking (287), we have λπi (N −1) = λπi (j0 ) . This implies that the claim is true
for j = N − 1. The lemma is proved.
A PPENDIX E
P ROOF
OF
L EMMA 13
Since there is only one vector in G1 , we only need to consider L ≥ 2. If ζ = 1 for λ(i0 ) , it is obvious that λ(i0 )
cannot be a conic combination of the other vectors in GL . Thus, we consider only λ(i0 ) ∈ GL such that ζ ≥ 2. We
prove the lemma by induction on L for L ≥ 2. We first check that the claim is true for L = 2. It is easy to see
from (108) that G2 = {(1, 0), (0, 1), (1, 1)}. Then f (1, 0) = f (0, 1) = (1, 0) and f (1, 1) = (2, 1). Since
(1, 1) = (1, 0) + (0, 1)
(301)
f2 (1, 1) > f2 (1, 0) + f2 (0, 1) ,
(302)
whereas
we see that (1, 1), f (1, 1) cannot be a conic combination of (1, 0), f (1, 0) and (0, 1), f (0, 1) . Thus, the
lemma is true for L = 2. For any L ≥ 3, the lemma will be proved by contradiction via the following proposition,
whose proof is given in Appendix F.
Proposition 1. If Lemma 13 is false for any L ≥ 3, then the lemma is false for L − 1.
By backward induction, if Lemma 13 is false for any L ≥ 3, then the lemma is false for L = 2. This is a
contradiction because we already have shown that the lemma is true for L = 2. This proves the lemma for all
L ≥ 2.
37
A PPENDIX F
P ROOF
OF
P ROPOSITION 1
Assume Lemma 13 is false for some L ≥ 3, i.e., for some i0 ∈ {1, 2, · · · , SL }, there exists (c1 , c2 , · · · , cSL ) ∈
RS+L
such that ci0 = 0 and
SL
X
ci λ(i) , f (λ(i) ) .
λ(i0 ) , f (λ(i0 ) ) =
(303)
i=1
Assume without loss of generality that λ(i0 ) ∈ GL0 . Since we assume at the beginning of Appendix E that
(i )
(i )
(i )
0
ζ ≥ 2 for λ(i0 ) , we can see from (108) that λ2 0 , λ3 0 , · · · , λL 0 ∈ GL−1
by construction. Let GL−1 =
n
o
(1)
(2)
(SL−1 )
λL−1 , λL−1 , · · · , λL−1
. Then there exists a unique j0 ∈ {1, 2, · · · , SL−1 } such that
(i )
(i )
(i )
(j0 )
(304)
= λ2 0 , λ3 0 , · · · , λL 0 .
λL−1
By Lemma 11, (303) implies that
SL
X
(i)
(i)
(i)
(i)
(i)
(i)
(j0 )
(j0 )
ci (λ2 , λ3 , · · · , λL ), f (λ2 , λ3 , · · · , λL ) .
λL−1
, f (λL−1
) =
(305)
i=1
o
n
(j0 )
(i)
(i)
(i)
(j )
, and
Let KL = {1, 2, · · · , SL }, IL = {i ∈ KL : ci 6= 0}, KL 0 = i ∈ KL : λ2 , λ3 , · · · , λL = λL−1
X
d0 =
ci .
(306)
(j )
i∈KL 0
In the proof of Theorem 2, we have shown that any vector in FL1 is a conic combination of the vectors in FL2 .
(j )
(i) (i)
(i)
S
Then for any i ∈ KL \KL 0 , there exists t1 , t2 , · · · , tSL−1 ∈ R+L−1 such that
SL−1
X (i) (j)
(j)
(i)
(i)
(i)
(i)
(i)
(i)
tj λL−1 , f (λL−1 ) .
(λ2 , λ3 , · · · , λL ), f (λ2 , λ3 , · · · , λL ) =
(307)
j=1
Substitute (306) and (307) into (305), we have
(j0 )
(j0 )
λL−1
, f (λL−1
)
(j0 )
(j0 )
, f (λL−1
) +
= d0 λL−1
= d0 +
(j )
i∈KL \KL 0
= d0 +
Thus,
1 − d0 −
X
(j )
(i)
ci tj0
ci
SL−1
X
(i)
tj
j=1
(j0 )
(j0 )
)
, f (λL−1
λL−1
+
(j )
(j0 )
(j0 )
(i)
) +
, f (λL−1
ci tj0 λL−1
(j0 )
(j0 )
(i)
) =
, f (λL−1
ci tj0 λL−1
X
j∈KL−1 \{j0 }
(j)
(j)
λL−1 , f (λL−1 )
X
(j )
i∈KL \KL 0
i∈KL \KL 0
i∈KL \KL 0
X
(j )
i∈KL \KL 0
X
X
X
ci
(j )
(j0 )
(i)
(i)
(i)
.
Proposition 2. There exists i ∈ IL such that λ2 , λ3 , · · · , λL 6= λL−1
(i)
tj
X
(j)
(j)
λL−1 , f (λL−1 )
(j )
i∈KL \KL 0
i∈KL \KL 0
X
j∈KL−1 \{j0 }
j∈KL−1 \{j0 }
X
(j)
(j)
(i)
ci tj λL−1 , f (λL−1 ) .
(j)
(j)
(i)
ci tj λL−1 , f (λL−1 ) .
(308)
(309)
38
The proof of Proposition 2 is given in Appendix G. The proposition implies that
(j )
IL ∩ (KL \KL 0 ) 6= ∅.
(310)
(j )
For any i ∈ IL ∩ (KL \KL 0 ), we can rewrite (307) as follows:
(i)
(i)
(i)
(i)
(i)
(i)
(λ2 , λ3 , · · · , λL ), f (λ2 , λ3 , · · · , λL )
X
(j)
(j)
(i)
(j0 )
(j0 )
(i)
tj λL−1 , f (λL−1 ) .
) +
, f (λL−1
= tj0 λL−1
(311)
j∈KL−1 \{j0 }
Since
(i)
(i)
(λ2 , λ3 , · · ·
(i)
, λL )
6=
(j0 )
λL−1
,
there must exist j ∈ KL−1 \{j0 } such that
(i)
tj > 0.
(312)
L−1
L−1
For any x, y ∈ R+
, define a binary relation ‘>’ by x > y if and only if (x − y) ∈ R+
, i.e., x is strictly
greater than y in at least one component (cf. (7)). Then for the RHS of (309), we have
X
X
(j)
(j)
(i)
ci tj λL−1 , f (λL−1 )
j∈KL−1 \{j0 }
X
=
(j )
i∈KL \KL 0
(j )
i∈KL \KL 0
X
=
(i)
X
ci
tj
(j)
(j)
λL−1 , f (λL−1
j∈KL−1 \{j0 }
(j )
i∈IL ∩(KL \KL 0 )
> 0,
(i)
X
ci
tj
j∈KL−1 \{j0 }
)
(j)
(j)
λL−1 , f (λL−1 )
(313)
(j)
where the inequality follows from (310), (312) and the fact that λL−1 > 0 for all j ∈ KL−1 \{j0 }. Then we can
see from (309) that
which implies that
1 − d0 −
X
(j )
i∈KL \KL 0
(j0 )
(j0 )
(i)
, f (λL−1
) > 0,
ci tj0 λL−1
X
1 − d0 −
(i)
ci tj0 > 0.
(314)
(315)
(j )
i∈KL \KL 0
For each j ∈ KL−1 \{j0 }, let
dj =
1
1 − d0 −
P
(j )
i∈KL \KL 0
(i)
c i tj .
(i)
ci tj0
(j )
i∈KL \KL 0
X
(316)
It is easy to see that dj ≥ 0 for all j ∈ KL−1 \{j0 }. By (313), there exists j ∈ KL−1 \{j0 } such that dj > 0. Upon
letting dj = 0 for j = j0 , by (309) and (316), we have
SL−1
(j0 )
(j0 )
λL−1
, f (λL−1
)
=
X
j=1
(j)
(j)
dj λL−1 , f (λL−1 ) .
This means that Lemma 13 is false for L − 1. The proposition is proved.
(317)
39
A PPENDIX G
P ROOF
OF
P ROPOSITION 2
Since λ(i0 ) ∈ GL0 , there exists a unique ηi0 ∈ {0, 1, · · · , L − 1} such that
L
1 X (i0 )
(318)
λ .
ηi0 j=2 j
(i )
(i )
(i )
= λ2 0 , λ3 0 , · · · , λL 0 . Since we assume at the beginning of Appendix E that
(i )
λ1 0 =
(j )
0
Recall from (304) that λL−1
ζ ≥ 2 for λ(i0 ) , we see that
(j )
0
λL−1
6= 0,
which implies that
PL
j=2
(319)
(i )
(i )
λj 0 > 0. Then we have ηi0 6= 0, otherwise λ1 0 = ∞ in (318). Thus,
ηi0 ∈
{1, 2, · · · , L − 1}. By Remark 2 following Lemma 7, we have
L
(j )
0
)=
fηi0 +1 (λ(i0 ) ) = fηi0 (λL−1
1 X (i0 )
λj .
ηi0 + 1
(320)
j=1
We now prove the proposition by contradiction. Assume that for all i ∈ IL ,
(i)
(i)
(i)
(j0 )
λ2 , λ3 , · · · , λL = λL−1
.
(321)
This means that for each i ∈ IL ,
(i)
(i)
(i)
(i)
(i)
(i )
(i )
(i )
λ(i) = λ1 , λ2 , λ3 , · · · , λL = λ1 , λ2 0 , λ3 0 , · · · , λL 0 .
(j )
0
Furthermore, since λ(i0 ) ∈ GL0 , we see from (108) that λL−1
(i)
(i)
0
GL−1
. Then by Lemma 12, we have λ1 = 0 or λπi (1) .
(322)
(i)
(i)
(i)
0
∈ GL−1
by construction, so that λ2 , λ3 , · · · , λL ∈
(i)
(i)
Let IL0 be the subset of IL such that i ∈ IL0 if and only if λ1 = 0. For i ∈ IL0 , it is easy to see that λπi (L) = 0.
Then upon noting that
by Theorem 1 we have
(i )
(i )
(i )
(i)
(i)
(i)
λπi (1) , λπi (2) , · · · , λπi (L−1) = λ2 0 , λ3 0 , · · · , λL 0 ,
(i)
fηi0 +1 (λ ) =
(323)
L−1
X (i)
1
min
λπi (j)
β∈{0,1,··· ,ηi0 } (ηi0 + 1) − β
j=β+1
=
L
X
1
(i )
λj 0
β∈{0,1,··· ,ηi0 } (ηi0 + 1) − β
min
j=β+2
=
(j0 )
)
fηi0 +1 (λL−1
<
0
),
fηi0 (λL−1
(j )
(324)
where the inequality follows from Lemma 5 in [4]. By (320), this implies that
L
fηi0 +1 (λ(i) ) <
(i)
(i)
X (i )
1
λ 0 .
ηi0 + 1 j=1 j
(325)
For i ∈ IL \IL0 , we have λ1 = λπi (1) . Since λ(i) ∈ GL , there exists a unique ηi ∈ {0, 1, · · · , L − 1} such that
(i)
λ1 =
L
1 X (i)
λ .
ηi j=2 j
(326)
40
(i)
(i)
(i)
From (319) and the assumption in (321), we have λ2 , λ3 , · · · , λL 6= 0. Then from (326), we have ηi 6= 0,
and thus ηi ∈ {1, 2, · · · , L − 1}. Since ci0 = 0 in (303), we have λ(i) 6= λ(i0 ) for all i ∈ IL and hence for all
(i)
(i )
i ∈ IL \IL0 . In light of (322), λ(i) 6= λ(i0 ) implies λ1 6= λ1 0 , and upon comparing (318) and (326), we see that
ηi 6= ηi0 .
(327)
Let IL1 = {i ∈ IL \IL0 : ηi > ηi0 } and IL2 = {i ∈ IL \IL0 : ηi < ηi0 }. Then we can see from (327) that
IL0 ∪ IL1 ∪ IL2 = IL .
(i)
(328)
(i )
For i ∈ IL1 , we have λ1 < λ1 0 , which is equivalent to
(i)
λ1 <
L
1 X (i)
λ .
ηi0 j=2 j
(329)
Thus, by Lemma 7 (i), we have
L
L
X (i )
1 X (i)
1
fηi0 +1 (λ ) =
λj <
λ 0 .
ηi0 + 1 j=1
ηi0 + 1 j=1 j
(i)
(i)
(330)
(i )
On the other hand, for i ∈ IL2 , we have λ1 > λ1 0 , which is equivalent to
(i)
λ1 >
L
1 X (i)
λ .
ηi0 j=2 j
(331)
Thus, by Lemma 7 (ii), we have
(j )
0
),
fηi0 +1 (λ(i) ) = fηi0 +1 (λL−1
(332)
which by (320) implies that
L
fηi0 +1 (λ(i) ) =
X (i )
1
λ 0 .
ηi0 + 1 j=1 j
(333)
From (303) and (328), we have
fηi0 +1 (λ
(i0 )
) =
SL
X
ci fηi0 +1 (λ(i) )
i=1
=
X
ci fηi0 +1 (λ(i) )
i∈IL
=
X
0
i∈IL
ci fηi0 +1 (λ(i) ) +
X
ci fηi0 +1 (λ(i) ) +
1
i∈IL
X
ci fηi0 +1 (λ(i) ).
(334)
2
i∈IL
Comparing (320) for fηi0 +1 (λ(i0 ) ) and (325), (330), and (333) for fηi0 +1 (λ(i) ), we see that both IL0 and IL1 must
be empty in order for the equality in (334) to hold, and hence
IL = IL2 .
(335)
41
For any i ∈ IL2 , since ηi < ηi0 and ηi ≥ 1, we see that ηi0 ≥ 2. Thus from (318), we have
L
L
L
1 1 X (i0 ) X (i0 )
1 X (i0 )
λj
=
+
λ
λ
ηi0 j=1 j
ηi0 ηi0 j=2 j
j=2
=
1
ηi0
X
L
1
(i )
λj 0
+1
ηi0
j=2
L
<
Then by Lemma 7 (i), (318) implies that
fηi0 (λ(i0 ) ) =
Since λ(i0 ) is ordered, by (318), we have
X (i )
1
λ 0 .
ηi0 − 1 j=2 j
L
L
X
1
1 X (i0 )
(i )
λj <
λ 0 .
ηi0 j=1
ηi0 − 1 j=2 j
(i )
(i )
λ2 0 ≤ λ1 0 =
which implies that
L
1 X (i0 )
λ ,
ηi0 j=2 j
(336)
(337)
(338)
L
(i )
λ2 0 ≤
Then by Lemma 7 (i), we have
X (i )
1
λ 0 .
ηi0 − 1 j=3 j
(339)
L
(j0 )
)
fηi0 −1 (λL−1
It follows from (337) and (340) that
X (i )
1
=
λ 0 .
ηi0 − 1 j=2 j
(j )
0
).
fηi0 (λ(i0 ) ) < fηi0 −1 (λL−1
(340)
(341)
For i ∈ IL2 , we have ηi0 ≥ ηi + 1. Then by Lemma 7 (ii), (326) implies that
(j )
0
).
fηi0 (λ(i) ) = fηi0 −1 (λL−1
(342)
Following (303), we have
(i )
(i )
(i )
λ2 0 , λ3 0 , · · · , λL 0
X (i) (i)
(i)
=
ci λ2 , λ3 , · · · , λL
i∈IL
=
X
2
i∈IL
=
(i)
(i)
(i)
ci λ2 , λ3 , · · · , λL
X
2
i∈IL
ci
(i )
(i )
(i )
λ2 0 , λ3 0 , · · · , λL 0 ,
(343)
(344)
where (343) follows from (335) and (344) follows from the assumption in (321). Thus we have
X
ci = 1.
(345)
Then from (342) and (345), we see that
X
X
(j0 )
(j0 )
),
) = fηi0 −1 (λL−1
ci fηi0 (λ(i) ) =
ci fηi0 −1 (λL−1
(346)
2
i∈IL
2
i∈IL
2
i∈IL
42
and it follows from (341) that
X
ci fηi0 (λ(i) ) > fηi0 (λ(i0 ) .
(347)
2
i∈IL
This is a contradiction to (303). Therefore, the assumption in (321) is false and the proposition is proved.
A PPENDIX H
P ROOF
OF
L EMMA 14
For any permutation ω on {1, 2, · · · , L} and any λ ∈ RL
+ , recall from the beginning of Section III that
ω(λ) = λω(1) , λω(2) , · · · , λω(L) .
(348)
Then for the ordered permutation π, we have π(λ) = λπ(1) , λπ(2) , · · · , λπ(L) .
If λ = π(λ), the lemma is immediate. Otherwise, let ω0 (i) = i for all i ∈ L so that ω0 (λ) = λ. Set t = 1 and
we sort λ in descending order by iteration as follows:
(i) Let it = min{i ∈ L : ωt−1 (i) 6= π(i)}. Let kt , jt be any indexes in L such that
π(kt ) = ωt−1 (it )
(349)
ωt−1 (jt ) = π(it ).
(350)
and
It is easy to check that kt > it and jt > it , which implies
λπ(it ) − λπ(kt ) ≥ 0
(351)
Rjt − Rit ≥ 0.
(352)
and
Let ωt (λ) = (λωt (1) , λωt (2) , · · · , λωt (L) ) be a permutation of ωt−1 (λ) where we switch λωt−1 (jt ) and λωt−1 (it ) ,
i.e.,
π(it ),
if i = it
ωt (i) = π(kt ),
if i = jt
ω (i), otherwise.
t−1
Then we have
L
X
λωt−1 (i) Ri −
=
λωt (i) Ri
i=1
i=1
=
L
X
(353)
λωt−1 (it ) Rit + λωt−1 (jt ) Rjt − λωt (it ) Rit + λωt (jt ) Rjt
λπ(kt ) Rit + λπ(it ) Rjt − λπ(it ) Rit + λπ(kt ) Rjt
=
Rit (λπ(kt ) − λπ(it ) ) + Rjt (λπ(it ) − λπ(kt ) )
=
(λπ(it ) − λπ(kt ) )(Rjt − Rit )
≥
0,
(354)
43
where the second equality follows from (349), (350) and (353), and the inequality follows from (351) and
(352).
(ii) If ωt (λ) = π(λ), return T = t and stop. Otherwise, let t = t + 1 and go back to step (i).
At the end of the iteration, ωT (λ) is sorted in the same order as π(λ), and we have
L
X
λi Ri −
λπ(i) Ri
i=1
i=1
=
L
X
L
X
λω0 (i) Ri −
L
T
X
X
t=1
≥
λωT (i) Ri
i=1
i=1
=
L
X
λωt−1 (i) Ri −
L
X
λωt (i) Ri
i=1
i=1
0.
!
(355)
This proves the lemma.
R EFERENCES
[1] J. R. Roche, R. W. Yeung, and K. P. Hau, “Symmetrical multilevel diversity coding,” IEEE Trans. Inf. Theory, vol. 43, pp. 1059–1064,
May 1997.
[2] R. W. Yeung and Z. Zhang, “Distributed source coding for satellite communications,” IEEE Trans. Inf. Theory, vol. 45, pp. 1111–1120,
May 1999.
[3] L. Song, R. W. Yeung, and N. Cai, “Zero-error network coding for acyclic networks,” IEEE Trans. Inf. Theory, vol. 49, pp. 3129–3139,
Dec 2003.
[4] R. W. Yeung and Z. Zhang, “On symmetrical multilevel diversity coding,” IEEE Trans. Inf. Theory, vol. 45, pp. 609–621, Mar. 1999.
[5] A. Balasubramanian, H. D. Ly, S. Li, T. Liu, and S. L. Miller, “Secure symmetrical multilevel diversity coding,” IEEE Trans. Inf. Theory,
vol. 59, pp. 3572–3581, June 2013.
[6] J. Jiang, N. Marukala, and T. Liu, “Symmetrical multilevel diversity coding and subset entropy inequalities,” IEEE Trans. Inf. Theory,
vol. 60, pp. 84–103, Jan. 2014.
[7] Z. Xiao, J. Chen, Y. Li, and J. Wang, “Distributed multilevel diversity coding,” IEEE Trans. Inf. Theory, vol. 61, pp. 6368–6384, Nov
2015.
[8] C. Tian and T. Liu, “Multilevel diversity coding with regeneration,” IEEE Trans. Inf. Theory, vol. 62, pp. 4833–4847, Sept 2016.
[9] S. Mohajer, C. Tian, and S. N. Diggavi, “Asymmetric multilevel diversity coding and asymmetric gaussian multiple descriptions,” IEEE
Trans. Inf. Theory, vol. 56, pp. 4367–4387, Sept 2010.
[10] T. S. Han, “Nonnegative entropy measures of multivariate symmetric correlations,” Inf. Control, vol. 36, pp. 133–156, 02 1978.
[11] M. Madiman and P. Tetali, “Information inequalities for joint distributions, with interpretations and applications,” IEEE Trans. Inf. Theory,
vol. 56, pp. 2699–2713, June 2010.
A PPENDIX I
TABLES
OF
N ON -R EDUNDANT λ
For L = 1, 2, · · · , 5, the vectors λ ∈ GL0 and the corresponding fα (λ) are listed in the following tables. The
L
P
parameter θ is the integer such that λ1 = θ1
λi .
i=2
44
λ
f1 (λ)
suffix
λ
f1 (λ)
f2 (λ)
θ
1
1
-
(1, 0)
1
0
0
(1)
(1, 1)
2
1
1
TABLE I
NON - REDUNDANT
CONSTRAINT FOR L
TABLE II
= 1.
NON - REDUNDANT CONSTRAINTS FOR
suffix
λ
f1 (λ)
f2 (λ)
f3 (λ)
θ
-
(1, 0, 0)
1
0
0
0
(1, 0)
(1, 1, 0)
2
1
0
1
(1, 1, 1)
3
3
2
1
2
4
2
1
1
(1, 1)
(2, 1, 1)
L = 2.
TABLE III
NON - REDUNDANT CONSTRAINTS FOR
L = 3.
suffix
λ
f1 (λ)
f2 (λ)
f3 (λ)
f4 (λ)
θ
-
(1, 0, 0, 0)
1
0
0
0
0
(1, 0, 0)
(1, 1, 0, 0)
2
1
0
0
1
(1, 1, 1, 0)
3
3
2
1
0
2
(2, 1, 1, 0)
4
2
1
0
1
(1, 1, 1, 1)
4
2
4
3
1
3
( 23 , 1, 1, 1)
9
2
9
4
3
2
1
2
(3, 1, 1, 1)
6
3
3
2
1
1
(2, 2, 1, 1)
6
3
2
1
2
8
4
2
1
1
(1, 1, 0)
(1, 1, 1)
(2, 1, 1)
(4, 2, 1, 1)
TABLE IV
NON - REDUNDANT CONSTRAINTS FOR
L = 4.
45
suffix
λ
f1 (λ)
f2 (λ)
f3 (λ)
f4 (λ)
f5 (λ)
θ
-
(1, 0, 0, 0, 0)
1
0
0
0
0
0
(1, 0, 0, 0) (1, 1, 0, 0, 0)
2
1
0
0
0
1
(1, 1, 1, 0, 0)
3
3
2
1
0
0
2
(2, 1, 1, 0, 0)
4
2
1
0
0
1
(1, 1, 1, 1, 0)
4
2
4
3
1
0
3
(1, 1, 1, 0) ( 3 , 1, 1, 1, 0)
2
9
2
9
4
3
2
1
0
2
(3, 1, 1, 1, 0)
6
3
3
2
1
0
1
(2, 2, 1, 1, 0)
6
3
2
1
0
2
(4, 2, 1, 1, 0)
8
4
2
1
0
1
(1, 1, 1, 1, 1)
5
5
2
5
3
5
4
1
4
( 34 , 1, 1, 1, 1)
16
3
8
3
16
9
4
3
1
3
(2, 1, 1, 1, 1)
6
3
2
4
3
1
2
(4, 1, 1, 1, 1)
8
4
2
4
3
1
1
( 32 , 23 , 1, 1, 1)
6
3
2
3
2
1
3
( 32 , 1, 1, 1) ( 9 , 3 , 1, 1, 1)
4 2
27
4
27
8
9
4
3
2
1
2
( 92 , 23 , 1, 1, 1)
9
9
2
9
4
3
2
1
1
(3, 3, 1, 1, 1)
9
9
2
3
3
2
1
2
(6, 3, 1, 1, 1)
12
6
3
3
2
1
1
(2, 2, 2, 1, 1)
8
4
8
3
2
1
3
(2, 2, 1, 1) (3, 2, 2, 1, 1)
9
9
2
3
2
1
2
(6, 2, 2, 1, 1)
12
6
3
2
1
1
(4, 4, 2, 1, 1)
12
6
4
2
1
2
(8, 4, 2, 1, 1)
16
8
4
2
1
1
(1, 1, 0, 0)
(2, 1, 1, 0)
(1, 1, 1, 1)
(3, 1, 1, 1)
(4, 2, 1, 1)
TABLE V
NON - REDUNDANT CONSTRAINTS FOR
L = 5.
| 7cs.IT
|
Efficient Exploration through Bayesian Deep Q-Networks
Kamyar Azizzadenesheli
1
Emma Brunskill 2 Animashree Anandkumar 3
arXiv:1802.04412v1 [cs.AI] 13 Feb 2018
Abstract
We propose Bayesian Deep Q-Network (BDQN),
a practical Thompson sampling based Reinforcement Learning (RL) Algorithm. Thompson sampling allows for targeted exploration in high dimensions through posterior sampling but is usually computationally expensive. We address this
limitation by introducing uncertainty only at the
output layer of the network through a Bayesian
Linear Regression (BLR) model. This layer can
be trained with fast closed-form updates and its
samples can be drawn efficiently through the
Gaussian distribution. We apply our method to
a wide range of Atari games in Arcade Learning
Environments. Since BDQN carries out more efficient exploration, it is able to reach higher rewards substantially faster than a key baseline, the
double deep Q network (DDQN).
1. Introduction
A central challenge in RL is how to design algorithms that
both scale to enormous or infinite state spaces, and efficiently balance exploration and exploitation in such environments. Much of the exciting advances in deep RL
that scale to enormous domains employ simple exploration
strategies such as ε-greedy, which are often highly inefficient. Though there is a large body of work on efficient
exploration relevant when the domain is small enough to
be represented with lookup tables for the value function,
there has been much less work on scaling exploration Several papers(Bellemare et al., 2016; Tang et al., 2016) that
do combine generalization and strategic exploration use
optimism-under-uncertainty, which involve explicit or implicit bonuses over rewards based on uncertainty over the
reward, dynamics or values.
An
alternative
to
(Brafman & Tennenholtz,
optimism-under-uncertainty
2003) is Thompson Sam-
1
University of California Irvine 2 Stanford University 3 Caltech,
Amazon. Correspondence to: Kamyar Azizzadenesheli <[email protected]>, Emma Brunskill <[email protected]>, Animashree Anandkumar <[email protected]>.
pling (TS) (Thompson, 1933). Thompson sampling is a
Bayesian approach which involves maintaining a prior
distribution over the environment models (reward and/or
dynamics), which is updated as observations are made
during the interaction with the environment. To choose
an action, a sample from the posterior belief is drawn and
an action is selected that maximizes the expected return
under the sampled belief. Interestingly, posterior sampling
for decision making has also been studied in the field of
psychology (Sanborn & Chater, 2016).
In the MDP setting this involves sampling a reward and
dynamics model and then performing MDP planning using the sampled models to compute an optimal action
for the current state (Strens, 2000; Osband et al., 2013).
Thompson sampling approaches have been observed to often empirically work significantly better than optimistic
approaches in contextual bandit settings(Chapelle & Li,
2011) and small MDPs(Osband et al., 2013) and still
maintains strong preserve state-of-the- art performance
bounds(Russo & Van Roy, 2014; Agrawal & Goyal, 2012;
Osband et al., 2013; Abbasi-Yadkori & Szepesvári, 2015).
In large MDPs, sampling a model and then performing
planning for that model is computationally intractable.
Therefore some form of function approximation is required to help scale the ideas of Thompson Sampling.
To help address this, (Osband et al., 2014) introduced randomized least-squares value iteration (RLSVI). RLSVI involves combining linear value function approximation with
Bayesian regression in order to be able to sample the value
function weights from a distribution. The authors prove
strong regret bounds for RLSVI when a tabular basis function set is used, but RLSVI is not scalable to large-scale
RL with deep neural networks. To try to combine the benefits of Thompson sampling style approaches with deep networks for generalization and scale, Osband et al. (2016) introduced a bootstrapped-ensemble approach that trains several models in parallel to approximate the posterior distribution. Other works suggest using a posterior over the parameters of each node in the network and employ a variational approximation (Lipton et al., 2016b) or use noisy
network (Fortunato et al., 2017). However, mostly these
approaches have lead to modest gains on the Atari benchmarks, not equaling some of the substantial benefits over
by combining optimism-under-uncertainty with deep neu-
Efficient Exploration through Bayesian Deep Q-Networks
ral networks(Bellemare et al., 2016).
Surprizingly in this paper we show that a simple approach
that extends the randomized least-squares value iteration
method (Osband et al., 2014) to deep neural networks can
yield substantial gains on Atari benchmarks. Specifically,
we combine a deep neural network with Bayesian linear
regression at the last layer of the network. Our work
is also related to a concurrently developed approach by
Levine et al. (2017) who perform least squares temporal
difference learning on top of a deep neural network, uses
ε-greedy exploration on top of the learned Q function, and
also demonstrate modest gains on 5 Atari benchmarks. Our
results show that performing Bayesian regression instead,
and sampling from the result, can yield a substantial benefit, indicating that it is not just the higher data efficiency
at the last layer, but that leveraging an explicit uncertainty
representation over the value function is of substantial benefit.
More specifically we introduce Bayesian deep Qnetworks (BDQN) which combines a Deep Q network
(DQN) (Mnih et al., 2013) with a Bayesian linear regression model on the last layer. Due to linearity and by choosing a Gaussian prior, we derive a closed-form analytical
update to the approximated posterior distribution over Q
functions. We can also draw samples efficiently from the
Gaussian distribution. Exploration is performed by sampling from the learned Gaussian posterior to instantiate the
Q values, and then the best action is selected. We test
BDQN on a wide range of Arcade Learning Environment
(Bellemare et al., 2013; Machado et al., 2017) Atari games,
and compare our results to our own implementation of
DDQN (Van Hasselt et al., 2016). BDQN and DDQN share
the same architecture, and follow same target objective,
and differ only in the way they are used to select actions:
DDQN uses ε-greedy and BDQN performs Bayesian linear
regression on the last layer, samples the parameters from
the resulting distributions, and selects the best action for
that sample. We also compare our results to the reported results from a number of the state-of-the-art approaches. Our
proposed approach has several benefits– simplicity and targeted exploration– and yields performance often substantially better than existing optimism-based and other state
of the art deep RL approaches.
2. Related Work
The complexity of the exploration-exploitation
trade-off has been deeply investigated in RL literature (Kearns & Singh, 2002; Brafman & Tennenholtz,
2003; Asmuth et al., 2009). Jaksch et al. (2010) investigates the regret analysis of MDPs where Optimism in Face
of Uncertainty (OFU) principle is deployed to guarantee a
high probability regret upper bound. Azizzadenesheli et al.
Table 1. We run both BDQN and DDQN for the same number of times steps, written in the last column. The first column
presents the score ratio of BDQN to DDQN after the steps provided in the last column. The second column is the score ratio of
BDQN after the number of steps in the last column compared to
the score of DDQN† , which is the reported scores of DDQN in
Van Hasselt et al. (2016) after running for 200M samples during
evaluation time where the ε = 0.001, and the third column is with
respect to Human score reported at Mnih et al. (2015). It is worth
noting that we do not design a evaluation phase for BDQN
BDQN
DDQN
BDQN
DDQN†
BDQN
H UMAN
Steps
Amidar
558%
103%
Alien
Assault
396%
2517%
Asteroids
Asterix
531%
207%
BeamRider
281%
BattleZone
Atlantis
80604%
292%
DemonAttack
114%
Centipede
BankHeist
211%
148%
CrazyClimber
ChopperCommand 14500%
295%
Enduro
112%
Pong
788%
103%
176%
1516%
385%
114%
253%
49413%
114%
178%
100%
122%
1576%
350%
100%
325%
43%
589%
108%
687%
150%
172%
11172%
326%
61%
100%
350%
732%
361%
226%
100M
100M
100M
100M
100M
70M
50M
40M
40M
40M
40M
40M
40M
30M
5M
Game
(2016a) deploys OFU in order to propose the high probability regret upper bound for Partially Observable MDPs
(POMDPs) using spectral methods (Anandkumar et al.,
2014). Furthermore, Bartók et al. (2014) tackles a general
case of partial monitoring games and provides minimax
regret guarantee which is polynomial in certain dimensions
of the problem.
In multi-arm bandit, there are compelling empirical pieces
of evidence that Thompson Sampling can provide better results than optimism-under-uncertainty approaches
(Chapelle & Li, 2011), while the state of the art performance bounds are preserved (Russo & Van Roy, 2014;
Agrawal & Goyal, 2012). A natural adaptation of this algorithm to RL, posterior sampling RL (PSRL), first proposed by Strens (2000) also shown to have good frequentist and Bayesian performance guarantees (Osband et al.,
2013; Abbasi-Yadkori & Szepesvári, 2015). Even though
the theoretical RL addresses the exploration and exploitation trade-offs, these problems are still prominent in empirical reinforcement learning research (Mnih et al., 2015;
Abel et al., 2016; Azizzadenesheli et al., 2016b). On the
empirical side, the recent success in the video games has
sparked a flurry of research interest. Following the success of Deep RL on Atari games (Mnih et al., 2015) and
the board game Go (Silver et al., 2017), many researchers
have begun exploring practical applications of deep re-
Efficient Exploration through Bayesian Deep Q-Networks
inforcement learning (DRL). Some investigated applications include, robotics (Levine et al., 2016), self-driving
cars (Shalev-Shwartz et al., 2016), and safety (Lipton et al.,
2016a).
much beyond modest gains on Atari games while BDQN
provides significant improvements in terms of both sample
complexity and final performance.
Inevitably for PSRL, the act of posterior sampling for policy or value is computationally intractable with large systems, so PSRL can not be easily leveraged to high dimensional problems. To remedy these failings Osband et al.
(2017) consider the use of randomized value functions to
approximate posterior samples for the value function in
a computationally efficient manner. They show that with
a suitable linear value function approximation, using the
approximated Bayesian linear regression for randomized
least-squares value iteration method can remain statistically
efficient (Osband et al., 2014) but still is not scalable to
large-scale RL with deep neural networks.
3. Thompson Sampling vs ε−greedy
To combat these shortcomings, Osband et al. (2016) suggests a bootstrapped-ensemble approach that trains several
models in parallel to approximate the posterior distribution. Other works suggest using a variational approximation to the Q-networks (Lipton et al., 2016b) or noisy network (Fortunato et al., 2017). However, most of these approaches significantly increase the computational cost of
DQN and neither approach produced much beyond modest gains on Atari games. Interestingly, Bayesian approach
as a technique for learning a neural network has been
deployed for object recognition and image caption generation where its significant advantage has been verified
Snoek et al. (2015).
In this work we present another alternative approach that
extends randomized least-squares value iteration method
(Osband et al., 2014) to deep neural networks: we approximate the posterior by a Bayesian linear regression only on
the last layer of the neural network. This approach has
several benefits, e.g. simplicity, robustness, targeted exploration, and most importantly, we find that this method
is much more effective than any of these predecessors in
terms of sample complexity and final performance.
Concurrently, Levine et al. (2017) proposes least squares
temporal difference which learns a linear model on the
feature representation in order to estimate the Q-function
while ε-greedy exploration is employed and improvement
on 5 tested Atari games is provided. Out of these 5 games,
one is common with our set of 15 games which BDQN outperform it by factor of 360% (w.r.t. the score reported
in their paper). Drop-out, as another randomized exploration method is proposed by Gal & Ghahramani (2016)
but Osband et al. (2016) investigates the sufficiency of the
estimated uncertainty and hardness in driving a suitable exploitation out of it. As stated before, in spite of the novelties proposed by the methods, mentioned in this section,
neither of them, including TS based approaches, produced
In this section, we enumerate a few benefits of TS over εgreedy strategies. We show how TS strategies exploit the
uncertainties and expected returns to design a randomized
exploration while ε−greedy strategies disregard all these
useful information for the exploration.
In order to make a balance between exploration and exploitation, TS explores actions with higher estimated return
with higher probability. In order to exploit the estimated
uncertainties, TS dedicates a higher chance to explore an
action if its uncertainty increases. Fig. 1(a) expresses the
agent’s estimated values and uncertainties for the available
actions at a given state x. While ε−greedy strategy mostly
focuses on the greedy action, action 1, the TS based strategy randomizes, mostly, over actions 1 through 4, utilizes
their approximated expected returns and uncertainties, and
with low frequency explores actions 5, 6. On the other hand,
ε−greedy strategy explores actions 5 and 6, the actions that
the RL agent is almost sure about their low expected returns, as frequent as other sub-greedy actions 2, 3, 4 which
increases its samples complexity. Moreover, a ε−greedy
strategy requires a deep network to approximate the value
of all the sub-greedy actions equally good, therefore, it dedicates the network capacity to accurately estimate the values of all the sub-greedy actions equally good, instead of focusing more on the actions with higher promising estimated
value. Therefore, it ends up with not accurate enough estimation of other good actions compared to the greedy action.
In a study of value-based deep RL, e.g. DQN, the network
is following a target value which is updated occasionally.
Therefore, TS based strategy should not estimate the posterior distribution which adaptively follows the target values.
A commonly used technique in deep RL is a moving window of replay buffer to store the recent experiences. The
TS based agent, after a few tries of actions 5 and 6, builds a
belief in the low return of these actions given the current target values, while it is possible that later on, the target value
suggests a high expected return of these actions. Since the
replay buffer is bounded moving window, lack of samples
of these actions pushes the posterior belief of these actions
to the prior belief, over time, and the agent tries them again
in order to update its belief. Fig. 1(b) shows that the lack of
samples for action 6 in the replay buffer, increases the uncertainty of this action and a randomized TS strategy starts
to explore them over. It means that due to adaptive change
of target value, respectively the objective, and limited replay buffer, the BDQN agent is never too confident about
Efficient Exploration through Bayesian Deep Q-Networks
(c)
(a)
(b)
Figure 1. A cartoon on TS vs ε-greedy. The red crosses are the target values, the diamonds are the mean of estimated Q-values with
blue intervals as uncertainties (e.g. c · variance). (a) ε-greedy strategy mostly chooses the greedy action, action 1 and explore actions
5, 6 as much as actions 2, 3, 4, while TS randomizes mostly over actions 1, 2, 3, 4 barely explore actions 5, 6. (b) BDQN computes the
posterior using recent experiences in the replay buffer. Therefore, lack of samples of action 6 increases the uncertainty on the estimated
value and BDQN explores it again. This further exploration is crucial since the target value changes over time. (c) maze.
the expected return of poor actions and keeps exploring
them once in a while.
In general, TS based strategy advances the explorationexploitation balance by making a trade-off between the expected returns and the uncertainties, while ε−greedy strategy ignores all of this information.
Another benefit of TS over ε-greedy can be described using Fig. 1(c). Consider a deterministic and episodic maze
game, with episode length H of the shortest pass from the
start to the destination. The agent is placed to the start point
at the beginning of each episode where the goal state is to
reach the destination and receive a reward of 1 otherwise
the reward is 0. Consider an agent, which is given a set
of Q-functions where the true Q-function is within the set
and is the most optimistic function in the set. The agent
is supposed to find a Q from this set which maximizes the
average return. It is worth noting that the agent task is to
find a good function from a function set.
In this situation, TS randomizes over the Q-functions with
high promising returns and relatively high uncertainty, including the true Q-function. When the TS agent picks the
true Q-function, it increases the posterior probability of this
Q-function because it matches the observation. When the
TS agent chooses other functions, they predict deterministically wrong values and the posterior update of those
functions set to zero. Therefore, the agent will not choose
these functions again, i.e. TS finds the true Q-function
by transferring the information through posterior update
which helps the agent to find the optimal Q very fast. For
ε-greedy agent, even though it chooses the true function at
the beginning (it is the optimistic one), at each time step, it
randomizes its action with the probability ε. Therefore, it
takes exponentially many trials in order to get to the target
in this game.
4. Preliminaries
An infinite horizon γ-discounted MDP M is a tuple
hX , A, T, R, γi, with state space X , action space A, and
the transition kernel T , accompanied with reward function
of R where 0 < γ ≤ 1. At each time step t, the environment is at a state xt , called current state, where the
agent needs to make a decision at under its policy. Given
the current state and action, the environment stochastically
proceed to a successor state xt+1 under probability distribution T (Xt+1 |xt , at ) := P(Xt+1 |xt , at ) and provides a
stochastic reward rt with mean of E[r|x = xt , a = at ] =
R(xt , at ). The agent objective is to optimize the overall expected discounted reward over its policy π := X → A,
a stochastic mapping from states to actions, π(a|x) :=
P(a|x).
∗
∗
η = η(π ) = max η(π) = max lim Eπ
π
π
N →∞
"
N
X
t
γ rt
t=0
#
(1)
The expectation in Eq. 1 is with respect to the randomness
in the distribution of initial state, transition probabilities,
stochastic rewards, and policy, under stationary distribution, where η ∗ , π ∗ are optimal average return and optimal
policy, respectively. Let Qπ (x, a) denote the average discounted reward under policy π starting off from state x and
taking action a in the first place.
Qπ (x, a) := lim Eπ
N →∞
"N
X
t=0
t
γ rt |x0 = x, a0 = a
#
For a given policy π and Markovian assumption of the
model, we can rewrite the equation for the Q functions as
Efficient Exploration through Bayesian Deep Q-Networks
follows:
Qπ (x, a) = R(x, a) + γ
X
T (x′ |x, a)π(a′ |x′ )Qπ (x′ , a′ )
x′ ,a′
(2)
To find the optimal policy, one can solve a linear programming problem in Eq. 1 or follow the corresponding Bellman equation Eq. 2 where both of the optimization methods solve the following
X
Q∗ (x, a) = R(x, a) + γ
T (x′ |x, a) max
Q∗ (x′ , a′ )
′
x′
a
∀x, a, Q∗ (x, a) = Qπ∗ (x, a) and the optimal policy is
a deterministic mapping from state to actions in A, i.e.
x → arg maxa Q∗ (x, a). In RL, we do not know the transition kernel and the reward function in advance, therefore,
we cannot solve the posed Bellman equation directly. In order to tackle this problem, the property of minimizing the
Bellman residual of a given Q-function
h
i
2
L(Q) = Eπ (Q(x, a) − r − γQ(x′ , a′ ))
(3)
has been proposed (Lagoudakis & Parr, 2003; Antos et al.,
2008). Here, the tuple (x, a, r, x′ ) consists of consecutive samples under behavioral policy π. Furthermore,
(Mnih et al., 2015) carries the same idea, and introduce
Deep Q-Network (DQN) where the Q-functions are parameterized by a deep network. To improve the quality of Q
estimate, they use back propagation on loss L(Q) using
the TD update (Sutton & Barto, 1998). In the following
we describe the setting used in DDQN. In order to reduce
the bias of the estimator, they introduce target network
Qtarget and target value y = r + γQtarget (x′ , â) where
â = arg maxa′ Q(x′ , a′ ) with a new loss L(Q, Qtarget )
i
h
2
(4)
L(Q, Qtarget ) = Eπ (Q(x, a) − y)
This regression problem minimizes the estimated loss
b
L(Q,
Qtarget ), which minimize the distance between the
Q and the target y. A DDQN agent, once in a while updates
the Qtarget network by setting it to Q network, pursues the
regression with the new target value and provides a biased
estimator of the target.
5. Bayesian Deep Q-Networks
We now show how we can extend the randomized leastsquares value iteration method (Osband et al., 2014) to
combine it with a deep neural network. The result can be
viewed as a coarse approximation to representing the uncertainty over the Q-function, which we use to guide exploration.
We utilize the DQN architecture, remove its last layer,
and directly build a Bayesian linear regression (BLR)
Algorithm 1 BDQN
1: Initialize parameter sets θ, θtarget , W , W target , and
Cov using a normal distribution.
2: Initialize replay buffer and set counter = 0
3: for episode = 1 to inf do
4:
Initialize x1 to the initial state of the environment
5:
for t = to the end of episode do
6:
if count mod T sample = 0 then
7:
sample W ∼ N (W target , Cov)
8:
end if
9:
Select action at = argmaxa′ W ⊤ φθ (xt ) a′
10:
Execute action at in environment, observing reward rt and successor state xt+1
11:
Store transition (xt , at , rt , xt+1 ) in replay buffer
12:
Sample a random minibatch of transitions (xτ ,
aτ , rτ ,xτ +1 ) from replay buffer
for terminal xτ +1 :
rτ
for
non-terminal
xτ +1 :
13:
yτ ←
h
i
target ⊤ target
r
+
W
φ
(x
)
where
τ
τ
+1
θ
â
⊤
â = argmaxa′ W φθ (xτ +1 ) a′
14:
θ ← θ − η · ∇θ (yτ − W ⊤ φθ (xτ ) aτ )2
15:
if count mod T target = 0 then
16:
set θtarget ← θ
17:
end if
18:
if count mod T Bayes target = 0 then
19:
Update W target and Cov
20:
end if
21:
count = count + 1
22:
end for
23: end for
(Rasmussen & Williams, 2006) on the output of the deep
network φθ (x) ∈ Rd , the feature representation layer,
parametrized by θ. We use BLR to efficiently approximate
the distribution over the Q-values where the uncertainty
over the values is captured. A common assumption in DNN
is that the feature representation is suitable for linear classification or regression (same assumption in DQN), therefore,
therefore building a linear model on the features a suitable
choice (as was done recently in (Levine et al., 2017)).
The Q-functions are approximated as a linear transformation of the deep neural network features, i.e. for a given
pair of state-action, Q(x, a) = φθ (x)⊤ wa , where wa ∈
Rd , ∀a ∈ A. Consequently, as mentioned in the previous
section, the target value is generated using target model.
The target model follows the same structure as the Q model,
and contains φtarget
(x) ∈ Rd , ∀x ∈ X denotes the feature
θ
representation of target network, and wtarget â , ∀â ∈ A denotes the target linear model applied on the target feature
representation. Inspired by DDQN, for a given tuple of experience (x, a, r, x′ ), the predicted value of pair (x, a) is
Efficient Exploration through Bayesian Deep Q-Networks
Q(x, a) = φθ (x)⊤ wa , while the target value is
y = r + γφtarget (x′ )wtarget â , ; â = argmaxa φθ (x′ )⊤ wa .
Therefore, by deploying BLR on the space of features, we
can approximate the posterior distribution of model parameter wa ∈ Rd , ∀a ∈ A, as well as the posterior distribution of the Q-functions using the corresponding target values. In Gaussian BLR models, in order to make the posterior update computationally tractable in a closed form a
common approximation is to make the prior and likelihood
choices as conjugates of each other. Therefore, for a given
pair of (x, a), the vector wa is drawn from a Gaussian prior
N (0, σ 2 ) and given wa , the target value is generated from
the following model;
y ∼ wa⊤ φ(x) + ǫ
where ǫ ∼ N (0, σǫ2 ) is an iid noise.
Therefore,
y|x, a, wa ∼ N (φ(x)⊤ wa , σǫ2 ). Moreover,
the
distribution
R
of the target value y is P (y| a) = wa P (y|wa ) P (wa ) dwa
which also has a closed form.
Given a experience replay buffer D = {xτ , aτ , yτ }D
τ =1 ,
we construct |A| (number of actions) disjoint datasets for
each action, Da , where D = ∪a∈A Da and Da is a set of
tuples xτ , aτ , yτ with the action aτ = a and cardinality
Da . We are interested in the approximated posterior distribution of wa , ∀a ∈ A and correspondingly the Q(x, a);
P(wa |Da ), ∀a ∈ A and P(Q(x, a)|Da ), ∀x ∈ X .
The following are the standard Bayesian linear regression
equations adjusted for our setting and we encourage readers
who are familiar with Bayesian linear regression to skip the
derivation. For each action a and the corresponding dataset
Da , we construct a matrix Φa ∈ Rd×Da , a concatenation
Da
a
,a
of feature column vectors {φ(xi )}D
i=1 , and ya ∈ R
concatenation of target values in set Da . Therefore the posterior distribution of wa is as follows:
wa ∼ N
−1
1
1
1
⊤
,
Ξ
=
Ξ
Φ
y
,
Ξ
Φ
Φ
+
I
a
a a a
a
a a
σǫ2
σǫ2
σ2
(5)
and I ∈ Rd is an identity matrix. The Q(x, a)|Da =
wa⊤ φ(x) where wa is drawn following the posterior distribution in Eq. 5. Since the prior and likelihood are conjugate
of each other we have the closed
form posterior distribution
P
t
of the discounted return, N
γ
rt |x0 = x, a0 = a, Da ,
t=0
is approximated as
N
1
⊤
⊤
φ(x) Ξa Φa ya , φ(x) Ξa φ(x)
σǫ2
(6)
As TS suggests, for the exploration, we exploit the expression in Eq. 5.At the decision time, we sample a wight vector wa for each action in order to have samples of Q-values.
Then we act optimally with respect to these sampled value
aTS = arg max wa⊤ φθ (x).
a
|A|
(7)
|A|
Let W = {wa }a=1 , respectively W target = {watarget }a=1 ,
|A|
and Cov = {Ξa }a=1 . In BDQN, the agent interacts with
the environment through applying the actions proposed by
TS , i.e. aTS . We utilize a notion of experience replay
buffer where the agent stores its recent experiences. The
agent draws W ∼ N (W target , Cov) (abbreviation for
sampling of vector wa for each action separately) every
T sample steps and act optimally with respect to the drawn
weights. During the inner loop of the algorithm, we draw a
minibatch of data from replay buffer and use loss
(yτ − W ⊤ φθ (xτ ) aτ )2
(8)
where
h
i
⊤
yτ :=rτ + W target φθtarget (xτ +1 )
â
⊤
′
(9)
â :=argmaxa W φθ (xτ +1 ) a′
and
⊤update the
weights of network: θ ← θ − η · ∇θ (yτ −
W φθ (xτ ) aτ )2 .
We update the target network every T target steps and set
θtarget to θ. With the period of T Bayes target the agent
updates its posterior distribution using a larger minibatch
of data drawn from replay buffer, set the watarget , ∀a ∈
A to the mean of the posterior, and sample wa , ∀A with
respect to the updated posterior. Algorithm 1 gives the full
description of BDQN.
6. Experiments
We apply BDQN on a variety of Atari games using
the Arcade Learning Environment (Bellemare et al., 2013)
through OpenAI Gym1 (Brockman et al., 2016). As a baseline, we run the DDQN algorithm and evaluate BDQN on
the measures of sample complexity and score. Furthermore,
all the implementations are coded in MXNet framework
(Chen et al., 2015).
Network architecture: The input to the network part of
BDQN is 4 × 84 × 84 tensor with a rescaled and averaged
over channels of the last four observations. The first convolution layer has 32 filters of size 8 with a stride of 4. The
second convolution layer has 64 filters of size 4 with stride
2. The last convolution layer has 64 filters of size 3 followed by a fully connected layer with size 512. We add a
BLR layer on top of this.
1
Each input frame is a pixel-max of the two consecutive
frames. We detailed the environment setting in the implementa-
Efficient Exploration through Bayesian Deep Q-Networks
0.6
2.5
2.0
1.5
1.0
0.5
0.0
0
1
2
3
Number of steps
4
5
4
3
2
BDQN
DDQN
1
1
2
3
Number of steps
4
1e7
1.0
1e8
2.0
1.5
1.0
0.5
0.2
0.4
0.6
0.8
r a zyl m b er
0.8
0.6
0.4
BDQN
DDQN
0.2
3
N um ber of st eps
4
0.6
Number of steps
0.8
4
3
0
4
3
2
1
0
0.2
0.4
0.6
0.8
E d u r o
0.8
0.6
0.4
0.2
0.0
0.5
1.0
1.5
2.0
2.5
3.0
Number of steps
4
1
0
5
1e6
BDQN
DDQN
0.6
0.4
0.2
0.0
0.0
0.5
1.0
1.5
2.0
2.5
3.0
0.6
0.4
0.2
0.0
1.5
2.0
2.5
N um ber of st eps
3.0
3.5
4.0
1e7
4
5
1e7
BDQN
DDQN
6
5
4
3
2
0.0
BDQN
DDQN
3
7
0.5
1.0
1.5
2.0
2.5
3.0
3.5
1e7
N um ber of st eps
Beam R der
1e4
0.8
1.0
2
Number of steps
C e t pede
8
3.5
h opper om m a d
0.5
1
1e7
N um ber of st eps
1.0
0
1e3
0.8
0.0
1e7
N um ber of st eps
3
1.0
1e5
BDQN
DDQN
1.0
2
Dem o n At t ack
1.2
1.0
1e8
N um ber of st eps
1.2
1
1e4
i
2
BDQN
DDQN
1.0
1e8
5
0.0
1e7
5
0
Ast er ix
0.0
1.0
2
0.4
1e3
1.2
1
0.2
BDQN
DDQN
6
1.0
1e8
N um ber of st eps
0
0.0
1e4
2.5
0.0
6
0
Number of episode
0.8
Alien
1e5
7
0.6
BDQN
DDQN
3.0
1e7
BankHeist
1e2
0.4
1e3
BDQN
DDQN
3.0
0.2
10
Aver age Rew ar d per ep sode
Atlantis
1e6
0.0
BDQN
DDQN
6
Aver age Rew ar d per ep sode
1.0
1e8
BattleZone
1e4
7
−20
0.0
Aver age Rew ar d per ep sode
Number of steps
0.8
0.2
Aver age Rew ar d per ep sode
0.6
20
−10
0.4
Aver age Rew ar d per episode
0.4
Pong
Average Reward per episode
0.8
Aver age Rew ar d per ep sode
0.2
Aver age Rew ar d per episode
0.0
0.0
Average Reward per episode
Average Reward per step
0
Average Reward per episode
1.0
0.2
1
BDQN
DDQN
1.2
0.4
2
Asteroids
1e4
1.4
0.6
3
Average Reward per episode
1.6
BDQN
DDQN
0.8
4
0
Assault
1e4
BDQN
DDQN
Average Reward per episode
Amidar
1e3
5
Aver age Rew ar d per ep sode
Average Reward per episode
6
1.0
0.8
0.6
0.4
BDQN
DDQN
0.2
0.0
0
1
2
3
4
N um ber of st eps
5
6
1e7
Figure 2. The efficient and targeted exploration of BDQN
Choice of hyper-parameters: For BDQN, we set the values of W target to the mean of the posterior distribution
over the weights of BLR with covariances Cov and draw
W from this posterior. For the fixed W and W target ,
we randomly initialize the parameters of network part of
BDQN, θ, and train it using RMSProp, with learning rate of
0.0025, and a momentum of 0.95, inspired by (Mnih et al.,
2015) where the discount factor is γ = 0.99, the number of steps between target updates T target = 10k steps,
and weights W are re-sampled from their posterior distribution every T sample steps. We update the network part of
BDQN every 4 steps by uniformly at random sampling a
mini-batch of size 32 samples from the replay buffer. We
update the posterior distribution of the weight set W every
T Bayes target using mini-batch of size B (if the size of replay buffer is less than B at the current step, we choose the
minimum of these two ), with entries sampled uniformly
form replay buffer. The experience replay contains the
1M most recent transitions. Further hyper-parameters are
equivalent to ones in DQN setting.
Furthermore, for the BLR part of BDQN, we have noise
variance σǫ , variance of prior over weights σ, sample size
B, posterior update period T Bayes target , and the posterior sampling period T sample . To optimize for this set of
hyper-parameters we set up a very simple, fast, cheap, and
non-exhaustive hyper-parameter tuning procedure using a
pre-trained DQN model for the game of Assault. The simplicity and cheapness of our hyper parameter tuning proves
the robustness and superiority of BDQN where the exhaustive hyper-parameter search is likely to provide even better
performance. The details of hyper parameters tuning is provided in Apx. A.
tion code
In order to compare the fairness in sample usage, we argue in Apx. A, that the network part of BDQN and its
corresponding part in DDQN observe the same number of
samples but the BLR part of BDQN uses 16 times fewer
samples compared to its corresponding last layer in DDQN,
Apx. A.
Baselines: We implemented DDQN and fix its architecture to match our BDQN implementation. We also aimed
to implement a couple other deep RL methods that employ
strategic exploration.
Unfortunately we encountered
several implementation challenges. To try to illustrate the
performance of our approach we instead extracted the best
reported results from a number of state-of-the-art deep RL
methods and include them in Table 2. Note that this is not
a perfect comparison, as sometimes there can be additional
details that are not included in the papers that mean that it
is hard to compare the reported results (an issue that has
been discussed extensively recently, e.g. (Henderson et al.,
2017) ).5 Here we tried to report final performance
(when those were reported or identifiable from plots). We
report results from bootstrapped DQN(Osband et al.,
2016),
count
based
exploration(Bellemare et al.,
2016), the Pixel and Reactor results that build on the
count based exploration(Ostrovski et al., 2017), and
NoisyNet(Fortunato et al., 2017).
5
To further reproducibility, we released our codes and trained
models https://github.com/kazizzad/BDQN-MxNet-Gluon. Since
DRL experiments are expensive, we also have released the
recorded arrays of returns, in order to make it possible for others to compare against BDQN, without running the experiments
again. Moreover, our Bootstrap DQN implementation is also
available https://github.com/kazizzad/Bootstrap-DQN
Efficient Exploration through Bayesian Deep Q-Networks
Table 2. Comparison of scores and sample complexities(scores in the first two columns are average of 100 consecutive episodes). The
sample complexity, SC, represents the number of samples the BDQN requires to bit the human score (Mnih et al., 2015)(“ − ” means
BDQN could not bit) and SC† is the number number of samples the BDQN requires to bit the score of DDQN † . The scores of DDQN
†
are borrowed from the original DDQN paper Van Hasselt et al. (2016) which are reported after running for 200M samples during
evaluation time where the ε = 0.001. Same for bootstrap DQN (Osband et al., 2016), CTS, Pixel, Reactor (Ostrovski et al., 2017)
where the scores are borrowed from their original papers. For NoisyNet (Fortunato et al., 2017) since there is no score for noisy DDQN
we reported the score of its closest model noisyDQN scores. The Human scores are borrowed from Mnih et al. (2015). It is worth noting
that we do not design a evaluation phase for BDQN.
Game
BDQN DDQN DDQN† Bootstrap2 NoisyNet3
Amidar
Alien
Assault
Asteroids
Asterix
BeamRider
BattleZone
Atlantis
DemonAttack
Centipede
BankHeist
CrazyClimber
ChopCmd4
Enduro
Pong
5.52k
3k
8.84k
14.1k
58.4k
8.7k
65.2k
3.24M
11.1k
7.3k
0.72k
124k
72.5k
1.12k
21
0.99k
2.9k
2.23k
0.56k
11k
4.2k
23.2k
39.7k
3.8k
6.4k
0.34k
84k
0.5k
0.38k
18.82
0.7k
2.9k
5.02k
0.93k
15.15k
7.6k
24.7k
64.76k
9.7k
4.1k
0.72k
102k
4.6k
0.32k
21
1.27k
2.44k
8.05k
1.03k
19.7k
23.4k
36.7k
99.4k
82.6k
4.55k
1.21k
138k
4.1k
1.59k
20.9
1.5k
2.9k
3.1k
2.1k
11.0
14.7k
11.9k
7.9k
26.7k
3.35k
0.64k
121k
5.3k
0.91k
21
Results: The results are provided in Fig. 2 and Table. 2.
BDQN performs best across the majority of games at the
stated number of samples, even typically performing much
better than several other methods when they are trained
for much longer. Note that comparisons to Bootstrap,
NoisyNet and CTS should be viewed lightly, since the reported results for those algorithms were generally when
trained for substantially longer (often 100-200M steps).
Reactor(Ostrovski et al., 2017) outperformed our BDQN
on three games, Alien, Atlantis and Enduro, when trained
for an identical number of time steps.
Note also that BDQN outperforms the optimism based approaches (Bellemare et al., 2016; Ostrovski et al., 2017) on
all other games we tried including Amidar, which they classify as one of the harder exploration games(Ostrovski et al.,
2017).
It is worth noting that the scores of DDQN are reported during the leaning phase (not evaluation phase). For example,
DDQN gives score of 18.82 during the learning phase, but
setting ε to zero, it mostly gives the score of 21. We also
report the number of samples (sample complexity (SC)) it
take from BDQN to reach human scores, SC, and DDQN †
scores, SC† , Apx. A.
For the game Atlantis, DDQN† gives score of 64.67k after 200M samples during evaluation time, while BDQN
reaches 3.24M after 40M samples. As it is been shown
in Fig. 2, BDQN saturates for Atlantis after 20M samples.
CTS
1.03k
1.9k
2.88k
3.95k
9.55k
7.0k
7.97k
1.8M
39.3k
5.4k
1.3k
112.9k
5.1k
0.69k
20.8
Pixel Reactor Human
0.62k
1.7k
1.25k
0.9k
1.4k
3k
10k
40k
1.3k
1.8k
0.42k
75k
2.5k
0.19k
17
1.18k
3.5k
3.5k
1.75k
6.2k
3.8k
45k
9.5M
7k
3.5k
1.1k
119k
4.8k
2.49k
20
1.7k
6.9k
1.5k
13.1k
8.5k
5.8k
38k
29k
3.4k
12k
0.72k
35.4k
9.9k
0.31k
9.3
SC
SC†
Step
22.9M 4.4M 100M
36.27M 100M
1.6M 24.3M 100M
58.2M 9.7M 100M
3.6M
5.7M 100M
4.0M
8.1M
70M
25.1M 14.9M 50M
3.3M
5.1M
40M
2.0M 19.9M 40M
4.2M
40M
2.1M 10.1M 40M
0.12M 2.1M
40M
4.4M
2.2M
40M
0.82M 0.8M
30M
1.2M
2.4M
5M
We realized that BDQN reaches the internal OpenAIGym
limit of max episode, where relaxing it improves score after 15M steps to 62M .
We observe that BDQN immediately learns significantly
better policies due to its targeted exploration in a much
shorter period of time. Since BDQN on game Atlantis
promise a big jump around time step 20M , we ran it five
more times in order to make sure it was not just a coincidence Apx. A Fig. 5. For the game Pong, we ran the experiment for a longer period but just plotted the beginning of it
in order to observe the difference. For some games, we did
not run the experiment to 100M samples since the reached
their plateau.
7. Conclusion
In this work we proposed BDQN, a practical TS based RL
algorithm which provides targeted exploration in a computationally efficient manner. It involved making simple modifications to the DDQN architecture by replacing the last
layer with a Bayesian linear regression. Under the Gaussian prior, we obtained fast closed-form updates for the
posterior distribution. We demonstrated significantly faster
training and much better performance in many games compared to the reported results of a wide number of state-ofthe-art baselines. Due to computational limitations we did
not try the algorithm on all games and it remains an interesting issue to further explore its performance, and combine it
Efficient Exploration through Bayesian Deep Q-Networks
with other advances in deep RL that can be easily extended.
In this work, for BDQN we randomize the last layer of the
model and use a Bayesian linear regression framework to
train it, and alternate it with training the other layers of the
network. An alternative approach is to train it end-to-end
using stochastic optimization approaches (Welling & Teh,
2011). This could significantly speed up training while retaining the computational efficiency of DDQN. In this work,
we have considered value based approaches in deep RL and
we plan to explore the advantages of TS based exploration
in policy gradient based approaches in future.
References
Abbasi-Yadkori, Yasin and Szepesvári, Csaba. Bayesian
optimal control of smoothly parameterized systems. In
UAI, pp. 1–11, 2015.
Abel, David, Agarwal, Alekh, Diaz, Fernando, Krishnamurthy, Akshay, and Schapire, Robert E. Exploratory
gradient boosting for reinforcement learning in complex
domains. arXiv, 2016.
Agrawal, Shipra and Goyal, Navin. Analysis of thompson
sampling for the multi-armed bandit problem. In COLT,
2012.
Anandkumar, Animashree, Ge, Rong, Hsu, Daniel,
Kakade, Sham M, and Telgarsky, Matus. Tensor decompositions for learning latent variable models. The Journal of Machine Learning Research, 15(1):2773–2832,
2014.
Antos, András, Szepesvári, Csaba, and Munos, Rémi.
Learning near-optimal policies with bellman-residual
minimization based fitted policy iteration and a single
sample path. Machine Learning, 2008.
Asmuth, John, Li, Lihong, Littman, Michael L, Nouri, Ali,
and Wingate, David. A bayesian sampling approach to
exploration in reinforcement learning. In Proceedings of
the Twenty-Fifth Conference on Uncertainty in Artificial
Intelligence, 2009.
Azizzadenesheli, Kamyar, Lazaric, Alessandro, and
Anandkumar, Animashree. Reinforcement learning of
pomdps using spectral methods. In Proceedings of the
29th Annual Conference on Learning Theory (COLT),
2016a.
Azizzadenesheli, Kamyar, Lazaric, Alessandro, and
Anandkumar, Animashree. Reinforcement learning in
rich-observation mdps using spectral methods. arXiv
preprint arXiv:1611.03907, 2016b.
Bartók, Gábor, Foster, Dean P, Pál, Dávid, Rakhlin, Alexander, and Szepesvári, Csaba. Partial monitoringclassification, regret bounds, and algorithms. Mathematics of
Operations Research, 2014.
Bellemare, Marc, Srinivasan, Sriram, Ostrovski, Georg,
Schaul, Tom, Saxton, David, and Munos, Remi. Unifying count-based exploration and intrinsic motivation. In
Advances in Neural Information Processing Systems, pp.
1471–1479, 2016.
Bellemare, Marc G, Naddaf, Yavar, Veness, Joel, and Bowling, Michael. The arcade learning environment: An
evaluation platform for general agents. J. Artif. Intell.
Res.(JAIR), 2013.
Brafman, Ronen I and Tennenholtz, Moshe. R-max-a general polynomial time algorithm for near-optimal reinforcement learning. The Journal of Machine Learning
Research, 3:213–231, 2003.
Brockman, Greg, Cheung, Vicki, Pettersson, Ludwig,
Schneider, Jonas, Schulman, John, Tang, Jie, and
Zaremba, Wojciech. Openai gym, 2016.
Chapelle, Olivier and Li, Lihong. An empirical evaluation
of thompson sampling. In Advances in neural information processing systems, 2011.
Chen, Tianqi, Li, Mu, Li, Yutian, Lin, Min, Wang, Naiyan,
Wang, Minjie, Xiao, Tianjun, Xu, Bing, Zhang, Chiyuan,
and Zhang, Zheng. Mxnet: A flexible and efficient machine learning library for heterogeneous distributed systems. arXiv, 2015.
Fortunato, Meire, Azar, Mohammad Gheshlaghi, Piot, Bilal, Menick, Jacob, Osband, Ian, Graves, Alex, Mnih,
Vlad, Munos, Remi, Hassabis, Demis, Pietquin, Olivier,
et al. Noisy networks for exploration. arXiv, 2017.
Gal, Yarin and Ghahramani, Zoubin. Dropout as a bayesian
approximation: Representing model uncertainty in deep
learning. In ICML, 2016.
Henderson, Peter, Islam, Riashat, Bachman, Philip, Pineau,
Joelle, Precup, Doina, and Meger, David. Deep reinforcement learning that matters. arXiv, 2017.
Jaksch, Thomas, Ortner, Ronald, and Auer, Peter. Nearoptimal regret bounds for reinforcement learning. Journal of Machine Learning Research, 2010.
Kearns, Michael and Singh, Satinder. Near-optimal reinforcement learning in polynomial time. Machine Learning, 49(2-3):209–232, 2002.
Lagoudakis, Michail G and Parr, Ronald. Least-squares
policy iteration. Journal of machine learning research, 4
(Dec):1107–1149, 2003.
Efficient Exploration through Bayesian Deep Q-Networks
Levine, Nir, Zahavy, Tom, Mankowitz, Daniel J, Tamar,
Aviv, and Mannor, Shie. Shallow updates for deep reinforcement learning. arXiv, 2017.
Russo, Daniel and Van Roy, Benjamin. Learning to optimize via posterior sampling. Mathematics of Operations
Research, 39(4):1221–1243, 2014.
Levine et al., Sergey. End-to-end training of deep visuomotor policies. JMLR, 2016.
Sanborn, Adam N and Chater, Nick. Bayesian brains without probabilities. Trends in cognitive sciences, 2016.
Lipton, Zachary C, Gao, Jianfeng, Li, Lihong, Chen, Jianshu, and Deng, Li. Combating reinforcement learning’s sisyphean curse with intrinsic fear. arXiv preprint
arXiv:1611.01211, 2016a.
Shalev-Shwartz, Shai, Shammah, Shaked, and Shashua,
Amnon. Safe, multi-agent, reinforcement learning for
autonomous driving. arXiv, 2016.
Lipton, Zachary C, Gao, Jianfeng, Li, Lihong, Li, Xiujun, Ahmed, Faisal, and Deng, Li. Efficient exploration
for dialogue policy learning with bbq networks & replay
buffer spiking. arXiv preprint arXiv:1608.05081, 2016b.
Silver, David, Schrittwieser, Julian, Simonyan, Karen,
Antonoglou, Ioannis, Huang, Aja, Guez, Arthur, Hubert,
Thomas, Baker, Lucas, Lai, Matthew, Bolton, Adrian,
et al. Mastering the game of go without human knowledge. Nature, 2017.
Machado, Marlos C, Bellemare, Marc G, Talvitie, Erik, Veness, Joel, Hausknecht, Matthew, and Bowling, Michael.
Revisiting the arcade learning environment: Evaluation
protocols and open problems for general agents. arXiv
preprint arXiv:1709.06009, 2017.
Snoek, Jasper, Rippel, Oren, Swersky, Kevin, Kiros,
Ryan, Satish, Nadathur, Sundaram, Narayanan, Patwary,
Mostofa, Prabhat, Mr, and Adams, Ryan. Scalable
bayesian optimization using deep neural networks. In
ICML, 2015.
Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David,
Graves, Alex, Antonoglou, Ioannis, Wierstra, Daan, and
Riedmiller, Martin. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013.
Strens, Malcolm. A bayesian framework for reinforcement
learning. In ICML, 2000.
Mnih, Volodymyr, Kavukcuoglu, Koray, Silver, David,
Rusu, Andrei A, Veness, Joel, Bellemare, Marc G,
Graves, Alex, Riedmiller, Martin, Fidjeland, Andreas K,
Ostrovski, Georg, et al. Human-level control through
deep reinforcement learning. Nature, 2015.
Sutton, Richard S and Barto, Andrew G. Reinforcement
learning: An introduction. MIT press Cambridge, 1998.
Tang, Haoran, Houthooft, Rein, Foote, Davis, Stooke,
Adam, Chen, Xi, Duan, Yan, Schulman, John, De Turck,
Filip, and Abbeel, Pieter.
Exploration:a study of
count-based exploration for deep reinforcement learning.
arXiv, 2016.
Osband, Ian, Russo, Dan, and Van Roy, Benjamin. (more)
efficient reinforcement learning via posterior sampling.
In Advances in Neural Information Processing Systems,
2013.
Thompson, William R. On the likelihood that one unknown
probability exceeds another in view of the evidence of
two samples. Biometrika, 1933.
Osband, Ian, Van Roy, Benjamin, and Wen, Zheng. Generalization and exploration via randomized value functions.
arXiv, 2014.
Van Hasselt, Hado, Guez, Arthur, and Silver, David. Deep
reinforcement learning with double q-learning. In AAAI,
2016.
Osband, Ian, Blundell, Charles, Pritzel, Alexander, and
Van Roy, Benjamin. Deep exploration via bootstrapped
dqn. In Advances in Neural Information Processing Systems, 2016.
Welling, Max and Teh, Yee W. Bayesian learning via
stochastic gradient langevin dynamics. In Proceedings
of the 28th International Conference on Machine Learning (ICML-11), pp. 681–688, 2011.
Osband, Ian, Russo, Daniel, Wen, Zheng, and Van Roy,
Benjamin. Deep exploration via randomized value functions. arXiv, 2017.
Ostrovski, Georg, Bellemare, Marc G, Oord, Aaron
van den, and Munos, Rémi. Count-based exploration
with neural density models. arXiv, 2017.
Rasmussen, Carl Edward and Williams, Christopher KI.
Gaussian processes for machine learning, volume 1.
MIT press Cambridge, 2006.
Efficient Exploration through Bayesian Deep Q-Networks
A. Appendix
Hyper-parameters tuning: For the BLR, we have noise variance σǫ , variance of prior over weights σ, sample size
B, posterior update period T Bayes target , and the posterior sampling period T sample . To optimize for this set of hyperparameters we set up a very simple, fast, and cheap hyper-parameter tuning procedure which proves the robustness of
BDQN. To find the first three, we set up a simple hyper-parameter search. We used a pretrained DQN model for the game
of Assault, and removed the last fully connected layer in order to have access to its already trained feature representation.
Then we tried combination of B = {T target , 10·T target }, σ = {1, 0.1, 0.001}, and σǫ = {1, 10} and test for 1000 episode
of the game. We set these parameters to their best B = 10 · T target , σ = 0.001, σ = 1.
The above hyper-parameter tuning is cheap and fast since it requires only a few times the B number of forwarding passes.
For the remaining parameter, we ran BDQN ( with weights randomly initialized) on the same game, Assault, for 5M time
target
target
steps, with a set of T Bayes target = {T target , 10 · T target } and T sample = { T 10 , T 100 } where BDQN performed
Bayes target
target
sample
better with choice of T
= 10 · T
. For both choices of T
, it performed almost equal where we
choose the higher one. We started off with the learning rate of 0.0025 and did not tune for that. Thanks to the efficient TS
exploration and closed form BLR, BDQN can learn a better policy in an even shorter period of time. In contrast, it is well
known for DQN based methods that changing the learning rate causes a major degradation in the performance, Apx. A.
The proposed hyper-parameter search is very simple where the exhaustive hyper-parameter search is likely to provide even
better performance.
Average Reward per episode
Learning rate: It is well known that DQN and DDQN are sensitive to the learning rate and change of learning rate can
degrade the performance to even worse than random policy. We tried the same learning rate as BDQN, 0.0025, for DDQN
and observed that its performance drops. Fig. 3 shows that the DDQN with higher learning rates learns as good as BDQN
at the very beginning but it can not maintain the rate of improvement and degrade even worse than the original DDQN.
Amidar
1e3
BDQN
DDQN
DDQN-10xlr
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0
1
2
3
Number of steps
4
5
1e7
Figure 3. Effect of learning rate on DDQN
Computational and sample cost comparison: For a given period of game time, the number of the backward pass in
both BDQN and DQN are the same where for BDQN it is cheaper since it has one layer (the last layer) less than DQN. In
the sense of fairness in sample usage, for example in duration of 10 · T Bayes target = 100k, all the layers of both BDQN
and DQN, except the last layer, sees the same number of samples, but the last layer of BDQN sees 16 times fewer samples
compared to the last layer of DQN. The last layer of DQN for a duration of 100k, observes 25k = 100k/4 (4 is back prob
period) mini batches of size 32, which is 16 · 100k, where the last layer of BDQN just observes samples size of B = 100k.
As it is mentioned in Alg. 1, to update the posterior distribution, BDQN draws B samples from the replay buffer and needs
to compute the feature vector of them. Therefore, during the duration of 100k decision making steps, for the learning
procedure, DDQN does 32 ∗ 25k of forward passes and 32 ∗ 25k of backward passes, while BDQN does same number of
backward passes (cheaper since there is no backward pass for the final layer) and 36 ∗ 25k of forward passes. One can
easily relax it by parallelizing this step along the main body of BDQN or deploying on-line posterior update methods.
Efficient Exploration through Bayesian Deep Q-Networks
Thompson sampling frequency: The choice of TS update frequency can be crucial from domain to domain. If one
chooses T sample too short, then computed gradient for backpropagation of the feature representation is not going to be
useful since the gradient get noisier and the loss function is changing too frequently. On the other hand, the network tries to
find a feature representation which is suitable for a wide range of different weights of the last layer, results in improper use
of model capacity. If the TS update frequency is too low, then it is far from being TS and losses randomized exploration
property. The current choice of T sample is suitable for a variety of Atari games since the length of each episode is in range
of O(T sample ) and is infrequent enough to make the feature representation robust to big changes.
For the RL problems with shorter horizon we suggest to introduce two more parameters, T̃ sample and W̃ where T̃ sample ,
the period that of W̃ is sampled our of posterior, is much smaller than T sample and W̃ is being used just for making TS
actions while W is used for backpropagation of feature representation. For game Assault, we tried using T̃ sample and W̃
but did not observe much a difference, and set them to T sample and W . But for RL setting with a shorter horizon, we
suggest using them.
Average Reward per episode
Further investigation in Atlantis: After removing the maximum episode length limit for the game Atlantis, BDQN gets
the score of 62M. This episode is long enough to fill half of the replay buffer and make the model perfect for the later
part of the game but losing the crafted skill for the beginning of the game. We observe in Fig. 4 that after losing the game
in a long episode, the agent forgets a bit of its skill and loses few games but wraps up immediately and gets to score of
30M . To overcome this issue, one can expand the replay buffer size, stochastically store samples in the reply buffer where
the later samples get stored with lowers chance, or train new models for the later parts of the episode. There are many
possible cures for this interesting observation and while we are comparing against DDQN, we do not want to advance
BDQN structure-wise.
Atlantis
1e7
BDQN
6
5
4
3
2
1
0
0.0
0.5
1.0
1.5
Number of steps
2.0
2.5
1e7
Average Reward per episode
Figure 4. BDQN on Atlantis after removing the limit on max of episode length hits the score of 62M in 16M samples.
Atlantis
1e6
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0
1
2
3
Number of steps
4
1e7
Figure 5. A couple of more runs of BDQN where the jump around 15M constantly happens
Efficient Exploration through Bayesian Deep Q-Networks
Further discussion on Reproducibility In Table. 2, we provide the scores of bootstrap DQN (Osband et al., 2016) and
NoisyNet6 (Fortunato et al., 2017) along side with BDQN. These score are directly copied from their original papers and
we did not make any change to them. We also desired to report the scores of count-based method (Ostrovski et al., 2017),
but unfortunately there is no table of score in that paper in order to provide them here.
In order to make it easier for the readers to compare against the results in Ostrovski et al. (2017), we visually approximated
their plotted curves for CT S, P ixel, P ixel, and Reactor, and added them to the Table. 2. We added these numbers just
for the convenience of the readers Surly we do not argue any scientific meaning for them and leave it to the readers to
interpret them.
Table. 2 shows a significant improvement of BDQN over these baselines by looking at Table. 2. Despite the simplicity
and negligible computation overhead of BDQN over DDQN, we can not scientifically claim that BDQN outperforms these
baselines by just looking at the scores in Table.2 because we are not aware of their detailed implementation as well as
environment detail. For example, in this work, we directly implemented DDQN by following the implementation details
mentioned in the original DDQN paper and the scores of our DDQN implementation during the evaluation time almost
matches the scores of DDQN reported in the original paper. But the reported scores of implemented DDQN in Osband et al.
(2016) are much different from the reported score in the original DDQN paper.
6
This work does not have scores of Noisy-net with DDQN objective function but it has Noisy-net with DQN objective which are the
scores reported in Table. 2
| 2cs.AI
|
Knowledge Matters: Importance of Prior Information for
Optimization
Çağlar Gülçehre
[email protected]
Département d’informatique et de recherche opérationnelle
Université de Montréal, Montréal, QC, Canada
arXiv:1301.4083v6 [cs.LG] 13 Jul 2013
Yoshua Bengio
[email protected]
Département d’informatique et de recherche opérationnelle
Université de Montréal, Montréal, QC, Canada
Editor: Not Assigned
Abstract
We explore the effect of introducing prior information into the intermediate level of deep
supervised neural networks for a learning task on which all the black-box state-of-the-art machine learning algorithms tested have failed to learn. We motivate our work from the hypothesis
that there is an optimization obstacle involved in the nature of such tasks, and that humans
learn useful intermediate concepts from other individuals via a form of supervision or guidance
using a curriculum. The experiments we have conducted provide positive evidence in favor of
this hypothesis. In our experiments, a two-tiered MLP architecture is trained on a dataset for
which each image input contains three sprites, and the binary target class is 1 if all three have
the same shape. Black-box machine learning algorithms only got chance on this task. Standard
deep supervised neural networks also failed. However, using a particular structure and guiding
the learner by providing intermediate targets in the form of intermediate concepts (the presence of each object) allows to nail the task. Much better than chance but imperfect results are
also obtained by exploring architecture and optimization variants, pointing towards a difficult
optimization task. We hypothesize that the learning difficulty is due to the composition of two
highly non-linear tasks. Our findings are also consistent with hypotheses on cultural learning
inspired by the observations of effective local minima (possibly due to ill-conditioning and the
training procedure not being able to escape what appears like a local minimum).
Keywords: Deep Learning, Neural Networks, Optimization, Evolution of Culture, Curriculum Learning, Training with Hints
1. Introduction
There is a recent emerging interest in different fields of science for cultural learning (Henrich
and McElreath, 2003) and how groups of individuals exchanging information can learn in ways
superior to individual learning. This is also witnessed by the emergence of new research fields
such as ”Social Neuroscience”. Learning from other agents in an environment by the means
of cultural transmission of knowledge with a peer-to-peer communication is an efficient and
natural way of acquiring or propagating common knowledge. The most popular belief on how
the information is transmitted between individuals is that bits of information are transmitted
by small units, called memes, which share some characteristics of genes, such as self-replication,
mutation and response to selective pressures (Dawkins, 1976).
1
This paper is based on the hypothesis (which is further elaborated in Bengio (2013a)) that
human culture and the evolution of ideas have been crucial to counter an optimization issue:
this difficulty would otherwise make it difficult for human brains to capture high level knowledge of the world without the help of other educated humans. In this paper machine learning
experiments are used to investigate some elements of this hypothesis by seeking answers for
the following questions: are there machine learning tasks which are intrinsically hard for a
lone learning agent but that may become very easy when intermediate concepts are provided
by another agent as additional intermediate learning cues, in the spirit of Curriculum Learning (Bengio et al., 2009b)? What makes such learning tasks more difficult? Can specific initial
values of the neural network parameters yield success when random initialization yield complete failure? Is it possible to verify that the problem being faced is an optimization problem or
with a regularization problem? These are the questions discussed (if not completely addressed)
here, which relate to the following broader question: how can humans (and potentially one day,
machines) learn complex concepts?
In this paper, results of different machine learning algorithms on an artificial learning task
involving binary 64×64 images are presented. In that task, each image in the dataset contains
3 Pentomino tetris sprites (simple shapes). The task is to figure out if all the sprites in the
image are the same or if there are different sprite shapes in the image. Several state-of-the-art
machine learning algorithms have been tested and none of them could perform better than
a random predictor on the test set. Nevertheless by providing hints about the intermediate
concepts (the presence and location of particular sprite classes), the problem can easily be solved
where the same-architecture neural network without the intermediate concepts guidance fails.
Surprisingly, our attempts at solving this problem with unsupervised pre-training algorithms
failed solve this problem. However, with specific variations in the network architecture or
training procedure, it is found that one can make a big dent in the problem. For showing the
impact of intermediate level guidance, we experimented with a two-tiered neural network, with
supervised pre-training of the first part to recognize the category of sprites independently of
their orientation and scale, at different locations, while the second part learns from the output
of the first part and predicts the binary task of interest.
The objective of this paper is not to propose a novel learning algorithm or architecture,
but rather to refine our understanding of the learning difficulties involved with composed tasks
(here a logical formula composed with the detection of object classes), in particular the training
difficulties involved for deep neural networks. The results also bring empirical evidence in favor
of some of the hypotheses from Bengio (2013a), discussed below, as well as introducing a
particular form of curriculum learning (Bengio et al., 2009b).
Building difficult AI problems has a long history in computer science. Specifically hard
AI problems have been studied to create CAPTCHA’s that are easy to solve for humans, but
hard to solve for machines (Von Ahn et al., 2003). In this paper we are investigating a difficult
problem for the off-the-shelf black-box machine learning algorithms.1
1.1 Curriculum Learning and Cultural Evolution Against Effective Local Minima
What Bengio (2013a) calls an effective local minimum is a point where iterative training
stalls, either because of an actual local minimum or because the optimization algorithm is
1. You can access the source code of some experiments presented in that paper and their hyperparameters from
here: https://github.com/caglar/kmatters
2
unable (in reasonable time) to find a descent path (e.g., because of serious ill-conditioning). In
this paper, it is hypothesized that some more abstract learning tasks such as those obtained
by composing simpler tasks are more likely to yield effective local minima for neural networks,
and are generally hard for general-purpose machine learning algorithms.
The idea that learning can be enhanced by guiding the learner through intermediate easier
tasks is old, starting with animal training by shaping (Skinner, 1958; Peterson, 2004; Krueger
and Dayan, 2009). Bengio et al. (2009b) introduce a computational hypothesis related to a
presumed issue with effective local minima when directly learning the target task: the good
solutions correspond to hard-to-find-by-chance effective local minima, and intermediate tasks
prepare the learner’s internal configuration (parameters) in a way similar to continuation methods in global optimization (which go through a sequence of intermediate optimization problems,
starting with a convex one where local minima are no issue, and gradually morphing into the
target task of interest).
In a related vein, Bengio (2013a) makes the following inferences based on experimental
observations of deep learning and neural network learning:
Point 1: Training deep architectures is easier when some hints are given about the function
that the intermediate levels should compute (Hinton et al., 2006; Weston et al., 2008;
Salakhutdinov and Hinton, 2009; Bengio, 2009). The experiments performed here expand
in particular on this point.
Point 2: It is much easier to train a neural network with supervision (where examples ar
provided to it of when a concept is present and when it is not present in a variety of
examples) than to expect unsupervised learning to discover the concept (which may also
happen but usually leads to poorer renditions of the concept). The poor results obtained
with unsupervised pre-training reinforce that hypothesis.
Point 3: Directly training all the layers of a deep network together not only makes it difficult to
exploit all the extra modeling power of a deeper architecture but in many cases it actually
yields worse results as the number of required layers is increased (Larochelle et al., 2009;
Erhan et al., 2010). The experiments performed here also reinforce that hypothesis.
Point 4: Erhan et al. (2010) observed that no two training trajectories ended up in the
same effective local minimum, out of hundreds of runs, even when comparing solutions as
functions from input to output, rather than in parameter space (thus eliminating from the
picture the presence of symmetries and multiple local minima due to relabeling and other
reparametrizations). This suggests that the number of different effective local minima
(even when considering them only in function space) must be huge.
Point 5: Unsupervised pre-training, which changes the initial conditions of the descent procedure, sometimes allows to reach substantially better effective local minima (in terms
of generalization error!), and these better local minima do not appear to be reachable
by chance alone (Erhan et al., 2010). The experiments performed here provide another
piece of evidence in favor of the hypothesis that where random initialization can yield
rather poor results, specifically targeted initialization can have a drastic impact, i.e., that
3
effective local minima are not just numerous but that some small subset of them are much
better and hard to reach by chance.2
Based on the above points, Bengio (2013a) then proposed the following hypotheses regarding
learning of high-level abstractions.
• Optimization Hypothesis: When it learns, a biological agent performs an approximate
optimization with respect to some implicit objective function.
• Deep Abstractions Hypothesis: Higher level abstractions represented in brains require deeper computations (involving the composition of more non-linearities).
• Local Descent Hypothesis: The brain of a biological agent relies on approximate local
descent and gradually improves itself while learning.
• Effective Local Minima Hypothesis: The learning process of a single human learner
(not helped by others) is limited by effective local minima.
• Deeper Harder Hypothesis: Effective local minima are more likely to hamper learning
as the required depth of the architecture increases.
• Abstractions Harder Hypothesis: High-level abstractions are unlikely to be discovered by a single human learner by chance, because these abstractions are represented by
a deep subnetwork of the brain, which learns by local descent.
• Guided Learning Hypothesis: A human brain can learn high level abstractions if
guided by the signals produced by other agents that act as hints or indirect supervision
for these high-level abstractions.
• Memes Divide-and-Conquer Hypothesis: Linguistic exchange, individual learning
and the recombination of memes constitute an efficient evolutionary recombination operator in the meme-space. This helps human learners to collectively build better internal
representations of their environment, including fairly high-level abstractions.
This paper is focused on “Point 1 ” and testing the “Guided Learning Hypothesis”, using
machine learning algorithms to provide experimental evidence. The experiments performed
also provide evidence in favor of the “Deeper Harder Hypothesis” and associated “Abstractions
Harder Hypothesis”. Machine Learning is still far beyond the current capabilities of humans,
and it is important to tackle the remaining obstacles to approach AI. For this purpose, the
question to be answered is why tasks that humans learn effortlessly from very few examples,
while machine learning algorithms fail miserably?
2. Recent work showed that rather deep feedforward networks can be very successfully trained when large
quantities of labeled data are available (Ciresan et al., 2010; Glorot et al., 2011a; Krizhevsky et al., 2012).
Nonetheless, the experiments reported here suggest that it all depends on the task being considered, since
even with very large quantities of labeled examples, the deep networks trained here were unsuccessful.
4
2. Culture and Optimization Difficulty
As hypothesized in the “Local Descent Hypothesis”, human brains would rely on a local approximate descent, just like a Multi-Layer Perceptron trained by a gradient-based iterative optimization. The main argument in favor of this hypothesis relies on the biologically-grounded
assumption that although firing patterns in the brain change rapidly, synaptic strengths underlying these neural activities change only gradually, making sure that behaviors are generally
consistent across time. If a learning algorithm is based on a form of local (e.g. gradient-based)
descent, it can be sensitive to effective local minima (Bengio, 2013a).
When one trains a neural network, at some point in the training phase the evaluation of
error seems to saturate, even if new examples are introduced. In particular Erhan et al. (2010)
find that early examples have a much larger weight in the final solution. It looks like the learner
is stuck in or near a local minimum. But since it is difficult to verify if this is near a true local
minimum or simply an effect of strong ill-conditioning, we call such a “stuck” configuration an
effective local minimum, whose definition depends not just on the optimization objective but
also on the limitations of the optimization algorithm.
Erhan et al. (2010) highlighted both the issue of effective local minima and a regularization effect when initializing a deep network with unsupervised pre-training. Interestingly, as
the network gets deeper the difficulty due to effective local minima seems to be get more pronounced. That might be because of the number of effective local minima increases (more like
an actual local minima issue), or maybe because the good ones are harder to reach (more like
an ill-conditioning issue) and more work will be needed to clarify this question.
As a result of Point 4 we hypothesize that it is very difficult for an individual’s brain to
discover some higher level abstractions by chance only. As mentioned in the “Guided Learning
Hypothesis” humans get hints from other humans and learn high-level concepts with the guidance of other humans3 . Curriculum learning (Bengio et al., 2009a) and incremental learning
(Solomonoff, 1989), are examples of this. This is done by properly choosing the sequence of
examples seen by the learner, where simpler examples are introduced first and more complex
examples shown when the learner is ready for them. One of the hypothesis on why curriculum
works states that curriculum learning acts as a continuation method that allows one to discover
a good minimum, by first finding a good minimum of a smoother error function. Recent experiments on human subjects also indicates that humans teach by using a curriculum strategy
(Khan et al., 2011).
Some parts of the human brain are known to have a hierarchical organization (i.e. visual
cortex) consistent with the deep architecture studied in machine learning papers. As we go from
the sensory level to higher levels of the visual cortex, we find higher level areas corresponding
to more abstract concepts. This is consistent with the Deep Abstractions Hypothesis.
Training neural networks and machine learning algorithms by decomposing the learning
task into sub-tasks and exploiting prior information about the task is well-established and
in fact constitutes the main approach to solving industrial problems with machine learning.
The contribution of this paper is rather on rendering explicit the effective local minima issue
and providing evidence on the type of problems for which this difficulty arises. This prior
information and hints given to the learner can be viewed as inductive bias for a particular task,
an important ingredient to obtain a good generalization error (Mitchell, 1980). An interesting
3. But some high-level concepts may also be hardwired in the brain, as assumed in the universal grammar
hypothesis (Montague, 1970), or in nature vs nurture discussions in cognitive science.
5
earlier finding in that line of research was done with Explanation Based Neural Networks
(EBNN) in which a neural network transfers knowledge across multiple learning tasks. An
EBNN uses previously learned domain knowledge as an initialization or search bias (i.e. to
constrain the learner in the parameter space) (O’Sullivan, 1996; Mitchell and Thrun, 1993).
Another related work in machine learning is mainly focused on reinforcement learning algorithms, based on incorporating prior knowledge in terms of logical rules to the learning
algorithm as a prior knowledge to speed up and bias learning (Kunapuli et al., 2010; Towell
and Shavlik, 1994).
As discussed in “Memes Divide and Conquer Hypothesis“ societies can be viewed as a
distributed computational processing systems. In civilized societies knowledge is distributed
across different individuals, this yields a space efficiency. Moreover computation, i.e. each
individual can specialize on a particular task/topic, is also divided across the individuals in the
society and hence this will yield a computational efficiency. Considering the limitations of the
human brain, the whole processing can not be done just by a single agent in an efficient manner.
A recent study in paleoantropology states that there is a substantial decline in endocranial
volume of the brain in the last 30000 years Henneberg (1988). The volume of the brain shrunk
to 1241 ml from 1502 ml (Henneberg and Steyn, 1993). One of the hypothesis on the reduction
of the volume of skull claims that, decline in the volume of the brain might be related to the
functional changes in brain that arose as a result of cultural development and emergence of
societies given that this time period overlaps with the transition from hunter-gatherer lifestyle
to agricultural societies.
3. Experimental Setup
Some tasks, which seem reasonably easy for humans to learn4 , are nonetheless appearing almost
impossible to learn for current generic state-of-art machine learning algorithms.
Here we study more closely such a task, which becomes learnable if one provides hints to
the learner about appropriate intermediate concepts. Interestingly, the task we used in our
experiments is not only hard for deep neural networks but also for non-parametric machine
learning algorithms such as SVM’s, boosting and decision trees.
The result of the experiments for varying size of dataset with several off-the-shelf black box
machine learning algorithms and some popular deep learning algorithms are provided in Table
1. The detailed explanations about the algorithms and the hyperparameters used for those
algorithms are given in the Appendix Section 5.2. We also provide some explanations about
the methodologies conducted for the experiments at Section 3.2.
3.1 Pentomino Dataset
In order to test our hypothesis, an artificial dataset for object recognition using 64×64 binary
images is designed5 . If the task is two tiered (i.e., with guidance provided), the task in the
first part is to recognize and locate each Pentomino object class6 in the image. The second
4. keeping in mind that humans can exploit prior knowledge, either from previous learning or innate knowledge.
5. The source code for the script that generates the artificial Pentomino datasets (Arcade-Universe) is available
at: https://github.com/caglar/Arcade-Universe. This implementation is based on Olivier Breuleux’s
bugland dataset generator.
6. A human learner does not seem to need to be taught the shape categories of each Pentomino sprite in order
to solve the task. On the other hand, humans have lots of previously learned knowledge about the notion of
shape and how central it is in defining categories.
6
(a) sprites, not all same type
(b) sprites, all of same type
Figure 1: Left (a): An example image from the dataset which has a different sprite type in it.
Right (b): An example image from the dataset that has only one type of Pentomino
object in it, but with different orientations and scales.
part/final binary classification task is to figure out if all the Pentominos in the image are of
the same shape class or not. If a neural network learned to detect the categories of each object
at each location in an image, the remaining task becomes an XOR-like operation between the
detected object categories. The types of Pentomino objects that is used for generating the
dataset are as follows:
Pentomino sprites N, P, F, Y, J, and Q, along with the Pentomino N2 sprite (mirror of
“Pentomino N” sprite), the Pentomino F2 sprite (mirror of “Pentomino F” sprite), and the
Pentomino Y2 sprite (mirror of “Pentomino Y” sprite).
Figure 2: Different classes of Pentomino shapes used in our dataset.
As shown in Figures 1(a) and 1(b), the synthesized images are fairly simple and do not
have any texture. Foreground pixels are “1” and background pixels are “0”. Images of the
training and test sets are generated iid. For notational convenience, assume that the domain
of raw input images is X, the set of sprites is S, the set of intermediate object categories is Y
for each possible location in the image and the set of final binary task outcomes is Z = {0, 1}.
Two different types of rigid body transformation is performed: sprite rotation rot(X, γ) where
Γ = {γ : (γ = 90 × φ) ∧ [(φ ∈ N), (0 ≤ φ ≤ 3)]} and scaling scale(X, α) where α ∈ {1, 2} is the
scaling factor. The data generating procedure is summarized below.
Sprite transformations: Before placing the sprites in an empty image, for each image x ∈ X,
a value for z ∈ Z is randomly sampled which is to have (or not) the same three sprite
shapes in the image. Conditioned on the constraint given by z, three sprites are randomly
7
selected sij from S without replacement. Using a uniform probability distribution over
all possible scales, a scale is chosen and accordingly each sprite image is scaled. Then
rotate each sprite is randomly rotated by a multiple of 90 degrees.
Sprite placement: Upon completion of sprite transformations, a 64×64 uniform grid is generated which is divided into 8×8 blocks, each block being of size 8×8 pixels, and randomly
select three different blocks from the 64=8×8 on the grid and place the transformed
objects into different blocks (so they cannot overlap, by construction).
Each sprite is centered in the block in which it is located. Thus there is no object translation
inside the blocks. The only translation invariance is due to the location of the block inside the
image.
A Pentomino sprite is guaranteed to not overflow the block in which it is located, and there
are no collisions or overlaps between sprites, making the task simpler. The largest possible
Pentomino sprite can be fit into an 8×4 mask.
3.2 Learning Algorithms Evaluated
Initially the models are cross-validated by using 5-fold cross-validation. With 40,000 examples,
this gives 32,000 examples for training and 8,000 examples for testing. For neural network
algorithms, stochastic gradient descent (SGD) is used for training. The following standard
learning algorithms were first evaluated: decision trees, SVMs with Gaussian kernel, ordinary
fully-connected Multi-Layer Perceptrons, Random Forests, k-Nearest Neighbors, Convolutional
Neural Networks, and Stacked Denoising Auto-Encoders with supervised fine-tuning. More
details of the configurations and hyper-parameters for each of them are given in Appendix
Section 5.2. The only better than chance results were obtained with variations of the Structured
Multi-Layer Perceptron described below.
3.2.1 Structured Multi-Layer Perceptron (SMLP)
The neural network architecture that is used to solve this task is called the SMLP (Structured
Multi-Layer Perceptron), a deep neural network with two parts as illustrated in Figure 5 and
7:
The lower part, P1NN (Part 1 Neural Network, as it is called in the rest of the paper),
has shared weights and local connectivity, with one identical MLP instance of the P1NN for
each patch of the image, and typically an 11-element output vector per patch (unless otherwise
noted). The idea is that these 11 outputs per patch could represent the detection of the sprite
shape category (or the absence of sprite in the patch). The upper part, P2NN (Part 2 Neural
Network) is a fully connected one hidden layer MLP that takes the concatenation of the outputs
of all patch-wise P1NNs as input. Note that the first layer of P1NN is similar to a convolutional
layer but where the stride equals the kernel size, so that windows do not overlap, i.e., P1NN can
be decomposed into separate networks sharing the same parameters but applied on different
patches of the input image, so that each network can actually be trained patch-wise in the case
where a target is provided for the P1NN outputs. The P1NN output for patch pi which is
extracted from the image x is computed as follows:
fθ (pi ) = g2 (V g1 (U pi + b) + c)
8
(1)
where pi ∈ Rd is the input patch/receptive field extracted from location i of a single image.
U ∈ Rdh ×d is the weight matrix for the first layer of P1NN and b ∈ Rhd is the vector of biases
for the first layer of P1NN. g1 (·) is the activation function of the first layer and g2 (·) is the
activation function of the second layer. In many of the experiments, best results were obtained
with g1 (·) a rectifying non-linearity (a.k.a. as RELU), which is max(0, X) (Jarrett et al.,
2009b; Nair and Hinton, 2010; Glorot et al., 2011a; Krizhevsky et al., 2012). V ∈ Rdh ×do is the
second layer’s weights matrix, such that and c ∈ Rdo are the biases of the second layer of the
P1NN, with do expected to be smaller than dh .
In this way, g1 (U pi + b) is an overcomplete representation of the input patch that can
potentially represent all the possible Pentomino shapes for all factors of variations in the patch
(rotation, scaling and Pentomino shape type). On the other hand, when trained with hints,
fθ (pi ) is expected to be the lower dimensional representation of a Pentomino shape category
invariant to scaling and rotation in the given patch.
In the experiments with SMLP trained with hints (targets at the output of P1NN), the
P1NN is expected to perform classification of each 8×8 non-overlapping patches of the original
64×64 input image without having any prior knowledge of whether that specific patch contains
a Pentomino shape or not. P1NN in SMLP without hints just outputs the local activations for
each patch, and gradients on fθ (pi ) are backpropagated from the upper layers. In both cases
P1NN produces the input representation for the Part 2 Neural Net (P2NN). Thus the input
representation of P2NN is the concatenated output of P1NN across all the 64 patch locations:
ho = [fθ (p0 ), ..., fθ (pi ), ..., fθ (pN ))] where N is the number of patches and the ho ∈ Rdi , di =
do × N . ho is the concatenated output of the P1NN at each patch.
There is a standardization layer on top of the output of P1NN that centers the activations
and performs divisive normalization by dividing by the standard deviation over a minibatch
of the activations of that layer. We denote the standardization function z(·). Standardization
makes use of the mean and standard deviation computed for each hidden unit such that each
hidden unit of ho will have 0 activation and unit standard deviation on average over the
minibatch. X is the set of pentomino images in the minibatch, where X ∈ Rdin ×N is a matrix
(i)
with N images. ho (xj ) is the vector of activations of the i-th hidden unit of hidden layer
ho (xj ) for the j-th example, with xj ∈ X.
µh(i) =
o
s
σh(i) =
1 X (i)
ho (xj )
N
(2)
xj ∈X
PN
j
(i)
(ho (xj ) − µh(i) )2
o
N
o
+
(3)
(i)
z(ho(i) (xj ))
=
ho (xj ) − µh(i)
o
max(σh(i) , )
(4)
o
where is a very small constant, that is used to prevent numerical underflows in the standard
deviation. P1NN is trained on each 8×8 patches extracted from the image.
ho is standardized for each training and test sample separately. Different values of were
used for SMLP-hints and SMLP-nohints.
The concatenated output of P1NN is fed as an input to the P2NN. P2NN is a feedforward
MLP with a sigmoid output layer using a single RELU hidden layer. The task of P2NN is to
perform a nonlinear logical operation on the representation provided at the output of P1NN.
9
3.2.2 Structured Multi Layer Perceptron Trained with Hints (SMLP-hints)
The SMLP-hints architecture exploits a hint about the presence and category of Pentomino
objects, specifying a semantics for the P1NN outputs. P1NN is trained with the intermediate
target Y , specifying the type of Pentomino sprite shape present (if any) at each of the 64 patches
(8×8 non-overlapping blocks) of the image. Because a possible answer at a given location can
be “none of the object types” i.e., an empty patch, yp (for patch p) can take one of the 11
possible values, 1 for rejection and the rest is for the Pentomino shape classes, illustrated in
Figure 2:
(
0
if patch p is empty
yp =
s ∈ S if the patch p contains a Pentomino sprite .
A similar task has been studied by Fleuret et al. (2011) (at SI appendix Problem 17), who
compared the performance of humans vs computers.
The SMLP-hints architecture takes advantage of dividing the task into two subtasks during
training with prior information about intermediate-level relevant factors. Because the sum of
the training losses decomposes into the loss on each patch, the P1NN can be pre-trained patchwise. Each patch-specific component of the P1NN is a fully connected MLP with 8×8 inputs
and 11 outputs with a softmax output layer. SMLP-hints uses the the standardization given
in Equation 3 but with = 0.
The standardization is a crucial step for training the SMLP on the Pentomino dataset, and
yields much sparser outputs, as seen on Figures 3 and 4. If the standardization is not used,
even SMLP-hints could not solve the Pentomino task. In general, the standardization step
dampens the small activations and augments larger ones(reducing the noise). Centering the
activations of each feature detector in a neural network has been studied in (Raiko et al., 2012)
and (Vatanen et al., 2013). They proposed that transforming the outputs of each hidden neuron
in a multi-layer perceptron network to have zero output and zero slope on average makes first
order optimization methods closer to the second order techniques.
By default, the SMLP uses rectifier hidden units as activation function, we found a significant boost by using rectification compared to hyperbolic tangent and sigmoid activation
functions. The P1NN has a highly overcomplete architecture with 1024 hidden units per patch,
and L1 and L2 weight decay regularization coefficients on the weights (not the biases) are respectively 1e-6 and 1e-5. The learning rate for the P1NN is 0.75. 1 training epoch was enough
for the P1NN to learn the features of Pentomino shapes perfectly on the 40000 training examples. The P2NN has 2048 hidden units. L1 and L2 penalty coefficients for the P2NN are
1e-6, and the learning rate is 0.1. These were selected by trial and error based on validation
set error. Both P1NN (for each patch) and P2NN are fully-connected neural networks, even
though P1NN globally is a special kind of convolutional neural network.
Filters of the first layer of SMLP are shown in Figure 6. These are the examples of the
filters obtained with the SLMP-hints trained with 40k examples, whose results are given in
Table 1. Those filters look very noisy but they work perfectly on the Pentomino task.
3.2.3 Deep and Structured Supervised MLP without Hints (SMLP-nohints)
SMLP-nohints uses the same connectivity pattern (and deep architecture) that is also used
in the SMLP-hints architecture, but without using the intermediate targets (Y ). It directly
predicts the final outcome of the task (Z), using the same number of hidden units, the same
10
Figure 3: Bar chart of concatenated softmax output activations ho of P1NN (11×64=704 outputs) in SMLP-hints before standardization, for a selected example. There are very
large spikes at each location for one of the possible 11 outcome (1 of K representation).
11
Figure 4: Softmax output activations ho of P1NN at SMLP-hints before standardization. There
are positive spiked outputs at the locations where there is a Pentomino shape. Positive and negative spikes arise because most of the outputs are near an average value.
Activations are higher at the locations where there is a pentomino shape.
12
Structured MLP
Architecture with
Hints
Second
Level
Neural
Network
Final Binary task
labels
Intermediate level
targets.
First
Level
Neural
Network
Figure 5: Structured MLP architecture, used with hints (trained in two phases, first P1NN,
bottom two layers, then P2NN, top two layers). In SMLP-hints, P1NN is trained
on each 8x8 patch extracted from the image and the softmax output probabilities of
all 64 patches are concatenated into a 64×11 vector that forms the input of P2NN.
Only U and V are learned in the P1NN and its output on each patch is fed into
P2NN. The first level and the second level neural networks are trained separately,
not jointly.
Figure 6: Filters of Structured MLP architecture, trained with hints on 40k examples.
13
connectivity and the same activation function for the hidden units as SMLP-hints. 120 hyperparameter values have been evaluated by randomly selecting the number of hidden units
from [64, 128, 256, 512, 1024, 1200, 2048] and randomly sampling 20 learning rates uniformly in
the log-domain within the interval of [0.008, 0.8]. Two fully connected hidden layers with 1024
hidden units (same as P1NN) per patch is used and 2048 (same as P2NN) for the last hidden layer, with twenty training epochs. For this network the best results are obtained with a
learning rate of 0.05.7
Structured MLP
Architecture without
Hints
Second
Level
Neural
Network
Final Binary task
labels
First
Level
Neural
Network
Figure 7: Structured MLP architecture, used without hints (SMLP-nohints). It is the same
architecture as SMLP-hints (Figure 5) but with both parts (P1NN and P2NN) trained
jointly with respect to the final binary classification task.
We chose to experiment with various SMLP-nohint architectures and optimization procedures, trying unsuccessfully to achieve as good results with SMLP-nohint as with SMLP-hints.
Rectifier Non-Linearity A rectifier nonlinearity is used for the activations of MLP hidden
layers. We observed that using piecewise linear nonlinearity activation function such as the
rectifier can make the optimization more tractable.
7. The source code of the structured MLP is available at the github repository: https://github.com/caglar/
structured_mlp
14
Figure 8: First layer filters learned by the Structured MLP architecture, trained without using hints on 447600 examples with online SGD and a sigmoid intermediate layer
activation.
Intermediate Layer The output of the P1NN is considered as an intermediate layer of the
SMLP. For the SMLP-hints, only softmax output activations have been tried at the intermediate
layer, and that sufficed to learn the task. Since things did not work nearly as well with
the SMLP-nohints, several different activation functions have been tried: softmax(·), tanh(·),
sigmoid(·) and linear activation functions.
Standardization Layer Normalization at the last layer of the convolutional neural networks
has been used occasionaly to encourage the competition between the hidden units. (Jarrett
et al., 2009a) used a local contrast normalization layer in their architecture which performs
subtractive and divisive normalization. A local contrast normalization layer enforces a local
competition between adjacent features in the feature map and between features at the same
spatial location in different feature maps. Similarly (Krizhevsky et al., 2012) observed that
using a local response layer that enjoys the benefit of using local normalization scheme aids
generalization.
Standardization has been observed to be crucial for both SMLP trained with or without hints. In both SMLP-hints and SMLP-nohints experiments, the neural network was not
able to generalize or even learn the training set without using standardization in the SMLP
intermediate layer, doing just chance performance. More specifically, in the SMLP-nohints architecture, standardization is part of the computational graph, hence the gradients are being
backpropagated through it. The mean and the standard deviation is computed for each hidden
unit separately at the intermediate layer as in Equation 4. But in order to prevent numerical
underflows or overflows during the backpropagation we have used = 1e − 8 (Equation 3).
The benefit of having sparse activations may be specifically important for the ill-conditioned
problems, for the following reasons. When a hidden unit is “off”, its gradient (the derivative of
the loss with respect to its output) is usually close to 0 as well, as seen here. That means that all
off-diagonal second derivatives involving that hidden unit (e.g. its input weights) are also near 0.
This is basically like removing some columns and rows from the Hessian matrix associated with
a particular example. It has been observed that the condition number of the Hessian matrix
(specifically, its largest eigenvalue) increases as the size of the network increases (Dauphin and
Bengio, 2013), making training considerably slower and inefficient (Dauphin and Bengio, 2013).
Hence one would expect that as sparsity of the gradients (obtained because of sparsity of the
activations) increases, training would become more efficient, as if we were training a smaller
sub-network for each example, with shared weights across examples, as in dropouts (Hinton
et al., 2012).
In Figure 9, the activation of each hidden unit in a bar chart is shown: the effect of
standardization is significant, making the activations sparser.
15
(a) Before standardization.
(b) After standardization.
Figure 9: Activations of the intermediate-level hidden units of an SLMP-nohints for a particular
examples (x-axis: hidden unit number, y-axis: activation value). Left (a): before
standardization. Right (b): after standardization.
In Figure 10, one can see the activation histogram of the SMLP-nohints intermediate layer,
showing the distribution of activation values, before and after standardization. Again the
sparsifying effect of standardization is very apparent.
(a) Before standardization.
(b) After standardization.
Figure 10: Distribution histogram of activation values of SMLP-nohints intermediate layer.
Left (a): before standardization. Right (b): after standardization.
In Figures 10 and 9, the intermediate level activations of SMLP-nohints are shown before
and after standardization. These are for the same SMLP-nohints architecture whose results
are presented on Table 1. For that same SMLP, the Adadelta (Zeiler, 2012) adaptive learning
16
rate scheme has been used, with 512 hidden units for the hidden layer of P1NN and rectifier
activation function. For the output of the P1NN, 11 sigmoidal units have been used while
P2NN had 1200 hidden units with rectifier activation function. The output nonlinearity of the
P2NN is a sigmoid and the training objective is the binary crossentropy.
Adaptive Learning Rates We have experimented with several different adaptive learning
rate algorithms. We tried rmsprop 8 , Adadelta (Zeiler, 2012), Adagrad (Duchi et al., 2010) and
a linearly (1/t) decaying learning rate (Bengio, 2013b). For the SMLP-nohints with sigmoid
activation function we have found Adadelta(Zeiler, 2012) converging faster to an effective local
minima and usually yielding better generalization error compared to the others.
3.2.4 Deep and Structured MLP with Unsupervised Pre-Training
Several experiments have been conducted using an architecture similar to the SMLP-nohints,
but by using unsupervised pre-training of P1NN, with Denoising Auto-Encoder (DAE) and/or
Contractive Auto-Encoders (CAE). Supervised fine-tuning proceeds as in the deep and structured MLP without hints. Because an unsupervised learner may not focus the representation
just on the shapes, a larger number of intermediate-level units at the output of P1NN has been
explored: previous work on unsupervised pre-training generally found that larger hidden layers
were optimal when using unsupervised pre-training, because not all unsupervised features will
be relevant to the task at hand. Instead of limiting to 11 units per patch, we experimented
with networks with up to 20 hidden (i.e., code) units per patch in the second-layer patch-wise
auto-encoder.
In Appendix 5.1 we also provided the result of some experiments with binary-binary RBMs
trained on 8×8 patches from the 40k training dataset.
In unsupervised pretraining experiments in this paper, both contractive auto-encoder (CAE)
with sigmoid nonlinearity and binary cross entropy cost function and denoising auto-encoder (DAE)
have been used. In the second layer, experiments were performed with a DAE with rectifier
hidden units utilizing L1 sparsity and weight decay on the weights of the auto-encoder. Greedy
layerwise unsupervised training procedure is used to train the deep auto-encoder architecture
(Bengio et al., 2007). In unsupervised pretraining experiments, tied weights have been used.
Different combinations of CAE and DAE for unsupervised pretraining have been tested, but
none of the configurations tested managed to learn the Pentomino task, as shown in Table 1.
3.3 Experiments with 1 of K representation
To explore the effect of changing the complexity of the input representation on the difficulty
of the task, a set of experiments have been designed with symbolic representations of the
information in each patch. In all cases an empty patch is represented with a 0 vector. These
representation can be seen as an alternative input for a P2NN-like network, i.e., they were fed
as input to an MLP or another black-box classifier.
The following four experiments have been conducted, each one using one using a different
input representation for each patch:
8. This is learning rate scaling method that is discussed by G. Hinton in his Video Lecture 6.5 - rmsprop: Divide
the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine
Learning, 2012.
17
Algorithm
SVM RBF
K Nearest Neighbors
Decision Tree
Randomized Trees
MLP
Convnet/Lenet5
Maxout Convnet
2 layer sDA
Struct. Supervised MLP w/o hints
Struct. MLP+CAE Supervised Finetuning
Struct. MLP+CAE+DAE, Supervised Finetuning
Struct. MLP+DAE+DAE, Supervised Finetuning
20k dataset
Training
Test
Error
Error
26.2
50.2
24.7
50.0
5.8
48.6
3.2
49.8
26.5
49.3
50.6
49.8
14.5
49.5
49.4
50.3
0.0
48.6
50.5
49.7
49.1
49.7
49.5
50.3
40k dataset
Training
Test
Error
Error
28.2
50.2
25.3
49.5
6.3
49.4
3.4
50.5
33.2
49.9
49.4
49.8
0.0
50.1
50.2
50.3
0.0
36.0
49.8
49.7
49.4
49.7
49.7
49.8
80k dataset
Training
Test
Error
Error
30.2
49.6
25.6
49.0
6.9
49.9
3.5
49.1
27.2
50.1
50.2
49.8
0.0
44.6
49.7
50.3
0.0
12.4
50.3
49.7
50.1
49.7
50.3
49.7
Struct. MLP with Hints
0.21
0
0
30.7
3.1
0.01
Table 1: The error percentages with different learning algorithms on Pentomino dataset with
different number of training examples.
Experiment 1-Onehot representation without transformations: In this experiment several trials have been done with a 10-input one-hot vector per patch. Each input corresponds to an object category given in clear, i.e., the ideal input for P2NN if a supervised
P1NN perfectly did its job.
Experiment 2-Disentangled representations: In this experiment, we did trials with 16
binary inputs per patch, 10 one-hot bits for representing each object category, 4 for
rotations and 2 for scaling, i.e., the whole information about the input is given, but it is
perfectly disentangled. This would be the ideal input for P2NN if an unsupervised P1NN
perfectly did its job.
Experiment 3-Onehot representation with transformations: For each of the ten object
types there are 8 = 4×2 possible transformations. Two objects in two different patches
are the considered “the same” (for the final task) if their category is the same regardless
of the transformations. The one-hot representation of a patch corresponds to the crossproduct between the 10 object shape classes and the 4×2 transformations, i.e., one out of
80=10×4 × 2 possibilities represented in an 80-bit one-hot vector. This also contains all
the information about the input image patch, but spread out in a kind of non-parametric
and non-informative (not disentangled) way, like a perfect memory-based unsupervised
learner (like clustering) could produce. Nevertheless, the shape class would be easier to
read out from this representation than from the image representation (it would be an OR
over 8 of the bits).
Experiment 4-Onehot representation with 80 choices: This representation has the same
1 of 80 one-hot representation per patch but the target task is defined differently. Two objects in two different patches are considered the same iff they have exactly the same 80-bit
onehot representation (i.e., are of the same object category with the same transformation
applied).
The first experiment is a sanity check. It was conducted with single hidden-layered MLP’s
with rectifier and tanh nonlinearity, and the task was learned perfectly (0 error on both training
and test dataset) with very few training epochs.
18
0.6
0.6
Training Error Rate
Test Error Rate
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0.00
100
200
300
400
500
600
700
800
Training Error Rate
Test Error Rate
0.00
900
(a) Training and Test Errors for Experiment 4
100
200
300
400
500
600
700
800
(b) Training and Test Errors for Experiment 3
Figure 11: Tanh MLP training curves. Left (a): The training and test errors of Experiment
3 over 800 training epochs with 100k training examples using Tanh MLP. Right
(b):The training and test errors of Experiment 4 over 700 training epochs with 100k
training examples using Tanh MLP.
The results of Experiment 2 are given in Table 2. To improve results, we experimented
with the Maxout non-linearity in a feedforward MLP (Goodfellow et al., 2013) with two hidden
layers. Unlike the typical Maxout network mentioned in the original paper, regularizers have
been deliberately avoided in order to focus on the optimization issue, i.e: no weight decay, norm
constraint on the weights, or dropout. Although learning from a disentangled representation is
more difficult than learning from perfect object detectors, it is feasible with some architectures
such as the Maxout network. Note that this representation is the kind of representation that
one could hope an unsupervised learning algorithm could discover, at best, as argued in Bengio
et al. (2012).
The only results obtained on the validation set for Experiment 3 and Experiment 4 are
shown respectively in Table 3 and Table 4. In these experiments a tanh MLP with two hidden
layers have been tested with the same hyperparameters. In experiment 3 the complexity of the
problem comes from the transformations (8=4×2) and the number of object types.
But in experiment 4, the only source of complexity of the task comes from the number of
different object types. These results are in between the complete failure and complete success
observed with other experiments, suggesting that the task could become solvable with better
training or more training examples. Figure 11 illustrates the progress of training a tanh MLP,
on both the training and test error, for Experiments 3 and 4. Clearly, something has been
learned, but the task is not nailed yet. On experiment 3 for both maxout and tanh the maxout
there was a long plateau where the training error and objective stays almost same. Maxout
did just chance on the experiment for about 120 iterations on the training and the test set.
But after 120th iteration the training and test error started decline and eventually it was able
to solve the task. Moreover as seen from the curves in Figure 11(a) and 11(b), the training
and test error curves are almost the same for both tasks. This implies that for onehot inputs,
whether you increase the number of possible transformations for each object or the number of
19
Learning Algorithm
SVM
Random Forests
Tanh MLP
Maxout MLP
Training Error
0.0
1.29
0.0
0.0
Test Error
35.6
40.475
0.0
0.0
Table 2: Performance of different learning algorithms on disentangled representation in Experiment 2.
Learning Algorithm
SVM
Random Forests
Tanh MLP
Maxout MLP
Training Error
11.212
24.839
0.0
0.0
Test Error
32.37
48.915
22.475
0.0
Table 3: Performance of different learning algorithms using a dataset with onehot vector and
80 inputs as discussed for Experiment 3.
object categories, as soon as the number of possible configurations is same, the complexity of
the problem is almost the same for the MLP.
3.4 Does the Effect Persist with Larger Training Set Sizes?
The results shown in this section indicate that the problem in the Pentomino task clearly
is not just a regularization problem, but rather basically hinges on an optimization problem.
Otherwise, we would expect test error to decrease as the number of training examples increases.
This is shown first by studying the online case and then by studying the ordinary training
case with a fixed size training set but considering increasing training set sizes. In the online
minibatch setting, parameter updates are performed as follows:
θt+1 = θt − ∆θt
PN
∆θt =
i
∇θt L(xt , θt )
N
(5)
(6)
where L(xt , θt ) is the loss incurred on example xt with parameters θt , where t ∈ Z + and
is the learning rate.
Ordinary batch algorithms converge linearly to the optimum θ∗ , however the noisy gradient
estimates in the online SGD will cause parameter θ to fluctuate near the local optima. However,
online SGD directly optimizes the expected risk, because the examples are drawn iid from the
ground-truth distribution (Bottou, 2010). Thus:
Z
L∞ = E[L(x, θ)] = L(x, θ)p(x)dx
(7)
x
20
Learning Algorithm
SVM
Random Forests
Tanh MLP
Training Error
4.346
23.456
0
Test Error
40.545
47.345
25.8
Table 4: Performance of different algorithms using a dataset with onehot vector and 80 binary
inputs as discussed in Experiment 4.
where L∞ is the generalization error. Therefore online SGD is trying to minimize the
expected risk with noisy updates. Those noisy updates have the effect of regularizer:
PN
∆ θt =
i
∇θt L(xt , θt )
= ∇θt L(x, θt ) + ξt
N
(8)
where ∇θt L(x, θt ) is the true gradient and ξt is the zero-mean stochastic gradient “noise”
due to computing the gradient over a finite-size minibatch sample.
We would like to know if the problem with the Pentomino dataset is more a regularization
or an optimization problem. An SMLP-nohints model was trained by online SGD with the randomly generated online Pentomino stream. The learning rate was adaptive, with the Adadelta
procedure (Zeiler, 2012) on minibatches of 100 examples. In the online SGD experiments, two
SMLP-nohints that is trained with and without standardization at the intermediate layer with
exactly the same hyperparameters are tested. The SMLP-nohints P1NN patch-wise submodel
has 2048 hidden units and the SMLP intermediate layer has 1152 = 64 × 18 hidden units. The
nonlinearity that is used for the intermediate layer is the sigmoid. P2NN has 2048 hidden units.
SMLP-nohints has been trained either with or without standardization on top of the output
units of the P1NN. The experiments illustrated in Figures 12 and 13 are with the same SMLP
without hints architecture for which results are given in Table 1. In those graphs only the results
for the training on the randomly generated 545400 Pentomino samples have been presented. As
shown in the plots SMLP-nohints was not able to generalize without standardization. Although
without standardization the training loss seems to decrease initially, it eventually gets stuck in
a plateau where training loss doesn’t change much.
Training of SMLP-nohints online minibatch SGD is performed using standardization in the
intermediate layer and Adadelta learning rate adaptation, on 1046000 training examples from
the randomly generated Pentomino stream. At the end of the training, test error is down to
27.5%, which is much better than chance but from from the score obtained with SMLP-hints
of near 0 error.
In another SMLP-nohints experiment without standardization the model is trained with the
1580000 Pentomino examples using online minibatch SGD. P1NN has 2048 hidden units and 16
sigmoidal outputs per patch. for the P1NN hidden layer. P2NN has 1024 hidden units for the
hidden layer. Adadelta is used to adapt the learning rate. At the end of training this SMLP,
the test error remained stuck, at 50.1%.
21
0.55
0.50
Test Error
0.45
0.40
0.35
0.30
0.25
0.200
SMLP with standardization
SMLP without standardization
100
200
300
400
500
Batch no
600
700
800
Figure 12: Test errors of SMLP-nohints with and without standardization in the intermediate
layer. Sigmoid as an intermediate layer activation has been used. Each tick (batch
no) in the x-axis represents 400 examples.
22
7
SMLP with standardization
SMLP without standardization
6
Training Loss
5
4
3
2
1
00
100
200
300
400
Batch no
500
600
700
Figure 13: Training errors of SMLP-nohints with and without standardization in the intermediate layer. Sigmoid nonlinearity has been used as an intermediate layer activation
function. The x-axis is in units of blocks of 400 examples in the training set.
23
3.4.1 Experiments with Increased Training Set Size
Here we consider the effect of training different learners with different numbers of training
examples. For the experimental results shown in Table 1, 3 training set sizes (20k, 40k and 80k
examples) had been used. Each dataset was generated with different random seeds (so they
do not overlap). Figure 14 also shows the error bars for an ordinary MLP with three hidden
layers, for a larger range of training set sizes, between 40k and 320k examples. The number of
training epochs is 8 (more did not help), and there are three hidden layers with 2048 feature
detectors. The learning rate we used in our experiments is 0.01. The activation function of the
MLP is a tanh nonlinearity, while the L1, L2 penalty coefficients are both 1e-6.
Table 1 shows that, without guiding hints, none of the state-of-art learning algorithms could
perform noticeably better than a random predictor on the test set. This shows the importance
of intermediate hints introduced in the SMLP. The decision trees and SVMs can overfit the
training set but they could not generalize on the test set. Note that the numbers reported in the
table are for hyper-parameters selected based on validation set error, hence lower training errors
are possible if avoiding all regularization and taking large enough models. On the training set,
the MLP with two large hidden layers (several thousands) could reach nearly 0% training error,
but still did not manage to achieve good test error.
In the experiment results shown in Figure 14, we evaluate the impact of adding more training
data for the fully-connected MLP. As mentioned before for these experiments we have used a
MLP with three hidden layers where each layer has 2048 hidden units. The tanh(·) activation
function is used with 0.05 learning rate and minibatches of size 200.
As can be seen from the figure, adding more training examples did not help either training
or test error (both are near 50%, with training error slightly lower and test error slightly
higher), reinforcing the hypothesis that the difficult encountered is one of optimization, not of
regularization.
Figure 14: Training and test error bar charts for a regular MLP with 3 hidden layers. There
is no significant improvement on the generalization error of the MLP as the new
training examples are introduced.
24
3.5 Experiments on Effect of Initializing with Hints
Initialization of the parameters in a neural network can have a big impact on the learning
and generalization (Glorot and Bengio, 2010). Previously Erhan et al. (2010) showed that
initializing the parameters of a neural network with unsupervised pretraining guides the learning towards basins of attraction of local minima that provides better generalization from the
training dataset. In this section we analyze the effect of initializing the SMLP with hints and
then continuing without hints at the rest of the training. For experimental analysis of hints
based initialization, SMLP is trained for 1 training epoch using the hints and for 60 epochs it
is trained without hints on the 40k examples training set. We also compared the same architecture with the same hyperparameters, against to SMLP-nohints trained for 61 iterations on
the same dataset. After one iteration of hint-based training SMLP obtained 9% training error
and 39% test error. Following the hint based training, SMLP is trained without hints for 60
epochs, but at epoch 18, it already got 0% training and 0% test error. The hyperparameters
for this experiment and the experiment that the results shown for the SMLP-hints in Table 1
are the same. The test results for initialization with and without hints are shown on Figure 15.
This figure suggests that initializing with hints can give the same generalization performance
but training takes longer.
Figure 15: Plots showing the test error of SMLP with random initialization vs initializing with
hint based training.
3.5.1 Further Experiments on Optimization for Pentomino Dataset
With extensive hyperparameter optimization and using standardization in the intermediate
level of the SMLP with softmax nonlinearity, SMLP-nohints was able to get 5.3% training and
25
6.7% test error on the 80k Pentomino training dataset. We used the 2050 hidden units for the
hidden layer of P1NN and 11 softmax output per patch. For the P2NN, we used 1024 hidden
units with sigmoid and learning rate 0.1 without using any adaptive learning rate method. This
SMLP uses a rectifier nonlinearity for hidden layers of both P1NN and P2NN. Considering that
architecture uses softmax as the intermediate activation function of SMLP-nohints. It is very
likely that P1NN is trying to learn the presence of specific Pentomino shape in a given patch.
This architecture has a very large capacity in the P1NN, that probably provides it enough
capacity to learn the presence of Pentomino shapes at each patch effortlessly.
An MLP with 2 hidden layers, each 1024 rectifier units, was trained using LBFGS (the
implementation from the scipy.optimize library) on 40k training examples, with gradients computed on batches of 10000 examples at each iteration. However, after convergence of training,
the MLP was still doing chance on the test dataset.
We also observed that using linear units for the intermediate layer yields better generalization error without standardization compared to using activation functions such as sigmoid,
tanh and RELU for the intermediate layer. SMLP-nohints was able to get 25% generalization error with linear units without standardization whereas all the other activation functions
that has been tested failed to generalize with the same number of training iterations without
standardization and hints. This suggests that using non-linear intermediate-level activation
functions without standardization introduces an optimization difficulty for the SMLP-nohints,
maybe because the intermediate level acts like a bottleneck in this architecture.
4. Conclusion and Discussion
In this paper we have shown an example of task which seems almost impossible to solve by
standard black-box machine learning algorithms, but can be almost perfectly solved when
one encourages a semantics for the intermediate-level representation that is guided by prior
knowledge. The task has the particularity that it is defined by the composition of two nonlinear sub-tasks (object detection on one hand, and a non-linear logical operation similar to
XOR on the other hand).
What is interesting is that in the case of the neural network, we can compare two networks
with exactly the same architecture but a different pre-training, one of which uses the known
intermediate concepts to teach an intermediate representation to the network. With enough
capacity and training time they can overfit but did not not capture the essence of the task, as
seen by test set performance.
We know that a structured deep network can learn the task, if it is initialized in the right
place, and do it from very few training examples. Furthermore we have shown that if one
pre-trains SMLP with hints for only one epoch, it can nail the task. But the exactly same
architecture which started training from random initialization, failed to generalize.
Consider the fact that even SMLP-nohints with standardization after being trained using
online SGD on 1046000 generated examples and still gets 27.5% test error. This is an indication
that the problem is not a regularization problem but possibly an inability to find a good effective
local minima of generalization error.
What we hypothesize is that for most initializations and architectures (in particular the
fully-connected ones), although it is possible to find a good effective local minimum of training
error when enough capacity is provided, it is difficult (without the proper initialization) to find a
good local minimum of generalization error. On the other hand, when the network architecture
26
is constrained enough but still allows it to represent a good solution (such as the structured
MLP of our experiments), it seems that the optimization problem can still be difficult and even
training error remains stuck high if the standardization isn’t used. Standardization obviously
makes the training objective of the SMLP easier to optimize and helps it to find at least a
better effective local minimum of training error. This finding suggests that by using specific
architectural constraints and sometimes domain specific knowledge about the problem, one can
alleviate the optimization difficulty that generic neural network architectures face.
It could be that the combination of the network architecture and training procedure produces a training dynamics that tends to yield into these minima that are poor from the point of
view of generalization error, even when they manage to nail training error by providing enough
capacity. Of course, as the number of examples increases, we would expect this discrepancy to
decrease, but then the optimization problem could still make the task unfeasible in practice.
Note however that our preliminary experiments with increasing the training set size (8-fold) for
MLPs did not reveal signs of potential improvements in test error yet, as shown in Figure 14.
Even using online training on 545400 Pentomino examples, the SMLP-nohints architecture was
still doing far from perfect in terms of generalization error (Figure 12).
These findings bring supporting evidence to the “Guided Learning Hypothesis” and “Deeper
Harder Hypothesis” from Bengio (2013a): higher level abstractions, which are expressed by
composing simpler concepts, are more difficult to learn (with the learner often getting in an
effective local minimum ), but that difficulty can be overcome if another agent provides hints
of the importance of learning other, intermediate-level abstractions which are relevant to the
task.
Many interesting questions remain open. Would a network without any guiding hint eventually find the solution with a enough training time and/or with alternate parametrizations?
To what extent is ill-conditioning a core issue? The results with LBFGS were disappointing but
changes in the architectures (such as standardization of the intermediate level) seem to make
training much easier. Clearly, one can reach good solutions from an appropriate initialization,
pointing in the direction of an issue with local minima, but it may be that good solutions are
also reachable from other initializations, albeit going through a tortuous ill-conditioned path
in parameter space. Why did our attempts at learning the intermediate concepts in an unsupervised way fail? Are these results specific to the task we are testing or a limitation of the
unsupervised feature learning algorithm tested? Trying with many more unsupervised variants and exploring explanatory hypotheses for the observed failures could help us answer that.
Finally, and most ambitious, can we solve these kinds of problems if we allow a community
of learners to collaborate and collectively discover and combine partial solutions in order to
obtain solutions to more abstract tasks like the one presented here? Indeed, we would like to
discover learning algorithms that can solve such tasks without the use of prior knowledge as
specific and strong as the one used in the SMLP here. These experiments could be inspired by
and inform us about potential mechanisms for collective learning through cultural evolutions
in human societies.
Acknowledgments
We would like to thank to the ICLR 2013 reviewers for their insightful comments, and NSERC,
CIFAR, Compute Canada and Canada Research Chairs for funding.
27
References
A. Ben-Hur and J. Weston. A user’s guide to support vector machines. Methods in Molecular
Biology, 609:223–239, 2010.
Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle. Greedy layer-wise training of deep
networks. In NIPS’2006, 2007.
Yoshua Bengio. Learning deep architectures for AI. Foundations and Trends in Machine
Learning, 2(1):1–127, 2009. Also published as a book. Now Publishers, 2009.
Yoshua Bengio. Evolving culture vs local minima. In Growing Adaptive Machines: Integrating
Development and Learning in Artificial Neural Networks, number also as ArXiv 1203.2990v1,
pages T. Kowaliw, N. Bredeche & R. Doursat, eds. Springer-Verlag, March 2013a. URL
http://arxiv.org/abs/1203.2990.
Yoshua Bengio. Practical recommendations for gradient-based training of deep architectures.
In K.-R. Müller, G. Montavon, and G. B. Orr, editors, Neural Networks: Tricks of the Trade.
Springer, 2013b.
Yoshua Bengio, Jerome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning.
In Léon Bottou and Michael Littman, editors, Proceedings of the Twenty-sixth International
Conference on Machine Learning (ICML’09). ACM, 2009a.
Yoshua Bengio, Jerome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning.
In ICML’09, 2009b.
Yoshua Bengio, Aaron Courville, and Pascal Vincent. Unsupervised feature learning and deep
learning: A review and new perspectives. Technical Report arXiv:1206.5538, U. Montreal,
2012. URL http://arxiv.org/abs/1206.5538.
Yoshua Bengio, Aaron Courville, and Pascal Vincent. Unsupervised feature learning and deep
learning: A review and new perspectives. IEEE Trans. Pattern Analysis and Machine Intelligence (PAMI), 2013.
James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: a CPU
and GPU math expression compiler. In Proceedings of the Python for Scientific Computing
Conference (SciPy), 2010.
Léon Bottou. Large-scale machine learning with stochastic gradient descent. In Proceedings of
COMPSTAT’2010, pages 177–186. Springer, 2010.
Leo Breiman. Random forests. Machine Learning, 45(1):5–32, 2001.
D. C. Ciresan, U. Meier, L. M. Gambardella, and J. Schmidhuber. Deep big simple neural nets
for handwritten digit recognition. Neural Computation, 22:1–14, 2010.
Yann Dauphin and Yoshua Bengio. Big neural networks waste capacity. Technical Report
arXiv:1301.3583, Universite de Montreal, 2013.
Richard Dawkins. The Selfish Gene. Oxford University Press, 1976.
28
J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and
stochastic optimization. Journal of Machine Learning Research, 12:2121–2159, 2010.
Dumitru Erhan, Yoshua Bengio, Aaron Courville, Pierre-Antoine Manzagol, Pascal Vincent,
and Samy Bengio. Why does unsupervised pre-training help deep learning? Journal of
Machine Learning Research, 11:625–660, February 2010.
François Fleuret, Ting Li, Charles Dubout, Emma K Wampler, Steven Yantis, and Donald
Geman. Comparing machines and humans on a visual categorization test. Proceedings of the
National Academy of Sciences, 108(43):17621–17625, 2011.
X. Glorot, A. Bordes, and Y. Bengio. Deep sparse rectifier neural networks. In AISTATS,
2011a.
Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward
neural networks. In JMLR W&CP: Proceedings of the Thirteenth International Conference
on Artificial Intelligence and Statistics (AISTATS 2010), volume 9, pages 249–256, May
2010.
Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In
JMLR W&CP: Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics (AISTATS 2011), April 2011b.
Ian J. Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio.
Maxout networks. In ICML, 2013.
Maciej Henneberg. Decrease of human skull size in the holocene. Human biology, pages 395–405,
1988.
Maciej Henneberg and Maryna Steyn. Trends in cranial capacity and cranial index in subsaharan africa during the holocene. American journal of human biology, 5(4):473–479, 1993.
J. Henrich and R. McElreath. The evolution of cultural evolution. Evolutionary Anthropology:
Issues, News, and Reviews, 12(3):123–135, 2003.
Geoffrey E. Hinton, Simon Osindero, and Yee Whye Teh. A fast learning algorithm for deep
belief nets. Neural Computation, 18:1527–1554, 2006.
Geoffrey E. Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Improving neural networks by preventing co-adaptation of feature detectors. Technical
report, arXiv:1207.0580, 2012.
C.W. Hsu, C.C. Chang, C.J. Lin, et al. A practical guide to support vector classification, 2003.
Kevin Jarrett, Koray Kavukcuoglu, Marc’Aurelio Ranzato, and Yann LeCun. What is the
best multi-stage architecture for object recognition? In Proc. International Conference on
Computer Vision (ICCV’09), pages 2146–2153. IEEE, 2009a.
Kevin Jarrett, Koray Kavukcuoglu, Marc’Aurelio Ranzato, and Yann LeCun. What is the best
multi-stage architecture for object recognition? In ICCV’09, 2009b.
29
Faisal Khan, Xiaojin Zhu, and Bilge Mutlu. How do humans teach: On curriculum learning and
teaching dimension. In Advances in Neural Information Processing Systems 24 (NIPS’11),
pages 1449–1457, 2011.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton. ImageNet classification with deep
convolutional neural networks. In Advances in Neural Information Processing Systems 25
(NIPS’2012). 2012.
Kai A. Krueger and Peter Dayan. Flexible shaping: how learning in small steps helps. Cognition,
110:380–394, 2009.
G. Kunapuli, K.P. Bennett, R. Maclin, and J.W. Shavlik. The adviceptron: Giving advice to
the perceptron. Proceedings of the Conference on Artificial Neural Networks In Engineering
(ANNIE 2010), 2010.
Hugo Larochelle, Yoshua Bengio, Jerome Louradour, and Pascal Lamblin. Exploring strategies
for training deep neural networks. Journal of Machine Learning Research, 10:1–40, 2009.
Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document
recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
T.M. Mitchell. The need for biases in learning generalizations. Department of Computer
Science, Laboratory for Computer Science Research, Rutgers Univ., 1980.
T.M. Mitchell and S.B. Thrun. Explanation-based neural network learning for robot control.
Advances in Neural information processing systems, pages 287–287, 1993.
R. Montague. Universal grammar. Theoria, 36(3):373–398, 1970.
V. Nair and G. E Hinton. Rectified linear units improve restricted Boltzmann machines. In
ICML’10, 2010.
L.B.J.H.F.R.A. Olshen and C.J. Stone. Classification and regression trees. Belmont, Calif.:
Wadsworth, 1984.
Joseph O’Sullivan. Integrating initialization bias and search bias in neural network learning,
1996.
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel,
P. Prettenhofer, R. Weiss, V. Dubourg, et al. Scikit-learn: Machine learning in python. The
Journal of Machine Learning Research, 12:2825–2830, 2011.
Gail B. Peterson. A day of great illumination: B. F. Skinner’s discovery of shaping. Journal of
the Experimental Analysis of Behavior, 82(3):317–328, 2004.
Tapani Raiko, Harri Valpola, and Yann LeCun. Deep learning made easier by linear transformations in perceptrons. In International Conference on Artificial Intelligence and Statistics,
pages 924–932, 2012.
Salah Rifai, Pascal Vincent, Xavier Muller, Xavier Glorot, and Yoshua Bengio. Contractive
auto-encoders: Explicit invariance during feature extraction. In ICML’2011, 2011.
30
Salah Rifai, Yoshua Bengio, Yann Dauphin, and Pascal Vincent. A generative process for sampling contractive auto-encoders. In Proceedings of the Twenty-nine International Conference
on Machine Learning (ICML’12). ACM, 2012. URL http://icml.cc/discuss/2012/590.
html.
R. Salakhutdinov and G.E. Hinton. Deep Boltzmann machines. In Proceedings of the Twelfth
International Conference on Artificial Intelligence and Statistics (AISTATS 2009), volume 8,
2009.
Burrhus F. Skinner. Reinforcement today. American Psychologist, 13:94–99, 1958.
R.J. Solomonoff. A system for incremental learning based on algorithmic probability. In Proceedings of the Sixth Israeli Conference on Artificial Intelligence, Computer Vision and Pattern
Recognition, pages 515–527. Citeseer, 1989.
G.G. Towell and J.W. Shavlik. Knowledge-based artificial neural networks. Artificial intelligence, 70(1):119–165, 1994.
Tommi Vatanen, Tapani Raiko, Harri Valpola, and Yann LeCun. Pushing stochastic gradient
towards second-order methods–backpropagation learning with transformations in nonlinearities. arXiv preprint arXiv:1301.3476, 2013.
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. Journal of Machine Learning Research, 11:3371–3408, December
2010.
Luis Von Ahn, Manuel Blum, Nicholas J Hopper, and John Langford. Captcha: Using hard
ai problems for security. In Advances in CryptologyEUROCRYPT 2003, pages 294–311.
Springer, 2003.
Jason Weston, Frédéric Ratle, and Ronan Collobert. Deep learning via semi-supervised embedding. In William W. Cohen, Andrew McCallum, and Sam T. Roweis, editors, Proceedings of the Twenty-fifth International Conference on Machine Learning (ICML’08),
pages 1168–1175, New York, NY, USA, 2008. ACM. ISBN 978-1-60558-205-4. doi:
10.1145/1390156.1390303.
Matthew D Zeiler.
Adadelta:
arXiv:1212.5701, 2012.
An adaptive learning rate method.
arXiv preprint
5. Appendix
5.1 Binary-Binary RBMs on Pentomino Dataset
We trained binary-binary RBMs (both visible and hidden are binary) on 8×8 patches extracted
from the Pentomino Dataset using PCD (stochastic maximum likelihood), a weight decay of
.0001 and a sparsity penalty9 . We used 256 hidden units and trained by SGD with a batch
size of 32 and a annealing learning rate (Bengio, 2013b) starting from 1e-3 with annealing rate
9. implemented as TorontoSparsity in pylearn2, see the yaml file in the repository for more details
31
1.000015. The RBM is trained with momentum starting from 0.5. The biases are initialized to
-2 in order to get a sparse representation. The RBM is trained for 120 epochs (approximately
50 million updates).
After pretraining the RBM, its parameters are used to initialize the first layer of an SMLPnohints network. As in the usual architecture of the SMLP-nohints on top of P1NN, there is
an intermediate layer. Both P1NN and the intermediate layer have a sigmoid nonlinearity, and
the intermediate layer has 11 units per location. This SMLP-nohints is trained with Adadelta
and standardization at the intermediate layer 10 .
RBM Test and Training Errors
0.5
Training Error
Test Error
Error percentage
0.4
0.3
0.2
0.1
0.00
5
10
15
20
25
Epoch
30
35
40
45
Figure 16: Training and test errors of an SMLP-nohints network whose first layer is pre-trained
as an RBM. Training error reduces to 0% at epoch 42, but test error is still chance.
5.2 Experimental Setup and Hyper-parameters
5.2.1 Decision Trees
We used the decision tree implementation in the scikit-learn (Pedregosa et al., 2011) python
package which is an implementation of the CART (Regression Trees) algorithm. The CART
algorithm constructs the decision tree recursively and partitions the input space such that the
samples belonging to the same category are grouped together (Olshen and Stone, 1984). We
used The Gini index as the impurity criteria. We evaluated the hyper-parameter configurations with a grid-search. We cross-validated the maximum depth (max depth) of the tree (for
preventing the algorithm to severely overfit the training set) and minimum number of samples
10. In our auto-encoder experiments we directly fed features to P2NN without standardization and Adadelta.
32
Figure 17: Filters learned by the binary-binary RBM after training on the 40k examples. The
RBM did learn the edge structure of Pentomino shapes.
Figure 18: 100 samples generated from trained RBM. All the generated samples are valid Pentomino shapes.
33
required to create a split (min split). 20 different configurations of hyper-parameter values were
evaluated. We obtained the best validation error with max depth = 300 and min split = 8.
5.2.2 Support Vector Machines
We used the “Support Vector Classifier (SVC)” implementation from the scikit-learn package
which in turn uses the libsvm’s Support Vector Machine (SVM) implementation. Kernelbased SVMs are non-parametric models that map the data into a high dimensional space and
separate different classes with hyperplane(s) such that the support vectors for each category
will be separated by a large margin. We cross-validated three hyper-parameters of the model
using grid-search: C, γ and the type of kernel(kernel type). C is the penalty term (weight
decay) for the SVM and γ is a hyper-parameter that controls the width of the Gaussian for
the RBF kernel. For the polynomial kernel, γ controls the flexibility of the classifier (degree
of the polynomial) as the number of parameters increases (Hsu et al., 2003; Ben-Hur and
Weston, 2010). We evaluated forty-two hyper-parameter configurations. That includes, two
kernel types: {RBF, P olynomial}; three gammas: {1e − 2, 1e − 3, 1e − 4} for the RBF kernel,
{1, 2, 5} for the polynomial kernel, and seven C values among: {0.1, 1, 2, 4, 8, 10, 16}. As a
result of the grid search and cross-validation, we have obtained the best test error by using the
RBF kernel, with C = 2 and γ = 1.
5.2.3 Multi Layer Perceptron
We have our own implementation of Multi Layer Perceptron based on the Theano (Bergstra
et al., 2010) machine learning libraries. We have selected 2 hidden layers, the rectifier activation
function, and 2048 hidden units per layer. We cross-validated three hyper-parameters of the
model using random-search, sampling the learning rates in log-domain, and selecting L1
and L2 regularization penalty coefficients in sets of fixed values, evaluating 64 hyperparameter
values. The range of the hyperparameter values are ∈ [0.0001, 1], L1 ∈ {0., 1e−6, 1e−5, 1e−4}
and L2 ∈ {0, 1e − 6, 1e − 5}. As a result, the following were selected: L1 = 1e − 6, L2 = 1e − 5
and = 0.05.
5.2.4 Random Forests
We used scikit-learn’s implementation of “Random Forests” decision tree learning. The Random Forests algorithm creates an ensemble of decision trees by randomly selecting for each tree
a subset of features and applying bagging to combine the individual decision trees (Breiman,
2001). We have used grid-search and cross-validated the max depth, min split, and number
of trees (n estimators). We have done the grid-search on the following hyperparameter values, n estimators ∈ {5, 10, 15, 25, 50}, max depth ∈ {100, 300, 600, 900}, and min splits ∈
{1, 4, 16}. We obtained the best validation error with max depth = 300, min split = 4 and
n estimators = 10.
5.2.5 k-Nearest Neighbors
We used scikit-learn’s implementation of k-Nearest Neighbors (k-NN). k-NN is an instancebased, lazy learning algorithm that selects the training examples closest in Euclidean distance
to the input query. It assigns a class label to the test example based on the categories of the
k closest neighbors. The hyper-parameters we have evaluated in the cross-validation are the
number of neighbors (k) and weights. The weights hyper-parameter can be either “uniform” or
34
“distance”. With “uniform”, the value assigned to the query point is computed by the majority
vote of the nearest neighbors. With “distance”, each value assigned to the query point is
computed by weighted majority votes where the weights are computed with the inverse distance
between the query point and the neighbors. We have used n neighbours ∈ {1, 2, 4, 6, 8, 12} and
weights ∈ {”unif orm”, ”distance”} for hyper-parameter search. As a result of cross-validation
and grid search, we obtained the best validation error with k = 2 and weights=“uniform”.
5.2.6 Convolutional Neural Nets
We used a Theano (Bergstra et al., 2010) implementation of Convolutional Neural Networks
(CNN) from the deep learning tutorial at deeplearning.net, which is based on a vanilla
version of a CNN LeCun et al. (1998). Our CNN has two convolutional layers. Following
each convolutional layer, we have a max-pooling layer. On top of the convolution-poolingconvolution-pooling layers there is an MLP with one hidden layer. In the cross-validation we
have sampled 36 learning rates in log-domain in the range [0.0001, 1] and the number of filters
from the range [10, 20, 30, 40, 50, 60] uniformly. For the first convolutional layer we used 9×9
receptive fields in order to guarantee that each object fits inside the receptive field. As a result
of random hyperparameter search and doing manual hyperparameter search on the validation
dataset, the following values were selected:
• The number of features used for the first layer is 30 and the second layer is 60.
• For the second convolutional layer, 7×7 receptive fields. The stride for both convolutional
layers is 1.
• Convolved images are downsampled by a factor of 2×2 at each pooling operation.
• The learning rate for CNN is 0.01 and it was trained for 8 epochs.
5.2.7 Maxout Convolutional Neural Nets
We used the pylearn2 (https://github.com/lisa-lab/pylearn2) implementation of maxout
convolutional networks (Goodfellow et al., 2013). There are two convolutional layers in the
selected architecture, without any pooling. In the last convolutional layer, there is a maxout
non-linearity. The following were selected by cross-validation: learning rate, number of channels
for the both convolution layers, number of kernels for the second layer and number of units and
pieces per maxout unit in the last layer, a linearly decaying learning rate, momentum starting
from 0.5 and saturating to 0.8 at the 200’th epoch. Random search for the hyperparameters
was used to evaluate 48 different hyperparameter configurations on the validation dataset. For
the first convolutional layer, 8×8 kernels were selected to make sure that each Pentomino shape
fits into the kernel. Early stopping was used and test error on the model that has the best
validation error is reported. Using norm constraint on the fan-in of the final softmax units
yields slightly better result on the validation dataset.
As a result of cross-validation and manually tuning the hyperparameters we used the following hyperparameters:
• 16 channels per convolutional layer. 600 hidden units for the maxout layer.
• 6x6 kernels for the second convolutional layer.
35
• 5 pieces for the convolution layers and 4 pieces for the maxout layer per maxout units.
• We decayed the learning rate by the factor of 0.001 and the initial learning rate is 0.026367.
But we scaled the learning rate of the second convolutional layer by a constant factor of
0.6.
• The norm constraint (on the incoming weights of each unit) is 1.9365.
Figure 19 shows the first layer filters of the maxout convolutional net, after being trained
on the 80k training set for 85 epochs.
Figure 19: Maxout convolutional net first layer filters. Most of the filters were able to learn
the basic edge structure of the Pentomino shapes.
5.2.8 Stacked Denoising Auto-Encoders
Denoising Auto-Encoders (DAE) are a form of regularized auto-encoder (Bengio et al., 2013).
The DAE forces the hidden layer to discover more robust features and prevents it from simply
learning the identity by reconstructing the input from a corrupted version of it (Vincent et al.,
2010). Two DAEs were stacked, resulting in an unsupervised transformation with two hidden
layers of 1024 units each. Parameters of all layers are then fine-tuned with supervised finetuning using logistic regression as the classifier and SGD as the gradient-based optimization
algorithm. The stochastic corruption process is binomial (0 or 1 replacing each input value,
with probability 0.2). The selected learning rate is 0 = 0.01 for the DAe and 1 = 0.1 for
supervised fine-tuning. Both L1 and L2 penalty for the DAEs and for the logistic regression
layer are set to 1e-6.
CAE+MLP with Supervised Finetuning: A regularized auto-encoder which sometimes
outperforms the DAE is the Contractive Auto-Encoder (CAE), (Rifai et al., 2012), which
penalizes the Frobenius norm of the Jacobian matrix of derivatives of the hidden units with
respect to the CAE’s inputs. The CAE serves as pre-training for an MLP, and in the supervised
fine-tuning state, the Adagrad method was used to automatically tune the learning rate (Duchi
et al., 2010).
After training a CAE with 100 sigmoidal units patch-wise, the features extracted on each
patch are concatenated and fed as input to an MLP. The selected Jacobian penalty coefficient
is 2, the learning rate for pre-training is 0.082 with batch size of 200 and 200 epochs of unsupervised learning are performed on the training set. For supervised finetuning, the learning
rate is 0.12 over 100 epochs, L1 and L2 regularization penalty terms respectively are 1e-4 and
1e-6, and the top-level MLP has 6400 hidden units.
36
Greedy Layerwise CAE+DAE Supervised Finetuning: For this experiment we stack a
CAE with sigmoid non-linearities and then a DAE with rectifier non-linearities during the pretraining phase. As recommended by Glorot et al. (2011b) we have used a softplus nonlinearity
for reconstruction, sof tplus(x) = log(1 + ex ). We used an L1 penalty on the rectifier outputs
to obtain a sparser representation with rectifier non-linearity and L2 regularization to keep the
non-zero weights small.
The main difference between the DAE and CAE is that the DAE yields more robust reconstruction whereas the CAE obtains more robust features (Rifai et al., 2011).
As seen on Figure 7 the weights U and V are shared on each patch and we concatenate the
outputs of the last auto-encoder on each patch to feed it as an input to an MLP with a large
hidden layer.
We used 400 hidden units for the CAE and 100 hidden units for DAE. The learning rate
used for the CAE is 0.82 and for DAE it is 9*1e-3. The corruption level for the DAE (binomial
noise) is 0.25 and the contraction level for the CAE is 2.0. The L1 regularization penalty for
the DAE is 2.25*1e-4 and the L2 penalty is 9.5*1e-5. For the supervised finetuning phase the
learning rate used is 4*1e-4 with L1 and L2 penalties respectively 1e-5 and 1e-6. The top-level
MLP has 6400 hidden units. The auto-encoders are each trained for 150 epochs while the whole
MLP is fine-tuned for 50 epochs.
Greedy Layerwise DAE+DAE Supervised Finetuning: For this architecture, we have
trained two layers of denoising auto-encoders greedily and performed supervised finetuning
after unsupervised pre-training. The motivation for using two denoising auto-encoders is the
fact that rectifier nonlinearities work well with the deep networks but it is difficult to train
CAEs with the rectifier non-linearity. We have used the same type of denoising auto-encoder
that is used for the greedy layerwise CAE+DAE supervised finetuning experiment.
In this experiment we have used 400 hidden units for the first layer DAE and 100 hidden
units for the second layer DAE. The other hyperparameters for DAE and supervised finetuning
are the same as with the CAE+DAE MLP Supervised Finetuning experiment.
37
| 9cs.NE
|
QSGD: Communication-Efficient SGD
via Gradient Quantization and Encoding
arXiv:1610.02132v4 [cs.LG] 6 Dec 2017
Dan Alistarh
IST Austria & ETH Zurich
[email protected]
Demjan Grubic
ETH Zurich & Google
[email protected]
Ryota Tomioka
Microsoft Research
[email protected]
Jerry Z. Li
MIT
[email protected]
Milan Vojnovic
London School of Economics
[email protected]
Abstract
Parallel implementations of stochastic gradient descent (SGD) have received significant research attention, thanks to its excellent scalability properties. A fundamental
barrier when parallelizing SGD is the high bandwidth cost of communicating gradient updates between nodes; consequently, several lossy compresion heuristics have
been proposed, by which nodes only communicate quantized gradients. Although
effective in practice, these heuristics do not always converge.
In this paper, we propose Quantized SGD (QSGD), a family of compression
schemes with convergence guarantees and good practical performance. QSGD
allows the user to smoothly trade off communication bandwidth and convergence
time: nodes can adjust the number of bits sent per iteration, at the cost of possibly
higher variance. We show that this trade-off is inherent, in the sense that improving
it past some threshold would violate information-theoretic lower bounds. QSGD
guarantees convergence for convex and non-convex objectives, under asynchrony,
and can be extended to stochastic variance-reduced techniques.
When applied to training deep neural networks for image classification and automated speech recognition, QSGD leads to significant reductions in end-to-end
training time. For instance, on 16GPUs, we can train the ResNet-152 network to
full accuracy on ImageNet 1.8× faster than the full-precision variant.
1
Introduction
The surge of massive data has led to significant interest in distributed algorithms for scaling computations in the context of machine learning and optimization. In this context, much attention has
been devoted to scaling large-scale stochastic gradient descent (SGD) algorithms [33], which can be
briefly defined as follows. Let f : Rn → R be a function which we want to minimize. We have access
to stochastic gradients ge such that E[e
g (x)] = ∇f (x). A standard instance of SGD will converge
towards the minimum by iterating the procedure
xt+1 = xt − ηt ge(xt ),
(1)
where xt is the current candidate, and ηt is a variable step-size parameter. Notably, this arises if
we are given i.i.d. data points X1 , . . . , Xm generated from an unknown distribution D, and a loss
function `(X, θ), which measures the loss of the model θ at data point X. We wish to find a model
θ∗ which minimizes f (θ) = EX∼D [`(X, θ)], the expected loss to the data. This framework captures
many fundamental tasks, such as neural network training.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
In this paper, we focus on parallel SGD methods, which have received considerable attention recently
due to their high scalability [6, 8, 13, 32]. Specifically, we consider a setting where a large dataset is
partitioned among K processors, which collectively minimize a function f . Each processor maintains
a local copy of the parameter vector xt ; in each iteration, it obtains a new stochastic gradient update
(corresponding to its local data). Processors then broadcast their gradient updates to their peers, and
aggregate the gradients to compute the new iterate xt+1 .
In most current implementations of parallel SGD, in each iteration, each processor must communicate
its entire gradient update to all other processors. If the gradient vector is dense, each processor will
need to send and receive n floating-point numbers per iteration to/from each peer to communicate
the gradients and maintain the parameter vector x. In practical applications, communicating the
gradients in each iteration has been observed to be a significant performance bottleneck [8, 35, 37].
One popular way to reduce this cost has been to perform lossy compression of the gradients [1, 3,
10, 11, 41]. A simple implementation is to simply reduce precision of the representation, which has
been shown to converge under convexity and sparsity assumptions [10]. A more drastic quantization
technique is 1BitSGD [35, 37], which reduces each component of the gradient to just its sign
(one bit), scaled by the average over the coordinates of ge, accumulating errors locally. 1BitSGD
was experimentally observed to preserve convergence [35], under certain conditions; thanks to the
reduction in communication, it enabled state-of-the-art scaling of deep neural networks (DNNs) for
acoustic modelling [37]. However, it is currently not known if 1BitSGD provides any guarantees,
even under strong assumptions, and it is not clear if higher compression is achievable.
Contributions. Our focus is understanding the trade-offs between the communication cost of dataparallel SGD, and its convergence guarantees. We propose a family of algorithms allowing for lossy
compression of gradients called Quantized SGD (QSGD), by which processors can trade-off the
number of bits communicated per iteration with the variance added to the process.
QSGD is built on two algorithmic ideas. The first is an intuitive stochastic quantization scheme:
given the gradient vector at a processor, we quantize each component by randomized rounding to a
discrete set of values, in a principled way which preserves the statistical properties of the original.
The second step is an efficient lossless code for quantized gradients, which exploits their statistical
properties to generate efficient encodings. Our analysis gives tight bounds on the precision-variance
trade-off induced by QSGD.
√
At one extreme of this trade-off, we can guarantee that each processor transmits
√ at most n(log n +
O(1)) expected bits per iteration, while increasing variance by at most a n multiplicative factor.
At the other extreme, we show that each processor can transmit ≤ 2.8n + 32 bits per iteration in
expectation, while increasing variance by a only a factor of 2. In particular, in the latter regime,
compared to full precision SGD, we use ≈ 2.8n bits of communication per iteration as opposed to
32n bits, and guarantee at most 2× more iterations, leading to bandwidth savings of ≈ 5.7×.
QSGD is fairly general: it can also be shown to converge, under assumptions, to local minima for nonconvex objectives, as well as under asynchronous iterations. One non-trivial extension we develop
is a stochastic variance-reduced [23] variant of QSGD, called QSVRG, which has exponential
convergence rate.
One key question is whether QSGD’s compression-variance trade-off is inherent: for instance, does
any algorithm guaranteeing at most constant variance blowup need to transmit Ω(n) bits per iteration?
The answer is positive: improving asymptotically upon this trade-off would break the communication
complexity lower bound of distributed mean estimation (see [44, Proposition 2] and [38]).
Experiments. The crucial question is whether, in practice, QSGD can reduce communication cost
by enough to offset the overhead of any additional iterations to convergence. The answer is yes.
We explore the practicality of QSGD on a variety of state-of-the-art datasets and machine learning
models: we examine its performance in training networks for image classification tasks (AlexNet,
Inception, ResNet, and VGG) on the ImageNet [12] and CIFAR-10 [25] datasets, as well as on
LSTMs [19] for speech recognition. We implement QSGD in Microsoft CNTK [3].
Experiments show that all these models can significantly benefit from reduced communication when
doing multi-GPU training, with virtually no accuracy loss, and under standard parameters. For example, when training AlexNet on 16 GPUs with standard parameters, the reduction in communication
time is 4×, and the reduction in training to the network’s top accuracy is 2.5×. When training an
LSTM on two GPUs, the reduction in communication time is 6.8×, while the reduction in training
2
time to the same target accuracy is 2.7×. Further, even computationally-heavy architectures such as
Inception and ResNet can benefit from the reduction in communication: on 16GPUs, QSGD reduces
the end-to-end convergence time of ResNet152 by approximately 2×. Networks trained with QSGD
can converge to virtually the same accuracy as full-precision variants, and that gradient quantization
may even slightly improve accuracy in some settings.
Related Work. One line of related research studies the communication complexity of convex
optimization. In particular, [40] studied two-processor convex minimization in the same model,
provided a lower bound of Ω(n(log n + log(1/))) bits on the communication cost of n-dimensional
convex problems, and proposed a non-stochastic algorithm for strongly convex problems, whose
communication cost is within a log factor of the lower bound. By contrast, our focus is on stochastic
gradient methods. Recent work [5] focused on round complexity lower bounds on the number of
communication rounds necessary for convex learning.
Buckwild! [10] was the first to consider the convergence guarantees of low-precision SGD. It gave
upper bounds on the error probability of SGD, assuming unbiased stochastic quantization, convexity,
and gradient sparsity, and showed significant speedup when solving convex problems on CPUs.
QSGD refines these results by focusing on the trade-off between communication and convergence.
We view quantization as an independent source of variance for SGD, which allows us to employ
standard convergence results [7]. The main differences from Buckwild! are that 1) we focus on the
variance-precision trade-off; 2) our results apply to the quantized non-convex case; 3) we validate
the practicality of our scheme on neural network training on GPUs. Concurrent work proposes
TernGrad [41], which starts from a similar stochastic quantization, but focuses on the case where
individual gradient components can have only three possible values. They show that significant
speedups can be achieved on TensorFlow [1], while maintaining accuracy within a few percentage
points relative to full precision. The main differences to our work are: 1) our implementation
guarantees convergence under standard assumptions; 2) we strive to provide a black-box compression
technique, with no additional hyperparameters to tune; 3) experimentally, QSGD maintains the same
accuracy within the same target number of epochs; for this, we allow gradients to have larger bit
width; 4) our experiments focus on the single-machine multi-GPU case.
We note that QSGD can be applied to solve the distributed mean estimation problem [24, 38] with an
optimal error-communication trade-off in some regimes. In contrast to the elegant random rotation
solution presented in [38], QSGD employs quantization and Elias coding. Our use case is different
from the federated learning application of [24, 38], and has the advantage of being more efficient to
compute on a GPU.
There is an extremely rich area studying algorithms and systems for efficient distributed large-scale
learning, e.g. [1, 3, 6, 10, 11, 21, 32, 39, 43]. Significant interest has recently been dedicated to
quantized frameworks, both for inference, e.g., [1, 17] and training [10, 16, 20, 35, 37, 42, 45]. In
this context, [35] proposed 1BitSGD, a heuristic for compressing gradients in SGD, inspired by
delta-sigma modulation [34]. It is implemented in Microsoft CNTK, and has a cost of n bits and two
floats per iteration. Variants of it were shown to perform well on large-scale Amazon datasets by [37].
Compared to 1BitSGD, QSGD can achieve asymptotically higher compression, provably converges
under standard assumptions, and shows superior practical performance in some cases.
2
Preliminaries
SGD has many variants, with different preconditions and guarantees. Our techniques are rather
portable, and can usually be applied in a black-box fashion on top of SGD. For conciseness, we will
focus on a basic SGD setup. The following assumptions are standard; see e.g. [7].
Let X ⊆ Rn be a known convex set, and let f : X → R be differentiable, convex, smooth, and
unknown. We assume repeated access to stochastic gradients of f , which on (possibly random) input
x, outputs a direction which is in expectation the correct direction to move in. Formally:
Definition 2.1. Fix f : X → R. A stochastic gradient for f is a random function ge(x) so that
E[e
g (x)] = ∇f (x). We say the stochastic gradient has second moment at most B if E[ke
g k22 ] ≤ B for
2
2
2
all x ∈ X . We say it has variance at most σ if E[ke
g (x) − ∇f (x)k2 ] ≤ σ for all x ∈ X .
Observe that any stochastic gradient with second moment bound B is automatically also a stochastic
gradient with variance bound σ 2 = B, since E[ke
g (x) − ∇f (x)k2 ] ≤ E[ke
g (x)k2 ] as long as
E[e
g (x)] = ∇f (x). Second, in convex optimization, one often assumes a second moment bound
3
Data: Local copy of the parameter vector x
1 for each iteration t do
2
Let g
eti be an independent stochastic gradient ;
i
i
M ← Encode(e
g (x)) //encode gradients ;
3
broadcast M i to all peers;
for each peer ` do
receive M ` from peer `;
4
5
6
`
`
g
b ← Decode(M ) //decode gradients ;
7
8
9
10 end
Figure 1: An illustration of generalized
stochastic quantization with 5 levels.
end
P
xt+1 ← xt − (ηt /K) K
b` ;
`=1 g
Algorithm 1: Parallel SGD Algorithm.
when dealing with non-smooth convex optimization, and a variance bound when dealing with smooth
convex optimization. However, for us it will be convenient to consistently assume a second moment
bound. This does not seem to be a major distinction in theory or in practice [7].
Given access to stochastic gradients, and a starting point x0 , SGD builds iterates xt given by Equation
(1), projected onto X , where (ηt )t≥0 is a sequence of step sizes. In this setting, one can show:
Theorem 2.1 ([7], Theorem 6.3). Let X ⊆ Rn be convex, and let f : X → R be unknown, convex,
and L-smooth. Let x0 ∈ X be given, and let R2 = supx∈X kx − x0 k2 . Let T > 0 be fixed. Given
repeated, independent access to stochastic gradients with variance
bound σ 2 for f , SGD with initial
q
point x0 and constant step sizes ηt =
"
E f
T
1X
xt
T t=0
1
L+1/γ ,
where γ =
R
σ
!#
r
− min f (x) ≤ R
x∈X
2
T
, achieves
LR2
2σ 2
+
.
T
T
(2)
Minibatched SGD. A modification to the SGD scheme presented above often observed in practice
is a technique known as minibatching. In minibatched SGD, updates are of the form xt+1 =
e t (xt ) = 1 Pm get,i , and where each get,i is an independent stochastic
e t (xt )), where G
ΠX (xt − ηt G
i=1
m
gradient for f at xt . It is not hard to see that if get,i are stochastic gradients with variance bound σ 2 ,
e t is a stochastic gradient with variance bound σ 2 /m. By inspection of Theorem 2.1, as
then the G
long as the first term in (2) dominates, minibatched SGD requires 1/m fewer iterations to converge.
Data-Parallel SGD. We consider synchronous data-parallel SGD, modelling real-world multi-GPU
systems, and focus on the communication cost of SGD in this setting. We have a set of K processors
p1 , p2 , . . . , pK who proceed in synchronous steps, and communicate using point-to-point messages.
Each processor maintains a local copy of a vector x of dimension n, representing the current estimate
of the minimizer, and has access to private, independent stochastic gradients for f .
In each synchronous iteration, described in Algorithm 1, each processor aggregates the value of x,
then obtains random gradient updates for each component of x, then communicates these updates
to all peers, and finally aggregates the received updates and applies them locally. Importantly, we
add encoding and decoding steps for the gradients before and after send/receive in lines 3 and 7,
respectively. In the following, whenever describing a variant of SGD, we assume the above general
pattern, and only specify the encode/decode functions. Notice that the decoding step does not
necessarily recover the original gradient ge` ; instead, we usually apply an approximate version.
When the encoding and decoding steps are the identity (i.e., no encoding / decoding), we shall refer
to this algorithm as parallel SGD. In this case, it is a simple calculation to see that at each processor,
if xt was the value of x that the processors held before iteration t, then the updated value of x by the
PK
end of this iteration is xt+1 = xt − (ηt /K) `=1 ge` (xt ), where each ge` is a stochatic gradient. In
particular, this update is merely a minibatched update of size K. Thus, by the discussion above, and
by rephrasing Theorem 2.1, we have the following corollary:
Corollary 2.2. Let X , f, L, x0 , and R be as in Theorem 2.1. Fix > 0. Suppose we run parallel
SGD on K processors, each with access to independent stochastic gradients with second moment
4
√
bound B, with step size ηt = 1/(L + K/γ), where γ is as in Theorem 2.1. Then if
"
!#
T
1X
2B L
2
,
, then E f
xt
T = O R · max
− min f (x) ≤ .
x∈X
K2
T t=0
(3)
In most reasonable regimes, the first term of the max in (3) will dominate the number of iterations
necessary. Specifically, the number of iterations will depend linearly on the second moment bound B.
3
Quantized Stochastic Gradient Descent (QSGD)
In this section, we present our main results on stochastically quantized SGD. Throughout, log denotes
the base-2 logarithm, and the number of bits to represent a float is 32. For any vector v ∈ Rn , we
let kvk0 denote the number of nonzeros of v. For any string ω ∈ {0, 1}∗ , we will let |ω| denote its
length. For any scalar x ∈ R, we let sgn (x) ∈ {−1, +1} denote its sign, with sgn (0) = 1.
3.1
Generalized Stochastic Quantization and Coding
Stochastic Quantization. We now consider a general, parametrizable lossy-compression scheme
for stochastic gradient vectors. The quantization function is denoted with Qs (v), where s ≥ 1 is
a tuning parameter, corresponding to the number of quantization levels we implement. Intuitively,
we define s uniformly distributed levels between 0 and 1, to which each value is quantized in a way
which preserves the value in expectation, and introduces minimal variance. Please see Figure 1.
For any v ∈ Rn with v 6= 0, Qs (v) is defined as
Qs (vi ) = kvk2 · sgn (vi ) · ξi (v, s) ,
(4)
where ξi (v, s)’s are independent random variables defined as follows. Let 0 ≤ ` < s be an
integer such that |vi |/kvk2 ∈ [`/s, (` + 1)/s]. That is, [`/s, (` + 1)/s] is the quantization interval
corresponding to |vi |/kvk2 . Then
(
|vi |
`/s
with probability 1 − p kvk
,
s
;
2
ξi (v, s) =
(` + 1)/s otherwise.
Here, p(a, s) = as − ` for any a ∈ [0, 1]. If v = 0, then we define Q(v, s) = 0.
The distribution of ξi (v, s) has minimal variance over distributions with support {0, 1/s, . . . , 1}, and
its expectation satisfies E[ξi (v, s)] = |vi |/kvk2 . Formally, we can show:
n
Lemma 3.1. For any
√ vector v ∈ R , we have that (i) E[Qs (v)] = v (unbiasedness),√(ii) E[kQs (v)−
vk22 ] ≤ min(n/s2 , n/s)kvk22 (variance bound), and (iii) E[kQs (v)k0 ] ≤ s(s + n) (sparsity).
Efficient Coding of Gradients. Observe that for any vector v, the output of Qs (v) is naturally
expressible by a tuple (kvk2 , σ, ζ), where σ is the vector of signs of the vi ’s and ζ is the vector
of integer values s · ξi (v, s). The key idea behind the coding scheme is that not all integer values
s · ξi (v, s) can be equally likely: in particular, larger integers are less frequent. We will exploit this
via a specialized Elias integer encoding [14].
Intuitively, for any positive integer k, its code, denoted Elias(k), starts from the binary representation
of k, to which it prepends the length of this representation. It then recursively encodes this prefix.
We show that for any positive integer k, the length of the resulting code has |Elias(k)| = log k +
log log k + . . . + 1 ≤ (1 + o(1)) log k + 1, and that encoding and decoding can be done efficiently.
Given a gradient vector represented as the triple (kvk2 , σ, ζ), with s quantization levels, our coding
outputs a string S defined as follows. First, it uses 32 bits to encode kvk2 . It proceeds to encode
using Elias recursive coding the position of the first nonzero entry of ζ. It then appends a bit denoting
σi and follows that with Elias(s · ξi (v, s)). Iteratively, it proceeds to encode the distance from the
current coordinate of ζ to the next nonzero, and encodes the σi and ζi for that coordinate in the
same way. The decoding scheme is straightforward: we first read off 32 bits to construct kvk2 , then
iteratively use the decoding scheme for Elias recursive coding to read off the positions and values of
the nonzeros of ζ and σ. The properties of the quantization and of the encoding imply the following.
Theorem 3.2. Let f : Rn → R be fixed, and let x ∈ Rn be arbitrary. Fix s ≥ 2 quantization
levels. If ge(x) is a stochastic gradient for f at x with second moment bound B, then Qs (e
g (x)) is a
5
stochastic gradient for f at x with variance bound min
√
n
n
B.
s2 , s
Moreover, there is an encoding
scheme so that in expectation, the number of bits to communicate Qs (e
g (x)) is upper bounded by
2
√
3
2(s + n)
√
3+
s(s + n) + 32.
+ o(1) log
2
s(s + n)
Sparse
levels 0, 1, and −1, the gradient density is
√ Regime. For the case s = 1, i.e., quantization
√
O(√n), while the second-moment blowup is ≤ n. Intuitively, this means√that we will employ
O( n log n) bits per iteration, while the convergence time is increased by O( n).
√
Dense Regime. The variance blowup is minimized to at most 2 for s = n quantization levels; in
this case, we devise a more efficient encoding which yields an order of magnitude shorter codes
compared to the full-precision variant. The proof of this statement is not entirely obvious, as it
exploits both the statistical properties of the quantization and the guarantees of the Elias coding.
Corollary 3.3. Let f, x, and ge(x) be as in Theorem 3.2. There is an encoding scheme for Q√n (e
g (x))
which in expectation has length at most 2.8n + 32.
3.2
QSGD Guarantees
Putting the bounds on the communication and variance given above with the guarantees for SGD
algorithms on smooth, convex functions yield the following results:
Theorem 3.4 (Smooth Convex QSGD). Let X , f, L, x0 , and R be as in Theorem 2.1. Fix > 0.
Suppose we run parallel QSGD with s quantization levels on K processors accessing indepen√
dent stochastic gradients with second moment bound B, with stepsize ηt = 1/(L + K/γ),
√
where γ is as in Theorem 2.1 with σ = B 0 , where B 0 = min sn2 , sn B. Then if T =
0
h P
i
T
L
2B
O R2 · max K
, then E f T1 t=0 xt − minx∈X f (x) ≤ . Moreover, QSGD re2,
2
√
+n)
√
quires 3 + 23 + o(1) log 2(s
(s2 + n) + 32 bits of communication per round. In the
2
s + n
√
special case when s = n, this can be reduced to 2.8n + 32.
QSGD is quite portable, and can be applied to almost any stochastic gradient method. For illustration,
we can use quantization along with [15] to get communication-efficient non-convex SGD.
Theorem 3.5 (QSGD for smooth non-convex optimization). Let f : Rn → R be a L-smooth
(possibly nonconvex) function, and let x1 be an arbitrary initial point. Let T > 0 be fixed, and
s > 0. Then there is a random stopping time R supported on {1, . . . , N } so that QSGD with
quantization level s, constant stepsizes η = O(1/L) and access
√ to stochastic gradients√of f with
L(f (x1 )−f ∗ )
min(n/s2 , n/s)B
1
second moment bound B satisfies L E k∇f (x)k22 ≤ O
+
.
N
L
Moreover, the communication cost is the same as in Theorem 3.4.
3.3
Quantized Variance-Reduced SGD
Assume we are given K processors, and a parameter m > 0, where each processor i P
has access to
m
1
functions {fim/K , . . . , f(i+1)m/K−1 }. The goal is to approximately minimize f = m
i=1 fi . For
P
P
(i+1)m/K−1
K
1
processor i, let hi = m j=im/K
fi be the portion of f that it knows, so that f = i=1 hi .
A natural question is whether we can apply stochastic quantization to reduce communication for
parallel SVRG. Upon inspection, we notice that the resulting update will break standard SVRG. We
resolve this technical issue, proving one can quantize SVRG updates using our techniques and still
obtain the same convergence bounds.
√
e
= Q(v, n), where Q(v, s) is defined as in Section 3.1. Given
Algorithm Description. Let Q(v)
arbitrary starting point x0 , we let y (1) = x0 . At the beginning of epoch p, each processor broadcasts ∇hi (y (p)
P),mthat is, the unquantized full gradient, from which the processors each aggregate
∇f (y (p) ) = i=1 ∇hi (y (p) ). Within each epoch, for each iteration t = 1, . . . , T , and for each
(p)
processor i = 1, . . . , K, we let ji,t be a uniformly random integer from [m] completely independent
from everything
else. Then, in iteration t in epochp, processor i broadcasts the update vector
(p)
(p)
e
ut,i = Q ∇fj (p) (xt ) − ∇fj (p) (y (p) ) + ∇f (y (p) ) .
i,t
i,t
6
Table 1: Description of networks, final top-1 accuracy, as well as end-to-end training speedup on 8GPUs.
Network
AlexNet
ResNet152
ResNet50
ResNet110
BN-Inception
VGG19
LSTM
Dataset
ImageNet
ImageNet
ImageNet
CIFAR-10
ImageNet
ImageNet
AN4
Params.
62M
60M
25M
1M
11M
143M
13M
Init. Rate
0.07
1
1
0.1
3.6
0.1
0.5
Top-1 (32bit)
59.50%
77.0%
74.68%
93.86%
81.13%
Top-1 (QSGD)
60.05% (4bit)
76.74% (8bit)
74.76% (4bit)
94.19% (4bit)
81.15 % (4bit)
Speedup (8 GPUs)
2.05 ×
1.56 ×
1.26 ×
1.10 ×
1.16× (projected)
2.25× (projected)
2× (2 GPUs)
PK
(p)
(p)
(p)
(p)
1
Each processor then computes the total update ut = K
i=1 ut,i , and sets xt+1 = xt − ηut .
P
(p)
T
At the end of epoch p, each processor sets y (p+1) = T1 t=1 xt . We can prove the following.
P
m
1
Theorem 3.6. Let f (x) = m
i=1 fi (x), where f is `-strongly convex, and fi are convex and
∗
L-smooth, for all i. Let x be the unique minimizer of f over Rn . Then, if η = O(1/L)
and T = O(L/`), then QSVRG with initial point y (1) ensures E f (y (p+1) ) − f (x∗ ) ≤
0.9p f (y (1) ) − f (x∗ ) , for any epoch p ≥ 1. Moreover, QSVRG with T iterations per epoch
requires ≤ (F + 2.8n)(T + 1) + F n bits of communication per epoch.
Discussion. In particular, this allows us to largely decouple the dependence between F and the
condition number of f in the communication. Let κ = L/` denote the condition number of f . Observe
that whenever F κ, the second term is subsumed by the first and the per epoch communication
is dominated by (F + 2.8n)(T + 1). Specifically, for any fixed , to attain accuracy we must
take F = O(log 1/). As long as log 1/ ≥ Ω(κ), which is true for instance in the case when
κ ≥ poly log(n) and ≥ poly(1/n), then the communication per epoch is O(κ(log 1/ + n)).
Gradient Descent. The full version of the paper [4] contains an application of QSGD to gradient
descent. Roughly, in this case, QSGD can simply truncate the gradient to its top components, sorted
by magnitude.
4
QSGD Variants
Our experiments will stretch the theory, as we use deep networks, with non-convex objectives. (We
have also tested QSGD for convex objectives. Results closely follow the theory, and are therefore
omitted.) Our implementations will depart from the previous algorithm description as follows.
First, we notice that the we can control the variance the quantization by quantizing into buckets
of a fixed size d. If we view each gradient as a one-dimensional vector v, reshaping tensors if
necessary, a bucket will be defined as a set of d consecutive vector values. (E.g. the ith bucket is the
sub-vector v[(i − 1)d + 1 : i · d].) We will quantize each bucket independently, using QSGD. Setting
d = 1 corresponds to no quantization (vanilla SGD), and d = n corresponds to full quantization,
as described in the previous section. It is easy to see that, using bucketing, the guarantees from
Lemma 3.1 will be expressed in terms of d, as opposed to the full dimension n. This provides a
knob by which we can control variance, at the cost of storing an extra scaling factor on every d
bucket values. As an example, if we use a bucket √
size of 512, and 4 bits, the variance increase
due to quantization will be upper bounded by only 512/24 ' 1.41. This provides a theoretical
justification for the similar convergence rates we observe in practice.
The second difference from the theory is that we will scale by the maximum value of the vector (as
opposed to the 2-norm). Intuitively, normalizing by the max preserves more values, and has slightly
higher accuracy for the same number of iterations. Both methods have the same baseline bandwidth
reduction because of lower bit width (e.g. 32 bits to 2 bits per dimension), but normalizing by the
max no longer provides any
√ sparsity guarantees. We note that this does not affect our bounds in the
regime where we use Θ( n) quantization levels per component, as we employ no sparsity in that
case. (However, we note that in practice max normalization also generates non-trivial sparsity.)
5
Experiments
Setup. We performed experiments on Amazon EC2 p2.16xlarge instances, with 16 NVIDIA K80
GPUs. Instances have GPUDirect peer-to-peer communication, but do not currently support NVIDIA
7
2.3x
1.6x
3.5x
Figure 2: Breakdown of communication versus computation for various neural networks, on 2, 4, 8, 16 GPUs,
2.0
1.5
Training loss
> 2x faster
2bit QSGD (d=128)
4bit QSGD (d=8192)
8bit QSGD (d=8192)
SGD
1.0
0.5
0.00
(a) AlexNet Accuracy versus Time.
300
600
900
Time (sec)
1200
(b) LSTM error vs Time.
1500
Test accuracy (%)
for full 32-bit precision versus QSGD 4-bit. Each bar represents the total time for an epoch under standard
parameters. Epoch time is broken down into communication (bottom, solid) and computation (top, transparent).
Although epoch time diminishes as we parallelize, the proportion of communication increases.
80
70
60
50
40
30
20
10
00
1bitSGD*
32bit
QSGD 4bit
QSGD 8bit
20
40
60
Epoch
80
100
120
(c) ResNet50 Accuracy.
Figure 3: Accuracy numbers for different networks. Light blue lines represent 32-bit accuracy.
NCCL extensions. We have implemented QSGD on GPUs using the Microsoft Cognitive Toolkit
(CNTK) [3]. This package provides efficient (MPI-based) GPU-to-GPU communication, and implements an optimized version of 1bit-SGD [35]. Our code is released as open-source [31].
We execute two types of tasks: image classification on ILSVRC 2015 (ImageNet) [12], CIFAR10 [25], and MNIST [27], and speech recognition on the CMU AN4 dataset [2]. For vision, we
experimented with AlexNet [26], VGG [36], ResNet [18], and Inception with Batch Normalization [22] deep networks. For speech, we trained an LSTM network [19]. See Table 1 for details.
Protocol. Our methodology emphasizes zero error tolerance, in the sense that we always aim to
preserve the accuracy of the networks trained. We used standard sizes for the networks, with hyperparameters optimized for the 32bit precision variant. (Unless otherwise stated, we use the default
networks and hyper-parameters optimized for full-precision CNTK 2.0.) We increased batch size
when necessary to balance communication and computation for larger GPU counts, but never past the
point where we lose accuracy. We employed double buffering [35] to perform communication and
quantization concurrently with the computation. Quantization usually benefits from lowering learning
rates; yet, we always run the 32bit learning rate, and decrease bucket size to reduce variance. We will
not quantize small gradient matrices (< 10K elements), since the computational cost of quantizing
them significantly exceeds the reduction in communication. However, in all experiments, more than
99% of all parameters are transmitted in quantized form. We reshape matrices to fit bucket sizes, so
that no receptive field is split across two buckets.
Communication vs. Computation. In the first set of experiments, we examine the ratio between
computation and communication costs during training, for increased parallelism. The image classification networks are trained on ImageNet, while LSTM is trained on AN4. We examine the cost
breakdown for these networks over a pass over the dataset (epoch). Figure 2 gives the results for
various networks for image classification. The variance of epoch times is practically negligible (<1%),
hence we omit confidence intervals.
Figure 2 leads to some interesting observations. First, based on the ratio of communication to
computation, we can roughly split networks into communication-intensive (AlexNet, VGG, LSTM),
and computation-intensive (Inception, ResNet). For both network types, the relative impact of
communication increases significantly as we increase the number of GPUs. Examining the breakdown
for the 32-bit version, all networks could significantly benefit from reduced communication. For
8
example, for AlexNet on 16 GPUs with batch size 1024, more than 80% of training time is spent on
communication, whereas for LSTM on 2 GPUs with batch size 256, the ratio is 71%. (These ratios
can be slightly changed by increasing batch size, but this can decrease accuracy, see e.g. [21].)
Next, we examine the impact of QSGD on communication and overall training time. (Communication
time includes time spent compressing and uncompressing gradients.) We measured QSGD with
2-bit quantization and 128 bucket size, and 4-bit and 8-bit quantization with 512 bucket size. The
results for these two variants are similar, since the different bucket sizes mean that the 4bit version
only sends 77% more data than the 2-bit version (but ∼ 8× less than 32-bit). These bucket sizes are
chosen to ensure good convergence, but are not carefully tuned.
On 16GPU AlexNet with batch size 1024, 4-bit QSGD reduces communication time by 4×, and
overall epoch time by 2.5×. On LSTM, it reduces communication time by 6.8×, and overall epoch
time by 2.7×. Runtime improvements are non-trivial for all architectures we considered.
Accuracy. We now examine how QSGD influences accuracy and convergence rate. We ran AlexNet
and ResNet to full convergence on ImageNet, LSTM on AN4, ResNet110 on CIFAR-10, as well as
a two-layer perceptron on MNIST. Results are given in Figure 5, and exact numbers are given in
Table 1. QSGD tests are performed on an 8GPU setup, and are compared against the best known
full-precision accuracy of the networks. In general, we notice that 4bit or 8bit gradient quantization
is sufficient to recover or even slightly improve full accuracy, while ensuring non-trivial speedup.
Across all our experiments, 8-bit gradients with 512 bucket size have been sufficient to recover or
improve upon the full-precision accuracy. Our results are consistent with recent work [30] noting
benefits of adding noise to gradients when training deep networks. Thus, quantization can be seen
as a source of zero-mean noise, which happens to render communication more efficient. At the
same time, we note that more aggressive quantization can hurt accuracy. In particular, 4-bit QSGD
with 8192 bucket size (not shown) loses 0.57% for top-5 accuracy, and 0.68% for top-1, versus full
precision on AlexNet when trained for the same number of epochs. Also, QSGD with 2-bit and 64
bucket size has gap 1.73% for top-1, and 1.18% for top-1.
One issue we examined in more detail is which layers are more sensitive to quantization. It appears
that quantizing convolutional layers too aggressively (e.g., 2-bit precision) can lead to accuracy loss
if trained for the same period of time as the full precision variant. However, increasing precision to
4-bit or 8-bit recovers accuracy. This finding suggests that modern architectures for vision tasks, such
as ResNet or Inception, which are almost entirely convolutional, may benefit less from quantization
than recurrent deep networks such as LSTMs.
Additional Experiments. The full version of the paper contains additional experiments, including a
full comparison with 1BitSGD. In brief, QSGD outperforms or matches the performance and final
accuracy of 1BitSGD for the networks and parameter values we consider.
6
Conclusions and Future Work
We have presented QSGD, a family of SGD algorithms which allow a smooth trade off between
the amount of communication per iteration and the running time. Experiments suggest that QSGD
is highly competitive with the full-precision variant on a variety of tasks. There are a number of
optimizations we did not explore. The most significant is leveraging the sparsity created by QSGD.
Current implementations of MPI do not provide support for sparse types, but we plan to explore
such support in future work. Further, we plan to examine the potential of QSGD in larger-scale
applications, such as super-computing. On the theoretical side, it is interesting to consider applications
of quantization beyond SGD.
The full version of this paper [4] contains complete proofs, as well as additional applications.
7
Acknowledgments
The authors would like to thank Martin Jaggi, Ce Zhang, Frank Seide and the CNTK team for their
support during the development of this project, as well as the anonymous NIPS reviewers for their
careful consideration and excellent suggestions. Dan Alistarh was supported by a Swiss National
Fund Ambizione Fellowship. Jerry Li was supported by the NSF CAREER Award CCF-1453261,
CCF-1565235, a Google Faculty Research Award, and an NSF Graduate Research Fellowship. This
work was developed in part while Dan Alistarh, Jerri Li and Milan Vojnovic were with Microsoft
Research Cambridge, UK.
9
References
[1] Martın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro,
Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. Tensorflow: Large-scale
machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467,
2016.
[2] Alex Acero. Acoustical and environmental robustness in automatic speech recognition, volume
201. Springer Science & Business Media, 2012.
[3] Amit Agarwal, Eldar Akchurin, Chris Basoglu, Guoguo Chen, Scott Cyphers, Jasha Droppo,
Adam Eversole, Brian Guenter, Mark Hillebrand, Ryan Hoens, et al. An introduction to
computational networks and the computational network toolkit. Technical report, Tech. Rep.
MSR-TR-2014-112, August 2014., 2014.
[4] Dan Alistarh, Demjan Grubic, Jerry Li, Ryota Tomioka, and Milan Vojnovic. QSGD:
Communication-efficient SGD via gradient quantization and encoding. arXiv preprint
arXiv:1610.02132, 2016.
[5] Yossi Arjevani and Ohad Shamir. Communication complexity of distributed convex learning
and optimization. In NIPS, 2015.
[6] Ron Bekkerman, Mikhail Bilenko, and John Langford. Scaling up machine learning: Parallel
and distributed approaches. Cambridge University Press, 2011.
[7] Sébastien Bubeck. Convex optimization: Algorithms and complexity. Foundations and Trends R
in Machine Learning, 8(3-4):231–357, 2015.
[8] Trishul Chilimbi, Yutaka Suzue, Johnson Apacible, and Karthik Kalyanaraman. Project adam:
Building an efficient and scalable deep learning training system. In OSDI, October 2014.
[9] Cntk brainscript file for alexnet. https://github.com/Microsoft/CNTK/tree/master/
Examples/Image/Classification/AlexNet/BrainScript. Accessed: 2017-02-24.
[10] Christopher M De Sa, Ce Zhang, Kunle Olukotun, and Christopher Ré. Taming the wild: A
unified analysis of hogwild-style algorithms. In NIPS, 2015.
[11] Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew
Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. In NIPS,
2012.
[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale
hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009.
IEEE Conference on, pages 248–255. IEEE, 2009.
[13] John C Duchi, Sorathan Chaturapruek, and Christopher Ré. Asynchronous stochastic convex
optimization. NIPS, 2015.
[14] Peter Elias. Universal codeword sets and representations of the integers. IEEE transactions on
information theory, 21(2):194–203, 1975.
[15] Saeed Ghadimi and Guanghui Lan. Stochastic first- and zeroth-order methods for nonconvex
stochastic programming. SIAM Journal on Optimization, 23(4):2341–2368, 2013.
[16] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning
with limited numerical precision. In ICML, pages 1737–1746, 2015.
[17] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149,
2015.
[18] 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, pages 770–778, 2016.
[19] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation,
9(8):1735–1780, 1997.
[20] Itay Hubara, Matthieu Courbariaux, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized
neural networks. In Advances in Neural Information Processing Systems, pages 4107–4115,
2016.
10
[21] Forrest N Iandola, Matthew W Moskewicz, Khalid Ashraf, and Kurt Keutzer. Firecaffe: nearlinear acceleration of deep neural network training on compute clusters. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, pages 2592–2600, 2016.
[22] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training
by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015.
[23] Rie Johnson and Tong Zhang. Accelerating stochastic gradient descent using predictive variance
reduction. In NIPS, 2013.
[24] Jakub Konečnỳ. Stochastic, distributed and federated optimization for machine learning. arXiv
preprint arXiv:1707.01155, 2017.
[25] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images,
2009.
[26] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep
convolutional neural networks. In Advances in neural information processing systems, pages
1097–1105, 2012.
[27] Yann LeCun, Corinna Cortes, and Christopher JC Burges. The mnist database of handwritten
digits, 1998.
[28] Mu Li, David G Andersen, Jun Woo Park, Alexander J Smola, Amr Ahmed, Vanja Josifovski,
James Long, Eugene J Shekita, and Bor-Yiing Su. Scaling distributed machine learning with
the parameter server. In OSDI, 2014.
[29] Xiangru Lian, Yijun Huang, Yuncheng Li, and Ji Liu. Asynchronous parallel stochastic gradient
for nonconvex optimization. In NIPS. 2015.
[30] Arvind Neelakantan, Luke Vilnis, Quoc V Le, Ilya Sutskever, Lukasz Kaiser, Karol Kurach,
and James Martens. Adding gradient noise improves learning for very deep networks. arXiv
preprint arXiv:1511.06807, 2015.
[31] Cntk implementation of qsgd. https://gitlab.com/demjangrubic/QSGD. Accessed: 201711-4.
[32] Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu. Hogwild: A lock-free
approach to parallelizing stochastic gradient descent. In NIPS, 2011.
[33] Herbert Robbins and Sutton Monro. A stochastic approximation method. The Annals of
Mathematical Statistics, pages 400–407, 1951.
[34] Richard Schreier and Gabor C Temes. Understanding delta-sigma data converters, volume 74.
IEEE Press, Piscataway, NJ, 2005.
[35] Frank Seide, Hao Fu, Jasha Droppo, Gang Li, and Dong Yu. 1-bit stochastic gradient descent
and its application to data-parallel distributed training of speech dnns. In INTERSPEECH, 2014.
[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556, 2014.
[37] Nikko Strom. Scalable distributed DNN training using commodity GPU cloud computing. In
INTERSPEECH, 2015.
[38] Ananda Theertha Suresh, Felix X Yu, H Brendan McMahan, and Sanjiv Kumar. Distributed
mean estimation with limited communication. arXiv preprint arXiv:1611.00429, 2016.
[39] Seiya Tokui, Kenta Oono, Shohei Hido, CA San Mateo, and Justin Clayton. Chainer: a nextgeneration open source framework for deep learning. In Proceedings of workshop on machine
learning systems (LearningSys), 2015.
[40] John N Tsitsiklis and Zhi-Quan Luo. Communication complexity of convex optimization.
Journal of Complexity, 3(3), 1987.
[41] Wei Wen, Cong Xu, Feng Yan, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Terngrad: Ternary gradients to reduce communication in distributed deep learning. arXiv preprint
arXiv:1705.07878, 2017.
[42] Hantian Zhang, Jerry Li, Kaan Kara, Dan Alistarh, Ji Liu, and Ce Zhang. Zipml: Training
linear models with end-to-end low precision, and a little bit of deep learning. In International
Conference on Machine Learning, pages 4035–4043, 2017.
11
[43] Sixin Zhang, Anna E Choromanska, and Yann LeCun. Deep learning with elastic averaging
sgd. In Advances in Neural Information Processing Systems, pages 685–693, 2015.
[44] Yuchen Zhang, John Duchi, Michael I Jordan, and Martin J Wainwright. Information-theoretic
lower bounds for distributed statistical estimation with communication constraints. In NIPS,
2013.
[45] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net:
Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint
arXiv:1606.06160, 2016.
Roadmap of the Appendix
A
Proof of Lemmas and Theorems
A.1
Proof of Lemma 3.1
The first claim obviously holds. We thus turn our attention to the second claim of the lemma. We first
note the following bound:
h
i
2
E[ξi (v, s)2 ] = E[ξi (v, s)]2 + E (ξi (v, s) − E[ξi (v, s)])
1
|vi |
v2
|vi |
= i 2 + 2p
,s
1−p
,s
kvk2
s
kvk2
kvk2
2
v
1
|vi |
≤ i 2 + 2p
,s .
kvk2
s
kvk2
Using this bound, we have
"
2 #
|vi |
E[kQ(v, s)k ] =
E
,s
kvk2
i=1
n
X
|vi |
|vi |2
1
p
,
s
+
≤ kvk22
kvk22
s2
kvk2
i=1
!
n
1 X
|vi |
= 1+ 2
p
,s
kvk22
s i=1
kvk2
a
n kvk1
≤ 1 + min 2 ,
kvk22
s skvk2
√
n
n
≤ 1 + min 2 ,
kvk22 .
s
s
2
n
X
kvk22 ξi
where (a) follows from thefact that p(a, s) ≤ 1 and p(a, s) ≤ as. This immediately implies that
E[kQ(v, s) − vk22 ] ≤ min
A.2
√
n
n
s2 , s
· kvk22 , as claimed.
A Compression Scheme for Qs Matching Theorem 3.2
In this section, we describe a scheme for coding Qs and provide an upper bound for the expected
number of information bits that it uses, which gives the bound in Theorem 3.2.
Observe that for any vector v, the output of Q(v, s) is naturally expressible by a tuple (kvk2 , σ, ζ),
where σ is the vector of signs of the vi ’s and ζ is the vector of ξi (v, s) values. With a slight abuse of
notation, let us consider Q(v, s) as a function from R \ {0} to Bs , where
Bs = {(A, σ, z) ∈ R × Rn × Rn : A ∈ R≥0 , σi ∈ {−1, +1}, zi ∈ {0, 1/s, . . . , 1}} .
12
We define a coding scheme that represents each tuple in Bs with a codeword in {0, 1}∗ according to
a mapping Codes : Bs → {0, 1}∗ .
To encode a single coordinate, we utilize a lossless encoding scheme for positive integers known as
recursive Elias coding or Elias omega coding.
Definition A.1. Let k be a positive integer. The recursive Elias coding of k, denoted Elias(k), is
defined to be the {0, 1} string constructed as follows. First, place a 0 at the end of the string. If k = 0,
then terminate. Otherwise, prepend the binary representation of k to the beginning of the code. Let
k 0 be the number of bits so prepended minus 1, and recursively encode k 0 in the same fashion. To
decode an recursive Elias coded integer, start with N = 1. Recursively, if the next bit is 0, stop, and
output N . Otherwise, if the next bit is 1, then read that bit and N additional bits, and let that number
in binary be the new N , and repeat.
The following are well-known properties of the recursive Elias code which are not too hard to prove.
Lemma A.1. For any positive integer k, we have
1. |Elias(k)| ≤ log k + log log k + log log log k . . . + 1 = (1 + o(1)) log k + 1.
2. The recursive Elias code of k can be encoded and decoded in time O(|Elias(k)|).
3. Moreover, the decoding can be done without previously knowing a bound on the size of k.
Given a tuple (A, σ, z) ∈ Bs , our coding outputs a string S defined as follows. First, it uses F bits to
encode A. It proceeds to encode using Elias recursive coding the position of the first nonzero entry
of z. It then appends a bit denoting σi and follows that with Elias(szi ). Iteratively, it proceeds to
encode the distance from the current coordinate of z to the next nonzero using c, and encodes the σi
and zi for that coordinate in the same way. The decoding scheme is also straightforward: we first
read off F bits to construct A, then iteratively use the decoding scheme for Elias recursive coding to
read off the positions and values of the nonzeros of z and σ.
We can now present a full description of our lossy-compression scheme. For any input vector v, we
first compute quantization Q(v, s), and then encode using Codes . In our notation, this is expressed
as v → Codes (Q(v, s)).
√
Lemma A.2. For any v ∈ Rn and s2 + n ≤ n/2, we have
2
√
3
2(s + n)
√
E[|Codes (Q(v, s))|] ≤ 3 + · (1 + o(1)) log
(s2 + n) .
2
2
s + n
This lemma together with Lemma 3.1 suffices to prove Theorem 3.2.
We first show a technical lemma about the behavior of the coordinate-wise coding function c on a
vector with bounded `p norm.
Lemma A.3. Let q ∈ Rd be a vector so that for all i, we have that qi is a positive integer, and
moreover, kqkpp ≤ ρ. Then
d
ρ
X
1 + o(1)
|Elias(qi )| ≤
log
+1 n.
p
n
i=1
Proof. Recall that for any positive integer k, the length of Elias(k) is at most (1 + o(1)) log k + 1.
Hence, we have
d
d
X
X
|Elias(qi )| ≤ (1 + o(1))
(log qi ) + d
i=1
i=1
n
1 + o(1) X
≤
(log(qip )) + d
p
i=1
!
n
(a) 1 + o(1)
1X p
≤
n log
q
+ +d
p
n i=1 i
ρ
1 + o(1)
≤
n log
+n
p
n
13
where (a) follows from Jensen’s inequality.
We can bound the number of information bits needed for our coding scheme in terms of the number
of non-zeroes of our vector.
Lemma A.4. For any tuple (A, σ, z) ∈ Bs , the string Codes (A, σ, z) has length of at most this
many bits:
2
n
1 + o(1)
s kzk22
F + (1 + o(1)) · log
+
log
+3
· kzk0 .
kzk0
2
kzk0
Proof. First, the float A takes F bits to communicate. Let us now consider the rest of the string. We
break up the string into a couple of parts. First, there is the subsequence S1 dedicated to pointing to
the next nonzero coordinate of z. Second, there is the subsequence S2 dedicated to communicating
the sign and c(zi ) for each nonzero coordinate i. While these two sets of bits are not consecutive
within the string, it is clear that they partition the remaining bits in the string. We bound the length of
these two substrings separately.
We first bound the length of S1 . Let i1 , . . . , ikzk0 be the nonzero coordinates of z. Then, from the
definition of Codes , it is not hard to see that S1 consists of the encoding of the vector
q (1) = (i1 , i2 − i1 , . . . , ikzk0 − ikzk0 −1 ) ,
where each coordinate of this vector is encoded using c. By Lemma A.3, since this vector has length
kzk0 and has `1 norm at most n, we have that
n
|S1 | ≤ (1 + o(1)) log
+ 1 kzk0 .
(5)
kzk0
We now bound the length of S2 . Per non-zero coordinate of z, we need to communicate a sign (which
takes one bit), and c(szi ). Thus by Lemma A.3, we have that
kzk0
|S2 | =
X
(1 + |Elias(szi )|)
j=1
≤ kzk0 +
(1 + o(1))
s2 kzk22
log
+ 1 kzk0 .
2
kzk0
(6)
Putting together (5) and (6) yields the desired conclusion.
We first need the following technical lemma about the number of nonzeros of Q(v, s) that we have in
expectation.
Lemma A.5. Let v ∈ Rn such that kvk2 6= 0. Then
E[kQ(v, s)k0 ] ≤ s2 +
√
n.
Proof. Let u = v/kvk2 . Let I(u) denote the set of coordinates i of u so that ui ≤ 1/s. Since
X
1≥
u2i ≥ (n − |I(u)|)/s2 ,
i6∈I(u)
we must have that s2 ≥ n − |I(u)|. Moreover, for each i ∈ I(u), we have that Qi (v, s) is nonzero
with probability ui , and zero otherwise. Hence
X
√
E[L(v)] ≤ n − |I(u)| +
ui ≤ s2 + kuk1 ≤ s2 + n .
i∈I(u)
14
Proof of Lemma A.2 Let Q(v, s) = (kvk2 , σ, ζ), and let u = v/kvk2 . Observe that we always
have that
2 (a) X
n
n
n
X
X
1
n
1
kζk22 ≤
ui +
=
2
1
+
,
(7)
≤ 2
u2i + 2
s
s2
s2
i=1
i=1
i=1
where (a) follows since (a + b)2 ≤ 2(a2 + b2 ) for all a, b ∈ R.
By Lemma A.4, we now have that
n
E[|Codes (Q(v, s)|] ≤F + (1 + o(1)) E kζk0 log
kζk0
2
1 + o(1)
s R(ζ)
+
E kζk0 log
+ 3 E[kζk0 ]
2
kζk0
n
≤F + (1 + o(1)) E kζk0 log
+
kζk0
!#
"
√
2 s2 + n
1 + o(1)
+ 3 s2 + n ,
E kζk0 log
2
kζk0
by (7) and Lemma A.5.
It is a straightforward verification that the function f (x) = x log Cx is concave for all C > 0.
Moreover, it is increasing up until x = C/2,√
and decreasing afterwards. Hence, by Jensen’s inequality,
Lemma A.5, and the assumption that s2 + n ≤ n/2, we have that
√
n
n
√
E kζk0 log
, and
≤ (s2 + n) log
kζk0
s2 + n
"
!#
2
√
2 s2 + n
2(s + n)
√
E kζk0 log
.
≤ (s2 + n) log
kζk0
s2 + n
Simplifying yields the expression in the Lemma.
A.3
A Compression Scheme for Qs Matching Theorem 3.3
For the case of the quantized SGD scheme that requires Θ(n) bits per iteration, we can improve the
constant factor in the bit length bound in√Theorem A.2 by using a different encoding of Q(v, s). This
corresponds to the regime where s = n, i.e., where the quantized update is not expected to be
sparse. In this case, there is no advantage gained by transmitting the location of the next nonzero,
since generally that will simply be the next coordinate of the vector. Therefore, we may as well
simply transmit the value of each coordinate in sequence.
Motivated by the above remark, we define the following alternative compression function. Define
Elias0 (k) = Elias(k + 1) to be a compression function on all nonnegative natural numbers. It is
easy to see that this is uniquely decodable. Let Code0s be the compression function which, on input
(A, σ, z), simply encodes every coordinate of z in the same way as before, even if it is zero, using
Elias0 . It is straightforward to show that this compression function is still uniquely decodable. Then,
just as before, our full quantization scheme is as follows. For any arbitrary vector v, we first compute
Q(v, s), and then encode using Code0s . In our notation, this is expressed as v → Code0s (Q(v, s)).
For this compression scheme, we show:
Lemma A.6. For any v ∈ Rn , we have
√
1 + o(1)
s2 + min(n, s n)
0
E[|Codes (Q(v, s))|] ≤ F +
log 1 +
+1 +2 n.
2
n
√
In particular, if s = n, then E[|Code0s (Q(v, s))|] ≤ F + 2.8n.
It is not hard to see that this is equivalent to the bound stated in Theorem 3.3.
We start by showing the following lemma.
15
Lemma A.7. For any tuple (A, σ, z) ∈ Bs , the string Code0s (A, σ, z) has length of at most this
many bits:
1 + o(1)
s2 kzk22
F+
log 1 +
+ 1 + 2 n.
2
n
Proof. The proof of this lemma follows by similar arguments as that of Lemma A.4. The main
differences are that (1) we do not need to encode the position of the nonzeros, and (2) we always
encode Elias(k + 1) instead of Elias(k). Hence, for coordinate i, we require 1 + Elias(szi + 1) bits,
since in addition to encoding zi we must also encode the sign. Thus the total number of bits may be
bounded by
F+
n
X
(Elias(zi + 1) + 1) = F + n +
i=1
≤F +n+
n
X
i=1
n
X
Elias(szi + 1)
[(1 + o(1)) log(szi + 1) + 1]
i=1
≤ F + 2n + (1 + o(1))
≤ F + 2n +
(a)
≤ F + 2n +
1 + o(1)
2
1 + o(1)
2
n
X
log(szi + 1)
i=1
n
X
log((szi + 1)2 )
i=1
n
X
(log(1 + s2 zi2 ) + log (2))
i=1
n
(b)
1 + o(1)
1X 2 2
≤ F + 2n +
n log 1 +
s zi
2
n i=1
!
!
+1
where (a) follows from basic properties of logarithms and (b) follows from the concavity of the
function x 7→ log(1 + x) and Jensen’s inequality. Simplifying yields the desired statement.
Proof of Lemma A.6 As in the proof of Lemma A.2, let Q(v, s) = (kvk2 , σ, ζ), and let u =
v/kvk2 . By Lemma A.7, we have
1 + o(1)
s2 R(ζ)
0
E[|Codes (Q(v, s)|] ≤ F +
E log 1 +
+1 +2 n
2
n
!
!
!
(a)
E s2 R(ζ)
1 + o(1)
log 1 +
+1 +2 n
≤ F+
2
n
√
(b)
1 + o(1)
s2 (1 + min(n/s2 , n/s)
log 1 +
+1 +2 n
≤F+
2
n
where (a) follows from Jensen’s inequality, and (b) follows from the proof of Lemma 3.1.
B
Quantized SVRG
Variance Reduction for Sums of Smooth Functions. One common setting in which SGD sees
application in machine learning is when
Pfmcan be naturally expressed as a sum of smooth functions.
1
Formally, we assume that f (x) = m
i=1 fi (x). When f can be expressed as a sum of smooth
functions, this lends itself naturally to SGD. This is because a natural stochastic gradient for f in
this setting is, on input x, to sample a uniformly random index i, and output ∇fi (x). We will also
impose somewhat stronger assumptions on f and f1 , f2 , . . . , fm , namely, that f is strongly convex,
and that each fi is convex and smooth.
16
Definition B.1 (Strong Convexity). Let f : Rn → R be a differentiable function. We say that f is
`-strongly convex if for all x, y ∈ Rn , we have
`
f (x) − f (y) ≤ ∇f (x)T (x − y) − kx − yk22 .
2
Observe that when ` = 0 this is the standard definition of convexity.
Note that it is well-known that even if we impose these stronger assumptions on f and f1 , f2 , . . . , fm ,
then by only applying SGD one still cannot achieve exponential convergence rates, i.e. error rates
which improve as exp(−T ) at iteration T . (Such a rate is known in the optimization literature as
linear convergence.) However, an epoch-based modification of SGD, known as stochastic variance
reduced gradient descent (SVRG) [23], is able to give such rates in this specific setting. We describe
the method below, following the presentation of Bubeck [7].
(p)
Background on SVRG. Let y (1) ∈ Rn be an arbitrary point. For p = 1, 2, . . . , P , we let x1 =
(p)
y (p) . Each p is called an epoch. Then, within epoch p, for t = 1, . . . , T , we let it be a uniformly
random integer from [m] completely independent from everything else, and we set:
(p)
(p)
(p)
xt+1 = xt − η ∇fi(p) (xt ) − ∇fi(p) (y (p) ) + ∇f (y (p) ) .
t
t
We then set
y
(p+1)
k
1 X (p)
=
x .
k i=1 t
With this iterative scheme, we have the following guarantee:
Pm
1
Theorem B.1 ([23]). Let f (x) = m
i=1 fi (x), where f is `-strongly convex, and fi are convex
∗
and L-smooth, for all i. Let x be the unique minimizer of f over Rn . Then, if η = O(1/L) and
T = O(L/`), we have
h
i
E f (y (p+1) ) − f (x∗ ) ≤ 0.9p f (y (1) ) − f (x∗ ) .
(8)
Quantized SVRG. In parallel SVRG, we are given K processors, each processor i having access
to fim/K , . . . , f(i+1)m/K−1 . The goal is the same as before: to approximately minimize f =
Pm
P(i+1)m/K−1
1
1
fi be the portion of f that it knows, so that
i=1 fi . For processor i, let hi = m
j=im/K
m
PK
f = i=1 hi .
A natural question is whether we can apply randomized quantization to reduce communication for
parallel SVRG. Whenever one applies our quantization functions to the gradient updates in SVRG,
the resulting update is no longer an update of the form used in SVRG, and hence the analysis for
SVRG does not immediately give any results in black-box fashion. Instead, we prove that despite
this technical issue, one can quantize SVRG updates using our techniques and still obtain the same
convergence bounds.
√
e
Let Q(v)
= Q(v, n), where Q(v, s) is defined as in Section 3.1. Our quantized SVRG updates are
as follows. Given arbitrary starting point x0 , we let y (1) = x0 . At the beginning of epoch p, each
processor broadcasts
(i+1)m/K−1
X
1
e
e (∇hi ) ,
Hp,i = Q
∇fi (y (p) ) = Q
m
j=im/K
Pm
from which the processors collectively form Hp = i=1 Hp,i without additional communication.
(p)
Within each epoch, for each iteration t = 1, . . . , T , and for each processor i = 1, . . . , K, we let ji,t
be a uniformly random integer from [m] completely independent from everything else. Then, in
iteration t in epoch p, processor i broadcasts the update vector
(p)
(p)
e ∇f (p) (x(p)
ut,i = Q
)
−
∇f
(y
)
+
H
.
(p)
p
t
j
j
i,t
i,t
17
(p)
Each processor then computes the total update for that iteration ut =
(p)
(p)
(p)
xt+1 = xt − ηut . At the end of epoch p, each processor sets y (p+1) =
PK
1
ut,i ,
K
PTi=1 (p)
1
t=1 xt .
T
and sets
Our main theorem is that this algorithm still converges, and is communication efficient:
Pm
1
Theorem B.2. Let f (x) = m
i=1 fi (x), where f is `-strongly convex, and fi are convex and
L-smooth, for all i. Let x∗ be the unique minimizer of f over Rn . Then, if η = O(1/L) and
T = O(L/`), then QSVRG with initial point y (1) ensures
h
i
E f (y (p+1) ) − f (x∗ ) ≤ 0.9p f (y (1) ) − f (x∗ ) .
(9)
Moreover, QSVRG with P epochs and T iterations per epoch requires ≤ P (F + 2.8n)(T + 1) bits
of communication per processor.
In particular, observe that when L/` is a constant, this implies that for all epochs p, we may
communicate O(pn) bits and get an error rate of the form (14). Up to constant factors, this matches
the lower bound given in [40].
Proof of Theorem G.2. By Theorem A.6, each processor transmits at most F + 2.8n bits per iteration, and then an additional F + 2.8n bits per epoch to communicate the Hp,i . Thus the claimed
communication bound follows trivially.
We now turn our attention to correctness. As with the case of quantized SGD, it is not hard to see that
the parallel updates are equivalent to minibatched updates, and serve only to decrease the variance of
the random gradient estimate. Hence, as before, for simplicity of presentation, we will consider the
effect of quantization on convergence rates on a single processor. In this case, the updates can be
written down somewhat more simply. Namely, in iteration t of epoch p, we have that
(p)
(p)
xt+1 = xt
(p)
where jt
(p)
et
− ηQ
(p)
e (p) (∇f (y))
∇fj (p) (xt ) − ∇fj (p) (y (p) ) + Q
t
,
t
e (p)
e (p) are all different, independent instances of Q.
e
is a random index of [m], and Q
t and Q
We follow the presentation in [7]. Fix an epoch p ≥ 1, and let E denote the expectation taken with
respect to the randomness within that epoch.
We will show that
"
#
T
h
i
1 X (p)
(p+1)
∗
− f (x∗ ) ≤ 0.9p f (y (1) ) − f (x∗ ) .
E f y
− f (x ) = E
xt
T t=1
This clearly suffices to show the theorem. Because we only deal with a fixed epoch, for simplicity
of notation,
on p in the notation. For t = 1, . . . , T , let
we shall proceed to drop the dependence
(p)
e
e
vt = Qt ∇fjt (xt ) − ∇fjt (y) − Q (∇f (y)) be the update in iteration t. It suffices to show the
following two equations:
Ejt ,Qet ,Qe [vt ] = ∇f (xt ) , and
Ejt ,Qet ,Qe kvt k2 ≤ C · L (f (xt ) − f (x∗ ) + f (y) − f (x∗ )) ,
18
where C is some universal constant. That the first equation is true follows from the unbiasedness of
e We now show the second. We have:
Q.
Ejt ,Qet ,Qe kvt k2 = Ejt ,Qe EQes kvt k2
(a)
2
e (p) (∇f (y))
≤ 2 Ejt ,Qe ∇fjt (xt ) − ∇fjt (y) + Q
h
i
(b)
2
e (p) (∇f (y))
≤ 4 Ejt k∇fjt (xt ) − ∇fjt (x∗ )k + 4 Ejt ,Qe ∇fjt (x∗ ) − ∇fjt (y) + Q
(c)
h
i
h
i
2
2
≤ 4 Ejt k∇fjt (xt ) − ∇fjt (x∗ )k + 4 Ejt ,Qe k∇fjt (x∗ ) − ∇fjt (y) + ∇f (y)k
2
(p)
e
+ 4 Ejt ,Qe ∇f (y) − Q (∇f (y))
(d)
h
i
h
i
2
2
≤ 4 Ejt k∇fjt (xt ) − ∇fjt (x∗ )k + 4 Ejt ,Qe k∇fjt (x∗ ) − ∇fjt (y) + ∇f (y)k
+ 8 k∇f (y)k
2
(e)
≤ 8L (f (xt ) − f (x∗ )) + 4L (f (y) − f (x∗ )) + 16L(f (y) − f (x∗ ))
≤ C · L (f (xt ) − f (x∗ ) + f (y) − f (x∗ )) ,
as claimed, for some positive constant C ≤ 16. Here (a) follows from Lemma 3.1, (b) and (c)
follow from the fact that (a + b)2 ≤ 2a2 + 2b2 for all scalars a, b, (d) follows from Lemma 3.1
and independence, and (e) follows from Lemma 6.4 in [7] and the standard fact that k∇f (y)k2 ≤
2L(f (y) − f (x∗ )) if f is `-strongly convex.
Plugging these bounds into proof structure in [7] yields the proof of G.1, as claimed.
Why does naive quantization not achieve this rate? Our analysis shows that quantized SVRG
achieves the communication efficient rate, using roughly 2.8 times as many bits per iteration, and
roughly C/2 = 8 times as many iterations. This may beg the question why naive quantization
schemes (say, quantizing down to 16 or 32 bits) fails. At a high level, this is because any such
quantization can inherently only achieve up to constant error, since the stochastic gradients are always
biased by a (small) constant. To circumvent this, one may quantize down to O(log 1/) bits, however,
this only matches the upper bound given by [40], and is off from the optimal rate (which we achieve)
by a logarithmic factor.
C
Quantization for Non-convex SGD
As stated previously, our techniques are portable, and apply easily to a variety of settings where
SGD is applied. As a demonstration of this, we show here how we may use quantization on top of
recent results which show that SGD converges to local minima when applied on smooth, non-convex
functions.
Throughout this paper, our theory only considers the case when f is a convex function. In many
interesting applications such as neural network training, however, the objective is non-convex, where
much less is known. However, there has been an interesting line of recent work which shows that SGD
at least always provably converges to a local minima, when f is smooth. For instance, by applying
Theorem 2.1 in [15], we immediately obtain the following convergence result for quantized SGD.
Let Qs be the quantization function defined in Section 3.1. Here we will only state the convergence
bound; the communication complexity per iteration is the same as in 3.1.
Theorem C.1. Let f : Rn → R be a L-smooth (possibly nonconvex) function, and let x1 be an
arbitrary initial point. Let T > 0 be fixed, and s > 0. Then there is a random stopping time
R supported on {1, . . . , N } so that QSGD with quantization function Qs , and constant stepsizes
η = O(1/L) and access to stochastic gradients of f with second moment bound B satisfies
!
p
√
2
∗)
L(f
(x
)
−
f
1
(1
+
min(n/s
,
n/s))B
1
E k∇f (x)k22 ≤ O
+
.
L
N
L
19
2
Table 2: Description of networks.
Network
AlexNet
BN-Inception
ResNet152
VGG19
ResNet110
LSTM
Dataset
ImageNet
ImageNet
ImageNet
ImageNet
CIFAR-10
AN4
Epochs
112
300
120
80
160
20
Parameters
62M
11M
60M
143M
1M
13M
Init. L. Rate
0.07
3.6
1
0.1
0.1
0.5
Minibatch size (2, 4, 8, 16 GPUs)
Varies (256, 512, 1024, 1024)
Varies (256, 256, 256, 1024)
Varies (32, 64, 128, 256)
Varies (64, 128, 256)
128
256
Observe that the only difference in the assumptions in [15] from what we generally assume is that
they assume a variance bound on the stochastic gradients, whereas we prefer a second moment bound.
Hence our result applies immediately to their setting.
Another recent result [29] demonstrates local convergence for SGD for smooth non-convex functions
in asynchronous settings. The formulas there are more complicated, so for simplicity we will not
reproduce them here. However, it is not hard to see that quantization affects the convergence bounds
there in a manner which is parallel to Theorem 3.2.
D
Asynchronous QSGD
We consider an asynchronous parameter-server model [28], modelled identically as in [29, Section 3].
In brief, the system consists of a star-shaped network, with a central parameter server, communicating
with worker nodes, which exchange information with the master independently and simultaneously.
Asynchrony consists of the fact that competing updates might be applied by the master to the shared
parameter (but workers always get a consistent version of the parameter).
In this context, the following follows from [29, Theorem 1]:
Theorem D.1. Let f : Rn → R be a L-smooth (possibly nonconvex) function, and let x1 be an arbitrary initial point. Assume unbiased stochastic gradients, with bounded variance σ 2 , and Lipschitzian
gradient with parameter L. Let K be the number of iterations, and M be the minibatch size. Further
assume that all the locations of the gradient updates {ξk,m }k=[K],m=[M ] are independent random
variables, and that the delay with which each update is applied is upper bounded by a parameter T .
Finally, assume that the steplength sequence {γk }k=[K] satisfies
LM γk + 2L2 M 2 T γk
T
X
γk+κ ≤ 1, ∀k = 1, 2, . . . .
κ=1
We then have the following ergodic convergence rate for the iteration of QSGD with quantization
PK
√
function Qs . Let γ = k=1 γk , and σs = (1 + min(n/s2 , n/s))σ. Then:
PK 2
Pk−1
2
2
2
2
∗
K
γ
M
L
+
2L
M
γ
γ
2(f
(x
)
−
f
(x
))
+
X
1
k
k
k=1
j=k−T j σs
γk E [k∇f (xk )k]
≤
.
γ
Mγ
k=1
E
Experiments
We now empirically validate our approach on data-parallel GPU training of deep neural networks.
Setup. We performed experiments on Amazon EC2 p2.16xlarge instances, using up to 16 NVIDIA
K80 GPUs. Instances have GPUDirect peer-to-peer communication, but do not currently support
NVIDIA NCCL extensions. We have implemented QSGD on GPUs using the Microsoft Cognitive
Toolkit (CNTK) [3]. This package provides efficient (MPI-based) GPU-to-GPU communication, and
implements an optimized version of 1bit-SGD [35]. Our code is released both as open-source and as
a docker instance.
We do not quantize small gradient matrices in QSGD, since the computational cost of quantizing
small matrices significantly exceeds the reduction in communication from quantization. However,
20
1bitSGD
32bit
8 GPUs
4 GPUs
1bitSGD
32bit
20
8 GPUs
2bit (d=128) 4bit (d=1024)
10
8
6
4
2
0
4 GPUs
1bitSGD
32bit
1bitSGD
32bit
8 GPUs
16 GPUs
5
40
35
30
25
20
15
10
5
0
ResNet50
2 GPUs
4 GPUs
10
0
16 GPUs
VGG19
2 GPUs
15
2bit (d=128) 4bit (d=1024)
Inception
2 GPUs
16 GPUs
Time per epoch (hours)
AlexNet
4 GPUs
Time per epoch (hours)
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
2 GPUs
Time per epoch (hours)
Time per epoch (hours)
Time per epoch (hours)
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
8 GPUs
2 GPUs
1bitSGD
2bit (d=128) 4bit (d=1024)
ResNet152
4 GPUs
32bit
8 GPUs
16 GPUs
2bit (d=128) 4bit (d=1024)
16 GPUs
2bit (d=128) 4bit (d=1024)
Figure 4: Breakdown of communication versus computation for various neural networks, on 2, 4, 8, 16
GPUs, for full 32-bit precision versus 1BitSGD versus QSGD 2-bit and 4-bit. Each bar represents the
total time for an epoch under standard parameters. Epoch time is broken down into communication
(bottom, solid color) and computation (top, transparent color). Notice that, although epoch time
usually diminishes as we increase parallelism, the proportion of communication cost increases.
in all experiments, at least 99% of all parameters are transmitted in quantized form. If required, we
reshape matrices to fit bucket sizes.
We execute two types of tasks: image classification on the ILSVRC (ImageNet) [12], CIFAR10 [25], and MNIST [27] datasets, and speech recognition on the CMU AN4 dataset [2]. For
vision, we experimented with AlexNet [26], VGG [36], ResNet [18], and Inception with Batch
Normalization [22] deep networks. For speech, we trained an LSTM network [19]. See Table 1.
We used standard sizes for the networks, with hyper-parameters optimized for the 32bit precision
variant.1 Full details for networks and experiments are given in the additional material. We increased
batch size when necessary to balance communication and computation for larger GPU counts, and
we employed double buffering [35] to perform communication and quantization concurrently with
the computation. Quantization usually benefits from lowering learning rates; yet, we always run the
32bit learning rate, and decrease bucket size to reduce variance if needed.
Communication vs. Computation. In the first set of experiments, we examine the ratio between
computation and communication costs during training, for increased parallelism. The image classification networks are trained on ImageNet, while LSTM is trained on AN4. We examine the cost
breakdown for these networks over a pass over the dataset (epoch). Figure 2 gives image classification
results. The variance of epoch times is practically negligible.
1
Unless otherwise stated, we use the default networks and hyper-parameters available in the open-source
CNTK 2.0.
21
Test accuracy (%)
> 2x faster
2.0
2bit QSGD (d=128)
4bit QSGD (d=8192)
8bit QSGD (d=8192)
SGD
Training loss
1.5
1.0
0.5
0.00
300
600
900
Time (sec)
1200
1500
(c) LSTM Accuracy versus Time.
20
40
60
2bit QSGD (d=128)
4bit QSGD (d=512)
8bit QSGD (d=512)
SGD
80 100 120 140 160
Epoch
(b) ResNet110 Accuracy on CIFAR.
Test accuracy (%)
(a) AlexNet Accuracy on ImageNet.
96
94
92
90
88
86
84
82
800
98
97
96
95
94
93
92
910
SGD
QSGD (d=256)
QSGD (d=1024)
QSGD (d=4096)
5
10
Epoch
15
20
(d) Two-Layer Perceptron Accuracy on
MNIST.
Figure 5: Accuracy numbers for different networks. The red lines in (a) represent final 32-bit
accuracy.
The data leads to some interesting observations. First, based on the ratio of communication to
computation, we can roughly split networks into communication-intensive (AlexNet, VGG, LSTM),
and computation-intensive (Inception, ResNet). For both network types, the relative impact of
communication increases significantly as we increase the number of GPUs. Examining the breakdown
for the 32-bit version, all networks could significantly benefit from reduced communication. For
example, for AlexNet on 16 GPUs with batch size 1024, more than 80% of training time is spent
on communication, whereas for LSTM on 2 GPUs with batch size 256, the proportion is 71%
communication.2
Next, we examine the impact of QSGD on communication and overall training time. (For QSGD,
communication time includes time spent compressing and uncompressing gradients.) We measured
QSGD with 2-bit quantization and 64 bucket size, and 4-bit quantization and 8192 bucket size. The
results for these two variants are similar, since the different bucket sizes mean that the 4bit version
only sends 77% more data than the 2-bit version (but ∼ 8× less than 32-bit). These bucket sizes are
chosen to ensure good convergence, but are not carefully tuned.
On 16GPU AlexNet with batch size 1024, 4-bit QSGD reduces communication time by 4×, and
overall epoch time by 2.5×. On LSTM, it reduces communication time by 6.8×, and overall epoch
time by 2.7×. Runtime improvements are non-trivial for all architectures we considered.
Accuracy. We now examine how QSGD influences accuracy and convergence rate. We ran AlexNet
to full convergence on ImageNet, LSTM on AN4, ResNet110 on CIFAR-10, as well as a two-layer
perceptron on MNIST. Results are given in Figure 5.
On ImageNet using AlexNet, 4-bit QSGD with 8192 bucket size converges to 59.22% top-1 error,
and 81.63% top-5 error. The gap from the 32bit version is 0.57% for top-5, and 0.68% for top-1 [9].
QSGD with 2-bit and 64 bucket size has gap 1.73% for top-1, and 1.18% for top-1. We note that we
did not tune bucket size, number of bits used, number of epochs or learning rate for this experiment.
2
These ratios can be improved by increasing batch size. However, increasing batch size further hurts
convergence and decreases accuracy, see also e.g. [21].
22
1
2
3
Data: Parameter vector x
procedure GradientDescent
for each iteration t do
Q(∇f (x)) ← Quantize(∇f (x))) //quantize gradient
x ← x − ηt Q(∇f (x)) //apply gradient
4
5
end
Algorithm 2: The gradient descent algorithm with gradient encoding.
On AN4 using LSTMs, 2-bit QSGD has similar convergence rate and the same accuracy as 32bit. It
is able to converge 3× faster to the target accuracy with respect to full precision, thanks to reduced
communication overheads. The 4-bit variant has the same convergence and accuracy, but is slightly
slower than 2-bit (by less than 10%).
On CIFAR-10, 2-bit QSGD applied to ResNet-110 drops about 1.22% top-1 accuracy points. However,
4-bit QSGD converges to the same accuracy as the original, whereas 8-bit QSGD improves accuracy
by 0.33%. We observe a similar result on MNIST, where 2-bit QSGD with buckets equal to the size
of hidden layers improves accuracy by 0.5%. These results are consistent with recent work [30]
noting benefits of added noise in training deep networks. Linear models on e.g. MNIST do not show
such improvements.
One issue we examined in more detail is which layers are more sensitive to quantization. It appears
that quantizing convolutional layers too aggressively (e.g., 2-bit precision) can lead to accuracy loss
if not trained further. However, increasing precision to 4-bit or 8-bit recovers accuracy. This finding
suggests that modern architectures for vision tasks, such as ResNet or Inception, which are almost
entirely convolutional, may benefit less from quantization than recurrent deep networks such as
LSTMs.
Comparison with 1BitSGD. We have also compared against the 1BitSGD algorithm of [35]. Before
discussing results, it is important to note some design choices made in the CNTK implementation
of 1BitSGD. For objects without dynamic dimensions, the first tensor dimension is the “row" while
the rest are flattened onto “columns." At the same time, 1BitSGD always quantizes per column. In
practice, this implies that quantization is often applied to a column of very small dimension (1–3),
especially in the case of networks with many convolutions. This has the advantage of having extremely
low variance, but does not yield any communication benefits. In fact, it can hurt performance due to
the cost of quantization. (By contrast, we reshape to quantize on large dimensions.)
Given this artefact, 1BitSGD is slower than even the 32bit version on heavily convolutional networks
such as ResNet and Inception. However, 1BitSGD matches the performance of 2-bit and 4-bit QSGD
on AlexNet, VGG, and LSTMs within 10%. In general, 1BitSGD attains very good accuracy (on par
with 32bit), probably since the more delicate convolutional layers are not quantized. QSGD has the
advantage of being able to perform quantization on the fly, without error accumulation: this saves
memory, since we do not need to allocate an additional model copy.
F
Quantized Gradient Descent: Description and Analysis
In this section, we consider the effect of lossy compression on standard (non-stochastic) gradient
descent. Since this procedure is not data-parallel, we will first have to modify the blueprint for the
iterative procedure, as described in Algorithm 2. In particular, we assume that, instead of directly
applying the gradient to the iterate xt+1 , the procedure first quantizes the gradient, before applying
it. This setting models a scenario where the model and the computation are performed by different
machines, and we wish to reduce the communication cost of the gradient updates.
We now give a quantization function tailored for gradient descent, prove convergence of gradient
descent with quantization, and then finally bound the length of the encoding.
23
The Quantization Function. We consider the following deterministic quantization function, inspired
by [35]. For any vector v ∈ Rn , let I(v) be the smallest set of indices of v such that
X
|vi | ≥ kvk.
i∈I(v)
Further, define Q(v) to be the vector
(
Q(v)i =
kvk
−kvk
0
if x ≥ 0 and i ∈ I(v);
if x < 0 and i ∈ I(v);
otherwise.
Practically, we preserve the sign for each index in I(v), the 2-norm of v, and cancel out all remaining
components of v.
Convergence Bound. We begin by proving some properties of our quantization function. We have
the following:
Lemma F.1. For all v ∈ Rn , we have
1. v T Q(v) ≥ kvk2 ,
√
2. |I(v)| ≤ n, and
√
3. kQ(v)k2 ≤ nkvk2 .
Proof. For the first claim, observe that v T Q(v) = kvk
P
i∈I(v)
|vi | ≥ kvk2 .
We now prove the second claim. Let v = (v1 , . . . , vn ), and without loss of generality, assume
that |vi | ≥ |vi+1 | for all i = 1, . . . , n − 1, so that the coordinates are in decreasing order. Then
I(v) = {1, . . . , D} for some D.
PD
√
√
We show that if D ≥ n then i=1 |vi | ≥ kvk, which shows that |I(v)| ≤ n. Indeed, we have
that
!2
D
D
X
X
X
vi2 +
|vi ||vj |
|vi |
=
i=1
i=1
≥
D
X
i6=j
i,j≤D
2
vi2 + (D2 − D)vD+1
.
i=1
On the other hand, we have
kvk2 =
D
X
vi2 +
i=1
≤
D
X
n
X
vi2
D+1
2
vi2 + (n − D)vD+1
i=1
and so we see that if D =
√
n, we must have
P
D
i=1
2
|vi | ≥ kvk2 , as claimed.
For the third claim, observe that kQ(v)k2 = kvk2 · |I(v)|; thus the claim follows from the previous
upper bound on the cardinality of I(v).
To establish convergence of the quantized method, we prove the following theorem.
∗
Theorem F.2. Let f : Rn → R be a `-strongly convex, L-smooth function, with
global
minimizer x ,
`√
and condition number κ = L/`. Then, for all step sizes η satisfying η ≤ O L2 n , for all T ≥ 1,
and all initial points x0 , we have
1
√
f (xT ) − f (x∗ ) ≤ exp −Ω
T (f (x0 ) − f (x∗ )) .
κ2 n
24
Proof. We first establish the following two properties:
Lemma F.3. Let f be `-strongly convex and L-smooth. Then,
1. for all x ∈ Rn ,
`
L
kx − x∗ k2 ≤ f (x) − f (x∗ ) ≤ kx − x∗ k2 .
2
2
2. for all x ∈ Rn ,
∇f (x)T Q(∇f (x)) ≥ `(f (x) − f (x∗ )) .
Proof. The first property follows directly from the definitions of strong convexity and smoothness.
We now show the second property. If x = x∗ the property trivially holds so assume that this does not
happen. By Lemma F.1, we have
∇f (x)T Q(∇f (x)) ≥ k∇f (x)k2 .
We then have
f (x) − f (x∗ ) ≤ ∇f (x)T (x − x∗ ) ≤ k∇f (x)k kx − x∗ k ,
where the first inequality follows from convexity, and the second from Cauchy-Schwartz. From
strong convexity we then have that 2` kx − x∗ k2 ≤ ∇f T (x)(x − x∗ ) from which we get that
`
x − x∗
kx − x∗ k ≤ ∇f (x)T
≤ k∇f (x)k ,
2
kx − x∗ k
where the last line follows since from self-duality of the 2-norm, we know that for all vectors v ∈ Rn ,
we have kvk2 = supkuk=1 v T u. Putting these two things together yields that
f (x) − f (x∗ ) ≤
2
2
2
k∇f (x)k ≤ ∇f (x)T Q(∇f (x)) ,
`
`
as claimed.
With all this in place, we can now complete the proof of the theorem. Fix t ≥ 0. By applying the
lemma, we have:
`
(10)
∇f (xt )T (xt+1 − xt ) = −η∇f (xt )T Q(∇f (xt )) ≤ −η (f (x) − f (x∗ )) .
2
Moreover, observe that, from standard properties of smooth functions [7], we have
1
k∇f (x)k2 ≤ f (x) − f (x∗ ) .
(11)
2L
Thus, we obtain the following chain of inequalities:
(a)
f (xt+1 ) − f (xt ) ≤ ∇f (xt+1 )T (xt+1 − xt )
T
= ∇f (xt )T (xt+1 − xt ) + (∇f (xt+1 ) − ∇f (xt )) (xt+1 − xt )
(b)
`
≤ −η (f (xt ) − f (x∗ )) + k∇f (xt+1 ) − ∇f (xt )k kxt+1 − xt k
2
(c)
`
≤ −η (f (xt ) − f (x∗ )) + Lkxt+1 − xt k2
2
`
= −η (f (xt ) − f (x∗ )) + η 2 LkQ(∇f (xt ))k2
2
(d)
√
`
≤ −η (f (xt ) − f (x∗ )) + η 2 L nk∇f (xt )k22
2
(e)
√
`
≤ −η (f (xt ) − f (x∗ )) + η 2 2L2 n(f (xt ) − f (x∗ ))
2
√
`
= −η + 2η 2 L2 n (f (xt ) − f (x∗ )) ,
2
25
where (a) follows from the convexity of f , (b) follows from Equation 10 and the Cauchy-Schwarz
inequality, (c) follows from the L-smoothness of f , (d) follows from Lemma F.1, and (e) follows
from Equation 11. By our choice of η, we know that the RHS of Equation 12 is negative. Hence, by
Lemma F.3 and the definition of η, we have
1
√
f (xt+1 ) − f (xt ) ≤ −Ω
(f (xT ) − f (x∗ )) .
(12)
κ2 n
Letting δt = f (xt ) − f (x∗ ), and observing that f (xt+1 ) − f (xt ) = δt+1 − δt , we see that Equation
12 is equivalent to the statement that
1
√
δt+1 ≤ 1 − Ω
δt .
κ2 n
Thus altogether we have
T
1
√
δ0
δT ≤ 1 − Ω
κ2 n
1
√
≤ exp −Ω
T δ0 ,
κ2 n
as claimed.
Encoding Length. We obtain the following:
Theorem F.4. Let v ∈ Rn . Then
√
|Code(Q(v))| ≤ n(log(n) + 1 + log(e)) + F.
G
Quantized SVRG
Variance Reduction for Sums of Smooth Functions. One common setting in which SGD sees
application in machine learning is when
Pfmcan be naturally expressed as a sum of smooth functions.
1
Formally, we assume that f (x) = m
i=1 fi (x). When f can be expressed as a sum of smooth
functions, this lends itself naturally to SGD. This is because a natural stochastic gradient for f in
this setting is, on input x, to sample a uniformly random index i, and output ∇fi (x). We will also
impose somewhat stronger assumptions on f and f1 , f2 , . . . , fm , namely, that f is strongly convex,
and that each fi is convex and smooth.
Definition G.1 (Strong Convexity). Let f : Rn → R be a differentiable function. We say that f is
`-strongly convex if for all x, y ∈ Rn , we have
`
f (x) − f (y) ≤ ∇f (x)T (x − y) − kx − yk22 .
2
Observe that when ` = 0 this is the standard definition of convexity.
Note that it is well-known that even if we impose these stronger assumptions on f and f1 , f2 , . . . , fm ,
then by only applying SGD one still cannot achieve exponential convergence rates, i.e. error rates
which improve as exp(−T ) at iteration T . (Such a rate is known in the optimization literature as
linear convergence.) However, an epoch-based modification of SGD, known as stochastic variance
reduced gradient descent (SVRG) [23], is able to give such rates in this specific setting. We describe
the method below, following the presentation of Bubeck [7].
(p)
Background on SVRG. Let y (1) ∈ Rn be an arbitrary point. For p = 1, 2, . . . , P , we let x1 =
(p)
y (p) . Each p is called an epoch. Then, within epoch p, for t = 1, . . . , T , we let it be a uniformly
random integer from [m] completely independent from everything else, and we set:
(p)
(p)
(p)
xt+1 = xt − η ∇fi(p) (xt ) − ∇fi(p) (y (p) ) + ∇f (y (p) ) .
t
t
We then set
y
(p+1)
k
1 X (p)
=
x .
k i=1 t
With this iterative scheme, we have the following guarantee:
26
Pm
1
Theorem G.1 ([23]). Let f (x) = m
i=1 fi (x), where f is `-strongly convex, and fi are convex
and L-smooth, for all i. Let x∗ be the unique minimizer of f over Rn . Then, if η = O(1/L) and
T = O(L/`), we have
h
i
E f (y (p+1) ) − f (x∗ ) ≤ 0.9p f (y (1) ) − f (x∗ ) .
(13)
Quantized SVRG. In parallel SVRG, we are given K processors, each processor i having access
to fim/K , . . . , f(i+1)m/K−1 . The goal is the same as before: to approximately minimize f =
Pm
P(i+1)m/K−1
1
1
fi be the portion of f that it knows, so that
i=1 fi . For processor i, let hi = m
j=im/K
m
PK
f = i=1 hi .
A natural question is whether we can apply randomized quantization to reduce communication for
parallel SVRG. Whenever one applies our quantization functions to the gradient updates in SVRG,
the resulting update is no longer an update of the form used in SVRG, and hence the analysis for
SVRG does not immediately give any results in black-box fashion. Instead, we prove that despite
this technical issue, one can quantize SVRG updates using our techniques and still obtain the same
convergence bounds.
√
e
Let Q(v)
= Q(v, n), where Q(v, s) is defined as in Section 3.1. Our quantized SVRG updates are
as follows. Given arbitrary starting point x0 , we let y (1) = x0 . At the beginning of epoch p, each
processor broadcasts
(i+1)m/K−1
X
1
e (∇hi ) ,
e
∇fi (y (p) ) = Q
Hp,i = Q
m
j=im/K
Pm
from which the processors collectively form Hp = i=1 Hp,i without additional communication.
(p)
Within each epoch, for each iteration t = 1, . . . , T , and for each processor i = 1, . . . , K, we let ji,t
be a uniformly random integer from [m] completely independent from everything else. Then, in
iteration t in epoch p, processor i broadcasts the update vector
(p)
(p)
e ∇f (p) (x(p)
ut,i = Q
)
−
∇f
(y
)
+
H
(p)
p .
t
j
j
i,t
i,t
(p)
Each processor then computes the total update for that iteration ut =
(p)
(p)
(p)
xt+1 = xt − ηut . At the end of epoch p, each processor sets y (p+1) =
PK
1
ut,i ,
K
PTi=1 (p)
1
t=1 xt .
T
and sets
Our main theorem is that this algorithm still converges, and is communication efficient:
Pm
1
Theorem G.2. Let f (x) = m
i=1 fi (x), where f is `-strongly convex, and fi are convex and
L-smooth, for all i. Let x∗ be the unique minimizer of f over Rn . Then, if η = O(1/L) and
T = O(L/`), then QSVRG with initial point y (1) ensures
h
i
E f (y (p+1) ) − f (x∗ ) ≤ 0.9p f (y (1) ) − f (x∗ ) .
(14)
Moreover, QSVRG with P epochs and T iterations per epoch requires ≤ P (F + 2.8n)(T + 1) bits
of communication per processor.
In particular, observe that when L/` is a constant, this implies that for all epochs p, we may
communicate O(pn) bits and get an error rate of the form (14). Up to constant factors, this matches
the lower bound given in [40].
Proof of Theorem G.2. By Theorem A.6, each processor transmits at most F + 2.8n bits per iteration, and then an additional F + 2.8n bits per epoch to communicate the Hp,i . Thus the claimed
communication bound follows trivially.
We now turn our attention to correctness. As with the case of quantized SGD, it is not hard to see that
the parallel updates are equivalent to minibatched updates, and serve only to decrease the variance of
the random gradient estimate. Hence, as before, for simplicity of presentation, we will consider the
27
effect of quantization on convergence rates on a single processor. In this case, the updates can be
written down somewhat more simply. Namely, in iteration t of epoch p, we have that
(p)
(p)
(p)
(p)
(p)
e (p)
e
xt+1 = xt − η Q
∇f
(x
)
−
∇f
(y
)
+
Q
(∇f
(y))
,
(p)
(p)
t
t
j
j
t
where
(p)
jt
is a random index of [m], and
t
e (p)
Q
t
e (p)
and Q
e
are all different, independent instances of Q.
We follow the presentation in [7]. Fix an epoch p ≥ 1, and let E denote the expectation taken with
respect to the randomness within that epoch.
We will show that
"
#
T
h
i
1 X (p)
(p+1)
∗
E f y
− f (x ) = E
xt
− f (x∗ ) ≤ 0.9p f (y (1) ) − f (x∗ ) .
T t=1
This clearly suffices to show the theorem. Because we only deal with a fixed epoch, for simplicity
of notation,
on p in the notation. For t = 1, . . . , T , let
we shall proceed to drop the dependence
(p)
e
e
vt = Qt ∇fjt (xt ) − ∇fjt (y) − Q (∇f (y)) be the update in iteration t. It suffices to show the
following two equations:
Ejt ,Qet ,Qe [vt ] = ∇f (xt ) , and
Ejt ,Qet ,Qe kvt k2 ≤ C · L (f (xt ) − f (x∗ ) + f (y) − f (x∗ )) ,
where C is some universal constant. That the first equation is true follows from the unbiasedness of
e We now show the second. We have:
Q.
Ejt ,Qet ,Qe kvt k2 = Ejt ,Qe EQes kvt k2
(a)
2
(p)
e
≤ 2 Ejt ,Qe ∇fjt (xt ) − ∇fjt (y) + Q (∇f (y))
h
i
(b)
2
e (p) (∇f (y))
≤ 4 Ejt k∇fjt (xt ) − ∇fjt (x∗ )k + 4 Ejt ,Qe ∇fjt (x∗ ) − ∇fjt (y) + Q
(c)
h
i
h
i
2
2
≤ 4 Ejt k∇fjt (xt ) − ∇fjt (x∗ )k + 4 Ejt ,Qe k∇fjt (x∗ ) − ∇fjt (y) + ∇f (y)k
2
(p)
e
+ 4 Ejt ,Qe ∇f (y) − Q (∇f (y))
(d)
h
i
h
i
2
2
≤ 4 Ejt k∇fjt (xt ) − ∇fjt (x∗ )k + 4 Ejt ,Qe k∇fjt (x∗ ) − ∇fjt (y) + ∇f (y)k
+ 8 k∇f (y)k
2
(e)
≤ 8L (f (xt ) − f (x∗ )) + 4L (f (y) − f (x∗ )) + 16L(f (y) − f (x∗ ))
≤ C · L (f (xt ) − f (x∗ ) + f (y) − f (x∗ )) ,
as claimed, for some positive constant C ≤ 16. Here (a) follows from Lemma 3.1, (b) and (c)
follow from the fact that (a + b)2 ≤ 2a2 + 2b2 for all scalars a, b, (d) follows from Lemma 3.1
and independence, and (e) follows from Lemma 6.4 in [7] and the standard fact that k∇f (y)k2 ≤
2L(f (y) − f (x∗ )) if f is `-strongly convex.
Plugging these bounds into proof structure in [7] yields the proof of G.1, as claimed.
Why does naive quantization not achieve this rate? Our analysis shows that quantized SVRG
achieves the communication efficient rate, using roughly 2.8 times as many bits per iteration, and
roughly C/2 = 8 times as many iterations. This may beg the question why naive quantization
schemes (say, quantizing down to 16 or 32 bits) fails. At a high level, this is because any such
quantization can inherently only achieve up to constant error, since the stochastic gradients are always
biased by a (small) constant. To circumvent this, one may quantize down to O(log 1/) bits, however,
this only matches the upper bound given by [40], and is off from the optimal rate (which we achieve)
by a logarithmic factor.
28
2
| 8cs.DS
|
arXiv:1710.07831v1 [cs.CV] 21 Oct 2017
A Generative Restricted Boltzmann Machine
Based Method for High-Dimensional Motion
Data Modeling
Siqi Nie∗
Ziheng Wang
Qiang Ji‡
†
October 24, 2017
Abstract
Many computer vision applications involve modeling complex spatiotemporal patterns in high-dimensional motion data. Recently, restricted
Boltzmann machines (RBMs) have been widely used to capture and represent spatial patterns in a single image or temporal patterns in several
time slices. To model global dynamics and local spatial interactions, we
propose to theoretically extend the conventional RBMs by introducing
another term in the energy function to explicitly model the local spatial
interactions in the input data. A learning method is then proposed to
perform efficient learning for the proposed model. We further introduce
a new method for multi-class classification that can effectively estimate
the infeasible partition functions of different RBMs such that RBM is
treated as a generative model for classification purpose. The improved
RBM model is evaluated on two computer vision applications: facial expression recognition and human action recognition. Experimental results
on benchmark databases demonstrate the effectiveness of the proposed
algorithm.
1
Introduction
Spatio-temporal patterns in high-dimensional motion data are crucial in many
recognition applications. For example, human action is the combination of the
body joint movements over a time interval. Facial expression is the result of the
facial landmark movements (Figure 1). Understanding the movement trajectories and modeling the underlying spatio-temporal patterns play an important
∗ Email:
[email protected].
[email protected].
‡ Email: [email protected]. Affiliation: Rensselaer Polytechnic Institute, USA.
† Email:
1
role in recognizing these actions, especially with the recent emergences of reliable
algorithms [1, 2] to estimate the positions of body joints and facial landmarks.
In this work, we are interested in developing a probabilistic model to capture the spatio-temporal pattenrs in high-dimensional time series for classification purpose. Many recent works develop novel models to capture the spatiotemporal dynamics [3, 4, 5, 6]. However, most models, such as hidden Markov
model (HMM) [5], dynamic Bayesian network (DBN) [7] and conditional random field (CRF) [8] are time-slice local models, which assume Markov property
and stationary transitions and hence can only capture local dynamics. The
local models suffer from two limitations. First, local dynamics may not represent a sequence well because it fails to mpdel the overall dynamics. Second, the
stationary transition assumption may not hold for many real-world applications.
Compared with time-sliced dynamic models, restricted Boltzmann machines
(RBMs) has shown strong capability of modeling joint distributions and therefore can capture the global patterns. In literature, RBMs have been successfully
applied to separately capture the spatial [9] or temporal [10] patterns in different types of data. In this work, we propose a variant of RBM that can capture
spatial and global temporal patterns simultaneously to comprehensively model
the high-dimensional motion data.
In a typical RBM, since there are no lateral connections among nodes in
each layer, input data are independent of each other given the states of hidden
layer. This assumption limits RBM’s representation power, since there exist
direct dependencies among input data. There are generally two types of data
interactions: interactions through latent variable and direct interactions independent of latent variable. For example, as stated in [11], soldiers on a parade
follow the commander’s order to march in some direction, but they “form a
neat rectangle by interacting with their neighbors”. This example illustrates
that soldiers’ behavior are determined by both the commander’s order (latent)
and the interactions with their neighbors. In this case, RBM is not effective in
modeling the local interaction that is independent of the latent variable. The
use of RBM for data representation and classification is further hindered by the
difficulty in comparing different RBMs due to the intractable computation of
the partition functions.
Allowing interaction among visible units can overcome this shortcoming.
We introduce restricted Boltzmann machine with local interactions (LRBM) to
capture both the global temporal patterns and local spatial interactions in the
input data. Specifically, we add a new pairwise potential term in the learning objective function of RBM to capture the local spatial interactions among
components of an input vector. To perform efficient learning for the model, we
replace the reconstruction procedure in the conventional Contrastive Divergence
[13] algorithm with a mean field approach.
For classification task, RBM is typically used for learning features, as the
input to a second stage classifier (e.g., SVM [14]). Typically one model is
trained for all classes. To obtain good features, deep structure is built and back
propagation is employed to carefully tune the parameters. In this work, we
use RBM as a generative model to capture the spatio-temporal patterns in the
2
Figure 1: Examples of spatiotemporal patterns in different applications: facial
expression recognition and human action recognition. Both spatial interactions
and temporal movements of landmarks define an expression or an action. Images
are from CK+ data set[2] and G3D data set [12], respectively.
data. RBM is used for data representation instead of feature learning. Given an
observation, the only output we get from an RBM model is the likelihood of the
model. For the recognition purpose, one model is trained for one class of input
data. To compare among different models, we employ a method to estimate
the relative partition function of a pair of RBMs for binary classification, and
a label ranking method is used to extend the binary classification to multi-class
classification.
To evaluate the performance of LRBM, we apply it to two areas related to
complex spatial and temporal patterns: facial expression recognition and human
action recognition. Experimental results on benchmark databases demonstrate
the effectiveness of the proposed model.
The rest of the paper is structured as follows. Section 2 presents an overview
of the related work. Section 3 introduces the LRBM to model the global temporal dynamics and spatial patterns of motion data, as well as the classification
method. We will then give the experimental results in 4. The paper is concluded
in section 5.
2
Related Work
Capturing and representing spatio-temporal structure in data is important for
many recognition and classification tasks. Research for capturing such patterns can be categorized into feature-based and model-based methods. The
most widely used spatio-temporal features include spatio-temporal interest point
(STIP) based features [3, 4] and optical flow based features [15]. These features
capture local appearance or motion patterns near the interest points or optical
3
flows. Although having been successfully applied to many applications, these
features generally focus more on local patterns.
Model-based methods include probabilistic graphical models such as Hidden Markov Models [5], Dynamic Bayesian Networks [16], Conditional Random
Fields [17], and their variants. While capable of simultaneously capturing both
spatial and temporal interactions, they can only capture the local spatial and
temporal interactions due to the underlying Markov assumption.
Restricted Boltzmann machines (RBMs) have been separately used for modeling spatial correlation or temporal correlation in the data in the last decade.
RBM was firstly introduced to learn deep features from handwritings to recognize digits [18]. In [9], Eslami et al. propose a Deep Belief Network to model the
shapes of horses and motorbikes. The samples from the model look realistic and
have a good generalization. A more complicated model, proposed by Nair and
Hinton [19], considers the spatial correlation among visible layer using a factored
3-way RBM, in which triple potentials are used to model the correlations among
pixels in natural images. The intuition is that in natural images, the intensity
of each pixel is approximately the average of its neighbors. Wu et al. [20] apply
the 3-way RBM to facial landmark tracking, and model the relationship between
posed faces and frontal faces, under varying facial expressions.
For dynamic data modeling, Taylor et al. [21] use a Conditional RBM
(CRBM) to model the temporal transitions in human body movements, and
reconstruct body movements. Nevertheless, like HMM, CRBM still models local dynamics by assuming n’th order Markov property.
The idea of using RBM to model global pattern is not new. In [22, 23],
RBM and CRF are combined for face labeling problem in a simgle image or
video sequences. Compared with these works, our work is different for several
reasons. First, our goal is to use RBM for data representation for multi-class
classification, while the goal of [22, 23] is MAP inference, which is to recover the
label for each superpixel. Second, we extend the conventional RBM to capture
the local shape and global temporal patterns in a unified model, while in [22, 23]
RBM is built on top of the hidden layer of the CRF, as the prior for the labels.
Third, our method learns all the parameters simultaneously, while their method
performs learning separately.
RBM and its variants have also been used for modeling motion data. For
example, Sutskever and Hinton [24] introduce a temporal RBM to model highdimensional sequences. Wang et al. [10] use RBM to capture the global dynamics of finger trace. However, it is limited to model the global dynamics for 1-D
data only.
To improve the representation power of RBM, semi-restricted Boltzmann
machine (SRBM) [11] is introduced to model the lateral interactions between
visible variables. The main property of SRBM is that given the hidden variables,
the visible layer forms a Markov random field. However, for high-dimensional
motion data, there will be too many parameters if every pair of visible units has
an interaction. In this work, we model the dynamic nature of data with fewer
parameters than a SRBM, which makes the learning process more efficient.
Besides feature extraction and shape modeling, RBMs have also been used
4
Hidden Layer
h
v
Visible Layer
Figure 2: Illustration of a standard RBM
for classification. Larochelle and Bengio [25] introduce a discriminative RBM
as a classifier by including the labels in visible layer, and make predictions by
comparing the likelihood of each label vector. In [26] discriminate RBM is
introduced to model vector inputs by duplicating the discriminative RBM and
adding constraints on the hidden layer. In all RBM related models discussed
above, one RBM is trained for all classes. For this work, in contrast, we build
one RBM for each class, and perform a multi-class classification task.
In this research, we propose to extend the standard RBM to simultaneously
capture the spatial and global temporal patterns in high-dimensional sequential
data and employ such RBM model to different classification tasks in computer
vision.
3
Data Spatio-Temporal Pattern Modeling
In this section, we will firstly give a brief introduction of restricted Boltzmann
machine, and then introduce the proposed RBM with Local Interaction (LRBM)
to model multi-dimensional motion data and its learning method. Finally we
present the method to use LRBMs as a set of pairwise classifier to perform
multi-class recognition.
3.1
Restricted Boltzmann Machine
A standard RBM is a generative model with two densely connected layers, one
visible layer to represent data and one latent layer to extract stochastic binary
features from data (Figure 2). Hidden units are connected to visible nodes
using symmetrically weighted connections to model their joint distribution. In
our work, we use Gaussian-Binary RBM, where the hidden units are binary and
the visible variables are assumed to follow normal distribution.
The energy function E(v, h) for each pair (v, h) is parameterized in Equation
5
1.
E (v, h) =
X (vi − ai )2
2σi2
i
−
X vi
X
wij hj −
bj hj ,
σj
i,j
j
(1)
where ai is the bias for visible unit vi , σi is the standard deviation of the
Gaussian distribution, which is typically 1 if we normalize the data, bj is the
bias for the hidden unit hj , wij is the weight of the link connecting vi and hj .
For every possible pair of visible and hidden vector, the network assigns a
probability:
1
(2)
p (v, h) = exp (−E (v, h)) ,
Z
where Z is the partition function. With continuous inputs, the Z is calculated
by integrating over all visible nodes and summing over all hidden unit configurations.
3.2
Proposed Model
Using standard RBM to model high-dimensional motion data has its limitations.
First, the interaction among the data is represented through latent variables,
which can be easily represented in direct connections. Second, if a single RBM
models the whole sequence, the spatial information in a time slice is treated the
same as the temporal information of one dimension. If one RBM models only
one time slice (as in [21]), the temporal information remains local.
In this work, we propose the restricted Boltzmann machine with local interaction (LRBM, Figure 3) to overcome such limitations. For d × nt sequential
data V = [v1 , v2 , . . . , vnt ], by allowing local interaction, we have the following
energy function,
E(V, h) =
nh
nt
X
1X
(vi − ai )T (vi − ai ) −
bj hj
2 i=1
j=1
−
nh
nt X
X
n
viT wij· hj −
i=1 j=1
t
1X
vT Uvi ,
2 i=1 i
(3)
where vi is a d-dimension vector representing input vector at time slice i, w
has the dimension of nt × nh × d, wij· is the weight vector connecting vi to a
hidden node hj . ai and bj have the same meaning as in Equation 1. U is a
d × d symmetric matrix with zeros on diagonal, modeling the correlation of each
vector vi .
The proposed energy function models two kinds of data interactions: interaction through latent variables and interaction directly among data. In highdimensional motion data, components in a single time slice (spatial information)
is better to be considered differently from components along the timeline (temporal information). Interactions between input data and hidden variables are
through weight w, which models global pattenrs, because every visible unit is
connected to every latent unit. Matrix U models direct interactions among
6
input data, as in Figure 3, representing local spatial patterns. This kind of interaction directly affects visible layer without going through latent layer, so it is
more effective to model some local spatial patterns. The elements in U are pairwise potentials of features in one time slice. Different shapes (spatial pattern) of
input will have different contributions to the energy function. Thus, U captures
spatial patterns among elements of input vector. With temporal information
captured via the hidden nodes, our work can capture global spatio-temporal
dynamics.
We assume the spatial relationship are constant throughout the whole sequence, so the parameters in U are shared in different frames, which means U is
invariant of i. Under this invariance assumption of U, the number of parameters
is significantly reduced.
From the energy function (Equation 3) and likelihood function (Equation
2), we can derive the probability of an hidden unit to be activated, given an
visible layer:
X
p(hj = 1|V) = sigmoid(bj +
viT wij· ) .
(4)
i
As one hidden unit is connected to all visible units, whether it is activated
or not depends stochastically on the visible layer. A hidden unit hj is activated
through Equation 4 when it detects some specific pattern in the visible layer.
The pattern is captured by the weights connecting each element in V to hj .
Therefore the hidden layer h is able to capture important global patterns of
V. Hence, the proposed model can simultaneously capture the global temporal
patterns (through w) and local shape patterns (through U).
LRBM is similar to the factored 3-way RBM [19] and SRBM [11] with respect
to modeling the interactions in the visible layer. However, there exist some
significant differences. First, in factored 3-way RBM, every pair of visible units
has a potential to represent the interaction, while in LRBM, units only interact
with neighbors in the time slice. Second, if the number of factors in 3-way
RBM or the data dimension in SRBM is high, the overall parameters are much
more than LRBM. In LRBM, the additional parameters are from matrix U,
with the d(d − 1)/2 parameters, where d is the dimension of the vector vi . This
would greatly reduce the computational load. Finally, both 3-way RBM and
semi-RBM are proposed to model the spatial patterns of a single natural image,
while LRBM models spatio-temporal patterns of sequential data.
The joint probability of V and h are the same as in Equation 2. The
likelihood of input data is computed by summing over all configurations of
hidden units:
1 X −E(V,h)
e
.
(5)
p(V) =
Z
h
3.3
Model Learning
In our work, one LRBM is trained for one class. The parameters include bias
for visible and hidden units ai and bj , weight between two layers w, and local
interaction matrix U. To learn the parameters, we seek to maximize the joint
7
ℎ𝑗
Hidden Layer
…
(𝑟)
(𝑠)
𝑣𝑡
𝒗1
𝑢𝑟𝑠
𝑣𝑡
…
𝒗𝑛 𝑡
𝒗𝑡
Slice 𝑡
Slice 1
Visible Layer
𝑤𝑖𝑗
Slice 𝑛𝑡
T
Figure 3: Illustration of the proposed LRBM model. Each hidden unit connects
to each visible unit to capture the global temporal patterns. (Some links are
omitted for brevity.) Connections within each time slice are to capture spatial
pattern. The local covariance is parameterized using symmetric matrix U, which
is consistent over time.
Q
probability of all training data D of a class, P (D) = V∈D p(V). Assume the
data has been normalized, so ai can be removed from the energy function. The
derivative of the log-likelihood of a training instance with respect to a parameter
is given below:
∂ log p(V)
∂wij
∂ log p(V)
∂bj
∂ log p(V)
∂urs
= hvi hj idata − hvi hj imodel ,
(6)
= hhj idata − hhj imodel ,
(7)
= h
nt
X
(r) (s)
vi vi idata
i
nt
X
(r) (s)
vi vi imodel ,
−h
(8)
i
where the angle brackets are used to denote expectations under the distribution
specified by the subscript that follows. urs is the element of U at position (r, s).
(r)
(s)
vi , vi are the rth and sth components of vector vi . This leads to a simple
learning rule: stochastic steepest ascent in the log-likelihood of training data.
Take wij as an example:
∆wij = (hvi hj idata − hvi hj imodel ) ,
(9)
where is the learning rate. hvi hj idata is easy to get from the data. hvi hj imodel
is much more difficult to compute due to the large dimension of hidden and vis8
ible layers. Hinton [27] proposes the CD algorithm to approximate hvi hj imodel
by using a reconstructed sample, which gives the following weight change:
∆wij = (hvi hj idata − hvi hj irecon ) .
(10)
Given the visible layer, the hidden units are independent with each other, so
sampling the states of hidden units can be performed in parallel using Equation
4.
Given the states of the hidden units, the visible units form a Markov Random
(r)
(s)
Field in which the pairwise interaction weight between vi and vi is urs . They
are no longer independent, so sampling cannot be done in parallel. However, we
can obtain the conditional probability of each node by fixing its neighbors. The
mean field algorithm is used to sample data from the model, and to reconstruct
visible layer when learning the parameters of LRBM. Specifically, for each vector
(s)
vi , each component vi is sampled by fixing all its neighbors. Once a component
is sampled, the vector is updated with the newly sampled component. This
procedure is repeated until all components have been updated, and thus we get
a reconstructed vector. The conditional probability of one visible node is given
in Equation 11:
X
X
(s)
(s)
(s)
(s)
p(vi |N(vi ), h) = N (ai +
hj wij· +
vk uks , 1) ,
(11)
j
(s)
(s)
vk ∈N(vi )
(s)
where N(vi ) are the neighbors of vi , N (µ, σ 2 ) denotes the Gaussian probability density function with mean µ and variance σ 2 . The superscript (s) means
the sth component of a vector.
Compared with distribution of visible units in standard RBM:
X
p(vi |h) = N (ai +
hj wij , 1) ,
(12)
j
the mean of the distribution of a visible node is similar in the first two terms,
except that in LRBM, it is modified by the pairwise potentials relating one node
to all its neighbors.
With the reconstructed data, we are able to calculate the approximate gradient of each parameter, and perform the CD learning procedure. As mentioned
in [28], the RBMs learn better if more steps of reconstruction are used before
collecting the statistics. In practice, we sample 10 times using the mean field
method before collecting the reconstructed data in each epoch.
Due to the non-convex property of the objective function, different parameter
initializations in RBM learning can end up with different models. For each class,
several candidate models are learned from different initializations, and we select
the one that can best discriminate current class from the others.
For example, after learning model M1 for class C1 , given instances from
all classes {I1 , I2 , · · · , IN }, we expect instance I1 to have greater likelihood on
model M1 than all other instances ({I2 , · · · , IN }). In practice, we can sort the
9
likelihoods of instances from all classes, and choose the candidate model that
minimizes the rank of the instances from class C1 .
Notice that the likelihood is calculated according to Equation 5, given a
single LRBM model, so for the comparison of the likelihoods, we don’t need to
estimate the intractable partition function, since it is merely a constant.
3.4
Multi-class Classification
For this work, RBM is used as a generative model for classification. A common
strategy for training generative model for classification is to train multiple models for different classes and then evaluate the likelihood of each model during
testing. In particular, given a set of properly learned LRBM’s {Mi }N
i=1 for N
classes, the basic idea for classification is to find the model that generates the
largest likelihood given an instance of data V.
i∗ = arg max p(V|Mi ) ,
i
(13)
where i∗ is the prediction of our classifier. The likelihood of a data instance is
nt
nt
X
1X
(vi − ai )T (vi − ai ) +
viT Uvi
2 i=1
i=1
!!
nh
nt
X
X
viT wij
+
log 1 + exp bj +
− log Z .
log p(V) = −
(14)
i=1
j=1
For brevity, we denote p(V) as
p(V) = g(V) − log Z .
(15)
g(V) can be computed directly. However the partition functions Z are intractable for RBMs with large numbers of hidden and visible units. One possible solution is the Annealed Importance Sampling [29] to estimate the partition
functions, and directly compare the likelihood. However, such estimation needs
too many samples to obtain a good estimation of the partition function. Instead,
for binary classification, Schmah et al. [30] propose a method to discriminatively
estimate the difference of log-partition functions of two RBMs:
cij = log Zi − log Zj .
(16)
We extend this method to multi-class classification. Ideally, cij = cik + ckj .
But since the partition function is not directly computed, there is a slight error
in the estimation. To address this issue, we perform a label ranking process [31]
to make the final decision using a set of pairwise classifiers.
In the simplest case, with all the pairwise classifiers fij , each prediction is
interpreted as a vote for a class, and the class with the highest votes is proposed
as a prediction. Alternatively, in confidence estimation, instead of a binary
10
result {0, 1}, a “soft” classifier is employed to map the difference in likelihood
into the unit interval [0, 1]:
Fij (V) =
1
.
1 + exp(−α(g(V|Mi ) − g(V|Mj ) − cij ))
(17)
Parameter α is searched in a range of [0.01, 100] to maximize the training
accuracy.
The output of such “soft” binary classifier can be interpreted as a confidence
value in the classification: the closer the output Fij to 1, the stronger the
decision of choosing class Ci is supported. Notice that Fij is not symmetrical.
It only gives the preference toward i between i and j. The preference toward
j is 1 − Fij . A valued preference relation matrix RV is defined for any query
instance V:
Fij (V)
if i < j
RV (i, j) =
.
(18)
1 − Fij (V)
if i > j
In our approach, we evaluate the confidence score by summing up all the
confidence value
X
SV (i) =
Rv (i, j) ,
(19)
j6=i
where the index i goes through 1 to N , meaning the confidence score for one
instance on different classes. The label of the model that associates with the
highest score is proposed as the final decision.
i∗ = arg max SV (i) .
(20)
i
In general, if we have a N -class classification problem,
pairwise classifiers are used to compute the confidence score.
4
n
2
= n(n − 1)/2
Experiments
We evaluate our algorithm on two different areas that involve high-dimensional
motion data: facial expression recognition and human action recognition. Two
benchmark data sets are used in our experiment: the extended Cohn-Kanade
data set (CK+) [2] and G3D data set [12]. We will also compare the proposed
LRBM model with other related methods.
4.1
Facial Expression Recognition
The Extended Cohn-Kanade data set (CK+) is a complete data set for action
units and emotion-specified expressions. In this work, based on a sequence
of landmarks on the human face, our goal is to classify it into one of seven
expressions: angry, disgust, fear, happy, sadness, surprise and contempt. CK+
data set includes 593 sequences from 123 subjects. From all the sequences, 327
are associated with expression labels. In our experiment, these 327 sequences
11
300
250
200
150
100
50
0
0
50
100
150
200
250
300
Figure 4: The 15 points (circled dots) we selected as the input data for each
time slice. Because of facial symmetry, we only select points on the left side of
the face, and points along the vertical center line.
Table 1: Confusion Matrix of the classification performance of LRBM on CK+
data set (%). For brevity, An, Di, Fe, Ha, Sa, Su, Co correspond to angry,
disgust, fear, happy, sadness, surprise, contempt, respectively.
An
Di
Fe
Ha
Sa
Su
Co
An
97.8
8.5
0.0
0.0
10.7
0.0
5.6
Di
0.0
89.8
0.0
0.0
3.6
0.0
0.0
Fe
0.0
0.0
84.0
0.0
3.6
0.0
0.0
Ha
0.0
1.7
8.0
100.0
0.0
0.0
11.1
Sa
0.0
0.0
0.0
0.0
78.6
1.2
11.1
Su
2.2
0.0
8.0
0.0
3.6
97.6
0.0
Co
0.0
0.0
0.0
0.0
0.0
1.2
72.2
are used. Landmarks are the positions of 68 facial points from a detection and
tracking procedure, which are provided by the database.
As the pose of each subject varies slightly over time, we apply a pose rectification procedure to make each face a frontal one. Then, using the detected eyes
and the interocular distance, we perform a geometric normalization by making
the size of facial area the same throughout all subjects, and moving the centers
of eyes to the same position. A smoothing filter is also used to reduce the noise
in the trajectory. To reduce the feature dimension while keeping the useful information, as shown in Figure 4, we omit the outline landmarks and some other
points, and select the more informative points, resulting in a 15-point feature
for each frame, which is a 30d vector. The features selected are reasonable due
to the property of facial symmetry.
12
Table 2: Classification accuracy of LRBM and other state-of-the-art approaches
(%)
An
Di
Fe
Ha
Sa
Su
Co
Avg.
4.1.1
Lucey et al. [2]
75.0
94.7
65.2
100.0
68.0
96.0
84.4
83.3
Wang et al. [32]
91.1
94.0
83.3
89.8
76.0
91.3
78.6
86.3
LRBM
97.8
89.8
84.0
100.0
78.6
97.6
72.2
88.6
Classification Performance
Intuitively, the spatial patterns are crucial in recognizing an expression. From
a single face, human can identify the expression without any difficulty. But in
order to explore the underlying facial dynamics as additional features for our
system, we choose 10 frames from each sequence, representing the neutral look,
intermediate expressions and the peak expression. Since our method captures
global temporal patterns, a good alignment of different sequences is important to
ensure recognition performance. Linear interpolation is used to get sequences
with fixed length. The size of hidden layer is set as 400. To compare with
baseline method, we use a leave-one-subject-out cross-validation configuration.
Each time we generate the testing data from sequences of one subject. 25 other
subjects form the validation set, and all the other subject form the training set.
The performance of the algorithm is given in Table 1. As we only use the
shape features (i.e. facial landmark positions), the comparison is only between
methods using the same features. Our method’s hit rates for each expression are:
Angry - 97.8%, Disgust - 89.8%, Fear - 84.0%, Happy - 100.0%, Sadness 78.6%,
Surprise - 97.6%, Contempt - 72.2%. The average accuracy of all the expressions
is 88.6%, which is more than 2% better than state-of-the-art method, as reported
in [32]. In Table 2, we list the comparison of our model with some recent
approaches. Our approach reaches the best accuracy on five expressions (angry,
fear, happy, sadness and surprise) among all the methods listed. Although
our model does not achieve the best accuracy for every expression, it does not
fall behind very much on Disgust, but it fails sometimes on Contempt, because
this is a subtle expression that our method needs more features to represent
it accurately. In CK+ data set, contempt expression has only 18 samples.
However, the average performance of the proposed method has been improved
a lot. Another thing to mention is that the method in [2] uses both shape
features and appearance features, while we only use the shape features.
If we do not consider the spatial interaction within each frame, the data
is simply aligned as a long vector to feed in the conventional RBM. Then the
13
Angry
Disgust
Fear
1
1
1
0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0
0
0.2
0.4
0.6
0.8
1
0
0
0.1
0.2
(a)
0.4
0.6
0.8
1
0
0
Happy
Sadness
1
0.9
0.9
0.9
0.8
0.8
0.8
0.7
0.7
0.7
0.6
0.6
0.6
0.5
0.5
0.5
0.4
0.4
0.4
0.3
0.3
0.3
0.2
0.2
0.2
0.1
0.1
0.4
(d)
0.6
0.8
1
0
0
0.6
0.8
1
0.8
1
Surprise
1
0.2
0.4
(c)
1
0
0
0.2
(b)
0.1
0.2
0.4
(e)
0.6
0.8
1
0
0
0.2
0.4
0.6
(f)
Figure 5: ROC curves of six emotion classifiers: (a) anger, (b) disgust, (c) fear,
(d) happy, (e) sadness, (f) surprise.
performance is 86.3%, which is less than the LRBM. This also proves the improvement of LRBM over RBM.
We also compare our method with 4 best available results to date, including
Time-series Kernels (Lorincz et al. [33]), spatio-temporal independent component analysis (Long et al. [34]), boosted dynamic features (Yang et al. [35]),
and non-negative matrix factorization techniques (Jeni et al. [36]). These methods are different from the baseline method in [2], because contempt emotion
is removed from the data set, and binary classification is performed for each
expression, so for the classification of one expression, the other five expressions are negative samples. Following the same experiment settings, the result
of classification is shown in Figure 5, and detailed comparison is in Table 3.
The classification performance of our algorithm is close to, or even better than
state-of-the-art method, with a better average accuracy. This demonstrate the
effectiveness of the proposed LRBM model as both a binary classifier and a
multi-class classifier.
14
Table 3: AUC values of LRBM and the competing methods: Time-series Kernels
(Lorincz et al. [33]), spatio-temporal independent component analysis (Long et
al. [34]), boosted dynamic features (Yang et al. [35]), and non-negative matrix
factorization techniques (Jeni et al. [36])
Method
Yang et al. [35]
Long et al. [34]
Jeni et al. [36]
Lorincz et al. [33]
LRBM
An
0.973
0.933
0.989
0.991
0.992
Di
0.941
0.988
0.998
0.994
0.995
Fe
0.916
0.964
0.977
0.987
0.995
Ha
0.991
0.993
0.998
0.999
0.999
Sa
0.978
0.991
0.994
0.995
0.997
Su
0.998
0.999
0.994
0.996
0.995
Avg.
0.966
0.978
0.992
0.994
0.996
Table 4: Confusion Matrix of the classification performance using the feature
learning method.
An
Di
Fe
Ha
Sa
Su
Co
4.1.2
An
88.9
11.9
0.0
0.0
17.9
0.0
5.6
Di
11.1
86.4
4.0
0.0
0.0
2.4
5.6
Fe
0.0
0.0
36.0
1.4
3.6
0.0
5.6
Ha
0.0
1.7
12.0
97.1
0.0
0.0
27.8
Sa
0.0
0.0
0.0
0.0
57.1
0.0
5.6
Su
0.0
0.0
48.0
0.0
10.7
97.6
5.6
Co
0.0
0.0
0.0
1.4
10.7
0.0
44.4
Comparison with Feature Learning Method
To comprehensively evaluate our method, we compare with a feature learning
method, since RBMs are typically used for feature learning. A single RBM
is trained on all classes of sequences using contrastive divergence. The size of
hidden layer is the same as in the generative model. In the test process, given
each observation sequence V, we compute the posterior probability P (hj |V) for
each hidden variable hj according to Equation. 4, as the feature representation.
This is reasonable because P (hj = 1|V) is the probability of a pattern being
activated. A linear multi-class SVM is trained on such features for classification.
The confusion matrix of this method is given in Table 4. For most expressions, the feature-based method is not as good as the generative model. The
average recognition accuracy is 83.8%, which is approximately 5% below the
performance of the generative model. This is reasonable, because the features
learned from the model cannot capture the spatial relationship in each frame,
which is specifically modeled in our model using the matrix U . If we add another latent layer, the classification performance only increases marginally by
about 1%.
15
PunchRight 100
PunchLeft
KickRight
KickLeft
Defend
100
100
100
100
Golf
SwingFore
25.0
75.0
10.0 80.0
10.0
90.0
Serve
Bowling
10.0
90.0
SwingBack
44.4
10.0
44.4
Aim 7.5
11.1
92.5
Walk
91.7 8.3
Run
100
Jump
33.3
Climb
100
Crouch
100
Drive
100
Wave
27.3
72.7
Flap
100
Clap
t
t
t
t
d
igh hLef Righ kLef efen
c
k
hR
D
Kic
nc Pun Kic
Pu
66.7
100
lf
e
k
g
e
Go gFor gBac Serv owlin
in
B
in
Sw Sw
k
Aim Wal
e
e
h
p
n
b
p
p
Ru Jum Clim rouc Driv Wav Fla Cla
C
Figure 6: The confusion matrix of the proposed method on G3D data set
4.2
Human Action Recognition
G3D data set is an action data set containing a range of gaming actions captured
by Microsoft Kinect. The data set contains 10 subjects performing 20 gaming
actions: punch right, punch left, kick right, kick left, defend, golf swing, tennis
swing forehand, tennis swing backhand, tennis serve, throw bowling ball, aim
and fire gun, walk, run, jump, climb, crouch, steer a car, wave, flap and clap.
Synchronized video, depth and skeleton data are available in this data set. We
only pick the skeleton data. To reduce the data dimension but keep the useful
information, we use the 3D location information of four dominant joints (i.e.
two hands and two feet). However the proposed approach can be applied to
modeling more joints if we have enough training data.
Before abstracting the features, we perform a normalization procedure to
minimize the effect brought by difference in subjects’ body shapes. Specifically,
we obtain the average bone lengths from all subjects in the training data, and
then change the skeleton tracking results in every frame for each subject with
the average shape. Therefore, every subject has the same body shape, and the
cross subject distinction is alleviated.
As the size of the visible layer of an RBM is fixed, linear interpolation is
performed to convert all sequences into the same length (20 frames for each
sequence). The 3D positions of the body joints along all three dimensions (x,
y and z) form the 240-dimension input for the LRBM model. In the training
phase, we set the size of hidden layer as 80.
16
Table 5: F1 score of LRBM and baseline model
Action
Fighting
Golf
Tennis
Bowling
FPS
Driving
Misc
Avg.
Bloom et al. [12]
70.46
83.37
56.44
80.78
53.57
84.24
78.21
72.44
LRBM
97.09
81.82
84.38
61.54
95.52
100.00
95.24
87.94
We use the action segmentation that is provided by the data and the same
experiment configuration as in [12]. The data set is split by subjects where
the first 4 subjects were used for training, 1 for validation and the remaining 5
subjects for testing.
4.2.1
Classification Performance
The confusion matrix is given in Figure 6. The overall accuracy of LRBM is
90.5%. Our model encounters some failures for the actions of ThrowBowlingBall
in Bowling category and Jump. In these actions, occasionally the body parts are
occluded from the Kinect sensor, which will lead to estimated positions of joints.
Then the positions of the two feet are significantly corrupted. As our method
depends fully on the trajectories of joints, corrupted tracking results have a huge
influence on the performance of LRBM. If we remove the pairwise potential in
the visible layer and make the model a standard RBM, the accuracy drops to
84%, which proves the better performance of LRBM than RBM, as expected,
due to modeling the spatial interactions. Comparison with the baseline model
[12] is given in Table 5. Other than Bowling, performance on golf action is
quite close to the baseline method, and on all the other actions, our approach
outperforms the baseline method. The overall accuracy is increased by 15.5%
in terms of F1 score.
One reason of the improvement is that the method in [12] is based on 3
frames, hence can only capture local dynamics, while the proposed method is
sequence based that can handle global dynamics.
To further demonstrate the effectiveness of global dynamic model over local
dynamic model, we implement a conditional RBM [21] and a hidden Markov
model for action recognition. For the conditional RBM model, each frame depends on 3 previous frames. 20 models are trained for 20 actions. Classification
is based on the likelihood of a target sequence on different models. Basically
the local dynamic model (CRBM or HMM) models one frame at a time and the
global model (RBM or LRBM) models all frames at the same time.
Similar to the expression recognition case, the posterior probability of the
17
Table 6: Recognition accuracy of different dynamic models. The feature learning method is denoted as n-FL-SVM for the feature learning nature and SVM
classifier, with n hidden layers.
Method
HMM
1-FL-SVM
2-FL-SVM
CRBM [21]
RBM
LRBM
Accuracy (%)
70.3
75.9
76.7
83.2
84.0
90.5
hidden variables P (hj |V) can be used as the features for classification. We
implement the feature-based method using linear SVM as the classifier. Again,
it does not perform well compare with other dynamic methods, mainly due
to the reason that it cannot capture the local spatial patterns in each frame.
Details are given in Table 6.
4.2.2
Handling Noisy and Missing Data
One advantage of generative model is that they can handle noisy or missing data
in the input. To demonstrate this point, we design two scenarios with randomly
selected noisy or missing data:
(a) Noisy data. For a randomly selected portion of data, we multiply the
ground truth data by a Gaussian noise 1 + N (0, 1), where N (0, 1) is the normal
distribution;
(b) Missing data. For a randomly selected portion of data, we assume they
are missing, and use the average of their neighbors as the approximation when
computing the likelihood.
The data under these scenarios are then fed into our generative model for
classification purpose. We vary the portion of noisy or missing data from 0 to
50%, and observe the change of the classification accuracy. The details are given
in Figure 7.
In scenario (a), with the portion of noisy data increase, the performance
almost decays linearly. In scenario (b), if the percentage of missing data is
small (less than 30%), the average of its neighborhood is a decent estimation
of the missing value, but with more data missing, the performance decays quite
significantly. This is because several consecutive frames are missing at the same
time, and the temporal information cannot be recovered effectively. One thing
to notice is that with as much as 30% noisy or missing data, our algorithm can
achieve 86% accuracy, only 4% below the noiseless situation, which demonstrates
the effectiveness of our algorithm to handle noisy or missing data.
18
95
Noisy
Missing
Classification accuracy
90
85
80
75
70
0
5
10
15
20
25
30
35
40
45
50
Percentage of noisy or missing data
Figure 7: The classification performance with noisy or missing data.
4.3
Complexity
In the learning procedure of LRBM, we compute the data-dependent expectation and model-dependent expectation using matrix multiplication, so the
computational complexity is linear with the size of the weight matrix (dnt × nh )
and iteration times (niter ). Thus, the computation complexity is O(dnt nh niter ).
For facial expression recognition task, we set the epoch to 250, and the average
running time for training one RBM is 1.77s, compared with 1.30s for standard RBM without local interaction. With 10 candidate models, it takes less
than 5 min to train all the models. Experiments are performed on a desktop
computer with an Intel i7 3.4GHz CPU and 8GB RAM. For testing, given a target sequence, we need to compute the unnormalized likelihood on each LRBM
model using Equation 14, with computational cost O(nt d2 + nh nt d). Given
the unnormalized likelihood and the relative partition functions, the classification needs n2 comparisons, therefore the overall classification complexity is
O(n(nt d2 + nh nt d) + n2 ), which is linear in the dimension of the input sequence.
If we linearly increase the length of the sequence by increasing the frame rate,
the complexity is also increased linearly.
For multi-class classification with n classes, the confidence value matrix SV
is built with complexity O(n2 ). In typical computer vision applications, the
number of classes cannot exceed the dimension of the data, which means n2 does
not contribute much in the complexity, compared to the first term. To reduce
the quadratic term O(n2 ) to a linear term O(n), Annealed Importance Sampling
[29] can be used to estimate the partition function, thus the likelihood can be
directly compared. However, the sampling method requires many samples to
get a good estimation of the partition function, and the overall classification
19
complexity is not reduced significantly.
Theoretically, the proposed method can be scaled up to data with thousands
of dimensions, but much more training data will be needed since the number of
parameters increases linearly with the size of input.
5
Conclusion
In this paper, we study classification problem with multi-dimensional sequence
data. To capture both the global dynamics and the local spatial interactions,
we extend the conventional restricted Boltzmann machine by adding pairwise
potentials in the energy function. An efficient mean field learning method is proposed to replace the reconstruction procedure in typical Contrastive Divergence
learning, to estimate the additional parameters. Also, a novel label ranking
approach of estimating the partition function of different LRBMs is presented
to perform multi-class classification task. Experiments on two areas that involve multi-dimensional sequence data are performed to evaluate our approach.
Results on two benchmark data sets prove the effectiveness of our algorithm.
Using the generative model for data representation, we do not need to deal
with a deep and complicated structure, as in feature learning, and do not need
to carefully tune the parameters. This is a major advantage of our method
compared with feature learning methods. However, one shortcoming is that is
that the data sequences have to be aligned for our global model to function
properly, and the classification complexity is quadratic in the number of classes.
We will address these issues in future work.
References
References
[1] J. Shotton, A. Fitzgibbon, M. Cook, T. Sharp, M. Finocchio, R. Moore,
A. Kipman, A. Blake, Real-time human pose recognition in parts from
single depth images, in: CVPR, IEEE, 2011, pp. 1297–1304.
[2] P. Lucey, J. F. Cohn, T. Kanade, J. Saragih, Z. Ambadar, I. Matthews,
The extended cohn-kanade dataset (ck+): A complete dataset for action
unit and emotion-specified expression, in: CVPR Workshops, IEEE, 2010,
pp. 94–101.
[3] I. Laptev, On space-time interest points, International Journal of Computer
Vision 64 (2-3) (2005) 107–123.
[4] I. Laptev, M. Marszalek, C. Schmid, B. Rozenfeld, Learning realistic human
actions from movies, in: CVPR, 2008, pp. 1–8.
[5] F. Lv, R. Nevatia, Recognition and segmentation of 3-d human action using
hmm and multi-class adaboost, ECCV (2006) 359–372.
20
[6] J. Wang, Z. Liu, Y. Wu, J. Yuan, Mining actionlet ensemble for action
recognition with depth cameras, in: CVPR, IEEE, 2012, pp. 1290–1297.
[7] I. Dagli, M. Brost, G. Breuel, Action recognition and prediction for driver
assistance systems using dynamic belief networks, in: Agent Technologies,
Infrastructures, Tools, and Applications for E-Services, Springer, 2003, pp.
179–194.
[8] C. Sminchisescu, A. Kanaujia, D. Metaxas, Conditional models for contextual human motion recognition, Computer Vision and Image Understanding 104 (2) (2006) 210–220.
[9] S. A. Eslami, N. Heess, J. Winn, The shape boltzmann machine: a strong
model of object shape, in: CVPR, IEEE, 2012, pp. 406–413.
[10] Z. Wang, G. Schalk, Q. Ji, Anatomically constrained decoding of finger
flexion from electrocorticographic signals, in: NIPS, 2011.
[11] S. Osindero, G. E. Hinton, Modeling image patches with a directed hierarchy of markov random fields, in: NIPS, 2007, pp. 1121–1128.
[12] V. Bloom, D. Makris, V. Argyriou, G3d: A gaming action dataset and
real time action recognition evaluation framework, in: CVPR Workshops,
IEEE, 2012, pp. 7–12.
[13] M. Carreira-Perpinan, G. Hinton, On contrastive divergence learning, in:
Artificial Intelligence and Statistics, Vol. 2005, 2005, p. 17.
[14] H. Lee, R. Grosse, R. Ranganath, A. Y. Ng, Convolutional deep belief
networks for scalable unsupervised learning of hierarchical representations,
in: Proceedings of the 26th Annual International Conference on Machine
Learning, ACM, 2009, pp. 609–616.
[15] R. Cutler, M. Turk, View-based interpretation of real-time optical flow for
gesture recognition, in: Automatic Face and Gesture Recognition, IEEE,
1998, pp. 416–421.
[16] G. Yang, Y. Lin, P. Bhattacharya, A driver fatigue recognition model based
on information fusion and dynamic bayesian network, Information Sciences
180 (10) (2010) 1942–1954.
[17] B. Packer, K. Saenko, D. Koller, A combined pose, object, and feature
model for action understanding, in: CVPR, IEEE, 2012, pp. 1378–1385.
[18] G. Hinton, R. Salakhutdinov, Reducing the dimensionality of data with
neural networks, Science 313 (5786) (2006) 504–507.
[19] V. Nair, G. Hinton, 3-d object recognition with deep belief nets, NIPS 22
(2009) 1339–1347.
21
[20] Y. Wu, Z. Wang, Q. Ji, Facial feature tracking under varying facial expressions and face poses based on restricted boltzmann machine, in: CVPR,
IEEE, 2013.
[21] G. Taylor, G. Hinton, S. Roweis, Modeling human motion using binary
latent variables, NIPS 19 (2007) 1345.
[22] A. Kae, K. Sohn, H. Lee, E. Learned-Miller, Augmenting crfs with boltzmann machine shape priors for image labeling, in: Computer Vision and
Pattern Recognition (CVPR), 2013 IEEE Conference on, IEEE, 2013, pp.
2019–2026.
[23] A. Kae, B. Marlin, E. Learned-Miller, The shape-time random field for
semantic video labeling, in: Computer Vision and Pattern Recognition
(CVPR), 2013 IEEE Conference on, IEEE, 2014.
[24] I. Sutskever, G. E. Hinton, Learning multilevel distributed representations
for high-dimensional sequences, in: Artificial Intelligence and Statistics,
2007, pp. 544–551.
[25] H. Larochelle, Y. Bengio, Classification using discriminative restricted
boltzmann machines, in: ICML, ACM, 2008, pp. 536–543.
[26] J. Louradour, H. Larochelle, Classification of sets using restricted boltzmann machines, in: UAI, IEEE, 2011.
[27] G. Hinton, Training products of experts by minimizing contrastive divergence, Neural computation 14 (8) (2002) 1771–1800.
[28] G. Hinton, A practical guide to training restricted boltzmann machines,
Momentum 9 (2010) 1.
[29] R. Salakhutdinov, I. Murray, On the quantitative analysis of deep belief
networks, in: Proceedings of the 25th international conference on Machine
learning, ACM, 2008, pp. 872–879.
[30] T. Schmah, G. E. Hinton, S. L. Small, S. Strother, R. S. Zemel, Generative
versus discriminative training of rbms for classification of fmri images, in:
NIPS, 2008, pp. 1409–1416.
[31] E. Hüllermeier, J. Fürnkranz, W. Cheng, K. Brinker, Label ranking by
learning pairwise preferences, Artificial Intelligence 172 (16) (2008) 1897–
1916.
[32] Z. Wang, S. Wang, Q. Ji, Capturing complex spatio-temporal relations
among facial muscles for facial expression recognition, in: CVPR, 2013.
[33] A. Lorincz, L. A. Jeni, Z. Szabó, J. F. Cohn, T. Kanade, Emotional expression classification using time-series kernels, in: Computer Vision and Pattern Recognition Workshops (CVPRW), 2013 IEEE Conference on, IEEE,
2013, pp. 889–895.
22
[34] F. Long, T. Wu, J. R. Movellan, M. S. Bartlett, G. Littlewort, Learning
spatiotemporal features by using independent component analysis with application to facial expression recognition, Neurocomputing 93 (2012) 126–
132.
[35] P. Yang, Q. Liu, D. N. Metaxas, Boosting encoded dynamic features for
facial expression recognition, Pattern Recognition Letters 30 (2) (2009)
132–139.
[36] L. A. Jeni, J. M. Girard, J. F. Cohn, F. De La Torre, Continuous au intensity estimation using localized, sparse facial feature space, in: Automatic
Face and Gesture Recognition (FG), 2013 10th IEEE International Conference and Workshops on, IEEE, 2013, pp. 1–7.
23
| 1cs.CV
|
2013 IEEE International Conference on Computer Vision
Slice Sampling Particle Belief Propagation
Oliver Müller, Michael Ying Yang, and Bodo Rosenhahn
Institute for Information Processing (TNT), Leibniz University Hannover, Germany
arXiv:1802.03275v1 [cs.CV] 9 Feb 2018
{omueller, yang, rosenhahn}@tnt.uni-hannover.de
Abstract
#1
···
Inference in continuous label Markov random fields is
a challenging task. We use particle belief propagation
(PBP) for solving the inference problem in continuous label space. Sampling particles from the belief distribution is
typically done by using Metropolis-Hastings (MH) Markov
chain Monte Carlo (MCMC) methods which involves sampling from a proposal distribution. This proposal distribution has to be carefully designed depending on the particular model and input data to achieve fast convergence. We
propose to avoid dependence on a proposal distribution by
introducing a slice sampling based PBP algorithm. The proposed approach shows superior convergence performance
on an image denoising toy example. Our findings are validated on a challenging relational 2D feature tracking application.
#467
···
···
Figure 1. Relational 2D feature tracking example.
pend on a proposal distribution which is difficult to tune.
We show the superiority of our method theoretically on a
simplified toy application on image denoising. Our findings are then verified on a complex 2D relational feature
tracking application as shown in Fig. 1. We furthermore
provide a publicly available database of image sequences
for feature tracking applications including manually labeled
groundtruth data [11].
The rest of the paper is organized as follows. Section 2
provides an overview of related work. Section 3 introduces notations and definitions used throughout the paper
and gives a short introduction to slice sampling. Our proposed approach is described in detail in Sect. 4. In Sect. 5
we present a thorough evaluation of our method compared
to the state-of-the-art and propose a 2D relational feature
tracking application. We conclude our findings in Sect. 6.
1. Introduction
Markov Random Fields (MRFs) are a powerful tool for
modeling relational dependencies among observations. Inference in such models is an inherent problem which has
been widely addressed in the past. MRFs, and hence its inference methods, can be classified in two categories: discretely and continuously labeled problems. Numerous optimization approaches for discrete labels have been proposed, from binary labeled Graph Cuts [4], to multi-label
tree reweighted message passing [17, 7]. In this paper, we
deal with continuous labeled MRFs where we use a particle belief propagation (PBP) approach [6]. The efficiency
of such particle based approaches highly depends on the
sampling scheme used to explore the label space. Previous
approaches use Metropolis-Hastings (MH) Markov chain
Monte Carlo (MCMC) methods for particle sampling. The
performance of these methods depends on a carefully designed proposal distribution.
Contributions. We propose a novel sampling technique
for PBP based on slice sampling [12]. This method exploits
the structure of the PBP message passing equations for direct sampling from the target distribution and does not de1550-5499/13 $31.00 © 2013 IEEE
DOI 10.1109/ICCV.2013.144
#377
2. Related Work
Most works on MRF optimization specialize on a discrete
label space [4, 17, 7]. Often such approaches are hard to
apply on tasks where a continuous label space would be a
more natural choice, such as feature tracking with relational
constraints [14, 9].
Loopy belief propagation is a prominent method using a
local message passing mechanism for coordinating the optimal labeling of neighboring nodes. These methods work
on discrete label spaces. The computational complexity is
O(n2 ) over the number of discrete labels n, making computations with many labels for approximating near-continuous
models intractable [16].
Recently, message passing approaches working in continuous rather than discrete label space were proposed
1129
b(xt )
t
Ms→t (xt )
Mt→s (xs )
s
with a unary potential function ψs (xs ) and a binary potential function ψs,t (xs , xt ). Then p(x) ∝ exp [−E(x)]
defines a Markov random field (MRF).
We consider the problem of computing the maximum
p(x )1 .
marginals: μ(xs ) = max
b(xs )
b(xs )
u
Graphical Model (exemplary)
mcmc sampling
xs
x |xs =xs
3.2. Max-Product Particle Belief Propagation
Figure 2. Particle Belief Propagation framework. Left: Message
passing mechanism. Right: MCMC particle sampling of the belief
b(xs ) with an exemplary MCMC sampling chain of one particle
(blue) and its corresponding histogram (red).
In the following we summarize the max-product particle
belief propagation algorithm [8, 3]. The energy term E(x)
is approximated by particles such that the label space Ls of
each node s in the MRF is represented by a set of particles
(1)
(p)
Ps = {xs , . . . , xs }, where p is the number of particles
(i)
per node. Then the estimated belief bns (xs ) or log disbe(i)
(i)
lief Bsn (xs ) = − log(bns (xs )) of node s at iteration n is
calculated as follows [3]:
(i)
(i)
(i)
n
Bsn (xs ) = ψs (xs ) + t∈Ns Mt→s
(xs ),
(2)
[6, 8, 13, 16]. These approaches use MCMC methods to
approximate the message distributions. To the best of our
knowledge, all previously proposed MCMC based belief
propagation methods use Metropolis-Hastings (MH) sampling. This sampling strategy consists of two steps: (a) sampling a candidate particle from an easy to sample proposal
distribution, and (b) accept or reject the candidate depending on a transition probability [18]. Applying this sampling
technique involves a careful design of the proposal distribution, which is a compromise between exploring the label
space (using a broad proposal distribution) and maximizing
the transition acceptance ratio (minimize sample moves) at
the same time.
Throughout the paper we show that considering alternative sampling techniques can be advantageous. We propose
to use slice sampling [12] instead of MH, rendering proposal
distribution selection obsolete in the context of PBP.
To demonstrate superior performance of our method on
a real world problem we propose a relational feature tracking application inspired by [9, 14] in the experiment section. Some related works such as [15, 5] propose to formulate feature tracking as a discrete labeling problem and use
global optimization algorithms (i.e. linear programming or
dynamic programming). Such approaches need some sort
of label pruning in order to keep computational complexity
low. Closely related methods use belief propagation combined with particle filtering [19, 9, 14], but still use proposal
distributions for particle perturbation which introduces sensible optimization parameter tuning.
n
where the messages Mt→s
(xs ) for xs ∈ Ps from node t to
node s are:
n−1
n
Mt→s
(xs ) = min [ψs,t (xs , xt )+Btn−1 (xt )−Ms→t
(xt )].
xt ∈Pt
(3)
Note that the log disbelief Bsn (xs ) and the messages
n
Mt→s
(xs ) can be calculated for all continuous values xs ∈
Ls rather than only on the particle set Ps . On the other
hand, the messages from node s to node t are approximated only using the particles xt from the particle set
(1)
(p)
Pt = {xt , . . . , xt } of node t.
Messages and log disbeliefs are calculated iteratively for
n = 1, . . . , N iterations. An estimate of the most likely
configuration can be obtained with
x̂s = arg min BsN (xs ).
xs
The main issue in PBP lies in how to sample new particles xns ∼ Bsn (xs ). Typically, the Metropolis-Hastings
(MH) MCMC method is used. This method requires a proposal distribution q where new particles can be easily sampled from. Typically a Gaussian function q = pσ with a
predefined standard deviation σ is used.
Figure 2 shows a schematic overview of the PBP framework. Algorithm 1 summarizes the Metropolis-Hastings
based max-product particle belief propagation algorithm
(MH - PBP).
Typically, q needs to be carefully adjusted to the true
belief distribution. This introduces a dependency on prior
knowledge about how the labels are distributed in the label space. In the following we propose to replace the MH
sampling step by a slice sampling approach which does not
depend on proposal distribution selection.
3. Definitions and Notation
3.1. Markov Random Field
Let V be a set of nodes and Ns ⊂ V the set of neighboring nodes to s ∈ V. For every node s there
is a label xs
from the label space Ls . The product L = s∈V Ls is the
space of configurations x = {xs }s∈V . A Markov random
field potential energy is given by:
E(x) =
ψs (xs ) +
ψs,t (xs , xt )
(1)
s∈V
(4)
1 Backtracking can be used to compute the
arg max p(x) from the max-marginals [8].
s∈V t∈Ns
x
1130
MAP -configuration
x∗ =
Algorithm 1 MH - PBP [8, 3]
p(x)
(i)
{xs }i=1,...,p ,
proposal distribuInput: Initial set of particles:
tion pσ
(i)
0
1: Initialize the messages Mt→s
(xs ) and log disbelief Bs0 (xs )
with zero ∀s, t
2: for BP iteration n = 1 to N do
3:
for each node s and each particle i = 1, . . . , p do
(i)0
(i)
4:
Initialize sampling chain xs
← xs
5:
for MCMC iteration m = 1, . . . , M do
(i)m
(i)m−1
6:
Sample x̄s
∼ pσ (x | xs
)
from proposal distribution pσ
(i)m
7:
Calc. belief Bsn (x̄s
) from Eqs. (2), (3)
8:
Sample u ∼ U[0,1] (u)
(i)m
(i)m
) − log(u) then
9:
if Bsn (x̄s
)<Bsn (xs
(i)m
(i)m
10:
Accept: xs
← x̄s
11:
end if
12:
end for
(i)
(i)M
13:
xs ← x s
14:
end for
15:
Normalize messages and beliefs
16: end for
xm−1
m
u1
..
.
m
uL
x
(5)
∼ q(x | u
) = UA (x),
∼ q(uL | xm−1 ) = U[0,fL (xm−1 )] (uL )
(10)
∼ q(x |
m
m
u1 , . . . , u L )
= UAm (x),
(11)
4. Slice Sampling Particle Belief Propagation
Our main contribution is presented in this section. We
propose to sample particles from the belief b(xs ) using slice
sampling rather than Metropolis-Hastings sampling. For
applying the slice sampler, the sampling interval A(i)m
needs to be determined for the ith particle of node s and
for the mth MCMC iteration which we can uniformly sam(i)m
from. The superscripts (i)m are
ple the particle xs
omitted in the following for better readability.
The goal is to determine the sampling interval A. Given
the potential functions ψs (xs ) and ψs,t (xs , xt ), it is assumed that the intervals
Aψs (ū) = {xs ; ψs (xs ) ≤ ū} and
Axψts,t (ū) = {xs ; ψs,t (xs , xt ) ≤ ū}
(12)
(13)
can be computed analytically. Note that computations are
done in negative log space, thus a slice interval {x ; f (x) ≥
u} is transformed to {x ; − log(f (x)) ≤ ū}, where ū is the
negative logarithm of a uniformly sampled value.
The final sampling interval A can be computed from
these intervals as shown below. If the intervals cannot be
computed analytically then an approximated interval à may
be still computed and rejection sampling can be applied [1].
Sampling is then done by uniformly drawing the auxiliary variable u (defining the slice) and given this, uniformly
drawing the new sample from an interval A defined over u
as follows:
x
(9)
m
given an initial sample x .
Note that in the PBP framework, there is a MCMC
(i)m
(i)
sampling chain {xs
}m=1,...,M for each particle xs .
MCMC sampling could be done using several sampling techniques such as Metropolis-Hastings (MH) or Gibbs sampling (provided the conditional distributions are easy to
sample from). Metropolis-Hastings sampling has the drawback of requiring a proposal distribution. Choosing the proposal distribution is very often a difficult task and introduces
a compromise between reducing the rejection rate and obtaining large random moves [1].
In slice sampling, an auxiliary variable u ∈ R is introduced and the target distribution q(x) is extended to
1 if u ∈ [0, q(x)]
(6)
q ∗ (x, u) =
0 otherwise
m
m
∼ q(u1 | xm−1 ) = U[0,f1 (xm−1 )] (u1 )
where Am = {x ; fl (x) ≥ ul , l = 1, . . . , L} [1].
The main difficulty lies in determining the interval A.
Fortunately it turns out, that in the max-product particle belief propagation framework the sampling interval A can be
determined efficiently as shown in the following section.
0
m
x
where UI is the uniform distribution over an interval I and
A = {x; q(x) ≥ um }. Figure 3 shows an exemplary slice
sampling step.
Assume that q(x) can be decomposed in L functions
L
fl (x) such that q(x) ∝ l=1 fl (x). Then we can sample
over q(x) by introducing L auxiliary variables u1 , . . . , uL :
In this section we briefly summarize the concept of slice
sampling [12, 1] which is defined in a general MCMC sampling framework. Suppose we are given a distribution q(x)
and want to sample from this distribution, i.e. MCMC sampling of M samples x1 , x2 , . . . , xM :
um ∼ q(u | xm−1 ) = U[0,q(xm−1 )] (u)
xm
Figure 3. Slice Sampling [12, 1]
3.3. MCMC Slice Sampling
xm ∼ q(x | xm−1 ),
A
um
(7)
(8)
1131
√
solution
Aφs (ū) = {xs : (xs −ds )2 ≤ ū} = [ds − ū, ds +
√
ū]. Similarly, the closed form √
solution for
√ φs,t (xs , xt ) =
(xs − xt )2 is Axφts,t (ū) = [xt − ū, xt + ū].
Multidimensional Bounds. In order to deal with multidimensional label spaces, i.e. Ls ∈ Rd for d > 1, we
propose to randomly sample one dimension in each MCMC
step and slice sample on this dimension while the other dimensions are held fixed.
Analytic Bounds Calculation. Assume the unary and/
or binary potential functions ψs and ψst are given as an analytic function. Then one can use standard computer algebra solvers for defining Aψs (u) and/or Axψts,t (u). We have
implemented our S - PBP framework in MATLAB® with MEX
and use the MATLAB® -MUPAD® interface to solve the inequalities automatically. This way no manual work has to
be done.
Algorithm 2 S - PBP
(i)
Input: Initial set of particles: {xs }i=1,...,p
(i)
0
1: Initialize the messages Mt→s
(xs ) and log disbelief Bs0 (xs )
with zero ∀s, t
2: for BP iteration n = 1 to N do
3:
for each node s and each particle i = 1, . . . , p do
(i)0
(i)
4:
Initialize sampling chain xs
← xs
5:
for MCMC iteration m = 1, . . . , M do
(i)m−1
6:
Sample ūl = Fl (xs
) − log(ul ) where
ul ∼ U[0,1] (u) for l = 0, . . . , |Ns |
7:
Compute A(i)m from Eqs. (15), (16), (17)
(i)m
8:
Sample x̄s
∼ UA(i)m (x)
(i)m
9:
Calc. belief Bsn (x̄s
) from Eqs. (2), (3)
(i)m
) ≤ ūl for l = 0, . . . , |Ns | then
10:
if Fl (x̄s
(i)m
(i)m
11:
Accept: xs
← x̄s
12:
end if
13:
end for
(i)
(i)M
14:
xs ← x s
15:
end for
16:
Normalize messages and beliefs
17: end for
5. Experiments
5.1. Image Denoising
For analyzing the random walk behaviour of our method
we have chosen the application of image denoising due to
its relatively simple model structure. The basic image denoising model is as follows:
The log disbelief can be decomposed as follows:
B(xs ) =
|Ns |
l=0
Fl (xs )
(14)
ψs (xs ) = θ1 (xs − ds )2 ,
with F0 (xs ) = ψs (xs ) and Fj (xs ) = Mt(j) →s (xs ) where
t(j) is the j-th neighbor of s. From this follows the decomposition of the sampling interval
A=
|Ns |
Al ,
l=0
with
Al = {x ; Fl (x) ≤ ūl }.
ψs,t (xs , xt ) = θ2 min{θ3 , (xs − xt )2 }.
For minimizing particle noise in the final estimation result an annealing scheme is used where the target belief
(i)
distribution is modified to bns (xs )1/Tn , where Tn = T0 ·
n/N
is the temperature at PBP iteration n, T0 is the
(TN /T0 )
start temperature, and TN the end temperature. Given this
annealing scheme the temperature is successively reduced
for each new iteration n.
The evaluation was done on an example image as shown
in Fig. 4. The training and testing sets each include 10
noisy image instances with Gaussian noise standard deviation σ = 0.05 (where image intensity ∈ [0, 1]). Training of
the parameter vector θ = {θ1 , θ2 , θ3 } is done by minimizK
(i)
1
(i)
ing the empirical risk R(θ) = K
i=1 L(xθ , y ) given
2
(i)
the loss function L(x, y) = x − y 2 where {y , d(i) } is
the training data pair with groundtruth y(i) and noisy obser(i)
vation d(i) . xθ is the MAP estimate given d(i) and the parameter θ. Learned parameters are θ1 = 0.756, θ2 = 1.170,
θ3 = 0.0059.
Comparing S - PBP with MH - PBP. We further compared the efficiency of the slice sampling method to the
Metropolis-Hastings sampling applied on the image denoising problem. For the experimental setup we use N =
100 PBP iterations, p = 5 particles, and a temperature
schedule of T0 = 1 to TN = 10−4 . An MCMC chain
of M = 500 samples is generated for each particle and
(15)
Using the definitions of Eqs. (2, 3), we obtain for Al :
A0 = Aψs (ū1 )
Aj = {xs ; Mt(j) →s (xs ) ≤ ūj }
= {xs ; min Gxj t (xs ) ≤ ūj }
xt ∈Pt
{xs ; Gxj t (xs ) ≤ ūj }
=
(16)
xt ∈Pt
=
xt ∈Pt
Axψts,t (ūj − Bt (xt ) + Ms→t (xt )),
(18)
(17)
n−1
(xt ).
where Gxj t (xs ) = ψs,t (xs , xt ) + Btn−1 (xt ) − Ms→t
This result shows that A only depends on the given intervals Aψs (ū) and Axψts,t (ū) which are defined by the unary
and binary potential functions ψs and ψs,t . Algorithm 2
summarizes the proposed method.
We further refer to the proposed technique as S - PBP
(slice sampling particle belief propagation).
Example. Consider a quadratic unary potential function
φs (xs ) = (xs − ds )2 . Then Aφs (ū) has the closed form
1132
Empirical Risk
Autocorrelation ρk
Figure 4. Denoising example: Groundtruth (left), noisy input example (middle left), reconstruction with MH - PBP (middle right),
reconstruction with our proposed S - PBP method (right).
0.4
1
0.8
0.6
0.4
0.2
0
−0.2
mh-pbp
s-pbp
1
5
10
Order k
15
20
Figure 6. Comparison of S - PBP and MH - PBP at different PBP iterations (dotted n = 30, dashed n = 50, and solid n = 70) using
an annealing schedule.
s-pbp
mh-pbp
0.3
0.2
0.1
5.2. Relational Feature Tracking
bp 0.1 0.3 0.5 0.7 0.9 1.0 1.5 2.0
s-p σ= σ= σ= σ= σ= σ= σ= σ=
mh mh mh mh mh mh mh mh
We propose to apply our S - PBP algorithm on a 2D relational feature tracking system inspired by [9, 14] as a more
complex application.
Figure 5. Comparison of the empirical risk for S - PBP and MH - PBP
with different proposal distributions.
5.2.1
in each PBP iteration. The iteration numbers are chosen
to be more than sufficiently large in order to guarantee
convergence and to collect statistical information in the
MCMC chains in steady-state situations. For the MH - PBP
proposal distribution the family of Gaussian distributions
pσ (x | xm−1 ) = (2πσ 2 )−0.5 · exp[−0.5(x − xm−1 )2 ·
σ −2 ] is used. In order to provide a fair comparison the proposal distribution is adapted to the current temperature by
using pσ (x | xm−1 )1/Tn instead.
Figure 5 shows a comparison of the empirical risk for
different MH - PBP proposal distributions. For σ > 0.7 the
empirical risk stays nearly at the same level and thus we
selected σ = 0.7 for further experiments. Another observation is that S - PBP outperforms MH - PBP in terms of minimal
empirical risk. This is because the reconstructed images
with MH - PBP have always much higher noise than images
reconstructed with S - PBP. This effect can be significantly
reduced by averaging over particles instead of only selecting the best one as stated in Eq. (4).
For comparing the random walk behavior of the MCMC
sampling chains from S - PBP and MH - PBP, the normalized
autocorrelation function
M −k m
(x
− x̄)(xm−k − x̄)
ρk = m=1M −k
,
(19)
m − x̄)2
m=1 (x
m
1
where x̄ = M
x , is used [18]. Only the last 50 %
of the MCMC chain is considered to skip any burn-in phase.
Figure 6 shows a comparison of the first 20 kth order autocorrelation of S - PBP and MH - PBP at different PBP iterations
n (and thus at different temperatures Tn ). It can be observed
that the MH - PBP method produces a much higher autocorrelation than the S - PBP method, thus the MCMC chain mixing
behaviour of S - PBP outperforms MH - PBP.
Tracker Model
The proposed feature tracker uses a pairwise MRF model.
The model is separated into two parts: (a) the unary potentials are derived from a feature patch matching model, and
(b) the binary potentials encode the relative positioning of
the features to each other. The label space of the MRF is the
space of feature poses including the local central patch position, patch rotation, and scale. The proposed MRF model
is as follows:
E(x) =
ψs (xs ) +
α · ψs,t (xs , xt ), (20)
s∈V
s∈V t∈Ns
where the unary potential function
ref
ψs (xs ) = χ2 (HOGIn (ps , os ) − HOGI ref (pref
s , os ))
(21)
is the Chi-square distance of HOG features [10] of a patch
at position ps ∈ R2 of the current image In and orientation os ∈ R2 , where xs = {ps , os } and a reference image
ref
I ref at reference position pref
s and orientation os . The orientation vector os encodes two aspects: the feature patch
rotation (rotation of os , i.e. atan2(os )) and feature patch
scale (length of os , i.e. os 2 ).
The binary potential ψs,t (xs , xt ) is as follows:
ψs,t ( · ) =
pt − ps − Rs dst 22 + ps − pt − Rt dts 22
2 · dst 22
(22)
ref
where dst(ts) = pref
t(s) −ps(t) and Rs(t) = [ox,s(t) , −oy,s(t) ;
oy,s(t) , ox,s(t) ] is a 2 × 2 rotation and scale matrix. The
proposed binary potential function models the surrounding of each feature point as a weak-perspective model and
1133
5.2.3
transforms its neighbor points (with respect to the reference
frame) according to a similarity transformation (consisting
of translation, rotation, and scaling).
The scalar parameter α > 0 is a weighting factor determining the “stiffness” of the feature mesh balancing between feature point independence (α → 0; i.e. multi-target
tracker) and rigid single object tracking.
Tracker Evaluation
Test sequences. We use four challenging test sequences
(PAPER 1, PAPER 2, FACEOCC 1, and FACEOCC 2) to evaluate
our proposed method. The self-made PAPER 1 and PAPER 2
sequences were chosen to challenge the methods on a fast
moving deformable object under major scale changes. The
sequences have a spatial resolution of 960 px × 540 px and
consist of 563 and 726 frames respectively. The captured
object (paper) is textured with patches of similar appearance and shape. The similar appearing features were chosen to stress the relational structure of our tracker model.
Thus the only way to distinguish the features is by considering the relative position of the feature patches to each other.
The PAPER 1 sequence consists of five feature patches with a
carefully chosen position pattern which allows unique identification of the features by only having knowledge about
the relative distances of the features to each other. The
PAPER 2 sequence is more challenging since the number of
features is increased to 70 and the features are arranged in
a grid structure allowing local relational ambiguities. The
FACEOCC 1 and FACEOCC 2 sequences from [2, 5] are designed for evaluating object trackers under major occlusions. The sequences have a spatial resolution of 352 px ×
288 px (FACEOCC 1) and 320 px × 240 px (FACEOCC 2) and
both consist of 888 frames each. While the FACEOCC 1
sequence has only slow object movements, but showing
substantial occlusions, the FACEOCC 2 sequence challenges
with fast movements, illumination changes, object rotation
and substantial occlusions. The sequences and tracking results are shown in Fig. 7.
Parameter selection. Parameter selection can be split
into two parts. The first part consists in MRF model parameter selection. Since the proposed model is relatively
robust to changes in α, we set α in an ad-hoc fashion for
each sequence as follows: α = 20 for PAPER 1 and PA PER 2 and α = 50 for FACEOCC 1 and FACEOCC 2. For the
HOG features we set the smallest scale pyramid resolution
to 50 px × 50 px. This leads to 3 scales for FACEOCC 1 and
FACEOCC 2 and 4 scales for PAPER 1 and PAPER 2.
The second part is parameter selection for the PBP framework. We use N = 20 PBP iterations and p = 10 particles
for each node. With this setting both algorithms (MH - PBP
and S - PBP) converge well. Since we compare the overall
sampling behaviour of the proposed method rather than the
belief propagation convergence behaviour selecting these
parameters should be uncritical.
Evaluation metrics. We consider the distance εtrack between the estimated feature position and the groundtruth
(manually labeled) position as a quality measure. From this
measure we derive two metrics: The rooted mean of squared
distances (RMSD) and a quantile box-plot (10%, 25%, 50%,
75%, and 90% quantiles). While the first metric is very
sensitive to outliers, the second metric provides more infor-
5.2.2 Tracker Pipeline
A practical application requires some common modifications of the basic tracker pipeline in Sect. 5.2.1. The modifications include an additional particle resampling step,
where for each frame the initial set of particles are sampled
(i)
with replacement from the set of particles {xs }i=1,...,p
(i)
from the previous frame with probability bN
s (xs ). For the
tracker to be able to deal with fast moving objects, a resolution pyramid approach is applied. The resolution pyramid is
only applied to the unary potential function, i.e. the feature
descriptor is a concatenation of HOG descriptors of patches
with the same center position but differing spatial resolution. For each resolution pyramid level (scale) the image is
downsampled by a factor of 0.5 using bicubic interpolation.
Slice sampling. For the slice sampling approach we
need to define the boundary functions Aψs (u) and Axψts,t (u).
Since ψs,t is given as an analytic function we can use
our automatic inequality solver as described in Sect. 4.
An analytic description of the unary potential is not available thus we have to define the boundary manually. We
choose to set Aψs (u) to the whole image space for ps ,
i.e. ps ∈ [1, W ] × [1, H], where W and H are the image width and height respectively, and to restrict os to
os ∈ [−10, 10] × [−10, 10]. This way it is ensured that the
sampling space is large enough. On the other hand, particles
sampled outside the true (sub-)bounds are automatically rejected by the algorithm.
Metropolis-Hastings sampling. In order to provide a
fair comparison of our slice sampling approach to the stateof-the-art MH - PBP approach, the design of the proposal distribution has to be done very carefully. We propose to use a
4D Gaussian distribution with a covariance matrix Σ combined with a suitable coordinate transformation to ensure a
well-mixing random walk behaviour. The label space can
be divided into two parts, the feature position ps ∈ R2 and
orthogonal feature transformation os ∈ R2 . The proposal
m
m−1
m−1
) = N (ps
, I2×2 ·
distribution for ps is p(ps | ps
σxy ), where N (μ, Σ) is a Gaussian pdf with mean μ and covariance Σ. I2×2 is the 2 × 2 identity matrix. The vector os
is sampled analogously, but in the polar coordinate system
with covariance matrix Σpolar = [σr2 , 0; 0, σφ2 ], where σr2 is
the variance for the radius and σφ2 the variance for the angle.
Finally we have to carefully tune the three parameters σxy ,
σr , and σφ .
1134
Figure 7. Datasets and tracking results for our proposed method: PAPER 1, PAPER 2, FACEOCC 1, FACEOCC 2 (from left to right). First two
rows: successful tracking; third row: tracking failure cases.
Note that the estimation error varies highly, where very high
values (usually > 15 px) indicate a tracking failure. In order to visualize both the performance differences for nearoptimal parameters and tracking failures, the error values
below and above the 15 px mark are shown with a differing
vertical axis scaling. In Fig. 9, only a comparison for PA PER 1 and FACEOCC 1 is shown. The other two sequences
perform similarly. It can be observed that the tracking performance of MH - PBP strongly depends on careful parameter selection. The parameter σxy has the highest impact on
the tracking performance and the optimal parameter value
varies strongly between sequences (σxy = 5 for PAPER 1
and σxy = 0.5 for FACEOCC 1). Selecting σxy is a compromise between allowing fast object motions and reducing
overall localization noise. Selecting σr and σφ has analogous effects on changes in object scaling and rotation. This
way one has to incorporate prior knowledge about the object motion in order to obtain good tracking results using
MH - PBP . Tracked sequences and further comparisons are
provided in the supplemental material.
The computational complexity for MH - PBP is O(N SpM
(1 + V p)) and for S - PBP is O(N SpM (3 + 2V p)) given the
number of PBP iterations N , nodes S, particles p, MCMC
iterations M and the average number of neighbors per node
V . This indicates a doubling of computation time of S PBP compared to MH - PBP which is due to the overhead introduced for computing the interval bounds A. A look at
the CPU times using fixed parameters for both algorithms
(M = 5, p = 10, N = 20) verifies this finding: FACEOCC:
0.69 s/frame (S - PBP) vs. 0.33 s/frame (MH - PBP) ; PAPER 2:
7.43 s/frame vs. 3.66 s/frame. Nevertheless we have shown
mation about the overall error distribution.
Discussion. The evaluation results comparing S - PBP
with MH - PBP using different MCMC iterations are shown in
Fig. 8. For MH - PBP, the MH sampling parameters {σxy ,
σr , σφ } are chosen (from the set {0.1, 0.2, 0.5, 1.0, 2.0,
5.0} × {0.01, 0.02, 0.05, 0.10, 0.20, 0.50} × {0.01, 0.02,
0.05, 0.10, 0.20, 0.50}) such that the RMSD is minimized.
Note that for S - PBP such parameter tuning is not necessary.
We have evaluated the tracking performance for different
MCMC iterations M = 2 to 5. The box plots in Fig. 8 show
that S - PBP outperforms or performs equally well as MH PBP for all tested sequences except for sequence PAPER 2
with only 2 (and 3) MCMC iterations where both methods
fail. This is mainly due to a much higher overall sampling
noise of the MH - PBP method compared to S - PBP. We observed that the sampling noise of S - PBP is much less than
with MH - PBP at feature positions with high confidence (i.e.
high belief). On the other hand the sampling noise of S - PBP
increases for uncertain feature positions. The RMSD in sequence PAPER 2 and FACEOCC 1 is higher for S - PBP than for
MH - PBP due to temporal tracking failures. These tracking
failures are caused by strong local deformations or by occlusions of many feature points. Typical tracking failures
are depicted in the bottom row of Fig. 7. It can be observed
in such cases that S - PBP leads to much higher tracking error
than MH - PBP due to broader particle sampling in uncertain
feature positions.
Figure 9 shows an evaluation of MH - PBP under differing
(non-optimal) sampling parameters. To this end, we vary
each of the three sampling parameters individually and let
the other two parameters stay fixed at their optimal values.
1135
compute the slice sampling bounds, provided the unary and
binary potentials are defined by analytic functions or can be
bounded by one. We showed on a toy example that S - PBP
outperforms MH - PBP in terms of MCMC chain mixing performance. Furthermore we showed that our approach performs equally well or better than MH - PBP on challenging
relational feature tracking sequences.
2
3
4
6
4
2
εtrack
100
εtrack
rmsd
paper1
40
20
0
50
mh-pbp s-pbp
M =2
5
M
mh-pbp s-pbp
M =5
2
3
4
50
0
5
M
4
3
2
1
0
εtrack
100
εtrack
rmsd
paper2
60
40
20
0
mh-pbp s-pbp
M =2
Acknowledgements The work is funded by the ERC-Starting
mh-pbp s-pbp
M =5
Grant (DYNAMIC MINVIP). The authors gratefully acknowledge
the support.
2
0
2
3
4
5
M
εtrack
4
εtrack
rmsd
faceocc1
6
4
2
0
4
2
0
mh-pbp s-pbp
M =2
References
[1] C. Andrieu, N. de Freitas, A. Doucet, and M. I. Jordan. An introduction to mcmc for machine learning. Machine Learning, 50(1-2):5–
43, 2003.
[2] B. Babenko, M.-H. Yang, and S. Belongie. Visual tracking with online multiple instance learning. In CVPR, pages 983–990, 2011.
[3] F. Besse, C. Rother, A. Fitzgibbon, and J. Kautz. Pmbp: Patchmatch
belief propagation for correspondence field estimation. In BMVC,
2012.
[4] Y. Boykov, O. Veksler, and R. Zabih. Fast approximate energy minimization via graph cuts. PAMI, 23:1222–1239, 2001.
[5] G. Duan, H. Ai, S. Cao, and S. Lao. Group tracking: Exploring
mutual relations for multiple object tracking. In ECCV (3), pages
129–143, 2012.
[6] A. Ihler and D. McAllester. Particle belief propagation. In AISTATS,
pages 256–263, 2009.
[7] V. Kolmogorov. Convergent tree-reweighted message passing for energy minimization. PAMI, 28:1568–1583, 2006.
[8] R. Kothapa, J. Pacheco, and E. B. Sudderth. Max-product particle
belief propagation. Technical report, Brown University, 2011.
[9] W.-C. Lin and Y. Liu. Tracking dynamic near-regular textures under
occlusion and rapid movements. In ECCV, pages 44–55, 2006.
[10] O. Ludwig, D. Delgado, V. Goncalves, and U. Nunes. Trainable
classifier-fusion schemes: An application to pedestrian detection. In
IEEE Intelligent Transportation Systems (ITSC), pages 1–6, 2009.
[11] O. Müller, M. Y. Yang, and B. Rosenhahn. http://www.tnt.
uni-hannover.de/papers/view_paper.php?id=996,
2013.
[12] R. M. Neal. Slice sampling. Ann. Statist., 31(3):705–767, 2003. With
discussions and a rejoinder by the author.
[13] J. Peng, T. Hazan, D. McAllester, and R. Urtasun. Convex maxproduct algorithms for continuous mrfs with applications to protein
folding. In ICML, 2011.
[14] M. Salzmann and R. Urtasun. Beyond feature points: Structured prediction for monocular non-rigid 3d reconstruction. In ECCV, pages
245–259. 2012.
[15] H. B. Shitrit, J. Berclaz, F. Fleuret, and P. Fua. Tracking multiple
people under global appearance constraints. ICCV, pages 137–144,
2011.
[16] E. B. Sudderth, A. T. Ihler, M. Isard, W. T. Freeman, and A. S.
Willsky. Nonparametric belief propagation. Communications of the
ACM, 53(10):95–103, 2010.
[17] M. J. Wainwright, T. S. Jaakkola, and A. S. Willsky. Map estimation
via agreement on trees: message-passing and linear programming.
IEEE Trans. Information Theory, 51:3697–3717, 2005.
[18] B. Walsh. Markov chain monte carlo and gibbs sampling. In Lecture
Notes for EEB 581 version 26, April 2004.
[19] J. Xue, N. Zheng, J. Geng, and X. Zhong. Tracking multiple visual
targets via particle-based belief propagation. IEEE Trans. Systems,
Man, and Cybernetics, Part B, 38(1):196 –209, 2008.
mh-pbp s-pbp
M =5
2
3
4
5
mh-pbp s-pbp
M =2
5
M
mh-pbp
10
εtrack
εtrack
rmsd
faceocc2
8
6
4
2
0
8
6
4
2
mh-pbp s-pbp
M =5
s-pbp
Figure 8. Relational feature tracker evaluation results showing the
overal RMSD (for MCMC iterations from 2 to 5) and box plots over
the error distance to groundtruth for selected MCMC iterations.
paper1
200
100
15
10
5
0
1 2 5 1020 50
σxy · 10−1
200
100
15
10
5
0
1 2 5 1020 50
σr · 10−2
200
100
15
10
5
0
1 2 5 1020 50
σφ · 10−2
faceocc1
50
30
15
10
5
0
1 2 5 1020 50
σxy · 10−1
50
30
15
10
5
0
1 2 5 1020 50
σr · 10−2
50
30
15
10
5
0
1 2 5 1020 50
σφ · 10−2
Figure 9. Optimal parameter evaluation for MH - PBP method (with
M = 5). The vertical axis shows the error distance to groundtruth
in px. Note that the vertical axis is stretched for error values lower
than 15 px in order to better visualize performance differences.
that S - PBP needs significant less MCMC iterations than MH PBP such that the computational overhead can be typically
well compensated.
6. Conclusion
We presented a novel particle belief propagation algorithm using slice sampling (S - PBP) instead of MetropolisHastings. We exploit the message passing equations to
1136
| 1cs.CV
|
arXiv:1504.00941v2 [cs.NE] 7 Apr 2015
A Simple Way to Initialize Recurrent Networks of
Rectified Linear Units
Quoc V. Le, Navdeep Jaitly, Geoffrey E. Hinton
Google
Abstract
Learning long term dependencies in recurrent networks is difficult due to vanishing and exploding gradients. To overcome this difficulty, researchers have developed sophisticated optimization techniques and network architectures. In this
paper, we propose a simpler solution that use recurrent neural networks composed
of rectified linear units. Key to our solution is the use of the identity matrix or its
scaled version to initialize the recurrent weight matrix. We find that our solution is
comparable to a standard implementation of LSTMs on our four benchmarks: two
toy problems involving long-range temporal structures, a large language modeling
problem and a benchmark speech recognition problem.
1 Introduction
Recurrent neural networks (RNNs) are very powerful dynamical systems and they are the natural
way of using neural networks to map an input sequence to an output sequence, as in speech recognition and machine translation, or to predict the next term in a sequence, as in language modeling.
However, training RNNs by using back-propagation through time [30] to compute error-derivatives
can be difficult. Early attempts suffered from vanishing and exploding gradients [15] and this meant
that they had great difficulty learning long-term dependencies. Many different methods have been
proposed for overcoming this difficulty.
A method that has produced some impressive results [23, 24] is to abandon stochastic gradient
descent in favor of a much more sophisticated Hessian-Free (HF) optimization method. HF operates
on large mini-batches and is able to detect promising directions in the weight-space that have very
small gradients but even smaller curvature. Subsequent work, however, suggested that similar results
could be achieved by using stochastic gradient descent with momentum provided the weights were
initialized carefully [34] and large gradients were clipped [28]. Further developments of the HF
approach look promising [35, 25] but are much harder to implement than popular simple methods
such as stochastic gradient descent with momentum [34] or adaptive learning rates for each weight
that depend on the history of its gradients [5, 14].
The most successful technique to date is the Long Short Term Memory (LSTM) Recurrent Neural
Network which uses stochastic gradient descent, but changes the hidden units in such a way that
the backpropagated gradients are much better behaved [16]. LSTM replaces logistic or tanh hidden
units with “memory cells” that can store an analog value. Each memory cell has its own input and
output gates that control when inputs are allowed to add to the stored analog value and when this
value is allowed to influence the output. These gates are logistic units with their own learned weights
on connections coming from the input and also the memory cells at the previous time-step. There is
also a forget gate with learned weights that controls the rate at which the analog value stored in the
memory cell decays. For periods when the input and output gates are off and the forget gate is not
causing decay, a memory cell simply holds its value over time so the gradient of the error w.r.t. its
stored value stays constant when backpropagated over those periods.
1
The first major success of LSTMs was for the task of unconstrained handwriting recognition [12].
Since then, they have achieved impressive results on many other tasks including speech recognition [13, 10], handwriting generation [8], sequence to sequence mapping [36], machine translation [22, 1], image captioning [38, 18], parsing [37] and predicting the outputs of simple computer
programs [39].
The impressive results achieved using LSTMs make it important to discover which aspects of the
rather complicated architecture are crucial for its success and which are mere passengers. It seems
unlikely that Hochreiter and Schmidhuber’s [16] initial design combined with the subsequent introduction of forget gates [6, 7] is the optimal design: at the time, the important issue was to find any
scheme that could learn long-range dependencies rather than to find the minimal or optimal scheme.
One aim of this paper is to cast light on what aspects of the design are responsible for the success of
LSTMs.
Recent research on deep feedforward networks has also produced some impressive results [19, 3]
and there is now a consensus that for deep networks, rectified linear units (ReLUs) are easier to train
than the logistic or tanh units that were used for many years [27, 40]. At first sight, ReLUs seem
inappropriate for RNNs because they can have very large outputs so they might be expected to be far
more likely to explode than units that have bounded values. A second aim of this paper is to explore
whether ReLUs can be made to work well in RNNs and whether the ease of optimizing them in
feedforward nets transfers to RNNs.
2 The initialization trick
In this paper, we demonstrate that, with the right initialization of the weights, RNNs composed
of rectified linear units are relatively easy to train and are good at modeling long-range dependencies. The RNNs are trained by using backpropagation through time to get error-derivatives for the
weights and by updating the weights after each small mini-batch of sequences. Their performance
on test data is comparable with LSTMs, both for toy problems involving very long-range temporal
structures and for real tasks like predicting the next word in a very large corpus of text.
We initialize the recurrent weight matrix to be the identity matrix and biases to be zero. This means
that each new hidden state vector is obtained by simply copying the previous hidden vector then
adding on the effect of the current inputs and replacing all negative states by zero. In the absence
of input, an RNN that is composed of ReLUs and initialized with the identity matrix (which we call
an IRNN) just stays in the same state indefinitely. The identity initialization has the very desirable
property that when the error derivatives for the hidden units are backpropagated through time they
remain constant provided no extra error-derivatives are added. This is the same behavior as LSTMs
when their forget gates are set so that there is no decay and it makes it easy to learn very long-range
temporal dependencies.
We also find that for tasks that exhibit less long range dependencies, scaling the identity matrix by
a small scalar is an effective mechanism to forget long range effects. This is the same behavior as
LTSMs when their forget gates are set so that the memory decays fast.
Our initialization scheme bears some resemblance to the idea of Mikolov et al. [26], where a part of
the weight matrix is fixed to identity or approximate identity. The main difference of their work to
ours is the fact that our network uses the rectified linear units and the identity matrix is only used for
initialization. The scaled identity initialization was also proposed in Socher et al. [32] in the context
of tree-structured networks but without the use of ReLUs. Our work is also related to the work of
Saxe et al. [31], who study the use of orthogonal matrices as initialization in deep networks.
3 Overview of the experiments
Consider a recurrent net with two input units. At each time step, the first input unit has a real value
and the second input unit has a value of 0 or 1 as shown in figure 1. The task is to report the sum of
the two real values that are marked by having a 1 as the second input [16, 15, 24]. IRNNs can learn
to handle sequences with a length of 300, which is a challenging regime for other algorithms.
2
Another challenging toy problem is to learn to classify the MNIST digits when the 784 pixels are
presented sequentially to the recurrent net. Again, the IRNN was better than the LSTM, having been
able to achieve 3% test set error compared to 34% for LSTM.
While it is possible that a better tuned LSTM (with a different architecture or the size of the hidden
state) would outperform the IRNN for the above two tasks, the fact that the IRNN performs as well
as it does, with so little tuning, is very encouraging, especially given how much simpler the model
is, compared to the LSTM.
We also compared IRNNs with LSTMs on a large language modeling task. Each memory cell of an
LSTM is considerably more complicated than a rectified linear unit and has many more parameters,
so it is not entirely obvious what to compare. We tried to balance for both the number of parameters
and the complexity of the architecture by comparing an LSTM with N memory cells with an IRNN
with four layers of N hidden units, and an IRNN with one layer and 2N hidden units. Here we find
that the IRNN gives results comparable to the equivalent LSTM.
Finally, we benchmarked IRNNs and LSTMs on a acoustic modeling task on TIMIT. As the tasks
only require a short term memory of the inputs, we used a the identity matrix scaled by 0.01 as
initialization for the recurrent matrix. Results show that our method is also comparable to LSTMs,
despite being a lot simpler to implement.
4 Experiments
In the following experiments, we compare IRNNs against LSTMs, RNNs that use tanh units and
RNNs that use ReLUs with random Gaussian initialization.
For IRNNs, in addition to the recurrent weights being initialized at identity, the non-recurrent
weights are initialized with a random matrix, whose entries are sampled from a Gaussian distribution with mean of zero and standard deviation of 0.001.
Our implementation of the LSTMs is rather standard and includes the forget gate. It is observed that
setting a higher initial forget gate bias for LSTMs can give better results for long term dependency
problems. We therefore also performed a grid search for the initial forget gate bias in LSTMs from
the set {1.0, 4.0, 10.0, 20.0}. Other than that we did not tune the LTSMs much and it is possible that
the results of LSTMs in the experiments can be improved.
In addition to LSTMs, two other candidates for comparison are RNNs that use the tanh activation
function and RNNs that use ReLUs with standard random Gaussian initialization. We experimented
with several values of standard deviation for the random initialization Gaussian matrix and found
that values suggested in [33] work well.
To train these models, we use stochastic gradient descent with a fixed learning rate and gradient
clipping. To ensure that good hyperparameters are used, we performed a grid search over several
learning rates α = {10−9 , 10−8 , ..., 10−1 } and gradient clipping values gc = {1, 10, 100, 1000} [9,
36]. The reported result is the best result over the grid search. We also use the same batch size of
16 examples for all methods. The experiments are carried out using the DistBelief infrastructure,
where each experiment only uses one replica [20, 4].
4.1 The Adding Problem
The adding problem is a toy task, designed to examine the power of recurrent models in learning
long-term dependencies [16, 15]. This is a sequence regression problem where the target is a sum of
two numbers selected in a sequence of random signals, which are sampled from a uniform distribution in [0,1]. At every time step, the input consists of a random signal and a mask signal. The mask
signal has a value of zero at all time steps except for two steps when it has values of 1 to indicate
which two numbers should be added. An example of the adding problem is shown in figure 1 below.
A basic baseline is to always predict the sum to have a value of 1 regardless of the inputs. This will
give the Mean Squared Error (MSE) around 0.1767. The goal is to train a model that achieves MSE
well below 0.1767.
3
Figure 1: An example of the “adding” problem, where the target is 1.2 which is the sum of 2nd and
the 7th numbers in the first sequence [24].
The problem gets harder as the length of the sequence T increases because the dependency between
the output and the relevant inputs becomes more remote. To solve this problem, the recurrent net
must remember the first number or the sum of the two numbers accurately whilst ignoring all of the
irrelevant numbers.
We generated a training set of 100,000 examples and a test set of 10,000 examples as we varied T .
We fixed the hidden states to have 100 units for all of our networks (LSTMs, RNNs and IRNNs).
This means the LSTMs had more parameters by a factor of about 4 and also took about 4 times as
much computation per timestep.
As we varied T , we noticed that both LSTMs and RNNs started to struggle when T is around 150.
We therefore focused on investigating the behaviors of all models from this point onwards. The
results of the experiments with T = 150, T = 200, T = 300, T = 400 are reported in figure 2
below (best hyperparameters found during grid search are listed in table 1).
Adding two numbers in a sequence of 150 numbers
Adding two numbers in a sequence of 200 numbers
0.8
0.8
LSTM
RNN + Tanh
RNN + ReLUs
IRNN
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0
1
2
3
4
5
6
7
8
Steps
LSTM
RNN + Tanh
RNN + ReLUs
IRNN
0.7
Test MSE
Test MSE
0.7
0
9
0
Adding two numbers in a sequence of 300 numbers
3
4
5
6
7
8
9
6
x 10
Adding two numbers in a sequence of 400 numbers
0.8
LSTM
RNN + Tanh
RNN + ReLUs
IRNN
0.7
0.6
0.6
0.5
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
1
2
3
4
5
Steps
6
7
LSTM
RNN + Tanh
RNN + ReLUs
IRNN
0.7
Test MSE
Test MSE
2
Steps
0.8
0
1
6
x 10
8
0
9
0
1
2
3
4
5
Steps
6
x 10
6
7
8
9
6
x 10
Figure 2: The results of recurrent methods on the “adding” problem for the case of T = 150 (top
left), T = 200 (top right), T = 300 (bottom left) and T = 400 (bottom right). The objective
function is the Root Mean Squared Error, reported on the test set of 10,000 examples. Note that
always predicting the sum to be 1 should give MSE of 0.1767.
The results show that the convergence of IRNNs is as good as LSTMs. This is given that each LSTM
step is more expensive than an IRNN step (at least 4x more expensive). Adding two numbers in a
sequence of 400 numbers is somewhat challenging for both algorithms.
4
T
150
200
300
400
LSTM
lr = 0.01, gc = 10, f b = 1.0
lr = 0.001, gc = 100, f b = 4.0
lr = 0.01, gc = 1, f b = 4.0
lr = 0.01, gc = 100, f b = 10.0
RNN + Tanh
lr = 0.01, gc = 100
N/A
N/A
N/A
IRNN
lr = 0.01, gc = 100
lr = 0.01, gc = 1
lr = 0.01, gc = 10
lr = 0.01, gc = 1
Table 1: Best hyperparameters found for adding problems after grid search. lr is the learning rate, gc
is gradient clipping, and f b is forget gate bias. N/A is when there is no hyperparameter combination
that gives good result.
4.2 MNIST Classification from a Sequence of Pixels
Another challenging toy problem is to learn to classify the MNIST digits [21] when the 784 pixels
are presented sequentially to the recurrent net. In our experiments, the networks read one pixel at a
time in scanline order (i.e. starting at the top left corner of the image, and ending at the bottom right
corner). The networks are asked to predict the category of the MNIST image only after seeing all
784 pixels. This is therefore a huge long range dependency problem because each recurrent network
has 784 time steps.
To make the task even harder, we also used a fixed random permutation of the pixels of the MNIST
digits and repeated the experiments.
All networks have 100 recurrent hidden units. We stop the optimization after it converges or when
it reaches 1,000,000 iterations and report the results in figure 3 (best hyperparameters are listed in
table 2).
Pixel−by−pixel MNIST
Pixel−by−pixel permuted MNIST
100
100
LSTM
RNN + Tanh
RNN + ReLUs
IRNN
90
80
90
80
70
Test Accuracy
Test Accuracy
70
60
50
40
60
50
40
30
30
20
20
10
0
LSTM
RNN + Tanh
RNN + ReLUs
IRNN
10
0
1
2
3
4
5
Steps
6
7
8
9
0
10
5
x 10
0
1
2
3
4
5
Steps
6
7
8
9
10
5
x 10
Figure 3: The results of recurrent methods on the “pixel-by-pixel MNIST” problem. We report the
test set accuracy for all methods. Left: normal MNIST. Right: permuted MNIST.
Problem
MNIST
permuted
MNIST
LSTM
lr = 0.01, gc = 1
f b = 1.0
lr = 0.01, gc = 1
f b = 1.0
RNN + Tanh
lr = 10−8 , gc = 10
RNN + ReLUs
lr = 10−8 , gc = 10
IRNN
lr = 10−8 , gc = 1
lr = 10−8 , gc = 1
lr = 10−6 , gc = 10
lr = 10−9 , gc = 1
Table 2: Best hyperparameters found for pixel-by-pixel MNIST problems after grid search. lr is the
learning rate, gc is gradient clipping, and f b is the forget gate bias.
The results using the standard scanline ordering of the pixels show that this problem is so difficult
that standard RNNs fail to work, even with ReLUs, whereas the IRNN achieves 3% test error rate
which is better than most off-the-shelf linear classifiers [21]. We were surprised that the LSTM did
not work as well as IRNN given the various initialization schemes that we tried. While it still possible that a better tuned LSTM would do better, the fact that the IRNN perform well is encouraging.
5
Applying a fixed random permutation to the pixels makes the problem even harder but IRNNs on
the permuted pixels are still better than LSTMs on the non-permuted pixels.
The low error rates of the IRNN suggest that the model can discover long range correlations in the
data while making weak assumptions about the inputs. This could be important to have for problems
when input data are in the form of variable-sized vectors (e.g. the repeated field of a protobuffer 1 ).
4.3 Language Modeling
We benchmarked RNNs, IRNNs and LSTMs on the one billion word language modelling dataset [2],
perhaps the largest public benchmark in language modeling. We chose an output vocabulary of
1,000,000 words.
As the dataset is large, we observed that the performance of recurrent methods depends on the size
of the hidden states: they perform better as the size of the hidden states gets larger (cf. [2]). We
however focused on a set of simple controlled experiments to understand how different recurrent
methods behave when they have a similar number of parameters. We first ran an experiment where
the number of hidden units (or memory cells) in LSTM are chosen to be 512. The LSTM is trained
for 60 hours using 32 replicas. Our goal is then to check how well IRNNs perform given the same
experimental environment and settings. As LSTM have more parameters per time step, we compared
them with an IRNN that had 4 layers and same number of hidden units per layer (which gives
approximately the same numbers of parameters).
We also experimented shallow RNNs and IRNNs with 1024 units. Since the output vocabulary is
large, we projected the 1024 hidden units to a linear layer with 512 units before the softmax. This
avoids greatly increasing the number of parameters.
The results are reported in table 3, which show that the performance of IRNNs is closer to the
performance of LSTMs for this large-scale task than it is to the performance of RNNs.
Methods
LSTM (512 units)
IRNN (4 layers, 512 units)
IRNN (1 layer, 1024 units + linear projection with 512 units before softmax)
RNN (4 layer, 512 tanh units)
RNN (1 layer, 1024 tanh units + linear projection with 512 units before softmax)
Test perplexity
68.8
69.4
70.2
71.8
72.5
Table 3: Performances of recurrent methods on the 1 billion word benchmark.
4.4 Speech Recognition
We performed Phoneme recognition experiments on TIMIT with IRNNs and Bidirectional IRNNs
and compared them to RNNs, LSTMs and Bidirectional LSTMs and RNNs. Bidirectional LSTMs
have been applied previously to TIMIT in [11]. In these experiments we generated phoneme alignments from Kaldi [29] using the recipe reported in [17] and trained all RNNs with two and five
hidden layers. Each model was given log Mel filter bank spectra with their delta and accelerations,
where each frame was 120 (=40*3) dimensional and trained to predict the phone state (1 of 180).
Frame error rates (FER) from this task are reported in table 4.
In this task, instead of the identity initialization for the IRNNs matrices we used 0.01I so we refer
to them as iRNNs. Initalizing with the full identity led to slow convergence, worse results and
sometimes led to the model diverging during training. We hypothesize that this was because in the
speech task similar inputs are provided to the neural net in neighboring frames. The normal IRNN
keeps integrating this past input, instead of paying attention mainly to the current input because it
has a difficult time forgetting the past. So for the speech task, we are not only showing that iRNNs
work much better than RNNs composed of tanh units, but we are also showing that initialization
with the full identity is suboptimal when long range effects are not needed. Mulitplying the identity
with a small scalar seems to be a good remedy in such cases.
1
https://code.google.com/p/protobuf/
6
Methods
RNN (500 neurons, 2 layers)
LSTM (250 cells, 2 layers)
iRNN (500 neurons, 2 layers)
RNN (500 neurons, 5 layers)
LSTM (250 cells, 5 layers)
iRNN (500 neurons, 5 layers)
Bidirectional RNN (500 neurons, 2 layers)
Bidirectional LSTM (250 cells, 2 layers)
Bidirectional iRNN (500 neurons, 2 layers)
Bidirectional RNN (500 neurons, 5 layers)
Bidirectional LSTM (250 cells, 5 layers)
Bidirectional iRNN (500 neurons, 5 layers)
Frame error rates (dev / test)
35.0 / 36.2
34.5 / 35.4
34.3 / 35.5
35.6 / 37.0
35.0 / 36.2
33.0 / 33.8
31.5 / 32.4
29.6 / 30.6
31.9 / 33.2
33.9 / 34.8
28.5 / 29.1
28.9 / 29.7
Table 4: Frame error rates of recurrent methods on the TIMIT phone recognition task.
In general in the speech recognition task, the iRNN easily outperforms the RNN that uses tanh units
and is comparable to LSTM although we don’t rule out the possibility that with very careful tuning
of hyperparameters, the relative performance of LSTMs or the iRNNs might change. A five layer
Bidirectional LSTM outperforms all the other models on this task, followed closely by a five layer
Bidirectional iRNN.
4.5 Acknowledgements
We thank Jeff Dean, Matthieu Devin, Rajat Monga, David Sussillo, Ilya Sutskever and Oriol Vinyals
for their help with the project.
References
[1] D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align
and translate. arXiv preprint arXiv:1409.0473, 2014.
[2] C. Chelba, T. Mikolov, M. Schuster, Q. Ge, T. Brants, and P. Koehn. One billion word benchmark for measuring progress in statistical language modeling. CoRR, abs/1312.3005, 2013.
[3] G. E. Dahl, D. Yu, L. Deng, and A. Acero. Context-dependent pre-trained deep neural networks
for large vocabulary speech recognition. IEEE Transactions on Audio, Speech, and Language
Processing - Special Issue on Deep Learning for Speech and Language Processing, 2012.
[4] J. Dean, G. S. Corrado, R. Monga, K. Chen, M. Devin, Q. V. Le, M. Z. Mao, M. A. Ranzato,
A. Senior, P. Tucker, K. Yang, and A. Y. Ng. Large scale distributed deep networks. In NIPS,
2012.
[5] J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and
stochastic optimization. The Journal of Machine Learning Research, 12:2121–2159, 2011.
[6] F. A. Gers, J. Schmidhuber, and F. Cummins. Learning to forget: Continual prediction with
LSTM. Neural Computation, 2000.
[7] F. A. Gers, N. N. Schraudolph, and J. Schmidhuber. Learning precise timing with lstm recurrent networks. The Journal of Machine Learning Research, 2003.
[8] A. Graves.
Generating sequences with recurrent neural networks.
arXiv:1308.0850, 2013.
arXiv preprint
[9] A. Graves. Generating sequences with recurrent neural networks. In Arxiv, 2013.
[10] A. Graves and N. Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In Proceedings of the 31st International Conference on Machine Learning, 2014.
[11] A. Graves, N. Jaitly, and A-R. Mohamed. Hybrid speech recognition with deep bidirectional
lstm. In IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU),, 2013.
7
[12] A. Graves, M. Liwicki, S. Fernández, R. Bertolami, H. Bunke, and J. Schmidhuber. A novel
connectionist system for unconstrained handwriting recognition. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2009.
[13] A. Graves, A-R. Mohamed, and G. Hinton. Speech recognition with deep recurrent neural
networks. In IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), 2013.
[14] G. Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 2012.
[15] S. Hochreiter, Y. Bengio, P. Frasconi, and J. Schmidhuber. Gradient flow in recurrent nets: the
difficulty of learning long-term dependencies. A Field Guide to Dynamical Recurrent Neural
Networks, 2001.
[16] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 1997.
[17] N. Jaitly. Exploring Deep Learning Methods for discovering features in speech signals. PhD
thesis, University of Toronto, 2014.
[18] R. Kiros, R. Salakhutdinov, and R. S. Zemel. Unifying visual-semantic embeddings with
multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014.
[19] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional
neural networks. In Advances in Neural Information Processing Systems, 2012.
[20] Q. V. Le, M. A. Ranzato, 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 International
Conference on Machine Learning, 2012.
[21] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document
recognition. Proceedings of the IEEE, 1998.
[22] T. Luong, I. Sutskever, Q. V. Le, O. Vinyals, and W. Zaremba. Addressing the rare word
problem in neural machine translation. arXiv preprint arXiv:1410.8206, 2014.
[23] J. Martens. Deep learning via Hessian-free optimization. In Proceedings of the 27th International Conference on Machine Learning, 2010.
[24] J. Martens and I. Sutskever. Learning recurrent neural networks with Hessian-Free optimization. In ICML, 2011.
[25] J. Martens and I. Sutskever. Training deep and recurrent neural networks with Hessian-Free
optimization. Neural Networks: Tricks of the Trade, 2012.
[26] T. Mikolov, A. Joulin, S. Chopra, M. Mathieu, and M. A. Ranzato. Learning longer memory
in recurrent neural networks. arXiv preprint arXiv:1412.7753, 2014.
[27] V. Nair and G. Hinton. Rectified Linear Units improve Restricted Boltzmann Machines. In
International Conference on Machine Learning, 2010.
[28] R. Pascanu, T. Mikolov, and Y. Bengio. On the difficulty of training recurrent neural networks.
arXiv preprint arXiv:1211.5063, 2012.
[29] D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann,
P. Motlicek, Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely. The kaldi speech
recognition toolkit. In IEEE 2011 Workshop on Automatic Speech Recognition and Understanding. IEEE Signal Processing Society, 2011.
[30] D. Rumelhart, G. E. Hinton, and R. J. Williams. Learning representations by back-propagating
errors. Nature, 323(6088):533–536, 1986.
[31] A. M. Saxe, J. L. McClelland, and S. Ganguli. Exact solutions to the nonlinear dynamics of
learning in deep linear neural networks. arXiv preprint arXiv:1312.6120, 2013.
[32] R. Socher, J. Bauer, C. D. Manning, and A. Y. Ng. Parsing with compositional vector grammars. In ACL, 2013.
[33] D. Sussillo and L. F. Abbott. Random walk intialization for training very deep networks. arXiv
preprint arXiv:1412.6558, 2015.
[34] I. Sutskever, J. Martens, G. Dahl, and G. Hinton. On the importance of initialization and
momentum in deep learning. In Proceedings of the 30th International Conference on Machine
Learning, 2013.
8
[35] I. Sutskever, J. Martens, and G. E. Hinton. Generating text with recurrent neural networks.
In Proceedings of the 28th International Conference on Machine Learning, pages 1017–1024,
2011.
[36] I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence learning with neural networks.
In NIPS, 2014.
[37] O. Vinyals, L. Kaiser, T. Koo, S. Petrov, I. Sutskever, and G. Hinton. Grammar as a foreign
language. arXiv preprint arXiv:1412.7449, 2014.
[38] O. Vinyals, A. Toshev, S. Bengio, and D. Erhan. Show and tell: A neural image caption
generator. arXiv preprint arXiv:1411.4555, 2014.
[39] W. Zaremba and I. Sutskever. Learning to execute. arXiv preprint arXiv:1410.4615, 2014.
[40] M. Zeiler, M. Ranzato, R. Monga, M. Mao, K. Yang, Q. V. Le, P. Nguyen, A. Senior, V. Vanhoucke, and J. Dean. On rectified linear units for speech processing. In IEEE Conference on
Acoustics, Speech and Signal Processing (ICASSP), 2013.
9
| 9cs.NE
|
A Class of Control Certificates to Ensure Reach-While-Stay
for Switched Systems
Hadi Ravanbakhsh and Sriram Sankaranarayanan
Department of Computer Science
University of Colorado, Boulder
Boulder, CO, USA
[email protected]
In this article, we consider the problem of synthesizing switching controllers for temporal properties
through the composition of simple primitive reach-while-stay (RWS) properties. Reach-while-stay
properties specify that the system states starting from an initial set I, must reach a goal (target) set G
in finite time, while remaining inside a safe set S. Our approach synthesizes switched controllers that
select between finitely many modes to satisfy the given RWS specification. To do so, we consider
control certificates, which are Lyapunov-like functions that represent control strategies to achieve the
desired specification. However, for RWS problems, a control Lyapunov-like function is often hard to
synthesize in a simple polynomial form. Therefore, we combine control barrier and Lyapunov functions with an additional compatibility condition between them. Using this approach, the controller
synthesis problem reduces to one of solving quantified nonlinear constrained problems that are handled using a combination of SMT solvers. The synthesis of controllers is demonstrated through a set
of interesting numerical examples drawn from the related work, and compared with the state-of-theart tool SCOTS. Our evaluation suggests that our approach is computationally feasible, and adds to
the growing body of formal approaches to controller synthesis.
1
Introduction
The problem of synthesizing switching controllers for reach-while-stay (RWS) specifications is examined
in this article. RWS properties are an important class, since we may decompose more complex temporal
specifications into a sequence of RWS specifications [11]. The plant model is a switched system that
consists of finitely many (controllable) modes, and the dynamics for each mode are specified using
ODEs. Furthermore, we consider nonlinear ODEs for each mode, including rational, trigonometric, and
exponential functions. The goal of the controller is to switch between the appropriate modes, so that the
resulting closed loop traces satisfy the specification.
RWS properties specify that a goal set G must be reached by all behaviors of the closed-loop system
while staying inside a safe set S. Specifically, the state of the system is assumed to be initialized to
any state in the set S. RWS properties include safety properties (stay inside a safe set S), reachability
properties (reach a goal set G), and “control-to-facet” problems [7, 8].
The controller synthesis is addressed in two phases: (a) formulating a control certificate whose existence guarantees the existence of a non-Zeno switching control law for the given RWS specification,
and (b) solving for a certificate of a particular form as a feasibility problem. The control certificates are
control Lyapunov-like functions which represent a strategy for the controller to satisfy the specifications.
Additionally, this strategy can be effectively implemented as a feedback law using a controller that respects min dwell time constraints. In the second phase, a counterexample guided inductive synthesis
(CEGIS) framework [26], — an approach that uses SMT solvers at its core — is used to discover such
control certificates. However, this procedure is used off the shelf, building upon the previous work of
D. Fisman, S. Jacobs (Eds.): Sixth Workshop on Synthesis (SYNT 2017)
EPTCS 260, 2017, pp. 44–61, doi:10.4204/EPTCS.260.6
c H. Ravanbakhsh & S. Sankaranarayanan
This work is licensed under the
Creative Commons Attribution License.
H. Ravanbakhsh & S. Sankaranarayanan
45
Ravanbakhsh et al. [22]. This procedure uses a specialized solver for finding a certificate of a given parametric form that handles quantified formulas by alternating between a series of quantifier free formulas
using existing SMT solvers [18, 3].
The contributions of the paper are as follows: first, we show that a straightforward formulation of
the control certificate for the RWS problem yields an exponential number of conditions, and hence can
be computationally infeasible. Next, we introduce a class of control certificates which (i) has a concise
logical structure that makes the problem of discovering the certificates computationally feasible; and (ii)
we show that such certificates yield corresponding switching strategies with a min-dwell time property
unlike the conventional control certificates. Next, we extend our approach to the initialized RWS (IRWS)
property that additionally restricts the set of initial conditions of the system using a class of “control
zero-ing” barrier functions [33, 36] . Also, a suitable formulation for these functions is provided within
our framework. Finally, we provide numerical examples to demonstrate the effectiveness of the method,
including comparisons with recently developed state-of-art automatic control synthesis tool SCOTS [25].
1.1
Related Work
The broader area of temporal logic synthesis seeks to synthesize formally guaranteed controllers from
the given plant model and specifications. The dominant approach is to build a discrete abstraction of
the given plant that is related to the original system [35, 15, 10, 25, 17]. Once a suitable abstraction is
found, these approaches use a systematic temporal logic-based controller design approach over the abstraction [32]. The properties of interest in these systems include the full linear temporal logic (LTL) and
an efficiently synthesizable subset such as GR(1) [35, 15]. These approaches differ in how the abstraction
can be constructed in a guaranteed manner. One class of approaches works by fixing a time step, gridding
the state-space, and simulating one point per cell [17, 10, 25, 37, 31]. The resulting abstraction, however,
is not always approximately bisimilar to the original system. Nevertheless, conditions such as open loop
incremental stability of the plant can be used to obtain bisimilarity [5]. Alternatively, the abstraction
can be built without time discretization [20, 15] by considering infeasible transitions. And furthermore,
the abstraction can be iteratively refined through a counter-example refinement scheme [19]. Our work
here does not directly focus on building abstractions. Rather, our focus is on deductive approaches for a
narrow class of temporal logic properties namely RWS properties. Using our approach, control systems
for richer properties can be built from solving a series of RWS problems.
Our approach is closely related to work of Habets et al. [6] and Kloetzer et al. [11]. In these methods,
an abstraction is obtained by solving local control-to-facet problems instead of reachability analysis.
However, continuous feedback is synthesized for each control-to-facet problem. The key difference in
this paper is that the control-to-facet problems themselves are solved using switching. Furthermore, we
consider initialized problems, where the initial states are also restricted to belong to a set. We find that
IRWS problems can often be realized through a controller even when the corresponding RWS problem
(for which the initial condition is not restricted) cannot be synthesized.
Another related class of solutions is based on synthesizing “a deductive proof of correctness” simultaneously with “a control strategy”. The goal of these approaches also consists of finding a control
certificate, which yields a (control) strategy to guarantee the property. This typically takes the form of
a control Lyapunov-like function. The idea of control Lyapunov functions goes back to Artstein [1] and
Sontag [27]. The problem of discovering a control Lyapunov function is usually formulated using bilinear matrix inequalities (BMI) [30]. Also, instead of solving such NP-hard problems, usually alternating
optimization (V-K iteration or policy iteration) is used to conservatively find a solution [30, 4].
Wongpiromsarn et al. [34] discuss verification of temporal logic properties using barrier certificates.
46
A Class of Control Certificates to Ensure Reach-While-Stay for Switched Systems
For synthesis, Xu et al. [36] discuss conditions for the so-called “control zeroing” barrier functions for
safety and their properties. They also, consider their combination with control Lyapunov functions. In
this article, we provide an alternative condition that is based on “exponential condition” barrier functions [12] and enforcing a compatibility condition between the control actions suggested by the control
barrier and control Lyapunov functions. Also, Dimitrova et. al. [2] have shown that control certificates
can be extended to address more complicated specifications i.e. parity games. While these results show
that constraint solving based methods can be applied on more complicated specification, no method of
finding such certificates is provided.
The use of SMT solvers in control synthesis has also been well-studied. Taly et. al [28, 29] use
a constraint solving approach to find control certificates for reachability and safety. They adapt a technique known as Counter-Example Guided Inductive Synthesis (CEGIS), originally proposed for program
synthesis [26], to solve the control problems using a combination of an SMT solver with numerical simulations. Ravanbakhsh et al. [22] propose a combination of SMT and SDP solvers for finding control
certificates. However, their method is only applicable to stability or simple reachability properties, involving the use of a single Lyapunov function. In a subsequent paper, their approach is extended to
handle disturbance inputs [23]. The use of SMT solvers to solve for Lyapunov-like functions is used in
our paper as well. However, this paper focuses on defining a more tractable class of control certificates
for RWS problems. Furthermore, we show how these problems can be composed for more complex
temporal objectives. In particular, our use of the CEGIS procedure is not a contribution of this paper.
Furthermore, in order to handle nonlinear systems and also to guarantee numerical soundness of these
solvers, we use the nonlinear SMT solver dReal [3].
Huang et. al. [9] also propose control certificates to solve the RWS problem for piecewise affine
systems, using SMT solvers. Their approach uses piecewise constant functions as control certificates
and partitions the state space into small enough cells in order to define such functions. By using this
technique, any function can be approximated, which makes the method relatively complete.
As mentioned earlier, past work by Habets et al. and Klutzier et al. [6, 11] build a finite abstraction
by repeatedly solving control-to-facet problems. These problems seek to find a feedback law inside a
polytope P that guarantees all the resulting trajectories exit P through a specific facet F of P. Habets et.
al. [7] show necessary and sufficient conditions for the existence of a control strategy for the control-tofacet problem on simplices. This condition is sufficient but not necessary for polytopes. They extract a
unique certificate from each problem instance and check whether the condition holds for the certificate.
Subsequently, Roszak et al. [24] and Helwa et al. [8] extend this approach and solve reachability to a set
of facets by introducing flow condition, which combined with invariant condition serves as a control certificate similar to those used in this paper. From the published results, these methods are more efficient,
but are only applicable to affine systems over polytopes. In contrast, the dynamics in this article can be
non-linear involving rational, trigonometric, and exponential functions. In this article, we demonstrate
that our method can be used to solve such problems and it can be integrated into other methods which
build an abstraction for the system.
2
2.1
Background
Notation
.
Given a function f (t), let f + (t) ( f − (t)) be the right (left) limit of f at t, and f (t) represent the right
derivative of f at time t. For a set S ⊆ Rn , ∂ S and int(S) are its boundary and interior, respectively.
H. Ravanbakhsh & S. Sankaranarayanan
47
Definition 1 (Nondegenerate Basic Semialgebraic Set):
A nondegenerate basic semialgebraic set K is a nonempty set defined by a conjunction polynomial
inequalities:
K : {x | pK,1 (x) ≤ 0 ∧ · · · ∧ pK,i (x) ≤ 0} ,
where x ∈ Rn . For each j ∈ [1, i], we define
HK, j = {x | x ∈ K ∧ pK, j (x) = 0} .
It is required that (a) each HK, j is nonempty, (b) the boundary ∂ K and the interior int(K) are given
W
V
by ij=1 HK, j and ij=1 pK, j (x) < 0, respectively, and (c) the interior is nonempty. We use “basic semialgebraic” and “nondegenerate basic semialgebraic” interchangeably.
2.2
Switched Systems
We consider continuous-time switched system plants, controlled by a memoryless controller that provides continuous-time switching feedback. The state of the plant P is defined by n continuous variables
x in a state space X ⊆ Rn , along with a finite set of modes Q = {q1 , . . . , qm }. The trace of the system
(q(t), x(t)) maps time to mode q(.) : R+ → Q, and state x(.) : R+ → X. The mode q ∈ Q is controlled
by an external switching input q(t). The state of the plant inside each mode evolves according to (time
invariant) dynamics:
.
x(t) = fq(t) (x(t)) ,
(1)
wherein fq : X → Rn is a Lipschitz continuous function over X, describing the vector field of the plant
for mode q.
The controller C is defined as a function K : Q × X → Q, which given the current mode and state of
the plant, decides the mode of the plant at the next time instant. Formally:
q+ (t) = K (q(t), x(t)).
(2)
The closed loop hP, C i produces traces (q(t), x(t)) defined jointly by equations (1) and (2). However, care must be taken to avoid Zenoness, wherein the controller can switch infinitely often in a finite
time interval. Such controllers are physically unrealizable. Therefore, we will additionally ensure that
the K function satisfies a minimum dwell time requirement that guarantees a minimum time δ > 0
between mode switches.
Definition 2 (Minimum Dwell Time): A controller C has a minimum dwell time δ > 0 with respect to a
plant P iff for all traces and for all switch times T (q(T ) 6= q+ (T ), the controller does not switch during
the times t ∈ [T, T + δ ): i.e, K (q(t), x(t)) = q+ (T ) for all t ∈ [T, T + δ ).
Once the function K is defined with a minimum dwell time guarantee, given initial mode (q(0)),
and initial state (x(0)), a unique trace is defined for the system.
Specifications: Generally, specifications describe desired sequences of plant states x(t) over time t ≥ 0
that we wish to control for. In this paper, we focus on reach-while-stay (RWS) specifications involving
three sets: initial set I ⊆ X, safe set S ⊆ X and goal set G ⊆ X.
Definition 3 (Initialized Reach-While-Stay (RWS) Specification): A trace x(t) for t ∈ [0, ∞) satisfies a
reach-while-stay (RWS) specification w.r.t sets hI, S, Gi iff whenever x(0) ∈ I, there exists a time T ≥ 0
s.t. for all t ∈ [0, T ), x(t) ∈ S, and x(T ) ∈ G.
48
A Class of Control Certificates to Ensure Reach-While-Stay for Switched Systems
In other words, whenever the system is initialized inside the set I, it stays inside the safe set S until
it reaches the goal set G. Alternatively, we may express the specification in temporal logic as I =⇒
(S U G), where U is the temporal operator “until”.
We will assume that set S is a compact basic semialgebraic set. Typical examples include polytopes defined by linear inequalities or ellipsoids, that can be easily checked for the properties such as
compactness and nondegeneracy. Also, sets I and G are compact semialgebraic sets.
The special case when I = S will be called uninitialized RWS. Such a property simply states that the
system initialized inside the set S continues to remain in S until it reaches a goal state x ∈ G at some
finite time instant T . This case is suitable for building a finite abstraction as mentioned in Sec. 1.
2.3
Control Certificates
Encoding verification and synthesis problems into (control) certificates, which are defined by a set of
conditions, is a standard approach. For example Lyapunov functions have been used for ensuring stability
and barrier functions are employed to reason about safety properties. However, these functions are not
usually known in advance. To discover such a function in the first place, we solve a constrained problem
in which certificates are parameterized. Usually, certificates are defined over polynomials with unknown
coefficients and the problem reduces to finding proper coefficients for polynomials [21, 2]. For example,
to find a Lyapunov function, first, a template for Lyapunov function V is chosen: V = ∑i cα xα , where
xα is a monomial with degree greater than zero. Then, solving the following
constrained problem
yields
.
.
a Lyapunov function for proving stability to origin: (∃c) (∀x 6= 0) V (x) > 0 ∧ V (x) < 0 , where V is
∇V. f (x). In these techniques, it is essential to define control certificate with a simple structure that can
be discovered automatically. In the subsequent we combine the certificates for safety and liveness to
obtain a certificate for RWS properties.
3
RWS for Basic Semialgebraic Safe Sets
In this section, we first focus on the uninitialized RWS problem (I = S) and provide solutions for the case
when S is a basic nondegenerate semialgebraic set (see Def. 1).
Let S be a nondegenerate basic semialgebraic sets, as in Def. 1. Let ∂ S be partitioned into nonempty
facets F1 , . . . , Flk . Each facet Fk is, in turn, defined by two sets of polynomial inequalities Fk< of inactive
V
V
constraints and Fk= of active constraints: Fk = { pS, j ∈Fk< pS, j (x) < 0 ∧ pS, j ∈Fk= pS, j (x) = 0}.
For each state on a facet and not in G, we require the existence of a mode q, whose vector field
points inside S. Additionally, we will require a certificate V to decrease everywhere in S \ G. For any
.
polynomial p, let pq : (∇p) · fq (x). By combining conditions for safety and liveness, one can obtain the
following conditions:
.
x ∈ int(S) \ G =⇒ (∃ q) Vq (x) < −ε
.
V
.
V
(x)
<
−ε
∧
p
(x)
<
−ε
q
q
x ∈ F1 \ G =⇒ (∃ q)
=
p∈F1
..
.
.
V
.
V
(x)
<
−ε
∧
p
(x)
<
−ε
q
q
.
x ∈ Flk \ G =⇒ (∃ q)
p∈F =
lk
(3)
H. Ravanbakhsh & S. Sankaranarayanan
49
The first condition in Eq. (3) states that V must strictly decrease everywhere in the set int(S) \ G. The
subsequent conditions treat each facet Fj of the set S and posit the existence of a mode q for each state
that causes the active constraints and the function V to decrease.
However, we note that as the number of state variables increases, the number of facets can be exponential in the number of inequalities that define S [8] . This poses a serious limitation to the applicability
of Eq. (3).
Our solution to this problem, is based partly on the idea of exponential barriers discussed by Kong
et al. [12]. Rather than force the vector field to point inwards at each facet, we simply ensure that each
polynomial inequality pS, j ≤ 0 that defines S, satisfies a decrease condition outside set G. Thus, Eq. (3)
is replaced by a simpler (relaxed) condition:
.
x ∈ S \ G =⇒ (∃ q) Vq (x) < −ε ∧
V
.
j
( pS, j,q (x) + λ pS, j (x)) < −ε
.
(4)
Here λ > 0 is a user specified parameter. This rule is a relaxation of (3). The rule is made stronger
for larger values of λ . However, larger values of λ can cause numerical difficulties in practice while
searching for a control certificate.
.
For safety constraints, we require pq to be numerically ≤ −ε mainly, to avoid numerical issues. This
.
can be restrictive for cases where pS, j,q is simply zero. To go around this, we define a set of facets
.
Jq = { j|(∃x) pS, j,q (x) > 0} for each mode q. Informally speaking, Jq is set of all facets for which change
of pS, j must be considered when mode q is selected. Because for each facet j ∈
/ Jq , pS, j will never
increase as long as mode q is selected. Then, the conditions become:
.
x ∈ S \ G =⇒ (∃ q) Vq (x) < −ε ∧
V
.
j∈Jq
(( pS, j,q (x) + λ pS, j (x)) < −ε) .
(5)
As mentioned earlier, the problem of control synthesis consists of two phases. The first phase deals
with the problem of finding a control certificate V (x) that satisfies (5). We use a counter-example guided
inductive synthesis (CEGIS) framework to find such certificates. In the second phase, a switching strategy
is extracted from the control certificate to design the final controller. We now examine each phase, in
turn.
3.1
Discovering Control Certificates
We now explain the CEGIS framework that searches for a suitable control certificate V . To synthesize a
control certificate, we start with a parametric form Vc (x) = V (c, x) : ∑Ni=1 ci gi (x) with some (nonlinear)
basis functions g1 (x), . . . , gN (x) chosen by the user, and unknown coefficients c : (c1 , . . . , cN ), s.t. c ∈ C
for a compact set C ⊆ RN . The certificate V is a linear function over c.
The constraints from Eq. (5) become as follows:
W .
V
.
(∃c ∈ C) (∀x ∈ X) x ∈ S \ G =⇒ q Vq < −ε ∧ j∈Jq ( pS, j,q (x) + λ pS, j (x) < −ε) .
(6)
The constraints in Eq. (6) has a complex quantifier alternation structure involving the ∃c quantifier
nested outside the ∀x quantifier. First, we note that Jq is computed separately and here we assume it is
given. Next, we modify an algorithm commonly used for program synthesis problems to the problem of
synthesizing the coefficients c ∈ C [26].
The counterexample guided inductive synthesis (CEGIS) approach has its roots in program synthesis,
wherein it was proposed as a general approach to solve ∃∀ constraints that arise in such problems [26].
50
A Class of Control Certificates to Ensure Reach-While-Stay for Switched Systems
The key idea behind the CEGIS approach is to find solutions to such constraints while using a satisfiability (feasibility) solver for quantifier-free formulas that check whether a given set of constraints without
quantifiers have a feasible solution.
Solvers like Z3 allow us to solve many different classes of constraints with extensive support for linear arithmetic constraints [18]. On the other hand, general purpose nonlinear delta-satisfiability solvers
like dReal, support the solving of quantifier-free nonlinear constraints involving polynomials, trigonometric, and rational functions [3]. However, the presence of quantifiers drastically increases the complexity of solving these constraints. Here, we briefly explain the idea of CEGIS procedure for ∃∀ constraints
of the form
(∃ c ∈ C) (∀ x ∈ X) Ψ(c, x).
Here, c represents the unknown coefficients of a control certificate and x represents the state variables
of the system. Our goal is to find one witness for c that makes the overall quantified formula true. The
overall approach constructs, maintains, and updates two sets iteratively:
1. Xi ⊆ X is a finite set of witnesses. This is explicitly represented as Xi = {x1 , . . . , xi }.
2. Ci ⊆ C is a (possibly infinite) subset of available candidates. This is implicitly represented by a
constraint ψi (c), s.t. Ci : {c ∈ C | ψi (c)}.
In the beginning, X0 = {} and ψ0 : true representing the set C0 : C.
At each iteration, we perform the following steps:
(a) Choose a candidate solution ci+1 ∈ Ci . This is achieved by checking the feasibility of the formula
ψi . Throughout this paper, we will maintain ψi as a linear arithmetic formula that involves boolean
combinations of linear inequality constraints. Solving these problems is akin to solving linear optimization problems involving disjunctive constraints. Although the complexity is NP-hard, solvers like Z3
integrate fast LP solvers with Boolean satisfiability solvers to present efficient solutions [18].
(b) Test the current candidate. This is achieved by testing the satisfiability of ¬Ψ(c, x) for fixed c = ci+1 .
In doing so, we obtain a set of nonlinear constraints over x. We wish to now check if it is feasible.
If ¬Ψ(ci+1 , x) has no feasible solutions, then Ψ(ci+1 , x) is true (valid) for all x. Therefore, we can
stop with c = ci+1 as the required solution for c.
Otherwise, if ¬Ψ(c, x) is feasible for some x = xi+1 , we add it back as a witness: Xi+1 : Xi ∪ {xi+1 }.
The formula ψi+1 is given by
ψi+1 : ψi ∧ Ψ(c, xi+1 ) .
Note that ψi+1 =⇒ ψi , and ci+1 is no longer a feasible point for ψi+1 . The set Ci+1 described by ψi+1
is:
Ci+1 : {c ∈ C | Ψ(c, xi ) holds for each xi ∈ Xi+1 } .
The CEGIS procedure either (i) runs forever, or (ii) terminates after i iterations with a solution c : ci ,
or (iii) terminates with a set of witness points Xi proving that no solution exists.
We now provide further details of the CEGIS procedure adapted to find a certificate that satisfies
Eq. (6). In the CEGIS procedure, the formula Ψ(c, x) will have the following form:
x ∈ R1 =⇒ ϕ1 (c, x)
x ∈ R =⇒ ϕ (c, x)
2
2
(7)
...
x ∈ R =⇒ ϕ (c, x) ,
Nj
Nj
H. Ravanbakhsh & S. Sankaranarayanan
51
and each ϕ j for j = 1, . . . , N j has the form
_^
k
p j,k,l (c, x) > 0 ,
(8)
l
where p j,k,l (c, x) is a function linear in c and possibly nonlinear in x, depending on the dynamics and
template used for the control certificate.
The CEGIS procedure involves two calls to solvers: (a) Testing satisfiability of ψi (c) and (b) Testing
the satisfiability of ¬Ψ(ci+1 , x). We shall discuss each of these problems in the following paragraphs.
Finding Candidate Solutions:
exists c ∈ C s.t.
Given a finite set of witnesses Xi , a solution exists for ψi+1 iff there
Nj
^ ^
!
x ∈ R j =⇒
x∈Xi j=1
_^
k
p j,k,l (c, x) > 0
,
l
and since p j,k,l is a linear function in c, such c can be found by solving a formula in Linear Arithmetic
Theory (L A ).
Finding Witnesses: Finding a witness for a given candidate solution ci involves checking the satisfiability ¬Ψ. Whereas Ψ is a conjunction of N j clauses, ¬Ψ is a disjunction of clauses. The jth clause in
¬Ψ (1 ≤ j ≤ N j ) has the form
^_
x ∈ Rj ∧
p j,k,l (ci , x) ≤ 0 .
(9)
k
l
We will test each clause separately for satisfiability. Assuming that p j,k,l is a general nonlinear
function over x, SMT solvers like dReal [3] can be used to solve this over a compact set R j . Numerical
SMT solvers like dReal can either conclude that the given formula is unsatisfiable or provide a solution
to a “nearby” formula that is δ close. The parameter δ is adjusted by the user. As a result, dReal can
correctly conclude that the current candidate yields a valid certificate. On the other hand, its witness may
not be a witness for the original problem. In this case, using the spurious witness may cause the CEGIS
procedure to potentially continue (needlessly) even when a solution ci has been found. Nevertheless, the
overall procedure produces a correct result whenever it terminates with an answer.
Example 1 This example is adopted from [19]. There are two variables and three control modes with
the dynamics given below:
.
x1
−x2 − 1.5x1 − 0.5x13
0
0
2
+ Bq , Bq1 =
, Bq2 =
, Bq3 =
.
. =
x2
x1
−x22 + 2
−x2
10
The goal is to reach the target set G : (x1 +0.75)2 +(x2 −1.75)2 ≤ 0.252 , a circle centered at (−0.75, 1.75),
as shown in Figure 1a, while staying in the safe region given by the rectangle S0 : [−2, 2] × [−2, 3]:
S0 : {x|(x1 + 2)(x1 − 2) ≤ 0 ∧ (x2 + 2)(x2 − 3) ≤ 0} .
First, we shift co-ordniates to transform (−0.75, 1.75) as the new origin. Then, we use a quadratic
template for V (c1 x12 + c2 x1 x2 + c3 x22 ) , ε = 1, λ = 5. The solution V is found in 5 iterations. Then, we
translate the function back to the original co-ordinates:
V (x1 , x2 ) :37.782349x12 − 2.009762x1 x2 + 60.190607x1 + 4.415093x22 − 16.960145x2 + 37.411604 .
52
A Class of Control Certificates to Ensure Reach-While-Stay for Switched Systems
Example 2 A unicycle [37] has three variables. x and y .are position of the car and θ is its angle. The
.
.
dynamics of the system is x = u1 cos(θ ) , y = u1 sin(θ ) , θ = u2 , where u1 and u2 are inputs. Assuming
a switched system, we consider u1 ∈ {−1, 0, 1} and u2 ∈ {−1, 0, 1}. The safe set is [−1, 1] × [−1, 1] ×
[−π, π] and the target facet is x = 1. We use a template that is linear in (x, y) and quadratic in θ . Using
ε = 0.1 and λ = 0.5, the following CLF is found after 22 iterations:
V (x) : −x − y − 0.5881θ + θ 2 − 0.1956θ x + θ y .
Example 3 This example is adopted from [7]. There are four variables and two control inputs. The
dynamic is as follows:
.
u1
x1 + x2 + 8
x1
x.2 −x2 + x3 + 1 −u2
. =
x3 −2x3 + 2x4 + 1 + −2u1 .
.
u2
−3x4 + 1
x4
The region of interest S is hyber-box [−1, 1]4 and the input belongs to set [0, 1] × [0, 2]. The goal is to
reach facet x1 = 1, while staying in S as the safe region.
First, we discretize the control input to model the system as a switched system. For this purpose, we
assume u1 ∈ {0, 1} and u2 ∈ {0, 0.5, 1, 1.5, 2}. Then, we use a linear template for the CLF (c1 x1 + c2 x2 +
c3 x3 + c4 x4 ) , ε = 0.1, λ = 5. CEGIS framework finds certificate V (x) : −0.13333344(x1 + x2 + x3 + x4 ).
3.2
Control Design
Thus far, we discussed the CEGIS framework for finding a control certificate. Extracting the K function
from the certificate is now considered. Given a control certificate V satisfying Eq. (5), the choice of a
switching mode is dictated by a function ηq (x) defined for each state x ∈ X and mode q ∈ Q as follows:
.
ηq (x) : max Vq (x), ηS,1,q (x), · · · , ηS,k,q (x) ,
.
where for all j ∈ Jq , ηS, j,q is pS, j,q + λ pS, j and for j ∈
/ Jq , ηS, j,q = −∞ or equivalently ηS, j,q = −L for
some large constant L.
The idea is that whenever (at time t) the controller chooses a mode q s.t. ηq (x(t)) < −ε, one can
guarantee that ηq (x(t)) < 0 holds for all t ∈ [T, T + δ ), for some minimum time δ . Therefore, for some
fixed εs (0 < εs < ε), the function K for any x ∈ S \ G can be defined as
q̂
if
η
(x)
≥
−ε
∧
η
(x)
<
−ε
q
s
q̂
K (q, x) :=
(10)
q otherwise .
In other words, the controller state persists in a given mode q until ηq (x) ≥ −εs . Then, given that
x ∈ S \ G, Eq. (5) will provide us a new control mode q̂ that satisfies ηq̂ (x) < −ε. This mode is chosen
as the next mode to switch to.
Example 4 Consider once again, the problem from Ex. 1. Using the defined function V (x1 , x2 ), Eq. (10)
yields a controller. Figure 1b shows some of the simulation traces of this closed loop system, demonstrating the RWS property.
H. Ravanbakhsh & S. Sankaranarayanan
53
(a)
(b)
Figure 1: (a) Region G for Example 1 is shown shaded in the center, and the vector fields for modes
q1 , q2 and q3 are shown in red, green and blue, respectively. Level-sets of V are shown with black dashed
lines. (b) Closed loop trajectories for Example 1 using the controller defined by Eq. (10). The segments
shown in colors red, green and blue correspond to the modes q1 , q2 and q3 , respectively.
We now establish the key result that provides a minimum dwell time guarantee.
Lemma 1 There exists a δ > 0 s.t. for all initial conditions x(T ) ∈ S \ G, if ηq (x(T )) < −ε, and if the
mode of the system is set to q at time T , then
(∀t ∈ [T, T + δ ]) (x(t) ∈ S \ G) =⇒ ηq (x(t)) ≤ −εs .
Proof Let T + δ be the earliest time instant, where ηq (x(T + δ )) ≥ −εs while at the same time
(∀t ∈ [T, T + δ ]) q(t) = q, x(t) ∈ S \ G .
At time T , ηq (x(T )) < −ε and at time T + δ , ηq (x(T + δ )) = −εs . Note that ηq (x) is defined as
max(α1 (x), . . . , αm (x)) for some smooth functions α1 , . . . , αm . As a result, Since S is a bounded set, and
p, fq , and V are bounded over S, there exists a constant Λ > 0 s.t.
.
(∀ x ∈ S) αi,q ≤ Λ .
Therefore, for each αi , we have
Z T +δ
αi (x(T + δ )) = αi (x(T )) +
.
αi,q (x(t))dt ≤ αi (x(T )) + Λδ .
t=T
As a result, we conclude that
ηq (x(T + δ )) = max αi (x(T + δ )) = α j∗ (x(T + δ )) ≤ α j∗ (x(T )) + Λδ ≤ ηq (T ) + Λδ .
i
(11)
54
A Class of Control Certificates to Ensure Reach-While-Stay for Switched Systems
Therefore, we can conclude −εs < −ε + Λδ =⇒
ε−εs
Λ
< δ and there exists a fixed δ >
ε−εs
Λ
> 0 s.t.
(∀t ∈ [T, T + δ )) ηq (x(t)) < −εs .
Eq. (10) gives a switching strategy which respects the. min-dwell time and as long as x(t) ∈ S, the
.
controller guarantees ηq(t) (x(t)) ≤ −εs . I.e. for all j ∈ Jq , Vq (x(t)) ≤ −εs and pS, j,q (x(t))+λ pS, j (x(t)) ≤
−εs .
Theorem 1 Given nondegenerate basic semialgebraic set S, a semialgebraic set G, and a function V
(satisfying Equation (5)), the control strategy defined by Eq. (10) respects the min-dwell time property
and guarantees the RWS property defined by S, G: S =⇒ SU G.
Proof As discussed, there
exists a controller which respects the min-dwell time property. Also, the
.
.
controller guarantees Vq (x) ≤ −εs and ( pS, j,q (x) + λ pS, j (x) ≤ −εs (for all j ∈ Jq ), as long as x ∈ S \ G.
Assume x(t) is on the boundary of S (and not in G) at some time t. Because S is assumed to be
a nondegenerate basic semialgebraic set, there exists at least one j s.t. pS, j (x(t)) = 0. If j ∈
/ Jq , by
.
definition, pS, j,q is negative for all states and pS, j,q remains ≤ 0 as long as mode q is selected. Otherwise
.
( j ∈ Jq ), we obtain pS, j,q (x(t)) ≤ −εs < 0. Therefore, there exits τ j > 0, s.t. s ∈ (t,t + τ j ), we conclude
that pS, j,q (x(s)) < 0. As a result, the trajectory cannot leave the set S.
Thus, the trace cannot leave S, unless it reaches G. Now, we show that the trajectory cannot stay
inside S \ G forever. By the construction of the controller, we can conclude
. time diverges (because the
controller respects the min-dwell time property) and that V decreases (Vq (x(t)) ≤ −εs ). However, the
value of V is bounded on bounded set S \ G. Therefore, x cannot remain in S \ G and the only possible
outcome for the trace is to reach G.
4
RWS for Semialgebraic Safe Set
As Habets et al [6] discussed, control-to-facet problems can be used to build an abstraction. Here,
we demonstrate that the method described so far can be integrated in this framework to tackle more
complicated problems. First, we briefly explain how the method works. For a more detailed discussion,
the reader can refer to [6] or [11].
First, state space is decomposed into polytopes according to the specifications. Here, we can use
basic semialgebraic sets instead of polytopes. Then, for each such a set u, we consider an abstract state
A (u). Furthermore, for each of its n − 1 dimensional facet F, a control-to-facet problem is solved. The
corresponding problem is to find a control strategy to reach F starting from u. If the control-to-facet
problem is solved successfully, then for each basic semialgebraic set v with a n − 1 dimensional facet
F 0 ⊆ F, an edge from A (u) to A (v) (with label/action F) is added to the abstraction. Also for each
basic semialgebraic set u, one can check if u is a control invariant to build self loops. However, for
RWS properties, self loops are redundant and we skip them here. After building the abstract system, we
use standard techniques to solve the problem for finite systems. If the problem could be solved for the
abstract system, then, one can design a controller.
First, for each abstract state A (u), there is at least one action F that agrees with the winning strategy
for the abstract system. Let that action be F (A (u)). The idea is to implement transition F (A (u)),
using controller Ku,F (A (u)) for the corresponding control-to-facet problem [6]. Formally, the controller
H. Ravanbakhsh & S. Sankaranarayanan
55
can be defined as follows:
Ku1 ,F (A (u1 )) (q, x)
.
K (q, x) = ..
K
(q, x)
x ∈ u1
(12)
x ∈ us .
us ,F (A (us ))
When x belongs to multiple sets, one can break the tie by some ordering, where states in the winning
set have priorities. It is worth mentioning that combining these controllers together, does not produce
any Zeno behavior as it is guaranteed that each abstract state is visited only once for RWS properties.
However, superdense switching is possible as two facets of a polytope can get arbitrarily close.
If one is interested in LTL properties (not just reach-while-stay) or min-dwell time property, one
possible solution is to use fat facets, where the target sets are n dimensional goal sets. This extends the
domain of the control-to-facet problem to adjacent basic semialgebraic sets as well. Also, it allows the
controller to continue using current sub-controller for some minimum time (if min-dwell time requirement is not met), before changing the sub-controller (at the switch time).
Example 5 Consider again the system from Example 1, with the addition of some obstacles [19]. More
precisely, as shown in Fig 2a, safe set is S = S0 \ (O1 ∪ O2 ). First, the safe set is decomposed into four
basic semialgebraic sets, which are shown with R0 to R3 in Fig. 2a.
R0 is the target set. Next, we build a transition
O
relation between four abstract states, representR
ing four basic semialgebraic sets. This is done
R
Converge
R
by solving seven RWS problems for basic semialLeft
Diverge
R
gebraic sets. For R1 to R0 , we use a quadratic
R
R
template for V , and for other problems, we use
Up
linear template. The abstract system is shown in
Left
R
O
Fig. 2b. Next, the problem is solved for the abR
stract system. The solution to the abstract system
(b)
is simple: if the state is in R2 , the controller uses
(a)
the left facet to reach R1 or R3 . Otherwise, if the
state is in R3 , the controller uses the upper facet Figure 2: (a) Schematic view of state decomposito reach R1 and finally, if the state is in R1 , the tion. (b) Finite abstraction for the original problem.
controller makes sure the state reaches R0 .
2
1
0
0
2
1
2
1
3
3
Example 6 This example is a path planning problem for the unicycle [25]. Projection of safe set on
x and y yields a maze. The target set is placed at the right bottom corner of the maze (Fig. 3). Using
specification-guided technique, we modeled the system with 53 polyhedra. Each polyhedron is treated
as a single state and a transition relation is built by solving 113 control-to-facet problems. Then, the
problem is solved over the finite graph. The total computation took 1484 seconds. The figure also shows
a single trajectory of the closed loop system.
Example 7 This example is similar to Example 6, except for the fact that there is no direct control over
the angular velocity. More precisely, only the angular acceleration
is controllable and the system would
.
.
.
.
have the following dynamics x = u1 cos(θ ), y =p
u1 sin(θ ), θ = ω, ω = u2 . Also, we assume ω ∈ [−1, 1].
By changing the coordinates one can use r = x2 + y2 , z1 = x cos(θ ) + y sin(θ ) and z2 = y cos(θ ) −
x sin(θ ) to define position and angle of the car(cf. [13] for details). Then, we use the following template
V (x, y, θ , ω) = c1 r2 + c2 z1 + c3 z2 ω + c4 ω 2 , where the origin is located just outside of the target facet.
Using this template, we find control certificates for all 113 control-to-facet problems in 5296 seconds.
56
A Class of Control Certificates to Ensure Reach-While-Stay for Switched Systems
Figure 3: Region G is shown shaded in Orange, and unsafe regions are shown in blue. An execution
trace of the car is shown for x and y variables.
5
Initialized Reach-While-Stay
So far, we discussed uninitialized RWS specifications (S =⇒ SU G). In these systems, we use boundary
of safe set as barrier. However, as pointed out by Lin et al. [14], this may not be the case. Now, we
consider the initialized problem for a given initial set I (I =⇒ SU G). To avoid technical difficulties, we
assume that I ⊆ int(S). The solution is to create a composite barrier that is formed by the boundary of S
as well as other a priori unknown barrier functions.
Barrier Functions: We recall that for a control barrier function [29, 36], the following conditions are
considered
x ∈ ∂ S =⇒ B(x) > 0
x ∈ I =⇒ B(x)
(13)
<0
x ∈ S =⇒
.
B(x) = 0 =⇒ (∃q)Bq (x) < −ε .
This ensures that B(x) = 0 is a barrier and ∂ S is unreachable. Eq. (13), combined with the smoothness
of B and fq ensures that as soon as the state is sufficiently “close” to the barrier, it is possible to choose
a control mode that ensures the local decrease of the B.
The condition in Equation (13) can be encoded into the CEGIS framework. However, the presence
of the equality B(x) = 0 poses practical problems. In particular, it requires for each candidate Bc , to find
a counterexample x s.t. Bc (x) 6= 0. Unfortunately, such an assertion is easy to satisfy, resulting in the
procedure always exceeding the maximum number of iterations permitted.
Again, we find that the following relaxation of the third condition is particularly effective in our
experiments
x ∈ ∂ S =⇒ B(x) > 0
x ∈ I =⇒ B(x) < 0
(14)
.
.
W
x ∈ S =⇒ q Bq (x) − λ B(x) < −ε ∨ Bq (x) + λ B(x) < −ε ,
for some constant λ .
Intuitively, by choosing λ = 0, the condition is similar to that of Lyapunov functions, whereas as
|λ | → ∞, the condition gets less conservative and in the limit, it is equivalent to the original condition.
H. Ravanbakhsh & S. Sankaranarayanan
57
In fact, for smaller |λ | CEGIS terminates faster, but at the cost of missing potential solutions. On the
other hand, using larger |λ |, is less conservative at the cost of CEGIS timing out. We also note that this
formulation is less conservative than the one introduced by Kong et al. [12] as our formulation uses two
exponential conditions which only forces decrease of value of B around B∗ = {x | B(x) = 0}.
To solve the RWS in general form, we define a finite set of barriers B with the following conditions:
x ∈ ∂ S =⇒ B∈B B(x) > 0
V
x ∈ I =⇒ B∈B B(x) < 0 .
W
(15)
.
Also for each mode q, Bq is defined as Bq = {B ∈ B|(∃x)Bq (x) > 0}. Then, existence of a proper
mode can be encoded as the following:
.
V
.
Bq (x) + λ B(x) < −ε∨
.
x ∈ S \ G =⇒
Vq (x) < −ε ∧ B∈Bq
.
(16)
Bq (x) − λ B(x) < −ε
Theorem 2 Given nondegenerate basic semialgebraic set S, semialgebraic sets I and G, function V ,
and a non-empty set of functions B (satisfying Equation (15) and (16)), there is a control strategy that
respects the min-dwell time property and guarantees the RWS property: I =⇒ SU G.
To simplify these constraints and reduce the number of unknowns, one can use some of pS,i ’s to fix
some of these barriers, which yields conditions similar to the ones used for the uninitialized problem.
This trick is demonstrated in the following example.
Example 8 This example is taken from [16], in which a DC-DC converter is modeled with two variables
i and v.
The system has two modes q1 and q2 , with the following
dynamics:
B2 = 0
(.
B4 = 0
S
i = 0.0167i + 0.3333
q1 : .
v = −0.0142v
(.
G
i = −0.0183i − 0.0663v + 0.3333
q2 : .
v = −0.0711i − 0.0142v .
The safe set is S : [0.65, 1.65] × [4.95, 5.95] and the goal set
is G : [1.25, 1.45] × [5.55, 5.75]. We assume initial set to be
I : [0.85, 0.95] × [5.15, 5.25] (Fig. 4). Then, we use 5 barriers
B0 , . . . , B4 . Using boundaries of S, we choose B1 , . . . , B4 as
follows:
B1 =0.65 − i + εb
B2 = 1.65 − i + εb
B3 =4.95 − v + εb
B4 = 5.95 − v + εb ,
B0 = 0
I
B5 = 0
B3 = 0
B1 = 0
Figure 4: The blue lines are the barriers
and the red lines are level-sets of the Lyapunov function.
where εb > 0 is small enough that I ⊂ int( 4i=1 Bi ). In this case, we choose εb = 0.01. Notice that
such εb always exists by the definition. Next, we assume B0 = V and both have the following template:
T
B0 = V : c1 (i − 1.35)2 + c2 (i − 1.35)(v − 5.65) + c3 (v − 5.65)2 − 1 .
This template is chosen in a way that V is a quadratic function with minimum value of −1 for the point
of interest i = 1.35, v = 5.65. So far, we used these tricks to reduce the number of unknowns for barriers.
58
A Class of Control Certificates to Ensure Reach-While-Stay for Switched Systems
Table 1: Results of Comparison with SCOTS on examples
Legend: n: # state variables, itr : # iterations, Time: total computation time, η: state discretization step,
τ: time step. All timings are in seconds and rounded, TO: timeout (> 10 hours).
Problem
ID
n
Ex. 5 2
Ex. 8 2
Ex. 6 3
Ex. 3 4
Ex. 7 4
η
0.162
0.012
0.22 ×0.1
0.03×0.13
0.12 ×0.052
SCOTS
τ
itr Time
0.12
18
0
1.0
106
1
0.3
404
989
0.005
48
304
0.3
TO
CEGIS
δ
Time
−4
10
3
10−4
39
10−4 1484
10−5
3
−4
10
5296
However, our method fails to find a certificate. Next, we add one more barrier (B5 ) to the formulation
and we use the following template for B5 : c4 (i − 0.9)2 + c5 (i − 0.9)(v − 5.2) + c6 (v − 5.05)2 − 1, which
is a quadratic function with minimum value of −1 for initial point i = 0.9, v = 5.2. This time, we can
successfully find a control certificate. The final barriers and level-sets of the Lyapunov function is shown
in Fig. 4.
Comparison: While abstraction based methods can provide a near optimal solution (are relatively complete), these methods can be computationally expensive. On the other hand, our method is a Lyapunovbased method and the solution is not necessarily (relatively) complete. For example, our approach assumes that control certificates with a given form (that is given as input by user) exist . As such, the
existence of such certificates is not guaranteed and thus, our approach lacks the general applicability of a
fixed-point based synthesis. Also, for initialized problems our method needs an initial set as input, while
for the abstraction based methods, maximum controllable region can be obtained without the need for
specifying the initial set. However, our method is relatively more scalable thanks to recent development
in SMT solvers. Here, for the sake of completeness, we provide a brief comparison with SCOTS toolbox [25] for the examples provided in this article. To compare Example 3, we use fat facet and assume
target set has a volume (otherwise, because of time discretization, SCOTS cannot find a solution). More
precisely, we use target set [1, 1.2] × [−1, 1]3 instead of [1, 1] × [−1, 1]3 .
All the experiments are ran on a laptop with Core i7 2.9 GHz CPU and 16GB of RAM. The results are
reported in Table 1. We also note that if we use larger values for SCOTS parameters, SCOTS fails to solve
these problems (initial set is not a subset of controllable region). Table 1 shows that SCOTS performs
much better for Example 5 and 8 for which there are only 2 state variables. For Example 6, both methods
have similar performances. And for Example. 3 and Example.7, which have 4 state variables, our method
is faster.
6
Conclusions
In this paper, given a switched system, we addressed controller synthesis problems for RWS with composite barriers. Specifically, we addressed uninitialized problems which are useful for building an abstraction, as well as initialized problems. For each problem, we provided sufficient conditions in terms of
“existence of a control certificate”. Also, we demonstrated that searching for a control certificate can be
encoded into constrained problems and solving these problems is computationally feasible. In the future,
H. Ravanbakhsh & S. Sankaranarayanan
59
we wish to investigate how the initialized RWS problems can be extended to be used along fixed-point
computation based techniques as it allows more flexible switching strategies.
Acknowledgments
This work was funded in part by NSF under award numbers SHF 1527075 and CPS 1646556. All
opinions expressed are those of the authors and not necessarily of the NSF.
References
[1] Zvi Artstein (1983): Stabilization with relaxed controls. Nonlinear Analysis: Theory, Methods &
Applications 7(11), pp. 1163 – 1173, doi:10.1016/0362-546X(83)90049-4. Available at http://www.
sciencedirect.com/science/article/pii/0362546X83900494.
[2] Rayna Dimitrova & Rupak Majumdar (2014): Deductive control synthesis for alternating-time logics. In:
2014 International Conference on Embedded Software, EMSOFT 2014, New Delhi, India, October 12-17,
2014, pp. 14:1–14:10, doi:10.1145/2656045.2656054.
[3] Sicun Gao, Soonho Kong & Edmund M. Clarke (2013): dReal: An SMT Solver for Nonlinear Theories over
the Reals. In: Automated Deduction - CADE-24 - 24th International Conference on Automated Deduction,
Lake Placid, NY, USA, June 9-14, 2013. Proceedings, pp. 208–214, doi:10.1007/978-3-642-38574-2 14.
Available at https://doi.org/10.1007/978-3-642-38574-2_14.
[4] L. El Ghaoui & V. Balakrishnan (1994): Synthesis of fixed-structure controllers via numerical optimization. In: Proceedings of 1994 33rd IEEE Conference on Decision and Control, 3, pp. 2678–2683 vol.3,
doi:10.1109/CDC.1994.411398.
[5] A. Girard, G. Pola & P. Tabuada (2010): Approximately Bisimilar Symbolic Models for Incrementally Stable Switched Systems. IEEE Transactions on Automatic Control 55(1), pp. 116–126,
doi:10.1109/TAC.2009.2034922.
[6] L. C. G. J. M. Habets, P. J. Collins & J. H. van Schuppen (2006): Reachability and control synthesis for
piecewise-affine hybrid systems on simplices. IEEE Transactions on Automatic Control 51(6), pp. 938–948,
doi:10.1109/TAC.2006.876952.
[7] L.C.G.J.M. Habets & J.H. van Schuppen (2004): A control problem for affine dynamical systems on a fulldimensional polytope. Automatica 40(1), pp. 21 – 35, doi:10.1016/j.automatica.2003.08.001. Available at
http://www.sciencedirect.com/science/article/pii/S0005109803002620.
[8] M. K. Helwa & M. E. Broucke (2011):
doi:10.1109/CDC.2011.6160866.
Monotonic reach control on polytopes, pp. 4741–4746.
[9] Z. Huang, Y. Wang, S. Mitra, G. E. Dullerud & S. Chaudhuri (2015): Controller synthesis with inductive
proofs for piecewise linear systems: An SMT-based algorithm. In: 2015 54th IEEE Conference on Decision
and Control (CDC), pp. 7434–7439, doi:10.1109/CDC.2015.7403394.
[10] Manuel Mazo Jr., Anna Davitian & Paulo Tabuada (2010): PESSOA: A Tool for Embedded Controller
Synthesis. In: Computer Aided Verification, 22nd International Conference, CAV 2010, Edinburgh, UK,
July 15-19, 2010. Proceedings, pp. 566–569, doi:10.1007/978-3-642-14295-6 49. Available at https:
//doi.org/10.1007/978-3-642-14295-6_49.
[11] M. Kloetzer & C. Belta (2008): A Fully Automated Framework for Control of Linear Systems
from Temporal Logic Specifications. IEEE Transactions on Automatic Control 53(1), pp. 287–297,
doi:10.1109/TAC.2007.914952.
[12] Hui Kong, Fei He, Xiaoyu Song, William N. N. Hung & Ming Gu (2013): Exponential-Condition-Based Barrier Certificate Generation for Safety Verification of Hybrid Systems. In: Computer Aided Verification - 25th
60
A Class of Control Certificates to Ensure Reach-While-Stay for Switched Systems
International Conference, CAV 2013, Saint Petersburg, Russia, July 13-19, 2013. Proceedings, pp. 242–257,
doi:10.1007/978-3-642-39799-8 17. Available at https://doi.org/10.1007/978-3-642-39799-8_17.
[13] Daniel Liberzon (2012): Switching in systems and control.
doi:10.1007/978-1-4612-0017-8.
Springer Science & Business Media,
[14] Z. Lin & M. E. Broucke (2007): Reachability and control of affine hypersurface systems on polytopes. In:
2007 46th IEEE Conference on Decision and Control, pp. 733–738, doi:10.1109/CDC.2007.4434805.
[15] J. Liu, N. Ozay, U. Topcu & R. M. Murray (2013): Synthesis of Reactive Switching Protocols
From Temporal Logic Specifications. IEEE Transactions on Automatic Control 58(7), pp. 1771–1785,
doi:10.1109/TAC.2013.2246095.
[16] Sebti Mouelhi, Antoine Girard & Gregor Gössler (2012): CoSyMA: A Tool for Controller Synthesis Using Multi-scale Abstractions. Research Report RR-8108, INRIA. Available at https://hal.inria.fr/
hal-00743982.
[17] Sebti Mouelhi, Antoine Girard & Gregor Gößler (2013): CoSyMA: a tool for controller synthesis using multiscale abstractions. In: Proceedings of the 16th international conference on Hybrid systems: computation and
control, HSCC 2013, April 8-11, 2013, Philadelphia, PA, USA, pp. 83–88, doi:10.1145/2461328.2461343.
[18] Leonardo Mendonça de Moura & Nikolaj Bjørner (2008): Z3: An Efficient SMT Solver. In: Tools and
Algorithms for the Construction and Analysis of Systems, 14th International Conference, TACAS 2008,
Held as Part of the Joint European Conferences on Theory and Practice of Software, ETAPS 2008, Budapest,
Hungary, March 29-April 6, 2008. Proceedings, pp. 337–340, doi:10.1007/978-3-540-78800-3 24. Available
at https://doi.org/10.1007/978-3-540-78800-3_24.
[19] P. Nilsson & N. Ozay (2014): Incremental synthesis of switching protocols via abstraction refinement. In:
53rd IEEE Conference on Decision and Control, pp. 6246–6253, doi:10.1109/CDC.2014.7040368.
[20] N. Ozay, J. Liu, P. Prabhakar & R. M. Murray (2013): Computing augmented finite transition systems to
synthesize switching protocols for polynomial switched systems. In: 2013 American Control Conference, pp.
6237–6244, doi:10.1109/ACC.2013.6580816.
[21] S. Prajna, A. Papachristodoulou & P. A. Parrilo (2002): Introducing SOSTOOLS: a general purpose sum of
squares programming solver. In: Proceedings of the 41st IEEE Conference on Decision and Control, 2002.,
1, pp. 741–746 vol.1, doi:10.1109/CDC.2002.1184594.
[22] H. Ravanbakhsh & S. Sankaranarayanan (2015): Counter-Example Guided Synthesis of control Lyapunov
functions for switched systems. In: 2015 54th IEEE Conference on Decision and Control (CDC), pp. 4232–
4239, doi:10.1109/CDC.2015.7402879.
[23] Hadi Ravanbakhsh & Sriram Sankaranarayanan (2016): Robust Controller Synthesis of Switched Systems Using Counterexample Guided Framework. In: Proceedings of the 13th International Conference on Embedded
Software, EMSOFT ’16, ACM, pp. 8:1–8:10, doi:10.1145/2968478.2968485.
[24] Bartek Roszak & Mireille E. Broucke (2006): Necessary and sufficient conditions for reachability on a
simplex. Automatica 42(11), pp. 1913 – 1918, doi:10.1016/j.automatica.2006.06.003. Available at http:
//www.sciencedirect.com/science/article/pii/S0005109806002445.
[25] Matthias Rungger & Majid Zamani (2016): SCOTS: A Tool for the Synthesis of Symbolic Controllers. In:
Proceedings of the 19th International Conference on Hybrid Systems: Computation and Control, HSCC
2016, Vienna, Austria, April 12-14, 2016, pp. 99–104, doi:10.1145/2883817.2883834.
[26] Armando Solar Lezama (2008): Program Synthesis By Sketching. Ph.D. thesis, EECS Department, University of California, Berkeley. Available at http://www2.eecs.berkeley.edu/Pubs/TechRpts/2008/
EECS-2008-177.html.
[27] Eduardo D. Sontag (1989): A universal construction of Artstein’s theorem on nonlinear stabilization. Systems
& Control Letters 13(2), pp. 117 – 123, doi:10.1016/0167-6911(89)90028-5. Available at http://www.
sciencedirect.com/science/article/pii/0167691189900285.
H. Ravanbakhsh & S. Sankaranarayanan
61
[28] Ankur Taly, Sumit Gulwani & Ashish Tiwari (2011): Synthesizing switching logic using constraint solving.
STTT 13(6), pp. 519–535, doi:10.1007/s10009-010-0172-8. Available at https://doi.org/10.1007/
s10009-010-0172-8.
[29] Ankur Taly & Ashish Tiwari (2010): Switching logic synthesis for reachability. In: Proceedings of the 10th
International conference on Embedded software, EMSOFT 2010, Scottsdale, Arizona, USA, October 24-29,
2010, pp. 19–28, doi:10.1145/1879021.1879025.
[30] Weehong Tan & Andrew Packard (2004): Searching for control Lyapunov functions using sums of squares
programming. In: Allerton conference on communication, control and computing, pp. 210–219.
[31] Y. Tazaki & J. i. Imura (2012): Discrete Abstractions of Nonlinear Systems Based on Error Propagation
Analysis. IEEE Transactions on Automatic Control 57(3), pp. 550–564, doi:10.1109/TAC.2011.2161789.
[32] Wolfgang Thomas, Thomas Wilke et al. (2002): Automata, logics, and infinite games: a guide to current
research. 2500, Springer Science & Business Media, doi:10.1007/3-540-36387-4.
[33] Peter Wieland & Frank Allgwer (2007): CONSTRUCTIVE SAFETY USING CONTROL BARRIER FUNCTIONS. IFAC Proceedings Volumes 40(12), pp. 462 – 467, doi:10.3182/20070822-3-ZA-2920.00076. Available at http://www.sciencedirect.com/science/article/pii/S1474667016355690. 7th IFAC
Symposium on Nonlinear Control Systems.
[34] T. Wongpiromsarn, U. Topcu & A. Lamperski (2016): Automata Theory Meets Barrier Certificates: Temporal Logic Verification of Nonlinear Systems. IEEE Transactions on Automatic Control 61(11), pp. 3344–3355,
doi:10.1109/TAC.2015.2511722.
[35] Tichakorn Wongpiromsarn, Ufuk Topcu, Necmiye Ozay, Huan Xu & Richard M. Murray (2011): TuLiP: A
Software Toolbox for Receding Horizon Temporal Logic Planning. In: Proceedings of the 14th International
Conference on Hybrid Systems: Computation and Control, HSCC ’11, ACM, New York, NY, USA, pp.
313–314, doi:10.1145/1967701.1967747.
[36] Xiangru Xu, Paulo Tabuada, Jessy W. Grizzle & Aaron D. Ames (2015): Robustness of Control Barrier Functions for Safety Critical Control**This work is partially supported by the National Science Foundation Grants
1239055, 1239037 and 1239085. IFAC-PapersOnLine 48(27), pp. 54 – 61, doi:10.1016/j.ifacol.2015.11.152.
Available at http://www.sciencedirect.com/science/article/pii/S2405896315024106. Analysis and Design of Hybrid Systems ADHS.
[37] M. Zamani, G. Pola, M. Mazo & P. Tabuada (2012): Symbolic Models for Nonlinear Control Systems Without Stability Assumptions. IEEE Transactions on Automatic Control 57(7), pp. 1804–1809,
doi:10.1109/TAC.2011.2176409.
| 3cs.SY
|
Subjective Knowledge Acquisition and Enrichment
Powered By Crowdsourcing
Rui Meng
Hao Xin
Lei Chen
Yangqiu Song
arXiv:1705.05720v1 [cs.DB] 16 May 2017
Department of Computer Science and Engineering, HKUST, Hong Kong SAR, China
{rmeng,hxinaa,leichen,yqsong}@cse.ust.hk
ABSTRACT
being “subjective” by workers [19], 63% of location-based queries
in mobile search are asking for subjective opinions [7], and need the
corresponding subjective knowledge as the query answers. For example, there might exist such queries, “popular American singers”
or “beautiful cities in Europe”, we refer the knowledge concerning popular singers and beautiful cities as the subjective knowledge. More specifically, subjective knowledge refers to the dominant opinion about whether a particular subjective property applies
to entities of a particular type [25]. For instance, given a pair consists of a subjective property1 and a type from a KB (subjective
property-type pair, ST pair), e.g., POPULAR and SINGER, we can
find a list of instances of the type SINGER from the KB, e.g., E LVIS
P RESLEY, where the dominant opinion of “whether Elvis Presley is
a popular singer” is a piece of subjective knowledge. As this kind
of information is missing in existing KBs, queries concerning such
information cannot be satisfied. Fortunately, crowdsourcing, which
has been recently proved to be successful for various human intrinsic tasks such as entity resolution [28], knowledge extraction [13],
translation [31], etc., provides a natural and reliable way of obtaining the subjective knowledge by collecting opinions from workers.
Many works have been done to perform KB enrichment, completion and population [9] [29] [4] [12] [11], but none of these works
focus on the subjective dimension. For subjective knowledge acquisition, the state-of-the-art approach is to use information extraction techniques to mine the text of Web contents [25]. However, it
only relies on machine-based technique and online Web data, and
does not consider to incorporate the wisdom of the crowd and existing KB information. Thus, the precision is far from satisfactory,
i.e. SURVEYOR has the precision of 77% [25]. To the best of our
knowledge, we are the first to leverage the collaborative knowledge
from both the crowd and existing KBs to perform subjective knowledge acquisition and KB enrichment in the subjective dimension.
Challenge. Leveraging the power of the crowd for knowledge
acquisition comes with the challenge of “How to resolve the conflict between large scale knowledge facts and the limited crowdsourcing resource?”. Real world KBs are often in very large scales,
e.g., YAGO has 2,747,873 entities and 292,898 types, DBpedia has
2,531,369 entities and 827 types, while each crowdsourcing operation is associated with a monetary cost and is somewhat timeconsuming. Therefore, it is infeasible and costly to ask the crowd
to carry the whole burden of subjective knowledge acquisition task.
In our system CoSKA, we make use of the knowledge in existing
KBs and the semantic relationship among subjective properties to
perform knowledge inference and based on the inference power,
the most beneficial questions are identified for crowdsourcing.
Framework. The input of CoSKA is a list of ST pairs mined
from the corpus and a KB. The output is a list of subjective knowl-
Knowledge bases (KBs) have attracted increasing attention due to
its great success in various areas, such as Web and mobile search.
Existing KBs are restricted to objective factual knowledge, such as
CITY POPULATION or FRUIT SHAPE , whereas, subjective knowledge, such as BIG CITY, which is commonly mentioned in Web and
mobile queries, has been neglected. Subjective knowledge differs
from objective knowledge in that it has no documented or observed
ground truth. Instead, the truth relies on people’s dominant opinion.
Thus, we can use the crowdsourcing technique to get opinion from
the crowd. In our work, we propose a system, called crowdsourced
subjective knowledge acquisition (CoSKA), for subjective knowledge acquisition powered by crowdsourcing and existing KBs. The
acquired knowledge can be used to enrich existing KBs in the subjective dimension which bridges the gap between existing objective
knowledge and subjective queries. The main challenge of CoSKA is
the conflict between large scale knowledge facts and limited crowdsourcing resource. To address this challenge, in this work, we define knowledge inference rules and then select the seed knowledge
judiciously for crowdsourcing to maximize the inference power under the resource constraint. Our experimental results on real knowledge base and crowdsourcing platform verify the effectiveness of
CoSKA system.
1.
INTRODUCTION
Motivation. In recent years, knowledge bases (KBs) have become increasingly popular and large-scale KBs have been constructed, such as Freebase [2], DBpedia [17], YAGO [10], KnowItAll [8], etc. The KBs encode information and knowledge of the
real world in a structured, machine-understandable way which can
empower various kinds of applications, especially Web and mobile
search. Despite of containing millions of knowledge facts on large
amount of entities and relations, the knowledge encoded by these
KBs is limited in objective dimension. In other words, existing KBs
have so far focused on encoding objective knowledge facts, which
are factual and observable, such as FRUIT SHAPE, MOVIE DIREC TOR and so forth. In contrast, many real world queries are subjective, e.g., around 20% of product-related queries are labeled as
1
1
Typically expressed as an adjective
Figure 1: Framework of CoSKA.
edge facts and enriched KB. CoSKA consists of three stages: ST
pair selection, crowdsourced ST pair applying, and knowledge inference. The details of the framework is shown in Figure 1.
1) ST Pair Selection: Given the large amount of ST pairs, we
need to identify the benefit of each pair and select them judiciously
for subsequent crowdsourced ST pair applying as the process needs
the involvement of crowd workers. We first define some subjective
knowledge inference rules. Then the ST pair selection problem is
formulated as a Maximum Knowledge Inference Problem. We show
that the problem is NP-hard and propose a diversity-aware forward
greedy algorithm for ST pair selection.
2) Crowdsourced ST Pair Applying: For each selected ST pair,
the task of subjective knowledge acquisition is to identify the opinions that whether the subjective property can be applied to the instances of the type powered by crowdsourcing, referred as crowdsourced ST pair applying. However, asking the crowd for every
instance is still too costly as a type could contain hundreds of thousands instances. In order to improve the scalability of the knowledge acquisition task, we formulate the crowdsourced ST pair applying as a binary classification problem. The objective knowledge
of instances in existing KBs is selected as the features. We adopt
a representative sampling strategy to sample a set of instances to
ask the crowd and the classifier is trained based on the collected
answers.
3) Knowledge Inference: After the crowdsourced ST pair applying process, we have acquired a set of subjective knowledge
facts. To further improve the scalability of our system and derive
more subjective knowledge facts, we perform knowledge inference
based on the subjective inference rules.
The acquired and inferred knowledge can be encoded into existing KBs to perform KB enrichment in the subjective dimension.
In summary, the contributions of our work are as follows:
Figure 2: An example of knowledge base.
the crowdsourced ST pair applying problem as a classification task and derive more knowledge facts based on the
crowdsourced seed knowledge.
• We conduct extensive experiments using real large-scale knowledge base and crowdsourcing platform and verify the effectiveness of CoSKA system.
The rest of the paper is organized as follows. In Section 2, we
introduce preliminaries and give the formal definitions of subjective
knowledge acquisition and enrichment. In Section 3, we present
the methodology for ST pair selection. In Section 4, we describe
the models for crowdsourced ST pair applying. The crowdsourcing
mechanism design is illustrated in Section 5. Section 6 shows the
experimental results on real KBs and crowdsourcing platform. The
related works are introduced in Section 7. We conclude our work
in Section 8.
2.
PROBLEM DEFINITION
A knowledge base is a repository of storing entities and relations
in a real world scenario. Similar with [16], knowledge base is formally defined as follows.
D EFINITION 1 (K NOWLEDGE BASE ). A knowledge base KB
is a tuple denoted by (E, L, R, P ), consisting of a collection of
entities E, literals L, relations R holding between entities, and
properties P holding between entities and literals. An entity e ∈ E
can be a class or an instance.
Figure 2 shows a toy example of a knowledge base. There are six
entities - three classes, e.g., “lawyer”, “politician”, and “president”,
and three instances, e.g., “Obama”, “Michelle ”, and “Gorge W.
Bush”; the date “1961-8-4” and string “Barack Obama” are literals;
there are three relations (“type”, “married”, and “subclassOf”) and
two kinds of properties (“birthDate” and “fullName”).
• We propose the problem of crowdsourced subjective knowledge acquisition and perform knowledge base enrichment in
the subjective dimension, which bridges the gap between the
subjective queries and existing knowledge bases encoding
only objective knowledge.
• We describe and implement our CoSKA system, consists of
ST pair selection, crowdsourced ST pair applying and knowledge inference, for crowd-powered subjective knowledge acquisition.
D EFINITION 2 (O BJECTIVE K NOWLEDGE ). Objective knowledge is a fact of triple < s, po , o >, where s is an entity in a knowledge base, p is an objective property, and o is either an entity or a
literal. Objective knowledge recording the real world facts, which
is factual and observable.
• We define subjective knowledge inference rules among ST
pairs and formulate the ST pair selection problem as a Maximum Knowledge Inference Problem. We prove the problem
is NP-hard and propose a diversity-aware forward greedy algorithm for ST pair selection.
As shown in the toy example of a KB of Figure 2, there are eight
objective knowledge facts, e.g. < Obama, birthDate, “1961-8-400 >
and < president, subclassOf, policitian >.
As mentioned in Section 1, the subjective knowledge refers to
the dominant opinion about whether a particular subjective property applies to entities of a particular type [25]. Therefore, the
combination of a subjective property and a certain type should be
• To further resolve the conflict between large scale knowledge
facts and the limited crowdsourcing resource, we formulate
2
figured out for subsequent subjective knowledge acquisition. We
define it as the ST pair (subjective property-type pair).
D EFINITION 3 (ST PAIR ). An ST pair consists of a subjective property and a type, which corresponds to a class entity in the
knowledge base, i.e. ST = (ps , T ). An ST pair for knowledge
acquisition task indicates that the subjective property ps can be
applied to the type T , namely, can be applied to the instances of
the type T .
Figure 3: ST Pair Extraction Pattern.
For example, an ST pair, ST = (big, City) indicates that the
entities of the City type have the subjective property of big. Same
for the ST pairs like (cute, Animal), (popular, Sport) and etc.
3.
ST PAIR SELECTION
In this section, we describe the ST pair selection problem concerning
the knowledge inference power. We first introduce the ST
D EFINITION 4 (S UBJECTIVE K NOWLEDGE ). A subjective knowlpair
extraction
method; then, we introduce the subjective resemble
edge fact is a triple denoted by < s, ST, l >, where s is an entity in
relationship among ST pairs and define the knowledge inference
a knowledge base, ST is an ST pair consists of a subjective proprules based on the relationship. We then formulate the ST pair
erty and a type, and l is a label with value either be true or f alse.
selection problem as a Maximum Knowledge Inference Problem
which is NP-hard and propose a diversity-aware forward greedy
The subjective knowledge has no ground truth, instead it has a
algorithm for ST pair selection.
dominant opinion which can be used to derive such knowledge.
For example, if most people hold the opinion that “New York is a
3.1 ST Pair Extraction
big city”, then we can derive a new subjective knowledge fact <
As described, the input of our system CoSKA is a set of ST pairs,
N ewY ork, (big, City), true >; otherwise, we will derive a new
and an ST pair consists of a subjective property which is usually an
subjective knowledge fact < N ewY ork, (big, City), f alse >.
“adjective” and a type which is usually a “noun phrase” and corD EFINITION 5 (ST PAIR A PPLYING ). Given an ST pair, ST =< responds to a type (class) entity in the KB. For example, a pair of
ps , T >, consists of a subjective property ps and a type T , and
(big, city) is an ST pair as the big is an adjective and city can be
a knowledge base KB, ST pair applying refers to the process of
mapped to a class entity in the knowledge base. In order to derive
deciding whether ps can be applied to the instances of type T
the commonly used ST pairs, we perform extraction from the news
in the KB, and the result is a list of subjective knowledge facts,
from New York Times. We use three years’ data which contains
F s = {F1 , F2 , · · · , Fm }, where Fi = {ei , ST, l}.
167,958 news and 582,898,171 sentences. We process the data using NLP tools to identify adjective and noun phrase pairs. Similar
Given a list of ST pairs, subjective knowledge acquisition refers
with work [25], we use the synthetic patterns to extract information,
to the process of performing ST pair applying for all the input ST
i.e. ST pairs, from matched sentences. The pattern we adopted
pairs. The derived knowledge can be encoded into the existing KB
is shown in Figure 3. For example, given a sentence of “Snakes
to perform KB enrichment in the subjective dimension.
are dangerous animals”, an ST pair of < dangerous, animal >
are extracted and the ST pair of < successf ul, f ilm > can be
D EFINITION 6 (S UBJECTIVE K NOWLEDGE E NRICHMENT ).
extracted from sentence “Titanic is the most successful film of all
Given a knowledge base KB consisting objective knowledge facts
time”.
O
F , a list of ST pairs, ST = {ST 1 , ST 2 , · · · , ST m }, the target
After extraction using the pattern, we map the type from the exis to enrich the KB with a list of subjective knowledge facts F S by
tracted pairs to the given knowledge bases through textual similarperforming subjective knowledge acquisition for the ST pairs.
ity and filter out pairs that have no mapped class entity. In total,
there are 40,582 mapped ST pairs with DBpedia.
We employ crowdsourcing for the subjective knowledge acquisition, specifically for the ST pair applying process. Due to the
3.2 ST Pair Selection
limited crowdsourcing resource, we need to crowdsource in an efIn order to reduce the number of ST pairs for crowdsourced subficient and productive manner. In other words, for crowdsourced
sequent
ST pair applying and identify the most productive ST pairs
subjective knowledge acquisition, our target is to maximize the acin terms of the knowledge inference power. We define the Subjecquired knowledge under the crowdsourcing budget. Next, we detive Resemble Relationship among ST pairs as follows:
fine the Crowdsourced Subjective Knowledge Acquisition problem.
D EFINITION
8 (ST PAIR S UBJECTIVE R ESEMBLE R ELATIONSHIP ).
D EFINITION 7 (C ROWDSOURCED S UBJECTIVE K NOWLEDGE ACQUISITION
).
Given two ST pairs , ST1 = (ps1 , T1 ), ST2 = (ps2 , T2 ), a knowlGiven a list of ST pairs, a knowledge base and a crowdsourcing
edge base KB and an object e, we define that ST1 and ST2 have
budget k (e.g. the number of crowdsourcing operation or monethe subjective resemble relationship on e, denoted as ST1 ≈e ST2
tary budget). The Crowdsourced Subjective Knowledge Acquisition
if the following condition satisfies:
(CoSKA) problem is to perform ST pair applying operations to acquire new subjective knowledge facts powered by crowdsourcing.
• Object e is an instance of both types , i.e., e ∈ Ikb (T1 ) ∧ e ∈
The target is to maximize the number of derived knowledge facts
Ikb (T2 ), where Ikb (T ) denotes the instances of type T in the
under the given budget k.
KB.
As described in Section 1, we propose a three stage approach for
• There exist a “subclassOf” relationship among two types in
CoSKA: ST pair selection, crowdsourced ST pair applying and knowlthe KB, denoted as < T1 , subclassOf, T2 >∈ F (KB)∨
edge inference. In next sections, we illustrate the details of each
stage.
< T2 , subclassOf, T1 >∈ F (KB).
3
• If the two subjective properties are the same, synonymous or
antonymous, denoted as ps1 ≈ ps2 . If two subjective properties are synonymous or same, we have ps1 ≈+ ps2 and
ps1 ≈− ps2 for antonymous.
• The weight of wij is the number of entities on which two corresponding ST pairs have the subjective resemble relationship, denoted as wij = |E| where, ∀e ∈ E, STi ≈e STj .
For the example shown in Example 1, we can have an ST graph
with four vertices and two edges: G = {{v1 , v2 , v3 , v4 }, {w12 , w34 }},
where vi corresponds to STi and w12 = 1, w34 = 2.
Our target of ST pair selection is to identify the most beneficial
ST pairs for subsequent crowdsourced ST pair applying to increase
the acquired subjective knowledge facts. In our work, based on the
ST pair subjective resemble relationship and knowledge inference
rules, we use the number of knowledge facts that can be inferred,
i.e. the inference power, to measure the beneficial of selected ST
pairs. Therefore, we formulate the ST pair selection problem as a
Maximum Knowledge Inference Problem.
Note that there are two kinds of subjective resemble relationships,
−
+
s
+
ST1 ≈+
e ST2 and ST1 ≈e ST2 , we have ST1 ≈e ST2 if p1 ≈
s
−
s
− s
p2 and ST1 ≈e ST2 if p1 ≈ p2 .
If two ST pairs have the subjective resemble relationship on an
entity, we can perform knowledge inference using the knowledge
inference rule:
L EMMA 1 (K NOWLEDGE I NFERENCE RULE ). If we have a
knowledge fact of F = {e, ST1 , l} and two ST pairs where ST1 =<
ps1 , T1 >, ST2 =< ps2 , T2 >:
0
• If ST1 ≈+
e ST2 , a new knowledge fact of F = {e, ST2 , l}
can be inferred
D EFINITION 10 (M AXIMUM K NOWLEDGE I NFERENCE P ROBLEM ).
Given a knowledge base KB and a set of ST pairs, P = {ST1 , ST2 , · · · , STm },
the target is to select k ST pairs to maximize the knowledge inference power.
0
0
• If ST1 ≈−
e ST2 , a new knowledge fact of F = {e, ST2 , l }
0
can be inferred, where l = ¬l
Based on the ST Graph Model defined in 9, the maximum knowledge acquisition problem is to select a set of nodes in the graph
that maximize the total edge weight induced by the nodes. We can
prove that the Maximum Knowledge Inference Problem is NP-hard
by a reduction Densest k-Subgraph problem.
Next, we illustrate the ST Pair Subjective Resemble Relationship
and the Knowledge Inference Rule through the following example.
E XAMPLE 1. Given a knowledge base KB, four ST pairs, i.e.,
ST1 =< old, P olitician >, ST2 =< young, P resident >
, ST3 =< big, City >, ST4 =< large, City > and four inT HEOREM 1. The Maximum Knowledge Inference Problem is
stances e1 ={“Hillary Clinton”}, e2 ={“Barack Obama”} , e3 ={“New NP-hard
York”} and e4 ={”Los Angeles”}. Referring to the knowledge in
P ROOF. Given an undirected graph G = (V, E), the Densest
KB, we have that: e1 is an instance of type “Politician”, e2 is an ink-Subgraph
(DkS) problem on G is the problem of finding a substance of type “Politician” and type “President”, and e3 , e4 are inset U ⊆ V of vertices of size k with the maximum induced averstances of the type “City”, denoted as T ype(e1 )={“Politician”},
T ype(e2 )={“Politician”,“President”}, T ype(e3 )=T ype(e4 )={“City”}.age degree. The average degree of the subgraph will be denoted
as 2|E(U )|/k. Here |E(U )| denotes the number of edges in the
Therefore, based on the Definition 8, we can have the following ST
subgraph induced by U . We construct the instance of ST graph as
pair subjective resemble relationships:
follows: Given |V | ST pairs, each corresponds to a vertex in G, two
1). ST1 ≈−
e2 ST2 , as “young” and “old” are antonymous of
ST pairs, STi , STj have the subjective resemble relationship on a
each other, e2 is an instance of both type “President” and
single object if there is an edge between to corresponding nodes,
“Politician” and “President” is a subclass of “Politician” ;
vi , vj in G. Given the parameter k, the maximum knowledge in∗
ference problem
+
P is to select k nodes S with maximum induced
2). ST3 ≈e3 ST4 , as “big” and “large” are synonymous and
edge weight, e∈E(S ∗ ) W (e) = |E(S ∗ )|. Therefore, under the
the type of two ST pairs is “City” and e3 is an instance of
same k, the optimal solution of the maximum knowledge acqui“City”;
sition problem is equivalent to that of Densest k-Subgraph prob3). Similarly, we also have ST3 ≈+
lem
e4 ST 4
Based on the ST pair subjective resemble relationships, we can
infer new knowledge facts using the inference rule according to
Lemma 1. If we have crowdsourced subjective knowledge facts
of Fcr = {< e2 , ST2 , Y ES >, < e3 , ST3 , Y ES >} (“Barack
Obama is a young president” and “New York is a big city”), we
can get Finf =< e2 , ST1 , N O > (“Barack Obama is NOT an old
politician”) and Finf =< e3 , ST4 , Y ES > (“New York is a large
city”).
The maximum knowledge acquisition problem is NP-hard, a backwardgreedy strategy, which repeatedly removes a vertex with the minimum weighted-degree in the remaining graph, until exactly k vern 2
) −
tices are left, has an worst case approximation ratio of [( 12 + 2k
−1
2
1
n
1
n
O(n 3 ), ( 2 + 2k ) + O( n )] for k in the range of [ 3 , n] and
[2( nk − 1) − O( k1 ), 2( nk − 1) + O( kn2 )] for k in the range of [0, n3 ),
where n is the number of vertex [1]. However, the backwardgreedy strategy is time consuming as it needs to iterate (|V | − k)
times; moreover, the strategy does not consider the knowledge diBased on the subjective resemble relationship and the inference
versity, i.e. knowledge about different types, when making decirule, we can construct a graph to model the ST pair subjective resions, and therefore may results in top ST pairs share the same
semble relationship and the knowledge inference power.
type. For example, in our experiment, we find that there are only
D EFINITION 9 (ST G RAPH M ODEL ). Given a knowledge base
two types from top 100 ST pairs by the backward-greedy stratKB = (E, L, R, P ) and a set of ST pairs P = {ST1 , ST2 , · · · , STn }, egy. In order to improve the efficiency and balance the subjective
we can construct a weighted graph G = {V, W}, where
knowledge over various types, we propose a diversity-aware forward greedy strategy for ST pair selection: each time we select the
• Each vertex in V corresponds to an ST pair of P.
pair with the maximum weight-degree, and add the pair to the result if the number of pairs with the same type does not exceed the
• There is an undirected edge wij between node vi (STi ) and
given threshold.
vj (STj ) if exists an instance e ∈ E(KB) and STi ≈e STj .
4
Algorithm 1: Diversity-aware Forward Greedy Selection
Input: ST Model Graph G = {V, W}, Parameter k and
threshold δ
Output: A set of vertices S
1 V ← ∅
2 while |V| ≤ k do
3
v ∗ ← arg maxv∈V W eightDegree(v)
4
T ← type(v ∗ )
5
if N um(V, T ) ≤ δ ∗ k then
6
V ← V ∪ v∗
7
8
V ← V\v ∗
Figure 4: Human Intelligent Task (HIT) interface.
return V
property can be applied. Furthermore, we list all the properties of
the given type in each HIT and let the crowd worker to select the
properties that would affect the decision. We compute the voting
for each feature, and retain those with voting number exceeds the
threshold (set through experimental studies) for further classification models. The HIT interface is shown in Figure 4. The illustrated
HIT is for ST pair (big, City), we include five instances of the type
“City” in each HIT, and ask the crowd to select the instances that
has the attribute of “big”. Also, there are properties related to instances of type “City” in the KB, e.g., Country, areaLand, foundingDate, e.t.c, we list the properties and let the crowd workers to
select relevant ones.
Agreement-based Answer Aggregation. After all the HITs are
answered, for each selected instance sample, we can collect a set
of yes or no answers of whether the given subjective property can
be applied to it. We compute the degree of the agreement on each
task:
The procedure of the diversity-aware forward greedy algorithm
is illustrated in Algorithm 1. There are k iterations (lines 2-7); in
each iteration, we first pick the vertex with the maximum weightdegree (line 3); next, we check the number of vertices of the same
type in the current result set, if the number dose not exceed the
threshold, the vertex is added into the result set (lines 4-6). The
complexity of Algorithm 1 is O(|V| · |W|).
4.
CROWDSOURCED ST PAIR APPLYING
For a given ST pair, the task of subjective knowledge acquisition is to identify the dominant opinion of whether the subjective
property can be applied to the instances of the type, referred as
crowdsourced ST pair applying. However, asking the crowd for every instance is too costly as a type in a KB could contain hundreds
of thousands instances. Therefore, we formulate the crowdsourced
ST pair applying as a binary classification problem taking advantage of the knowledge in the KBs. For each instance, we decide
whether the ST pair can be applied by the classification result.
However, we do not have any labeled data for training the classifier. Therefore, we select a set of seed instances and ask the crowd
to collect the corresponding subjective knowledge facts. We take
the crowdsourced samples as the training data and train the classifier using the features extracted from the KB. As each type in a
KB would have a set of properties/relations, we extract these properties/relations and list them in the crowdsourcing tasks, the crowd
workers are also asked to mark which of the properties would affect the decision about whether the ST pair applies to the instances.
Then, we filter out the properties/relations with votes less than a
threshold and training the classifier on the remaining properties.
We adopt a representative sampling to sample the instances which
explores the clustering structure of the large amount of unlabeled
data and query the representative samples, i.e. samples from different clusters, as the training data. In our work, we cluster the
instances using the knowledge from existing KBs and sample k instances for each ST pair.
5.
A(I, ST ) =
X
1
V (I, wi )
|W | w ∈W
(1)
i
where V (I, wi ) = 1 if the worker answer is yes and 0 otherwise.
The degree agreement of properties is computed in a similar way,
and we retain the properties with agreement score at least θP as our
subsequent classification features
The agreement score evaluates the confidence of the collected
opinion among workers over the random answer. With the given
threshold θA (which is set through experimental study, as the degree of agreement would vary with different ST pairs [25]), we
derive the dominant opinion, denoted as (DO(I, ST )) of whether
an ST pair ST applies to the instance I:
(
DO(I, ps ) =
yes
no
if A(I, ST ) − 0.5 ≥ θA
otherwise
(2)
According to Equation (2), we would obtain the positive opinion over the ST pair applies to an instance if the majority of the
opinions is positive and this positive opinion has a high agreement
(larger than θA ).
CROWDSOURCING MECHANISM DESIGN
6.
EXPERIMENTS
In this section, we evaluate CoSkA on real knowledge base and
crowdsourcing platform with extracted ST pairs. We describe the
experimental setup in Section 6.1. Section 6.2 compares different methods for ST selection; Section 6.3 verifies the proposed approaches for crowdsourced ST pair applying; Section 6.4 shows the
test results of the proposed knowledge inference approach.
In this section, we first describe the human intelligent task (HIT)
interface of CoSKA, then we introduce the answer aggregation strategy.
HIT Interface. Given an ST pair, we need to obtain the knowledge of whether a subjective property can be applied to the instances of the type. In order to reduce the cost, we design the task
as a multiple choice question where each question contains 5 instances and the crowd worker is asked to select those that the given
6.1
5
Experimental Setup
Table 1: Statistics of DBpedia
Dataset
#Facts
#Entities #Classes
DBpedia 26,797,299 2,531,369 827
Factor
θA
θP
Classifier
Knowledge Base. We adopt DBpedia, which contains millions
of knowledge facts (restricted to objective knowledge), classes (types)
and instances as the KB in our experiments. The KB is represented
as text files containing a list of triples of facts. The statistics DBpedia are given in Table 1. The KB offers information for mapping
extracted ST pair to the KB types and subjective knowledge inference.
Crowdsourcing Platform. We use the real crowdsorucing platform, Amazon Mechanical Turk (AMT) as the platform to conduct
the subjective knowledge acquisition tasks. As mentioned in Section 5, each HIT is designed as a multiple choice question, each
question is assigned to 5 workers and each worker would get a reward of $0.02 for answering the task. In addition, we would pay for
the AMT platform $0.01 for each assignment. In our experimental
settings, for each crowdsourced ST pair applying task, we would
sample up to 200 instances. As illustrated in Figure 4, we include
= 40 HITs
5 instances in each HIT, therefore there are totally 200
5
which cost $6.
6.2
ST Pairs
For crowdsourced ST pair applying, we need to ask the crowd
for a set of seed subjective knowledge facts and train the classifier
based on the collected samples and features. In our experiments,
we select 5 ST pairs through Div-FGreedy algorithm as test cases
to evaluate the accuracy of our approach 2 . There are following
configurations for the task: answer aggregation parameter θA , feature selection parameter θP and classification models, the settings
are illustrated in Table 2, where we mark our default settings in bold
font. As we have no ground truth, we use 5-fold cross validation to
test the performance of our approach.
Effect of Answer Aggregation Parameter. There are two parameters in terms of the answer aggregation: θA for opinion aggregation and θP for feature selection. We first fix θA = 0.1 (in
our settings, θA = 0.1 means at lease 60% workers select the instance to have the given property), and vary the value of θP from
0.1∼0.5 to compare the classification accuracy. The results are
shown in Figure 6 3 . From the results, we can observe that the classification accuracy would change with various θP value, the reason is that different θP have different filter power, i.e. with larger
value of θP , there would be less features remaining. Overall, the
θP = 0.3 achieves best performance: from 0.1∼0.3, there is an increasing trend of the accuracy for three pairs (old,Building), (experienced,Athlete) and (popular,film), and for pair (cute,Animal), the
accuracy does not have much difference; for larger values (0.3∼0.5),
the accuracy would remain approximately the same. The reason
is that the performance would change with different features (remained properties), and with lower value, less properties would
be filtered therefore might retain those irrelevant properties as the
training features.
Next, we check the effect of θA on the classification accuracy.
We set the θP to 0.3 and vary the values of θA according to Table 2,
the classification accuracy results are illustrated in Figure 7. From
the results we can observe that the classification models achieve
best performance with the value of θA equals to 0.1, and would decrease as the value of θA increases. The reason is that with larger
θA , we have stronger restriction for deriving the dominant opinion
of whether the property applies to an instance. For example, in
our settings, when θ = 0.5, we would only obtain the opinion that
the property applies to the instance if all the 5 workers gives the
“yes” answer, which might results in missing some positive training samples and affect the classification performance. Overall, the
classification models achieve best performance with the value of
θA to 0.1. Therefore, we set the default value of θA to 0.1.
Effect of Classification Model. From the Figures 6 and 7,
we can find that the performance of different classification models varies with different ST pairs. For pair (cute,Animal) different
ST Pair Selection
As illustrated in the Section 3, we adopt a diversity-based forward greedy algorithm in our work for ST pair selection (DivFGreedy). To evaluate the efficiency and the effectiveness of the
propose algorithm, we use three metrics: 1). Induced Edge Weight
of ST pairs, which indicates the inference power of selected ST
pairs; 2). The number of different types of the ST pairs, which is
used to evaluate the knowledge diversity; 3). Running time, which
is recorded to demonstrate the efficiency of the algorithm. For comparison, we implement three other algorithms: backward greedy
selection algorithm (BGreedy), forward greedy selection algorithm
(FGreedy) and random selection algorithm (Random). We vary the
number of selected ST pairs from 10∼100, and fix the threshold
for the diversity-based forward greedy algorithm (Div-FGreedy) to
0.1 (the value of δ can be changed to satisfy the various diversity
demand as the Div-FGreedy strategy can derive ST pairs with at
least θ1 types), the results are shown in Figure 5.
From Figure 5(a), we can find that the Random algorithm cannot achieve a good result, and the FGreedy and the Div-FGreedy
algorithm outperform the BGreedy algorithm. We can observe that
in our experiments, the FGreedy strategy has the best performance
in terms of the inference power (induced edge weight). However,
from Figure 5(b), we can see that the FGreedy algorithm would favor pairs with the same type as it does not consider the knowledge
diversity, e.g. for the FGreedy algorithm, there would be only 5
types out of 100 selected ST pairs. The BGreedy strategy also has
the same problem, e.g. the BGreedy only has 2 types out of 100
selected ST pairs. For the Div-FGreedy strategy, we have 16 types
out of 100 selected ST pairs with the threshold set to 0.1. We can
see that the Random strategy can select pairs with larger type numbers as it selects ST pairs randomly. However, it does not consider
the inference power when selecting ST pairs and thus results in ST
pairs with quite low inference power as shown in Figure 5(a). For
the running time shown in Figure 5(c), we can observe that except
for the BGreedy algorithm, all other three algorithms are quite efficient. To conclude, considering all three evaluation metrics, the
Div-FGreedy algorithm can achieve a good inference power, guarantee the ST pair type diversity and is quite efficient.
6.3
Table 2: Parameter Setting
Setting
0.1, 0.3, 0.5
0.1, 0.2, 0.3, 0.4, 0.5
AdaBoost (AD), Decision Tree (DT), RBF-SVM,
Nearest Neighbors (NN), Random Forest (RF)
(big,City), (experienced, Athlete), (cute,Animal),
(old,Building), (popular,Film)
2
Note that the workflow of crowdsourced ST pair applying is same
for each ST pair, to acquire more knowledge facts, we can perform
crowdsourced ST pair applying for a larger number of ST pairs
3
Note that due to the space limit, we only show the results for four
pairs and the result of (big,City) pair is summarized in Table 5
Crowdsourced ST Pair Applying
6
80
BGreedy
Random
Div-FGreedy
FGreedy
1
0.5
0
60
Number of ST Pairs
(a) Induced Edge Weight
BGreedy
Random
Div-FGreedy
FGreedy
40
20
0
10 20 30 40 50 60 70 80 90 100
2.5
RunningTime(ms)
1.5
×108
Number of Types
Induced Edge Weight
2
×105
2
1.5
1
0.5
0
10 20 30 40 50 60 70 80 90 100
BGreedy
Random
Div-FGreedy
FGreedy
10 20 30 40 50 60 70 80 90 100
Number of ST Pairs
Number of ST Pairs
(b) Number of Types
(c) Running Time
Figure 5: ST pair Selection Comparison.
Table 3: Crowdsourced ST Pair Applying Performance
ST Pair
Classification Model Accuracy
(big,City)
DT
0.925
(experienced, Athlete)
RBF-SVM
0.80
(cute,Animal)
RBF-SVM/RF/DT
0.685
(old,Building)
RBF-SVM
0.89
(popular,Film)
DT
0.79
Table 4: Subjective Resemble Relationship
ST Pair
(big,City)
(experienced, Athlete)
(cute,Animal)
(old,Building)
(popular,Film)
Table 5: Knowledge Inference Performance
models have similar performance; for pairs (old,Building) and (experienced,Athlete) the RBF-SVM achieves the best performance
whereas for (popular,film), the DT model outperforms other approaches. We summarize the results of crowdsourced ST pair applying with the five pairs with all the parameters setting to the default value, the results are presented in Table 3. Overall, we can find
that the RBF-SVM and DT model can achieve good performance
for different ST pairs.
To justify the effectiveness of our approach for subjective knowledge acquisition, we compare our results with the state-of-the-art
technique, Surveyor, proposed in [25]. The reported accuracy of
the approach by Surveyor is 77%. Compared our results with the
Surveyor approach, we can observe that except for the ST pair
(cute,Animal), our approach can achieve better results, i.e. the pair
(popular,Film) achieves accuracy of 79% and all the other three
pairs can achieve the accuracy of over 80%. Therefore, our approach can perform accurate and scalable subjective knowledge acquisition with a low crowdsourcing budget (with up to 40 HITs and
$6 for each ST pair).
6.4
Subjective Resemble Relationship Pairs
(small,City) , (big, Settlement) , (large,Place)
(experienced, SoccerPlayer) , (trained, Boxer)
(lovely, Animal) , (lovely,Species)
(old, Hotel) , (new, Museum)
(popular, Work) , (neglected, Film)
ST Pair
(big,City)
(experienced, Athlete)
(cute,Animal)
(old,Building)
(popular,Film)
#Seed Facts
10,354
499
4,096
233
272
#Inferred Facts
93,186
3,488
12,284
1,398
1,632
Accuracy
0.92
0.83
0.76
0.93
0.87
with those presented in Table 3, we can observe that the inferred
knowledge of each ST pair has close but higher accuracy than the
facts acquired by crowdsourced ST pair applying, which confirms
that our proposed knowledge inference process does not introduce
significant noises and verifies the high quality of our knowledge inference rules. To conclude, our knowledge inference approach can
help to derive more high quality knowledge facts compared with
that only using the crowd and classification models in the crowdsourced ST pair applying process.
7.
Knowledge Inference
RELATED WORK
In this section, we discuss the works related to subjective knowledge acquisition, knowledge base enrichment and crowdsourcing.
Subjective knowledge acquisition is closely related to works that
associating properties with entities. Some works have been conducted for commonsense knowledge acquisition [18] [15] [14] [24].
WebChild [23] presents a method for automatically constructing a
large commonsense knowledge base, it contains triples that connect
nouns with adjectives via fine-grained relations. Entitytagger [5],
presented by Chakrabarti et al., automatically associate descriptive
phrases, referred to as etags (entity tags), to each entity. Instead
of subjective properties, these works focus on the less controversial and more objective properties, which is not related to obtaining
dominant opinion. The most similar work is SURVEYOR, which
mines the dominant opinion on the web content of whether a subjective property applies to a type. However, they does not consider
to use the existing information in knowledge base and resorting to
the crowd for subjective knowledge acquisition.
After the crowdsourced ST pair applying process, we have acquired a set of subjective knowledge facts, either collected from
the crowd or obtained through the classification model. Then, we
can perform knowledge inference to acquire more knowledge facts.
Some resemble relationship of the selected ST pairs are shown in
Table 4. In order to verify the effectiveness of our knowledge inference approach, we evaluate two metrics: the number of inferred
facts and the accuracy of inferred facts. As we do not have the
ground truth, we sample 100 facts for each pair and verify the correctness manually. We ask three students to label whether the fact is
correct or not and derive the answer by majority voting. The results
of the knowledge inference performance are shown in Table 5.
From the results, we can observe that, on the one hand, large
amount of knowledge facts could be inferred through the subjective knowledge inference approach; on the other hand, the inferred
knowledge facts have high accuracy. Compare the accuracy results
7
0.85
0.95
0.76
0.8
0.74
0.8
0.9
0.68
0.66
AD
DT
RBF-SVM
NN
RF
0.64
0.62
0.6
0.1
0.2
0.3
0.4
0.85
0.8
AD
DT
RBF-SVM
NN
RF
0.75
0.7
0.1
0.5
0.2
0.3
0.4
0.75
0.7
0.65
0.1
0.5
(a) Pair (cute,Animal)
0.2
0.3
0.4
0.75
0.7
(b) Pair (old,Building)
AD
DT
RBF-SVM
NN
RF
0.65
0.6
0.1
0.5
0.2
0.3
0.4
0.5
Value of θP
Value of θP
Value of θP
Value of θP
AD
DT
RBF-SVM
NN
RF
Accuracy
0.7
Accuracy
Accuracy
Accuracy
0.72
(c) Pair (experienced,Athlete)
(d) Pair (popular,Film)
Figure 6: Results on varying Feature Selection Parameter (θP ).
0.9
0.5
0.45
0.85
0.9
AD
DT
RBF-SVM
NN
RF
0.85
AD
DT
RBF-SVM
NN
RF
0.85
0.8
0.8
0.75
0.75
0.4
AD
DT
RBF-SVM
NN
RF
0.8
Accuracy
Accuracy
0.6
0.55
Accuracy
AD
DT
RBF-SVM
NN
RF
Accuracy
0.7
0.65
0.75
0.7
0.7
0.35
0.3
0.1
0.3
Value of θA
(a) Pair (cute,Animal)
0.5
0.7
0.1
0.3
0.65
0.1
0.5
Value of θA
0.3
0.5
Value of θA
(b) Pair (old,Building)
(c) Pair (experienced,Athlete)
0.65
0.1
0.3
0.5
Value of θA
(d) Pair (popular,Film)
Figure 7: Results on varying Opinion Aggregation Parameter (θA ).
among ST pairs and perform knowledge inference to derive more
knowledge facts. We formulate the ST pair selection problem as a
Maximum Knowledge Inference Problem which is NP-hard and we
propose a diversity-aware forward greedy algorithm for ST pair selection. The crowdsourced ST pair applying problem is formulated
as a classification task to further improve the system scalability.
Experimental results on real knowledge base and crowdsourcing
platform verify that our system, CoSKA, could derive large amount
accurate subjective knowledge facts with a comparative low crowdsourcing cost.
Knowledge base enrichment, completion and population have
been widely studied. There are two mainstreams: internal methods,
which use only the knowledge contained in the knowledge base to
predict missing information [27] [9]; external methods,which use
sources of knowledge such as text corpora or other knowledge base
to add new knowledge facts [30] [21] [12] [11]. However, these
works are limited to add objective knowledge and neglect subjective knowledge. Moreover, they do not consider to make use of
a natural source of knowledge, the crowd, to complete/enrich the
existing knowledge base.
Recently, the increasing popularity of crowdsourcing brings new
trend to leverage the power of the crowd in knowledge acquisition, data integration and many other applications. Kondreddi et
al. [13] proposes a hybrid approach that combines information extraction technique with human computation for knowledge acquisition. Marta et al. [22] presents a hybrid-genre workflow for games
in crowdsourced knowledge acquisition process. Works [6] [3] [20]
present approaches that use the wisdom of crowd to perform taxonomy construction. Crowdsourcing also proved to have good performance in applications such as entity resolution [28] [26], schema
matching [33], translation [32] and so forth.
9.
REFERENCES
[1] Y. Asahiro, K. Iwama, H. Tamaki, and T. Tokuyama.
Greedily finding a dense subgraph. Journal of Algorithms,
34(2):203–221, 2000.
[2] K. D. Bollacker, C. Evans, P. Paritosh, T. Sturge, and
J. Taylor. Freebase: a collaboratively created graph database
for structuring human knowledge. In SIGMOD, pages
1247–1250, 2008.
[3] J. Bragg, Mausam, and D. S. Weld. Crowdsourcing
multi-label classification for taxonomy creation. In
8. CONCLUSION
Proceedings of the First AAAI Conference on Human
Computation and Crowdsourcing, HCOMP 2013, November
In our work, we propose a system Crowdsourced subjective knowledge
7-9, 2013, Palm Springs, CA, USA, 2013.
acquisition (CoSKA), for subjective knowledge acquisition pow[4]
L. Bühmann and J. Lehmann. Pattern based knowledge base
ered by crowdsourcing and existing KBs. The acquired knowledge
enrichment. In The Semantic Web - ISWC 2013 - 12th
can be encoded into existing KBs to perform KB enrichment in
International Semantic Web Conference, Sydney, NSW,
the subjective dimension which can bridge the gap between existAustralia, October 21-25, 2013, Proceedings, Part I, pages
ing objective knowledge and the subjective queries. Our CoSKA
33–48, 2013.
system, consists of three stages: ST pair selection, Crowdsourced
ST pair applying and knowledge inference. To resolve the con[5] K. Chakrabarti, S. Chaudhuri, T. Cheng, and D. Xin.
flict between large scale knowledge facts and the limited crowdEntitytagger: automatically tagging entities with descriptive
sourcing resource, we define subjective knowledge inference rules
phrases. In Proceedings of the 20th International Conference
8
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
on World Wide Web, WWW 2011, Hyderabad, India, March
28 - April 1, 2011 (Companion Volume), pages 19–20, 2011.
L. B. Chilton, G. Little, D. Edge, D. S. Weld, and J. A.
Landay. Cascade: crowdsourcing taxonomy creation. In
2013 ACM SIGCHI Conference on Human Factors in
Computing Systems, CHI ’13, Paris, France, April 27 - May
2, 2013, pages 1999–2008, 2013.
M. Choy, J. Lee, G. Gweon, and D. Kim. Glaucus:
Exploiting the wisdom of crowds for location-based queries
in mobile environments. In Proceedings of the Eighth
International Conference on Weblogs and Social Media,
ICWSM 2014, Ann Arbor, Michigan, USA, June 1-4, 2014.,
2014.
O. Etzioni, M. J. Cafarella, D. Downey, S. Kok, A. Popescu,
T. Shaked, S. Soderland, D. S. Weld, and A. Yates.
Web-scale information extraction in knowitall: (preliminary
results). In WWW, pages 100–110, 2004.
L. A. Galárraga, C. Teflioudi, K. Hose, and F. M. Suchanek.
AMIE: association rule mining under incomplete evidence in
ontological knowledge bases. In 22nd International World
Wide Web Conference, WWW ’13, Rio de Janeiro, Brazil,
May 13-17, 2013, pages 413–422, 2013.
J. Hoffart, F. M. Suchanek, K. Berberich, and G. Weikum.
YAGO2: A spatially and temporally enhanced knowledge
base from wikipedia. Artif. Intell., 194:28–61, 2013.
H. Ji, T. Cassidy, Q. Li, and S. Tamang. Tackling
representation, annotation and classification challenges for
temporal knowledge base population. Knowl. Inf. Syst.,
41(3):611–646, 2014.
H. Ji and R. Grishman. Knowledge base population:
Successful approaches and challenges. In The 49th Annual
Meeting of the Association for Computational Linguistics:
Human Language Technologies, Proceedings of the
Conference, 19-24 June, 2011, Portland, Oregon, USA, pages
1148–1158, 2011.
S. K. Kondreddi, P. Triantafillou, and G. Weikum.
Combining information extraction and human computing for
crowdsourced knowledge acquisition. In ICDE, pages
988–999, 2014.
Y.-L. Kuo, J. Hsu, and F. Shih. Contextual commonsense
knowledge acquisition from social content by
crowd-sourcing explanations. In Proceedings of the Fourth
AAAI Workshop on Human Computation, pages 18–24, 2012.
Y.-L. Kuo and J. Y.-j. Hsu. Resource-bounded
crowd-sourcing of commonsense knowledge. In IJCAI
Proceedings-International Joint Conference on Artificial
Intelligence, volume 22, page 2470, 2011.
S. Lacoste-Julien, K. Palla, A. Davies, G. Kasneci,
T. Graepel, and Z. Ghahramani. Sigma: simple greedy
matching for aligning large knowledge bases. In The 19th
ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD 2013, Chicago, IL, USA,
August 11-14, 2013, pages 572–580, 2013.
J. Lehmann, R. Isele, M. Jakob, A. Jentzsch, D. Kontokostas,
P. N. Mendes, S. Hellmann, M. Morsey, P. van Kleef,
S. Auer, and C. Bizer. Dbpedia - A large-scale, multilingual
knowledge base extracted from wikipedia. Semantic Web,
pages 167–195, 2015.
H. Liu and P. Singh. Conceptneta practical commonsense
reasoning tool-kit. BT technology journal, 22(4):211–226,
2004.
J. McAuley and A. Yang. Addressing complex and subjective
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
9
product-related queries with customer reviews. In
Proceedings of the 25th International Conference on World
Wide Web, WWW 2016, Montreal, Canada, April 11 - 15,
2016, pages 625–635, 2016.
R. Meng, Y. Tong, L. Chen, and C. C. Cao. CrowdTC:
Crowdsourced taxonomy construction. In 2015 IEEE
International Conference on Data Mining, ICDM 2015,
Atlantic City, NJ, USA, November 14-17, 2015, pages
913–918, 2015.
A. G. Nuzzolese, A. Gangemi, V. Presutti, and P. Ciancarini.
Type inference through the analysis of wikipedia links. In
WWW2012 Workshop on Linked Data on the Web, Lyon,
France, 16 April, 2012, 2012.
M. Sabou, A. Scharl, and M. Föls. Crowdsourced knowledge
acquisition: Towards hybrid-genre workflows. Int. J.
Semantic Web Inf. Syst., 9(3):14–41, 2013.
N. Tandon, G. de Melo, F. M. Suchanek, and G. Weikum.
Webchild: harvesting and organizing commonsense
knowledge from the web. In Seventh ACM International
Conference on Web Search and Data Mining, WSDM 2014,
New York, NY, USA, February 24-28, 2014, pages 523–532,
2014.
N. Tandon, G. de Melo, and G. Weikum. Acquiring
comparative commonsense knowledge from the web. In
Proceedings of the Twenty-Eighth AAAI Conference on
Artificial Intelligence, July 27 -31, 2014, Québec City,
Québec, Canada., pages 166–172, 2014.
I. Trummer, A. Y. Halevy, H. Lee, S. Sarawagi, and
R. Gupta. Mining subjective properties on the web. In
SIGMOD, pages 1745–1760, 2015.
N. Vesdapunt, K. Bellare, and N. N. Dalvi. Crowdsourcing
algorithms for entity resolution. PVLDB, 7(12):1071–1082,
2014.
J. Völker and M. Niepert. Statistical schema induction. In
The Semantic Web: Research and Applications - 8th
Extended Semantic Web Conference, ESWC 2011, Heraklion,
Crete, Greece, May 29-June 2, 2011, Proceedings, Part I,
pages 124–138, 2011.
J. Wang, T. Kraska, M. J. Franklin, and J. Feng. CrowdER:
Crowdsourcing entity resolution. PVLDB, 5(11):1483–1494,
2012.
R. West, E. Gabrilovich, K. Murphy, S. Sun, R. Gupta, and
D. Lin. Knowledge base completion via search-based
question answering. In 23rd International World Wide Web
Conference, WWW ’14, Seoul, Republic of Korea, April 7-11,
2014, pages 515–526, 2014.
R. West, E. Gabrilovich, K. Murphy, S. Sun, R. Gupta, and
D. Lin. Knowledge base completion via search-based
question answering. In 23rd International World Wide Web
Conference, WWW ’14, Seoul, Republic of Korea, April 7-11,
2014, pages 515–526, 2014.
O. Zaidan and C. Callison-Burch. Crowdsourcing
translation: Professional quality from non-professionals. In
ACL, pages 1220–1229, 2011.
O. F. Zaidan and C. Callison-Burch. Crowdsourcing
translation: Professional quality from non-professionals. In
Proceedings of the 49th Annual Meeting of the Association
for Computational Linguistics: Human Language
Technologies-Volume 1, pages 1220–1229, 2011.
C. J. Zhang, L. Chen, H. V. Jagadish, and C. C. Cao.
Reducing uncertainty of schema matching via
crowdsourcing. PVLDB, 6(9):757–768, 2013.
| 2cs.AI
|
Trace-Based Run-Time Analysis of
Message-Passing Go Programs
arXiv:1709.01588v4 [cs.PL] 30 Jan 2018
Martin Sulzmann and Kai Stadtmüller
Faculty of Computer Science and Business Information Systems
Karlsruhe University of Applied Sciences
Moltkestrasse 30, 76133 Karlsruhe, Germany
[email protected]
[email protected]
Abstract. We consider the task of analyzing message-passing programs
by observing their run-time behavior. We introduce a purely librarybased instrumentation method to trace communication events during
execution. A model of the dependencies among events can be constructed
to identify potential bugs. Compared to the vector clock method, our
approach is much simpler and has in general a significant lower runtime overhead. A further advantage is that we also trace events that
could not commit. Thus, we can infer more alternative communications.
This provides the user with additional information to identify potential
bugs. We have fully implemented our approach in the Go programming
language and provide a number of examples to substantiate our claims.
1
Introduction
We consider run-time analysis of programs that employ message-passing. Specifically, we consider the Go programming language [4] which integrates messagepassing in the style of Communicating Sequential Processes (CSP) [6] into a C
style language. We assume the program is instrumented to trace communication
events that took place during program execution. Our objective is to analyze
program traces to assist the user in identifying potential concurrency bugs.
Motivating Example In Listing 1.1 we find a Go program implementing a system
of newsreaders. The main function creates two synchronous channels, one for each
news agency. Go supports (a limited form of) type inference and therefore no
type annotations are required. Next, we create one thread per news agency via
the keyword go. Each news agency transmits news over its own channel. In Go,
we write ch <- "REUTERS" to send value "REUTERS" via channel ch. We write
<-ch to receive a value via channel ch. As we assume synchronous channels,
both operations block and only unblock once a sender finds a matching receiver.
We find two newsreader instances. Each newsreader creates two helper threads
that wait for news to arrive and transfer any news that has arrived to a common
channel. The intention is that the newsreader wishes to receive any news whether
it be from Reuters or Bloomberg. However, there is a subtle bug (to be explained
shortly).
func reuters ( ch chan string ) { ch <- " REUTERS " } // r !
func bloomberg ( ch chan string ) { ch <- " BLOOMBERG " } // b !
func newsReader ( rCh chan string , bCh chan string ) {
ch := make ( chan string )
go func () { ch <- ( < - rCh ) }()
// r ?; ch !
go func () { ch <- ( < - bCh ) }()
// b ?; ch !
x := <- ch
// ch ?
}
func main () {
reutersCh := make ( chan string )
bloombergCh := make ( chan string )
go reuters ( reutersCh )
go bloomberg ( bloombergCh )
go newsReader ( reutersCh , bloombergCh ) // N1
newsReader ( reutersCh , bloombergCh )
// N2
}
Listing 1.1. Message passing in Go
Trace-Based Run-Time Verification We only consider finite program runs and
therefore each of the news agencies supplies only a finite number of news (exactly
one in our case) and then terminates. During program execution, we trace communication events, e.g. send and receive, that took place. Due to concurrency, a
bug may not manifest itself because a certain ‘bad’ schedule is rarely taken in
practice.
Here is a possible trace resulting from a ‘good’ program run.
r!; N1.r?; N1.ch!; N1.ch?; b!; N2.b?; N2.ch!; N2.ch?
We write r! to denote that a send event via the Reuters channel took place.
As there are two instances of the newsReader function, we write N1.r? to denote that a receive event via the local channel took place in case of the first
newsReader call. From the trace we can conclude that the Reuters news was
consumed by the first newsreader and the Bloomberg news by the second newsreader.
Here is a trace resulting from a bad program run.
r!; b!; N1.r?; N1.b?; N1.ch!; N1.ch?; DEADLOCK
The helper thread of the first newsreader receives the Reuters and the Bloomberg
news. However, only one of these messages will actually be read (consumed). This
is the bug! Hence, the second newsreader gets stuck and we encounter a deadlock.
The issue is that such a bad program run may rarely show up. So, the question is
how can we assist the user based on the trace information resulting from a good
program run? How can we infer that alternative schedules and communications
may exist?
2
Event Order via Vector Clock Method A well-established approach is to derive
a partial order among events. This is usually achieved via a vector of (logical)
clocks. The vector clock method was independently developed by Fidge [1] and
Mattern [8]. For the above good program run, we obtain the following partial
order among events.
r! < N1.r?
N1.r? < N1.ch!
N1.ch! < N1.ch?
b! < N2.b?
N2.b? < N2.ch! (1)
N2.ch! < N2.ch? (2)
For example, (1) arises because N2.ch! happens (sequentially) after N2.b? For
synchronous send/receive, we assume that receive happens after send. See (2).
Based on the partial order, we can conclude that alternative schedules are possible. For example, b! could take place before r!. However, it is not clear how
to infer alternative communications. Recall that the issue is that one of the
newsreaders may consume both news messages. Our proposed method is able
to clearly identify this issue and has the advantage to require a much simpler
instrumentation We discuss these points shortly. First, we take a closer look at
the details of instrumentation for the vector clock method.
Vector clocks are a refinement of Lamport’s time stamps [7]. Each thread
maintains a vector of (logical) clocks of all participating partner threads. For
each communication step, we advance and synchronize clocks. In pseudo code,
the vector clock instrumentation for event sndR.
vc [ r e u t e r s T h r e a d ]++
ch <- ( " R E U T E R S" , vc , vcCh )
vc ’ := max ( vc , <- vcCh )
We assume that vc holds the vector clock. The clock of the Reuters thread is
incremented. Besides the original value, we transmit the sender’s vector clock
and a helper channel vcCh. For convenience, we use tuple notation. The sender’s
vector clock is updated by building the maximum among all entries of its own
vector clock and the vector clock of the receiving party. The same vector clock
update is carried out on the receiver side.
Our Method We propose a much simpler instrumentation and tracing method
to obtain a partial order among events. Instead of a vector clock, each thread
traces the events that might happen and have happened. We refer to them as
pre and post events. In pseudo code, our instrumentation for sndR looks like
follows.
pre ( hash ( ch ) , " ! " )
ch <- ( " R E U T E R S" , t h r e a d I d)
post ( hash ( ch ) , " ! " )
The bang symbol (‘!’) indicates a send operation. Function hash builds a hash
index of channel names. The sender transmits its thread id number to the receiver. This is the only intra-thread overhead. No extra communication link is
necessary.
3
Here are the traces for individual threads resulting from the above good
program run.
R:
N1_helper1:
N1_helper2:
N1:
B:
N2_helper1:
N2_helper2:
N2:
pre(r!); post(r!)
pre(r?); post(R#r?); pre(ch1!); post(ch1!)
pre(b?)
pre(ch1?); post(N1_helper1#ch1?)
pre(b!); post(b!)
pre(r?)
pre(b?); post(B#b?); pre(ch2!); post(ch2!)
pre(ch2?); post(N2_helper2#ch2?)
We write pre(r!) to indicate that a send via the Reuters channel might happen.
We write post(R#r?) to indicate that a receive has happened via thread R. The
partial order among events is obtained by a simple post-processing phase where
we linearly scan through traces. For example, within a trace there is a strict
order and therefore
N2_helper2:
pre(b?); post(B#b?); pre(ch2!); post(ch2!)
implies N2.b? < N2.ch!. Across threads we check for matching pre/post events.
Hence,
R:
N1_helper1:
pre(r!); post(r!)
pre(r?); post(R#r?); ...
implies r! < N1.r?. So, we obtain the same (partial order) information as the
vector clock approach but with less overhead.
The reduction in terms of tracing overhead compared to the vector clock
method is rather drastic assuming a library-based tracing scheme with no access
to the Go run-time system. For each communication event we must exchange
vector clocks, i.e. n additional (time stamp) values need to be transmitted where
n is the number of threads. Besides extra data to be transmitted, we also require
an extra communication link because the sender requires the receivers vector
clock. In contrast, our method incurs a constant tracing overhead. Each sender
transmits in addition its thread id. No extra communication link is necessary.
This results in much less run-time overhead as we will see later.
The vector clock tracing method can be improved assuming we extend the
Go run-time system. For example, by maintaining a per-thread vector clock
and having the run-time system carrying out the exchange of vector clocks for
each send/receive communication. There is still the O(n) space overhead. Our
method does not require any extension of the Go run-time system to be efficient
and therefore is also applicable to other languages that offer similar features as
found in Go.
A further advantage of our method is that we also trace (via pre) events that
could not commit (post is missing). Thus, we can easily infer alternative communications. For example, for R: pre(r!); ... there is the alternative match
4
N2_helper1: pre(r?). Hence, instead of r! < N1.r? also r! < N2.r? is possible. This indicates that one newsreader may consume both news message. The
vector clock method, only traces events that could commit, post events in our
notation. Hence, the above alternative communication could not be derived.
Contributions Compared to earlier works based on the vector clock method, we
propose a much more light-weight and more informative instrumentation and
tracing scheme. Specifically, we make the following contributions:
– We give a precise account of our run-time tracing method (Section 3) for
message-passing as found in the Go programming language (Section 2) where
for space reasons we only formalize the case of synchronous channels and
selective communications.
– A simple analysis of the resulting traces allows us to detect alternative schedules and communications (Section 4). For efficiency reasons, we employ a directed dependency graph to represent happens-before relations (Section 4.1).
– We show that vector clocks can be easily recovered based on our tracing
method (Section 5). We also discuss the pros and cons of both methods for
analysis purposes.
– Our tracing method can be implemented efficiently as a library. We have fully
implemented the approach supporting all Go language features dealing with
message-passing such as buffered channels, select with default or timeout
and closing of channels (Section 6).
– We provide experimental results measuring the often significantly lower overhead of our method compared to the vector clock method assuming based
methods are implemented as libraries (Section 6.2).
The online version of this paper contains an appendix with further details.1
2
Message-Passing Go
Syntax For brevity, we consider a much simplified fragment of the Go programming language. We only cover straight-line code, i.e. omitting procedures, if-thenelse etc. This is not an onerous restriction as we only consider finite program
runs. Hence, any (finite) program run can be represented as a program consisting
of straight-line code only.
Definition 1 (Program Syntax).
x, y, . . .
i, j, . . .
b
::= x | i | hash(x) | head(b) | last(b) | bs | tid
bs
::= [] | b : bs
e, f ::= x ← b | y :=← x
c
::= y := b | y := makeChan | go p | select [ei ⇒ pi ]i∈I
p, q, r ::= [] | c : p
1
https://arxiv.org/abs/1709.01588
5
Variables, Channel Names
Integers
Expressions
Transmit/Receive
Commands
Program
For our purposes, values are integers or lists (slices in Go terminology). For
lists we follow Haskell style notation and write b : bs to refer to a list with head
element b and tail bs. We can access the head and last element in a list via
primitives head and last. We often write [b1 , . . . , bn ] as a shorthand b1 : · · · : [].
Primitive tid yields the thread id number of the current thread. We assume that
the main thread always has thread id number 1 and new thread id numbers are
generated in increasing order. Primitive hash() yields a unique hash index for
each variable name. Both primitives show up in our instrumentation.
A program is a sequence of commands where commands are stored in a list.
Primitive makeChan creates a new synchronous channel. Primitive go creates a
new go routine (thread). For send and receive over a channel we follow Go notation. We assume that a receive is always tied to an assignment. For assignment
we use symbol := to avoid confusion with the mathematical equality symbol =.
In Go, symbol := declares a new variable with some initial value. We also use
:= to overwrite the value of existing variables. As a message passing command
we only support selective communication via select. Thus, we can fix the bug in
our newsreader example.
func n e w s R e a d e r F i x e d ( rCh chan string , bCh chan string ) {
ch := make ( chan string )
select {
case x := <- rCh :
case x := <- bCh :
}
}
The select statement guarantees that at most one news message will be consumed
and blocks if no news are available. In our simplified language, we assume that
the x ← b command is a shorthand for select [x ← b ⇒ []]. For space reasons, we
omit buffered channels, select paired with a default/timeout case and closing of
channels. All three features are fully supported by our implementation.
Trace-Based Semantics The semantics of programs is defined via a small-step
operational semantics. The semantics keeps track of the trace of channel-based
communications that took place. This allows us to relate the traces obtained by
our instrumentation with the actual run-time traces.
We support multi-threading via a reduction relation
T
(S, [i1 ♯p1 , . . . , in ♯pn ]) =
⇒ (S ′ , [j1 ♯q1 , . . . , jn ♯qn ]).
We write i♯p to denote a program p that runs in its own thread with thread id i.
We use lists to store the set of program threads. The state of program variables,
before and after execution, is recorded in S and S ′ . We assume that threads
share the same state. Program trace T records the sequence of communications
that took place during execution. We write x! to denote a send operation on
channel x and x? to denote a receiver operation on channel x. The semantics
of expressions is defined in terms a big-step semantics. We employ a reduction
6
relation (i, S) ⊢ b ⇓ v where S is the current state, b the expression and v the
result of evaluating b. The formal details follow.
Definition 2 (State).
v ::= x | i | [] | vs
Values
vs ::= [] | v : vs
s ::= v | Chan
Storables
S ::= () | (x 7→ s) | S ⊳ S State
A state S is either empty, a mapping, or an override of two states. Each state
maps variables to storables. A storable is either a plain value or a channel. Variable names may appear as values. In an actual implementation, we would identify
the variable name by a unique hash index. We assume that mappings in the right
operand of the map override operator ⊳ take precedence. They overwrite any
mappings in the left operand. That is, (x 7→ v1 ) ⊳ (x 7→ v2 ) = (x 7→ v2 ).
Definition 3 (Expression Semantics (i, S) ⊢ b ⇓ v).
S(x) = v
(i, S) ⊢ b ⇓ v (i, S) ⊢ bs ⇓ vs
(i, S) ⊢ j ⇓ j (i, S) ⊢ [] ⇓ []
(i, S) ⊢ x ⇓ v
(i, S) ⊢ b : bs ⇓ v : vs
(i, S) ⊢ b ⇓ v : vs
(i, S) ⊢ b ⇓ [v1 , . . . , vn ]
(i, S) ⊢ tid ⇓ i (i, S) ⊢ hash(x) ⇓ x
(i, S) ⊢ head(b) ⇓ v
(i, S) ⊢ last(b) ⇓ vn
T
Definition 4 (Program Execution (S, P ) =
⇒ (S ′ , Q)).
i♯p
Single program thread
P, Q ::= [] | i♯p : P
Program threads
t
:= i♯x! | i ← j♯x? Send and receive event
T
::= [] | t : T
Trace
[]
We write (S, P ) =
⇒ (S ′ , Q) as a shorthand for (S, P ) =
⇒ (S ′ , Q).
Definition 5 (Single Step).
(Terminate) (S, i♯[] : P ) ⇒
= (S, P )
(Assign)
(MakeChan)
(i, S) ⊢ b ⇓ v S ′ = S ⊳ (y 7→ v)
(S, i♯(y := b : p) : P ) =
⇒ (S ′ , i♯p : P )
S ′ = S ⊳ (y 7→ Chan
(S, i♯(y := makeChan : p) : P ) =
⇒ (S ′ , i♯p : P )
7
Definition 6 (Multi-Threading and Synchronous Message-Passing).
(Go)
i 6∈ {i1 , . . . , in }
(S, i1 ♯(go p : p1 ) : P ) =
⇒ (S, i♯p : i1 ♯p1 : P )
∃l ∈ J, m ∈ K.el = x ← b fm = y :=← x S(x) = Chan
(i1 , S) ⊢ b ⇓ v S ′ = S ⊳ (y 7→ v)
(Sync) (S, i1 ♯(select [ej ⇒ qj ]j∈J : p1 ) : i2 ♯(select [fk ⇒ rk ]k∈K : p2 ) : P )
[i1 ♯x!,i2 ←i1 ♯x?]
==========⇒
(S ′ , i1 ♯(ql ++ p1 ) : i2 ♯(rm ++ p2 ) : P )
Definition 7 (Scheduling).
(Schedule)
π permutation on {1, . . . , n}
(S, [i1 ♯p1 , . . . , in ♯pn ]) =
⇒ (S, [π(i1 )♯pπ(1) , . . . , π(in )♯pπ(n) ])
T
(Closure)
3
(S, P ) =
⇒ (S ′ , P ′ )
T′
(S ′ , P ′ ) =⇒ (S ′′ , P ′′ )
T ++ T ′
(S, P ) =====⇒ (S ′′ , P ′′ )
Instrumentation and Run-Time Tracing
For each message passing primitive (send/receive) we log two events. In case
of send, (1) a pre event to indicate the message is about to be sent, and (2) a
post event to indicate the message has been sent. The treatment is analogous for
receive. In our instrumentation, we write x! to denote a single send event and
x? to denote a single receive event. These notations are shorthands and can be
expressed in terms of the language described so far. We use ≡ to define shortforms and their encodings. We define x! ≡ [hash(x), 1] and x? ≡ [hash(x), 0].
That is, send is represented by the number 1 and receive by the number 0.
As we support non-deterministic selection, we employ a list of pre events
to indicate that one of several events may be chosen For example, pre([x!, y?])
indicates that there is the choice among sending over channel x and receiving over
channel y. This is again a shorthand notation where we assume pre([b1 , . . . , bn ]) ≡
[0, b1 , . . . , bn ].
A post event is always singleton as at most one of the possible communications is chosen. As we also trace communication partners, we assume that the
sending party transmits its identity, the thread id, to the receiving party. We
write post (i♯x?) to denote reception via channel x where the sender has thread
id i. In case of a post send event, we simply write post (x!). The above are yet
again shorthands where i♯x? ≡ [hash(x), 0, i] and post (b) ≡ [1, b].
Pre and post events are written in a fresh thread local variable, denoted by
xtid where tid refers to the thread’s id number. At the start of the thread the
variable is initialized by xtid := []. Instrumentation ensures that pre and post
8
events are appropriately logged. As we keep track of communication partners, we
must also inject and project messages with additional information (the sender’s
thread id).
We consider instrumentation of select [x ← 1 ⇒ [], y :=← x ⇒ [z ← y]]. We
assume the above program text is part of a thread with id number 1. We nondeterministically choose between a send an receive operation. In case of receive,
the received value is further transmitted. Instrumentation yields the following.
[x1 := x1 ++ pre([x!, x?]),
select [x ← [tid, 1] ⇒ [x1 := x1 ++ post (x!)],
y ′ :=← x ⇒ [x1 := x1 ++ post (head(y ′ )♯x?), y := last(y ′ ),
z ← [tid, y]]]
We first store the pre events, either a read or send via channel x. The send is
instrumented by additionally transmitting the senders thread id. The post event
for this case simply logs that a send took place. Instrumentation of receive is
slightly more involved. As senders supply their thread id, we introduce a fresh
variable y ′ . Via head(y ′ ) we extract the senders thread id to properly record
the communication partner in the post event. The actual value transmitted is
accessed via last(y ′ ).
Definition 8 (Instrumentation of Programs). We write instr(p) = q to
denote the instrumentation of program p where q is the result of instrumentation.
Function instr (·) is defined by structural induction on a program. We assume a
similar instrumentation function for commands.
instr([])
instr(c : p)
= []
= instr (c) : instr (p)
instr(y := b)
= [y := b]
instr(y := makeChan)
= [y := makeChan]
instr(go p)
= [go ([xtid := [] ++ instr (p)])]
instr(select [ei ⇒ pi ]i∈{1,...,n} ) = [xtid := xtid ++ [pre([retr(e1 ), . . . , retr (en )])],
select [instr (ei ⇒ pi )]i∈{1,...,n} ]
instr(x ← b ⇒ p)
= x ← [tid, b] ⇒ (xtid := xtid ++ [post (x!)]) ++ instr (p)
instr(y :=← x ⇒ p)
= y ′ :=← x ⇒ [xtid := xtid ++ [post (head(y ′ )♯x?)],
y := last(y ′ )] ++ instr(p)
retr(x ← b) = x!
retr(y =← x) = x?
Run-time tracing proceeds as follows. We simply run the instrumented program and extract the local traces connected to variables xtid . We assume that
thread id numbers are created during program execution and can be enumerated
by 1 . . . n for some n > 0 where thread id number 1 belongs to the main thread.
Definition 9 (Run-Time Tracing). Let p and q be programs such that instr(p) =
T
q. We consider a specific instrumented program run where ((), [1♯[x1 := []] ++ q]) =
⇒
9
(S, 1♯[] : P ) for some S, T and P . Then, we refer to T as p’s actual run-time
trace. We refer to the list [1♯S(x1 ), . . . , n♯S(xn )] as the local traces obtained via
the instrumentation of p.
Command x1 := [] is added to the instrumented program to initialize the trace
of the main thread. Recall that main has thread id number 1. This extra step
is necessary because our instrumentation only initializes local traces of threads
generated via go. The final configuration (S, 1♯[] : P ) indicates that the main
thread has run to full completion. This is a realistic assumption as we assume
that programs exhibit no obvious bug during execution. There might still be
some pending threads, in case P differs from the empty list.
4
Trace Analysis
We assume that the program has been instrumented and after some program
run we obtain a list of local traces. We show that the actual run-time trace
can be recovered and we are able to point out alternative behaviors that could
have taken place. Alternative behaviors are either due alternative schedules or
different choices among communication partners.
We consider the list of local traces [1♯S(x1 ), . . . , n♯S(xn )]. Their shape can
be characterized as follows.
Definition 10 (Local Traces).
U, V
L
as
M
::= [] | i♯L : U
::= [] | pre(as) : M
::= [] | x! : as | x? : as
::= [] | post (x!) : L | post (i♯x?) : L
We refer to U = [1♯L1 , . . . , n♯Ln ] as a residual list of local traces if for each
Li either Li = [] or Li = [pre(. . . )].
To recover the communications that took place we check for matching pre and
post events recorded in the list of local traces. For this purpose, we introduce a
T
relation U =
⇒ V to denote that ‘replaying’ of U leads to V where communications
T took place. Valid replays are defined via the following rules.
10
T
Definition 11 (Replay U =
⇒ V ).
L1 = pre([. . . , x!, . . . ]) : post (x!) : L′1
L2 = pre([. . . , x?, . . . ]) : post (i1 ♯x?) : L′2
(Sync)
[i1 ♯x!,i2 ←i1 ♯x?]
i1 ♯L1 : i2 ♯L2 : U ==========⇒ i1 ♯L′1 : i2 ♯L′2 : U
(Schedule)
π permutation on {1, . . . , n}
[]
[i1 ♯L1 , . . . , in ♯Ln ] =
⇒ [iπ(1) ♯Lπ(1) , . . . , iπ(n) ♯Lπ(n) ]
T′
T
(Closure)
U=
⇒ U ′ U ′ =⇒ U ′′
T ++ T ′
U =====⇒ U ′′
Rule (Sync) checks for matching communication partners. In each trace, we must
find complementary pre events and the post events must match as well. Recall
that in the instrumentation the sender transmits its thread id to the receiver.
Rule (Schedule) shuffles the local traces as rule (Sync) only considers the two
leading local traces. Via rule (Closure) we perform repeated replay steps.
We can state that the actual run-time trace can be obtained via the replay reT
lation U =
⇒ V but further run-time traces are possible. This is due to alternative
schedules.
Proposition 1 (Replay Yields Run-Time Traces). Let p be a program
and q its instrumentation where for a specific program run we observe the actual behavior T and the list [1♯L1 , . . . , n♯Ln ] of local traces. Let T = {T ′ |
T′
[1♯L1 , . . . , n♯Ln ] =⇒ 1♯[] : U for some residual U }. Then, we find that T ∈ T
T′
and for each T ′ ∈ T we have that ((), p) =⇒ (S, 1♯[] : P ) for some S and P .
Definition 12 (Alternative Schedules). We say [1♯L1 , . . . , n♯Ln ] contains
T′
alternative schedules iff the cardinality of the set {T ′ | [1♯L1 , . . . , n♯Ln ] =⇒ 1♯[] :
U for some residual U } is greater than one.
We can also check if even further run-time traces might have been possible
by testing for alternative communications.
Definition 13 (Alternative Communications). We say [1♯L1 , . . . , n♯Ln ] contains alternative matches iff for some i, j, x, L, L′ we have that (1) Li = pre([. . . , x!, . . . ]) :
L, (2) Lj = pre([. . . , x?, . . . ]) : L′ , and (3) if L = post (x!) : L′′ for some L′′
then L′ 6= post (j♯x?) : L′′′ for any L′′′ .
We say U = [1♯L1 , . . . , n♯Ln ] contains alternative communications iff U
T
contains alternative matches or there exists T and V such that U =
⇒ V and V
contains alternative matches.
The alternative match condition states that a sender could synchronize with
a receiver (see (1) and (2)) but this synchronization did not take place (see (3)).
For an alternative match to result in an alternative communication, the match
must be along a possible run-time trace.
11
[x := makeChan, y := makeChan,
go [z := (← y)6 ], go [(y ← 1)4 , (x ← 1)5 ], go [(x ← 1)3 ],
x := (← x)1 , x := (← x)2 ]
x!|3
[4♯[pre((y?)6 ), post (3♯(y?)6 )],
3♯[pre((y!)4 ), post ((y!)4 ), pre((x!)5 ), post ((x!)5 )],
2♯[pre((x!)3 ), post ((x!)3 )],
1♯[pre((x?)1 ), post (2♯(x?)1 ), pre((x?)2 ), post (4♯(x?)3 )]]
x?|1
x?|2
x!|5
y!|4
y?|6
Fig. 1: Dependency Graph among Events
4.1
Dependency Graph for Efficient Trace Analysis
Instead of replaying traces to check for alternative schedules and communications, we build a dependency graph where the graph captures the partial order
among events. It is much more efficient to carry out the analysis on the graph
than replaying traces. Figure 1 shows a simple example.
We find a program that makes use of two channels and four threads. For
reference, send/receive events are annotated (as subscript) with unique numbers.
We omit the details of instrumentation and assume that for a specific program
run we find the list of given traces on the left. Pre events consist of singleton lists
as there is no select. Hence, we write pre((y?)6 ) as a shorthand for pre([(y?)6 ]).
Replay of the trace shows that the following locations synchronize with each
other: (4, 6), (3, 1) and (5, 2). This information as well as the order among events
can be captured by a dependency graph. Nodes are obtained by a linear scan
through the list of traces. To derive edges, we require another scan for each
element in a trace as we need to find pre/post pairs belonging to matching
synchronizations. This results overall in O(m ∗ m) for the construction of the
graph where m is the number of elements found in each trace. To avoid special
treatment of dangling pre events (with not subsequent post event), we assume
that some dummy post events are added to the trace.
Definition 14 (Construction of Dependency Graph). Each node corresponds to a send or a receive operation in the program text. Edges are constructed
by observing events recorded in the list of traces. We draw a (directed) edge among
nodes if either
– the pre and post events of one node precede the pre and post events of another
node in the trace, or
– the pre and post events belonging to both nodes can be synchronized. See rule
(Sync) in Definition 11. We assume that the edge starts from the node with
the send operation.
Applied to our example, this results in the graph on the right. See Figure 1.
For example, x!|3 denotes a send communication over channel x at program
12
location 3. As send precedes receive we find an edge from x!|3 to x?|1. In general,
there may be several initial nodes. By construction, each node has at most one
outgoing edge but may have multiple incoming edges.
The trace analysis can be carried out directly on the dependency graph.
To check if one event happens-before another event we seek for a path from
one event to the other. This can be done via a depth-first search and takes
time O(v + e) where v is the number of nodes and e the number of edges. Two
events are concurrent if neither happens-before the other. To check for alternative
communications, we check for matching nodes that are concurrent to each other.
By matching we mean that one of the nodes is a send and the other is a receive
over the same channel. For our example, we find that x!|5 and x?|1 represents an
alternative communication as both nodes are matching and concurrent to each
other.
To derive (all) alternative schedules, we perform a backward traversal of the
graph. Backward in the sense that we traverse the graph by moving from children
to parent node. We start with some final node (no outgoing edge). Each node
visited is marked. We proceed to the parent if all children are marked. Thus,
we guarantee that the happens-before relation is respected. For our example,
suppose we visit first y?6. We cannot visit its parent y!4 until we have visited x?2
and x!5. Via a (backward) breadth-first search we can ‘accumulate’ all schedules.
5
Comparison to Vector Clock Method
Via a simple adaptation of the Replay Definition 11 we can attach vector clocks
to each send and receive event. Hence, our tracing method strictly subsumes the
vector clock method as we are also able to trace events that could not commit.
Definition 15 (Vector Clock).
cs ::= [] | n : cs
For convenience, we represent a vector clock as a list of clocks where the first
position belongs to thread 1 etc. We write cs[i] to retrieve the i-th component
in cs. We write inc(i, cs) to denote the vector clock obtained from cs where
all elements are the same but at index i the element is incremented by one.
We write max(cs1 , cs2 ) to denote the vector clock where we per-index take the
greater element. We write ics to denote thread i with vector clock cs. We write
i♯x!cs to denote a send over channel x in thread i with vector clock cs. We write
i ← j♯x?cs to denote a receive over channel x in thread i from thread j with
vector clock cs.
Definition 16 (From Trace Replay to Vector Clocks).
(Sync)
L1 = pre([. . . , x!, . . . ]) : post (x!) : L′1
L2 = pre([. . . , x?, . . . ]) : post (i1 ♯x?) : L′2
cs = max(inc(i1 , cs1 ), inc(i2 , cs2 ))
[i1 ♯x!cs ,i2 ←i1 ♯x?cs ]
cs2
′
cs
′
1
ics
============⇒ ics
1 ♯L1 : i2 ♯L2 : U
1 ♯L1 : i2 ♯L2 : U =
13
Like the construction of the dependency graph, the (re)construction of vector
clocks takes time O(m ∗ m) where m is the number of elements found in each
trace.
To check for an alternative communication, the vector clock method seeks
for matching events. This incurs the same (quadratic in the size of the trace)
cost as for our method. However, the check that these two events are concurrent
to each other can be performed more efficiently via vector clocks. Comparison
of vector clocks takes time O(n) where n is the number of threads. Recall that
our graph-based method requires time O(v + e) where v is the number of nodes
and e the number of edges. The number n is smaller than v + e.
However, our dependency graph representation is more efficient in case of
exploring alternative schedules. In case of the vector clock method, we need to
continuously compare vector clocks whereas we only require a (backward) traversal of the graph. We believe that the dependency graph has further advantages
in case of user interaction and visualization as it is more intuitive to navigate
through the graph. This is something we intend to investigate in future work.
6
Implementation
We have fully integrated the approach laid out in the earlier sections into the
Go programming language and have built a prototype tool. We give an overview
of our implementation which can be found here [5]. A detailed treatment of all
of Go’s message-passing features can be found in the extended version of this
paper.
6.1
Library-Based Instrumentation and Tracing
We use a pre-processor to carry out the instrumentation as described in Section 3.
In our implementation, each thread maintains an entry in a lock-free hashmap
where each entry represents a thread (trace). The hashmap is written to file
either at the end of the program or when a deadlock occurs. We currently do
not deal with the case that the program crashes as we focus on the detection of
potential bugs in programs that do not show any abnormal behavior.
6.2
Measurement of Run-Time Overhead Library-Based Tracing
We measure the run-time overhead of our method against the vector clock
method. Both methods are implemented as libraries assuming no access to the
Go run-time system. For experimentation we use three programs where each program exercises some of the factors that have an impact on tracing. For example,
dynamic versus static number of threads and channels. Low versus high amount
of communication among threads.
The Add-Pipe (AP) example uses n threads where the first n − 1 threads
receive on an input channel, add one to the received value and then send the
14
3ms
6ms
243ms
PS100
C2000
107ms
476ms
5901ms
8ms
7ms
1475ms
PS250
910ms
AP21
6229ms
909ms
Default
3ms
17ms
C1000 44ms
AP51
1191ms
0ms
Pre-Post
2141ms
3825ms
7500ms
0ms
VC
7500ms
Fig. 2: Performance overhead using Pre/Post vs Vector clocks(VC) in ms.
new value on their output channel to the next thread. The first thread sends the
initial value and receives the result from the last thread.
In the Primesieve (PS) example, the communication among threads is similar to the Add-Pipe example. The difference is that threads and channels are
dynamically generated to calculate the first n prime numbers. For each found
prime number a ‘filter’ thread is created. Each thread has an input channel to
receive new possible prime numbers v and an output channel to report each number for which v mod prime 6= 0 where prime is the prime number associated
with this filter thread. The filter threads are run in a chain where the first thread
stores the prime number 2.
The Collector (C) example creates n threads that produce a number which
is then sent to the main thread for collection. This example has much fewer
communications compared to the other examples but uses a high number of
threads.
Figure 2 summarizes our results. Results are carried out on some commodity
hardware (Intel i7-6600U with 12 GB RAM, a SSD and Go 1.8.3 running on
Windows 10 was used for the tests). Our results show that a library-based implementation of the vector clock method does not scale well for examples with a dynamic number of threads and/or a high amount communication among threads.
See examples Primesieve and Add-Pipe. None of the vector clock optimizations [3] apply here because of the dynamic number of threads and channels. Our
method performs much better. This is no surprise as we require less (tracing)
data and no extra communication links. We believe that the overhead can still be
further reduced as access to the thread id in Go is currently rather cumbersome
and expensive.
7
Conclusion
One of the challenges of run-time verification in the concurrent setting is to
establish a partial order among recorded events. Thus, we can identify potential
bugs due to bad schedules that are possible but did not take place in some specific
program run. Vector clocks are the predominant method to achieve this task. For
15
example, see work by Vo [11] in the MPI setting and work by Tasharofi [10] in
the actor setting. There are several works that employ vector clocks in the shared
memory setting For example, see Pozniansky’s and Schuster’s work [9] on data
race detection. Some follow-up work by Flanagan and Freund [2] employs some
optimizations to reduce the tracing overhead by recording only a single clock
instead of the entire vector. We leave to future work to investigate whether
such optimizations are applicable in the message-passing setting and how they
compare to existing optimizations such as [3].
We have introduced a novel tracing method that has much less overhead
compared to the vector clock method. Our method can deal with all of Go’s
message-passing language features and can be implemented efficiently as a library.
We have built a prototype that can automatically identify alternative schedules
and communications. In future work we plan to conduct some case studies and
integrate heuristics for specific scenarios, e.g. reporting a send operation on a
closed channel etc.
Acknowledgments
We thank some HVC’17 reviewers for their constructive feedback on an earlier
version of this paper.
References
1. C. J. Fidge. Timestamps in message-passing systems that preserve the partial
ordering. 10(1):56–66, 1987.
2. C. Flanagan and S. N. Freund. Fasttrack: Efficient and precise dynamic race
detection. In Proc. of PLDI ’09, pages 121–133. ACM, 2009.
3. V. K. Garg, C. Skawratananond, and N. Mittal. Timestamping messages and events
in a distributed system using synchronous communication. Distributed Computing,
19(5-6):387–402, 2007.
4. The Go programming language. https://golang.org/.
5. Trace-based
run-time
analysis
of
message-passing
Go
programs.
https://github.com/KaiSta/gopherlyzer-GoScout.
6. C. A. R. Hoare. Communicating sequential processes. Commun. ACM, 21(8):666–
677, Aug. 1978.
7. L. Lamport. Time, clocks, and the ordering of events in a distributed system.
Communications of the ACM, 21(7):558–565, 1978.
8. F. Mattern. Virtual time and global states of distributed systems. In Parallel and
Distributed Algorithms, pages 215–226. North-Holland, 1989.
9. E. Pozniansky and A. Schuster. Multirace: efficient on-the-fly data race detection
in multithreaded C++ programs. Concurrency and Computation: Practice and
Experience, 19(3):327–340, 2007.
10. S. Tasharofi. Efficient testing of actor programs with non-deterministic behaviors.
PhD thesis, University of Illinois at Urbana-Champaign, 2013.
11. A. Vo. Scalable Formal Dynamic Verification of Mpi Programs Through Distributed
Causality Tracking. PhD thesis, University of Utah, 2011. AAI3454168.
16
func A ( x chan int ) {
x <- 1
// A1
}
func bufferedChan () {
x := make ( chan int ,1)
go A ( x )
x <- 1
// A2
<- x
}
func closedChan () {
x := make ( chan int )
go A ( x )
go B ( x )
close ( x )
}
func B ( x chan int ) {
<- x
}
func selDefault () {
x := make ( chan int )
go A ( x )
select {
case <-x :
// A3
fmt . Println ( " received from x " )
default :
fmt . Println ( " default " )
}
}
Fig. 3: Further Go Features
A
A.1
Further Go Message-Passing Features
Overview
Besides selective synchronous message-passing, Go supports some further message passing features that can be easily dealt with by our approach and are fully
supported by our implementation. Figure 3 shows such examples where we put
the program text in two columns.
Buffered Channels Go also supports buffered channels where send is asynchronous
assuming sufficient buffer space exists. See function buffered in Figure 3. Depending on the program run, our analysis reports that either A1 or A2 are
alternative matches for the receive operation.
In terms of the instrumentation and tracing, we treat each asynchronous
send as if the send is executed in its own thread. This may lead to some slight
inaccuracies. Consider the following variant.
func b u f f e r e d 2 () {
x := make ( chan int ,1)
x <- 1
go A ( x )
<-x
}
// B1
// B2
// B3
Our analysis reports that B2 and B3 form an alternative match. However, in the
Go semantics, buffered messages are queued. Hence, for every program run the
only possibility is that B1 synchronizes with B3. B3 never takes place! As our
17
main objective is bug finding, we argue that this loss of accuracy is justifiable.
How to eliminate such false positives is subject of future work.
Select with default/timeout Another feature in Go is to include a default/timeout
case to select. See selDefault in Figure 3. The purpose is to avoid (indefinite)
blocking if none of the other cases are available. For the user it is useful to find
out if other alternatives are available in case the default case is selected. The
default case applies for most program runs. Our analysis reports that A1 and
A3 are an alternative match.
To deal with default/timeout we introduce a new post event post (select ).
To carry out the analysis in terms of the dependency graph, each subtrace
. . . , pre([. . . , select , . . . ]), post (select ), . . . creates a new node. Construction of
edges remains unchanged.
Closing of Channels Another feature in Go is the ability to close a channel. See
closedChan in Figure 3. Once a channel is closed, each send on a closed channel
leads to failure (the program crashes). On the other hand, each receive on a
closed channel is always successful, as we receive a dummy value. A run of is
successful if the close operation of the main thread happens after the send in
thread A. As the close and send operations happen concurrently, our analysis
reports that the send A1 may take place after close.
For instrumentation/tracing, we introduce event close(x). It is easy to identify a receive on a closed channel, as we receive a dummy thread id. So, for
each subtrace [. . . , pre([. . . , x?, . . . ]), post (i♯x?), . . . ] where i is a dummy value
we draw an edge from close(x) to x?.
Here are the details of how to include buffered channels, select and closing
of channels.
A.2
Buffered Channels
Consider the following Go program.
x := make ( chan , 2)
x <- 1
// E1
x <- 1
// E2
<- x
// E3
<- x
// E4
We create a buffer of size 2. The two send operations will then be carried out
asynchronously and the subsequent receive operations will pick up the buffered
values. We need to take special care of buffered send operations. If we would
treat them like synchronous send operations, their respective pre and post events
would be recorded in the same trace as the pre and post events of the receive
operations. This would have the consequence that our trace analysis does not
find out that events E1 and E2 happen before E3 and E4.
Our solution to this issue is to treat each send operation on a buffered channel
as if the send operation is carried out in its own thread. Thus, our trace analysis
18
is able to detect that E1 and E2 take place before E3 and E4. This is achieved by
marking each send on a buffered channel in the instrumentation. After tracing,
pre and post events will then be moved to their own trace. From the viewpoint
of our trace analysis, a buffered channel then appears as having infinite buffer
space. Of course, when running the program a send operation may still block if
all buffer space is occupied.
Here are the details of the necessary adjustments to our method. During instrumentation/tracing, we simply record if a buffered send operation took place.
The only affected case in the instrumentation of commands (Definition 8) is
x ← b ⇒ p. We assume a predicate isBuffered(·) to check if a channel is buffered
or not. In terms of the actual implementation this is straightforward to implement. We write postB (x, n) to indicate a buffered send operation via x where n
is a fresh thread id. We create fresh thread id numbers via tidB.
Definition 17 (Instrumentation of Buffered Channels). Let x be a buffered
channel.
instr(x ← b ⇒ p)
| isBuffered(x) = x ← [n, b] ⇒ (xtid := xtid ++ [postB (x!n, ])) ++ instr (p)
where n = tidB
| otherwise
= x ← [tid, b] ⇒ (xtid := xtid ++ [post (x!)]) ++ instr (p)
The treatment of buffered channels has no overhead on the instrumentation
and tracing. However, we require a post-processing phase where marked events
will be then moved to their own trace. This can be achieved via a linear scan
through each trace. Hence, requires time complexity O(k) where k is the overall
size of all (initially recorded) traces. For the sake of completeness, we give below
a declarative description of post-processing in terms of relation U ⇒ V .
Definition 18 (Post-Processing for Buffered Channels U ⇒ V ).
(MovePostB)
L = pre(as) : postB (x!, n) : L′
i♯L : U ⇒ i♯L′ : n♯[pre(as), postB (x!, n)] : U
L = pre(as) : post (a) : L′
(a = x! ∨ a = j♯x?)
(Shift)
i♯L′ : U ⇒ i♯L′′ : U ′
i♯L : U ⇒ i♯pre(as) : post (a) : L′′ : U ′
(Schedule)
π permutation on {1, . . . , n}
[i1 ♯L1 , . . . , in ♯Ln ] ⇒ [iπ(1) ♯Lπ(1) , . . . , iπ(n) ♯Lπ(n) ]
(Closure)
U ⇒ U ′ U ′ ⇒ U ′′
U ⇒ U ′′
19
Subsequent analysis steps will be carried out on the list of traces obtained
via post-processing.
There is some space for improvement. Consider the following program text.
func A ( x chan int ) {
x <- 1
// A1
}
func b u f f e r e d 2 () {
x := make ( chan int ,1)
x <- 1
go A ( x )
<-x
}
// B1
// B2
// B3
Our analysis (for some program run) reports that B2 and B3 is an alternative
match. However, in the Go semantics, buffered messages are queued. Hence, for
every program run the only possibility is that B1 synchronizes with B3. B3 never
takes place. As our main objective is bug finding, we can live with this inaccuracy.
We will investigate in future work how to eliminate this false positive.
B
Select with default/timeout
In terms of the instrumentation/tracing, we introduce a new special post event
post (select ). For the trace analysis (Definition 11), we require a new rule.
[i♯select]
(Default/Timeout) i♯pre([. . . ]) : post (select ) : L : U =====⇒ i♯L : U
This guarantees that in case default or timeout is chosen, select acts as if asynchronous.
The dependency graph construction easily takes care of this new feature. For
each default/timeout case we introduce a node. Construction of edges remains
unchanged.
C
Closing of Channels
For instrumentation/tracing of the close(x) operation on channel x, we introduce
a special pre and post event. Our trace analysis keeps track of closed channels. As
a receive on a closed channel yields some dummy values, it is easy to distinguish
this case from the regular (Sync). Here are the necessary adjustments to our
replay relation from Definition 11.
C ::= [] | i♯close(x) : C
20
[]
(Close) (i♯pre(close(x)) : post (close(x)) : L : U | C) =
⇒ (i♯L : U | i♯close(x) : C)
(RcvClosed)
Q = j♯close(x) : Q′
[j♯close(x),i←j♯x?]
(i♯pre([. . . , x?, . . . ]) : post (j ′ ♯x?) : L : U | Q) ============⇒ (i♯L : U | Q)
For the construction of the dependency graph, we create a node for each close
statement. For each receive on a closed channel x at program location l, we draw
an edge from close(x) to x?|l.
D
Codes used for the Experimental results
Add-Pipe
func add1 ( in chan int) chan int {
out := make ( chan int )
go func () {
for {
n := <- in
out <- n + 1
}
}()
return out
}
func main () {
in := make ( chan int)
c1 := add1 ( in )
for i := 0; i < 19; i ++ {
c1 = add1 ( c1 )
}
for n := 1; n < 1000; n ++ {
in <- n
<- c1
}
}
Primesieve
func g e n e r a t e( ch chan int ) {
for i := 2; ; i ++ {
ch <- i
21
}
}
func filter ( in chan int , out chan int , prime int ) {
for {
tmp := <- in
if tmp % prime != 0 {
out <- tmp
}
}
}
func main () {
ch := make ( chan int)
go g e n e r a t e( ch )
for i := 0; i < 100; i ++ {
prime := <- ch
ch1 := make ( chan int )
go filter ( ch , ch1 , prime )
ch = ch1
}
}
Collector
func c o l l e c t( x chan int , v int ) {
x <- v
}
func main () {
x := make ( chan int )
for i := 0; i < 1000; i ++ {
go c o l l e c t(x , i )
}
for i := 0; i < 1000; i ++ {
<-x
}
}
22
x!|2
x?|4
x!|4’
x?|3
x!|2’
x?|4’’
x!|4’
x?|3
x!|2
x?|4
| 6cs.PL
|
Deep Mean Field Games for Learning Optimal Behavior Policy of
Large Populations
Jiachen Yang1 , Xiaojing Ye2 , Rakshit Trivedi1 , Huan Xu3 , and Hongyuan Zha1
1
College of Computing, Georgia Institute of Technology
Department of Mathematics and Statistics, Georgia State University
3
School of Industrial and Systems Engineering, Georgia Institute of Technology
arXiv:1711.03156v1 [cs.LG] 8 Nov 2017
2
Abstract
We consider the problem of representing a large population’s behavior policy that drives the evolution of the
population distribution over a discrete state space. A discrete time mean field game (MFG) is motivated as an interpretable model founded on game theory for understanding the aggregate effect of individual actions and predicting
the temporal evolution of population distributions. We achieve a synthesis of MFG and Markov decision processes
(MDP) by showing that a special MFG is reducible to an MDP. This enables us to broaden the scope of mean field
game theory and infer MFG models of large real-world systems via deep inverse reinforcement learning. Our method
learns both the reward function and forward dynamics of an MFG from real data, and we report the first empirical
test of a mean field game model of a real-world social media population.
1
Introduction
Nothing takes place in the world whose meaning is not that of some maximum or minimum.
(Leonhard Euler)
Major global events involving large populations, such as the wave of protests during the Arab Spring, the Black Lives
Matter movement, and the controversy over fake news during the 2016 U.S. presidential election, provide significant
impetus for devising new models that account for macroscopic population behavior resulting from the aggregation of
decisions and actions taken by all individuals (Howard et al., 2011; Anderson and Hitlin, 2016; Silverman, 2016). Just
as physical systems behave according to the principle of least action, to which Euler’s statement alludes, population
behavior emerging from individual actions may also be optimal with respect to some objective. The influential role
of social media in modern mass movements lends plausibility to this hypothesis (Perrin, 2015), since the availability
of information enables individuals to plan and act based on their observations of the global population state. For
example, a population’s behavior directly affects the ranking of a set of trending topics on social media, represented
by the global population distribution over topics, while each user’s observation of this global state influences their
choice of the next topic in which to participate, thereby contributing to future population behavior (Twitter, 2017).
In general, this phenomenon is present in any system where the distribution of a large population over a set of states
is observable (or partially observable) by the population itself, whose implicit behavior policy is informed by their
observations. This motivates multiple criteria for a model of population behavior:
1. The model captures the dependency between population distribution and their behavior policy.
2. It is explainable via a notion of a reward optimized by the aggregate decisions of all individuals.
3. It enables prediction of future distribution over a state space given measurements at previous times, and can be
learned from real data.
We present a mean field game (MFG) approach to address the modeling and prediction criteria. Mean field games
originated as a branch of game theory that provides tractable models of large agent populations, by considering the
limit of N -player games as N tends to infinity (Lasry and Lions, 2007). In this limit, an agent population is represented
1
via their distribution over a state space, the mutual influence between individual agents becomes infinitesimal, and each
agent’s optimal strategy is informed by a reward that is a function of the population distribution and their aggregate
actions. In its most general form, MFG represents a class of stochastic differential equations that can be specialized to
model the production of economic resources (Guéant et al., 2011), opinion dynamics in social networks (Bauso et al.,
2016), and the adoption of competing technologies by consumer populations (Lachapelle et al., 2010). Representing
agents as a distribution means that MFG is scalable to arbitrary population sizes, enabling it to simulate real-world
phenomenon such as the Mexican wave in stadiums (Guéant et al., 2011).
As the model detailed in Section 3 will show, MFG naturally addresses the modeling criteria in our problem context by
overcoming limitations of alternative predictive methods. For example, time series analysis builds predictive models
from data, but these models may not provide insight into the motivations that produce a population’s behavior policy,
since they do not consider the behavior as the result of optimization of a reward function. Alternatively, methods
that employ the underlying population network structure have assumed that nodes are only influenced by a local
neighborhood, do not include a representation of a global state, and may face difficulty in explaining events as the
result of uncontrolled implicit optimization (Farajtabar et al., 2015; De et al., 2016). MFG is unique as a descriptive
model whose solution tells us how a system naturally behaves according to its underlying optimal control policy. This
is the essential insight that enables us to draw a connection with the framework of Markov decision processes (MDP)
and reinforcement learning (RL) (Sutton and Barto, 1998). The crucial difference from a traditional MDP viewpoint is
that we frame the problem as MFG model inference via MDP policy optimization: we infer the implicit optimization
that the system performs on its own accord, by solving an associated MDP without externally controlling the system.
MFG offers a computationally tractable framework for adapting inverse reinforcement learning (IRL) methods (Ng and
Russell, 2000; Ziebart et al., 2008; Finn et al., 2016), with flexible neural networks as function approximators, to learn
complex reward functions that explain behavior of arbitrarily large populations. In the other direction, RL enables us
to devise a data-driven method for solving an MFG model of a real-world system. While research on the theory of
MFG has progressed rapidly in recent years, with some examples of numerical simulation of synthetic toy problems,
there is a conspicuous absence of scalable methods for empirical validation (Lachapelle et al., 2010; Achdou et al.,
2012; Bauso et al., 2016). Therefore, while we show how MFG is well-suited for the specific problem of modeling
population behavior, we also demonstrate a general data-driven approach to MFG inference via a synthesis of MFG
and MDP.
Our main contributions are the following. We propose a data-driven approach to learn an MFG model along with its
reward function, showing that research in MFG need not be confined to toy problems with artificial reward functions.
Specifically, we derive a discrete time graph-state MFG from general MFG and provide detailed interpretation in a
real-world setting (Section 3). Then we prove that a special case can be reduced to an MDP and show that finding
an optimal policy and reward function in the MDP is equivalent to inference of the MFG model (Section 4). Using
our approach, we empirically validate an MFG model of population’s activity distribution on social media (Section 5).
The learned MFG model shows significantly better predictive performance compared to baselines and offers insights
on population behavior. Our synthesis of MFG with MDP has potential to open new research directions for both fields.
2
Related work
Mean field games originated in the work of Lasry and Lions (2007), and independently as stochastic dynamic games
in Huang et al. (2006), both of which proposed mean field problems in the form of differential equations for modeling
problems in economics and analyzed the existence and uniqueness of solutions. Guéant et al. (2011) provided a survey
of MFG models and discussed various applications in continuous time and space, such as a model of population
distribution that informed the choice of application in our work. Even though the MFG framework is agnostic towards
the choice of cost function (i.e. negative reward), prior work make strong assumptions on the cost in order to attain
analytic solutions. We take a view that the dynamics of any game is heavily impacted by the reward function, and
hence we propose methods to learn the MFG reward function from data.
Discretization of MFGs in time and space have been proposed (Gomes et al., 2010; Achdou et al., 2012; Guéant,
2015), serving as the starting point for our model of population distribution over discrete topics; while these early
work analyze solution properties and lack empirical verification, we focus on algorithms for attaining solutions in
real-world settings. Related to our application case, prior work by Bauso et al. (2016) analyzed the evolution of
2
opinion dynamics in multi-population environments, but they imposed a Gaussian density assumption on the initial
population distribution and restrictions on agent actions, both of which limit the generality of the model and are not
assumed in our work. There is a collection of work on numerical finite-difference methods for solving continuous
mean field games (Achdou et al., 2012; Lachapelle et al., 2010; Carlini and Silva, 2014). These methods involve
forward-backward or Newton iterations that are sensitive to initialization and have inherent computational challenges
for large real-valued state and action spaces, which limit these methods to toy problems and cannot be scaled to
real-world problems. We overcome these limitations by showing how the MFG framework enables adaptation of RL
algorithms that have been successful for problems involving unknown reward functions in large real-world domains.
In reinforcement learning, there are numerous value- and policy-based algorithms employing deep neural networks as
function approximators for solving MDPs with large state and action spaces (Mnih et al., 2013; Silver et al., 2014;
Lillicrap et al., 2015). Even though there are generalizations to multi-agent settings (Hu et al., 1998; Littman, 2001;
Lowe et al., 2017), the MDP and Markov game frameworks do not easily suggest how to represent systems involving
thousands of interacting agents whose actions induce an optimal trajectory through time. In our work, mean field game
theory is the key to framing the modeling problem such that RL can be applied.
In the area of inverse reinforcement learning (Ng and Russell, 2000), the maximum entropy IRL framework has proved
successful at learning unknown reward functions from expert demonstrations in situations involving human and robotic
agency (Ziebart et al., 2008; Boularias et al., 2011; Kalakrishnan et al., 2013). This probabilistic framework can be
augmented with deep neural networks for learning complex reward functions from demonstration samples (Wulfmeier
et al., 2015; Finn et al., 2016). Our MFG model enables us to extend the sample -based IRL algorithm in Finn et al.
(2016) to the problem of learning a reward function under which a large population’s behavior is optimal, and we
employ a neural network to process MFG states and actions efficiently.
3
Mean field games
We begin with an overview of a continuous-time mean field games over graphs, and derive a general discrete-time
graph-state MFG (Guéant, 2015). Then we give a detailed presentation of a discrete-time MFG over a complete graph,
which will be the focus for the rest of this paper.
3.1
Mean field games on graphs
Let G = (V, E) be a directed graph, where the vertex set V = {1, . . . , d} represents d possible states of each agent,
and E ⊆ V × V is the edge set consisting of all possible direct transition between states (i.e., a agent can hop from
i to j only if (i, j) ∈ E). For each node i ∈ V, define Vi+ := {j : (j, i) ∈ E}, Vi− := {j : (i, j) ∈ E}, and
V̄i+ := Vi+ ∪ {i} and V̄i− := Vi− ∪ {i}. Let πi (t) be the density (proportion) of agent population in state i at time t,
and π(t) := (π1 (t), . . . , πd (t)). Population dynamics are generated by right stochastic matrices P (t) ∈ S(G), where
S(G) := S1 (G) × · · · × Sd (G) and each row Pi (t) belongs to Si (G) := {p ∈ ∆d−1 | supp(p) ⊂ V̄i− } where ∆d−1 is
the simplex in Rd . Moreover, we have a value function Vi (t) of state i at time t, and a reward function ri (π(t), Pi (t))
1
, quantifying the instantaneous reward for agents in state i taking transitions with probability Pi (t) when the current
distribution is π(t). We are mainly interested in a discrete time graph state MFG, which is derived from a continuous
time MFG by the following proposition. Appendix A provides a derivation from the continuous time MFG.
Proposition 1. Under a semi-implicit discretization scheme with unit time step labeled by n, the backward HamiltonJacobi-Bellman (HJB) equation and the forward Fokker-Planck equation for each i ∈ {1, . . . , d} and n = 0, . . . , N −1
in a discrete time graph state MFG are given by:
X
n n+1
(HJB)
Vin = maxPin ∈Si (G) ri (π n , Pin ) +
P
V
(1)
ij j
j∈V̄i−
X
n n
(Fokker-Planck)
πin+1 =
Pji
πj
(2)
+
j∈V̄i
1 We
here consider a rather special formulation where the reward function ri only depends on the overall population distribution π(t) and the
choice Pi the players in state i made.
3
3.2
Discrete time MFG over complete graph
Proposition 1 shows that a discrete time MFG given in Gomes et al. (2010) can be seen as a special case of a discrete
time graph state MFG with a complete graph (such that S(G) = ∆d−1 × · · · × ∆d−1 (d of ∆d−1 )). We focus on the
complete graph in this paper, as the methodology can be readily applied to general directed graphs. While Section 4
will show a connection between MFG and MDP, we note here that a “state” in the MFG sense is a node in V and not
an MDP state. 2 We now interpret the model using the example of evolution of user activity distribution over topics
on social media, to provide intuition and set the context for our real-world experiments in Section 5. Independent of
any particular interpretation, the MFG approach is generally applicable to any problem where population size vastly
outnumbers a set of discrete states.
• Population distribution π n ∈ ∆d−1 for n = 0, . . . , N − 1. Each π n is a discrete probability distribution over d
topics, where πin is the fraction of people who posted on topic i at time n. Although a person may participate in
more than one topic within a time interval, normalization can be enforced by a small time discretization or by using
a notion of “effective population size”, defined as population size multiplied by the max participation count of any
person during any time interval. π 0 is a given initial distribution.
• Transition matrix P n ∈ S(G). Pijn is the probability of people in topic i switching to topic j at time n, so we refer
to Pin as the action of people in topic i. P n generates the forward equation
d
X
Pijn πin
(3)
πjn+1 =
i=1
n
, Pin )
Pd
n
n
n
j=1 Pij rij (π , Pi ),
• Reward ri (π
:=
for i ∈ {1, . . . , d}. This is the reward received by people in topic
i who choose action Pin at time n, when the distribution is π n . In contrast to previous work, we learn the reward
function from data (Section 4.1). The only assumption we make is that reward for i depends only on Pin , not on the
entire P n . This is a causality assumption that actions by people in j 6= i have no instantaneous effect on the reward
for people in topic i. 3
• Value function V n ∈ Rd . Vin is the expected maximum total reward of being in topic i at time n. A terminal
value V N −1 is given, which we set to zero to avoid making any assumption on the problem structure beyond what
is contained in the learned reward function.
• Average reward ei (π, P, V ), for i ∈ {1, . . . , d} and V ∈ Rd and P ∈ S(G). This is the average reward received
by agents at topic i when the current distribution is π, action P is chosen, and the subsequent expected maximum
total reward is V . It is defined as:
d
X
ei (π, P, V ) =
Pij (rij (π, P ) + Vj )
(4)
j=1
Intuitively, agents want to act optimally in order to maximize their expected total average reward.
For P ∈ S(G) and a vector q ∈ Si (G), define P(P, i, q) to be the matrix equal to P , except with the i-th row replaced
by q. Then a Nash maximizer is defined as follows:
Definition 1. A right stochastic matrix P ∈ S(G) is a Nash maximizer of e(π, P, V ) if, given a fixed π ∈ ∆d−1 and a
fixed V ∈ Rd , for any i ∈ {1, . . . , d} and any q ∈ Si (G), there is
ei (π, P, V ) ≥ ei (π, P(P, i, q), V )
(5)
The rows of P form a Nash equilibrium set of actions, since for any topic i, the people in topic i cannot increase their
reward by unilaterally switching their action from Pi to any q. Under Definition 1, the value function of each topic i
at each time n satisfies the optimality criteria:
X
d
n+1
n
n
n
Vi = max
qj rij (π , P(P , i, q)) + Vj
(6)
q∈Si (G)
j=1
A solution of the MFG is a sequence of pairs {(π n , V n )}n=0,...,N satisfying optimality criteria (6) and forward equation (3).
2 Section
4 explains that the population distribution π is the appropriate definition of an MDP state.
this assumption is removed, there is a resemblance between the discrete time MFG and a Markov game in a continuous state and continuous
action space (Littman, 2001; Hu et al., 1998). See Appendix G for a discussion.
3 If
4
4
Inference of MFG via MDP optimization
A Markov decision process is a well-known framework for optimization problems. We focus on the discrete time MFG
in Section 3.2 and prove a reduction to a single-agent finite-horizon deterministic MDP, whose state trajectory under
an optimal policy coincides with the forward evolution of the MFG. This leads to the essential insight that solving
the optimization problem of a single-agent MDP is equivalent to solving the inference problem of an MFG. This
connection will enable us to apply efficient inverse RL methods, using measurements of real population trajectories,
to learn an MFG model along with its reward function in Section 4.1. The MDP is constructed as follows:
Definition 2. A single-agent finite-horizon deterministic MDP for a discrete time MFG over a complete graph is
defined as:
• States: π n ∈ ∆d−1 , the population distribution at time n.
• Actions: P n ∈ S(G), the transition probability matrix at time n.
Pd
Pd
• Reward: R(π n , P n ) := i=1 πin j=1 Pijn rij (π n , Pin )
Pd
• Finite-horizon state transition, given by Eq (3): ∀n ∈ {0, . . . , N − 1} : πjn+1 = i=1 Pijn πin .
Theorem 2. The value function of a solution to the discrete time MFG over a complete graph defined by optimality
criteria (6) and forward equation (3) is a solution to the Bellman optimality equation of the MDP in Definition 2.
Proof. Since rij depends on P n only through row Pin , optimality criteria 6 can be written as
X
X
Vin = max
Pij rij (π n , Pi ) +
Pij Vjn+1 .
Pi ∈Si (G)
j
(7)
j
We now define V ∗ (π n ) as follows and show that it is the value function of the constructed MDP in Definition 2 by
verifying that it satisfies the Bellman optimality equation:
V ∗ (π n ) :=
d
X
i=1
πin Vin =
d
X
πin max
Pi ∈Si (G)
i=1
= max
X
d
P ∈S(G)
= max
P ∈S(G)
= max
X
d
P ∈S(G)
πin
i=1
Pij rij (π n , Pi ) +
j=1
d
X
Pij Vjn+1
(8)
j=1
n
Pij rij (π , Pi ) +
d
d
X
X
j=1
j=1
R(π n , P ) +
d
X
d
X
πjn+1 Vjn+1
!
Pij πin
Vjn+1
(9)
i=1
(10)
j=1
∗
n
R(π , P ) + V (π
n+1
)
(11)
which is the Bellman optimality equation for the MDP in Definition 2.
Corollary 1. Given a start state π 0 , the state trajectory under the optimal policy of the MDP in Definition 2 is
equivalent to the forward evolution part of the solution to the MFG.
Proof. Under the optimal policy, equations 11 and 8 are satisfied, which means the matrix P generated by the optimal
policy at any state π n is the Nash maximizer matrix. Therefore, the state trajectory {π n }n=0,...,N −1 is the forward
part of the MFG solution.
4.1
Reinforcement learning solution for MFG
MFG provides a general framework for addressing the problem of modeling population dynamics, while the new
connection between MFG and MDP enables us to apply inverse RL algorithms to solve the MDP in Definition 2 with
5
unknown reward. In contrast to previous MFG research, most of which impose reward functions that are quadratic in
actions and logarithmic in the state distribution (Guéant, 2009; Lachapelle et al., 2010; Bauso et al., 2016), we learn
a reward function using demonstration trajectories measured from actual population behavior, to attain a succinct and
data-driven representation of the motivation behind population dynamics.
We leverage the MFG forward dynamics (Eq 3) in a sample-based IRL method based on the maximum entropy IRL
framework (Ziebart et al., 2008). From this probabilistic viewpoint, we minimize the relative entropy between a probability distribution p(τ ) over a space of trajectories T := {τi }i and a distribution q(τ ) from which demonstrated expert
trajectories are generated (Boularias et al., 2011). This is related to a path integral IRL formulation, where the likelihood of measured optimal trajectories is evaluated only using trajectories generated from their local neighborhood,
rather than uniformly over the whole trajectory space (Kalakrishnan et al., 2013). Specifically, making no assumption
on the true distribution of optimal demonstration other than matching of reward expectation, we posit that demonstration trajectories τi = (π 0 , P 1 , . . . , π N −1 , P N −1 )i are sampled from the maximum entropy distribution (Jaynes,
1957):
1
(12)
p(τ ) = exp(RW (τ ))
Z
P
where RW (τ ) = n RW (π n , P n ) is the sum of reward of single state-action pairs over a trajectory τ , and W are the
parameters of the reward function approximator (derivation in Appendix E). Intuitively, this means that trajectories
with higher reward are exponentially more likely to be sampled. Given M sampleR trajectories τj ∈ Dsamp from k
distributions F1 (τ ), . . . , Fk (τ ), an unbiased estimator of the partition function Z = exp(RW (τ ))dτ using multiple
P
1
importance sampling is Ẑ := M
τj zj exp(RW (τj )) (Owen and Zhou, 2000), where importance weights are zj :=
1 P
−1
(derivation in Appendix F). Each action matrix P is sampled from a stochastic policy Fk (P ; π, θ)
k Fk (τj )
k
(overloading notation with F (τ )), where π is the current state and θ the policy parameter. The negative log likelihood
of L demonstration trajectories τi ∈ Ddemo is:
X
X
1
1
L(W ) = −
RW (τi ) + log
zj exp(RW (τj ))
(13)
L
M
τi ∈Ddemo
τj ∈Dsamp
We build on Guided Cost Learning (GCL) in Finn et al. (2016) (Alg 1) to learn a deep neural network approximation
of RW (π, P ) via stochastic gradient descent on L(W ), and learn a policy F (P ; π, θ) using a simple actor-critic
algorithm (Sutton and Barto, 1998). In contrast to GCL, we employ a combination of convolutional neural nets
and fully-connected layers to process both the action matrix P and state vector π efficiently in a single architecture
(Appendix C), analogous to how Lillicrap et al. (2015) handle image states in Atari games. Due to our choice of
policy parameterization (described below), we also set importance weights to unity for numerical stability. These
implementation choices result in successful learning of a reward representation (Fig 1).
Our forward MDP solver (Alg 2) performs gradient ascent on the expected value Eθ [π 0 ] w.r.t. policy parameter θ, to
find successively improved stochastic policies Fk (P ; π, θ). We construct the joint distribution F (P ; π, θ) informed
by domain knowledge about human population behavior on social media, but this does not reduce the generality of
the MFG framework since it is straightforward to employ flexible policy and value networks in a DDPG algorithm
when intuition is not available (Silver et al., 2014; Lillicrap et al., 2015). Our joint distribution is d instances of a
d-dimensional Dirichlet distribution, each parameterized by an αi ∈ Rd+ . Each row Pi can be sampled from
f (Pi1 , . . . , Pid ; α1i , . . . , αdi ) =
d
i
1 Y
(Pij )αj −1
i
B(α ) j=1
(14)
where B(·) is the Beta function and αji is defined using the softplus function αji (π, θ) := ln(1 + exp{θ(πj − πi )}),
which is a monotonically increasing function of the population density difference πj −πi . In practice, a constant scaling
Qd
factor c ∈ R can be applied to α for variance reduction. Finally, we let F (P n ; π n , θ) = i=1 f (Pin ; αi (π n , θ)) denote
the parameterized policy, from which P n is sampled based on π n , and whose logarithmic gradient ∇θ ln(F ) can be
used in a policy gradient algorithm. We employ variance reduction by learning the value function using a linear
function approximation V̂ (π; w), containing all components of π up to second-order, with parameter w (Konda and
Tsitsiklis, 2000).
6
Reward density for training demo and generated transitions
Demo (test)
Generated
3.5
0.3
2
States
2.5
3
2.0
0.2
1
0.1
1.5
0.0
1.0
1
0.1
2
0.5
0
0.1
0.0
0.1
Reward
0.2
0.3
(a) Reward densities on train set
0.4
0.5
0.4
0
3.0
Density
Density
4
Reward of state-action pairs
Reward density for test demo and generated transitions
Demo (train)
Generated
0.2
0.1
0.0
0.1
Reward
0.2
0.3
(b) Reward densities on test set
0.4
0
1
Actions
2
(c) Reward of state-action pairs
Figure 1: (a) JSD between train demo and generated transitions is 0.130. (b) JSD between test demo and generated
transitions is 0.017. (c) Reward of state-action pairs. States: large negative mass gradient from π1 to πd (S0), less
negative gradient (S1), uniform (S2). Actions: high probability transitions to smaller indices (A0), uniform transition
(A1), row-reverse of A0 (A2).
5
Experiments
We demonstrate the effectiveness of our method with two sets of experiments: (i) recovery of an interpretable reward
function and (ii) prediction of population trajectory over time. Our experiment matches the discrete time mean field
game given in Section 3.2: we use data representing the activity of a Twitter population consisting of 406 users. We
model the evolution of the population distribution over d = 15 topics and N = 16 time steps (9am to midnight)
each day for 27 days. The sequence of state-action pairs {(π n , P n )}n=0,...,N −1 measured on each day shall be
called a demonstration trajectory. Although the set of topics differ semantically each day, indexing topics in order
of decreasing initial popularity suffices for identifying the topic sets across all days. As explained earlier, the MFG
framework can model populations of arbitrarily large size, and we find that our chosen population is sufficient for
extracting insights on population behavior. For evaluating performance on trajectory prediction, we compare MFG
with two baselines:
VAR. Vector autoregression of order 18 trained on 21 demonstration trajectories.
RNN. Recurrent neural network with a single fully-connected layer and rectifier nonlinearity.
We use Jenson-Shanon Divergence (JSD) as metric to report all our results. Appendix D provides comprehensive
implementation details.
5.1
Interpretation of reward function
Our method learned a representation of the implicit reward optimized by population behavior, which we evaluated
using four sets of state-action pairs acquired from: 1. all train demo trajectories; 2. trajectories generated by the
learned policy given initial states π 0 of train trajectories; 3. all test demo trajectories; 4. trajectories generated by the
learned policy given initial states π 0 of test trajectories. We find three distinct modes in the density of reward values for
both the train group of sets 1 and 2 (Fig 1a) and the test group of sets 3 and 4 (Fig 1b). Although we do not have access
to a ground truth reward function, the low JSD values of 0.13 and 0.017 between reward distributions for demo and
generated state-action pairs show generalizability of the learned reward function. We further investigated the reward
landscape with nine state-action pairs (Figure 1c), and find that the mode with highest rewards is attained by pairing
states that have large mass in topics having high initial popularity (S0) with action matrices that favor transition to
topics with higher density (A0). On the other hand, uniformly distributed state vectors (S2) attain the lowest rewards,
while states with a small negative mass gradient from topic 1 to topic d (S1) attain medium rewards.
5.2
Trajectory prediction
The primary hypothesis to test is that real user populations act near-optimally on social media, just as the MFG
approach assumes rational agents. Fig 2a (log scale) shows that MFG has 58% smaller error than VAR when evaluated
N −1
N −1
on the JSD between generated and measured final distributions JSD(πgenerated
, πmeasured
), and 40% smaller error when
7
Demonstration actions
JSD (log scale)
Average test error over 6 days
MFG
VAR
RNN
10
1
10
2
Final distribution
0
Difference
0
1.0
0.8
5
5
10
10
Entire trajectory (averaged)
0
5
10
0.6
0.4
0.2
0
(a) Prediction error
5
0.0
10
(b) Action matrices
Figure 2: (a) Test error on final distribution and mean over entire trajectory (log scale). MFG: (2.9e-3, 4.9e-3), VAR:
(7.0e-3, 8.1e-3), RNN: (0.58, 0.57). (b) heatmap of action matrix P ∈ R15×15 averaged element-wise over demo train
set, and absolute difference between average demo action matrix and average matrix generated from learned policy.
Topic 2 measurement and predictions
1.00
0.05
0.99
0.04
Topic 2 popularity
Topic 0 popularity
Topic 0 measurement and predictions
0.98
0.97
0.96
0.95
0
1
2
3
Day
4
test data
MFG (test)
VAR (test)
5
6
test data
MFG (test)
VAR (test)
RNN (test)
0.03
0.02
0.01
0.00
0
(a) Topic 0 test trajectory
1
2
3
Day
4
5
6
(b) Topic 2 test trajectory
Figure 3: (a) Measured and predicted trajectory of topic 0 popularity over test days for MFG and VAR (RNN outside
range and not shown). (b) Measured and predicted trajectory of topic 2 popularity over test days for all methods.
PN −1
n
n
evaluated on the average JSD over all hours in a day N1 n=0 JSD(πgenerated
, πmeasured
). Both measures were averaged
over M = 6 held-out test trajectories. It is worth emphasizing that learning the MFG model required only the initial
population distribution of each day in the training set, while VAR and RNN used the distributions over all hours
of each day. Even with much fewer training samples, MFG achieves excellent prediction performance because it
represents the underlying optimization processes conducted by large populations, unlike the simple models of VAR
and RNN. As shown by sample trajectories for topic 0 and 2 in Figures 3, and the average transition matrices in
Figure 2b, MFG correctly represents the fact that the real population tends to congregate to topics with higher initial
popularity (i.e. lower topic indices), and that the popularity of topic 0 becomes more dominant across time in each
day. The small real-world dataset size, and the fact that RNN mainly learns state transitions without accounting
for actions, could be contributing factors to lower performance of RNN compared to MFG. We acknowledge that our
design of policy parameterization, although informed by domain knowledge, introduced bias and resulted in noticeable
differences between demonstration and generated transition matrices. This can be addressed using deep policy and
value networks, since the MFG framework is agnostic towards choice of policy representation.
5.3
Insights
The learned reward function reveals that a social media population favors states characterized by a highly non-uniform
distribution with negative mass gradient in decreasing order of topic popularity, as well as transitions that increase
this imbalance. The high prediction accuracy of the learned policy provides evidence that population behavior can be
understood and modeled as the result of population-level optimization with respect to a reward function.
8
6
Conclusion
We have motivated and demonstrated a data-driven method to solve a mean field game model of population evolution,
by proving a connection to Markov decision processes and building on methods in reinforcement learning. Our
method is scalable to arbitrarily large populations, because the MFG framework represents population density rather
than individual agents, while the representations are linear in the number of MFG states and quadratic in the transition
matrix. Our real-world experiments show that MFG is a powerful framework for learning both the underlying reward
function being optimized by a real world population and a policy that is able to predict future population trajectories
more accurately than alternatives. Even with a simple policy parameterization designed via some domain knowledge,
our method attains superior performance on test data. It motivates exploration of flexible neural networks for more
complex applications.
An interesting extension is to develop an efficient method for solving the discrete time MFG in a more general setting,
where the reward at each state i is coupled to the full population transition matrix. Our work also opens the path to a
variety of real-world applications, such as a synthesis of MFG with models of social networks at the level of individual
connections to construct a more complete model of social dynamics, and mean field models of interdependent systems
that may display complex interactions via coupling through global states and reward functions.
References
Achdou, Y., Camilli, F., and Capuzzo-Dolcetta, I. (2012). Mean field games: numerical methods for the planning
problem. SIAM Journal on Control and Optimization, 50(1), 77–109.
Anderson, M. and Hitlin, P. (2016). Social Media Conversations About Race. Pew Research Center.
Bauso, D., Pesenti, R., and Tolotti, M. (2016). Opinion dynamics and stubbornness via multi-population mean-field
games. Journal of Optimization Theory and Applications, 170(1), 266–293.
Boularias, A., Kober, J., and Peters, J. (2011). Relative entropy inverse reinforcement learning. In Proceedings of the
Fourteenth International Conference on Artificial Intelligence and Statistics, pages 182–189.
Carlini, E. and Silva, F. J. (2014). A fully discrete semi-lagrangian scheme for a first order mean field game problem.
SIAM Journal on Numerical Analysis, 52(1), 45–67.
De, A., Valera, I., Ganguly, N., Bhattacharya, S., and Rodriguez, M. G. (2016). Learning and forecasting opinion
dynamics in social networks. In Advances in Neural Information Processing Systems, pages 397–405.
Farajtabar, M., Wang, Y., Rodriguez, M. G., Li, S., Zha, H., and Song, L. (2015). Coevolve: A joint point process
model for information diffusion and network co-evolution. In Advances in Neural Information Processing Systems,
pages 1954–1962.
Finn, C., Levine, S., and Abbeel, P. (2016). Guided cost learning: Deep inverse optimal control via policy optimization.
In International Conference on Machine Learning, pages 49–58.
Gomes, D. A., Mohr, J., and Souza, R. R. (2010). Discrete time, finite state space mean field games. Journal de
mathématiques pures et appliquées, 93(3), 308–328.
Guéant, O. (2009). A reference case for mean field games models. Journal de mathématiques pures et appliquées,
92(3), 276–294.
Guéant, O. (2015). Existence and uniqueness result for mean field games with congestion effect on graphs. Applied
Mathematics & Optimization, 72(2), 291–303.
Guéant, O., Lasry, J.-M., and Lions, P.-L. (2011). Mean field games and applications. In Paris-Princeton lectures on
mathematical finance 2010, pages 205–266. Springer.
Howard, P. N., Duffy, A., Freelon, D., Hussain, M. M., Mari, W., and Maziad, M. (2011). Opening closed regimes:
what was the role of social media during the arab spring?
9
Hu, J., Wellman, M. P., et al. (1998). Multiagent reinforcement learning: theoretical framework and an algorithm. In
ICML, volume 98, pages 242–250. Citeseer.
Huang, M., Malhamé, R. P., Caines, P. E., et al. (2006). Large population stochastic dynamic games: closed-loop
mckean-vlasov systems and the nash certainty equivalence principle. Communications in Information & Systems,
6(3), 221–252.
Jaynes, E. T. (1957). Information theory and statistical mechanics. Physical review, 106(4), 620.
Kalakrishnan, M., Pastor, P., Righetti, L., and Schaal, S. (2013). Learning objective functions for manipulation. In
Robotics and Automation (ICRA), 2013 IEEE International Conference on, pages 1331–1336. IEEE.
Konda, V. R. and Tsitsiklis, J. N. (2000). Actor-critic algorithms. In Advances in neural information processing
systems, pages 1008–1014.
Lachapelle, A., Salomon, J., and Turinici, G. (2010). Computation of mean field equilibria in economics. Mathematical Models and Methods in Applied Sciences, 20(04), 567–588.
Lasry, J.-M. and Lions, P.-L. (2007). Mean field games. Japanese journal of mathematics, 2(1), 229–260.
Lillicrap, T. P., Hunt, J. J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., and Wierstra, D. (2015). Continuous
control with deep reinforcement learning. arXiv preprint arXiv:1509.02971.
Littman, M. L. (2001). Value-function reinforcement learning in markov games. Cognitive Systems Research, 2(1),
55–66.
Lowe, R., Wu, Y., Tamar, A., Harb, J., Abbeel, P., and Mordatch, I. (2017). Multi-agent actor-critic for mixed
cooperative-competitive environments. arXiv preprint arXiv:1706.02275.
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., and Riedmiller, M. (2013). Playing
atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602.
Ng, A. Y. and Russell, S. (2000). Algorithms for inverse reinforcement learning. In in Proc. 17th International Conf.
on Machine Learning, pages 663–670. Morgan Kaufmann.
Owen, A. and Zhou, Y. (2000). Safe and effective importance sampling. Journal of the American Statistical Association, 95(449), 135–143.
Perrin, A. (2015). Social Media Usage: 2005-2015. Pew Research Center.
Silver, D., Lever, G., Heess, N., Degris, T., Wierstra, D., and Riedmiller, M. (2014). Deterministic policy gradient
algorithms. In Proceedings of the 31st International Conference on Machine Learning (ICML-14), pages 387–395.
Silverman, C. (2016). This analysis shows how viral fake election news stories outperformed real news on facebook.
Sutton, R. S. and Barto, A. G. (1998). Introduction to Reinforcement Learning. MIT Press, Cambridge, MA, USA,
1st edition.
Twitter (2017). Faqs about trends on twitter.
Wulfmeier, M., Ondruska, P., and Posner, I. (2015). Maximum entropy deep inverse reinforcement learning. arXiv
preprint arXiv:1507.04888.
Ziebart, B. D., Maas, A. L., Bagnell, J. A., and Dey, A. K. (2008). Maximum entropy inverse reinforcement learning.
In AAAI, volume 8, pages 1433–1438. Chicago, IL, USA.
10
A
Proof of Propostion 1
Given the definitions in Section 3.1, a mean field game is defined by a Hamilton-Jacobi-Bellman (HJB) equation
evolving backwards in time and a Fokker-Planck equation evolving forward in time. The continuous-time HamiltonJacobi-Bellman (HJB) equation on G is
X
0
Vi (t) = − maxPi
Pij (t)(Vj (t) − Vi (t)) + ri (π(t), Pi (t))
(15)
−
j∈V̄i
where ri (π, Pi ) is the reward function, and Vi (t) is the value function of state i at time t. Note that the reward function
ri (π(t), Pi (t)) is often presented as −ci (π(t), Pi (t)) for some cost function ci (π(t), Pi (t)) in the MFG context, and
similarly for Vi (t). In addition, we set ri (π(t), Pi (t)) = −∞ if Pi (t) ∈
/ Si (G) (i.e. P (t) must be a valid transition
matrix). For any fixed π(t), let Hi (π(t), ·) be the Legendre transform of ci (π(t), ·) defined by
Hi (π(t), ·) = maxPi {h·, Pi i − ci (π(t), Pi )} = maxPi {h·, Pi i + ri (π(t), Pi )}
(16)
Then the HJB equation (15) is an analogue to the backward equation in mean field games
Vi0 (t) + Hi (π(t), [Vj (t) − Vi (t)]j∈V̄ − ) = 0
(17)
i
−
where [Vj (t) − Vi (t)]j∈V̄ − ∈ R|V̄i | is the dual variable of Pi . We can discretize (15) using a semi-implicit scheme
i
with unit time step labeled by n to obtain
X
n+1
n+1
n
n
n
Vin+1 − Vin = − maxPi
P
(V
−
V
)
+
r
(π
,
P
)
(18)
i
ij
i
−
j
i
j∈V̄i
Rearranging (18) yields the discrete time HJB equation over a graph (19)
X
n
n
n
n n+1
Vi = maxPi ri (π , Pi ) +
Pij Vj
−
(19)
The forward evolving Fokker-Planck equation for the continuous-time graph-state MFG is given by
X
X
πi0 (t) =
Qji (t)πj (t) −
Qij (t)πi (t)
+
−
(20)
j∈V̄i
j∈Vi
j∈Vi
where Qji (t) = ∂ui Hj (π(t), [Vk (t) − Vj (t)]k∈V̄ − )
j
(21)
where ∂ui Hj (π, u) is the partial derivative w.r.t. the coordinate corresponding to the i-th index of the argument
−
u ∈ R|V̄j | . We can set Qji (t) = 0 for all (j, i) ∈
/ E, so that Q(t) := [Qji (t)] can be regarded as the d-by-d
infinitesimal generator matrix of states π(t), and hence (20) can be written as π 0 (t) = π(t)Q(t), where π(t) ∈ Rd is
a row vector. Then an Euler discretization of (20) with unit time step reduces to π n+1 − π n = π n Qn , which can be
written as
X
n n
πin+1 =
Pji
πj
(22)
+
j∈V̄i
where Pijn := Qnij + δij . If the graph G is complete, meaning E = {(i, j) : 1 ≤ i, j ≤ d}, then the summation is
taken over j = 1, . . . , d. For ease of presentation, we only consider the complete graph in this paper, as all derivations
can be carried out similarly for general directed graphs. A solution of a mean field game defined by (19) and (22) is a
collection of Vin and πin for i = 0, . . . , d − 1 and n = 0, . . . , N − 1.
11
B
Algorithms
We learn a reward function and policy using an adaptation of GCL (Finn et al., 2016) in Alg 1 and a simple actor-critic
Alg 2 (Sutton and Barto, 1998) as a forward RL solver.
Algorithm 1 Guided cost learning
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
procedure G UIDED COST LEARNING
Initialize F0 (P ; π, θ) as random policy and reward network weights W 0
for iteration 1 to I do
Generate sample trajectories Dtraj from Fk (P ; π, θ)
Dsamp ← Dsamp ∪ Dtraj
while Avgπi ,Pi ∼Ddemo (RW t (πi , Pi ) − RW t−1 (πi , Pi )) > dR do
Sample demonstration D̂demo ⊂ Ddemo from expert demonstration
Sample D̂samp ⊂ Dsamp
W t+1 ← W t − ∇L(W t ) using D̂demo and D̂samp
end while
Run Alg 2 on Dtraj for improved Fk+1
end for
return Final reward function RW (π, P ) and policy F (P ; π, θ)
end procedure
Algorithm 2 Actor-critic algorithm for MFG
Input: Generative model F (P ; π, θ), value function V̂ (π; w), training data {π 0 }M days
Output: Policy parameter θ, value function parameter w
1: procedure ACTOR - CRITIC -MFG(F, V̂ , {π 0 }M days , β, ξ, RW )
2:
initialize θ and w
3:
for episodes s = 1, . . . , S do
4:
Sample initial distribution π 0 from {π 0 }M days
5:
for time step n = 0, . . . , N − 1 do
6:
Sample action P n ∼ F (P ; π n , θ)
7:
Generate π n+1 using Eq 3
8:
Receive reward RW (π n , P n )
9:
δ ← R + V̂ (π n+1 ; w) − V̂ (π n ; w)
10:
w ← w + ξδ∇w V̂ (π n ; w)
11:
θ ← θ + βδ∇θ log(F (P ; π n , θ))
12:
end for
13:
end for
14: end procedure
C
Reward network
Our reward network uses two convolutional layers to process the 15 × 15 action matrix P , which is then flattened and
concatenated with the state vector π and processed by two fully-connected layers regularized with L1 and L2 penalties
and dropout (probability 0.6). The first convolutional layer zero-pads the input into a 19×19 matrix and convolves one
filter of kernel size 5 × 5 with stride 1 and applies a rectifier nonlinearity. The second convolutional layer zero-pads its
input into a 17 × 17 matrix and convolves 2 filters of kernel size 3 × 3 with stride 1 and applies a rectifier nonlinearity.
The fully connected layers have 8 and 4 hidden rectifier units respectively, and the output is a single fully connected
tanh unit. All layers were initialized using the Xavier normal initializer in Tensorflow.
12
D
Experiment details
By default, Twitter users in a certain geographical region primarily see the trending topics specific to that region
(Twitter, 2017). This experiment focused on the population and trending topics in the city of Atlanta in the U.S. state
of Georgia. First, a set of 406 active users were collected to form the fixed population. This was done by collecting
a set of high-visibility accounts in Atlanta (e.g. the Atlanta Falcons team), gathering all Twitter users who follow
these accounts, filtering for those whose location was set to Atlanta, and filtering for those who responded to least two
trending topics within four days.
Data collection proceeded as follows for 27 days: at 9am of each day, a list of the top 14 trending topics on Twitter
in Atlanta was recorded; for each hour until midnight, for each topic, the number of users who responded to the topic
within the past hour was recorded. Whether or not a user responded to a topic was determined by checking for posts
by the user containing unique words for that topic; the “hashtag” convention of trending topics on Twitter reduces
the likelihood of false positives. The hourly count of people who did not respond to any topic was recorded as the
count for a “null topic”. Although some users may respond to more than one topic within each hour, the data shows
that this is negligible, and a shorter time interval can be used to reduce this effect. The result of data collection is
a set of trajectories, one trajectory per day, where each trajectory consists of hourly measurements of the population
distribution over d = 15 topics over N = 16 hours.
The training set consists of trajectories {π 0,m , . . . , π N −1,m }m=1,...,M over the first M = 21 days. MFG uses the
initial distribution π 0 of each day for training (Alg 2 line 4), while VAR and RNN use all measured distributions.
RNN method employs a simple recurrent unit with ReLU as nonlinear activation and weight matrix of dimension
d × d. Table 1 shows parameters of Alg 2 and 1.
Table 1: Parameters
E
Parameter
Use
Value
S
β
ξ
c
dR
θfinal
max actor-critic episodes
critic learning rate
actor learning rate
αji scaling factor
Adam optimizer learning rate for reward
convergence threshold for reward iteration
learned policy parameter
4000
O(1/s)
O(1/s ln ln s)
1e4
1e-4
1e-4
8.64
Maximum entropy distribution
Given a finite set of trajectories {τi }i , where each trajectory is a sequence of state-action pairs τi = (si1 , ai1 , . . . , ).
Suppose
each trajectory τi has an unknown probability pi . The entropy of the probability distribution is H =
P
− i pi ln(pi ). In the continuous case, we write the differential entropy:
Z
H = − p(τ ) ln(p(τ ))dτ
where p(·) is the probability density we want to derive. The constraints are:
Z
r(τ )p(τ )dτ = E[r(τ )] = µr
Z
p(τ )dτ = 1
13
The first constraint says: the expected reward over all trajectories is equal to an empirical measurement µr . We write
the Lagrangian L:
Z
Z
Z
r(τ )p(τ )dτ − µr
L = − p(τ ) ln(p(τ ))dτ − λ1
p(τ )dτ − 1 − λ2
For L to be stationary, the Euler-Lagrange equation with integrand denoted by L says
∂L
=0
∂p
since L does not depend on
dp
dτ .
Hence
Z
−λ2 r(τ )
λ1 = ln
e
dτ − 1
Z
−λ2 r(τ )
e
p(τ ) = exp − ln
dτ − λ2 r(τ ) =
where Z :=
F
R
1
e−λ2 r(τ )
Z(λ2 )
e−λ2 r(τ ) dτ . Then the constant λ2 is determined by:
Z
Z
1
e−λ2 r(τ ) r(τ )dτ
µr = p(τ )r(τ )dτ =
Z(λ2 )
∂
=−
ln(Z(λ2 ))
∂λ2
Multiple importance sampling
We show how multiple importance sampling (Owen and Zhou, 2000) can beR used to estimate the partition function in
the maximum entropy IRL framework. The problem is to estimate Z := f (x)dx. Let p1 , . . . , pm be m proposal
distributions, with nj samples from the j-th proposal distribution, so that samples can be denoted Xij for i = 1, . . . , nj
and j = 1, . . . , m. Let wj (x) for j = 1, . . . , m satisfy
0 ≤ wj (x) ≤
m
X
wj (x) = 1
j=1
Then define the estimator
Ẑ =
nj
m
X
f (Xij )
1 X
wj (Xij )
n
pj (Xij )
j=1 j i=1
Let S(pj ) = {x | pj (x) > 0} be the support of pj and S(wj ) = {x | wj (x) > 0} be the support of wj , and let them
satisfy S(wj ) ⊂ S(pj ). Under these assumptions:
Z
E[Ẑ] = f (x)dx = Z
In particular, choose
nj pj (x)
wj (x) := Pm
k=1 nk pk (x)
Then the estimate becomes
Ẑ =
nj
m X
X
f (Xji )
Pm
k=1 nk pk (x)
j=1 i=1
m nj
=
1 XX
f (X )
Pm nkji
n j=1 i=1 k=1 n pk (x)
14
Pm
where n =
j=1 nj is the total count of samples. Further assuming that samples are drawn uniformly from all
proposal distributions, so that nj = nk = n/m for all j, k ∈ {1, . . . , m}, the expression for Ẑ reduces to the form
used in Eq 13:
f (x)
1 X
Pm
Ẑ =
1
n
k=1 pk (x)
m
all samples
G
A comparison of mean field games and multi-agent MDPs
In this section, we discuss the reason that the general MFG is not reducible to a collection of distinct single-agent
MDPs, and also not equivalent to a multi-agent MDP. Assume the case of a general reward function rij (π n , P n ) that
depends on the full Nash maximizer matrix P n .
G.1
Collection of single-agent MDPs
Consider each topic i as a separate entity associated with a value, rather than subsuming it into an average (as is the
case in Section 4). In order to assign a value to each topic, each tuple (i, π n ) must be defined as a state, which leads to
the problem: since a state requires specification of π n , and state transitions depend on the actions for all other topics,
the action at each topic is not sufficient for fully specifying the next state. More formally, consider a value function on
the state:
X
X
n
n
n
n T n
V (i, π ) = max
qj rij (π , P(P , i, q)) +
qj V (j, (P ) π )
(23)
q∈Si (G)
j
j
Superficially, this resembles the Bellman optimality equation for the value function in a single-agent stochastic MDP,
where s is a state, a is an action, R is an immediate reward, and P (s0 |s, a) is the probability of transition to state s0
from state s, given action a:
X
V ∗ (s) = max{R(s, a) +
P (s0 |s, a)V ∗ (s0 )}
(24)
a
s0
In the second summation, qj can be interpreted as a transition probability, conditioned on the fact that the current topic
is i. The action q selected in the state (i, π n ) induces a stochastic transition to a next topic j, but the next distribution
π n+1 is given by the deterministic forward equation π n+1 = (P n )T π n , where P n is the true Nash maximizer matrix.
This means that qj does not completely specify the next state (j, π n+1 ), and there is a formal difference between
P (s0 |s, a)V∗ (s0 ) and qj V (j, (P n )T π n ). Also notice that the Bellman equation sums over all possible next states s0 ,
but that is not the case for equation 23, because a full state specification requires (j, π).
G.2
Multi-agent MDP
The following explanation shows that an exact reduction from the MFG to a multi-agent MDP (i.e. Markov game)
is not possible. A discrete state space discrete action space multi-agent MDP is defined by d agents moving within
a set S of environment states; a collection {A1 , . . . , Ad } of sets of actions, one for each agent; a transition function
P (s0 |s, a1 , . . . , ad ) giving the probability of the environment transitioning to state s0 given that the current state is s
and agents choose actions ā , (a1 , . . . , ad ); a collection of reward functions {Ri (s, a1 , . . . , ad )}i , one for each agent;
and a discount factor γ in the infinite horizon case.
One natural attempt to reduce the MFG to a Markov game is to let the set of π n (with appropriate discretization) be the
state space, limit the set of actions to some discretization of the simplex, and consider each topic to be a single agent.
Now, the agent representing topic i is no longer identified with the set of people who selected topic i: topics have fixed
15
labels for all time, so an agent who accumulates reward for topic i must be associated only with topic i. Therefore, the
value function for agent i is defined only in terms of itself, never depending on the value function of agents j 6= i:
!
XY
X
µ
µ
µj (āj |s) Ri (s, ā) + γ
Vi (s) =
P (s0 |s, ā)Vi (s0 )
(25)
ā
s0
j
where µ := (µ1 , . . . , µd ) is a set of stationary policies of all agents. However, recall that the MFG equation for Vin
explicitly depends on Vjn+1 of all topics j, which would require a different form such as the following:
Viµ (π n )
=
d
X Y
P ∈S(G) j=1
!
n
µj (Pj |π )
X
n
Pik rik (π , P ) +
k
X
Pik Vkµ (P T π n )
(26)
k
where the last summation involves the value function Vkµ for all k. This difference is exactly the reason that the MFG
equations cannot be reduced to the equations of a standard Markov game.
16
| 5cs.CE
|
Generators of the pro-p Iwahori and Galois representations
arXiv:1611.06084v1 [math.NT] 18 Nov 2016
Christophe Cornut1 and Jishnu Ray2
1
CNRS - Institut de Mathématiques de Jussieu - Paris Rive Gauche,
4, place Jussieu, 75252 Paris Cedex 05, France,
[email protected]
2
Département de Mathématiques, Université Paris-Sud 11,
91405 Orsay Cedex, France,
[email protected]
Abstract
For an odd prime p, we determine a minimal set of topological generators of the pro-p
Iwahori subgroup of a split reductive group G over Zp . In the simple adjoint case and for any
sufficiently large regular prime p, we also construct Galois extensions of Q with Galois group
between the pro-p and the standard Iwahori subgroups of G.
1
Introduction
Let p be an odd prime, let G be a split reductive group over Zp , fix a Borel subgroup B = U ⋊ T
of G with unipotent radical U ⊳ B and maximal split torus T ⊂ B. The Iwahori subgroup I and
pro-p-Iwahori subgroup I(1) ⊂ I of G(Zp ) are defined [13, 3.7] by
I = {g ∈ G(Zp ) : red(g) ∈ B(Fp )},
I(1) = {g ∈ G(Zp ) : red(g) ∈ U(Fp )}.
where ‘red’ is the reduction map red: G(Zp ) → G(Fp ). The subgroups I and I(1) are both open
subgroups of G(Zp ). Thus I = I(1) ⋊ Ttors and T(Zp ) = T (1) × Ttors where T (1) and Ttors are
respectively the pro−p and torsion subgroups of T(Zp ). Following [3] (who works with G = GLn ),
we construct in section 2 a minimal set of topological generators for I(1).
More precisely, let M = X ∗ (T) be the group of characters of T, R `
⊂ M the set of roots of T
in g = Lie(G), ∆ ⊂ R the set of simple roots with respect to B, R = c∈C Rc the decomposition
of R into irreducible components, ∆c = ∆ ∩ Rc the simple roots in Rc , αc,max the highest positive
root in Rc . We let D ⊂ C be the set of irreducible components of type G2 and for d ∈ D, we
denote by δd ∈ Rd,+ the sum of the two simple roots in ∆d . We denote by M ∨ = X∗ (T) the group
of cocharacters of T, by ZR∨ the subgroup spanned by the coroots R∨ ⊂ M ∨ and we fix a set of
representatives S ⊂ M ∨ for an Fp -basis of
(M ∨ /ZR∨ ) ⊗ Fp = ⊕s∈S Fp · s ⊗ 1.
1
2
We show (see theorem 2.4.1):
Theorem. The following elements form a minimal set of topological generators of the pro-pIwahori subgroup I(1) of G = G(Qp ):
1. The semi-simple elements {s(1 + p) : s ∈ S} of T (1),
2. For each c ∈ C, the unipotent elements {xα (1) : α ∈ ∆c },
3. For each c ∈ C, the unipotent element x−αc,max (p),
4. (If p = 3) For each d ∈ D, the unipotent element xδd (1).
This result generalizes Greenberg [3] proposition 5.3, see also Schneider and Ollivier ([9], proposition
3.64, part i) for G = SL2 .
Let Tad be the image of T in the adjoint group Gad of G. The action of Gad on G induces an ac˜ of I(1) with a structure
tion of Tad (Zp ) on I and I(1) and the latter equips the Frattini quotient I(1)
ad
ad
ad
of Fp [Ttors ]-module, where Ttors is the torsion subgroup of T (Zp ) (cf. section 2.12). Any element
tors
β in ZR = M ad = X ∗ (Tad ) induces a character β : Tad
→ F×
p and we denote by Fp (β) the corad
responding simple (1-dimensional) Fp [Ttors ]-module. With these notations, the theorem implies that
ad
˜ is isomorphic to
Corollary. The Fp [Ttors
]-module I(1)
F♯S
⊕
⊕
F
(α)
⊕
⊕
F
(−α
)
⊕
⊕
F
(δ
)
if
p
=
3
.
α∈∆
p
c∈C
p
c,max
d∈D
p
c
p
Here ♯S is the cardinality of S. Suppose from now on in this introduction that G is simple and
of adjoint type. Then:
˜ is multiplicity free unless p = 3 and G is of type A1 , Bℓ
Corollary The Fp [Ttors ]-module I(1)
or Cℓ (ℓ ≥ 2), F4 or G2 .
Let now K be a Galois extension of Q, Σp the set of primes of K lying above p. Let M be the
compositum of all finite p-extensions of K which are unramified outside Σp , a Galois extension over
Q. Set Γ = Gal(M/K), Ω = Gal(K/Q) and Π = Gal(M/Q). We say that K is p-rational if Γ is a
free pro−p group, see [6]. The simplest example is K = Q, where Γ = Π is also abelian and M is
the cyclotomic Zp -extension of Q. Other examples of p-rational fields are Q(µp ) where p is a regular
prime.
Assume K is a p-rational, totally complex, abelian extension of Q and (p − 1) · Ω = 0. Then
Greenberg in [3] constructs a continuous homomorphism
ρ0 : Gal(M/Q) → GLn (Zp )
such that ρ0 (Γ) is the pro-p Iwahori subgroup of SLn (Zp ), assuming that there exists n distinct
characters of Ω, trivial or odd, whose product is the trivial character.
In section 3, we are proving results which show the existence of p-adic Lie extensions of Q where
the Galois group corresponds to a certain specific p-adic Lie algebra. More precisely, for p-rational
fields, we construct continuous morphisms with open image ρ : Π → I such that ρ(Γ) = I(1). We
3
show in corollary 3.3.1 that
Corollary Suppose that K is a p-rational totally complex, abelian extension of Q and (p−1)·Ω =
0. Assume also that if p = 3, our split simple adjoint group G is not of type A1 , Bℓ or Cℓ (ℓ ≥ 2),
F4 or G2 . Then there is a morphism ρ : Π → I such that ρ(Γ) = I(1) if and only if there is
morphism ρ : Ω → Ttors such that the characters α ◦ ρ : Ω → F×
p for α ∈ {∆ ∪ −αmax } are all
S
distinct and belong to Ω̂odd .
Here Ω̂Sodd is a subset of the characters of Ω with values in F×
p (it is defined after proposition
3.2.1). Furthermore assuming K = Q(µp ) we show the existence of such a morphism ρ : Ω → Ttors
provided that p is a sufficiently large regular prime (cf. section 3.2):
Corollary There is a constant c depending only upon the type of G such that if p > c is a regular prime, then for K = Q(µp ), M, Π and Γ as above, there is a continuous morphism ρ : Π → I
with ρ(Γ) = I(1).
The constant c can be determined from lemmas 3.4.1, 3.4.2 and remark 3.4.3.
In section 2, we find a minimal set of topological generators of I(1) and study the structure
ad
˜
of I(1)
as an Fp [Ttors
]-module. In section 3, assuming our group G to be simple and adjoint, we
discuss the notion of p-rational fields and construct continuous morphisms ρ : Π → I with open
image.
We would like to thank Marie-France Vignéras for useful discussions and for giving us the
reference [9]. We are also deeply grateful to Ralph Greenberg for numerous conversations on this
topic.
2
Topological Generators of the pro-p Iwahori
This section is organized as follows. In sections (2.1 − 2.3) we introduce the notations, then section
2.4 states our main result concerning the minimal set of topological generators of I(1) (see theorem
2.4.1) with a discussion of the Iwahori factorisation in section 2.5. Its proof for G simple and simply
connected is given in sections (2.6 − 2.10), where section 2.10 deals with the case of a group of type
G2 . The proof for an arbitrary split reductive group over Zp is discussed in sections (2.11 − 2.14).
In particular, section 2.14 establishes the minimality of our set of topological generators. Finally,
ad
˜
in section 2.15 we study the structure of the Frattini quotient I(1)
of I(1) as an Fp [Ttors
]-module
and determine the cases when it is multiplicity free.
2.1
1]
Let p be an odd prime, G be a split reductive group over Zp . Fix a pinning of G [11, XXIII
(T, M, R, ∆, (Xα )α∈∆ ) .
Thus T is a split maximal torus in G, M = X ∗ (T) is its group of characters,
g = g0 ⊕ ⊕α∈R gα
is the weight decomposition for the adjoint action of T on g = Lie(G), ∆ ⊂ R is a basis of the root
system R ⊂ M and for each α ∈ ∆, Xα is a Zp -basis of gα .
4
2.2
We denote by M ∨ = X∗ (T) the group of cocharacters of T, by α∨ the coroot associated to
α ∈ R and by R∨ ∈ M ∨ the set of all such coroots. We expand (Xα )α∈∆ to a Chevalley system
(Xα )α∈R of G [11, XXIII 6.2]. For α ∈ R, we denote by Uα ⊂ G the corresponding unipotent
group, by xα : Ga,Zp → Uα the isomorphism given by xα (t) = exp(tXα ). The height h(α) ∈ Z of
α ∈ R is the sum of the coefficients of α in the basis ∆ of R. Thus R+ = h−1 (Z>0 ) is the set of
positive roots in R, corresponding to a Borel subgroup B = U ⋊ T of G with unipotent radical U.
We let C be the set of irreducible components of R, so that
a
a
a
R=
Rc , ∆ =
∆c , R+ =
Rc,+
c∈C
c∈C
c∈C
with Rc irreducible, ∆c = ∆ ∩ Rc is a basis of Rc and Rc,+ = R+ ∩ Rc is the corresponding set of
positive roots in Rc . We denote by αc,max ∈ Rc,+ the highest root of Rc . We let D ⊂ C be the set
of irreducible components of type G2 and for d ∈ D, we denote by δd ∈ Rd,+ the sum of the two
simple roots in ∆d .
2.3
Since G is smooth over Zp , the reduction map
red : G(Zp ) → G(Fp )
is surjective and its kernel G(1) is a normal pro-p-subgroup of G(Zp ). The Iwahori subgroup I and
pro-p-Iwahori subgroup I(1) ⊂ I of G(Zp ) are defined [13, 3.7] by
I = {g ∈ G(Zp ) : red(g) ∈ B(Fp )} ,
I(1) = {g ∈ G(Zp ) : red(g) ∈ U(Fp )} .
Thus I(1) is a normal pro-p-sylow subgroup of I which contains U(Zp ) and
I/I(1) ≃ B(Fp )/U(Fp ) ≃ T(Fp ).
Since T(Zp ) ։ T(Fp ) is split by the torsion subgroup Ttors ≃ T(Fp ) of T(Zp ),
T(Zp ) = T (1) × Ttors
and
I = I(1) ⋊ Ttors
where
T (1) = T(Zp ) ∩ I(1) = ker (T(Zp ) → T(Fp ))
is the pro-p-sylow subgroup of T(Zp ). Note that
T (1) = Hom (M, 1 + pZp ) = M ∨ ⊗ (1 + pZp ),
Ttors = Hom (M, µp−1) = M ∨ ⊗ F×
p.
2.4
Let S ⊂ M ∨ be a set of representatives for an Fp -basis of
(M ∨ /ZR∨ ) ⊗ Fp = ⊕s∈S Fp · s ⊗ 1.
Theorem 2.4.1. The following elements form a minimal set of topological generators of the prop-Iwahori subgroup I(1) of G = G(Qp ):
1. The semi-simple elements {s(1 + p) : s ∈ S} of T (1).
2. For each c ∈ C, the unipotent elements {xα (1) : α ∈ ∆c }.
3. For each c ∈ C, the unipotent element x−αc,max (p).
4. (If p = 3) For each d ∈ D, the unipotent element xδd (1).
5
2.5
By [11, XXII 5.9.5] and its proof, there is a canonical filtration
U = U1 ⊃ U2 ⊃ · · · ⊃ Uh ⊃ Uh+1 = 1
of U by normal subgroups such that for 1 ≤ i ≤ h, the product map (in any order)
Y
Uα → U
h(α)=i
factors through Ui and yields an isomorphism of group schemes
Y
≃
Uα −→ Ui , Ui = Ui /Ui+1 .
h(α)=i
By [11, XXII 5.9.6] and its proof,
Ui (R) = Ui (R)/Ui+1 (R)
for every Zp -algebra R. It follows that the product map
Y
Uα × Ui+1 → Ui
h(α)=i
is an isomorphism of Zp -schemes and by induction, the product map
Y
Y
Y
Uα ×
Uα × · · · ×
Uα → U
h(α)=1
h(α)=2
h(α)=h
is an isomorphism of Zp -schemes. Similarly, the product map
Y
Y
Y
Uα ×
Uα × · · · ×
Uα → U−
h(α)=−h
h(α)=−h+1
h(α)=−1
is an isomorphism of Zp -schemes, where U− is the unipotent radical of the Borel subgroup B− =
U− ⋊ T opposed to B with respect to T. Then by [11, XXII 4.1.2], there is an open subscheme Ω
of G (the “big cell”) such that the product map
U− × T × U → G
is an open immersion with image Ω. Plainly, B = U ⋊ T is a closed subscheme of Ω. Thus by
definition of I, I ⊂ Ω(Zp ) and therefore any element of I (resp. I(1)) can be written uniquely as a
product
Y
Y
Y
Y
xα (aα ) × · · · ×
xα (aα ) × t ×
xα (aα ) × · · · ×
xα (aα )
h(α)=−h
h(α)=−1
h(α)=1
h(α)=h
where aα ∈ Zp for α ∈ R+ , aα ∈ pZp for α ∈ R− = −R+ and t ∈ T(Zp ) (resp. T (1)). This is
the Iwahori decomposition of I (resp. I(1)). If I + is the group spanned by {xα (Zp ) : α ∈ R+ } and
I − is the group spanned by {xα (pZp ) : α ∈ R− }, then I + = U(Zp ), I − ⊂ U− (Zp ) and every x ∈ I
(resp. I(1)) has a unique decomposition x = u− tu+ with u± ∈ I ± and t ∈ T(Zp ) (resp. t ∈ T (1)).
6
∨
∨
2.6
Suppose first that G is semi-simple and simply connected.
` Then M = ZR , thus S = ∅.
Moreover, everything splits according to the decomposition R = Rc :
Y
Y
Y
Y
Y
G=
Gc , T =
Tc , B =
Bc , I =
Ic and I(1) =
Ic (1).
To establish the theorem in this case, we may thus furthermore assume that G is simple. From now
on until section 2.11, we therefore assume that
G is (split) simple and simply connected.
2.7
As a first step, we show that
Lemma 2.7.1. The group generated by I + and I − contains T (1).
Proof. Since G is simply connected,
Y
α∨ :
α∈∆
Y
Gm,Zp → T
α∈∆
is an isomorphism, thus
Tc (1) =
Y
α∨ (1 + pZp ).
α∈∆
Now for any α ∈ ∆, there is a unique morphism [11, XX 5.8]
fα : SL(2)Zp → G
such that for every u, v ∈ Zp and x ∈ Z×
p,
fα
1 u
0 1
= xα (u),
fα
1 0
v 1
= x−α (v) and fα
x 0
0 x−1
= α∨(x).
Since for every x ∈ 1 + pZp [11, XX 2.7],
x 0
1 −x−1
1
0
1 1
1
0
=
0 x−1
0
1
x−1 1
0 1
x−1 − 1 1
in SL(2)(Zp ), it follows that α∨ (1 + pZp ) is already contained in the subgroup of G(Zp ) generated
by xα (Z×
p ) and x−α (pZp ). This proves the lemma.
2.8
Recall from [11, XXI 2.3.5] that for any pair of non-proportional roots α 6= ±β in R, the
set of integers k ∈ Z such that β + kα ∈ R is an interval of length at most 3, i.e. there are integers
r ≥ 1 and s ≥ 0 with r + s ≤ 4 such that
R ∩ {β + Zα} = {β − (r − 1)α, · · · , β + sα}.
The above set is called the α-chain through β and any such set is called a root chain in R. Let
k−k : R → R+ be the length function on R.
7
Proposition 2.8.1. Suppose kαk ≤ kβk. Then for any u, v ∈ Ga the commutator
[xβ (v) : xα (u)] = xβ (v)xα (u)xβ (−v)xα (−u)
is given by the following table, with (r, s) as above:
(r, s)
(−, 0)
(1, 1)
(1, 2)
(1, 3)
(2, 1)
(2, 2)
(3, 1)
[xβ (v) : xα (u)]
1
xα+β (±uv)
xα+β (±uv) · x2α+β (±u2 v)
xα+β (±uv) · x2α+β (±u2 v) · x3α+β (±u3 v) · x3α+2β (±u3 v 2 )
xα+β (±2uv)
xα+β (±2uv) · x2α+β (±3u2 v) · xα+2β (±3uv 2 )
xα+β (±3uv)
The signs are unspecified, but only depend upon α and β.
Proof. This is [11, XXIII 6.4].
Corollary 2.8.2. If r + s ≤ 3 and α + β ∈ R (i.e. s ≥ 1), then for any a, b ∈ Z, the subgroup of
G generated by xα (pa Zp ) and xβ (pb Zp ) contains xα+β (pa+b Zp ).
Proof. This is obvious if (r, s) = (1, 1) or (2, 1) (using p 6= 2 in the latter case). For the only
remaining case where (r, s) = (1, 3), note that
[xβ (v) : xα (u)][xβ (w 2 v) : xα (uw −1)]−1 = xα+β (±uv(1 − w)).
×
Since p 6= 2, we may find w ∈ Z×
p with (1 − w) ∈ Zp . Our claim easily follows.
Lemma 2.8.3. If R contains any root chain of length 3, then G is of type G2 .
Proof. Suppose that the α-chain through β has length 3. By [11, XXI 3.5.4], there is a basis ∆′
of R such that α ∈ ∆′ and β = aα + bα′ with α′ ∈ ∆′ , a, b ∈ N. The root system R′ spanned by
∆′ = {α, α′ } [11, XXI 3.4.6] then also contains an α-chain of length 3. By inspection of the root
systems of rank 2, for instance in [11, XXIII 3], we find that R′ is of type G2 . In particular, the
Dynkin diagram of R contains a triple edge (linking the vertices corresponding to α and α′ ), which
implies that actually R = R′ is of type G2 .
2.9
We now establish our theorem 2.4.1 for a group G which is simple and simply connected,
but not of type G2 .
Lemma 2.9.1. The group I + is generated by {xα (Zp ) : α ∈ ∆}.
Proof. Let H ⊂ I + be the group spanned by {xα (Zp ) : α ∈ ∆}. We show by induction on h(γ) ≥ 1
that xγ (Zp ) ⊂ H for every γ ∈ R+ . If h(γ) = 1, γ already belongs to ∆ and there is nothing to
prove. If h(γ) > 1, then by [1, VI.1.6 Proposition 19], there is a simple root α ∈ ∆ such that
β = γ − α ∈ R+ . Then h(β) = h(γ) − 1, thus by induction xβ (Zp ) ⊂ H. Since also xα (Zp ) ⊂ H,
xγ (Zp ) ⊂ H by Corollary 2.8.2.
Lemma 2.9.2. The group generated by I + and x−αmax (pZp ) contains I − .
8
Proof. Let H ⊂ I be the group spanned by I + and x−αmax (pZp ). We show by descending induction
on h(γ) ≥ 1 that x−γ (pZp ) ⊂ H for every γ ∈ R+ . If h(γ) = h(αmax ), then γ = αmax and there is
nothing to prove. If h(γ) < h(αmax ), then by [1, VI.1.6 Proposition 19], there is a pair of positive
roots α, β such that β = γ + α. Then h(β) = h(γ) + h(α) > h(γ), thus by induction x−β (pZp ) ⊂ H.
Since also xα (Zp ) ⊂ H, x−γ (pZp ) ⊂ H by Corollary 2.8.2.
Remark 2.9.3. From the Hasse diagrams in [10], it seems that in the previous proof, we may
always require α to be a simple root.
Proof. (Of theorem 2.4.1 for G simple, simply connected, not of type G2 ) By lemma 2.7.1, 2.9.1,
2.9.2 and the Iwahori decomposition of section 2.5, I(1) is generated by
{xα (Zp ) : α ∈ ∆} ∪ {x−αmax (pZp )}
thus topologically generated by
{xα (1) : α ∈ ∆} ∪ {x−αmax (p)} .
None of these topological generators can be removed: the first ones are contained in I + ( I(1), and
all of them are needed to span the image of
Y
I(1) ։ U(Fp ) ։ U1 (Fp ) ≃
Uα (Fp ),
α∈∆
a surjective morphism that kills x−αmax (p).
2.10
Let now G be simple of type G2 , thus ∆ = {α, β} with kαk < kβk and
R+ = {α, β, β + α, β + 2α, β + 3α, 2β + 3α}.
The whole root system looks like this:
3α + 2β
α+β
β
2α + β
3α + β
α
−α
−3α − β
−β
−2α − β
−α − β
−3α − 2β
Lemma 2.10.1. The group generated by I + and x−2β−3α (pZp ) contains I − .
Proof. Let H ⊂ I(1) be the group generated by I + and x−2β−3α (pZp ). Then, for every u, v ∈ Zp ,
H contains
[x−2β−3α (pv) : xβ (u)] =
[x−2β−3α (pv) : xβ+3α (u)] =
[x−2β−3α (pv) : xβ+2α (u)] =
x−β−3α (±puv)
x−β (±puv)
x−β−α (±puv) · xα (±pu2 v) · xβ+3α (±pu3 v) · x−β (±p2 u3 v 2 )
9
It thus contains x−β−3α (pZp ), x−β (pZp ) and x−β−α (pZp ), along with
[x−β−3α (pv) : xα (u)] =
[x−β−3α (pv) : xβ+2α (u)] =
x−β−2α (±puv) · x−β−α (±pu2 v) · x−β (±pu3 v) · x−2β−3α (±p2 u3 v 2 )
x−α (±puv) · xβ+α (±pu2 v) · x2β+3α (±pu3 v) · xβ (±p2 u3 v 2 )
It therefore also contains x−β−2α (pZp ) and x−α (pZp ).
The filtration (Ui )i≥1 of U in section 2.5 induces a filtration
I + = I1+ ⊃ · · · ⊃ I5+ ⊃ I6+ = 1
of I + = U(Zp ) by normal subgroups Ii+ = Ui (Zp ) whose graded pieces
+
+
I i = Ui (Zp ) = Ii+ /Ii+1
are free Zp -modules, namely
+
+
I 1 = Zp · xα ⊕ Zp · xβ ,
I 2 = Zp · xα+β
+
+
+
I 3 = Zp · x2α+β ,
I 4 = Zp · x3α+β ,
I 5 = Zp · x3α+2β
where xγ is the image of xγ (1). The commutator defines Zp -linear pairings
+
+
+
[−, −]i,j : I i × I j → I i+j
with [y, x]j,i = −[x, y]i,j , [x, x]i,i = 0 and, by Proposition 2.8.1,
[xβ , xα ] = ±xα+β , [xα+β , xα ] = ±2x2α+β , [x2α+β , xα ] = ±3x3α+β ,
[xα+β , x2α+β ] = ±x3α+2β and [xβ , x3α+β ] = ±x2α+2β
Let H be the subgroup of I + generated by xα (Zp ) and xβ (Zp ) and denote by Hi its image in
+
I + /Ii+1
= Gi . Then H1 = G1 , H2 contains [xβ , xα ] = ±xα+β thus H2 = G2 , H3 contains [xα+β , xα ] =
±2x2α+β thus H3 = G3 since p 6= 2, H4 contains [x2α+β , xα ] = ±3x3α+β thus H4 = G4 if p 6= 3, in
which case actually H = H5 = G5 = I + since H always contains [xα+β , x2α+β ] = ±x3α+2β .
If p = 3, let us also consider the exact sequence
+
0 → J4 → G4 → I 1 → 0
The group J4 = I2+ /I5+ is commutative, and in fact again a free Z3 -module:
J4 = (U2 /U5 )(Zp ) = Z3 x̃α+β ⊕ Z3 x̃2α+β ⊕ Z3 x3α+β
+
where x̃γ is the image of xγ (1). The action by conjugation of I 1 on J4 is given by
1
1
xβ 7→
1
xα 7→ ±2 1
±3 ±3 1
1
in the indicated basis of J4 . The Z3 -submodule H4′ = H4 ∩ J4 of J4 satisfies
H4′ + Z3 x3α+β = J4
and 3x3α+β ∈ H4′ .
10
Naming signs ǫi ∈ {±1} in formula (1, 3) of proposition 2.8.1, we find that H4′ contains
ǫ1 uv · x̃α+β + ǫ2 u2 v · x̃2α+β + ǫ3 u3 v · x3α+β
for every u, v ∈ Z3 . Adding these for v = 1 and u = ±1, we obtain
x̃2α+β ∈ H4′ .
It follows that H4′ actually contains the following Z3 -submodule of J4 :
J4′ = {a · x̃α+β + b · x̃2α+β + c · x3α+β : a, b, c ∈ Z3 , ǫ1 a ≡ ǫ3 c mod 3} .
Now observe that J4′ is a normal subgroup of G4 , and the induced exact sequence
+
0 → J4 /J ′ 4 → G4 /J4′ → I 1 → 0
+
is an abelian extension of I 1 ≃ Z23 by J4 /J4′ ≃ F3 . Since H4 /J4′ is topologically generated by two
+
elements and surjects onto I 1 , it actually defines a splitting:
G4 /J4′ = H4 /J4′ ⊕ J4 /J4′ .
Thus H4′ = J4′ , H4 is a normal subgroup of G4 , H is a normal subgroup of I + and
I + /H ≃ G4 /H4 ≃ J4 /J4′ ≃ F3
is generated by the class of xα+β (1) or x3α+β (1). We have shown:
Lemma 2.10.2. The group I + is spanned by xα (Zp ) and xβ (Zp ) plus xα+β (1) if p = 3.
Proof. (Of theorem 2.4.1 for G simple of type G2 ) By lemma 2.7.1, 2.10.1, 2.10.2 and the Iwahori
decomposition of section 2.5, the pro-p-Iwahori I(1) is generated by xα (Zp ), xβ (Zp ), x−2β−3α (pZp ),
along with xα+β (1) if p = 3. It is therefore topologically generated by xα (1), xβ (1), x−2β−3α (p),
along with xα+β (1) if p = 3. The surjective reduction morphism I(1) ։ U(Fp ) ։ U1 (Fp ) shows
that the first two generators can not be removed. The third one also can not, since all the others
belong to the closed subgroup I+ ( I(1). Finally, suppose that p = 3 and consider the extension
1 → U2 /U5 → U/U5 → U/U1 → 1
With notations as above, the reduction of
J4′ ⊂ J4 = U2 (Z3 )/U5 (Z3 ) = (U2 /U5 )(Z3 )
is a normal subgroup Y of X = (U/U5 )(F3 ) with quotient X/Y ≃ F33 . The surjective reduction
morphism
I(1) ։ U(F3 ) ։ U(F3 )/U5 (F3 ) = X ։ X/Y
then kills x−2β−3α (p). The fourth topological generator xα+β (1) of I(1) thus also can not be removed,
since the first two certainly do not span X/Y ≃ F33 .
11
2.11
We now return to an arbitrary split reductive group G over Zp . Let
Gsc ։ Gder ֒→ G ։ Gad
be the simply connected cover Gsc of the derived group Gder of G, and the adjoint group π : G ։
Gad of G. Then
ad
ad
ad
ad
ad
T , M , R , ∆ , Xα α∈∆ad = π(T), ZR, R, ∆, (π(Xα ))α∈∆
is a pinning of Gad and this construction yields a bijection between pinnings of G and pinnings of
Gad . Applying this to Gsc or Gder , we obtain pinnings
sc
sc
sc
sc
sc
der
der
der
der
der
T , M , R , ∆ , (Xα )α∈∆sc
and
T , M , R , ∆ , Xα α∈∆sc
for Gsc and Gder : all of the above constructions then apply to Gad , Gsc or Gder , and we will denote
with a subscript ad, sc or der for the corresponding objects. For instance, we have a sequence of
Iwahori (resp. pro-p-Iwahori) subgroups
I sc → I der ֒→ I → I ad
2.12
and I sc (1) → I der (1) ֒→ I(1) → I ad (1).
The action of G on itself by conjugation factors through a morphism
Ad : Gad → Aut(G).
For b ∈ Bad (Fp ), Ad(b)(BFp ) = BFp and Ad(b)(UFp ) = UFp . We thus obtain an action of the
Iwahori subgroup I ad of Gad = Gad (Qp ) on I or I(1). Similar consideration of course apply to Gsc
and Gder , and the sequence
I sc (1) → I der (1) ֒→ I(1) → I ad (1)
ad
is equivariant for these actions of I ad = I ad (1) ⋊ Ttors
.
2.13
Let J be the image of I sc (1) → I(1), so that J is a normal subgroup of I. From the
compatible Iwahori decompositions for I(1) and I sc (1) in section 2.5, we see that T (1) ֒→ I(1)
induces a T ad -equivariant isomorphism
T (1)/T (1) ∩ J → I(1)/J.
Since the inverse image of T(Zp ) in Gsc (Zp ) equals Tsc (Zp ) and since also
T sc (1) = Tsc (Zp ) ∩ I sc (1),
we see that T (1) ∩ J is the image of T sc (1) → T (1). Also, the kernel of I sc (1) → I(1) equals
Z ∩ I sc (1) where
Z = ker(Gsc → G)(Zp ) = ker(Tsc → T)(Zp ).
Therefore Z ∩ I sc (1) is the kernel of T sc (1) → T (1), which is trivial since Z is finite and T sc (1) ≃
Hom(M sc , 1 + pZp ) has no torsion. We thus obtain exact sequences
1 → T sc (1) → T (1) → Q → 0
∩
∩
k
sc
1 → I (1) → I(1) → Q → 0
where the cokernel Q is the finitely generated Zp -module
Q = (M ∨ /ZR∨ ) ⊗ (1 + pZp ) .
12
Remark 2.13.1. If G is simple, then M ∨ /ZR∨ is a finite group of order c, with c | ℓ + 1 if G is of
type Aℓ , c | 3 if G is of type E6 and c | 4 in all other cases. Thus Q = 0 and I sc (1) = I(1) unless
G is of type Aℓ with p | c | ℓ + 1 or p = 3 and G is adjoint of type E6 . In these exceptional cases,
M ∨ /ZR∨ is cyclic, thus Q ≃ Fp .
2.14
It follows that I(1) is generated by I sc (1) and s(1 + pZp ) for s ∈ S, thus topologically
generated by I sc (1) and s(1 + p) for s ∈ S. In view of the results already established in the simply
connected case, this shows that the elements listed in (1 − 4) of Theorem 2.4.1 indeed form a set of
topological generators for I(1).
None of the semi-simple elements in (1) can be removed: they are all needed to generate the
above abelian quotient Q of I(1) which indeed kills the unipotent generators in (2 − 4). Likewise,
none of the unipotent elements in (2) can be removed: they are all needed to generate the abelian
quotient
Y
Uα (Fp )
I(1) ։ U(Fp ) ։ U1 (Fp ) ≃
α∈∆
which kills the other generators in (1), (3) and (4).Q
One checks easily using the Iwahori decomposi−
tion of I(1) and the product decomposition U = c∈C U−
c that none of the unipotent elements in
(3) can be removed. Finally if p = 3 and d ∈ D, the central isogeny Gsc → Gad induces an isomorad
phism Gsc
d → Gd between the simple (simply connected and adjoint) components corresponding
to d, thus also an isomorphism between the corresponding pro-p-Iwahori’s Idsc (1) → Idad (1). In
particular, the projection I(1) → I ad (1) ։ Idad (1) is surjective. Composing it with the projection
Idad (1) ։ F33 constructed in section 2.10, we obtain an abelian quotient I(1) ։ F33 that kills all of
our generators except xα (1), xβ (1) and xα+β (1) where ∆d = {α, β}. In particular, the generator
xα+β (1) from (4) is also necessary. This finishes the proof of Theorem 2.4.1.
2.15
ad
The action of I ad = I ad (1) ⋊ Ttors
on I(1) induces an Fp -linear action of
ad
Ttors
= Hom M ad , µp−1 = Hom ZR, F×
p
˜ of I(1). Our minimal set of topological generators of I(1) reduces to
on the Frattini quotient I(1)
˜
˜ made of eigenvectors for the action of T ad . We denote
an eigenbasis of I(1),
i.e. an Fp -basis of I(1)
tors
ad
by Fp (α) the 1-dimensional representation of Ttors
on Fp defined by α ∈ ZR. We thus obtain:
ad
˜
Corollary 2.15.1. The Fp [Ttors
]-module I(1)
is isomorphic to
⊕
⊕
F
(δ
)
if
p
=
3
.
F♯S
⊕
⊕
F
(α)
⊕
⊕
F
(−α
)
d∈D p c
α∈∆ p
c∈C p
c,max
p
Here ♯S denotes the cardinality of the set S. The map α 7→ Fp (α) yields a bijection between
ad
ZR/(p − 1)ZR and the isomorphism classes of simple Fp [Ttors
]-modules. In particular some of the
simple modules in the previous corollary may happen to be isomorphic. For instance if G is simple
of type Bℓ and p = 3, then −αmax ≡ α mod 2 where α ∈ ∆ is a long simple root. An inspection of
the tables in [1] yields the following:
ad
˜ is multiplicity free unless p = 3 and G
Corollary 2.15.2. If G is simple, the Fp [Ttors
]-module I(1)
is of type A1 , Bℓ or Cℓ (ℓ ≥ 2), F4 or G2 .
In the next section we use this result to construct Galois representations landing in I ad with
image containing I ad (1).
13
3
The Construction of Galois Representations
Let G be a split simple adjoint group over Zp and let I(1) and I = I(1) ⋊ Ttors be the corresponding
Iwahori groups, as defined in the previous section. We want here to construct Galois representations
of a certain type with values in I with image containing I(1). After a short review of p-rational
fields in section 3.1, we establish a criterion for the existence of our representations in sections 3.2
and 3.3 and finally give some examples in section 3.4.
3.1
Let K be a number field, r2 (K) the number of complex primes of K, Σp the set of primes of K
lying above p, M the compositum of all finite p-extensions of K which are unramified outside Σp , M ab
the maximal abelian extension of K contained in M, and L the compositum of all cyclic extensions
of K of degree p which are contained in M or M ab . If we let Γ denote Gal(M/K), then Γ is a pro-p
group, Γab ∼
= Gal(L/K) is
= Γab /pΓab ∼
= Gal(M ab /K) is the maximal abelian quotient of Γ, and Γ̃ ∼
the Frattini quotient of Γ.
Definition A number field K is p-rational if the following equivalent conditions are satisfied:
(1) rank Zp (Γab ) = r2 (K) + 1 and Γab is torsion-free as a Zp -module,
(2) Γ is a free pro-p group with r2 (K) + 1 generators,
(3) Γ is a free pro-p group.
The equivalence of (1), (2) and (3) follows from [6], see also proposition 3.1 and the discussion before remark 3.2 of [3]. There is a considerable literature concerning p-rational fields, including [8], [4].
Examples:
(1) Suppose that K is a quadratic field and that either p ≥ 5 or p = 3 and is unramified in
K/Q. If K is real, then K is p-rational if and only if p does not divide the class number of K and
the fundamental unit of K is not a p-th power in the completions Kv of K at the places v above
p. On the other hand, if K is complex and p does not divide the class number of K, then K is a
p-rational field (cf. proposition 4.1 of [3]). However, there are p-rational complex K’s for which p
divides the class number (cf. chapter 2, section 1, p. 25 of [7]). For similar results, see also [2] and
[5] if K is complex.
(2) Let K = Q(µp ). If p is a regular prime, then K is a p-rational field (cf. [12], see also [3],
proposition 4.9 for a shorter proof).
3.2
Suppose that K is Galois over Q and p-rational with p ∤ [K : Q].
Since K is Galois over Q, so is M and we have an exact sequence
1→Γ→Π→Ω→1
(3.2.1)
where Ω = Gal(K/Q) and Π = Gal(M/Q). Conjugation in Π then induces an action of Ω on
the Frattini quotient Γ̃ = Gal(L/K) of Γ. Any continuous morphism ρ : Π → I maps Γ to I(1)
˜
If
and induces a morphism ρ : Ω → I/I(1) = Ttors and a ρ-equivariant morphism ρ̃ : Γ̃ → I(1).
ρ(Γ) = I(1), then ρ̃ is also surjective. Suppose conversely that we are given the finite data
ρ : Ω → Ttors
˜
and ρ̃ : Γ̃ ։ I(1).
14
Then as Ω has order prime to p, the Schur-Zassenhaus theorem ([14], proposition 2.3.3) implies that
the exact sequence 3.2.1 splits. The choice of a splitting Π ≃ Γ ⋊ Ω yields a non-canonical action of
Ω on Γ which lifts the canonical action of Ω on the Frattini quotient Γ̃. By [3], proposition 2.3, ρ̃ lifts
to a continuous Ω-equivariant surjective morphism ρ′ : Γ ։ I(1), which plainly gives a continuous
morphism
ρ = (ρ′ , ρ) : Π ≃ Γ ⋊ Ω → I = I(1) ⋊ Ttors
˜
Thus:
inducing ρ : Ω → Ttors and ρ̃ : Γ̃ ։ I(1).
Proposition 3.2.1. Under the above assumptions on K, there is a continuous morphism ρ : Π → I
such that ρ(Γ) = I(1) if and only if there is a morphism ρ : Ω → Ttors such that the induced Fp [Ω]˜ is a quotient of Γ̃.
module ρ∗ I(1)
˜
˜
as an
The Frattini quotient I(1)
is an Fp [Ttors ]-module and by the map ρ, we can consider I(1)
∗˜
Fp [Ω]-module which we denote by ρ I(1).
3.3
Suppose now that
A(K): K is a totally complex abelian (thus CM) Galois extension of Q which is p-rational of
degree [K : Q] | p − 1.
Let Ω̂ be the group of characters of Ω with values in F×
p , Ω̂odd ⊂ Ω̂ the subset of odd characters
(those taking the value −1 on complex conjugation), and χ0 ∈ Ω̂ the trivial character. Then by [3]
proposition 3.3,
Γ̃ = ⊕χ∈Ω̂odd ∪{χ0 } Fp (χ)
as an Fp [Ω]-module. In particular, Γ̃ is multiplicity free. Suppose therefore also that the Fp [Ttors ]˜ is multiplicity free, i.e. by corollary 2.15.2,
module I(1)
B(G): If p = 3, then G is not of type A1 , Bℓ or Cℓ (ℓ ≥ 2), F4 or G2 .
For S as in section 2.4, we define
Ω̂Sodd
(
Ω̂odd ∪ χ0 , if S = ∅
=
Ω̂odd ,
if S =
6 ∅.
Note that S = ∅ unless G if of type Aℓ with p | ℓ + 1 or G is of type E6 with p = 3, in which
both cases S is a singleton. We thus obtain:
Corollary 3.3.1. Under the assumptions A(K) on K and B(G) on G, there is a morphism
ρ : Π → I such that ρ(Γ) = I(1) if and only if there is morphism ρ : Ω → Ttors such that the
S
characters α ◦ ρ : Ω → F×
p for α ∈ ∆ ∪ {−αmax } are all distinct and belong to Ω̂odd .
15
3.4 Some examples. Write ∆ = {α1 , ..., αℓ } and αmax = n1 α1 +· · ·+nℓ αℓ using the conventions
of the tables in [1]. In this part we suppose that p is a regular (odd) prime and take K = Q(µp ), so
that K is p-rational and Ω = Z/(p − 1)Z.
Lemma 3.4.1. Suppose G is of type Aℓ , Bℓ , Cℓ or Dℓ and p ≥ 2l + 3 (resp. p ≥ 2l + 5) if p ≡ 1
mod 4 (resp. p ≡ 3 mod 4). Then we can find distinct characters φ1 , ..., φℓ+1 ∈ Ω̂odd ∪ χ0 such that
φn1 1 φn2 2 · · · φnℓ ℓ φℓ+1 = χ0 . Furthermore, if G is of type Aℓ and ℓ is odd, then one can even choose the
characters φ1 , ..., φℓ+1 to be inside Ω̂odd .
Proof. Since Ω is (canonically) isomorphic to Z/(p − 1)Z, ♯Ω̂odd = p−1
and there are exactly [ p−1
]
2
4
p−1
−1
−1
pairs of characters {χ, χ } with χ 6= χ in Ω̂odd . The condition on p is equivalent to ℓ ≤ 2[ 4 ]−1.
If G is of type Aℓ , then αmax = α1 + · · · + αℓ . If ℓ is even and 2ℓ ≤ [ p−1
], then we can pick 2ℓ
4
distinct pairs of odd characters {χ, χ−1 } as above for {φ1 , · · · , φℓ } and set φℓ+1 = χ0 . If ℓ is odd
≤ [ p−1
], then we can choose ℓ+1
distinct such pairs for the whole set {φ1 , · · · , φℓ+1 }.
and ℓ+1
2
4
2
If G is of type Dℓ (with ℓ ≥ 4), then αmax = α1 + 2α2 + ... + 2αℓ−2 + αℓ−1 + αℓ . Now if ℓ is odd
we can pick ℓ+1
such pairs {χ, χ−1 }, one for {φℓ−1, φℓ }, another pair for {φ1 , φℓ+1} and ℓ−3
such
2
2
ℓ
pairs for {φ2 , ..., φℓ−2 }. If ℓ is even, we let φ2 be the trivial character, and we can choose 2 such
pairs of characters {χ, χ−1 }, one pair for {φ1 , φℓ−1}, another pair for {φℓ , φℓ+1} and ℓ−4
such pairs
2
p−1
for {φ3 , ..., φℓ−2 }. So the inequality that we will need is 4 ≤ ℓ ≤ 2[ 4 ] − 1.
If G is of type Bℓ (with ℓ ≥ 2), then αmax = α1 + 2α2 + ... + 2αℓ . If ℓ is odd then we pick ℓ+1
2
pairs of characters {χ, χ−1 }; one pair for {φ1 , φℓ+1 } and ℓ−1
such
pairs
for
{φ
,
...,
φ
}.
If
ℓ
is
even
2
ℓ
2
then we need 2ℓ pairs of {χ, χ−1 }; one pair for {φ1 , φℓ+1} and ℓ−2
such pairs for {φ3 , ..., φℓ } and we
2
p−1
let φ2 be the trivial character. So in this case we need 3 ≤ ℓ ≤ 2[ 4 ] − 1.
The remaining Cℓ case is analogous.
P
Lemma 3.4.2. Suppose G is of type E6 , E7 , E8 , F4 or G2 and p ≥ ℓi=1 (2i − 1)ni + 2ℓ. Then we
can find distinct characters φ1 , ..., φℓ+1 ∈ Ω̂odd such that φn1 1 φn2 2 · · · φnℓ ℓ φℓ+1 = χ0 .
Proof. The choice of a generator ξ of F×
p yields an isomorphism Z/(p − 1)Z ≃ Ω̂, mapping i to χi
and
odd for i = 1, · · · , ℓ and φℓ+1 = χ−r where r =
Pℓ 1 + 2Z/(p − 1)Z to Ω̂odd . Set φi = χ2i−1 ∈ Ω̂P
ℓ
show
that
h
=
φ
∈ Ω̂odd and plainly
n
·
(2i
−
1).
The
tables
in
[1]
i=1 ni is odd, thus also
i=1 i
Pℓ ℓ+1
P
ℓ
nℓ
n1
φ1 · · · φℓ φℓ+1 = 1. If p ≥ i=1 (2i − 1)ni + 2ℓ, the elements {2i − 1, − i=1 ni · (2i − 1); i ∈ [1, ℓ]}
are all distinct modulo p − 1, which proves the lemma.
P
Remark 3.4.3. For G of type E6 , E7 , E8 , F4 or G2 , the tables in [1] show that the constant ℓi=1 (2i−
1)ni + 2ℓ of lemma 3.4.2 is 79, 127, 247, 53, 13 respectively.
Corollary 3.4.4. There is a constant c depending only upon the type of G such that if p > c
is a regular prime, then for K = Q(µp ), M, Π and Γ as above, there is a continuous morphism
ρ : Π → I with ρ(Γ) = I(1).
In conclusion, we have determined a minimal set of topological generators of the pro-p Iwahori
subgroup of a split reductive groups over Zp (theorem 2.4.1) and used it to study the structure of
ad
˜
the Frattini quotient I(1)
as an Fp [Ttors
]-module (corollary 2.15.1). Then we have used corollary
˜ is multiplicity free (see corollary 2.15.2). Furthermore in proposition
2.15.1 to determine when I(1)
3.2.1 and corollary 3.3.1, assuming p-rationality, we have shown that we can construct Galois
representations if and only if we can find a suitable list of distinct characters in Ω, the existence of
which is discussed in section 3.4 under the assumption K = Q(µp ), for any sufficiently large regular
prime p (see corollary 3.4.4).
16
References
[1]
N. Bourbaki, Éléments 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).
[2]
S. Fujii, On the maximal pro-p extension unramified outside p of an imaginary quadratic field,
Osaka J. Math. 45, (2008), 41-60.
[3]
R. Greenberg, Galois representation with open image, Annales mathématiques du Québec, Volume 40, Issue 1, (2016), 83-119.
[4]
J. F. Jaulent, T. Nguyen Quang Do, Corps p-rationnels, corps p-réguliers, et ramification restreinte, Séminaire de Théorie des Nombres de Bordeaux, (1987-88), Exposé 10, 10-01 10-26.
[5]
J. Minardi, Iwasawa modules for Zdp -extensions of algebraic number fields, University of Washington Ph. D. thesis, (1986).
[6]
A. Movahhedi, T. Nguyen Quang Do, Sur l’arithmétique des corps de nombres p-rationnels, Prog.
Math. 81, Birkhauser, (1990), 155-200.
[7]
A. Movahhedi, Sur les p-extensions des corps p-rationnels, Thèse de doctorat en Mathématiques,
Paris 7, (1988).
[8]
A. Movahhedi, Sur les p-extensions des corps p-rationnels, Math. Nach, 149, (1990), 163-176.
[9]
R. Ollivier, P. Schneider, A canonical torsion theory for pro-p Iwahori-Hecke modules,
https://arxiv.org/pdf/1602.00738v1.pdf, (2016).
[10] C. M. Ringel, The (n−1)-antichains in a root poset of width n, http://arxiv.org/abs/1306.1593v1,
(2013).
[11] Séminaire de Géométrie Algébrique du Bois Marie - 1962-64 - Schémas en groupes - (SGA 3)
Philippe Gille and Patrick Polo, editors.
[12] I. R. Shafarevich, Extensions with given points of ramification, Amer. Math. Soc, Translations
59, (1966), 128-149.
[13] J. Tits, Reductive groups over local fields, In Automorphic forms, representations and L-functions
(Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore., 1977), Part 1, Proc. Sympos.
Pure Math., XXXIII, pages 29-69. Amer. Math. Soc., Providence, R.I., (1979).
[14] J. S. Wilson, Profinite Groups, Oxford Science Publications, London Mathematical Society
Monographs, New Series 19, (2005).
| 4math.GR
|
1
Fixed-time cluster synchronization for complex
networks via pinning control
arXiv:1509.03350v1 [cs.SY] 6 Sep 2015
Xiwei Liu, Member, IEEE, and Tianping Chen, Senior Member, IEEE
Abstract—In this paper, the fixed-time cluster synchronization
problem for complex networks via pinning control is discussed.
Fixed-time synchronization has been a hot topic in recent years,
which means that the network can achieve synchronization in
finite-time and the settling time is bounded by a constant for any
initial values. To realize the fixed-time cluster synchronization,
a simple distributed protocol by pinning control technique is
designed, whose validity is rigorously proved, and some sufficient criteria for fixed-time cluster synchronization are also
obtained. Especially, when the cluster number is one, the cluster
synchronization becomes the complete synchronization problem;
when the intrinsic dynamics for each node is missed, the fixedtime cluster synchronization becomes the fixed-time cluster (or
complete) consensus problem; when the network has only one
node, the coupling term between nodes will disappear, and
the synchronization problem becomes the simplest master-slave
case, which also includes the stability problem for nonlinear
systems like neural networks. All these cases are also discussed.
Finally, numerical simulations are presented to demonstrate the
correctness of obtained theoretical results.
Index Terms—Complex networks, cluster synchronization,
fixed-time, finite-time, pinning control.
I. I NTRODUCTION
Synchronization of complex networks [1]-[11] has been a
hot topic in recent decades, and it is usually described by the
following continuous-time ordinary differential equation:
ẋi (t) = f (xi (t)) + gi (x1 (t), · · · , xN (t)), i = 1, · · · , N
n
where xi ∈ R denotes the state of agent i; the first term
f (xi (t)) denotes the intrinsic dynamics of agent i, when f =
0, the synchronization model becomes the consensus model;
and the second term gi (x1 (t), · · · , xN (t)) means the diffusive
coupling from agent i’s neighbours. Although each agent just
needs to get the local information of its neighbours, under the
above algorithm, the whole network can display a collective
behavior—-synchronization, i.e.,
lim kxi (t) − xj (t)k = 0; i, j = 1, 2, · · · , N
t→∞
(1)
This work is jointly supported by the National Natural Sciences Foundation
of China (Nos. 61203149, 61273211 and 61233016), the National Basic
Research Program of China (973 Program) under Grant No. 2010CB328101,
“Chen Guang” project supported by Shanghai Municipal Education Commission and Shanghai Education Development Foundation under Grant No.
11CG22, and the Fundamental Research Funds for the Central Universities.
Xiwei Liu is with Department of Computer Science and Technology,
Tongji University, and with the Key Laboratory of Embedded System and
Service Computing, Ministry of Education, Shanghai 200092, China. E-mail:
[email protected]; [email protected]
Corresponding Author Tianping Chen is with the School of Mathematical/Computer Sciences, Fudan University, 200433, Shanghai, P.R. China. Email: [email protected]
where k · k denotes some norm. It is a fundamental research
topic in decentralized control, and has broad applications in
cooperative control of unmanned air vehicles (UAVs), formation control of mobile robots, etc. The most popular model in
the synchronization literature is the linear coupling protocol
ẋi (t) = f (xi (t)) + c
N
X
aij (xj (t) − xi (t)), i = 1, · · · , N
j=1
and the nonlinear coupling protocol
ẋi (t) = f (xi (t)) + c
N
X
aij φ(xj (t), xi (t)), i = 1, · · · , N
j=1
(Infinite-time vs Finite-time) Under these coupling protocols, the sufficient criteria for asymptotic synchronization or
exponential synchronization can be obtained. Synchronization
rate is an important performance indicator for protocol. However, for the above two types of synchronization speed, the
disadvantage is that the completely same of each node can
never occur in finite-time, i.e., there does not exist a constant
T called the settling time, which may depend on the initial
values, such that for any i, j = 1, 2, · · · , N
lim kxi (t) − xj (t)k = 0; and xi (t) = xj (t), ∀t ≥ T. (2)
t→T
In many applications as robotics, a standard problem in system
theory is to develop controllers which drive a system to
a given position as fast as possible [12]. Moreover, chaos
synchronization plays an important role in the literature, if the
synchronization does not realize in a finite-time, for example,
exponential synchronization is realized, then the coupling
protocol or the external controllers should be always added
on the network, since if they are cancelled, from the property
of chaotic oscillator, a small error may cause a large difference
between nodes. Furthermore, finite-time synchronization can
lead to better system performances in the disturbance rejection
and robustness against uncertainty. Based on above reasons,
an investigation of finite-time synchronization under new
coupling protocols is important both in theoretical analysis
and real applications.
Next, we review some progress in the finite-time literature.
Finite-time synchronization (or consensus) heavily relies on
the development of finite-time stabilization theory. Finitetime stabilization problems have been studied mostly in the
contexts of optimality, controllability, and deadbeat control
for several decades. These control laws are usually timevarying, discontinuous, or even depending directly on the
2
initial conditions of considered systems, for example, in [13],
the author considered the following controllers
ẋ = −
grad(f )(x)
, and ẋ = −sgn(grad(f )(x)),
kgrad(f )(x)k2
by using the theory of Filippov, the author proved the finitetime stability and applied it on the network consensus problem.
Recently, finite-time stability and finite-time stabilization via
continuous time-invariant feedback have been studied. For
example, [14] studied the finite-time stability of a homogeneous systems and designed the bounded continuous finitetime stabilizing feedback laws for the double integrators; [15]
studied global finite-time control of robot systems through
state feedback and output feedback control; [16] discussed
the finite-time stability of continuous systems, investigated its
sensitivity to perturbations and rigorously set up a general
framework for finite-time stabilization, many papers’ works
were based on this excellent result; [17]-[18] proposed two
new sufficient conditions for local finite-time stability and
designed an observer for a class of homogeneous systems with
Lipschitz nonlinearity.
As for the finite-time synchronization/consensus literature,
in [19], the author proposed a new class of finite-time consensus protocol:
X
α
X
ẋi (t) = βsig
Wij (xj − xi )
+γ
Wij (xj − xi ),
j∈Ni
j∈Ni
where 0 < α < 1, β > 0, γ ≥ 0, they proved that if the
network had a spanning tree, then the above protocol can
realize the finite-time consensus. [20] investigated the finitetime consensus under the protocol:
X
ẋi (t) =
aij sig(xj − xi )αij (t) , 0 < αij (t) < 1,
j∈Ni
they proved that if the sum of time intervals, in which the
interaction topology was connected, was sufficiently large,
the above protocol can realize finite-time consensus for both
bidirected and unidirected networks. [21] considered the finitetime weighted average consensus of time-varying topology
with respect to the monotonic function g under the protocol:
n
X
1
aij (t)sig(h(xj ) − h(xi ))α , 0 < α < 1,
ẋi (t) =
ωi dg/dxi j=1
when ωi = 1, then the above protocol was the one in [22].
[23] considered the finite-time χ-consensus models,
ẋi (t) =
1
∂χ
|
| ∂x
i
n
X
aij (t)ψ(sig(xj − xi )α ),
0 ≤ α < 1,
α
aij (t)[ψ(xα
j ) − ψ(xi )],
0 ≤ α < 1,
j=1
and
ẋi (t) =
1
∂χ
|
| ∂x
i
n
X
j=1
using the homogeneity property of ψ and the Lyapunov function, the authors finally proved the finite-time consensus for
undirected (the first model) and directed (the second model)
networks. With the similar analysis, [24] discussed the finitetime consensus of the leader-following multi-agent systems
with jointly-reachable leader and switching jointly-reachable
leader for the first-order model
X
aij (t)φ(sig(xj (t) − xi (t))α )
ẋi (t) =
vj ∈Ni (t)
− bi (t)φ2 (sig(xi (t) − x0 )α ),
0 ≤ α < 1,
and for the second-order model
ẋi (t) =vi (t),
X
v̇i (t) =
aij (t)[ϕ1 (sig(xj (t) − xi (t))α1 )
vj ∈Ni (t)
+ ϕ2 (sig(vj (t) − vi (t))α2 )]
− bi (t)[ϕ3 (sig(xj (t) − x0 (t))α1 )
+ ϕ4 (sig(vj (t) − v0 (t))α2 ))],
2α1
where ẋ0 (t) = v0 (t), v̇0 (t) = 0 and 0 < α1 < 1, α2 = 1+α
.
1
Moreover, in [25], by using the tools from homogeneous
theory, the authors investigated finite-time consensus with
second-order integrators based on both relative position and
relative velocity measurements as
q̇i (t) =pi (t),
α1
X
N
aij (qi − qj )
ṗi (t) = − k1 tanh sig
j=1
− k2 tanh sig
X
N
2α1
1+α1 ,
j=1
α2
,
aij (pi − pj )
where 0 < α1 < 1, α2 =
and they also considered the
case with only the relative position measurements obtained. In
[26], the optimal finite-time stabilization problem was considered, and they designed a new switching protocol covering
both continuous control (0 < α < 1) and discontinuous
one (α = 0), which can shorten the stabilization time;
with the same aim, the authors in [27] also considered the
optimal consensus time for multi-agents systems and proposed
a new switching protocol. Moreover, in [28], the distributed
finite-time containment control algorithm for double-integrator
multi-agent systems with multiple dynamic or stationary leaders was proposed and its validity was also rigorously proved.
(Finite-time vs Fixed-time) Although many finite-time results are obtained in the above short review, the settling time
heavily depends on the initial conditions, which limits the
practical applications, since the knowledge of initial conditions is not available in advance. Therefore, to overcome
this drawback, in [29], a new concept called the fixed-time
stability is proposed, if it is globally finite-time stable and the
settling time function is bounded for any initial values. The
technique is adding extra terms on the previous finite-time
model, which will be stated more carefully in the next section.
Following this streamline of dealing with fixed-time stability,
many new approaches and results are obtained, see [30]-[35]
and references therein. For example, [33] and [34] proposed
two new protocols to realize the fixed-time stabilization, and
the upper bounds for the settling time were also estimated;
moreover, they also applied these new protocols on fixed-time
consensus problem for multi-agent systems. [35] proposed a
3
fixed-time terminal sliding-mode control protocol for a class
of second-order nonlinear systems in the presence of uncertainties and perturbations. In [36], the authors considered the
finite-time consensus for multi-agent systems with cooperative
and antagonistic interactions, and they proved that the states
of all agents can reach agreement in a finite-time regarding
consensus values that were the same in modulus but may not
be the same in sign.
Although many papers have considered the finite-time and
fixed-time consensus problem, the discussions about synchronization and consensus have a great difference because of the
existence of intrinsic dynamics. There are few works to discuss
the finite-time and fixed-time synchronization. For example,
[37] investigated the finite-time complete synchronization,
while [38] studied the fixed-time complete synchronization.
On the other hand, cluster synchronization is a more practical
phenomenon than the complete synchronization, which is
significant in biological sciences, communication engineering,
etc. The cluster synchronization means that nodes in the same
cluster can achieve complete synchronization, while nodes
in different clusters have different dynamical behaviors. Of
course, when all the nodes lie in the same cluster, then the cluster synchronization becomes the complete synchronization.
There were some paper considering the cluster synchronization
exponentially like [7], [8], [40]-[42], and [43] also considered
the finite-time cluster synchronization of Markovian switching
complex networks with stochastic perturbations by adding the
linear negative feedback controllers.
To our best knowledge, there are no papers considering
the fixed-time cluster synchronization via pinning control,
which will be the subject of this paper.
This paper is organized as follows. In Section II, some
necessary definitions, lemmas, assumptions and notations are
given. A simple coupling protocol to ensure the fixed-time
cluster synchronization via pinning control is also proposed.
In Section III, we first rigorously prove the effectiveness of
proposed protocol, then the fixed-time complete synchronization is also carefully discussed. Furthermore, some numerical
simulations are given in Section IV to show the correctness
of obtained theoretical results. Finally, a conclusion and some
discussion about future work are presented in Section V.
II. P RELIMINARIES
In this section, we present some definitions, lemmas, and
notations, which will be useful throughout this paper. First,
we will give some general results on the dynamical systems.
If a nonnegative function V (t) satisfies
V̇ (t) = −αµ(V (t)),
where α > 0, functions µ(V (t)) > 0, V (t) > 0; µ(0) = 0.
Because V̇ (t) > 0, therefore, V (t) is decreasing.
Define the function V1 (t) as follows:
Z s
V1 (s) = −α−1
µ−1 (V )dV,
V (0)
then,
t = V1 (V (t)),
(3)
and
V1− (t) = V (t),
(4)
V1−
where
is the inverse function of V .
If µ(s) = s, then
Z V (t)
V (t)
−1
,
t = V1 (V (t)) = −α
V −1 dV = −α−1 log
V (0)
V (0)
and
V (t) = V (0)e−αt .
If µ(s) = sp , p 6= 1, then
V1 (s) =
−1
(s1−p − V 1−p (0)),
α(1 − p)
which implies
1
V (t) = [α(p − 1)t + V 1−p (0)] 1−p .
In case p < 1, then V (t) = 0, if
t≥
V 1−p (0)
.
α(1 − p)
On the other hand, in case p > 1,
V (t) =
1
1
[α(p − 1)t + V 1−p (0)] p−1
.
As direct consequences, we show some useful lemmas on
the finite-time and fixed-time stability.
Lemma 1: (See [16]) If a nonnegative function V (t) satisfies
V̇ (t) ≤ −αV p (t), 0 < p < 1
(5)
where α > 0. Then, V (t) ≡ 0, if
t≥
V 1−p (0)
.
α(1 − p)
(6)
Many papers investigating the finite-time stability or consensus are based on this result, see [19]-[27]. Moreover, the
other forms like V̇ (t) ≤ −αV p (t) ± kV (t) are also proposed
and proved to realize finite-time stability in [17] and [18], here
we omit them, interested readers can refer to these works.
However, the settling time depends on the initial value V (0),
which in many cases cannot be obtained. Therefore, in our
recent paper, a general criteria for the fixed-time stability is
proposed, which can be stated as follows.
Lemma 2: (See [39]) If a nonnegative function V (t) satisfies
−αV p (t), 0 < p < 1 ; if 0 < V < 1
V̇ (t) ≤
(7)
−βV q (t), q > 1
; if V ≥ 1
where α > 0, β > 0. Then, V (t) ≡ 0, if
t≥
1
1
+
.
α(1 − p) β(q − 1)
(8)
Remark 1: As a direct result, if a nonnegative function V (t)
satisfies
V̇ (t) ≤ −αV p (t) − βV q (t),
(9)
4
where α > 0, β > 0, 0 < p < 1, q > 1. Then, V (t) ≡ 0, if
t≥
1
1
+
.
α(1 − p) β(q − 1)
(10)
Remark 2: From the above two lemmas, one can conclude that for the finite-time stability, only one term like
−V (t)p , 0 < p < 1 can realize this aim; however, to realize
fixed-time stability, except for the term for finite-time stability,
one should add an extra term −V (t)q , q > 1, whose role can
be regarded as pulling the system into the region with norm
less than 1 in a fixed-time.
Remark 3: Some other criteria to ensure the fixed-time
stability, like
V̇ (t) ≤ −(αV p (t) + βV q (t))k ,
(11)
where α > 0, β > 0. p > 0, q > 0, k > 0, 0 < pk <
1, qk > 1, which is presented in [29]; or
1
1
V̇ (t) ≤ −αV 1− 2µ (t) − βV 1+ 2µ (t), µ > 1,
(12)
m
n
p
(t) − βV q (t),
(13)
where m, n, p, q are all positive odd integers satisfying m > n
and p < q, which is proposed in [34], can all be regarded the
special case of Lemma 2. Of course, utilizing the concrete
form of (12) and (13), the authors can obtain a more exact
estimation of the settling time.
Since the dealing with different norms happens in the fixedtime and finite-time literature, we first introduce a powerful
lemma.
Lemma 3: (See [44]) For any vector z ∈ Rn and 0 < r < l,
the following norm equivalence property holds:
1/r
1/l X
X
n
n
r
l
,
(14)
|zi |
≤
|zi |
i=1
i=1
and
n
1X
|zi |l
n i=1
1/l
ΘAǫ + ATǫ Θ < 0.
Especially, when A ∈ A2 and ǫ < 0, we can choose Φ = I,
where I is the identity matrix with appropriate dimensions,
such that matrix Aǫ is negative definite, i.e., its eigenvalues
are all negative and can be sorted as
0 > λ1 (Aǫ ) ≥ λ2 (Aǫ ) ≥ λ3 (Aǫ ) ≥ · · · ≥ λN (Aǫ ).
which is proposed in [30]; or
V̇ (t) ≤ −αV
P
1) aij ≥ 0, i 6= j, aii = − N
j=1,j6=i aij , i = 1, · · · , N
2) A is irreducible.
Furthermore, if A ∈ A1 and aij = aji , i 6= j, then we say
matrix A belong to class A2, denoted as A ∈ A2.
Obviously, if the Laplacian matrix for the graph G is A,
then A ∈ A1 means that the graph is a strongly connected
directed graph; and A ∈ A2 means that the graph is a strongly
connected undirected graph.
Lemma 4: (See [5]) Suppose A ∈ A1 and ǫ <
0, then, there exists a positive definite diagonal matrix
Θ = diag(θ1 , · · · , θN ), such that the matrix Aǫ = A +
diag{ǫ, 0, · · · , 0} is Lyapunov stable, i.e.,
≥
1/r
X
n
1
.
|zi |r
n i=1
(15)
Remark 4: The first inequality (14) is called the Jensen
inequality, whose proof can be found in P. 4, [44], and this
inequality is also commonly used in the finite-time literature.
On the other hand, the second inequality (15) is very useful
to deal with the newly added term like −V (t)q , q > 1, whose
proof can be found in P. 26, [44]. The above two inequalities
can be combined in the norm form as follows:
1
1
(16)
kzkl ≤ kzkr ≤ n r − l kzkl ,
Pn
P
n
where kzkr = ( i=1 |zi |r )1/r and kzkl = ( i=1 |zi |l )1/l .
Next, we will introduce some definitions and lemmas about
the cluster synchronization in complex networks.
For a complex network of N nodes, suppose its graph is
G = {V, E}, where V represents the vertex set numbered by
{1, · · · , N }, and E denotes the edge set with e(i, j) ∈ E if
and only if there is an edge from vertex j to i.
Definition 1: (See [8]) Matrix A = (aij ) ∈ RN ×N is said
to belong to class A1, denoted as A ∈ A1, if
(17)
Lemma 5: (See [6]) Suppose A = (aij )N ×N ∈ A2,
then for any two vectors X = (x1 , · · · , xN )T and Y =
(y1 , · · · , yN )T , we have
X
aij (xj − xi )(yj − yi ).
(18)
X T AY = −
j>i
Definition 2: (See [8]) Matrix A = (aij ) ∈ RN1 ×N2 is said
to belong to class
PN2 A3, denoted as A ∈ A3, if its each row-sum
is zero, i.e., j=1
aij = 0, i = 1, . . . , N1 .
Remark 5: In fact, we can also assume that each row-sum
is a non-zero constant, but for convenience, we assume the
row-sum is zero, which means that the interactions between
nodes can be cooperative (when the element is positive) and
competitive (when the element is negative).
In order to investigate the cluster synchronization, we assume the set of nodes in the network can be divided into m
clusters, i.e., {1, · · · , N } = C1 ∪ C2 ∪ · · · ∪ Cm , where
C1 = {1, · · · , r1 }, C2 = {r1 + 1, · · · , r2 }, · · · ,
Ck = {rk−1 + 1, · · · , rk }, · · · , Cm = {rm−1 + 1, · · · , N }.
(19)
Now, using the above definitions of matrices, we can define
a new type of coupling matrix A for the cluster synchronization analysis.
Definition 3: Suppose A ∈ RN ×N is symmetric, the indexes {1, · · · , N } can be divided into m clusters as defined
in (19), and the following form holds
A11 A12 · · · A1m
A21 A22 · · · A2m
(20)
A= .
..
.. ,
..
..
.
.
.
Am1
Am2
···
Amm
where Aij ∈ R(ri −ri−1 )×(rj −rj−1 ) , r0 = 0, Aii ∈ A2 and
Aij ∈ A3, i, j ∈ {1, · · · , m}. Then the matrix A is said to
belong to class A4, denoted as A ∈ A4.
Remark 6: In fact, the coupling matrix for cluster synchronization is also defined in [8], but in that paper, the matrix
5
A can be asymmetric. The reason for the requirement in
the above definition is for the convenience of later analysis.
Moreover, in [42], the authors also investigate another type of
cluster synchronization, i.e., A ∈ A1 and Aij ∈ A3, obviously,
they are different, since in this type of coupling matrix, the
interactions between nodes are only cooperative ones.
Cluster synchronization means that: each vertex in the same
cluster has the same individual node dynamic, while nodes
in different clusters are nonidentical, which can guarantee
that the trajectories are apparently distinguishing when cluster
synchronization is reached. In [42], the authors use different
intrinsic dynamics to guarantee the final cluster synchronization, while [7]-[8] use the pinning control technique [5] to
realize the final cluster synchronization. Now, for the complex
network with clusters defined in (19), we give the following
definition of fixed-time cluster synchronization.
Definition 4: For the network with (19), the fixed-time
cluster synchronization is said to be realized, if there exists
a time T independent on the initial values, such that for any
initial values, node i can converge to the target trajectory sk (t),
which belongs to the k-th cluster Ck , k = 1, 2, · · · , m, i.e.,
lim kxi (t) − sk (t)k = 0, and xi (t) = sk (t), ∀t ≥ T, (21)
t→T
where target trajectories in different clusters are different with
each other, i.e., sk1 (t) 6= sk2 (t), k1 , k2 ∈ {1, · · · , m}.
Now, we are in the position to propose the coupling protocols for the fixed-time cluster synchronization via pinning
control. Without loss of generality, we assume the controllers
are added on just the first node of each cluster. Therefore, we
list the protocol for cluster Ck , k = 1, · · · , m as follows:
ẋrk−1 +1 (t) = f (xrk−1 +1 (t))
X
p
+α
ark−1 +1,j sig xj (t) − xrk−1 +1 (t)
j∈Ck
+
X
sig
p
X
j∈Ck
+
X
ark−1 +1,j (xj (t) − xrk−1 +1 (t))
X
brk−1 +1,j (xj (t) − xrk−1 +1 (t))
j∈Ck′
k′ 6=k
+β
X
sig
j∈Ck′
k′ 6=k
−ǫ1 sigp (xrk−1 +1 (t) − sk (t))
−ǫ2 sigq (xrk−1 +1 (t) − sk (t)),
k′ 6=k
+β
j∈Ck′
X
j∈Ck
+
X
k′ 6=k
X
(22)
j∈Ck′
(26)
Especially, when p = 1, sig(x) = x.
Remark 7: In fact, we have made many simple assumptions. For example, as for the QUAD condition (25), one
can assume that there exists a diagonal matrix ∆ =
diag(δ1 , · · · , δn ) and δi > 0, i = 1, · · · , n, such that
(x − y)T (f (x) − f (y)) ≤ (x − y)T ∆(x − y). Moreover,
as for the coupling term, one can also use a function Ψ(x)
instead of x, where Ψ(x) = (ψ1 (x1 ), · · · , ψn (xn ))T , therefore
in this case, the nonlinear function includes the linear one
with Ψ(x) = Γx, where Γ is a positive diagonal matrix.
Furthermore, the parameters α, β, ǫ1 , ǫ2 defined in (22) and
(23) can also be different values or vectors which are functions
of the cluster k as: α(k), β(k), ǫ1 (k), ǫ2 (k).
In all, to simply the process and emphasis on the key role of
terms sigp (·), 0 < p < 1 and sigq (·), q > 1 for the fixed-time
synchronization, we adopt the assumptions as given, interested
readers can generalize these cases themselves.
Lemma 6: (Young’s inequality) Suppose a, b, u, v are all
positive scalars, and u1 + v1 = 1, u > 1, v > 1, then
(27)
RESULTS
At first, we define the synchronization error in cluster
Ck , k = 1, 2, · · · , m as:
f˜(ei ) = f (xi (t)) − f (sk (t)).
(28)
(29)
Therefore, the dynamics of the synchronization error in
cluster Ck , k = 1, · · · , m is given as follows:
bij sig xj (t) − xi (t)
sigq
sigp (x) = (sign(x1 )|x1 |p , · · · , sign(xn )|xn |p )T .
where i = rk−1 + 1, · · · , rk and sk (t) is the target trajectory
defined by (24). For i ∈ Ck , define
aij (xj (t) − xi (t))
bij (xj (t) − xi (t)) ,
For any vector x = (x1 , · · · , xn )T , the function sigp (x) :
R → Rn is defined as:
n
ei (t) = xi (t) − sk (t),
q
X
(x − y)T (f (x) − f (y)) ≤ δ(x − y)T (x − y), δ > 0. (25)
A. Cluster synchronization with m > 1
j∈Ck
sigp
(24)
where targets in different clusters are different, i.e., sk1 (t) 6=
sk2 (t), k1 6= k2 ∈ {1, · · · , m}.
Function f : Rn → Rn is used to describe the intrinsic
behavior of each isolated node, which satisfies the following
QUAD condition:
III. M AIN
ẋi (t) =f (xi (t))
X
p
aij sig xj (t) − xi (t)
+α
+
k = 1, 2, · · · , m,
bv
au
+
u
v
u
where “=” holds if and only if a = bv .
and for i = rk−1 + 2, · · · , rk ,
X
ṡk (t) = f (sk (t)),
ab ≤
q
brk−1 +1,j sig xj (t) − xrk−1 +1 (t)
q
where xi = (x1i , · · · , xni )T ∈ Rn , scalars α > 0, β > 0, ǫ1 >
0, ǫ2 > 0, 0 < p < 1, q > 1, and the target trajectory sk (t) in
cluster Ck is governed by
(23)
ėrk−1 +1 (t) = f˜(erk−1 +1 (t))
X
p
+α
ark−1 +1,j sig ej (t) − erk−1 +1 (t)
j∈Ck
6
+
X
sigp
X
j∈Ck
+
X
ark−1 +1,j ej (t)
X
brk−1 +1,j ej (t)
j∈Ck′
k′ 6=k
+β
X
Theorem 1: For coupled systems (22) and (23), if A ∈ A4,
α − γ1 − 2δ > 0 and β − γ2 − 2δ > 0, then the fixed-time
cluster synchronization can be achieved with the settling time
defined as:
brk−1 +1,j sigq ej (t) − erk−1 +1 (t)
sigq
k′ 6=k
j∈Ck′
p
Tmax =
−ǫ1 sig (erk−1 +1 (t)) − ǫ2 sigq (erk−1 +1 (t)),
(30)
and for i = rk−1 + 2, · · · , rk ,
ėi (t) = f˜(ei (t))
X
X
X
sigp
aij ej (t)
+α
aij sigp ej (t) − ei (t) +
j∈Ck
+β
X
j∈Ck
X
X
sigq
bij sigq ej (t) − ei (t) +
bij ej (t) ,
j∈Ck′
k′ 6=k
(31)
Before giving the main result, we define some new matrices.
Suppose the coupling matrices A and B in (30) and (31)
satisfies that A ∈ A4 and B ∈ A4 in Definition 3 with the
same cluster partition defined in (19). For k = 1, · · · , m,
we define Akk = (aij )(rk −rk−1 )×(rk −rk−1 ) as: for i, j =
1, · · · , rk − rk−1 ,
2
a 1+p
i 6= j
rk−1 +i,rk−1 +j ;
aij =
(32)
2
1+p
− Pr′k −rk−1
a
i=j
′;
′
j =1,j 6=i
rk−1 +i,rk−1 +j
We also define B kk = (bij )(rk −rk−1 )×(rk −rk−1 ) as: for i, j =
1, · · · , rk − rk−1 ,
2
b 1+q
i 6= j
rk−1 +i,rk−1 +j ;
(33)
bij =
2
1+q
− Pr′k −rk−1
b
i=j
′;
′
j =1,j 6=i rk−1 +i,rk−1 +j
Obviously, Akk ∈ A2 and B kk ∈ A2, k = 1, · · · , m.
Moreover, define
2
Âkk = −2Akk + diag{(2ǫ1 α−1 ) 1+p , 0, · · · , 0},
B̂kk = −2B kk + diag{(2ǫ2 β −1 )
2
1+q
, 0, · · · , 0}.
(35)
ρ1 =
min λmin (Âkk ), ρ2 = min λmin (B̂kk );
k=1,··· ,m
k=1,··· ,m
X
m
(rk − rk−1 )(rk − rk−1 − 1)
+m ;
N =n
2
(36)
a=
max
i∈Ck ,j∈Ck′ ,k6=k′
1+p
ρ1 2 ,
|aij |,
r=
p
β = βN
b=
1−q
2
2
q−1
2
max
i∈Ck ,j∈Ck′ ,k6=k′
(37)
γ1 = a r(N n)
2
,
q
γ2 = b r2
1+q
2
=
.
V̇ (t)
rk
m
X
X
T
ei (t) ėi (t) =
+α
j∈Ck
X
+β
f˜(ei (t))
k′ 6=k
j∈C
k′ 6=k
j∈Ck′
′
Xk
X
q
q
sig
bij sig ej (t) − ei (t) +
bij ej
=
− ǫ1 sigp (erk−1 +1 (t)) − ǫ2 sigq (erk−1 +1 (t))
ei (t)T f˜(ei (t))
k=1 i=rk−1 +1
rk
m
X
X
T
ei (t)
+α
k=1 i=rk−1 +1
rk
m
X
X
T
aij sig ej (t) − ei (t)
p
sig
p
j∈Ck
T
ei (t)
k=1 i=rk−1 +1
m
X
X
X
X
aij ej (t)
X
bij ej (t)
j∈Ck′
k′ 6=k
k=1 i=rk−1 +1
rk
X
m
X
X
ei (t)T
+β
X
j∈Ck
ei (t)
k=1 i=rk−1 +1
rk
m
X
X
+
ei (t)
X
X
p
p
sig
aij sig ej (t) − ei (t) +
aij ej
j∈Ck
m
X
+
eTrk−1 +1
k=1
rk
m
X
X
+
T
k=1 i=rk−1 +1
k=1 i=rk−1 +1
X
rk
X
m
X
bij sigq ej (t) − ei (t)
sig
k′ 6=k
q
j∈Ck′
erk−1 +1 (t)T sigp (erk−1 +1 (t))
−ǫ2
m
X
erk−1 +1 (t)T sigq (erk−1 +1 (t))
=Ṽ1 (t) + Ṽ2 (t) + Ṽ3 (t) + Ṽ4 (t) + Ṽ5 (t) + Ṽ6 (t) + Ṽ7 (t).
k=1,··· ,m
1+p
2
Pk
where Vk (t) = ri=r
ei (t)T ei (t) = Ek (t)T Ek (t), and
k−1 +1
T
Ek (t) = (erk−1 +1 (t) , · · · , erk (t)T )T , k = 1, · · · , m.
Differentiating it, we have
k=1
|bij |;
max [N − (rk − rk−1 )];
1−p
2
(40)
k=1
k−1
k=1
1+q
ρ2 2 ;
rk
m
m
1X X
1X
ei (t)T ei (t) =
Vk (t),
2
2
i=r
+1
k=1
−ǫ1
k=1
p−1
2
V (t) =
(34)
Since Akk ∈ A2 and B kk ∈ A2, according to Lemma
4, matrices Âkk and B̂kk are positive definite, and we use
λmin (Âkk ) and λmin (B̂kk ) to denote the smallest eigenvalue
of Âkk and B̂kk , respectively. Denote positive scalars
α = α2
where δ is as defined in (25), and positive parameters
α, β, γ1 , γ2 are defined in (37) and (38).
Proof: Define the Lyapunov function as
j∈Ck′
k′ 6=k
2
2
,
+
(α − γ1 − 2δ)(1 − p) (β − γ2 − 2δ)(q − 1)
(39)
According to (25), one can get
(38)
Now, with the above notations, we will prove the following
main theorem.
Ṽ1 (t) ≤ δ
m
X
rk
X
k=1 i=rk−1 +1
ei (t)T ei (t) = 2δV (t).
(41)
7
Using (14) in Lemma 3, let r = (1 + p)/2 ∈ (0, 1), l = 1
and r < l, and combining with Lemma 5, one can get that
Ṽ2 (t) + Ṽ6 (t)
m
αX X
T
p
=
aij (ei (t) − ej (t)) sig ej (t) − ei (t)
2
k=1 i,j∈Ck
m
X
− ǫ1
=−
α
2
− ǫ1
T
erk−1 +1 (t)
k=1
n X
m X
X
p
sig (erk−1 +1 (t))
2
elrk−1 +1 (t)2
X
aij ej (t)
aij elj (t)|p
j∈Ck′
|eli (t)|1+p
1+p
p|elj (t)|1+p
+
1+p
− (rk − rk−1 )]
1+p
2
k=1 i∈Ck l=1
i,j∈Ck
2
1+p
|eli (t)||
m X X
n
X
l
2
≤a r
|ei (t)|
aij |elj (t) − eli (t)|2
X
j∈Ck′
p
2
1+p
−1
k=1 k′ 6=k i=rk−1 +1
rk
n X
m X
X
X
l=1 k=1 k′ 6=k i∈Ck j∈Ck′
n X
m X
X
|eli (t)|1+p [N
=ap
l=1 k=1 i∈Ck
|elrk−1 +1 (t)|1+p
l=1 k=1
ei (t)T sigp
≤a
l=1 k=1 i,j∈Ck
n m X
α XX
rk
X
l=1 k=1 k′ 6=k i=rk−1 +1
m X X X
n X
X
p
l=1 k=1
≤−
Ṽ3 (t)
m X
X
=
≤
aij |elj (t) − eli (t)|1+p
n X
m
X
According to the Young’s inequality in Lemma 6, we have
1+p
2
p
≤a r(N n)
+ (2ǫ1 α )
m
α X
=−
Ek (t)T [(−2Akk ) ⊗ I]Ek (t)
2
1−p
2
X
m X X
n
|eli (t)|2
k=1 i∈Ck l=1
1+p
2
= γ1 V (t)
1+p
2
.
(44)
Similarly,
k=1
T
−1
+ Ek (t) [diag{(2ǫ1 α
)
2
1+p
1+p
2
, 0, · · · , 0} ⊗ I]Ek (t)
m
1+p
2
α X
Ek (t)T [Âkk ⊗ I]Ek (t)
=−
2
k=1
X
1+p
m
2
α
T
λmin (Âkk )Ek (t) Ek (t)
≤−
2
k=1
X
1+p
m
2
α
T
ρ1
Ek (t) Ek (t)
≤−
2
k=1
1+p
2
1+p
α
= −αV (t) 2 .
=−
2ρ1 V (t)
2
≤
≤−
β
N
2
2
l=1 k=1
(42)
q
X
j∈Ck′
X
|eli (t)||
bij ej (t)
bij elj (t)|q
j∈Ck′
m X X
n
X
l
2
≤b r
|ei (t)|
q
|eli (t)|1+q
1+q
q|elj (t)|1+q
+
1+q
1+q
2
≤b r
X
m X X
n
k=1 i∈Ck l=1
|eli (t)|2
1+q
2
= γ2 V (t)
1+q
2
.
(45)
Therefore, from (41)-(45), we have
1+p
V̇ (t) ≤ 2δV (t) − (α − γ1 )V (t) 2 − (β − γ2 )V (t)
(
1+p
V (t) < 1
−(α − γ1 − 2δ)V (t) 2 ;
≤
1+q
2
−(β − γ2 − 2δ)V (t) ;
V (t) ≥ 1
1+q
2
(46)
According to Lemma 2, one can get that the fixed-time
synchronization is finally realized, and the settling time can
be also obtained as (39). The proof is completed.
Remark 8: (Cluster consensus) Obviously, when f (·) = 0,
then the above fixed-time cluster synchronization problem
becomes the fixed-time cluster consensus problem. In this case,
only if α − γ1 > 0 and β − γ2 > 0, then the fixed-time cluster
consensus can be achieved with the settling time defined as:
2
1+q
2
X
1+q
m
2
β 1−q
T
2
Ek (t) [B̂kk ⊗ I]Ek (t)
=− N
2
k=1
1+q
2
1−q
1+q
β
2
≤− N
2ρ2 V (t)
= −βV (t) 2 .
2
k=1 i∈Ck l=1
q
i,j∈Ck
2
ei (t) sig
k=1 k′ 6=k i=rk−1 +1
rk
n X
m X
X
X
l=1 k=1 k′ 6=k i∈Ck j∈Ck′
bij1+q |elj (t) − eli (t)|2
+ (2ǫ2 β −1 ) 1+q elrk−1 +1 (t)2
T
≤b
Ṽ4 (t) + Ṽ7 (t)
n m
β XX X
=−
bij |elj (t) − eli (t)|1+q
2
−
rk
X
l=1 k=1 k′ 6=k i=rk−1 +1
n m
q XX X X X
Similarly, using (15) in Lemma 3, let l = (1+q)/2 > 1 = r,
and combining with Lemma 5, one can get that
l=1 k=1 i,j∈Ck
n X
m
X
ǫ2
|elrk−1 +1 (t)|1+q
l=1 k=1
X
n X
m X
1−q
Ṽ5 (t)
m X
X
=
Tmax =
(43)
2
2
.
+
(α − γ1 )(1 − p) (β − γ2 )(q − 1)
(47)
Remark 9: Let us analysis the role of parameters for fixedtime cluster synchronization. For the network model (30)
8
and (31), since the network topology is constant with time,
therefore, parameters γ1 and γ2 in (38) and δ in (25) are
constants. If we let ǫ1 = αω1 and ǫ2 = αω2 , then matrices Âkk
and B̂kk in (34) and (35) are constant matrices independent
of the parameters α and β, so ρ1 and ρ2 in (36) are constants;
therefore, according to condition (39), one can get that the
larger of parameters α and β, the network can realize the
fixed-time cluster synchronization more quickly.
Remark 10: Expect for the protocol (22) and (23), another
two possible ways to realize fixed-time cluster synchronization
are replacing the terms
X
X
sigp
aij (xj (t) − xi (t))
k′ 6=k
j∈Ck′
and
X
sig
k′ 6=k
q
X
j∈Ck′
bij (xj (t) − xi (t))
in (22) and (23) by terms
X X
sigp
aij (xj (t) − xi (t))
+β
N
X
j=1
bij sigq xj (t) − xi (t) ,
(49)
where xi ∈ Rn , α > 0, β > 0, 0 < p < 1, q > 1, and the
target trajectory s(t) satisfies that:
ṡ(t) = f (s(t)).
(50)
Similarly, we define A = (aij )N ×N as: for i, j = 1, · · · , N ,
2
a 1+p
i 6= j
ij ;
(51)
aij =
2
1+p
− PN
a
; i=j
k=1,k6=i
ik
We also define B = (bij )N ×N as: for i, j = 1, · · · , N ,
2
b 1+q ;
i 6= j
ij
bij =
2
P
N
1+q
−
k=1,k6=i bik ; i = j
(52)
Moreover, define
2
 = −2A + diag{(2ǫ1 α−1 ) 1+p , 0, · · · , 0},
k′ 6=k j∈Ck′
B̂ = −2B + diag{(2ǫ2 β
−1
)
2
1+q
, 0, · · · , 0}.
(53)
(54)
and
sig
q
X X
k′ 6=k j∈Ck′
bij (xj (t) − xi (t)) ,
or by linearly coupling terms as
X X
aij (xj (t) − xi (t))
Since matrices  and B̂ are positive definite, we can use
λmin (Â) and λmin (B̂) to denote the smallest eigenvalue of
 and B̂, respectively. Let N = n( N (N2−1) + 1), and
α = α2
p−1
2
(λmin (Â))
1+p
2
, β = βN
1−q
2
2
q−1
2
(λmin (B̂))
1+q
2
.
(55)
k′ 6=k j∈Ck′
and
X X
bij (xj (t) − xi (t)).
k′ 6=k j∈Ck′
Obviously, the combinations of these protocols are also feasible. Since there are no great differences in proving their
validity for fixed-time cluster synchronization (for the linear
coupling, there have been many results), here we just present
these protocols, interested readers can complete the proofs.
B. Complete synchronization with m = 1
When m = 1, i.e., all the nodes lie in the same cluster, then
the cluster synchronization problem becomes the complete
synchronization with pinning control. In this case, the network
with N strongly connected nodes can be described as:
N
X
a1j sigp xj (t) − x1 (t)
ẋ1 (t) =f (x1 (t)) + α
Theorem 2: For coupled systems (48) and (49), if A ∈ A2,
α − 2δ > 0 and β − 2δ > 0, then the fixed-time complete
synchronization can be achieved with the settling time defined
as:
Tmax =
+β
b1j sigq
j=1
xj (t) − x1 (t)
− ǫ1 sigp (x1 (t) − s(t)) − ǫ2 sigq (x1 (t) − s(t)), (48)
ẋ1 (t)
=α
N
X
j=1
N
X
q
b1j sig xj − x1
a1j sig xj − x1 + β
p
j=1
− ǫ1 sigp (x1 (t) − s) − ǫ2 sigq (x1 (t) − s),
and for i = 2, · · · , N ,
and for i = 2, · · · , N ,
ẋi (t) =f (xi (t)) + α
(56)
where δ is as defined in (25), and positive parameters α, β are
defined in (55).
Since the proof process is the same with that in Theorem
1, here we omit it.
Remark 11: The network topology can be easily relaxed to
be detail-balanced [20], i.e., there exist some scalars ωi > 0,
such that ωi aij = ωj aji .
Corollary 1: (Complete consensus) For the following complex network,
j=1
N
X
2
2
,
+
(α − 2δ)(1 − p) (β − 2δ)(q − 1)
N
X
j=1
p
aij sig xj (t) − xi (t)
ẋi (t) =α
N
X
j=1
N
X
q
bij sig xj − xi ,
aij sig xj − xi + β
p
j=1
9
where xi ∈ R, α > 0, β > 0, 0 < p < 1, q > 1, and the
target trajectory s ∈ R can be any constant scalar, the fixedtime complete consensus can be realized, i.e., xi (t) = s, i =
1, · · · , N when
2
2
,
+
α(1 − p) β(q − 1)
t ≥ Tmax =
(57)
where positive parameters α, β are defined in (55).
When the number of network node N = 1, then the network
becomes the simplest master-slave coupled systems as:
p
q
ẋ(t) =f (x(t)) − ǫ1 sig (x(t) − s(t)) − ǫ2 sig (x(t) − s(t)),
(58)
where x(t) ∈ Rn , 0 < p < 1, q > 1, and the target trajectory
s(t) is defined by (50).
Theorem 3: For coupled systems (58), if α − 2δ > 0 and
β − 2δ > 0, then the fixed-time cluster synchronization can
be achieved with the settling time defined as:
Tmax =
2
2
,
+
(α − 2δ)(1 − p) (β − 2δ)(q − 1)
(59)
1+p
2
β = ǫ2 n
,
1−q
2
2
1+q
2
1
(W1 + W1T + εW2 W2T + ε−1 W3T W3 ) ≤ δI,
2
then under the following control
ẋ(t) =W1 x(t) + W2 Φ(x(t)) + J
− ǫ1 sigp (x(t) − x⋆ ) − ǫ2 sigq (x(t) − x⋆ ),
(63)
(64)
the fixed-time stability can be realized, and the settling time
is defined by (59).
Proof: Denote f (x) = W1 x(t)+ W2 φ(x(t))+ J, one just
needs to prove that the QUAD condition (25) holds. Denote
e(t) = x(t) − x⋆ , and Φ̃(e(t)) = Φ(x(t)) − Φ(x⋆ ), one can
get
(x(t) − x⋆ )T (f (x(t)) − f (x⋆ ))
=(x(t) − x⋆ )T [W1 (x(t) − x⋆ ) + W2 (Φ(x(t)) − Φ(x⋆ ))]
=e(t)T [W1 e(t) + W2 Φ̃(e(t))]
where δ is as defined in (25), and
α = ǫ1 2
need to control the system (62) to the desired state x⋆ , where
x⋆ is the equilibrium point of (62), W1 x⋆ +W2 Φ(x⋆ )+J = 0.
Corollary 2: For the nonlinear system (62), if there exist
two scalars δ > 0 and ε > 0, such that
.
(60)
Proof: Define the Lyapunov function
≤e(t)T W1 e(t) + εe(t)T W2 W2T e(t) + ε−1 Φ̃(e(t))T Φ̃(e(t))
1
≤ e(t)T W1 + W1T + εW2 W2T + ε−1 W3T W3 e(t)
2
≤δe(t)T e(t).
n
V (t) =
1X l 2
1
e(t)T e(t) =
e (t) ,
2
2
(61)
l=1
1
where e(t) = x(t) − s(t) = (e (t), · · · , en (t))T ∈ Rn .
Differentiating it, we have
V̇ (t)
=e(t)T [f˜(e(t)) − ǫ1 sigp (e(t)) − ǫ2 sigq (e(t))]
l=1
≤2δV (t) − ǫ1
l=1
1+p
l=1
l
2
e (t)
1+p
2
1+p
− ǫ2 n
1−q
2
X
n
l
1−q
=2δV (t) − ǫ1 2 2 V (t) 2 − ǫ2 n 2 2
(
1+p
V (t) < 1
−(α − 2δ)V (t) 2 ;
≤
1+q
2
−(β − 2δ)V (t) ;
V (t) ≥ 1
1+q
2
V (t)
2
e (t)
l=1
1+q
2
1+q
2
IV. N UMERICAL
According to Lemma 2, one can get that the fixed-time
synchronization is finally realized, and the settling time can
be also obtained as (59).
Next, we will apply the above theorem on the fixedtime stabilization of equilibrium point for nonlinear systems,
including neural networks.
Consider the nonlinear function described by:
ẋ(t) = W1 x(t) + W2 Φ(x(t)) + J
W1 + W1T
) + kW2 kkW3 k.
(65)
2
Remark 13: In [26], the stabilization of nonlinear systems
is also investigated, but they concern on finite-time stabilization, while in this paper we investigate the fixed-time
stabilizaPn
tion. Of course, one can also use V (t) = 21 i=1 ξi ei (t)2 , ξi >
0, i = 1, · · · , n to investigate the fixed-time stabilization of
the above Corollary, but in order to keep in accordance with
Theorem 3, here we don’t consider the parameters ξi .
δ ≥ λmax (
≤2δV (t) − ǫ1 |e(t)|1+p − ǫ2 |e(t)|1+q
n
n
X
X
1+p
1+q
=2δV (t) − ǫ1
(el (t)2 ) 2 − ǫ2
(el (t)2 ) 2
X
n
Then, according to Theorem 3, we can get the conclusion.
Remark 12: As for the existence of δ in condition (63), one
can first convert it to the matrix form, then use the Matlab
toolbox Linear Matrix Inequality (LMI) to solve it.
On the other hand, one can also estimate the term as:
e(t)T W2 Φ̃(e(t)) ≤ kW2 kkW3 ke(t)T e(t), in this case,
(62)
where x(t) ∈ Rn is the state vector, Φ(x) : Rn → Rn is
the nonlinear function satisfying kΦ(x) − Φ(y)k ≤ kW3 (x −
y)k, ∀x, y ∈ Rn , and J is the external disturbance vector. If we
EXAMPLES
In this section, a simple numerical example is given to
demonstrate the correctness of obtained theoretical results.
Consider a network of five agents, and the original dynamical behavior x(t) of each node is described by is a 3-D neural
network satisfying:
ẋ(t) = f (x(t)) = −x(t) + W Φ(x(t)),
(66)
where x(t) = (x1 (t), x2 (t), x3 (t))T , Φ(x(t)) = (φ(x1 (t)),
φ(x2 (t)), φ(x3 (t)))T , φ(v) = (|v + 1| − |v − 1|)/2, and
1.25 −3.2 −3.2
W = −3.2
1.1 −4.4 .
−3.2
4.4
1
10
+β
X
j=4,5
p
1.5
X
b3j xj (t)
b3j sigq xj (t) − x3 (t) + sigq
j=1,2
q
− αsig (x3 (t) − s2 (t)) − βsig (x1 (t) − s2 (t));
1
ẋi (t) = f (xi (t))
X
X
p
p
aij xj (t)
+α
aij sig xj (t) − xi (t) + sig
0.5
0
j=1,2
j∈C2 ,j6=i
−0.5
+β
−1
j∈C2 ,j6=i
−1.5
−2
−1
0
1
2
0
−1
−2
1
This neural network has a double-scrolling chaotic attractor,
see Fig. 1. For this function f (·), from (65), one can easily
get that δ = 6.1 can make the QUAD condition (25) hold.
Suppose the network is partitioned into two clusters as C1 =
{1, 2} and C2 = {3, 4, 5}. As for the coupling matrix, we
assume that
A11 A12
,
A=B=
A21 A22
where
−1
1
1 −1
A21 = AT12 ,
, A12 =
A22
X
q
q
bij xj (t)
bij sig xj (t) − xi (t) + sig
j=1,2
i = 4, 5
(67)
2
Fig. 1. Chaotic attractor of 3-D neural network (66) with initial values
(0.4, 0.1, −0.2)T .
A11 =
X
−0.1
0.3 −0.2
,
0.1 −0.3
0.2
−2
1
1
= 1 −2
1 .
1
1 −2
Assume the pinning controllers are added on the first node
of each cluster, i.e., only node 1 and 3 are pinned. In this case,
the network can be described by the following equations:
where s1 (t) and s2 (t) are two different trajectories described
by (66) with initial values s1 (0) = (0.4, 0.1, −0.2)T and
s2 (0) = (0.1, 0.1, 0.1)T , respectively.
Now, for p = 0.5 and q = 2, according to notations in (34)
and (35), one can get that
4
4.5198
2
3
Â11 = −2A11 + diag{2 , 0} =
,
2 −2
6.5198 −2 −2
4
−2
4 −2 ,
Â22 = −2A22 + diag{2 3 , 0, 0} =
−2 −2
4
2
3.5874
2
,
B̂11 = −2B11 + diag{2 3 , 0} =
2 −2
5.5874 −2 −2
2
−2
4 −2 ,
Â22 = −2A22 + diag{2 3 , 0, 0} =
−2 −2
4
so parameters in (36)-(38) are: ρ1 = 0.6395, ρ2 = 0.4445,
N = 18, and
α = 0.6013α, β = 0.0988β, γ1 = 5.4385, γ2 = 0.5091. (68)
According to Theorem 1, one can get that if α − γ1 − 2δ > 0
and β − γ2 − 2δ > 0, then α ≥ 29.3339 and β > 128.5617.
Define an index
v
u 2
5
X
uX
kx − s (t)k2
(69)
kx − s (t)k2 +
E(t) = t
i
i=1
ẋ1 (t) = f (x1 (t))
X
5
a1j xj (t)
+ α · a12 sigp x2 (t) − x1 (t) + sigp
j=3
X
5
q
q
b1j xj (t)
+ β · b12 sig x2 (t) − x1 (t) + sig
j=3
− αsigp (x1 (t) − s1 (t)) − βsigq (x1 (t) − s1 (t));
ẋ2 (t) = f (x2 (t))
X
5
a2j xj (t)
+ α · a21 sigp x1 (t) − x2 (t) + sigp
i
1
2
i=3
for cluster synchronization error. Let α = 30 and β = 130,
then one can get the settling time defined by (39) is 7.3956,
while the real settling time is about 0.1735, which is far less
than the theoretical value, see Fig. 2.
In fact, when parameters α and β are less than the calculated
theoretical values, the fixed-time cluster synchronization can
also be realized. In the following, we discuss the fixed-time
cluster synchronization under different values of parameters,
like α, β, p, q, see Fig. 3-Fig. 6. All these simulations can
coincide with our previous analysis about fixed-time cluster
synchronization.
j=3
X
5
q
q
b2j xj (t)
+ β · b21 sig x1 (t) − x2 (t) + sig
j=3
ẋ3 (t) = f (x3 (t))
X
X
p
p
a3j xj (t)
a3j sig xj (t) − x3 (t) + sig
+α
j=4,5
j=1,2
V. C ONCLUSION
In this paper, we study the fixed-time cluster synchronization problem, while previous works often concentrate on
fixed-time consensus problems without control or finite-time
complete synchronization/consensus problems. We first design
a new distributed protocol which can be used to realize the
11
4.5
7
p=0.5
p=0.1
p=0.001
4
6
3.5
5
3
2.5
E(t)
E(t)
4
2
3
1.5
2
1
0.5
1
0
0
0
0.1
0.2
0.3
0.4
0.5
Time t
0.6
0.7
0.8
0.9
0
0.2
0.4
0.6
0.8
1
Fig. 2. Fixed-time cluster synchronization for network (67) with the settling
time Tmax = 0.1375 for α = 30 and β = 130.
1
Time t
1.2
1.4
1.6
1.8
2
Fig. 5. Fixed-time cluster synchronization for network (67) with α = 5,
β = 10, q = 2 and p = 0.5, 0.1, 0.001 respectively. One can find that,
the smaller the p is, the faster the cluster synchronization can be realized.
Moreover, the times used for the network error from the initial values to 1
are almost the same, since the parameter q in these cases are the same.
7
α=1
α=5
α=10
α=15
α=20
α=25
α=30
6
5
4.5
q=1.5
q=2
q=2.5
4
3.5
4
E(t)
3
2.5
E(t)
3
2
2
1.5
1
1
0
0
0.2
0.4
0.6
0.8
1
Time t
1.2
1.4
1.6
1.8
2
0.5
0
Fig. 3. Fixed-time cluster synchronization for network (67) under β = 1
and α = 1, 5, 10, 15, 20, 25, 30 respectively. Obviously, the larger the α is,
the faster the cluster synchronization can be realized.
0
0.2
0.4
0.6
0.8
1
Time t
1.2
1.4
1.6
1.8
2
Fig. 6. Fixed-time cluster synchronization for network (67) with α = 5,
β = 10, p = 0.5 and q = 1.5, 2, 2.5 respectively. One can find that, the
settling times are almost the same, since the parameter p in these cases are
the same.
7
β=1
β=10
β=50
β=100
β=150
6
5
E(t)
4
3
2
1
0
0
0.2
0.4
0.6
0.8
1
Time t
1.2
1.4
1.6
1.8
2
Fig. 4. Fixed-time cluster synchronization for network (67) under α = 5
and β = 1, 10, 50, 100, 150 respectively. Obviously, the larger the β is, the
faster the cluster synchronization can be realized.
fixed-time cluster synchronization, then rigorous proofs are
given to show the validity of this new protocol. Moreover,
when all the nodes in the network lie in the same cluster,
it becomes the fixed-time complete synchronization problem,
which is also carefully discussed, because it can contain the
master-slave coupled case and the stability of the nonlinear
systems case, which is applied on the investigation of the
fixed-time stabilization of the equilibrium in neural networks.
Finally, some numerical simulations are presented to show the
correctness of our obtained theoretical results.
It is hoped that this paper may shed some light on the
study of fixed-time cluster synchronization via pinning control.
However, there are still many challenging problems to be investigated in the next step, for example: 1) the unified form for
general fixed-time cluster synchronization and finding its key
property; 2) the fixed-time cluster synchronization for directed
12
networks, which is more general in the real world, while in
this paper only undirected network topology is discussed; 3)
the fixed-time cluster synchronization for networks with time
delays; 4) the fixed-time cluster synchronization for networks
with adaptive coupling mechanisms.
R EFERENCES
[1] L. M. Pecora and T. L. Carroll, “Master stability functions for synchronized coupled systems,” Phys. Rev. Lett., vol. 80, no. 10, pp. 2109-2112,
1998.
[2] W. L. Lu and T. P. Chen, “New approach to synchronization analysis
of linearly coupled ordinary differential systems,” Phys. D, Nonlinear
Phenomena, vol. 213, no. 2, pp. 214-230, Jan. 2006.
[3] C. W. Wu and L. O. Chua, “Synchronization in an array of linearly
coupled dynamical systems,” IEEE Trans. Circuits Syst. I, Fundam.
Theory Appl., vol. 42, no. 8, pp. 430-447, Aug. 1995.
[4] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of
agents with switching topology and time-delays,” IEEE Trans. Autom.
Control, vol. 49, no. 9, pp. 1520-1533, Sep. 2004.
[5] T. P. Chen, X. W. Liu, and W. L. Lu, “Pinning complex networks by a
single controller,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 54, no.
6, pp. 1317-1326, Jun. 2007.
[6] T. P. Chen and Z. M. Zhu, “Exponential synchronization of nonlinear
coupled dynamical networks,” Internat. J. Bifur. Chaos, vol. 17, no. 3,
pp. 999-1005, Mar. 2007.
[7] W. Wu, W. J. Zhou, and T. P. Chen, “Cluster synchronization of linearly
coupled complex networks under pinning control,” IEEE Trans. Circuits
Syst. I, Reg. Papers, vol. 56, no. 4, pp. 829-839, Apr. 2009.
[8] X. W. Liu and T. P. Chen, “Cluster synchronization in directed networks
via intermittent pinning control,” IEEE Trans. Neural Netw., vol. 22, no.
7, pp. 1009-1020, Jul. 2011.
[9] Y. R. Liu, Z. D. Wang, J. L. Liang, and X. H. Liu, “Synchronization
of coupled neutral-type neural networks with jumping-mode-dependent
discrete and unbounded distributed delays,” IEEE Trans. Cybern., vol.
43, no. 1, pp. 102-114, Feb. 2013.
[10] Y. Tang, H. J. Gao, W. Zou, and J. Kurths, “Distributed synchronization
in networks of agent systems with nonlinearities and random switchings,”
IEEE Trans. Cybern., vol. 43, no. 1, pp. 358-370, Feb. 2013.
[11] X. W. Liu and T. P. Chen, “Synchronization of nonlinear coupled
networks via aperiodically intermittent pinning control,” IEEE Trans.
Neural Netw. Learn. Syst., vol. 26, no. 1, pp. 113-126, Jan. 2015.
[12] V. T. Haimo, “Finite time controllers,” SIAM J. Control Optim., vol. 24,
no. 6, pp. 760-770, Jul. 1986.
[13] J. Cortés, “Finite-time convergent gradient flows with applications to
network consensus,” Automatica, vol. 42, no. 11, pp. 1993-2000, Nov.
2006.
[14] S. P. Bhat and D. S. Bernstein, “Continuous finite-time stabilization of
the translational and rotational double integrators,” IEEE Trans. Autom.
Control, vol. 43, no. 5, pp. 678-682, May 1998.
[15] Y. G. Hong, Y. S. Xu and J. Huang, “Finite-time control for robot
manipulators,” Syst. Control Lett., vol. 46, no. 4, pp. 243-253, Jul. 2002.
[16] S. P. Bhat and D. S. Bernstein, “Finite-time stability of continuous
autonomous systems,” SIAM J. Control Optim., vol. 38, no. 3, pp. 751766, Mar. 2000.
[17] Y. J. Shen and X. H. Xia, “Semi-global finite-time observers for
nonlinear systems,” Automatica, vol. 44, no. 12, pp. 3152-3156, Dec.
2008.
[18] Y. J. Shen and Y. H. Huang, “Uniformly observable and globally
lipschitzian nonlinear systems admit global finite-time observers,” IEEE
Trans. Autom. Control, vol. 54, no. 11, pp. 2621-2625, Nov. 2009.
[19] F. Xiao, L. Wang, J. Chen, and Y. P. Gao, “Finite-time formation control
for multi-agent systems,” Automatica, vol. 45, no. 11, pp. 2605-2611, Nov.
2009.
[20] L. Wang and F. Xiao, “Finite-time consensus problem for network of
dynamic agents,” IEEE Trans. Autom. Control, vol. 55, no. 4, pp. 950-955,
Apr. 2010.
[21] F. C. Jiang and L. Wang, “Finite-time weighted average consensus with
respect to a monotonic function and its application,” Syst. Control Lett.,
vol. 60, no. 9, pp. 718-725, Sep. 2011.
[22] F. C. Jiang and L. Wang, “Finite-time information consensus for multiagent systems with fixed and switching topologies,” Physica D, vol. 238,
no. 16, pp. 1550-1560, Aug. 2009.
[23] X. L. Wang and Y. G. Hong, “Distributed finite-time χ-consensus
algorithms for multi-agent systems with variable coupling topology,” J.
Syst. Sci. Complex, vol. 23, no. 2, pp. 209-218, Apr. 2010.
[24] F. L. Sun, J. C. Chen, Z. H. Guan, L. Ding, and T. Li, “Leaderfollowing finite-time consensus for multi-agent systems with jointlyreachable leader,” Nonlinear Anal.-Real, vol. 13, no. 5, pp. 2271-2284,
Oct. 2012.
[25] Y. Zhao, Z. S. Duan, and G. H. Wen, “Finite-time consensus for secondorder multi-agent systems with saturated control protocols,” IET Control
Theory Appl., vol. 9, no. 3, pp. 312-319, Feb. 2015.
[26] X. Y. Liu, D. W. C. Ho, W. W. Yu, and J. D. Cao, “A new switching
design to finite-time stabilization of nonlinear systems with applications
to neural networks,” Neural Networks, vol. 57, pp. 94-102, Sep. 2014.
[27] X. Y. Liu, J. Lam, W. W. Yu, and G. Chen, “Finite-time consensus of
multiagent systems with a switching protocol,” IEEE Trans. Neural Netw.
Learn. Syst., DOI: 10.1109/TNNLS.2015.2425933.
[28] X. Y. Wang, S. H. Li, and P. Shi, “Distributed Finite-time containment
control for double-integrator multiagent systems,” IEEE Trans. Cybern.,
vol. 44, no. 9, pp. 1518-1528, Sep. 2014.
[29] A. Polyakov, “Nonlinear feedback design for fixed-time stabilization of
linear control systems,” IEEE Trans. Autom. Control, vol. 57, no. 8, pp.
2106-2110, Aug. 2012.
[30] S. Parsegov, A. Polyakov, and P. Shcherbakov, “Nonlinear fixed-time
control protocol for uniform allocation of agents on a segment,” in Proc.
51st IEEE Conf. on Decision and Control, Maui, Hawaii, USA, Dec.
2012, pp. 7732-7737.
[31] S. Parsegov, A. Polyakov, and P. Shcherbakov, “Fixed-time consensus
algorithm for multi-agent systems with integrator dynamics,” in Proc. 4th
IFAC Workshop Distrbuted Estimation and Control in Networked System,
Koblenz, Germany, Sep. 2013, pp. 110-115.
[32] A. Polyakov, D. Efimov, and W. Perruquetti, “Finite-time and fixed-time
stabilization: Implicit Lyapunov function approach,” Automatica, vol. 51,
pp. 332-340, Jan. 2015.
[33] Z. Y. Zuo and L. Tie, “A new class of finite-time nonlinear consensus
protocols for multi-agent systems,” Int. J. Control, vol. 87, no. 2, pp.
363-370, Feb. 2014.
[34] Z. Y. Zuo and L. Tie, “Distributed robust finite-time nonlinear
consensus protocols for multi-agent systems,” Int. J. Syst. Sci., doi:
10.1080/00207721.2014.925608.
[35] Z. Y. Zuo, “Non-singular fixed-time terminal sliding mode control of
non-linear systems,” IET Control Theory and Applications, vol. 9, no. 4,
pp. 545-552, Feb. 2015.
[36] D. Y. Meng, Y. M. Jia, and J. P. Du, “Finite-time consensus for
multiagent systems with cooperative and antagonistic interactions,” IEEE
Trans. Neural Netw. Learn. Syst., DOI: 10.1109/TNNLS.2015.2424225
[37] Y. Chen and J. H. Lv, “Finite time synchronization of complex dynamical
networks,” J. Sys. Sci. & Math. Scis., vol. 29, no. 10, pp. 1419-1430, Oct.
2009. (in Chinese)
[38] Y. J. Zhou and C. Y. Sun, “Fixed time synchronization of complex
dynamical networks,” in Proc. 2015 Chinese Intelligent Automation
Conference, Lecture Notes in Electrical Engineering, vol. 338, pp. 163170, Mar. 2015.
[39] W. L. Lu, X. W. Liu, and T. P. Chen, “Finite-time stability and
synchronization with finite-time,” arXiv:1507.07160v1, submitted.
[40] Z. J. Ma, Z. R. Liu, and G. Zhang, “A new method to realize cluster
synchronization in connected chaotic networks,” Chaos, vol. 16, no. 2, p.
023103, Jun. 2006.
[41] X. W. Liu and T. P. Chen, “Cluster synchronization for linearly coupled
complex networks,” J. Industrial Manag. Optim., vol. 7, no. 1, pp. 87-101,
Feb. 2011.
[42] W. L. Lu, B. Liu, and T. P. Chen, “Cluster synchronization in networks
of coupled nonidentical dynamical systems,” Chaos, vol. 20, no. 1, p.
013120, Mar. 2010.
[43] W. X. Cui, J. A. Fang, W. B. Zhang, and X. Wang, “Finite-time
cluster synchronisation of Markovian switching complex networks with
stochastic perturbations,” IET Control Theory Appl., vol. 8, no. 1, pp.
30-41, Jan. 2014.
[44] G. Hardy, J. Littlewood, and G. Polya, Inequalities, 2nd ed., Cambridge
University Press, Cambridge, 1952.
| 3cs.SY
|
Published as a conference paper at ICLR 2015
N EURAL M ACHINE T RANSLATION
BY J OINTLY L EARNING TO A LIGN AND T RANSLATE
Dzmitry Bahdanau
Jacobs University Bremen, Germany
arXiv:1409.0473v7 [cs.CL] 19 May 2016
KyungHyun Cho
Yoshua Bengio∗
Université de Montréal
A BSTRACT
Neural machine translation is a recently proposed approach to machine translation. Unlike the traditional statistical machine translation, the neural machine
translation aims at building a single neural network that can be jointly tuned to
maximize the translation performance. The models proposed recently for neural machine translation often belong to a family of encoder–decoders and encode
a source sentence into a fixed-length vector from which a decoder generates a
translation. In this paper, we conjecture that the use of a fixed-length vector is a
bottleneck in improving the performance of this basic encoder–decoder architecture, and propose to extend this by allowing a model to automatically (soft-)search
for parts of a source sentence that are relevant to predicting a target word, without
having to form these parts as a hard segment explicitly. With this new approach,
we achieve a translation performance comparable to the existing state-of-the-art
phrase-based system on the task of English-to-French translation. Furthermore,
qualitative analysis reveals that the (soft-)alignments found by the model agree
well with our intuition.
1
I NTRODUCTION
Neural machine translation is a newly emerging approach to machine translation, recently proposed
by Kalchbrenner and Blunsom (2013), Sutskever et al. (2014) and Cho et al. (2014b). Unlike the
traditional phrase-based translation system (see, e.g., Koehn et al., 2003) which consists of many
small sub-components that are tuned separately, neural machine translation attempts to build and
train a single, large neural network that reads a sentence and outputs a correct translation.
Most of the proposed neural machine translation models belong to a family of encoder–
decoders (Sutskever et al., 2014; Cho et al., 2014a), with an encoder and a decoder for each language, or involve a language-specific encoder applied to each sentence whose outputs are then compared (Hermann and Blunsom, 2014). An encoder neural network reads and encodes a source sentence into a fixed-length vector. A decoder then outputs a translation from the encoded vector. The
whole encoder–decoder system, which consists of the encoder and the decoder for a language pair,
is jointly trained to maximize the probability of a correct translation given a source sentence.
A potential issue with this encoder–decoder approach is that a neural network needs to be able to
compress all the necessary information of a source sentence into a fixed-length vector. This may
make it difficult for the neural network to cope with long sentences, especially those that are longer
than the sentences in the training corpus. Cho et al. (2014b) showed that indeed the performance of
a basic encoder–decoder deteriorates rapidly as the length of an input sentence increases.
In order to address this issue, we introduce an extension to the encoder–decoder model which learns
to align and translate jointly. Each time the proposed model generates a word in a translation, it
(soft-)searches for a set of positions in a source sentence where the most relevant information is
concentrated. The model then predicts a target word based on the context vectors associated with
these source positions and all the previous generated target words.
∗
CIFAR Senior Fellow
1
Published as a conference paper at ICLR 2015
The most important distinguishing feature of this approach from the basic encoder–decoder is that
it does not attempt to encode a whole input sentence into a single fixed-length vector. Instead, it encodes the input sentence into a sequence of vectors and chooses a subset of these vectors adaptively
while decoding the translation. This frees a neural translation model from having to squash all the
information of a source sentence, regardless of its length, into a fixed-length vector. We show this
allows a model to cope better with long sentences.
In this paper, we show that the proposed approach of jointly learning to align and translate achieves
significantly improved translation performance over the basic encoder–decoder approach. The improvement is more apparent with longer sentences, but can be observed with sentences of any
length. On the task of English-to-French translation, the proposed approach achieves, with a single
model, a translation performance comparable, or close, to the conventional phrase-based system.
Furthermore, qualitative analysis reveals that the proposed model finds a linguistically plausible
(soft-)alignment between a source sentence and the corresponding target sentence.
2
BACKGROUND : N EURAL M ACHINE T RANSLATION
From a probabilistic perspective, translation is equivalent to finding a target sentence y that maximizes the conditional probability of y given a source sentence x, i.e., arg maxy p(y | x). In
neural machine translation, we fit a parameterized model to maximize the conditional probability
of sentence pairs using a parallel training corpus. Once the conditional distribution is learned by a
translation model, given a source sentence a corresponding translation can be generated by searching
for the sentence that maximizes the conditional probability.
Recently, a number of papers have proposed the use of neural networks to directly learn this conditional distribution (see, e.g., Kalchbrenner and Blunsom, 2013; Cho et al., 2014a; Sutskever et al.,
2014; Cho et al., 2014b; Forcada and Ñeco, 1997). This neural machine translation approach typically consists of two components, the first of which encodes a source sentence x and the second
decodes to a target sentence y. For instance, two recurrent neural networks (RNN) were used by
(Cho et al., 2014a) and (Sutskever et al., 2014) to encode a variable-length source sentence into a
fixed-length vector and to decode the vector into a variable-length target sentence.
Despite being a quite new approach, neural machine translation has already shown promising results.
Sutskever et al. (2014) reported that the neural machine translation based on RNNs with long shortterm memory (LSTM) units achieves close to the state-of-the-art performance of the conventional
phrase-based machine translation system on an English-to-French translation task.1 Adding neural
components to existing translation systems, for instance, to score the phrase pairs in the phrase
table (Cho et al., 2014a) or to re-rank candidate translations (Sutskever et al., 2014), has allowed to
surpass the previous state-of-the-art performance level.
2.1
RNN E NCODER –D ECODER
Here, we describe briefly the underlying framework, called RNN Encoder–Decoder, proposed by
Cho et al. (2014a) and Sutskever et al. (2014) upon which we build a novel architecture that learns
to align and translate simultaneously.
In the Encoder–Decoder framework, an encoder reads the input sentence, a sequence of vectors
x = (x1 , · · · , xTx ), into a vector c.2 The most common approach is to use an RNN such that
ht = f (xt , ht−1 )
(1)
and
c = q ({h1 , · · · , hTx }) ,
n
where ht ∈ R is a hidden state at time t, and c is a vector generated from the sequence of the
hidden states. f and q are some nonlinear functions. Sutskever et al. (2014) used an LSTM as f and
q ({h1 , · · · , hT }) = hT , for instance.
1
We mean by the state-of-the-art performance, the performance of the conventional phrase-based system
without using any neural network-based component.
2
Although most of the previous works (see, e.g., Cho et al., 2014a; Sutskever et al., 2014; Kalchbrenner and
Blunsom, 2013) used to encode a variable-length input sentence into a fixed-length vector, it is not necessary,
and even it may be beneficial to have a variable-length vector, as we will show later.
2
Published as a conference paper at ICLR 2015
The decoder is often trained to predict the next word yt0 given the context vector c and all the
previously predicted words {y1 , · · · , yt0 −1 }. In other words, the decoder defines a probability over
the translation y by decomposing the joint probability into the ordered conditionals:
p(y) =
T
Y
p(yt | {y1 , · · · , yt−1 } , c),
(2)
t=1
where y = y1 , · · · , yTy . With an RNN, each conditional probability is modeled as
p(yt | {y1 , · · · , yt−1 } , c) = g(yt−1 , st , c),
(3)
where g is a nonlinear, potentially multi-layered, function that outputs the probability of yt , and st is
the hidden state of the RNN. It should be noted that other architectures such as a hybrid of an RNN
and a de-convolutional neural network can be used (Kalchbrenner and Blunsom, 2013).
3
L EARNING TO A LIGN AND T RANSLATE
In this section, we propose a novel architecture for neural machine translation. The new architecture
consists of a bidirectional RNN as an encoder (Sec. 3.2) and a decoder that emulates searching
through a source sentence during decoding a translation (Sec. 3.1).
3.1
D ECODER : G ENERAL D ESCRIPTION
In a new model architecture, we define each conditional probability
in Eq. (2) as:
yt-1
yt
p(yi |y1 , . . . , yi−1 , x) = g(yi−1 , si , ci ),
s t-1
st
(4)
where si is an RNN hidden state for time i, computed by
+
si = f (si−1 , yi−1 , ci ).
It should be noted that unlike the existing encoder–decoder approach (see Eq. (2)), here the probability is conditioned on a distinct
context vector ci for each target word yi .
The context vector ci depends on a sequence of annotations
(h1 , · · · , hTx ) to which an encoder maps the input sentence. Each
annotation hi contains information about the whole input sequence
with a strong focus on the parts surrounding the i-th word of the
input sequence. We explain in detail how the annotations are computed in the next section.
The context vector ci is, then, computed as a weighted sum of these
annotations hi :
ci =
Tx
X
αij hj .
(5)
αt,1
αt,2
αt,T
αt,3
h1
h2
h3
hT
h1
h2
h3
hT
x1
x2
x3
xT
Figure 1: The graphical illustration of the proposed model
trying to generate the t-th target word yt given a source
sentence (x1 , x2 , . . . , xT ).
j=1
The weight αij of each annotation hj is computed by
exp (eij )
,
αij = PTx
k=1 exp (eik )
(6)
where
eij = a(si−1 , hj )
is an alignment model which scores how well the inputs around position j and the output at position
i match. The score is based on the RNN hidden state si−1 (just before emitting yi , Eq. (4)) and the
j-th annotation hj of the input sentence.
We parametrize the alignment model a as a feedforward neural network which is jointly trained with
all the other components of the proposed system. Note that unlike in traditional machine translation,
3
Published as a conference paper at ICLR 2015
the alignment is not considered to be a latent variable. Instead, the alignment model directly computes a soft alignment, which allows the gradient of the cost function to be backpropagated through.
This gradient can be used to train the alignment model as well as the whole translation model jointly.
We can understand the approach of taking a weighted sum of all the annotations as computing an
expected annotation, where the expectation is over possible alignments. Let αij be a probability that
the target word yi is aligned to, or translated from, a source word xj . Then, the i-th context vector
ci is the expected annotation over all the annotations with probabilities αij .
The probability αij , or its associated energy eij , reflects the importance of the annotation hj with
respect to the previous hidden state si−1 in deciding the next state si and generating yi . Intuitively,
this implements a mechanism of attention in the decoder. The decoder decides parts of the source
sentence to pay attention to. By letting the decoder have an attention mechanism, we relieve the
encoder from the burden of having to encode all information in the source sentence into a fixedlength vector. With this new approach the information can be spread throughout the sequence of
annotations, which can be selectively retrieved by the decoder accordingly.
3.2
E NCODER : B IDIRECTIONAL RNN FOR A NNOTATING S EQUENCES
The usual RNN, described in Eq. (1), reads an input sequence x in order starting from the first
symbol x1 to the last one xTx . However, in the proposed scheme, we would like the annotation
of each word to summarize not only the preceding words, but also the following words. Hence,
we propose to use a bidirectional RNN (BiRNN, Schuster and Paliwal, 1997), which has been
successfully used recently in speech recognition (see, e.g., Graves et al., 2013).
→
−
A BiRNN consists of forward and backward RNN’s. The forward RNN f reads the input sequence
→
−
→
−
as it is ordered (from x1 to xTx ) and calculates a sequence of forward hidden states ( h 1 , · · · , h Tx ).
←
−
The backward RNN f reads the sequence in the reverse order (from xTx to x1 ), resulting in a
←
−
←
−
sequence of backward hidden states ( h 1 , · · · , h Tx ).
→
−
We obtain an annotation for each word xj by concatenating the forward hidden state h j and the
h→
←
−
− ←
− > i>
backward one h j , i.e., hj = h >
. In this way, the annotation hj contains the summaries
j ; hj
of both the preceding words and the following words. Due to the tendency of RNNs to better
represent recent inputs, the annotation hj will be focused on the words around xj . This sequence
of annotations is used by the decoder and the alignment model later to compute the context vector
(Eqs. (5)–(6)).
See Fig. 1 for the graphical illustration of the proposed model.
4
E XPERIMENT S ETTINGS
We evaluate the proposed approach on the task of English-to-French translation. We use the bilingual, parallel corpora provided by ACL WMT ’14.3 As a comparison, we also report the performance of an RNN Encoder–Decoder which was proposed recently by Cho et al. (2014a). We use
the same training procedures and the same dataset for both models.4
4.1
DATASET
WMT ’14 contains the following English-French parallel corpora: Europarl (61M words), news
commentary (5.5M), UN (421M) and two crawled corpora of 90M and 272.5M words respectively,
totaling 850M words. Following the procedure described in Cho et al. (2014a), we reduce the size of
the combined corpus to have 348M words using the data selection method by Axelrod et al. (2011).5
We do not use any monolingual data other than the mentioned parallel corpora, although it may be
possible to use a much larger monolingual corpus to pretrain an encoder. We concatenate news-test3
http://www.statmt.org/wmt14/translation-task.html
Implementations are available at https://github.com/lisa-groundhog/GroundHog.
5
Available online at http://www-lium.univ-lemans.fr/˜schwenk/cslm_joint_paper/.
4
4
Published as a conference paper at ICLR 2015
30
BLEU score
25
Figure 2: The BLEU scores
of the generated translations
on the test set with respect
to the lengths of the sentences. The results are on
the full test set which includes sentences having unknown words to the models.
20
15
RNNsearch-50
RNNsearch-30
RNNenc-50
RNNenc-30
10
5
0
0
10
20
30
40
50
60
Sentence length
2012 and news-test-2013 to make a development (validation) set, and evaluate the models on the test
set (news-test-2014) from WMT ’14, which consists of 3003 sentences not present in the training
data.
After a usual tokenization6 , we use a shortlist of 30,000 most frequent words in each language to
train our models. Any word not included in the shortlist is mapped to a special token ([UNK]). We
do not apply any other special preprocessing, such as lowercasing or stemming, to the data.
4.2
M ODELS
We train two types of models. The first one is an RNN Encoder–Decoder (RNNencdec, Cho et al.,
2014a), and the other is the proposed model, to which we refer as RNNsearch. We train each model
twice: first with the sentences of length up to 30 words (RNNencdec-30, RNNsearch-30) and then
with the sentences of length up to 50 word (RNNencdec-50, RNNsearch-50).
The encoder and decoder of the RNNencdec have 1000 hidden units each.7 The encoder of the
RNNsearch consists of forward and backward recurrent neural networks (RNN) each having 1000
hidden units. Its decoder has 1000 hidden units. In both cases, we use a multilayer network with a
single maxout (Goodfellow et al., 2013) hidden layer to compute the conditional probability of each
target word (Pascanu et al., 2014).
We use a minibatch stochastic gradient descent (SGD) algorithm together with Adadelta (Zeiler,
2012) to train each model. Each SGD update direction is computed using a minibatch of 80 sentences. We trained each model for approximately 5 days.
Once a model is trained, we use a beam search to find a translation that approximately maximizes the
conditional probability (see, e.g., Graves, 2012; Boulanger-Lewandowski et al., 2013). Sutskever
et al. (2014) used this approach to generate translations from their neural machine translation model.
For more details on the architectures of the models and training procedure used in the experiments,
see Appendices A and B.
5
R ESULTS
5.1
Q UANTITATIVE R ESULTS
In Table 1, we list the translation performances measured in BLEU score. It is clear from the table
that in all the cases, the proposed RNNsearch outperforms the conventional RNNencdec. More
importantly, the performance of the RNNsearch is as high as that of the conventional phrase-based
translation system (Moses), when only the sentences consisting of known words are considered.
This is a significant achievement, considering that Moses uses a separate monolingual corpus (418M
words) in addition to the parallel corpora we used to train the RNNsearch and RNNencdec.
6
7
We used the tokenization script from the open-source machine translation package, Moses.
In this paper, by a ’hidden unit’, we always mean the gated hidden unit (see Appendix A.1.1).
5
It
should
be
noted
that
the
marine
environment
is
the
least
known
of
environments
.
<end>
The
agreement
on
the
European
Economic
Area
was
signed
in
August
1992
.
<end>
Published as a conference paper at ICLR 2015
L'
accord
sur
la
zone
économique
européenne
a
été
signé
en
août
1992
.
<end>
Il
convient
de
noter
que
l'
environnement
marin
est
le
moins
connu
de
l'
environnement
.
<end>
(b)
"
This
will
change
my
future
with
my
family
,
"
the
man
said
.
<end>
Destruction
of
the
equipment
means
that
Syria
can
no
longer
produce
new
chemical
weapons
.
<end>
(a)
La
destruction
de
l'
équipement
signifie
que
la
Syrie
ne
peut
plus
produire
de
nouvelles
armes
chimiques
.
<end>
"
Cela
va
changer
mon
avenir
avec
ma
famille
"
,
a
dit
l'
homme
.
<end>
(c)
(d)
Figure 3: Four sample alignments found by RNNsearch-50. The x-axis and y-axis of each plot
correspond to the words in the source sentence (English) and the generated translation (French),
respectively. Each pixel shows the weight αij of the annotation of the j-th source word for the i-th
target word (see Eq. (6)), in grayscale (0: black, 1: white). (a) an arbitrary sentence. (b–d) three
randomly selected samples among the sentences without any unknown words and of length between
10 and 20 words from the test set.
One of the motivations behind the proposed approach was the use of a fixed-length context vector
in the basic encoder–decoder approach. We conjectured that this limitation may make the basic
encoder–decoder approach to underperform with long sentences. In Fig. 2, we see that the performance of RNNencdec dramatically drops as the length of the sentences increases. On the other hand,
both RNNsearch-30 and RNNsearch-50 are more robust to the length of the sentences. RNNsearch50, especially, shows no performance deterioration even with sentences of length 50 or more. This
superiority of the proposed model over the basic encoder–decoder is further confirmed by the fact
that the RNNsearch-30 even outperforms RNNencdec-50 (see Table 1).
6
Published as a conference paper at ICLR 2015
Model
RNNencdec-30
RNNsearch-30
RNNencdec-50
RNNsearch-50
RNNsearch-50?
Moses
5.2
5.2.1
All
13.93
21.50
17.82
26.75
28.45
33.30
No UNK◦
24.19
31.44
26.71
34.16
36.15
35.63
Table 1: BLEU scores of the trained models computed on the test set. The second and third columns
show respectively the scores on all the sentences and,
on the sentences without any unknown word in themselves and in the reference translations. Note that
RNNsearch-50? was trained much longer until the
performance on the development set stopped improving. (◦) We disallowed the models to generate [UNK]
tokens when only the sentences having no unknown
words were evaluated (last column).
Q UALITATIVE A NALYSIS
A LIGNMENT
The proposed approach provides an intuitive way to inspect the (soft-)alignment between the words
in a generated translation and those in a source sentence. This is done by visualizing the annotation
weights αij from Eq. (6), as in Fig. 3. Each row of a matrix in each plot indicates the weights
associated with the annotations. From this we see which positions in the source sentence were
considered more important when generating the target word.
We can see from the alignments in Fig. 3 that the alignment of words between English and French
is largely monotonic. We see strong weights along the diagonal of each matrix. However, we also
observe a number of non-trivial, non-monotonic alignments. Adjectives and nouns are typically
ordered differently between French and English, and we see an example in Fig. 3 (a). From this
figure, we see that the model correctly translates a phrase [European Economic Area] into [zone
économique européen]. The RNNsearch was able to correctly align [zone] with [Area], jumping
over the two words ([European] and [Economic]), and then looked one word back at a time to
complete the whole phrase [zone économique européenne].
The strength of the soft-alignment, opposed to a hard-alignment, is evident, for instance, from
Fig. 3 (d). Consider the source phrase [the man] which was translated into [l’ homme]. Any hard
alignment will map [the] to [l’] and [man] to [homme]. This is not helpful for translation, as one
must consider the word following [the] to determine whether it should be translated into [le], [la],
[les] or [l’]. Our soft-alignment solves this issue naturally by letting the model look at both [the] and
[man], and in this example, we see that the model was able to correctly translate [the] into [l’]. We
observe similar behaviors in all the presented cases in Fig. 3. An additional benefit of the soft alignment is that it naturally deals with source and target phrases of different lengths, without requiring a
counter-intuitive way of mapping some words to or from nowhere ([NULL]) (see, e.g., Chapters 4
and 5 of Koehn, 2010).
5.2.2
L ONG S ENTENCES
As clearly visible from Fig. 2 the proposed model (RNNsearch) is much better than the conventional
model (RNNencdec) at translating long sentences. This is likely due to the fact that the RNNsearch
does not require encoding a long sentence into a fixed-length vector perfectly, but only accurately
encoding the parts of the input sentence that surround a particular word.
As an example, consider this source sentence from the test set:
An admitting privilege is the right of a doctor to admit a patient to a hospital or
a medical centre to carry out a diagnosis or a procedure, based on his status as a
health care worker at a hospital.
The RNNencdec-50 translated this sentence into:
Un privilège d’admission est le droit d’un médecin de reconnaı̂tre un patient à
l’hôpital ou un centre médical d’un diagnostic ou de prendre un diagnostic en
fonction de son état de santé.
7
Published as a conference paper at ICLR 2015
The RNNencdec-50 correctly translated the source sentence until [a medical center]. However, from
there on (underlined), it deviated from the original meaning of the source sentence. For instance, it
replaced [based on his status as a health care worker at a hospital] in the source sentence with [en
fonction de son état de santé] (“based on his state of health”).
On the other hand, the RNNsearch-50 generated the following correct translation, preserving the
whole meaning of the input sentence without omitting any details:
Un privilège d’admission est le droit d’un médecin d’admettre un patient à un
hôpital ou un centre médical pour effectuer un diagnostic ou une procédure, selon
son statut de travailleur des soins de santé à l’hôpital.
Let us consider another sentence from the test set:
This kind of experience is part of Disney’s efforts to ”extend the lifetime of its
series and build new relationships with audiences via digital platforms that are
becoming ever more important,” he added.
The translation by the RNNencdec-50 is
Ce type d’expérience fait partie des initiatives du Disney pour ”prolonger la durée
de vie de ses nouvelles et de développer des liens avec les lecteurs numériques qui
deviennent plus complexes.
As with the previous example, the RNNencdec began deviating from the actual meaning of the
source sentence after generating approximately 30 words (see the underlined phrase). After that
point, the quality of the translation deteriorates, with basic mistakes such as the lack of a closing
quotation mark.
Again, the RNNsearch-50 was able to translate this long sentence correctly:
Ce genre d’expérience fait partie des efforts de Disney pour ”prolonger la durée
de vie de ses séries et créer de nouvelles relations avec des publics via des
plateformes numériques de plus en plus importantes”, a-t-il ajouté.
In conjunction with the quantitative results presented already, these qualitative observations confirm our hypotheses that the RNNsearch architecture enables far more reliable translation of long
sentences than the standard RNNencdec model.
In Appendix C, we provide a few more sample translations of long source sentences generated by
the RNNencdec-50, RNNsearch-50 and Google Translate along with the reference translations.
6
6.1
R ELATED W ORK
L EARNING TO A LIGN
A similar approach of aligning an output symbol with an input symbol was proposed recently by
Graves (2013) in the context of handwriting synthesis. Handwriting synthesis is a task where the
model is asked to generate handwriting of a given sequence of characters. In his work, he used a
mixture of Gaussian kernels to compute the weights of the annotations, where the location, width
and mixture coefficient of each kernel was predicted from an alignment model. More specifically,
his alignment was restricted to predict the location such that the location increases monotonically.
The main difference from our approach is that, in (Graves, 2013), the modes of the weights of the
annotations only move in one direction. In the context of machine translation, this is a severe limitation, as (long-distance) reordering is often needed to generate a grammatically correct translation
(for instance, English-to-German).
Our approach, on the other hand, requires computing the annotation weight of every word in the
source sentence for each word in the translation. This drawback is not severe with the task of
translation in which most of input and output sentences are only 15–40 words. However, this may
limit the applicability of the proposed scheme to other tasks.
8
Published as a conference paper at ICLR 2015
6.2
N EURAL N ETWORKS FOR M ACHINE T RANSLATION
Since Bengio et al. (2003) introduced a neural probabilistic language model which uses a neural network to model the conditional probability of a word given a fixed number of the preceding words,
neural networks have widely been used in machine translation. However, the role of neural networks has been largely limited to simply providing a single feature to an existing statistical machine
translation system or to re-rank a list of candidate translations provided by an existing system.
For instance, Schwenk (2012) proposed using a feedforward neural network to compute the score of
a pair of source and target phrases and to use the score as an additional feature in the phrase-based
statistical machine translation system. More recently, Kalchbrenner and Blunsom (2013) and Devlin
et al. (2014) reported the successful use of the neural networks as a sub-component of the existing
translation system. Traditionally, a neural network trained as a target-side language model has been
used to rescore or rerank a list of candidate translations (see, e.g., Schwenk et al., 2006).
Although the above approaches were shown to improve the translation performance over the stateof-the-art machine translation systems, we are more interested in a more ambitious objective of
designing a completely new translation system based on neural networks. The neural machine translation approach we consider in this paper is therefore a radical departure from these earlier works.
Rather than using a neural network as a part of the existing system, our model works on its own and
generates a translation from a source sentence directly.
7
C ONCLUSION
The conventional approach to neural machine translation, called an encoder–decoder approach, encodes a whole input sentence into a fixed-length vector from which a translation will be decoded.
We conjectured that the use of a fixed-length context vector is problematic for translating long sentences, based on a recent empirical study reported by Cho et al. (2014b) and Pouget-Abadie et al.
(2014).
In this paper, we proposed a novel architecture that addresses this issue. We extended the basic
encoder–decoder by letting a model (soft-)search for a set of input words, or their annotations computed by an encoder, when generating each target word. This frees the model from having to encode
a whole source sentence into a fixed-length vector, and also lets the model focus only on information
relevant to the generation of the next target word. This has a major positive impact on the ability
of the neural machine translation system to yield good results on longer sentences. Unlike with
the traditional machine translation systems, all of the pieces of the translation system, including
the alignment mechanism, are jointly trained towards a better log-probability of producing correct
translations.
We tested the proposed model, called RNNsearch, on the task of English-to-French translation. The
experiment revealed that the proposed RNNsearch outperforms the conventional encoder–decoder
model (RNNencdec) significantly, regardless of the sentence length and that it is much more robust to the length of a source sentence. From the qualitative analysis where we investigated the
(soft-)alignment generated by the RNNsearch, we were able to conclude that the model can correctly align each target word with the relevant words, or their annotations, in the source sentence as
it generated a correct translation.
Perhaps more importantly, the proposed approach achieved a translation performance comparable to
the existing phrase-based statistical machine translation. It is a striking result, considering that the
proposed architecture, or the whole family of neural machine translation, has only been proposed
as recently as this year. We believe the architecture proposed here is a promising step toward better
machine translation and a better understanding of natural languages in general.
One of challenges left for the future is to better handle unknown, or rare words. This will be required
for the model to be more widely used and to match the performance of current state-of-the-art
machine translation systems in all contexts.
9
Published as a conference paper at ICLR 2015
ACKNOWLEDGMENTS
The authors would like to thank the developers of Theano (Bergstra et al., 2010; Bastien et al.,
2012). We acknowledge the support of the following agencies for research funding and computing
support: NSERC, Calcul Québec, Compute Canada, the Canada Research Chairs and CIFAR. Bahdanau thanks the support from Planet Intelligent Systems GmbH. We also thank Felix Hill, Bart van
Merriénboer, Jean Pouget-Abadie, Coline Devin and Tae-Ho Kim.
R EFERENCES
Axelrod, A., He, X., and Gao, J. (2011). Domain adaptation via pseudo in-domain data selection.
In Proceedings of the ACL Conference on Empirical Methods in Natural Language Processing
(EMNLP), pages 355–362. Association for Computational Linguistics.
Bastien, F., Lamblin, P., Pascanu, R., Bergstra, J., Goodfellow, I. J., Bergeron, A., Bouchard, N.,
and Bengio, Y. (2012). Theano: new features and speed improvements. Deep Learning and
Unsupervised Feature Learning NIPS 2012 Workshop.
Bengio, Y., Simard, P., and Frasconi, P. (1994). Learning long-term dependencies with gradient
descent is difficult. IEEE Transactions on Neural Networks, 5(2), 157–166.
Bengio, Y., Ducharme, R., Vincent, P., and Janvin, C. (2003). A neural probabilistic language model.
J. Mach. Learn. Res., 3, 1137–1155.
Bergstra, J., Breuleux, O., Bastien, F., Lamblin, P., Pascanu, R., Desjardins, G., Turian, J., WardeFarley, D., and Bengio, Y. (2010). Theano: a CPU and GPU math expression compiler. In
Proceedings of the Python for Scientific Computing Conference (SciPy). Oral Presentation.
Boulanger-Lewandowski, N., Bengio, Y., and Vincent, P. (2013). Audio chord recognition with
recurrent neural networks. In ISMIR.
Cho, K., van Merrienboer, B., Gulcehre, C., Bougares, F., Schwenk, H., and Bengio, Y. (2014a).
Learning phrase representations using RNN encoder-decoder for statistical machine translation.
In Proceedings of the Empiricial Methods in Natural Language Processing (EMNLP 2014). to
appear.
Cho, K., van Merriënboer, B., Bahdanau, D., and Bengio, Y. (2014b). On the properties of neural
machine translation: Encoder–Decoder approaches. In Eighth Workshop on Syntax, Semantics
and Structure in Statistical Translation. to appear.
Devlin, J., Zbib, R., Huang, Z., Lamar, T., Schwartz, R., and Makhoul, J. (2014). Fast and robust
neural network joint models for statistical machine translation. In Association for Computational
Linguistics.
Forcada, M. L. and Ñeco, R. P. (1997). Recursive hetero-associative memories for translation. In
J. Mira, R. Moreno-Dı́az, and J. Cabestany, editors, Biological and Artificial Computation: From
Neuroscience to Technology, volume 1240 of Lecture Notes in Computer Science, pages 453–462.
Springer Berlin Heidelberg.
Goodfellow, I., Warde-Farley, D., Mirza, M., Courville, A., and Bengio, Y. (2013). Maxout networks. In Proceedings of The 30th International Conference on Machine Learning, pages 1319–
1327.
Graves, A. (2012). Sequence transduction with recurrent neural networks. In Proceedings of the
29th International Conference on Machine Learning (ICML 2012).
Graves, A. (2013). Generating sequences with recurrent neural networks. arXiv:1308.0850
[cs.NE].
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, pages 273–278.
10
Published as a conference paper at ICLR 2015
Hermann, K. and Blunsom, P. (2014). Multilingual distributed representations without word alignment. In Proceedings of the Second International Conference on Learning Representations (ICLR
2014).
Hochreiter, S. (1991). Untersuchungen zu dynamischen neuronalen Netzen. Diploma thesis, Institut
für Informatik, Lehrstuhl Prof. Brauer, Technische Universität München.
Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8),
1735–1780.
Kalchbrenner, N. and Blunsom, P. (2013). Recurrent continuous translation models. In Proceedings
of the ACL Conference on Empirical Methods in Natural Language Processing (EMNLP), pages
1700–1709. Association for Computational Linguistics.
Koehn, P. (2010). Statistical Machine Translation. Cambridge University Press, New York, NY,
USA.
Koehn, P., Och, F. J., and Marcu, D. (2003). Statistical phrase-based translation. In Proceedings
of the 2003 Conference of the North American Chapter of the Association for Computational
Linguistics on Human Language Technology - Volume 1, NAACL ’03, pages 48–54, Stroudsburg,
PA, USA. Association for Computational Linguistics.
Pascanu, R., Mikolov, T., and Bengio, Y. (2013a). On the difficulty of training recurrent neural
networks. In ICML’2013.
Pascanu, R., Mikolov, T., and Bengio, Y. (2013b). On the difficulty of training recurrent neural
networks. In Proceedings of the 30th International Conference on Machine Learning (ICML
2013).
Pascanu, R., Gulcehre, C., Cho, K., and Bengio, Y. (2014). How to construct deep recurrent neural
networks. In Proceedings of the Second International Conference on Learning Representations
(ICLR 2014).
Pouget-Abadie, J., Bahdanau, D., van Merriënboer, B., Cho, K., and Bengio, Y. (2014). Overcoming
the curse of sentence length for neural machine translation using automatic segmentation. In
Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation. to appear.
Schuster, M. and Paliwal, K. K. (1997). Bidirectional recurrent neural networks. Signal Processing,
IEEE Transactions on, 45(11), 2673–2681.
Schwenk, H. (2012). Continuous space translation models for phrase-based statistical machine
translation. In M. Kay and C. Boitet, editors, Proceedings of the 24th International Conference on
Computational Linguistics (COLIN), pages 1071–1080. Indian Institute of Technology Bombay.
Schwenk, H., Dchelotte, D., and Gauvain, J.-L. (2006). Continuous space language models for
statistical machine translation. In Proceedings of the COLING/ACL on Main conference poster
sessions, pages 723–730. Association for Computational Linguistics.
Sutskever, I., Vinyals, O., and Le, Q. (2014). Sequence to sequence learning with neural networks.
In Advances in Neural Information Processing Systems (NIPS 2014).
Zeiler, M. D. (2012).
[cs.LG].
ADADELTA: An adaptive learning rate method.
11
arXiv:1212.5701
Published as a conference paper at ICLR 2015
A
M ODEL A RCHITECTURE
A.1
A RCHITECTURAL C HOICES
The proposed scheme in Section 3 is a general framework where one can freely define, for instance,
the activation functions f of recurrent neural networks (RNN) and the alignment model a. Here, we
describe the choices we made for the experiments in this paper.
A.1.1
R ECURRENT N EURAL N ETWORK
For the activation function f of an RNN, we use the gated hidden unit recently proposed by Cho
et al. (2014a). The gated hidden unit is an alternative to the conventional simple units such as an
element-wise tanh. This gated unit is similar to a long short-term memory (LSTM) unit proposed
earlier by Hochreiter and Schmidhuber (1997), sharing with it the ability to better model and learn
long-term dependencies. This is made possible by having computation paths in the unfolded RNN
for which the product of derivatives is close to 1. These paths allow gradients to flow backward
easily without suffering too much from the vanishing effect (Hochreiter, 1991; Bengio et al., 1994;
Pascanu et al., 2013a). It is therefore possible to use LSTM units instead of the gated hidden unit
described here, as was done in a similar context by Sutskever et al. (2014).
The new state si of the RNN employing n gated hidden units8 is computed by
si = f (si−1 , yi−1 , ci ) = (1 − zi ) ◦ si−1 + zi ◦ s̃i ,
where ◦ is an element-wise multiplication, and zi is the output of the update gates (see below). The
proposed updated state s̃i is computed by
s̃i = tanh (W e(yi−1 ) + U [ri ◦ si−1 ] + Cci ) ,
where e(yi−1 ) ∈ Rm is an m-dimensional embedding of a word yi−1 , and ri is the output of the
reset gates (see below). When yi is represented as a 1-of-K vector, e(yi ) is simply a column of an
embedding matrix E ∈ Rm×K . Whenever possible, we omit bias terms to make the equations less
cluttered.
The update gates zi allow each hidden unit to maintain its previous activation, and the reset gates ri
control how much and what information from the previous state should be reset. We compute them
by
zi = σ (Wz e(yi−1 ) + Uz si−1 + Cz ci ) ,
ri = σ (Wr e(yi−1 ) + Ur si−1 + Cr ci ) ,
where σ (·) is a logistic sigmoid function.
At each step of the decoder, we compute the output probability (Eq. (4)) as a multi-layered function (Pascanu et al., 2014). We use a single hidden layer of maxout units (Goodfellow et al., 2013)
and normalize the output probabilities (one for each word) with a softmax function (see Eq. (6)).
A.1.2
A LIGNMENT M ODEL
The alignment model should be designed considering that the model needs to be evaluated Tx × Ty
times for each sentence pair of lengths Tx and Ty . In order to reduce computation, we use a singlelayer multilayer perceptron such that
a(si−1 , hj ) = va> tanh (Wa si−1 + Ua hj ) ,
where Wa ∈ Rn×n , Ua ∈ Rn×2n and va ∈ Rn are the weight matrices. Since Ua hj does not
depend on i, we can pre-compute it in advance to minimize the computational cost.
8
Here, we show the formula of the decoder. The same formula can be used in the encoder by simply
ignoring the context vector ci and the related terms.
12
Published as a conference paper at ICLR 2015
A.2
A.2.1
D ETAILED D ESCRIPTION OF THE M ODEL
E NCODER
In this section, we describe in detail the architecture of the proposed model (RNNsearch) used in the
experiments (see Sec. 4–5). From here on, we omit all bias terms in order to increase readability.
The model takes a source sentence of 1-of-K coded word vectors as input
x = (x1 , . . . , xTx ), xi ∈ RKx
and outputs a translated sentence of 1-of-K coded word vectors
y = (y1 , . . . , yTy ), yi ∈ RKy ,
where Kx and Ky are the vocabulary sizes of source and target languages, respectively. Tx and Ty
respectively denote the lengths of source and target sentences.
First, the forward states of the bidirectional recurrent neural network (BiRNN) are computed:
(
→
−
→
−
−
−
→
−
(1 − →
z i ) ◦ h i−1 + →
z i ◦ h i , if i > 0
hi =
0
, if i = 0
where
−
i
→
−
−
→
→
− h− →
h i = tanh W Exi + U →
r i ◦ h i−1
−
−
→
→
− →
→
−
z i =σ W z Exi + U z h i−1
−
−
→
→
− →
→
−
r i =σ W r Exi + U r h i−1 .
−
→ −
→ −
→
→
− →
− →
−
E ∈ Rm×Kx is the word embedding matrix. W , W z , W r ∈ Rn×m , U , U z , U r ∈ Rn×n are
weight matrices. m and n are the word embedding dimensionality and the number of hidden units,
respectively. σ(·) is as usual a logistic sigmoid function.
←
−
←
−
The backward states ( h 1 , · · · , h Tx ) are computed similarly. We share the word embedding matrix
E between the forward and backward RNNs, unlike the weight matrices.
We concatenate the forward and backward states to to obtain the annotations (h1 , h2 , · · · , hTx ),
where
" →
− #
hi
hi = ←
(7)
−
hi
A.2.2
D ECODER
The hidden state si of the decoder given the annotations from the encoder is computed by
si =(1 − zi ) ◦ si−1 + zi ◦ s̃i ,
where
s̃i = tanh (W Eyi−1 + U [ri ◦ si−1 ] + Cci )
zi =σ (Wz Eyi−1 + Uz si−1 + Cz ci )
ri =σ (Wr Eyi−1 + Ur si−1 + Cr ci )
E is the word embedding matrix for the target language. W, Wz , Wr ∈ Rn×m , U, Uz , Ur ∈ Rn×n ,
and C, Cz , Cr ∈ Rn×2n are weights. Again, m and n are the word embedding dimensionality
and the
of hidden units, respectively. The initial hidden state s0 is computed by s0 =
number
←
−
tanh Ws h 1 , where Ws ∈ Rn×n .
The context vector ci are recomputed at each step by the alignment model:
ci =
Tx
X
αij hj ,
j=1
13
Published as a conference paper at ICLR 2015
Model
RNNenc-30
RNNenc-50
RNNsearch-30
RNNsearch-50
RNNsearch-50?
Updates (×105 )
8.46
6.00
4.71
2.88
6.67
Epochs
6.4
4.5
3.6
2.2
5.0
Hours
109
108
113
111
252
GPU
TITAN BLACK
Quadro K-6000
TITAN BLACK
Quadro K-6000
Quadro K-6000
Train NLL
28.1
44.0
26.7
40.7
36.7
Dev. NLL
53.0
43.6
47.2
38.1
35.2
Table 2: Learning statistics and relevant information. Each update corresponds to updating the
parameters once using a single minibatch. One epoch is one pass through the training set. NLL is
the average conditional log-probabilities of the sentences in either the training set or the development
set. Note that the lengths of the sentences differ.
where
exp (eij )
αij = PTx
k=1 exp (eik )
eij =va> tanh (Wa si−1 + Ua hj ) ,
0
0
and hj is the j-th annotation in the source sentence (see Eq. (7)). va ∈ Rn , Wa ∈ Rn ×n and
0
Ua ∈ Rn ×2n are weight matrices. Note that the model becomes RNN Encoder–Decoder (Cho
→
−
et al., 2014a), if we fix ci to h Tx .
With the decoder state si−1 , the context ci and the last generated word yi−1 , we define the probability
of a target word yi as
p(yi |si , yi−1 , ci ) ∝ exp yi> Wo ti ,
where
>
ti = max t̃i,2j−1 , t̃i,2j j=1,...,l
and t̃i,k is the k-th element of a vector t̃i which is computed by
t̃i =Uo si−1 + Vo Eyi−1 + Co ci .
Wo ∈ RKy ×l , Uo ∈ R2l×n , Vo ∈ R2l×m and Co ∈ R2l×2n are weight matrices. This can be understood as having a deep output (Pascanu et al., 2014) with a single maxout hidden layer (Goodfellow
et al., 2013).
A.2.3
M ODEL S IZE
For all the models used in this paper, the size of a hidden layer n is 1000, the word embedding
dimensionality m is 620 and the size of the maxout hidden layer in the deep output l is 500. The
number of hidden units in the alignment model n0 is 1000.
B
B.1
T RAINING P ROCEDURE
PARAMETER I NITIALIZATION
←
− ←
− ←
− →
− →
−
→
−
We initialized the recurrent weight matrices U, Uz , Ur , U , U z , U r , U , U z and U r as random orthogonal matrices. For Wa and Ua , we initialized them by sampling each element from the Gaussian
distribution of mean 0 and variance 0.0012 . All the elements of Va and all the bias vectors were initialized to zero. Any other weight matrix was initialized by sampling from the Gaussian distribution
of mean 0 and variance 0.012 .
B.2
T RAINING
We used the stochastic gradient descent (SGD) algorithm. Adadelta (Zeiler, 2012) was used to
automatically adapt the learning rate of each parameter ( = 10−6 and ρ = 0.95). We explicitly
14
Published as a conference paper at ICLR 2015
normalized the L2 -norm of the gradient of the cost function each time to be at most a predefined
threshold of 1, when the norm was larger than the threshold (Pascanu et al., 2013b). Each SGD
update direction was computed with a minibatch of 80 sentences.
At each update our implementation requires time proportional to the length of the longest sentence in
a minibatch. Hence, to minimize the waste of computation, before every 20-th update, we retrieved
1600 sentence pairs, sorted them according to the lengths and split them into 20 minibatches. The
training data was shuffled once before training and was traversed sequentially in this manner.
In Tables 2 we present the statistics related to training all the models used in the experiments.
C
T RANSLATIONS OF L ONG S ENTENCES
Source
Reference
RNNenc-50
RNNsearch-50
Google
Translate
Source
Reference
RNNenc-50
RNNsearch-50
Google
Translate
Source
Reference
RNNenc-50
RNNsearch-50
Google
Translate
An admitting privilege is the right of a doctor to admit a patient to a hospital or a medical centre
to carry out a diagnosis or a procedure, based on his status as a health care worker at a hospital.
Le privilège d’admission est le droit d’un médecin, en vertu de son statut de membre soignant
d’un hôpital, d’admettre un patient dans un hôpital ou un centre médical afin d’y délivrer un
diagnostic ou un traitement.
Un privilège d’admission est le droit d’un médecin de reconnaı̂tre un patient à l’hôpital ou un
centre médical d’un diagnostic ou de prendre un diagnostic en fonction de son état de santé.
Un privilège d’admission est le droit d’un médecin d’admettre un patient à un hôpital ou un
centre médical pour effectuer un diagnostic ou une procédure, selon son statut de travailleur des
soins de santé à l’hôpital.
Un privilège admettre est le droit d’un médecin d’admettre un patient dans un hôpital ou un
centre médical pour effectuer un diagnostic ou une procédure, fondée sur sa situation en tant
que travailleur de soins de santé dans un hôpital.
This kind of experience is part of Disney’s efforts to ”extend the lifetime of its series and build
new relationships with audiences via digital platforms that are becoming ever more important,”
he added.
Ce type d’expérience entre dans le cadre des efforts de Disney pour ”étendre la durée de
vie de ses séries et construire de nouvelles relations avec son public grâce à des plateformes
numériques qui sont de plus en plus importantes”, a-t-il ajouté.
Ce type d’expérience fait partie des initiatives du Disney pour ”prolonger la durée de vie de
ses nouvelles et de développer des liens avec les lecteurs numériques qui deviennent plus complexes.
Ce genre d’expérience fait partie des efforts de Disney pour ”prolonger la durée de vie de ses
séries et créer de nouvelles relations avec des publics via des plateformes numériques de plus
en plus importantes”, a-t-il ajouté.
Ce genre d’expérience fait partie des efforts de Disney à “étendre la durée de vie de sa série et
construire de nouvelles relations avec le public par le biais des plates-formes numériques qui
deviennent de plus en plus important”, at-il ajouté.
In a press conference on Thursday, Mr Blair stated that there was nothing in this video that might
constitute a ”reasonable motive” that could lead to criminal charges being brought against the
mayor.
En conférence de presse, jeudi, M. Blair a affirmé qu’il n’y avait rien dans cette vidéo qui puisse
constituer des ”motifs raisonnables” pouvant mener au dépôt d’une accusation criminelle contre
le maire.
Lors de la conférence de presse de jeudi, M. Blair a dit qu’il n’y avait rien dans cette vidéo qui
pourrait constituer une ”motivation raisonnable” pouvant entraı̂ner des accusations criminelles
portées contre le maire.
Lors d’une conférence de presse jeudi, M. Blair a déclaré qu’il n’y avait rien dans cette vidéo qui
pourrait constituer un ”motif raisonnable” qui pourrait conduire à des accusations criminelles
contre le maire.
Lors d’une conférence de presse jeudi, M. Blair a déclaré qu’il n’y avait rien dans cette vido
qui pourrait constituer un ”motif raisonnable” qui pourrait mener à des accusations criminelles
portes contre le maire.
Table 3: The translations generated by RNNenc-50 and RNNsearch-50 from long source sentences
(30 words or more) selected from the test set. For each source sentence, we also show the goldstandard translation. The translations by Google Translate were made on 27 August 2014.
15
| 9cs.NE
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.